Method and apparatus for generating robot motion adjustment instructions based on faults

Abstract

This application discloses a method and apparatus for generating robot motion adjustment commands based on faults. The method includes: acquiring the robot's current joint motion data, current joint torque measurements, and current multidimensional physical field data at the current moment; determining a set of predicted current joint torque values ​​based on the current joint motion data; determining the robot's current fault characteristics based on the predicted current joint torque values ​​and the current joint torque measurements; determining a semantic description of the robot's fault based on the current multidimensional physical field data, the current fault characteristics, and a target knowledge graph; and determining a sequence of joint drive commands for the robot over a predetermined time period, where the predetermined time period is the time period following the current moment. This application solves the technical problem in related technologies where the determination of joint drive command sequences for robots during faults is inaccurate.

Description

Technical Field

[0001] This application relates to the field of robotics technology, and more specifically, to a method and apparatus for generating fault-based robot motion adjustment instructions. Background Technology

[0002] For fault warning of quadruped robots in complex environments such as cable tunnels, related technologies mainly rely on a single motor model to extract fault features and determine abnormal states by comparing the deviation between the model's predicted values ​​and the measured values. Building on this, some related technologies attempt to introduce physical field data collected by multiple sensors for fusion processing and apply preset fault-tolerant control strategies to adjust motion based on the identified fault types.

[0003] The above-mentioned related technologies have the following limitations: (1) The prediction accuracy of a single motor model is easily affected by environmental changes and model errors, making it difficult to guarantee the accuracy of fault feature extraction and unable to effectively distinguish between real faults and environmental interference; (2) The fusion of physical field data collected by multiple sensors mostly stays at the numerical or feature level, lacking deep association of different physical field data at the fault semantic level, making it difficult to accurately identify fault types and locations; (3) The fault warning result is only used to trigger the preset fault-specific control strategy, and the control effect of the generated motion adjustment command cannot be fed back to optimize the fault warning process, resulting in the entire fault warning and motion adjustment command generation process lacking adaptive optimization capability.

[0004] The intertwining and coupling of the above problems leads to the technical problem of inaccurate determination of joint drive command sequence in robots during malfunctions.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides a method and apparatus for generating robot motion adjustment commands based on faults, so as to at least solve the technical problem in the related art that the determination result of joint drive command sequence of robot in the event of a fault is inaccurate.

[0007] According to one aspect of the embodiments of this application, a method for generating robot motion adjustment commands based on faults is provided, comprising: acquiring current joint motion data, current joint torque measurement values, and current multidimensional physical field data of the robot at the current moment, wherein the physical field data includes at least acoustic signature data, vibration data, electromagnetic field data, and energy consumption data; determining a set of predicted current joint torque values ​​of the robot based on the current joint motion data; determining current fault characteristics of the robot based on the set of predicted current joint torque values ​​and the current joint torque measurement values; determining a fault semantic description of the robot based on the current multidimensional physical field data, the current fault characteristics, and a target knowledge graph, wherein the target knowledge graph includes fault mode nodes, intermediate state nodes, physical field feature nodes, and the relationships between nodes, wherein intermediate state nodes represent the internal mechanism degradation or transitional damage of the robot during the fault evolution process, and physical field feature nodes represent abnormal phenomena that can be directly mapped according to the current multidimensional physical field data; and determining a sequence of joint drive commands for the robot in a predetermined time period based on the fault semantic description, wherein the predetermined time period is the time period after the current moment.

[0008] According to another aspect of the embodiments of this application, a fault-based robot motion adjustment command generation device is provided, comprising: a data acquisition module, configured to acquire current joint motion data, current joint torque measurement values, and current multidimensional physical field data of the robot at the current moment, wherein the physical field data includes at least acoustic signature data, vibration data, electromagnetic field data, and energy consumption data; a first determination module, configured to determine a set of predicted current joint torque values ​​of the robot based on the current joint motion data; a second determination module, configured to determine the current fault characteristics of the robot based on the set of predicted current joint torque values ​​and the current joint torque measurement values; a third determination module, configured to determine a fault semantic description of the robot based on the current multidimensional physical field data, the current fault characteristics, and a target knowledge graph, wherein the target knowledge graph includes fault mode nodes, intermediate state nodes, physical field feature nodes, and the relationships between nodes, wherein intermediate state nodes characterize the internal mechanism degradation or transitional damage of the robot during the fault evolution process, and physical field feature nodes characterize abnormal phenomena that can be directly mapped according to the current multidimensional physical field data; and a fourth determination module, configured to determine a sequence of joint drive commands for the robot during a predetermined time period based on the fault semantic description, wherein the predetermined time period is the time period after the current moment.

[0009] According to another aspect of the embodiments of this application, a non-volatile storage medium is provided, which stores multiple instructions, the instructions being adapted for a fault-based robot motion adjustment instruction generation method to be loaded by a processor and any one of them executed.

[0010] According to another aspect of the embodiments of this application, an electronic device is provided, including: one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any one of the fault-based robot motion adjustment instruction generation methods.

[0011] According to another aspect of the embodiments of this application, a computer program product is provided, which, when executed on a data processing device, is adapted to perform the steps of a fault-based robot motion adjustment instruction generation method.

[0012] In this embodiment, the robot acquires its current joint motion data, current joint torque measurement value, and current multidimensional physical field data at the current moment. The physical field data includes at least acoustic signature data, vibration data, electromagnetic field data, and energy consumption data. Based on the current joint motion data, a set of predicted current joint torque values ​​is determined. Based on the predicted current joint torque value set and the current joint torque measurement value, the robot's current fault characteristics are determined. Based on the current multidimensional physical field data, the current fault characteristics, and a target knowledge graph, a semantic description of the robot's fault is determined. The target knowledge graph includes fault mode nodes, intermediate state nodes, physical field feature nodes, and the relationships between nodes. Intermediate state nodes represent the robot's internal mechanism degradation or transitional damage during the fault evolution process, and physical field feature nodes represent abnormal phenomena that can be directly mapped from the current multidimensional physical field data. Based on the semantic description of the fault, a sequence of joint drive commands for the robot within a predetermined time period is determined, where the predetermined time period is the time period after the current moment. The goal is to determine the current fault characteristics based on the robot's current joint motion data and current joint torque measurements, combine them with the target knowledge graph and current multidimensional physical field data, determine the robot's fault semantic description, and determine the robot's joint drive command sequence based on the fault semantic description. This achieves the technical effect of improving the accuracy of the joint drive command sequence determination result when the robot is in faulty, thereby solving the technical problem of inaccurate joint drive command sequence determination result when the robot is in faulty in related technologies. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0014] Figure 1 This is a flowchart of a fault-based robot motion adjustment instruction generation method according to an embodiment of this application;

[0015] Figure 2This is a flowchart of an optional fault-based robot motion adjustment instruction generation method provided according to an embodiment of this application;

[0016] Figure 3 This is a flowchart of an optional motion parameter adjustment method provided according to an embodiment of this application;

[0017] Figure 4 This is a schematic diagram of a fault-based robot motion adjustment instruction generation device according to an embodiment of this application;

[0018] Figure 5 This is a structural diagram of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] It should be noted that the information (including but not limited to target knowledge graphs) and data (including but not limited to current joint motion data, current joint torque measurements, and current multidimensional physical field data) collected in this application are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, and necessary confidentiality measures have been taken. This process does not violate public order and good morals, and corresponding operation entry points are provided for users to choose whether to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding operation entry points for them to choose whether to agree to or refuse the automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0022] For ease of description, the following explains some of the nouns or terms used in the embodiments of this application:

[0023] Artificial intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain better results.

[0024] A quadruped robot is a legged mobile robot designed based on biomimetic principles. It mimics the limb structure and movement of quadrupeds such as dogs and felines, using four legs as its locomotion mechanism. As an important branch of legged robots, quadruped robots have stronger load-bearing capacity and movement stability compared to bipedal robots, and are structurally simpler and more efficiently controlled compared to hexapods and octagonals. Its core technical architecture typically includes a joint drive system, a multimodal environmental perception system, and motion control algorithms. It drives leg movement through joint motors, hydraulic or pneumatic components, and uses gyroscopes, accelerometers, and visual sensors to perceive its own posture and external environment. Based on gait planning and balance control algorithms, it achieves various movement modes such as walking, trotting, and jumping. Quadruped robots can stably traverse unstructured complex terrains such as stairs, gravel, sand, and steep slopes, exhibiting outstanding terrain adaptability and obstacle-crossing capabilities. They are widely used in industrial inspection, emergency rescue, and scientific research and education.

[0025] According to an embodiment of this application, a method embodiment for generating robot motion adjustment instructions based on faults is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0026] Figure 1This is a flowchart of a fault-based robot motion adjustment instruction generation method according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0027] Step S102: Obtain the robot's current joint motion data, current joint torque measurement value, and current multidimensional physical field data at the current moment. The physical field data includes at least voiceprint data, vibration data, electromagnetic field data, and energy consumption data.

[0028] As can be understood, current joint motion data refers to the real-time motion state information of each joint of the robot during actual operation, collected at the current moment, such as the current joint angle, current joint angular velocity, current joint angular acceleration, and current joint current. This data is typically collected by sensors installed at the joints to reflect the robot's current posture and dynamic behavior. Current multidimensional physical field data encompasses environmental and robot-specific state information collected from different physical dimensions at the current moment. Specifically, it includes: acoustic signature data, vibration data, electromagnetic field data, and energy consumption data. Acoustic signature data is used to capture the sound characteristics of the robot during operation, such as motor noise and friction sounds; vibration data is used to monitor the vibration patterns of the robot's structure or components, such as abnormal shaking; electromagnetic field data is used to detect abnormal electromagnetic radiation from motors or circuits; and energy consumption data is used to assess the energy consumption of the robot's motors or the entire robot system, such as abnormal power fluctuations.

[0029] Optionally, current joint motion data can be acquired by installing simple angle and current sensors at each joint of the quadruped robot. For example, joint angles can be measured using potentiometers or Hall effect sensors, and motor current can be measured using current transformers. Various physical field sensing data can be collected by deploying independent microphones, single-axis accelerometers, simple magnetic field detectors, and ammeters near the quadruped robot's body and joints to collect acoustic signature data, vibration data, electromagnetic field data, and energy consumption data, respectively. The acquisition of the long-term cross-calibration history of multiple heterogeneous models can be achieved by manually recording the predicted and measured joint torque values ​​of each heterogeneous motor model after each fault-free operation of the quadruped robot, and calculating their average deviation and standard deviation as the historical fluctuation pattern of the historical deviation values ​​of each heterogeneous motor model. Simultaneously, the statistical distribution of the average absolute error between the predicted joint torque values ​​of each heterogeneous motor model can be recorded as the predictive consistency distribution characteristics of each heterogeneous motor model within a historical fault-free period (e.g., the first historical time period).

[0030] Step S104: Based on the current joint motion data, determine the current joint torque prediction value set of the robot;

[0031] It is understandable that, based on the current joint motion data, multiple heterogeneous motor models are used to obtain a set of predicted joint torque values ​​for the robot. By using multiple heterogeneous motor models to predict the robot's joint torque, the limitations of a single heterogeneous motor model being susceptible to environmental interference or its own errors can be overcome, thus improving the comprehensiveness and robustness of joint torque prediction.

[0032] In one optional embodiment, determining the robot's current joint torque prediction set based on current joint motion data includes: determining the robot's current joint angle, current joint angular velocity, and current joint current based on the current joint motion data; obtaining a first joint torque prediction value using a first heterogeneous motor model based on the current joint angle, current joint angular velocity, and current joint current, wherein the first heterogeneous motor model is a physical mechanism model constructed based on motor dynamics equations; obtaining a second joint torque prediction value using a second heterogeneous motor model based on the current joint angle, current joint angular velocity, and current joint current, wherein the second heterogeneous motor model has pre-learned the correlation between the current joint angle, current joint angular velocity, and current joint current and the second joint torque prediction value; obtaining a third joint torque prediction value using a third heterogeneous motor model based on the current joint current, wherein the third heterogeneous motor model is a pre-constructed lookup table model characterizing the mapping relationship between joint current and joint torque prediction values; and obtaining the current joint torque prediction value set based on the first joint torque prediction value, the second joint torque prediction value, and the third joint torque prediction value.

[0033] It can be understood that the current joint angle, current joint angular velocity, and current joint current are extracted from the current joint motion data. These three parameters are then input into a first heterogeneous motor model and a second heterogeneous motor model, respectively, to obtain the first joint torque prediction value output by the first heterogeneous motor model and the second joint torque prediction value output by the second heterogeneous motor model. Based on the current joint current, a third heterogeneous motor model is used to obtain the third joint torque prediction value. These three joint torque prediction values ​​constitute the robot's current joint torque prediction value set. Specifically, the first heterogeneous motor model is a physical mechanism model built based on motor dynamics equations; the second heterogeneous motor model pre-learns the correlation between the current joint angle, current joint angular velocity, and current joint current and the second joint torque prediction value, for example, through a temporal learning model built based on a long short-term memory network; and the third heterogeneous motor model is a pre-built lookup table model representing the mapping relationship between the joint current and the joint torque prediction values. By using the first heterogeneous motor model, the second heterogeneous motor model, and the third heterogeneous motor model for parallel prediction, we can simultaneously consider the physical rationality of the prediction results, the ability to capture complex dynamic evolution laws, and the real-time response characteristics with extremely low latency, thus obtaining a set of current joint torque prediction values ​​that combines high confidence, strong adaptability, and high real-time performance.

[0034] Step S106: Based on the current set of predicted joint torque values ​​and the current measured joint torque values, determine the current fault characteristics of the robot;

[0035] It is understandable that fault characteristics refer to quantitative indicators that can characterize the robot's potential faults or abnormal states, extracted by comprehensively analyzing the current joint torque prediction value set and the current joint torque measurement value.

[0036] In one optional embodiment, determining the robot's current fault characteristics based on the current joint torque prediction set and the current joint torque measurement value includes: acquiring the robot's historical prediction consistency distribution characteristics and historical fluctuation patterns within a first historical time period, wherein the robot is fault-free within the first historical time period, the historical environmental conditions corresponding to the first historical time period are the same as the current environmental conditions corresponding to the current moment, the historical prediction consistency distribution characteristics characterize the statistical distribution and statistical characteristics of the dispersion of the historical joint torque prediction set within the first historical time period, and the historical fluctuation patterns are obtained based on statistical analysis of the historical deviation value set within the first historical time period, the historical deviation value set being based on the historical joint torque prediction set within the first historical time period and the historical fluctuation patterns within the first historical time period. The following steps are taken: First, the historical joint torque measurements are used to determine the current set of predicted joint torque values. Second, based on the current set of predicted joint torque values ​​and the current measured joint torque values, a current set of deviation values ​​is determined. Third, based on the current set of deviation values ​​and historical fluctuation patterns, a deviation feature set is determined. Fourth, based on the current set of predicted joint torque values, a current prediction consistency coefficient is determined, whereby the current prediction consistency coefficient characterizes the consistency between the predicted torque values ​​of the first, second, and third joints in the current set of predicted joint torque values. Fifth, based on the current prediction consistency coefficient and historical prediction consistency distribution characteristics, a consistency deviation feature is determined, whereby the consistency deviation feature characterizes the degree of deviation of the current prediction consistency coefficient from the historical prediction consistency distribution characteristics. Finally, based on the deviation feature set and the consistency deviation feature, the current fault characteristics are determined.

[0037] It is understandable that obtaining the historical prediction consistency distribution characteristics and historical fluctuation patterns of the robot in the first historical time period is crucial. The historical environmental conditions corresponding to the first historical time period are the same as the current environmental conditions corresponding to the current moment. The historical prediction consistency distribution characteristics represent the statistical distribution and statistical characteristics of the dispersion of the historical joint torque prediction value set in the first historical time period. The dispersion can be represented by standard deviation, variance, coefficient of variation (the ratio of standard deviation to mean), or mean absolute error. The statistical distribution can be represented by probability density function, cumulative distribution function, etc. The historical fluctuation patterns are obtained based on the statistical analysis of the historical deviation value set in the first historical time period, such as statistical quantities like mean, standard deviation, variance, skewness, kurtosis, etc., as well as the probability density function of the fitted deviation. The deviation is calculated based on the current set of predicted joint torque values ​​and the current measured joint torque values ​​to determine the current deviation value set. This current deviation value set includes a first deviation value, a second deviation value, and a third deviation value. The first deviation value is obtained from the predicted first joint torque value and the current measured joint torque value; the second deviation value is obtained from the predicted second joint torque value and the current measured joint torque value; and the third deviation value is obtained from the predicted third joint torque value and the current measured joint torque value. Based on the current deviation value set and historical fluctuation patterns, a deviation feature set is obtained. The consistency between the predicted first, second, and third joint torque values ​​is determined based on the current set of predicted joint torque values, resulting in a current prediction consistency coefficient. Based on the current prediction consistency coefficient and historical prediction consistency distribution characteristics, the degree of deviation of the current prediction consistency coefficient from the historical prediction consistency distribution characteristics is determined, resulting in a consistency deviation feature. The deviation feature set and the consistency deviation feature are concatenated, and the concatenated features are then subjected to dimensionality reduction and normalization to obtain the robot's current fault characteristics. The method of determining the current deviation value set is adopted. Based on the historical joint torque prediction value set and the historical joint torque measurement value set of the first historical time period, the historical deviation value set is determined, including the first historical deviation value sets corresponding to multiple first historical moments included in the first historical time period. By introducing historical prediction consistency distribution characteristics and historical fluctuation patterns that match the current environmental conditions, interference caused by environmental changes can be eliminated, effectively distinguishing between real faults and inherent model errors and environmental noise, thereby generating current fault characteristics with high robustness and environmental adaptability.

[0038] Optionally, the historical prediction consistency distribution characteristics and historical fluctuation patterns can be obtained using the following methods: Encoders and current sensors deployed at each joint of the quadruped robot are used to collect real-time joint angle data, current joint angular velocity data, and current joint current data as real-time joint motion data. Acoustic fingerprint data, triaxial vibration data (i.e., vibration data), electromagnetic field strength data (i.e., electromagnetic field data), and instantaneous power and current waveforms of each motor (i.e., energy consumption data) are simultaneously collected by an array of acoustic fingerprint sensors, an array of accelerometers, an array of electromagnetic field sensors, and an energy consumption monitoring module deployed near the quadruped robot's body and each joint, respectively, as multi-physical field sensing data. Historical joint torque prediction value sequences, historical joint torque measurement value sequences, and corresponding historical environmental condition labels for each heterogeneous motor model are retrieved from local storage or a cloud database during multiple fault-free runs of the quadruped robot within its historical operating cycle. Statistical analysis is performed on the historical deviation value sequences between the historical joint torque prediction value sequences and the historical joint torque measurement value sequences of each heterogeneous motor model to obtain the historical fluctuation patterns (i.e., historical fluctuation patterns) of the historical deviation values ​​of each heterogeneous motor model under various historical environmental conditions. Further, the dispersion analysis of the historical joint torque prediction values ​​of multiple heterogeneous motor models at the same moment in the historical operating cycle was carried out to obtain the historical prediction consistency distribution characteristics of each heterogeneous motor model under various historical environmental conditions.

[0039] Optionally, statistical analysis of the historical deviation sequence between the historical joint torque prediction value sequence and the historical joint torque measurement value sequence in historical fault-free data can establish the "normal" deviation range and distribution characteristics of each heterogeneous motor model under specific environmental conditions. This fluctuation pattern can distinguish between the inherent prediction error of each heterogeneous motor model, normal fluctuations caused by environmental factors, and abnormal deviations caused by actual faults, thereby improving the sensitivity and accuracy of fault early warning. This can be achieved by calculating the mean, standard deviation, variance, skewness, kurtosis, and other statistical measures of its historical deviation values ​​(historical joint torque measurement value minus the corresponding historical joint torque prediction value) for each heterogeneous motor model and each environmental condition. Furthermore, the probability density function of the deviation (such as a Gaussian distribution, Laplace distribution, etc.) can be fitted to obtain the historical fluctuation pattern. Another approach is to use non-parametric statistical methods, such as kernel density estimation (KDE) or empirical cumulative distribution function (ECDF), to directly learn the distribution characteristics from the historical deviation value data, avoiding pre-defined assumptions about the data distribution. At the same time, these statistics can be dynamically updated using methods such as sliding windows or exponentially weighted moving averages to adapt to slow changes in environmental conditions.

[0040] Optionally, when multiple heterogeneous motor models predict the same physical quantity, the difference in their predicted values ​​(dispersion) can reflect the collaborative working state of the multiple heterogeneous motor models and the degree of consensus on system behavior. Under normal circumstances, although these models are heterogeneous, their prediction results should remain consistent within a certain range. By analyzing this historical consistency distribution characteristic, a "normal" consistency benchmark can be established. When the real-time prediction consistency coefficient deviates from this benchmark, it may indicate that an anomaly has occurred in a certain model or system. This can be achieved by: calculating the standard deviation, variance, coefficient of variation (the ratio of standard deviation to mean), or mean absolute error, etc., of the historical joint torque prediction values ​​of multiple heterogeneous motor models at the same time to measure the dispersion. Then, statistical analysis is performed on the dispersion indices at different times to obtain the corresponding statistical distribution. Based on this statistical distribution and its statistical characteristics, the historical prediction consistency distribution characteristics are obtained. Another approach is to use information theory metrics such as information entropy, Kullback-Leibler divergence, or Jensen-Shannon divergence to quantify the differences in the distribution of historical joint torque predictions among multiple heterogeneous motor models, thereby obtaining a more robust measure of dispersion. Then, the distributions of these information theory metrics are modeled to form a consistent distribution characteristic of historical predictions.

[0041] Optionally, the historical deviation values, historical fluctuation patterns, and historical prediction consistency distribution characteristics matching the current environmental conditions can be retrieved from the historical data record set using the following method: Based on real-time joint motion data and synchronously collected ambient temperature, humidity, and tunnel section markers, the current environmental conditions of the quadruped robot are identified. In the historical data record set, the long-term mutual calibration history of the multi-heterogeneous motor model matching the current environmental conditions is retrieved. The historical fluctuation patterns of the historical deviation values ​​corresponding to the current environmental conditions are extracted from the long-term mutual calibration history of the multi-heterogeneous motor model. The historical prediction consistency distribution characteristics corresponding to the current environmental conditions are extracted from the long-term mutual calibration history of the multi-heterogeneous motor model.

[0042] In one optional embodiment, a deviation feature set is determined based on the current deviation value set and historical fluctuation patterns, including: determining the historical deviation mean and historical deviation standard deviation of the historical deviation value set based on historical fluctuation patterns; determining the current difference set based on the current deviation value set and the historical deviation mean; determining the deviation multiple set based on the current difference set and the historical deviation standard deviation; determining the dynamic confidence set based on the deviation multiple set; and determining the deviation feature set based on the dynamic confidence set and the current deviation value set.

[0043] It can be understood that the historical deviation mean and historical deviation standard deviation of the historical deviation value set for the first historical period are extracted from the historical fluctuation pattern. The differences between the first, second, and third deviation values ​​in the current deviation value set and the historical deviation mean are determined respectively, resulting in the current difference set. The current difference set includes the first difference, the second difference, and the third difference. The first difference is the difference between the first deviation value and the historical deviation mean; the second difference is the difference between the second deviation value and the historical deviation mean; and the third difference is the difference between the third deviation value and the historical deviation mean. The deviation multiples are calculated by comparing the first, second, and third differences in the current set of differences with the historical standard deviations. This yields a set of deviation multiples, specifically a first, second, and third deviation multiple. The first, second, and third deviation multiples are calculated by comparing the first difference with the historical standard deviation, respectively. Based on these deviation multiples, a dynamic confidence set is determined. This set includes a first, second, and third dynamic confidence level. The first, second, and third dynamic confidence levels are determined based on the first, second, and third deviation multiples. The dynamic confidence level decreases monotonically as the deviation multiple increases. Based on the dynamic confidence set and the current deviation value set, a deviation feature set is determined. This deviation feature set includes a first deviation feature, a second deviation feature, and a third deviation feature. The first deviation feature is obtained by multiplying the first dynamic confidence level and the first deviation value; the second deviation feature is obtained by multiplying the second dynamic confidence level and the second deviation value; and the third deviation feature is obtained by multiplying the third dynamic confidence level and the third deviation value. By using the historical deviation mean and historical deviation standard deviation to determine the deviation multiple set, and generating a dynamic confidence set accordingly, adaptive weighted calibration of the current deviation value set can effectively suppress the interference of instantaneous noise and abnormal fluctuations on the fault features, improving the robustness and realism of the current fault features in complex environments.

[0044] In one optional embodiment, determining the current prediction consistency coefficient based on the current set of joint torque prediction values ​​includes: determining a historical consensus center value, wherein the historical consensus center value is determined based on the historical set of joint torque prediction values ​​for a second historical time period; determining a deviation distance set based on the current set of joint torque prediction values ​​and the historical consensus center value, wherein the deviation distance set includes a first deviation distance between a first joint torque prediction value and a historical consensus center value, a second deviation distance between a second joint torque prediction value and a historical consensus center value, and a third deviation distance between a third joint torque prediction value and a historical consensus center value; and determining the current prediction consistency coefficient based on the deviation distance set.

[0045] It is understandable that the historical consensus center value is determined based on the historical joint torque prediction value set of the second historical time period. Based on the current joint torque prediction value set and the historical consensus center value, a deviation distance set is determined, which includes a first deviation distance, a second deviation distance, and a third deviation distance. The robot is fault-free during the second historical time period, and the historical environmental conditions corresponding to the second historical time period and the current environmental conditions corresponding to the current moment, as well as the robot's historical motion phase corresponding to the second historical time period and the robot's current motion phase corresponding to the current moment, are highly matched. The first deviation distance, the second deviation distance, and the third deviation distance are weighted and fused to obtain the current prediction consistency coefficient. By introducing the historical consensus center value as a reference benchmark, quantifying the deviation of the joint torque prediction values ​​of each heterogeneous motor model from the historical normal state, and performing weighted fusion, the interference of environmental condition changes on the consistency assessment between models can be eliminated, generating a highly robust current prediction consistency coefficient.

[0046] Optionally, the historical consensus center value can be determined as follows: Obtain the current joint angle and current joint angular velocity from real-time joint motion data (i.e., current joint motion data), and determine the current motion phase of the quadruped robot based on the current joint angle and current joint angular velocity. Retrieve a historical time window (second historical time period) that matches the current environmental conditions and the current motion phase. Extract the historical joint torque prediction value sequences output by the first heterogeneous motor model, the second heterogeneous motor model, and the third heterogeneous motor model during the historical fault-free period from the historical time window. Calculate the median of the historical joint torque prediction values ​​of the first heterogeneous motor model, the second heterogeneous motor model, and the third heterogeneous motor model at the same moment to obtain the instantaneous median value corresponding to each moment. Then, perform a moving average filter on the instantaneous median value corresponding to each moment to obtain the historical consensus center value corresponding to the current motion phase.

[0047] In one optional embodiment, the consistency deviation feature is determined based on the current prediction consistency coefficient and the historical prediction consistency distribution characteristics, including: constructing a baseline membership function based on the historical prediction consistency distribution characteristics; determining the joint angular velocity fluctuation amplitude based on the current joint angle in the current joint motion data and the historical joint angle in the third historical time period; correcting the baseline membership function based on the joint angular velocity fluctuation amplitude to obtain a target membership function; obtaining an initial membership degree based on the current prediction consistency coefficient and the target membership function; and correcting the initial membership degree based on the deviation feature set to obtain the consistency deviation feature.

[0048] It can be understood that a baseline membership function is constructed using the historical prediction consistency distribution characteristics as the benchmark form of the membership function. The joint angular velocity fluctuation amplitude is determined based on the current joint angle in the current joint motion data and the historical joint angles in the third historical time period. The third historical time period includes multiple second historical moments corresponding to the first historical joint angles. The baseline membership function is corrected using the joint angular velocity fluctuation amplitude, and the confidence interval boundaries of the baseline membership function are dynamically adjusted to obtain the target membership function. The current prediction consistency coefficient is input into the target membership function for mapping, resulting in an initial membership degree characterizing the degree to which the current prediction consistency state belongs to the historical fault-free mode. The overall deviation direction is determined based on the positive and negative signs corresponding to the first, second, and third deviation features in the deviation feature set, and the initial membership degree is corrected using the overall deviation direction to obtain the consistency deviation feature. By dynamically adjusting the confidence interval boundary of the benchmark membership function in conjunction with the fluctuation amplitude of joint angular velocity to adapt to real-time operating conditions, and by introducing the overall deviation direction to correct the initial membership, it is possible to comprehensively capture the deviation of the consistency state and the deviation trend, and generate a consistency deviation feature that can accurately reflect the consistency distribution characteristics of the current prediction consistency coefficient relative to the historical prediction consistency distribution characteristics.

[0049] Optionally, the consistency deviation feature can be determined as follows: Based on the historical prediction consistency distribution characteristics, a historical consistency probability density function matching the current environmental conditions is determined as the baseline form of the target membership function, thus obtaining the baseline membership function. Based on real-time joint motion data, the joint angular velocity fluctuation amplitude is determined, and the confidence interval boundary of the baseline membership function is dynamically adjusted to generate an adaptive membership function (i.e., the target membership function) associated with the current motion state. The real-time prediction consistency coefficient is input into the adaptive membership function for mapping to obtain the initial membership degree of the current prediction consistency state belonging to the historical normal pattern. The positive and negative signs of each weighted deviation component in the multi-source deviation feature are obtained, and the initial membership degree is corrected according to the overall deviation direction indicated by the positive and negative signs to generate the target membership degree, which is then used as the consistency deviation feature.

[0050] Optionally, the multi-source deviation features and consistency deviation features can be concatenated as follows, and the concatenated features can be dimensionality reduced and normalized to obtain the current fault features. The multi-source deviation features and consistency deviation features are concatenated along their feature dimensions to generate an initial concatenated feature vector. The variance contribution rate of each dimension in the initial concatenated feature vector is calculated, and the principal component dimensions are selected from the initial concatenated feature vector based on the variance contribution rate to obtain a dimensionality-reduced compressed feature vector. The dynamic confidence scores of each weighted deviation component in the multi-source deviation features are obtained, and the corresponding dimensions in the compressed feature vector are adaptively weighted according to the target confidence score. The adaptively weighted compressed feature vector is mapped to a unified numerical range to obtain the current fault features.

[0051] Step S108: Based on the current multidimensional physical field data, current fault features and target knowledge graph, determine the robot's fault semantic description. The target knowledge graph includes fault mode nodes, intermediate state nodes, physical field feature nodes and the relationships between nodes. Intermediate state nodes represent the robot's internal mechanism degradation or transitional damage during the fault evolution process. Physical field feature nodes represent abnormal phenomena that can be directly mapped based on the current multidimensional physical field data.

[0052] It is understandable that by constructing a target knowledge graph containing intermediate state nodes, it is possible to reveal the evolution path of a fault from its appearance to its essence, thereby generating a fault semantic description that is both interpretable and insightful into its underlying mechanisms, and achieving accurate identification of fault types and locations.

[0053] Optionally, the target knowledge graph is a structured knowledge representation used to store and manage knowledge related to robot faults. It consists of different types of nodes and the relationships connecting these nodes. Nodes include fault mode nodes, intermediate state nodes, and physical field feature nodes. Fault mode nodes represent known fault types, such as motor overload or joint jamming; intermediate state nodes characterize the robot's internal mechanism degradation or transitional damage during fault evolution, such as "accelerated bearing wear," "partial discharge in windings," "poor lubrication," and "motor overheating"; physical field feature nodes represent features extracted from physical field data such as acoustic signature data, vibration data, electromagnetic field data, and energy consumption data, such as abnormal noise frequencies and specific vibration modes. Relationships define the semantic connections between these nodes; for example, a certain physical field feature may lead to a certain fault mode.

[0054] Optionally, fault semantic description refers to a high-level, interpretable vector representation of fault information generated by deep semantic fusion of current fault features and current multidimensional physical field data with historical fault features. It is not merely a collection of numerical values ​​or features, but rather a clear description of the fault type, possible location, evolution trend, and its performance in different physical fields, thus providing richer semantic information for determining subsequent joint drive command sequences.

[0055] In one optional embodiment, the robot's fault semantic description is determined based on current multidimensional physical field data, current fault features, and a target knowledge graph. This includes: determining a timing guidance signal based on current fault features and historical fault features from a fourth historical time period, wherein the timing guidance signal characterizes the evolution direction and rate of the fault features; extracting features from the current multidimensional physical field data to obtain an initial multiphysics synchronization feature sequence; determining a fault evolution path based on the timing guidance signal and the target knowledge graph, wherein the fault evolution path starts from a physical field feature node and ends at a fault mode node; reordering and aligning the initial multiphysics synchronization feature sequence based on the node association order of the fault evolution path to obtain a target multiphysics synchronization feature sequence; and determining the fault semantic description based on the target multiphysics synchronization feature sequence.

[0056] It is understandable that, based on the current fault characteristics and the historical fault characteristics of the fourth historical time period, a time-series guidance signal characterizing the evolution direction and rate of the fault characteristics is determined. The fourth historical time period includes multiple first historical fault characteristics corresponding to third historical moments. Feature extraction is performed on the current multi-dimensional physical field data to obtain an initial multi-physics synchronization feature sequence, which includes the physical field synchronization features corresponding to the current multi-dimensional physical field data. Based on the time-series guidance signal and the target knowledge graph, a fault evolution path, including multiple intermediate state nodes, is determined from the physical field feature node to the fault mode node. Based on the node association order of the fault evolution path (i.e., the connection order between nodes, representing the causal logical relationship and temporal evolution order between nodes), the initial multi-physics synchronization feature sequence is reordered and phase-aligned to obtain the target multi-physics synchronization feature sequence. The reordering and phase alignment of the initial multi-physics synchronization feature sequence refers to rearranging and aligning the physical field synchronization features corresponding to the current multi-dimensional physical field data in the initial multi-physics synchronization feature sequence on the time axis according to the logical causal order and temporal sequence of each node in the fault evolution path. This ensures that the current multi-dimensional physical field data strictly matches the logical stage of fault evolution in terms of temporal sequence and semantics. Based on the target multi-physics synchronization feature sequence, a fault semantic description is determined. By extracting a fault evolution path with causal logic and temporal sequence from the target knowledge graph based on temporal guidance signals, and reordering and phase aligning the initial multi-physics synchronization feature sequence accordingly, it is possible to ensure that the physical field synchronization features corresponding to the current multi-dimensional physical field data strictly match the inherent logic of the fault evolution path in terms of temporal sequence and semantics, generating a highly reliable fault semantic description that accurately reflects the robot's fault type, location, and evolution mechanism.

[0057] Optionally, the semantic description of the fault can be determined as follows: Based on the historical fault characteristics and current fault characteristics of the fourth historical time period, a temporal evolution trend analysis is performed on the fault characteristics to obtain the dynamic evolution direction of the fault characteristics within the fourth historical time period, and the dynamic evolution direction is used as the temporal guidance signal for cross-physical field fusion. Using the temporal guidance signal as a constraint, fault evolution paths associated with the dynamic evolution direction are extracted from the target knowledge graph. The fault evolution path includes sequentially associated physical field feature nodes, intermediate state nodes, and fault mode nodes. Multi-resolution temporal decomposition is performed on various physical field sensing data (i.e., current multi-dimensional physical field data) to obtain the evolution components of acoustic signature data, vibration data, electromagnetic field data, and energy consumption data at different time scales (i.e., the feature extraction process of current multi-dimensional physical field data), thereby determining the initial multi-physical field synchronization feature sequence. Based on the node association order of the fault evolution path, the initial multi-physical field synchronization feature sequence is reordered and phase-aligned to generate a target multi-physical field synchronization feature sequence that matches the fault evolution path. The target multiphysics synchronization feature sequence is sequentially input into the path attention network along the fault evolution path. Using the node embeddings of each node in the fault evolution path as path guidance, the semantic matching degree between the target multiphysics synchronization feature sequence and the current node is calculated node by node. The semantic matching degrees of each node are accumulated along the path to generate a fault semantic description that integrates path evolution information and multiphysics synchronization features.

[0058] Optionally, fault features are a quantitative representation of the potential fault states of the quadruped robot. These features integrate differences between heterogeneous motor models and the degree of deviation from historical normal states, serving as a preliminary basis for fault judgment. Temporal evolution trend analysis involves analyzing fault features within the fourth historical time period to reveal their patterns and directions of change over time. Its purpose is to capture the dynamic process of a fault's development from its inception, rather than merely judging it at a single instant. The dynamic evolution direction refers to the changing trend of fault features within the fourth historical time period, such as whether the deviation continuously increases, decreases, fluctuates, or remains stable. It quantifies the "direction" and "speed" of fault feature changes over time. The temporal guidance signal encodes the aforementioned dynamic evolution direction and serves as guidance information for subsequent cross-physics semantic fusion processes. Its role is to ensure that the semantic fusion process focuses on the physical field information most relevant to the current fault evolution trend, improving the targeting and accuracy of the fusion.

[0059] Optionally, the dynamic evolution direction of fault characteristics within the fourth historical time period can be determined as follows: Based on the historical fault characteristics of the fifth historical time period, a temporal evolution trend analysis is performed to obtain a baseline temporal evolution trajectory. Based on this baseline trajectory, an evolutionary baseline manifold under normal conditions is fitted, where the quadruped robot is fault-free during the fifth historical time period. Based on the historical and current fault characteristics of the fourth historical time period, a temporal evolution trend analysis is performed to obtain a real-time evolution trajectory. The evolutionary baseline manifold is compared with the real-time evolution trajectory, and the instantaneous offset angle and offset rate of the real-time evolution trajectory relative to the baseline manifold are calculated. Based on the instantaneous offset angle and offset rate, a dynamic evolution direction vector within the fourth historical time period is fitted.

[0060] Optionally, the evolutionary baseline manifold can be determined as follows: Multiple historical evolution trajectories of the quadruped robot's fault characteristics across several historical fault-free periods are determined, and these trajectories are time-axis aligned and amplitude normalized to obtain a set of aligned, standardized trajectories. Phase space reconstruction is performed on each trajectory in the standardized trajectory set to obtain a sequence of phase points in a high-dimensional phase space. Local linear embedding is performed on all phase point sequences in the high-dimensional phase space to extract low-dimensional manifold coordinates that characterize the intrinsic structure of the fault characteristic evolution, resulting in an initial evolutionary manifold. Principal curve fitting is performed on the initial evolutionary manifold to obtain a one-dimensional continuous curve passing through the central region of the initial evolutionary manifold coordinates, which serves as the evolutionary baseline manifold.

[0061] Optionally, using time-series guiding signals as constraints means that when extracting fault evolution paths from the target knowledge graph, the dynamic evolution direction of fault features must be considered to select fault evolution paths that conform to the actual fault development trend, avoiding the extraction of irrelevant or erroneous fault evolution paths. The target knowledge graph is a structured knowledge base containing physical field feature nodes, intermediate state nodes, and fault mode nodes related to quadruped robots, as well as their interrelationships. It is typically represented in graph form, where nodes represent entities (e.g., "joint jamming," "motor overheating," "high-frequency vibration"), and edges represent relationships between entities (e.g., "leads to," "manifests as," "related to"). Its function is to provide domain expert knowledge to guide fault diagnosis and reasoning. A fault evolution path refers to a sequence starting from one or more physical field feature nodes, passing through a series of intermediate state nodes, and arriving at one or more fault mode nodes. This path describes the gradual evolution of a fault from its initial physical manifestation to its fault type, providing a logical framework for understanding and interpreting how multiple physical field sensing data collectively point to a specific fault. In the target knowledge graph, physical field feature nodes represent abnormal phenomena or characteristics observable by various physical field sensors, such as acoustic signature data, vibration data, electromagnetic field data, and energy consumption data. Examples include "high-frequency vibration," "abnormal current fluctuations," and "abnormal acoustic signature noise." Intermediate state nodes represent transitional stages or internal mechanisms in the fault evolution process. These states may not be directly observable by a single physical field sensor but can be inferred through reasoning or the fusion of multiple physical field sensor data. Examples include "accelerated bearing wear," "partial discharge in windings," "poor lubrication," and "motor overheating." Fault mode nodes represent fault types that need to be identified and addressed. Examples include "joint jamming," "motor overload," and "reducer damage."

[0062] Optionally, a method for extracting fault evolution paths associated with the dynamic evolution direction from the target knowledge graph, using a time-series guidance signal as a constraint, specifically includes: generating a corresponding direction quantization representation based on the dynamic evolution direction contained in the time-series guidance signal, and mapping the direction quantization representation to the embedding space of the physical field feature nodes in the target knowledge graph to obtain a direction query embedding. Based on the similarity between the direction query embedding and the embeddings of each physical field feature node in the target knowledge graph, recalling a set of initial physical field feature nodes that match the dynamic evolution direction from the target knowledge graph. Using each physical field feature node in the set of initial physical field feature nodes as the path starting point, searching along a preset association type in the target knowledge graph to obtain multiple candidate propagation paths from each path starting point to the fault mode node. Based on the evolution rate contained in the time-series guidance signal, filtering the multiple candidate propagation paths for path time consistency, and selecting the candidate propagation path with the highest overall semantic similarity to the dynamic evolution direction among the remaining candidate propagation paths as the fault evolution path associated with the dynamic evolution direction.

[0063] Optionally, multiple physical field sensing data refer to acoustic signature data, vibration data, electromagnetic field data, and energy consumption data synchronously collected by the quadruped robot during its operation in the cable tunnel. These data reflect the operating status of the quadruped robot from different dimensions. Multi-resolution temporal decomposition refers to decomposing the original physical field sensing data into components with different frequency and amplitude characteristics at different time scales. Its function is to capture the manifestation of faults at different time scales. For example, high-frequency impacts in vibration data may indicate bearing damage, while low-frequency fluctuations may indicate structural loosening. For example, wavelet basis functions at different scales can be used to decompose the signal to obtain detailed and approximate components in different frequency bands, thereby achieving multi-resolution analysis. Alternatively, complex signals can be adaptively decomposed into a series of intrinsic mode functions (IMFs), each IMF representing the oscillation mode of the signal at a specific time scale. Or, the signal can be converted to the frequency domain, and then different bandpass filters can be designed to extract signal components in different frequency ranges, and then inversely transformed back to the time domain. Evolutionary components (i.e., the corresponding physical field synchronization features) refer to the subsequences or features extracted from the original physical field sensing data after multi-resolution temporal decomposition, representing their dynamic changes at a specific time scale. Reordering and phase alignment refer to rearranging and synchronizing the evolutionary components of different physical fields according to the logical order and temporal relationship of nodes in the extracted fault evolution path. This ensures that the physical field data from different physical fields can be fused according to the logical order of the fault evolution path and maintain temporal consistency for subsequent semantic matching. For example, if the nodes in the target knowledge graph contain temporal information or relative temporal relationships, the starting points or key event points of different physical field data can be aligned based on this information. Alternatively, the similarity between the evolutionary components of different physical fields can be calculated, and the optimal time offset can be determined by maximizing the similarity, thereby achieving phase alignment. Furthermore, event markers related to specific nodes in the fault evolution path can be identified in different physical field data, and these events can be used as synchronization points for alignment. The target multiphysics synchronization feature sequence refers to a sequence that combines evolutionary components obtained from different physics and decomposed at different time scales after reordering and phase alignment. This sequence is consistent with the fault evolution path in both time and semantics, providing a structured input for subsequent path attention network processing.

[0064] Optionally, the semantic matching degree can be determined as follows: Extract the node embedding of the current node sequentially from the fault evolution path, and extract the physical field synchronization features corresponding to the node type from the target multi-physics synchronization feature sequence according to the node type of the current node. Project the physical field synchronization features onto the semantic space where the node embedding is located to obtain the feature vector to be matched. Calculate the cosine similarity between the feature vector to be matched and the node embedding, as the initial semantic matching degree of the current node. Obtain the historical matching degree (i.e., the path-guided semantic matching degree corresponding to the predecessor node of the current node in the fault evolution path, and use the historical matching degree as the path context weight to weight and correct the initial semantic matching degree of the current node, obtaining the path-guided semantic matching degree of the current node. Associate and store the path-guided semantic matching degree of the current node with the feature vector to be matched, as the historical matching degree referenced in the calculation of subsequent nodes, and repeat the above steps sequentially along the fault evolution path until the path-guided semantic matching degree calculation of all nodes is completed.

[0065] Optionally, a method for generating a fault semantic description that integrates path evolution information and multi-physics synchronization features by accumulating the semantic matching degrees of each node along the fault evolution path includes: sequentially obtaining the path-guided semantic matching degree of each node along the fault evolution path, and extracting the physical field synchronization features corresponding to each node from the target multi-physics synchronization feature sequence. The physical field synchronization features corresponding to each node are arranged according to the order of the nodes in the fault evolution path to obtain a path alignment feature sequence. The path-guided semantic matching degree of each node is used as a weight coefficient to perform a weighted summation of the physical field synchronization features at corresponding positions in the path alignment feature sequence to obtain a path accumulation vector. The path accumulation vector is concatenated with the path-guided semantic matching degree sequence, and the concatenated vector is input into a fully connected layer for feature fusion, outputting a fault semantic description that integrates path evolution information and the target multi-physics synchronization feature sequence.

[0066] Step S110: Based on the fault semantic description, determine the sequence of joint drive commands for the robot in a predetermined time period, where the predetermined time period is the time period after the current moment.

[0067] It is understandable that by generating a sequence of joint drive commands within a predetermined time period based on fault semantic description, it is possible to achieve forward planning and adaptive adjustment of robot motion parameters, thereby executing fault-tolerant control in the early stages of a fault, ensuring that the robot can maintain motion stability in complex environments and prevent the fault from worsening further.

[0068] In one optional embodiment, determining the robot's joint drive command sequence over a predetermined time period based on the fault semantic description includes: selecting multiple candidate fault-specific control strategies from multiple fault-specific control strategies based on the semantic similarity between the fault semantic description and multiple fault-specific control strategy vectors; determining a customized control parameter combination for the robot based on the control parameters corresponding to the multiple candidate fault-specific control strategies; correcting the gait cycle parameter and joint swing amplitude in the customized control parameter combination based on the current joint motion data and joint torque margin assessment value to obtain a corrected control parameter combination, wherein the joint torque margin assessment value is determined based on the current joint torque measurement value, and the gait cycle parameter refers to the time required for the robot to complete a complete gait cycle; determining the robot's motion adjustment commands based on the corrected control parameter combination; and determining the joint drive command sequence based on the motion adjustment commands.

[0069] It is understandable that the similarity analysis is performed between the fault semantic description and multiple fault-specific control strategy vectors. Based on the semantic similarity between the fault semantic description and the multiple fault-specific control strategy vectors, multiple candidate fault-specific control strategies are selected from the multiple fault-specific control strategies, and then the control parameters corresponding to each candidate fault-specific control strategy are obtained. The control parameters corresponding to the multiple candidate fault-specific control strategies are reconstructed by interpolation based on semantic similarity weighting to obtain a customized control parameter combination for the robot. Based on the current joint motion data and joint torque margin assessment values, the gait period parameters and joint swing amplitudes (i.e., the original gait period parameters and the original joint swing amplitudes) in the customized control parameter combination are corrected to obtain a corrected control parameter combination. Based on the corrected control parameter combination, dynamic matching is performed using a preset gait primitive library to obtain motion adjustment commands that adapt to the robot's current fault characteristics and remaining motion capabilities. Based on the motion adjustment commands, the joint drive command sequence of the robot in a predetermined time period is determined. By generating customized control parameter combinations through semantic similarity-based filtering and interpolation reconstruction, and then combining the current joint motion data and joint torque margin evaluation values ​​for adaptive correction and dynamic matching of gait primitives, motion adjustment commands that highly fit the current fault semantics and meet the robot's physical safety constraints are generated. This enables precise joint drive control of the robot while ensuring the robot's motion stability.

[0070] Optionally, motion adjustment instructions refer to instructions generated by a unified fault-tolerant decision model based on a fault semantic description, used to modify the motion parameters of the quadruped robot in real time. These instructions are designed to enable the quadruped robot to adjust its gait, posture, or joint movements when a fault occurs, in order to maintain the stability and functionality of the quadruped robot and prevent the fault from worsening or causing task failure.

[0071] Optionally, motion adjustment commands can be determined as follows: The fault semantic description is input into the policy routing network of the unified fault-tolerant decision model. The policy routing network maps the fault semantic description to a policy embedding space composed of multiple fault-specific control policies, outputting the semantic similarity distribution between the fault semantic description and each fault-specific control policy. Based on the semantic similarity distribution, the core control parameters (i.e., control parameters) of the top N policies (i.e., multiple candidate fault-specific control policies) that are semantically closest to the current fault semantic description are extracted from the multiple fault-specific control policies. The core control parameters of the top N policies are then reconstructed using semantic similarity-weighted interpolation in the policy embedding space to generate a customized control parameter combination corresponding to the current fault semantic description, where N is a positive integer. The customized control parameter combination is input into the motion adaptation correction module of the unified fault-tolerant decision model. The motion adaptation correction module adaptively scales the original gait period parameters and original joint swing amplitude in the customized control parameter combination based on the current joint motion data and joint torque margin evaluation values ​​of the quadruped robot, obtaining a corrected control parameter combination (corrected control parameter combination). The modified control parameter combination is dynamically matched with the preset gait primitive library. Based on the matching result, the corresponding gait primitive is selected, and then motion adjustment instructions that adapt to the current fault state and remaining motion capacity are generated.

[0072] Optionally, the unified fault-tolerant decision model is an intelligent decision-making system that integrates multiple fault-specific control strategies. Through knowledge fusion, it integrates and optimizes response strategies for different fault scenarios, forming a general framework capable of adaptively generating motion adjustment commands based on the current fault semantic description. The unified fault-tolerant decision model can be a simple rule engine that directly matches preset fault-specific control strategies (i.e., candidate fault-specific control strategies, such as "left front leg gait adjustment strategy") based on keywords in the fault semantic description (e.g., "left front leg motor overload"), and extracts their core control parameters (e.g., gait cycle, joint swing amplitude) to construct a customized control parameter combination for the quadruped robot. The motion adaptive correction module of the unified fault-tolerant decision model can perform simple linear scaling on the gait cycle parameter and joint swing amplitude in the customized control parameter combination based on the quadruped robot's current joint motion data (e.g., current speed, posture) and joint torque margin assessment value. For example, if the joint torque margin assessment value is small, the swing amplitude is reduced. The modified control parameter combination (i.e., the modified control parameter combination) is matched with a preset gait primitive library (e.g., containing several fixed gaits such as walking, jogging, and slow walking), the closest gait primitive is selected, and motion adjustment instructions are generated.

[0073] Optionally, the customized control parameter combination can be determined as follows: Extract the core control parameter vectors corresponding to each fault-specific control strategy from the first N strategies, and obtain the strategy embedding vectors of the first N strategies in the strategy embedding space. Map the fault semantic description to the strategy embedding space to obtain the query embedding point corresponding to the current fault semantic description, and calculate the distance between the query embedding point and the strategy embedding vectors corresponding to the first N strategies to obtain the spatial distance between each fault-specific control strategy and the query embedding point. Use the reciprocal of the spatial distance as a basis function to perform radial basis function interpolation on the core control parameter vectors corresponding to the first N strategies to construct a parameter interpolation function with the query embedding point as the independent variable. Substitute the query embedding point into the parameter interpolation function to calculate the parameter vector corresponding to the position of the current fault semantic description in the strategy embedding space, and use the parameter vector as the customized control parameter combination.

[0074] Optionally, the corrected control parameter combination can be determined as follows: Based on the difference between the actual torque output value of each joint at the current moment (i.e., the current joint torque measurement) and the torque limit value, calculate the current torque margin of each joint, and aggregate the current torque margins of each joint to obtain a global torque margin evaluation value (i.e., the joint torque margin evaluation value). Based on the global torque margin evaluation value, determine the adjustable range of the gait cycle parameter and the adjustable range of the joint swing amplitude. Use the upper limit of the adjustable range of the gait cycle parameter as the target gait cycle, and the lower limit of the adjustable range of the joint swing amplitude as the target swing amplitude, to construct an initial scaling target vector. Determine the tracking error between the actual motion trajectory and the desired motion trajectory of the quadruped robot at the current moment based on the current joint motion data. Calculate the stability margin of the current motion state of the quadruped robot based on the tracking error, and use this stability margin as a scaling factor to dynamically adjust the initial scaling target vector to obtain the target scaling target vector. The original gait cycle parameters and original joint swing amplitude in the customized control parameter combination are linearly interpolated to the target scaling target vector to generate the adaptively scaled target gait cycle parameters and target joint swing amplitude. Based on the adaptively scaled target gait cycle parameters and target joint swing amplitude, the corrected control parameter combination is determined.

[0075] Optionally, the joint drive command sequence can be determined and the motion parameters of the quadruped robot adjusted in real time as follows: The motion adjustment commands are parsed to extract the adaptively scaled target gait cycle parameters, target joint swing amplitude, and target body posture parameters, which are used as target values ​​for the motion parameters to be adjusted. The actual gait cycle parameters, actual joint swing amplitude parameters, and actual body posture parameters of the quadruped robot at the current moment are obtained. Based on the differences between the actual gait cycle parameters, actual joint swing amplitude parameters, and actual body posture parameters and the target values ​​of the motion parameters, adjustment step sizes for each motion parameter are generated. Based on the adjustment step sizes of each motion parameter and combined with preset motion smoothing constraints, a continuous transition trajectory from the current motion parameters to the target values ​​of the motion parameters (i.e., the trajectory formed by the joint drive command sequence of the quadruped robot over a predetermined time period) is generated. The joint drive commands of the quadruped robot are updated hourly according to the continuous transition trajectory. During the update process, the tracking error between the actual feedback position and the desired position of each joint is monitored in real time. When the tracking error exceeds a preset threshold, the adjustment is paused and the current motion parameters are maintained.

[0076] Through the above steps S102 to S110, the goal is to determine the robot's fault semantic description by combining the current fault characteristics determined based on the robot's current joint motion data and current joint torque measurement values ​​with the target knowledge graph and current multidimensional physical field data, and to determine the robot's joint drive command sequence based on the fault semantic description. This achieves the technical effect of improving the accuracy of the joint drive command sequence determination result when the robot is in faulty, thereby solving the technical problem of inaccurate joint drive command sequence determination result when the robot is in faulty in related technologies.

[0077] Based on the above embodiments and optional embodiments, this application proposes an implementation method for an optional fault-based robot motion adjustment instruction generation method. The following describes this implementation method in detail with the example of a quadruped robot inspecting a cable tunnel.

[0078] Figure 2 This is a flowchart of an optional fault-based robot motion adjustment instruction generation method provided according to an embodiment of this application, such as... Figure 2 As shown, the steps of this optional implementation include:

[0079] Step S1: Obtain real-time joint motion data (i.e., current joint motion data) of the quadruped robot running in the cable tunnel, synchronously collected multiple physical field sensor data (i.e., current multidimensional physical field data), and the long-term mutual calibration historical state of the multiple heterogeneous motor models (including the first heterogeneous motor model, the second heterogeneous motor model, and the third heterogeneous motor model) formed by the quadruped robot in the historical operating cycle (i.e., the first historical time period). The multiple physical field sensor data include acoustic data, vibration data, electromagnetic field data, and energy consumption data. The long-term mutual calibration historical state of the multiple heterogeneous models includes the prediction consistency distribution characteristics of each heterogeneous motor model in the historical fault-free period and the historical fluctuation law of the historical deviation value of each heterogeneous motor model (i.e., historical fluctuation law).

[0080] Current joint motion data can be acquired by installing simple angle and current sensors at each joint of the quadruped robot. For example, joint angles can be measured using potentiometers or Hall effect sensors, and motor current can be measured using current transformers. Various physical field sensing data can be collected by deploying independent microphones, single-axis accelerometers, simple magnetic field detectors, and ammeters near the quadruped robot's body and joints to collect acoustic signature data, vibration data, electromagnetic field data, and energy consumption data, respectively. Obtaining the long-term cross-calibration history of multiple heterogeneous models can be achieved by manually recording the predicted and measured joint torque values ​​of each heterogeneous motor model after each fault-free operation of the quadruped robot, and calculating their average deviation and standard deviation as the historical fluctuation pattern of the historical deviation values ​​of each heterogeneous motor model. Simultaneously, the statistical distribution of the average absolute error between the predicted joint torque values ​​of each heterogeneous motor model can be recorded as the predictive consistency distribution characteristics of each heterogeneous motor model within a historical fault-free period (e.g., the first historical time period).

[0081] The long-term cross-calibration history of multiple heterogeneous motor models refers to the state information formed by predicting joint torques using multiple heterogeneous motor models and performing long-term cross-validation and calibration of these historical joint torque predictions during the fault-free operation period of the quadruped robot (e.g., the first historical time period). It includes two main aspects: First, the prediction consistency distribution characteristics, which characterize the statistical distribution and characteristics of the dispersion (e.g., represented by mean absolute error) between the prediction results of each heterogeneous motor model under normal operating conditions (i.e., during the fault-free operation of the quadruped robot). Second, the historical fluctuation patterns of the historical deviation values ​​of each heterogeneous motor model, recording the statistical characteristics of the historical deviation values ​​between the historical joint torque prediction values ​​and the historical joint torque measurement values ​​of each heterogeneous motor model under normal operating conditions.

[0082] Step S1 above proposes acquiring real-time joint motion data of the quadruped robot running in a cable tunnel, synchronously collecting various physical field sensor data, and the long-term cross-calibration historical state of multiple heterogeneous motor models to support fault early warning. However, in its implementation, the accuracy and completeness of data acquisition are insufficient, which may lead to inaccurate determination of current fault characteristics. Specifically, the real-time joint motion data acquisition method is not clearly defined, easily introducing noise. The lack of a synchronous acquisition mechanism for various physical field sensor data causes time mismatch. The long-term cross-calibration historical state of multiple heterogeneous motor models is not deeply analyzed in conjunction with environmental conditions, failing to provide a reliable historical benchmark. These shortcomings reduce the reliability of fault early warning.

[0083] To address this, a method is further proposed for acquiring real-time joint motion data of a quadruped robot operating in a cable tunnel, synchronously collecting various physical field sensor data, and long-term cross-calibration of historical states using multiple heterogeneous motor models formed during the quadruped robot's historical operating cycle. Specifically, this includes:

[0084] Step S11: The encoders and current sensors deployed on each joint of the quadruped robot collect the current joint angle data, current joint angular velocity data and current joint current data in real time, which are used as the real-time joint motion data.

[0085] This step aims to accurately acquire joint motion data for each joint of the quadruped robot. Encoders are used to measure the rotation angle of the joints, thus forming current joint angle data, reflecting the instantaneous position of the joint. Current sensors are used to measure the current driving the motors of each joint, forming current joint current data, which is directly related to the joint's output torque and load condition. Current joint angular velocity data can be obtained by performing time differentiation on the current joint angle data, or by directly measuring it using a dedicated angular velocity sensor. This real-time, high-precision joint motion data is fundamental to assessing the quadruped robot's health status and identifying potential faults. Implementation methods may include: installing a high-resolution photoelectric encoder or magnetic encoder on the rotation axis of each joint to acquire current joint angle data by counting pulses or reading absolute position signals. Simultaneously, a Hall effect current sensor or shunt resistor is connected in series in the power supply line of each joint drive motor to monitor the motor's current joint current data in real time. Another approach is to use an integrated intelligent joint module. This module integrates a high-precision encoder, a current sensor, and a microcontroller. It can directly output pre-processed front joint angle data, current joint angular velocity data, and current joint current data. It communicates with the main controller via a high-speed bus (such as EtherCAT (Ethernet for Control Automation Technology) or CANopen (CAN application protocol)) to ensure the real-time performance and synchronization of the data.

[0086] Step S12: By deploying an array of acoustic fingerprint sensors, an array of accelerometers, an array of electromagnetic field sensors, and an energy consumption monitoring module near the body and joints of the quadruped robot, acoustic fingerprint data, triaxial vibration data (i.e., vibration data), electromagnetic field strength data (i.e., electromagnetic field data), and the instantaneous power and current waveforms of each motor (i.e., energy consumption data) are collected synchronously as multiple physical field sensing data.

[0087] This step aims to comprehensively and synchronously acquire multi-dimensional physical field information about the quadruped robot's operating environment and its own state. An acoustic sensor array is used to capture acoustic features generated during the quadruped robot's operation, such as motor noises and joint friction sounds; these sound signals are often early signs of mechanical failure. An accelerometer array measures the triaxial vibration signals of the quadruped robot's body and joints; changes in vibration modes can indicate problems such as structural loosening, bearing wear, or imbalance. An electromagnetic field sensor array monitors the electromagnetic environment around the quadruped robot and the electromagnetic radiation generated by its own motors; abnormal electromagnetic field strength may be related to motor failure or wiring problems. The energy consumption monitoring module records the instantaneous power and current waveforms of each motor in real time; abnormal fluctuations in energy consumption or distortions in the current waveform are important indicators of motor overload, efficiency degradation, or internal short circuits. The synchronous acquisition mechanism of these sensor arrays and modules ensures time alignment between data from different physical fields, providing a reliable foundation for subsequent cross-physical field fusion analysis. The implementation can be achieved in several ways: The acoustic sensor array can utilize a MEMS (Micro-Electro-Mechanical Systems) microphone array, using beamforming technology to focus sound sources in a specific area. The accelerometer array can employ a triaxial MEMS accelerometer, deployed at key structural points. The electromagnetic field sensor array can use Hall effect sensors or fluxgate sensors. The energy consumption monitoring module can be integrated into the motor driver, outputting power and current data in real time. All sensors acquire data through a unified clock synchronization mechanism (such as PTP (Precision Time Protocol)). Another implementation method uses multifunctional integrated sensor nodes, each containing acoustic, vibration, and electromagnetic field sensing units. These nodes are synchronously sampled via a high-speed data acquisition card, transmitting the raw data to an edge computing unit for preliminary processing and timestamping. The energy consumption monitoring module operates independently, but its data stream is aligned with the sensor node's data stream through a unified time synchronization server.

[0088] Step S13: Read from local storage or cloud database the historical joint torque prediction value sequence, historical joint torque measurement value sequence, and corresponding historical environmental condition labels of each heterogeneous motor model recorded during multiple fault-free runs of the quadruped robot in the historical operating cycle.

[0089] This step aims to obtain the long-term cross-calibration history of multiple heterogeneous motor models used to build historical benchmarks. This historical data (historical joint torque prediction sequences and historical joint torque measurement sequences) is a key reference for assessing whether the current state of the quadruped robot is abnormal. The historical joint torque prediction sequence of the heterogeneous motor models records the predicted joint torque results of each heterogeneous motor model under historical fault-free conditions, reflecting the model's behavior pattern under normal operating conditions. The historical joint torque measurement sequence provides real joint torque data for calibrating and verifying the accuracy of model predictions. Recording environmental condition labels (such as temperature, humidity, terrain type, load conditions, etc.) allows historical data to be associated with specific operating environments, enabling the retrieval of the most relevant historical benchmarks for the current environmental conditions in subsequent analyses, avoiding misjudgments caused by environmental changes. This can be achieved by installing a high-capacity solid-state drive (SSD) inside the quadruped robot as local storage, periodically writing data to it. Simultaneously, data is uploaded to a cloud database via wireless communication modules (such as 5G (5th Generation Mobile Networks) or Wi-Fi (Wireless Fidelity)) to achieve redundant data backup and remote access. Another approach is to employ a distributed data storage architecture, where the quadruped robot stores only recent data locally, while historical long-term data is synchronized in real-time to an enterprise-level data center or public cloud platform via an edge gateway. Big data technologies are then used for indexing and management to ensure efficient retrieval.

[0090] Step S14: Perform statistical analysis on the historical deviation value sequence between the historical joint torque prediction value sequence and the historical joint torque measurement value sequence of each heterogeneous motor model to obtain the historical fluctuation law of the historical deviation value of each heterogeneous motor model under various historical environmental conditions.

[0091] This step aims to quantify the prediction accuracy and stability of various heterogeneous motor models under different environmental conditions. By statistically analyzing the historical deviation sequence between the historical joint torque prediction value sequence and the historical joint torque measurement value sequence in historical fault-free data, the "normal" deviation range and distribution characteristics of each heterogeneous motor model under specific environmental conditions can be established. This fluctuation pattern can distinguish between the inherent prediction error of each heterogeneous motor model, normal fluctuations caused by environmental factors, and abnormal deviations caused by actual faults, thereby improving the sensitivity and accuracy of fault early warning. The implementation can include: for each heterogeneous motor model and each environmental condition, calculating the mean, standard deviation, variance, skewness, kurtosis, and other statistical measures of its historical deviation values ​​(historical joint torque measurement value minus the corresponding historical joint torque prediction value), and fitting the probability density function of the deviation (such as Gaussian distribution, Laplace distribution, etc.) to obtain the historical fluctuation pattern. Another approach is to use nonparametric statistical methods, such as kernel density estimation (KDE) or empirical cumulative distribution function (ECDF), to directly learn the distribution characteristics from historical deviation data, avoiding pre-defined assumptions about the data distribution. Simultaneously, methods such as sliding windows or exponentially weighted moving averages can be used to dynamically update these statistics to adapt to slow changes in environmental conditions.

[0092] Step S15: Further analyze the dispersion of the historical joint torque prediction values ​​of multiple heterogeneous motor models at the same moment in the historical operating cycle to obtain the historical prediction consistency distribution characteristics of each heterogeneous motor model under various historical environmental conditions.

[0093] This step aims to assess the level of consistency in prediction results among heterogeneous motor models under historical fault-free conditions. When multiple heterogeneous motor models predict the same physical quantity, the differences in their predicted values ​​(dispersion) can reflect the collaborative working state of the multiple heterogeneous motor models and the degree of consensus on system behavior. Under normal circumstances, although these models are heterogeneous, their prediction results should remain consistent within a certain range. By analyzing this historical consistency distribution characteristic, a "normal" consistency benchmark can be established. When the real-time prediction consistency coefficient deviates from this benchmark, it may indicate that an anomaly has occurred in a certain model or system. This can be achieved by: calculating the standard deviation, variance, coefficient of variation (the ratio of standard deviation to mean), or mean absolute error of the historical joint torque prediction values ​​of multiple heterogeneous motor models at the same time to measure the dispersion. Then, statistical analysis is performed on the dispersion indices at different times to obtain the corresponding statistical distribution. Based on this statistical distribution and its statistical characteristics, the historical prediction consistency distribution characteristics are obtained. Another approach is to use information theory metrics such as information entropy, Kullback-Leibler divergence, or Jensen-Shannon divergence to quantify the differences in the distribution of historical joint torque predictions among multiple heterogeneous motor models, thereby obtaining a more robust measure of dispersion. Then, the distributions of these information theory metrics are modeled to form a consistent distribution characteristic of historical predictions.

[0094] Step S16: The historical fluctuation pattern of prediction deviation and the distribution characteristics of historical prediction consistency are used together as the long-term mutual calibration history of the multi-heterogeneous motor model.

[0095] This step aims to integrate the two key historical statistical information sets mentioned above to form a comprehensive, multi-dimensional historical benchmark. The historical deviation value fluctuation pattern focuses on the predictive accuracy of individual heterogeneous motor models and their stability under different environmental conditions, while the historical prediction consistency distribution characteristics focus on the synergy and consensus among multiple heterogeneous motor models. Combining the two allows for a precise characterization of the quadruped robot's historical fault-free state from both "individual performance" and "group consensus" perspectives. This comprehensive historical state reference provides richer and more reliable background information for subsequent real-time fault feature extraction, enabling the system to more accurately distinguish between normal fluctuations, model errors, and actual faults. This can be achieved by packaging the historical deviation value fluctuation pattern (such as the mean, standard deviation, and probability density function parameters of each model's deviation) and the historical prediction consistency distribution characteristics (such as the mean, standard deviation, distribution function parameters, and probability density function of the consistency coefficient) into a structured data object or tuple and storing it in a database for subsequent querying and use. Another approach is to construct a multidimensional feature vector containing key statistical indicators such as the historical fluctuation patterns of historical deviation values ​​and the distribution characteristics of historical prediction consistency. This multidimensional feature vector is then modeled using machine learning models (such as Gaussian mixture models or autoencoders) to learn its distribution manifold under normal conditions, thereby forming a more abstract and robust representation of the long-term mutual calibration history of multi-heterogeneous motor models.

[0096] Through steps S1-S16, the acquisition of real-time joint motion data and various physical field sensor data achieves high precision and good synchronization, avoiding noise introduction and time mismatch issues. Simultaneously, through in-depth statistical analysis of fault-free historical data, combined with environmental conditions, a long-term cross-calibration historical state of multiple heterogeneous models is constructed, incorporating historical deviation value fluctuation patterns and historical prediction consistency distribution characteristics. This provides a comprehensive, reliable, and environmentally adaptive historical reference benchmark for subsequent fault feature extraction. This improves the accuracy and reliability of fault early warning, enabling the system to more effectively identify real faults and distinguish environmental interference, laying a solid data foundation for early warning of quadruped robots in complex cable tunnel environments.

[0097] Step S2: Based on real-time joint motion data and the long-term mutual calibration history of the multi-heterogeneous motor model, the current fault characteristics are obtained by calculating the deviation of the current joint torque prediction set of the multi-heterogeneous motor model, evaluating the consistency of the current prediction, and analyzing the deviation of the distribution characteristics of the consistency of the current prediction and the consistency of the historical prediction.

[0098] Current joint motion data can be input into pre-defined multi-heterogeneous motor models (e.g., simplified physical models based on Newton's second law and models based on simple neural networks). These models output predicted joint torque values ​​at the current moment, resulting in a set of predicted joint torque values. Then, a simple difference calculation is performed between the predicted joint torque value of each heterogeneous motor model and the measured joint torque value at the current moment, yielding multiple real-time deviation values ​​(e.g., a first deviation value, a second deviation value, and a third deviation value), thus obtaining a set of current deviation values. The dispersion of the current joint torque prediction value set from the multi-heterogeneous motor models is calculated, for example, by calculating its variance or standard deviation, to obtain the real-time prediction consistency coefficient (i.e., the current prediction consistency coefficient). Based on the real-time prediction consistency coefficient and the historical prediction consistency distribution characteristics (e.g., a statistical distribution constructed based on the historical prediction consistency coefficients of a first historical time period), consistency deviation characteristics are obtained. The current deviation value set and consistency deviation characteristics are then simply concatenated to form current fault characteristics that incorporate historical state references.

[0099] Step S2 above proposes to obtain the current fault features based on the historical state reference obtained by long-term mutual calibration of real-time joint motion data and multiple heterogeneous motor models to realize fault early warning. However, in this process, real-time deviation calculation and consistency evaluation may be misjudged due to environmental changes or the errors of the heterogeneous motor models themselves, resulting in inaccurate or non-robust extraction of current fault features, and inability to effectively distinguish between real faults and environmental interference.

[0100] To address this, a method is further proposed that integrates historical data from real-time joint motion data with the long-term cross-calibration of multiple heterogeneous motor models. This method involves calculating the deviation of the current joint torque prediction set from the multiple heterogeneous motor models, evaluating the consistency of the current predictions, and analyzing the deviation between the consistency of the current predictions and the historical predictions. Specifically, this method yields current fault characteristics that incorporate historical data.

[0101] Step S21: Real-time joint motion data is input in parallel to multiple pre-trained heterogeneous motor models, and each heterogeneous motor model outputs the predicted value of joint torque at the current moment.

[0102] This step aims to synchronously predict the joint torque of a quadruped robot using multiple heterogeneous motor models with different modeling principles and data sources. The heterogeneous motor models can include analytical models based on physical equations (i.e., the first heterogeneous motor model), data-driven models based on machine learning (such as neural networks and support vector machines) (i.e., the second heterogeneous motor model), and models based on empirical rules or lookup tables (i.e., the third heterogeneous motor model). During the training phase, these models utilize historical operating data and / or theoretical knowledge for parameter optimization and structure learning to ensure their predictive capabilities under different environmental conditions. Real-time joint motion data can include current joint angles, current joint angular velocities, current joint angular accelerations, and current joint currents, which are synchronously input into each heterogeneous motor model. Each heterogeneous motor model independently calculates and outputs its predicted value for the current joint torque based on its internal logic and parameters. This parallel processing approach provides multi-perspective prediction results, offering rich information for subsequent deviation calculations and consistency assessments.

[0103] Step S21 above proposes using multiple heterogeneous motor models to obtain joint torque prediction values ​​in order to improve the accuracy of current fault feature extraction. However, in its implementation, if the heterogeneous motor model type and input method are not specifically defined, the joint torque prediction values ​​may not fully cover different dimensions such as physical constraints, dynamic evolution and fast response, resulting in insufficient accuracy of current fault feature extraction.

[0104] To address this, a method is further proposed that real-time joint motion data be input in parallel to multiple pre-trained heterogeneous motor models, with each model outputting a predicted joint torque value for the current moment. Specifically, this includes:

[0105] Step S211: Extract the current joint angle, current joint angular velocity, and current joint current from the real-time joint motion data as a shared input feature set.

[0106] Real-time joint motion data refers to the dynamic information about the joint state of a quadruped robot acquired in real time through sensors during actual operation. This typically includes joint angles, current joint angular velocity, joint acceleration, and joint current or torque driving the joints. Its purpose is to provide an accurate description of the current motion state of the quadruped robot's joints, serving as the basis for assessing the robot's health and predicting its behavior. In addition to collecting joint angles, angular velocities, and joint currents through encoders and current sensors deployed at each joint, joint torque data can also be collected through force sensors deployed at the joints, or the joint motion state can be inferred from the kinematic model using an Inertial Measurement Unit (IMU). The shared input feature set refers to a unified data set extracted from the real-time joint motion data and used as input to multiple heterogeneous motor models. It includes the current joint angle, current joint angular velocity, and current joint current at the current moment. Its purpose is to provide a consistent and comprehensive input basis for different heterogeneous motor models, ensuring that each model makes predictions at the same time and under the same motion state, thus facilitating subsequent comparison and fusion of the prediction results from each heterogeneous motor model. In addition to the three types of features mentioned above, the shared input feature set can also include environmental parameters such as joint acceleration and temperature, or the overall posture information of the quadruped robot.

[0107] Step S212: The shared input feature set is input in parallel into the first heterogeneous motor model, the second heterogeneous motor model, and the third heterogeneous motor model. The first heterogeneous motor model is a physical mechanism model built based on the motor dynamics equations. The second heterogeneous motor model is a time-series learning model built based on a long short-term memory network. The third heterogeneous motor model is a lookup table model built based on offline calibration data.

[0108] The physical mechanism model is a mathematical model established based on the physical laws and engineering principles of motor operation. By modeling the electromagnetic, mechanical, and thermodynamic processes inside the motor, it can accurately describe the torque output characteristics of the motor under different inputs. The first heterogeneous motor model provides torque predictions that conform to physical constraints, ensuring the physical rationality of the prediction results. Besides being based on classical DC or AC motor dynamics equations (such as Kirchhoff's laws and Newton's second law), it can also be constructed using finite element analysis (FEA) for refined motor modeling, or by combining empirical parameters for correction. The time-series learning model is a machine learning model capable of processing and predicting time-series data. Long Short-Term Memory (LSTM) networks, as a special type of recurrent neural network (RNN), are particularly adept at capturing long-term dependencies and dynamic patterns in time-series data. The second heterogeneous motor model utilizes current joint motion data combined with historical shared input feature sets from historical time-series windows to learn the complex dynamic evolution of joint motion, thereby predicting future torque. Besides LSTM, other advanced temporal learning architectures such as Gated Recurrent Units (GRUs), Transformer models, or Deep Convolutional Networks (DCNs) can be used to build such models. The lookup table model is a mapping model based on pre-measured offline calibration data. It stores the correspondence between input parameters (such as joint current) and output parameters (such as joint torque) in a table, and quickly retrieves the predicted value at runtime by looking up the table. The second heterogeneous motor model provides extremely fast torque prediction to meet real-time requirements. In addition to directly storing the current-torque correspondence table, methods such as multinomial fitting and spline interpolation can be used to smooth and functionalize the offline calibration data to improve lookup accuracy and generalization ability.

[0109] Step S213: Solve the physical equations of the shared input feature set using the first heterogeneous motor model, and output the first joint torque prediction value that conforms to the motor dynamics constraints.

[0110] Physical equation solving refers to using mathematical methods and numerical calculation techniques to solve the set of dynamic equations contained in the physical mechanism model, thereby calculating the motor's torque output based on a shared input feature set. Its purpose is to ensure that the predicted torque value of the first joint strictly follows physical laws. Solving methods can include analytical methods, numerical integration methods, or iterative solutions based on optimization algorithms.

[0111] Step S214: The shared input feature set and its associated historical time window are nonlinearly mapped using the second heterogeneous motor model to output the predicted value of the second joint torque that captures the dynamic evolution law.

[0112] Nonlinear mapping refers to the process by which a temporal learning model transforms its input into a predicted value of the second joint torque through its internal nonlinear activation function and multi-layer network structure. Its purpose is to capture the complex nonlinear relationship between input and output, thereby better reflecting the dynamic characteristics of joint motion. Besides the gating mechanism and activation function within LSTM, this can be achieved by introducing attention mechanisms, residual connections, or multi-head self-attention mechanisms to enhance the model's nonlinear expressive power.

[0113] Step S215: The current joint current in the shared input feature set is mapped by a lookup table using the third heterogeneous motor model, and the predicted value of the third joint torque with the fastest response speed is output.

[0114] Lookup table mapping refers to a lookup table model that searches for the corresponding torque output value in a pre-built lookup table based on the input current joint current value. Its purpose is to provide a fast and direct prediction of the third joint torque. Besides direct hash table lookup or binary search, it can also be implemented using an index-based fast lookup algorithm, or by storing interpolation coefficients in the lookup table to perform linear or nonlinear interpolation during the lookup, thereby improving prediction accuracy.

[0115] Steps S21-S215 above, by defining the specific types and input / output methods of three heterogeneous motor models, solve the problems of model diversity and robustness when obtaining joint torque prediction values, thereby generating a more comprehensive and reliable set of current joint torque prediction values, providing a solid foundation for subsequent fault feature extraction. Specifically, the current joint angle, current joint angular velocity, and current joint current at the current moment are extracted from real-time joint motion data as a shared input feature set, ensuring that all heterogeneous motor models are based on the same real-time data input, avoiding prediction deviations caused by inconsistent inputs, and providing a unified basis for parallel processing of heterogeneous motor models. The shared input feature set is input in parallel into the first heterogeneous motor model, the second heterogeneous motor model, and the third heterogeneous motor model. By inputting in parallel into three different types—physical mechanism model, time series learning model, and lookup table model—model diversity is achieved, covering different aspects such as physical constraints, time series dynamics, and fast response. The first heterogeneous motor model is a physical mechanism model built based on motor dynamics equations. This model is based on physical principles, ensuring that the first joint torque prediction value conforms to motor dynamics constraints, providing a physically reasonable torque estimate. The second heterogeneous motor model is a time-series learning model built on a long short-term memory network. This model performs a nonlinear mapping on the shared input feature set and its associated historical time-series feature set, capturing the dynamic evolution of joint motion and adapting to time-related changes. The third heterogeneous motor model is a lookup table model built on offline calibration data. This model performs a lookup table mapping on the current joint current in the shared input feature set, providing low-latency predictions and ensuring real-time response. The first heterogeneous motor model outputs a first joint torque prediction value that conforms to motor dynamics constraints, enhancing the physical reliability and constraint of the prediction. The second heterogeneous motor model outputs a second joint torque prediction value that captures dynamic evolution, improving adaptability to dynamic environmental changes. The third heterogeneous motor model outputs the fastest-responding third joint torque prediction value, ensuring real-time performance and efficiency. These features work synergistically to generate diverse joint torque prediction values, combining physical, dynamic, and speed advantages to improve the overall robustness and accuracy of the prediction.

[0116] In some embodiments described above, this application proposes retrieving historical fluctuation patterns of prediction deviations and historical prediction consistency distribution characteristics from long-term cross-calibration historical states of multiple heterogeneous models to generate fault features that incorporate historical state references. However, during implementation, due to dynamic changes in environmental conditions, historical state data may not accurately match current operating conditions, leading to a disconnect between the retrieved historical fluctuation patterns of prediction deviations and historical prediction consistency distribution characteristics and the real-time environment. This affects the accuracy of fault feature extraction, making it impossible to effectively distinguish between real faults and environmental interference, thus reducing the reliability of the early warning system.

[0117] Step S22: Retrieve historical deviation values, historical fluctuation patterns, and historical prediction consistency distribution characteristics from the historical data record set (which stores the historical long-term mutual calibration history of the quadruped robot's multi-heterogeneous motor models acquired in historical time periods) that match the current environmental conditions.

[0118] This step aims to retrieve the most relevant historical cross-calibration status of multiple heterogeneous motor models from a pre-stored historical data set, based on the current environmental conditions of the quadruped robot. The historical cross-calibration status of the multiple heterogeneous motor models can be stored in a local database or a cloud server. Its content includes the statistical characteristics (i.e., historical fluctuation patterns of historical deviation values) (such as mean, variance, and distribution type) of the historical deviation values ​​of each heterogeneous motor model under different environmental conditions (e.g., temperature, humidity, load type, operating mode, etc.) and the statistical distribution of prediction consistency among models (i.e., historical prediction consistency distribution characteristics). The retrieval process can match the current environmental sensor data with the environmental condition tags in the historical records to obtain the most relevant historical reference data. The retrieval mechanism can employ index-based fast lookup or similarity-based fuzzy retrieval. For example, the system can pre-cluster the historical status data to form typical patterns representing different environmental conditions. When the real-time environmental condition (i.e., the current environmental condition) occurs, the system determines the most matching historical state by calculating the distance to these typical patterns and extracts the corresponding historical fluctuation patterns of historical deviation values ​​and historical prediction consistency distribution characteristics.

[0119] Step S22 above proposes retrieving historical deviation values, historical fluctuation patterns, and historical prediction consistency distribution characteristics from the historical data record set to generate current fault features that integrate historical state references. However, in its implementation, due to the dynamic changes in environmental conditions, it may not accurately match the current environmental conditions. This leads to a long-term disconnect between the retrieved historical states of the multi-heterogeneous motor models and the real-time environment, thus affecting the accuracy of current fault feature extraction, failing to effectively distinguish between real faults and environmental interference, and reducing the reliability of fault warnings.

[0120] To address this, a method is further proposed to retrieve historical deviation values, historical fluctuation patterns, and historical prediction consistency distribution characteristics from historical data records that match current environmental conditions. Specifically, this includes:

[0121] Step S221: Based on real-time joint motion data and synchronously collected ambient temperature data, humidity data, and tunnel section markings, identify the current environmental conditions of the quadruped robot.

[0122] Real-time joint motion data refers to the joint angles, angular velocities, and joint currents collected in real time by encoders and current sensors deployed at each joint during the quadruped robot's operation within a cable tunnel. This data accurately reflects the quadruped robot's current posture, speed, and joint load, providing fundamental information for assessing its operational status. The synchronously acquired ambient temperature data refers to the temperature value measured in real time within the quadruped robot's operating environment. The concept here is to quantify the thermodynamic state of the environment in which the quadruped robot exists. Possible implementation methods include: using a thermistor sensor to infer temperature by measuring resistance changes; or using a resistance temperature detector to measure the temperature by utilizing the characteristic of metal resistance changing with temperature; or using a thermocouple to obtain temperature by measuring the thermoelectric potential difference between different metal junctions. The humidity data refers to the air humidity value measured in real time within the quadruped robot's operating environment. The concept here is to quantify the water vapor content in the environment. Possible implementation methods include: using a capacitive humidity sensor to reflect humidity by measuring changes in dielectric constant; or using a resistive humidity sensor to measure the temperature by measuring the characteristic of conductive polymer resistance changing with humidity. Humidity can also be calculated by measuring the temperature difference between the dry and wet bulbs using the principle of wet-bulb and dry-bulb thermometers. The tunnel section identifier refers to the marking information used to uniquely identify the specific geographical location or logical area of ​​the cable tunnel where the quadruped robot is currently located; its concept is to provide precise spatial context. Possible implementations include: deploying Radio Frequency Identification (RFID) tags at specific locations in the tunnel, with the quadruped robot acquiring the tag via a reader. Alternatively, using the quadruped robot's own Simultaneous Localization and Mapping (SLAM) technology combined with a pre-drawn tunnel map to infer the current section in real time. Pre-set milestones or visual markers can also be used, with section information obtained through image recognition technology.

[0123] Based on this, identifying the current environmental condition of the quadruped robot aims to comprehensively utilize real-time joint motion data, ambient temperature data, humidity data, and tunnel section markers to classify or quantify the current environmental condition in which the quadruped robot is located. Its purpose is to provide accurate matching conditions for subsequent long-term cross-calibration historical state retrieval of multiple heterogeneous motor models. Possible implementation methods include: employing a rule-based expert system to pre-define the environmental condition types corresponding to different combinations of environmental parameters; or using machine learning classification algorithms, such as support vector machines or neural networks, to train on multi-dimensional input data and automatically identify the current environmental condition; or using clustering analysis methods to group similar operating environment parameters into the same environmental condition.

[0124] Step S222: In the historical data record set, retrieve the long-term mutual calibration history of the multi-heterogeneous motor model that matches the current environmental conditions.

[0125] This step aims to filter historical state data of multiple heterogeneous motor models that are highly similar to or completely match the current environmental conditions from a pre-stored historical data set. The purpose is to ensure that the subsequently extracted historical reference data is environmentally relevant. Possible implementation methods include: building an index based on environmental condition tags in the historical data set for fast and accurate queries; or using fuzzy search based on similarity metrics to calculate the similarity between the current environmental conditions and the environmental condition tags in the historical data set, and selecting records with similarity scores above a threshold. Alternatively, data structures such as hash tables or B-trees can be used to optimize the storage and retrieval efficiency of the historical data set.

[0126] Step S223: Extract the historical deviation values ​​and historical fluctuation patterns corresponding to the current environmental conditions from the long-term mutual calibration history of the multi-heterogeneous motor models.

[0127] This step aims to extract statistical characteristics or temporal evolution patterns of the historical deviations between the predicted and measured joint torque values ​​of each heterogeneous motor model during its past trouble-free operation, from the retrieved long-term cross-calibration history of multiple heterogeneous motor models matched to the current environmental conditions. Its purpose is to provide an accurate historical benchmark for the dynamic confidence weighting of real-time deviation values. Possible implementation methods include directly reading pre-calculated and stored statistical parameters such as the mean, standard deviation, and variance of the deviations of each heterogeneous motor model under specific environmental conditions from the long-term cross-calibration history of the multiple heterogeneous motor models.

[0128] Step S224: Extract the historical prediction consistency distribution features corresponding to the current environmental conditions from the long-term mutual calibration history of the multi-heterogeneous motor model.

[0129] This step aims to obtain the statistical distribution characteristics of the consistency among the historical joint torque prediction values ​​of multiple heterogeneous motor models during past fault-free operation from the retrieved long-term cross-calibration history of multiple heterogeneous motor models. Its purpose is to provide a normal historical reference distribution for real-time deviation value consistency assessment. Possible implementation methods include directly reading the probability density function parameters or cumulative distribution functions of the historical prediction consistency coefficients of each heterogeneous motor model under specific environmental conditions, which are pre-calculated and stored from the long-term cross-calibration history of multiple heterogeneous motor models.

[0130] Through steps S221-S224, the problem of mismatch between the long-term mutual calibration historical state of multiple heterogeneous motor models and the current environmental conditions caused by dynamic changes in environmental conditions is effectively solved. Specifically, by comprehensively utilizing real-time joint motion data, ambient temperature data, humidity data, and tunnel section markers, the current environmental conditions of the quadruped robot can be identified. Based on this, the long-term mutual calibration historical state of multiple heterogeneous motor models that highly matches the current environmental conditions is intelligently retrieved from the historical data record set, thereby ensuring that the historical deviation values, historical fluctuation patterns, and historical prediction consistency distribution characteristics of the retrieved heterogeneous motor models are specific to the current environmental conditions. This makes fault feature extraction no longer limited to a single, static historical reference, but can dynamically adapt to changes in environmental conditions, improving the accuracy and reliability of fault feature extraction. By providing historical references that accurately match the current environmental conditions, the system can more effectively distinguish between real fault signals and model prediction fluctuations caused by environmental factors, thereby reducing false alarm rates and false negative rates, and enhancing the robustness and practicality of the fault early warning system.

[0131] Step S23: Calculate the deviation between the predicted torque values ​​of each joint (i.e., the current set of predicted joint torque values, including the predicted torque values ​​of the first joint, the second joint, and the third joint) and the actual measured torque values ​​of the joints at the current moment (i.e., the current measured torque values), determine multiple real-time deviation values ​​(including the first deviation value, the second deviation value, and the third deviation value), obtain the current deviation value set, and perform confidence weighting on each real-time deviation value according to the historical fluctuation pattern of the historical deviation values ​​to generate multi-source deviation features calibrated by historical states (i.e., the deviation feature set, including the first deviation feature, the second deviation feature, and the third deviation feature).

[0132] This step aims to quantify the difference between the predicted results and actual observations for each heterogeneous motor model. Deviation calculation can employ various mathematical methods, such as direct subtraction (predicted value - measured value), calculating absolute error, relative error, or squared error. For each heterogeneous motor model, its output predicted joint torque is compared with the actual measured joint torque directly measured by a sensor (such as a torque sensor) at the same time, thus obtaining the real-time deviation value for that heterogeneous motor model. In some cases, deviation calculation can also consider time series lags or leads, for example, by aligning the predicted and measured value sequences using methods such as Dynamic Time Warping (DTW) before calculating the deviation, to address any potential minor time synchronization errors.

[0133] This process generates multi-source deviation features calibrated based on historical state by weighting real-time deviation values ​​according to their historical fluctuation patterns. This involves assigning different weights to multiple current real-time deviation values ​​based on the inherent deviation characteristics exhibited by each heterogeneous motor model during historical fault-free operation, reflecting the reliability of their anomaly levels. Confidence weighting can be determined based on the historical fluctuation patterns of real-time and historical deviation values ​​(e.g., the historical deviation mean and standard deviation). For example, if a real-time deviation value falls within the historical normal fluctuation range (e.g., the historical deviation mean plus or minus twice the historical deviation standard deviation), a higher dynamic confidence level is assigned. If the deviation is large, a lower dynamic confidence level is assigned. The weighting function can be a Gaussian function, a Sigmoid function, or other nonlinear mapping functions, whose output value represents the reliability or anomaly level of the real-time deviation value. Another approach is based on the probability density distribution of the historical deviation value set. Real-time deviation values ​​can be mapped to their cumulative probability or probability density value in the probability density distribution of the historical deviation value set, which can then be used as the dynamic confidence level. In this way, the inherent deviation characteristics of each heterogeneous motor model under different environmental conditions can be taken into account, making the generated deviation characteristics more robust and reducing false alarms.

[0134] The above step S23 proposes to obtain the current fault characteristics based on deviation calculation and consistency assessment. However, in its implementation process, the real-time deviation value may be affected by instantaneous noise or anomalies, and the historical fluctuation pattern is not fully considered, resulting in the generated deviation characteristics being not robust enough and unable to effectively distinguish between real faults and environmental interference.

[0135] To address this, a method is further proposed that uses confidence weighting of each real-time deviation value based on the historical fluctuation pattern of the historical deviation value to generate multi-source deviation features calibrated by historical states. Specifically, this method includes:

[0136] Step S231: Extract the mean and standard deviation of the historical deviation of each heterogeneous motor model under the current environmental conditions from the historical fluctuation pattern of the historical deviation value.

[0137] This step aims to provide a quantified historical normal fluctuation range benchmark for real-time deviation values. By analyzing the historical deviation value set data of each heterogeneous motor model operating without failure under specific environmental conditions, its average deviation level (which can be represented by the historical deviation mean) and the dispersion of the deviation (which can be represented by the historical deviation standard deviation) can be statistically determined. These statistics reflect the inherent characteristics and environmental adaptability of the heterogeneous motor model under normal operating conditions. In practice, this can be achieved by filtering historical data based on environmental condition labels (such as ambient temperature, humidity, load, etc.), and then calculating the arithmetic mean of the historical deviation value set sequence (e.g., within the first historical time period) for each filtered heterogeneous motor model as the historical deviation mean, and calculating its standard deviation as the historical deviation standard deviation. Alternatively, a sliding window average and standard deviation calculation can be used to dynamically statistically analyze the historical data to better adapt to subtle changes in environmental conditions, or a Gaussian Mixture Model (GMM) can be used to model the historical deviation distribution and extract its mean and standard deviation parameters.

[0138] Step S232: For each heterogeneous motor model, calculate the difference between the real-time deviation value and the historical average deviation value to obtain the current difference set. Then, calculate the ratio between the current difference set and the historical standard deviation to obtain the deviation multiple for each heterogeneous motor model, and obtain the deviation multiple set.

[0139] This step quantifies the degree of deviation of the current real-time deviation value relative to the historical normal fluctuation range. By calculating the difference between the real-time deviation value and the historical mean deviation, the direction and magnitude of the deviation can be understood. Then, comparing this difference with the historical standard deviation standardizes the deviation, making it comparable and intuitively reflecting whether the real-time deviation value exceeds the typical range of historical normal fluctuations. In implementation, for each heterogeneous motor model, its current real-time deviation value is acquired, and then the difference is obtained by direct subtraction. The absolute value of this difference is then divided by the corresponding historical standard deviation to obtain the deviation multiple. Alternatively, statistical methods such as Z-score or T-score can be used to calculate the deviation multiple, where Z-score = (real-time deviation value - historical mean deviation) / historical standard deviation. This method not only considers the degree of deviation but also retains the direction (positive or negative).

[0140] Step S233: Based on the deviation multiple corresponding to each heterogeneous motor model, a dynamic confidence level is assigned to each real-time deviation value through a pre-set confidence mapping function to obtain a dynamic confidence level set, wherein the dynamic confidence level decreases monotonically as the deviation multiple increases.

[0141] The core of this step lies in dynamically adjusting the weight of the real-time deviation value in subsequent fusion based on its degree of anomaly (measured by the deviation factor). The larger the deviation of the real-time deviation value from the historical normal range (i.e., the larger the deviation factor), the lower the probability that it is a genuine fault signal (or the higher the probability that it is considered noise), thus assigning a lower dynamic confidence level. Conversely, the smaller the deviation, the higher the dynamic confidence level. This dynamic adjustment helps suppress the influence of transient noise and sporadic anomalies. In implementation, the preset confidence mapping function can be a Sigmoid function or a Gaussian function, with the deviation factor as input and a dynamic confidence value between 0 and 1 as output. For example, when the deviation factor is close to 0, the dynamic confidence level is close to 1. As the deviation factor increases, the dynamic confidence level monotonically decreases and approaches 0. Alternatively, a piecewise linear function or an exponentially decaying function can be used as the confidence mapping function. For example, a threshold can be set; when the deviation factor is less than this threshold, the dynamic confidence level is 1. When it exceeds this threshold, the dynamic confidence level decays exponentially or linearly.

[0142] Step S234: Perform a dot product operation between each real-time deviation value and its corresponding dynamic confidence level to obtain multiple weighted deviation components calibrated with historical fluctuation range.

[0143] This step directly applies the previously calculated dynamic confidence level to the real-time deviation value, thereby "calibrating" the real-time deviation. Through dot product operations, the contribution of real-time deviation values ​​with low dynamic confidence levels (i.e., those deviating significantly from the historical normal range or considered noise) to the current fault characteristics is weakened, while real-time deviation values ​​with high dynamic confidence levels (i.e., fluctuating within the historical normal range) retain their original information, thus generating more robust weighted deviation components that better reflect the true fault state. In implementation, for each heterogeneous motor model, its current real-time deviation value is directly multiplied by its corresponding dynamic confidence level to obtain the weighted deviation components for that heterogeneous motor model. Alternatively, the sign of the real-time deviation value can be processed before dot product operations to ensure that the weighting operation does not change its positive or negative direction, only adjusting its magnitude, to better reflect the actual physical meaning of the deviation.

[0144] Step S235: Multiple weighted deviation components are spliced ​​together in a fixed order according to multiple heterogeneous motor models to obtain multi-source deviation characteristics.

[0145] This step aims to integrate weighted deviation components from various heterogeneous motor models, calibrated for historical fluctuation ranges, into a unified multi-source deviation feature vector. By concatenating them in a fixed order, the dimensionality and semantic consistency of the multi-source deviation feature vector are ensured, facilitating subsequent fault feature analysis and processing. This concatenated vector contains deviation information from multiple heterogeneous motor models at the current moment, calibrated with dynamic confidence levels, and is a multi-source deviation feature incorporating historical state references. In implementation, the order of the heterogeneous motor models is predefined (e.g., the weighted deviation components of the first heterogeneous motor model first, followed by the second), and then the weighted deviation components of each heterogeneous motor model are sequentially arranged and combined into a one-dimensional multi-source deviation feature vector. Alternatively, the weighted deviation components can be concatenated with the corresponding heterogeneous motor model ID or type code to form a more semantically informative multi-source deviation feature, or a matrix format can be used, where rows represent heterogeneous motor models and columns represent weighted deviation components.

[0146] Through steps S231-S235, the problem of real-time deviation values ​​being susceptible to transient noise or anomalies, resulting in insufficient robustness of current fault characteristics, is effectively solved. Specifically, by extracting the historical deviation mean and standard deviation from the historical fluctuation patterns of historical deviation values, a quantified historical normal fluctuation range benchmark is provided for the real-time deviation values. Based on this, the difference between the real-time deviation value and the historical deviation mean is calculated and compared with the historical deviation standard deviation to obtain the deviation factor. This allows the abnormality of the real-time deviation value to be standardized and objectively measured. Furthermore, based on this deviation factor, a dynamic confidence level is assigned to each real-time deviation value through a preset confidence mapping function. The dynamic confidence level monotonically decreases as the deviation factor increases, enabling the system to intelligently identify and mitigate the impact of real-time deviation values ​​that may be caused by transient noise or occasional anomalies and are severely inconsistent with the historical normal fluctuation range. Each real-time deviation value is multiplied by its corresponding dynamic confidence level to generate a weighted deviation component calibrated with the historical fluctuation range, and these components are then concatenated to form a multi-source deviation feature. This processing method not only integrates the prediction information of multiple heterogeneous motor models into the generated current fault features, but more importantly, it effectively filters out noise interference by introducing the historical fluctuation patterns of historical deviation values ​​for confidence weighting. This improves the robustness and accuracy of the current fault features, thereby enabling a more reliable distinction between real faults and environmental interference, and providing a more solid foundation for subsequent fault diagnosis and early warning.

[0147] Step S24: Calculate the dispersion of multiple joint torque prediction values ​​in the current joint torque prediction value set to obtain the real-time prediction consistency coefficient (current prediction consistency coefficient).

[0148] This step aims to assess the consistency or dispersion of the predicted joint torques from different heterogeneous motor models at the current time. Dispersion can be calculated using various statistical metrics, such as variance, standard deviation, range, coefficient of variation, or interquartile range. For the predicted joint torques output by multiple heterogeneous motor models at the same time, the dispersion of these predictions is calculated to quantify their level of consistency. Lower dispersion indicates high consistency among the heterogeneous motor models, while higher dispersion indicates significant discrepancies. Alternatively, methods based on information entropy or fuzzy set theory can be used to assess consistency. For example, the predicted joint torques of each heterogeneous motor model can be treated as a fuzzy set, and the similarity or overlap between these fuzzy sets can be calculated to obtain a coefficient reflecting overall consistency.

[0149] Step S24 above proposes to calculate the degree of dispersion of multiple joint torque prediction values ​​in the current joint torque prediction value set to obtain the real-time prediction consistency coefficient. However, in its implementation process, due to the lack of historical consensus as a reference benchmark, the consistency assessment may be inaccurate, easily affected by environmental interference, and unable to effectively distinguish between real faults and normal fluctuations.

[0150] To address this, a method is further proposed to calculate the dispersion of multiple joint torque prediction values ​​in the current joint torque prediction value set, thereby obtaining a real-time prediction consistency coefficient. Specifically, this includes:

[0151] Step S241: Obtain the predicted value of the first joint torque output by the first heterogeneous motor model, the predicted value of the second joint torque output by the second heterogeneous motor model, and the predicted value of the third joint torque output by the third heterogeneous motor model.

[0152] This step aims to provide multi-source, heterogeneous input data for subsequent discreteness calculations. By integrating prediction results from heterogeneous motor models with different principles or architectures (e.g., physical mechanism models, time-series learning models, lookup table models), the inherent limitations or errors of a single model can be avoided, thus laying a more solid data foundation for consistency evaluation. These historical joint torque predictions can be generated in real time by the various heterogeneous motor models running in parallel and directly provided to the consistency evaluation module, or obtained from the models through a data aggregation layer.

[0153] Step S242: Determine the historical consensus center value that matches the current environmental conditions. The historical consensus center value is used to characterize the equilibrium point among the historical joint torque prediction values ​​of each heterogeneous motor model during the historical fault-free operation period (e.g., the second historical time period).

[0154] This step aims to introduce a "consensus" based on historical fault-free data as an evaluation benchmark. The historical consensus center value is not a simple average, but rather an ideal state or equilibrium point that the historical joint torque predictions of various heterogeneous motor models should approach under specific environmental conditions, obtained through long-term cross-calibration and statistical analysis. This value can be pre-stored in a database and retrieved based on real-time environmental conditions (such as temperature, humidity, tunnel sections, etc.). Alternatively, it can be dynamically generated based on historical data and current environmental conditions using a lightweight prediction model.

[0155] In step S242, it is proposed to determine the historical consensus center value as a reference for calculating the real-time prediction consistency coefficient. However, in its implementation, due to the dynamic changes in environmental conditions and motion phase, the accuracy of the determined historical consensus center value is insufficient, which in turn affects the reliability of the current fault feature extraction.

[0156] In response, a method for determining historical consensus central values ​​that match current environmental conditions is further proposed, specifically including:

[0157] Step S2421: Obtain the current joint angle and current joint angular velocity contained in the real-time joint motion data, and determine the current motion phase of the quadruped robot based on the current joint angle and current joint angular velocity.

[0158] This step aims to accurately identify the current motion state of the quadruped robot, providing contextual information for subsequent historical data matching. One approach is to pre-define a motion phase state machine, which divides the quadruped robot's motion into different discrete phases (e.g., swinging phase, supporting phase, transition phase) based on specific thresholds or combinations of joint angles and angular velocities. Another approach is to train a machine learning-based model, such as a recurrent neural network (RNN) or a hidden markov model (HMM), which can learn and identify the current motion phase from real-time current joint angles and angular velocities.

[0159] Step S2422: Retrieve the historical time window (second historical time period) that matches the current environmental conditions and the current motion phase.

[0160] This step aims to filter out a subset of historical data that is highly correlated with the current environmental conditions and the current motion phase, ensuring the validity of historical consensus center values. One approach is to construct a multi-dimensional index structure, such as a KD-tree or R-tree, where historical data is indexed as a combination of environmental conditions (e.g., temperature, humidity, tunnel segment) and motion phase. The retrieval process will perform efficient nearest neighbor searches or range queries within this index structure by querying the current environmental conditions and the current motion phase. Another approach is to define a comprehensive similarity metric function that combines the distance of environmental parameters (e.g., Euclidean distance) and the degree of matching of motion phases (e.g., classification matching or phase distance), and then retrieve historical time windows where the comprehensive similarity exceeds a preset threshold.

[0161] Step S2423: Extract the historical joint torque prediction value sequence output by the first heterogeneous motor model, the second heterogeneous motor model, and the third heterogeneous motor model during the historical fault-free period from the historical time window.

[0162] This step aims to obtain historical joint torque prediction values ​​generated by various heterogeneous motor models within a historical time window, serving as the raw data for constructing the historical consensus center value. One implementation involves including references to specific data blocks stored in a local or cloud database within the metadata of the historical time window. The system then directly reads the historical joint torque prediction values ​​for each heterogeneous motor model within the corresponding time period based on these references. Another implementation involves dynamically extracting the required historical joint torque prediction values ​​from the database by sending a query request with the time range, heterogeneous motor model identifier, and fault-free label to the data storage system.

[0163] Step S2424: Calculate the median of the historical joint torque prediction values ​​of the first heterogeneous motor model, the second heterogeneous motor model, and the third heterogeneous motor model at the same time to obtain the instantaneous median value corresponding to each time. Then, perform a moving average filter on the instantaneous median value corresponding to each time to obtain the historical consensus center value corresponding to the current motion phase.

[0164] The median calculation aims to reduce the impact of outliers predicted by individual heterogeneous motor models on the historical consensus center value, providing a more robust measure of central tendency. One implementation involves collecting historical joint torque predictions from three heterogeneous motor models for each point in time within the historical time window, sorting these three values, and taking the median as the instantaneous median for that moment. Another implementation utilizes a specialized median lookup algorithm, such as the fast selection algorithm, to calculate the instantaneous median value for each moment more efficiently. Moving average filtering is used to smooth the sequence of instantaneous median values ​​within the historical time window, eliminating short-term fluctuations and thus obtaining a more stable and representative historical consensus center value. One implementation uses a simple moving average filter to average the instantaneous median values ​​within a preset window length. Another implementation uses an exponential moving average (EMA) filter, which assigns higher weight to recent data, maintaining smoothness while being more sensitive to potential trend changes.

[0165] Through steps S2421-S2424, the problem of insufficient accuracy of historical consensus center values ​​caused by dynamic changes in environmental conditions and motion phases can be effectively solved. By acquiring the current joint angle and angular velocity in real time to accurately determine the current motion phase, the retrieval of historical data can fully consider the robot's current specific motion mode, avoiding errors introduced by motion mismatch. When retrieving historical time windows, matching is performed using both current environmental conditions and current motion phase, greatly improving the relevance and representativeness of the selected historical data, ensuring that the historical consensus center value is determined based on fault-free historical data highly consistent with the current operating environment and motion mode. The median of the historical joint torque prediction values ​​for multiple heterogeneous motor models is calculated, effectively suppressing the interference of prediction bias or outliers that may exist in a single heterogeneous motor model on the consensus result, enhancing the robustness of instantaneous consensus. Furthermore, a moving average filter is used to smooth the instantaneous median value, eliminating short-term noise and fluctuations, generating a stable and accurate historical consensus center value corresponding to the current motion phase. This historical consensus center value, which has undergone refined matching and robust processing, serves as a reference benchmark for calculating the real-time prediction consistency coefficient. It can more accurately reflect the degree of consistency or deviation between the current set of joint torque prediction values ​​and the historical normal pattern, thereby improving the reliability and accuracy of the current fault feature extraction.

[0166] Step S243: Calculate the first deviation distance between the predicted torque value of the first joint and the historical consensus center value, the second deviation distance between the predicted torque value of the second joint and the historical consensus center value, and the third deviation distance between the predicted torque value of the third joint and the historical consensus center value, respectively, to obtain the deviation distance set.

[0167] This step aims to quantify the deviation of the real-time joint torque prediction value of each heterogeneous motor model from its historical normal behavior. By calculating the distance between the joint torque prediction value of each heterogeneous motor model and the historical consensus center value, it is possible to independently assess whether the current performance of each heterogeneous motor model conforms to the "normal" pattern shown in its fault-free historical data. This deviation distance can be calculated using various mathematical metrics such as Euclidean distance, absolute error, and relative error; the specific choice can be adjusted according to the actual application scenario and the requirements for error sensitivity.

[0168] Step S244: Weighted fusion of the first deviation distance, the second deviation distance, and the third deviation distance to generate a real-time prediction consistency coefficient with the historical consensus center value as a reference.

[0169] This step aims to integrate the independent deviation information of various heterogeneous motor models into a comprehensive index. Weighted fusion allows assigning different weights to the deviation distances of each heterogeneous motor model based on its characteristics, reliability, or performance under specific environmental conditions. For example, under certain environmental conditions, the physical mechanism model may be more reliable, and its deviation distance can be given a higher weight. This weighted fusion can be implemented through linear weighted summation, nonlinear function mapping, or machine learning-based fusion algorithms to generate a single real-time prediction consistency coefficient. This coefficient comprehensively reflects the overall consistency level of the current set of joint torque predictions from the current multi-heterogeneous motor models relative to the historical consensus center value.

[0170] Through steps S241-S244, the historical consensus center value is introduced as a reference benchmark, improving the stability and reliability of the real-time prediction consistency coefficient and solving the problem of inaccurate consistency assessment under changing environmental conditions. Specifically, the joint torque prediction values ​​of multiple heterogeneous motor models are obtained, providing a multi-source data foundation for the assessment and avoiding misjudgments caused by the deviation of a single model. A historical consensus center value matching the current environmental conditions is determined, establishing a reliable equilibrium point reference, ensuring historical continuity in the consistency assessment and enhancing its anti-interference capability in dynamic environments. The deviation distance between each joint torque prediction value and the historical consensus center value is calculated separately, quantifying the deviation of each heterogeneous motor model's joint torque prediction value relative to the historical normal state, rather than directly comparing the differences between heterogeneous motor models, reducing noise impact and improving the accuracy of deviation detection. The deviation distances are weighted and fused, comprehensively considering the prediction characteristics of each heterogeneous motor model, to generate a comprehensive real-time prediction consistency coefficient. This coefficient, with the historical consensus center value as a reference benchmark, more realistically reflects the degree of consistency between the current state and the historical normal mode, thus providing a more robust input for subsequent current fault characteristic analysis. This evaluation method based on historical consensus central values ​​enables the system to effectively distinguish between normal fluctuations caused by environmental changes and abnormal deviations caused by actual faults, thereby improving the accuracy and robustness of fault warnings.

[0171] Step S25: Substitute the real-time prediction consistency coefficient into the target membership function defined by the historical prediction consistency distribution characteristics, calculate the membership degree of the current prediction consistency state relative to the historical normal pattern, and obtain the consistency deviation characteristics.

[0172] This step aims to compare the current level of consistency predicted among heterogeneous motor models with the normal consistency pattern under historical fault-free conditions to quantify the degree of anomaly in the current state. Historical prediction consistency distribution characteristics can include a probability density function (PDF) or a cumulative distribution function (CDF) (i.e., a statistical distribution), which describes the distribution of historical prediction consistency coefficients during historical fault-free periods. The target membership function can be constructed based on this historical distribution; for example, defining the central region of the historical distribution as high membership and the tails far from the center as low membership. Inputting the real-time prediction consistency coefficients into the target membership function yields the degree to which they belong to the historical normal pattern. The target membership function can also be a trigonometric function, trapezoidal function, or Gaussian function based on fuzzy logic, with parameters set according to the statistical characteristics (such as mean, standard deviation, and upper and lower limits) of the statistical distribution in the historical prediction consistency distribution characteristics. By calculating the output value of the real-time prediction consistency coefficients on the target membership function, the similarity or deviation between the current consistency state and the historical normal pattern can be quantified, thereby generating consistency deviation characteristics.

[0173] Step S25 above proposes to evaluate the deviation of the current predicted consistency state from the historical normal pattern by calculating the real-time prediction consistency coefficient and substituting it into the target membership function, thereby generating consistency deviation features for current fault feature extraction. However, in its implementation, due to the dynamic changes in environmental conditions and motion states, a fixed target membership function may not be able to adapt to the evaluation needs under different environmental conditions, resulting in the evaluation results being distorted by motion interference. At the same time, relying solely on the real-time prediction consistency coefficient itself ignores the overall influence of the overall deviation direction of multi-source deviation features, which may fail to fully capture fault tendencies, leading to misjudgment or missed detection.

[0174] To address this, a further proposed method is to substitute the real-time prediction consistency coefficient into the target membership function defined by the historical prediction consistency distribution characteristics, calculate the membership degree of the current prediction consistency state relative to the historical normal pattern, and obtain the consistency deviation characteristics. Specifically, this method includes:

[0175] Step S251: Based on the historical prediction consistency distribution characteristics, determine the historical consistency probability density function that matches the current environmental conditions, and use it as the benchmark form of the target membership function to obtain the benchmark membership function.

[0176] This step aims to provide a statistical benchmark for constructing the target membership function based on historical fault-free data and current environmental conditions. The historical consistency probability density function characterizes the statistical distribution of the dispersion of historical joint torque predictions among multiple heterogeneous motor models under specific environmental conditions. Its role is to ensure that subsequent evaluations of real-time prediction consistency can refer to a "normal" pattern consistent with current environmental conditions, thereby avoiding misjudgments caused by environmental changes. In practice, various methods can be employed. For example, the statistical data on the dispersion of predictions from each heterogeneous model during historical fault-free periods can be parametrically fitted, such as to a Gaussian, beta, or gamma distribution, to determine the historical consistency probability density function. Alternatively, non-parametric methods, such as kernel density estimation (KDE), can be used to directly estimate the historical consistency probability density function from historical data to capture more complex distribution patterns.

[0177] Step S252: Based on real-time joint motion data, determine the fluctuation amplitude of joint angular velocity, and then dynamically adjust the confidence interval boundary of the baseline membership function to generate an adaptive membership function (i.e., target membership function) associated with the current motion state.

[0178] This step aims to address the problem of evaluation distortion of fixed membership functions under dynamic motion conditions. Joint angular velocity fluctuation amplitude is a crucial indicator of the intensity and stability of quadruped robot motion. When a quadruped robot performs vigorous movements, even under normal conditions, significant fluctuations may occur between the prediction results of different heterogeneous motor models, causing the current prediction consistency coefficient to deviate from the historical average. By dynamically adjusting the confidence interval boundary of the baseline membership function, the "normal" consistency range can be adaptively expanded or contracted according to the motion state. For example, when the joint angular velocity fluctuation amplitude is large, the confidence interval can be appropriately widened to allow for greater consistency fluctuations. Conversely, when the motion is stable, the confidence interval is tightened to improve sensitivity to minor anomalies. Specifically, a scaling factor positively correlated with the joint angular velocity fluctuation amplitude can be designed. This factor is used to adjust the standard deviation or quantile range of the baseline membership function, thereby dynamically changing the shape and boundary of the baseline membership function. Another approach is to directly adjust the predefined confidence threshold in the baseline membership function based on different joint angular velocity fluctuation amplitude ranges using preset fuzzy rules or lookup table mechanisms.

[0179] Step S253: Input the real-time prediction consistency coefficient into the adaptive membership function for mapping to obtain the initial membership degree of the current prediction consistency state belonging to the historical normal mode.

[0180] This step is crucial for assessing the "normality" of the real-time prediction consistency state. The real-time prediction consistency coefficient quantifies the degree of similarity between the prediction results of various heterogeneous motor models at the current moment. This coefficient is input into the aforementioned generated adaptive membership function, which outputs a value between 0 and 1 based on its defined probability distribution or fuzzy rules; this is the initial membership degree. A higher value indicates that the current prediction consistency state is closer to the historical normal pattern. Conversely, a lower value indicates a greater deviation. For example, if the adaptive membership function is a probability density function, the initial membership degree can be the probability density value of the real-time prediction consistency coefficient on that function. If it is a fuzzy membership function, the initial membership degree represents the degree to which the real-time prediction consistency coefficient belongs to the "normal consistency" set.

[0181] Step S254: Obtain the positive and negative signs of each weighted deviation component in the multi-source deviation feature, and correct the initial membership degree according to the overall deviation direction indicated by the positive and negative signs to generate the target membership degree, and use the target membership degree as the consistency deviation feature.

[0182] This step introduces information on the overall deviation direction to compensate for the potential bias in evaluation that may result from relying solely on real-time prediction consistency coefficients. Each weighted deviation component in the multi-source deviation features carries a positive or negative sign, indicating whether the prediction result of the corresponding heterogeneous motor model is too high or too low relative to the current joint torque measurement. By analyzing these signs, it is possible to determine whether there is an overall deviation direction (i.e., the overall deviation direction), for example, all heterogeneous motor models tending to over-predict or under-predict. This overall deviation direction is often a signal of early failures, even if the real-time prediction consistency coefficient itself is still within the "normal" range. When revising the initial membership, the initial membership can be appropriately adjusted based on the strength and consistency of the overall deviation direction. For example, if all heterogeneous motor models are found to consistently deviate in a certain direction (such as too high or too low), even if their consistency is acceptable, the initial membership can be appropriately reduced to reflect this potential abnormal trend. Specifically, the number or weighted sum of positive and negative weighted bias components can be counted. If the weighted bias component in a certain direction is dominant, a penalty factor is applied to the initial membership degree to reduce it, thereby generating a more diagnostically meaningful target membership degree.

[0183] Through steps S251-S254, the problems of inaccurate evaluation using traditional fixed membership functions and neglecting the overall deviation direction by relying solely on real-time prediction consistency coefficients in dynamically changing environmental conditions and motion states are effectively solved. By extracting a historical consistency probability density function matching the current environmental conditions from the historical prediction consistency distribution characteristics to determine the benchmark membership function, it is ensured that the benchmark membership function can be constructed for the specific current operating environment, avoiding the inadequacy of general membership functions under changing environmental conditions, thereby improving the accuracy of the evaluation. Dynamically adjusting the confidence interval boundary of the benchmark membership function based on the joint angular velocity fluctuation amplitude allows the benchmark membership function to adapt to changes in the robot's motion state in real time, effectively reducing evaluation deviations and false alarms caused by motion disturbances and enhancing the ability to capture instantaneous anomalies. By acquiring the positive and negative signs of each weighted deviation component in the multi-source deviation features and correcting the initial membership degree according to the overall deviation direction indicated by them, the trend information of the weighted deviation components is integrated into the consistency deviation features. This avoids the one-sidedness of unidirectional evaluation and enables the consistency deviation features to reflect the potential fault tendency of the system more comprehensively and in detail. As a result, the reliability and diagnostic accuracy of the current fault feature extraction of the quadruped robot are improved, providing a more solid foundation for subsequent fault warning and fault-tolerant control.

[0184] Step S26: Concatenate the multi-source deviation features and the consistency deviation features, and perform dimensionality reduction and normalization on the concatenated features to obtain the current fault features.

[0185] Feature concatenation of multi-source deviation features and consistency deviation features refers to combining two different types of fault indication information into a more comprehensive feature vector. Feature concatenation is a common feature fusion method, which involves concatenating two or more feature vectors in terms of dimensions. For example, if the multi-source deviation feature is an N-dimensional vector and the consistency deviation feature is an M-dimensional vector, then the concatenated feature vector will be an (N+M)-dimensional vector. This operation preserves the original information of both features and provides a more comprehensive input for subsequent analysis. Before concatenation, the two features can be preprocessed, such as by feature scaling or standardization, to ensure that they are comparable in numerical range and to avoid any one feature dimension contributing too much to the overall feature vector.

[0186] Dimensionality reduction and normalization of cascaded features to obtain the current fault features refers to optimizing the combined high-dimensional feature vector to remove redundant information, reduce computational complexity, and map it to a uniform numerical range for subsequent fault identification and analysis. Dimensionality reduction can employ techniques such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), t-SNE (t-Distributed Stochastic Neighbor Embedding), or autoencoders. These methods aim to extract the most important information from high-dimensional features, remove redundancy and noise, and thus obtain a more compact and representative low-dimensional feature vector. Dimensionality reduction helps improve the efficiency and robustness of subsequent fault classification or identification algorithms. Normalization typically maps the feature vector values ​​to a specific range, such as [0,1] or [-1,1], or performs Z-score standardization (mean 0, standard deviation 1). Common normalization methods include Min-Max normalization, Z-score normalization, or L2 norm normalization. Normalization helps eliminate dimensional differences between different feature dimensions and prevents certain features from dominating subsequent processing due to their large numerical range.

[0187] The above step S26 proposes to concatenate multi-source deviation features and consistency deviation features to generate current fault features. However, in this process, the concatenated feature vectors have high dimensionality, redundant information, unbalanced feature importance, and inconsistent numerical ranges, resulting in inaccurate extraction of current fault features and affecting the accuracy of subsequent fault warnings.

[0188] To address this, a method is further proposed that involves concatenating multi-source deviation features and consistency deviation features, and then performing dimensionality reduction and normalization on the concatenated features to obtain the current fault features. Specifically, this method includes:

[0189] Step S261: Concatenate the multi-source deviation features and the consistency deviation features along the feature dimension to generate an initial concatenated feature vector.

[0190] This step aims to integrate fault indication information from different sources. For example, the multi-source deviation feature could be a vector, where each component represents the weighted deviation between the predicted and actual results of a heterogeneous motor model, while the consistency deviation feature could be a scalar or vector characterizing the degree of consistency deviation between the predicted results of multiple heterogeneous motor models. This is achieved by directly concatenating the features along the feature dimension, for example, combining the deviation feature vector Vdev (length M) and the consistency deviation feature vector Vcon (length N) into [Vdev, Vcon], forming an initial concatenated feature vector of length M+N. Alternatively, a type identifier can be appended to each feature component, forming a structure of [(value1, type1), (value2, type2), ...], to clarify its source and semantics. This concatenation operation ensures that all potential fault information is considered, laying the foundation for subsequent comprehensive analysis.

[0191] Step S262: Calculate the variance contribution rate of each dimension in the initial cascaded feature vector, and select the principal component dimensions from the initial cascaded feature vector according to the variance contribution rate to obtain the dimensionality-reduced compressed feature vector.

[0192] This step aims to address the potential issues of excessive dimensionality and information redundancy in the initial concatenated feature vectors. Specifically, Principal Component Analysis (PCA) can be performed on historically collected datasets of initial concatenated feature vectors. By calculating the variance contribution rate of each principal component (i.e., the proportion of its eigenvalue to the sum of all eigenvalues), principal components with a cumulative variance contribution rate reaching a preset threshold (e.g., 95%) are selected as new feature dimensions. Then, the current initial concatenated feature vectors are projected onto these selected principal component spaces, resulting in dimensionality-reduced compressed feature vectors. Alternatively, the variance of each dimension in the initial concatenated feature vectors can be calculated independently, and the vectors can be sorted according to variance. The top K dimensions with the largest variances are selected as principal component dimensions, and the corresponding eigenvalues ​​are directly extracted to form the compressed feature vectors. This dimensionality reduction effectively removes redundant information and noise from the initial concatenated feature vectors, reduces the feature dimensionality, thereby reducing the computational burden of subsequent processing and improving the robustness of the features.

[0193] Step S263: Obtain the dynamic confidence of each weighted bias component in the multi-source bias feature, and adaptively weight the corresponding dimension in the compressed feature vector according to the target confidence.

[0194] This step introduces a target confidence weight adjustment mechanism to reflect the reliability and importance of different feature components under the current environmental conditions. The target confidence is calculated based on the deviation of each weighted bias component from its historical normal fluctuation range. The higher the target confidence, the more likely the corresponding weighted bias component is to indicate a true fault. In implementation, for each dimension of the compressed feature vector derived from multi-source bias features, its value can be directly multiplied by its corresponding target confidence, i.e., feature_weighted = feature_original × dynamic_confidence, where feature_weighted is the weighted bias component, feature_original is the weighted bias component, and dynamic_confidence is the target confidence. Alternatively, a confidence mapping function can be designed, such as weight_factor = 1 + alpha × dynamic_confidence, where weight_factor is the adjusted target confidence, alpha is the adjustment coefficient, and then the corresponding dimension of the compressed feature vector is multiplied by this weight_factor. This adaptive weighting mechanism can dynamically adjust the importance of each weighted bias component in the current fault characteristics based on the historical fluctuation patterns and target confidence levels. This strengthens more reliable and indicative feature dimensions, while appropriately weakening feature dimensions that are more affected by noise or uncertainty. This balances the distribution of feature importance and improves the ability to distinguish current fault characteristics.

[0195] Step S264: Map the adaptively weighted compressed feature vector to a unified numerical range to obtain the current fault features.

[0196] This step aims to eliminate differences in units and numerical ranges between different feature dimensions, ensuring that all feature dimensions have an equal chance of contributing in subsequent processing. Specifically, a Min-Max normalization method can be used to linearly scale the data for each feature dimension to the interval [0,1], i.e., x_normalized = (x - min_val) / (max_val - min_val), where min_val and max_val are the minimum and maximum values ​​of the feature dimension in historical data, x_normalized is the normalized feature dimension, and x is the feature dimension before normalization. Another method is Z-score normalization, which transforms the data for each feature dimension into a distribution with a mean of 0 and a standard deviation of 1, i.e., x_normalized = (x - mean_val) / std_val, where mean_val and std_val are the mean and standard deviation of the feature dimension in historical data. By normalizing, the differences in dimensions and numerical ranges between different feature dimensions are eliminated, ensuring that all feature dimensions have an equal chance of contributing to the subsequent fault identification or classification model. This avoids the problem of certain feature dimensions dominating the decision-making process due to their large values, and further enhances the stability and comparability of the current fault features.

[0197] Through steps S261-S264, a current fault feature with moderate dimensionality, rich information, balanced importance, and stable values ​​can be generated. This improves the accuracy and robustness of the current fault feature, providing high-quality input for subsequent cross-physics semantic fusion and fault semantic description generation based on the current fault feature. This, in turn, enhances the accuracy and reliability of fault early warning for quadruped robots in complex environments such as cable tunnels.

[0198] Through steps S21-S26, the robustness and accuracy of current fault characteristics in complex environments can be enhanced by fusing long-term cross-calibration historical states of multiple heterogeneous models, thus solving the problem that current fault characteristics may be affected by environmental interference. Real-time joint motion data is input in parallel to multiple pre-trained heterogeneous motor models to output joint torque prediction values. Long-term cross-calibration historical states of multiple heterogeneous models matching the current environmental conditions are retrieved. This ensures the diversity and environmental relevance of predictions through parallel processing of multiple models and retrieval of historical states, providing a stable foundation for subsequent analysis and avoiding the amplification of errors from a single model. The deviation between the predicted and actual measured joint torque values ​​is calculated to obtain real-time deviation values. Based on the historical fluctuation patterns of these deviation values, confidence-weighted multi-source deviation features are generated. This allows for the calibration of real-time deviation values ​​using historical fluctuation patterns, reducing the impact of environmental noise and improving the reliability of the deviation features. The real-time prediction consistency coefficient is obtained by calculating the dispersion of multiple joint torque prediction values. This coefficient is then substituted into a target membership function defined based on historical prediction consistency distribution characteristics to calculate the membership degree, thus obtaining the consistency deviation feature. This quantifies the deviation of model consistency from historical normal patterns, enhancing the sensitivity of fault detection and preventing misjudgments. The multi-source deviation feature and the consistency deviation feature are concatenated and normalized to obtain the fault feature, integrating deviation and consistency information to generate a more compact and standardized feature representation, facilitating subsequent fault semantic description. Overall, these steps effectively distinguish between real faults and environmental interference, improving the accuracy and robustness of current fault feature extraction and providing a solid foundation for early warning of quadruped robots in complex environments such as cable tunnels.

[0199] Step S3: Based on the current fault characteristics, multiple physical field sensing data, and the historical deviation values ​​and historical fluctuation patterns of various heterogeneous motor models, cross-physical field semantic fusion is carried out through the target knowledge graph to generate a fault semantic description that integrates multi-source verification information and model historical behavior reference. The target knowledge graph includes fault mode nodes, physical field feature nodes, intermediate state nodes, and the relationship between the three types of nodes.

[0200] The timing guidance signal can be determined by observing the direction of change of fault features in the most recent time steps (i.e., the fourth historical time period), such as the trend of continuous increase or decrease in the feature values. The target knowledge graph can be a simple, pre-constructed graph containing a small number of fault mode nodes, physical field feature nodes, and intermediate state nodes, as well as the pre-defined direct relationships between them. Based on the timing guidance signal, such as continuous increase in deviation, physical field feature nodes (e.g., "abnormal vibration") and fault mode nodes (e.g., "motor wear") associated with "continuous increase" can be directly searched from the target knowledge graph to form a simple fault evolution path. Simple time-domain feature extraction (e.g., calculating mean and peak values) is performed on various physical field sensing data, and the data is arranged according to the order of nodes in the fault evolution path to obtain the target multi-physics synchronization feature sequence. Each feature segment (i.e., physical field synchronization feature) in the target multi-physics synchronization feature sequence is matched with the text description of the corresponding node in the fault evolution path using keywords, and the semantic matching degree is calculated. These semantic matching degrees are then simply accumulated to form a preliminary fault semantic description.

[0201] Step S3 above proposes to generate a fault semantic description by guiding cross-physical field semantic fusion through target knowledge graph. However, in this process, there is a lack of effective utilization of the temporal evolution trend of fault features and extraction of fault evolution path, resulting in insufficient depth of multi-source information fusion and inability to reliably identify fault type and location.

[0202] To address this, a method is further proposed that, based on current fault characteristics, multiple physical field sensor data, and historical deviation values ​​and fluctuation patterns of various heterogeneous motor models, cross-physical field semantic fusion is guided by a target knowledge graph to generate a fault semantic description that integrates multi-source verification information and model historical behavior references. Specifically, this method includes:

[0203] Step S31: Based on the historical fault characteristics and current fault characteristics of the fourth historical time period, perform time-series evolution trend analysis on the fault characteristics to obtain the dynamic evolution direction of the fault characteristics in the fourth historical time period, and use the dynamic evolution direction as the time-series guidance signal for cross-physical field fusion.

[0204] Fault features are a quantitative representation of the potential fault states of a quadruped robot. They integrate information about the differences between heterogeneous motor models and the degree of deviation from historical normal states, serving as a preliminary basis for fault judgment. Temporal evolution trend analysis refers to analyzing fault features within a fourth historical time period to reveal their patterns and directions of change over time. Its role is to capture the dynamic process of a fault's development from its inception to its current state, rather than merely judging it at a single instant. For example, statistical methods such as moving averages, exponential smoothing, and Kalman filtering can be used to process the fault feature sequence, extracting its short-term and long-term trend components and calculating the slope or acceleration of the trend. Alternatively, temporal models such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or Transformers can be trained to learn the evolution patterns of fault features under normal and abnormal states and predict their future evolution direction. Furthermore, typical fault evolution patterns (such as slow rise, rapid fall, periodic fluctuations, etc.) can be predefined or learned from historical data, and then the fault feature sequence can be matched with these patterns to identify the current evolution trend. The dynamic evolution direction refers to the changing trend of fault characteristics within the fourth historical time period, such as whether the deviation continues to increase, decrease, fluctuate, or remain stable. It quantifies the "direction" and "speed" of fault characteristic changes over time. The time-series guidance signal encodes the aforementioned dynamic evolution direction and serves as guidance information for the subsequent cross-physics semantic fusion process. Its role is to ensure that the semantic fusion process can focus on the physical field information most relevant to the current fault evolution trend, thereby improving the targeting and accuracy of the fusion.

[0205] Step S31 above proposes to perform temporal evolution trend analysis based on historical fault characteristics and current fault characteristics of the fourth historical time period in order to generate a temporal guidance signal for cross-physical field fusion. However, in this process, due to environmental interference, the temporal analysis may not be accurate enough, resulting in inaccurate guidance signals and difficulty in effectively distinguishing between real faults and environmental interference, thereby affecting the reliability of subsequent cross-physical field semantic fusion.

[0206] To address this, a further proposal is made to analyze the temporal evolution trend of fault characteristics based on historical and current fault characteristics within the fourth historical time period, thereby obtaining the dynamic evolution direction of fault characteristics within the fourth historical time period. Specifically, this includes:

[0207] Step S311: Perform temporal evolution trend analysis based on the historical fault characteristics of the fifth historical time period to obtain the baseline temporal evolution trajectory. Fit the baseline evolution manifold under normal conditions based on the baseline temporal evolution trajectory. The quadruped robot is fault-free during the fifth historical time period.

[0208] This step aims to establish a stable reference baseline, unaffected by current environmental noise. The baseline temporal evolution trajectory refers to the sequence of fault characteristics changing over time during the normal operation of the quadruped robot. Extracting these trajectories can be achieved by dividing historical fault characteristic data into time windows and cleaning the data. For example, historical fault characteristic data marked as operating under fault-free conditions can be retrieved and filtered from local storage or cloud databases, and then reconstructed as a time series. Fitting an evolutionary baseline manifold based on the baseline temporal evolution trajectory aims to capture the intrinsic structure and evolutionary patterns of fault characteristics under normal conditions. One approach is to use dimensionality reduction techniques such as Principal Component Analysis (PCA) or Independent Component Analysis (ICA) to map high-dimensional historical fault characteristic temporal data to a low-dimensional space. Then, manifold learning algorithms (such as Isomap (Isometric Mapping) and LLE (Locally Linear Embedding)) are used to fit a manifold structure that represents the evolutionary pattern of normal fault characteristics. Another approach is to use Gaussian mixture models (GMM) or kernel density estimation (KDE) to model the probability distribution of the time-series trajectory of historical fault characteristics, and define the region with high probability density as the baseline manifold for deviation evolution under normal conditions.

[0209] The above step S311 proposes to determine the baseline time-series evolution trajectory for analyzing fault characteristics. However, in its implementation, if the baseline time-series evolution trajectory is not fitted accurately, it may not be able to accurately characterize the internal structure of fault characteristic evolution under normal conditions. It is easily affected by noise and anomalies, resulting in inaccurate dynamic evolution direction vector analysis, which in turn affects the reliability of subsequent fault warning.

[0210] To address this, a method is further proposed that uses historical fault characteristics from the fifth historical time period to analyze the temporal evolution trend, obtain a baseline temporal evolution trajectory, and fit the baseline evolutionary manifold under normal conditions based on the baseline temporal evolution trajectory. Specifically, this method includes:

[0211] Step S3111: Determine multiple historical evolution trajectories of the fault characteristics of the quadruped robot in multiple historical fault-free periods, and perform time axis alignment and amplitude normalization on the multiple historical evolution trajectories to obtain a set of aligned standardized trajectories.

[0212] This step aims to collect sufficient and representative historical normal fault feature data to comprehensively characterize the evolutionary behavior of fault features under fault-free conditions. By extracting multiple historical evolution trajectories from multiple historical fault-free periods, different operating conditions, environmental conditions, and time periods can be covered, thereby ensuring the robustness and generalization ability of the subsequent evolutionary baseline manifold construction. For example, after each fault-free operation task, the quadruped robot can upload its internally stored fault feature data to a cloud database and tag it as "fault-free," and then extract it in batches from the database according to the tag and time range. Alternatively, a continuously updated historical fault feature data buffer can be maintained in local storage. When the system confirms that the current operating state is fault-free, the latest fault feature data is appended to the buffer, and the fault feature data in the buffer is periodically filtered and archived.

[0213] Based on this, multiple historical evolutionary trajectories undergo time axis alignment and amplitude normalization to obtain an aligned standardized trajectory set. This step aims to eliminate temporal misalignment and numerical scale differences between different historical evolutionary trajectories caused by factors such as start time, sampling frequency, running speed, or environmental disturbances, ensuring that all trajectories are compared and analyzed under a unified spatiotemporal reference. Time axis alignment can employ the Dynamic Time Warping (DTW) algorithm, which minimizes the distance between two historical evolutionary trajectories by finding the optimal matching path, thereby achieving nonlinear time alignment. Amplitude normalization can use Min-Max normalization to linearly scale the data to the [0,1] interval. Alternatively, time axis alignment can also employ synchronization methods based on key event points (such as the start point of joint movement cycles or specific gait phase transition points) to align key event points of different historical evolutionary trajectories. Amplitude normalization can use Z-score normalization to convert the data into a distribution with a mean of 0 and a standard deviation of 1.

[0214] Step S3112: Reconstruct the phase space of each trajectory in the standardized trajectory set to obtain the phase point sequence of each trajectory in the high-dimensional phase space.

[0215] Phase space reconstruction is a nonlinear time series analysis technique aimed at recovering the dynamic evolution trajectory of a system in a multidimensional state space from one-dimensional observation data. By combining the current fault features at the current moment with historical fault features from several past moments into a high-dimensional vector (i.e., phase points), the inherent dynamic structure and potential chaotic characteristics of the system can be revealed, providing richer geometric information for subsequent manifold learning. For example, phase space reconstruction can be performed using the delayed embedding method guided by Takens' theorem, selecting appropriate embedding dimensions and time delays to reconstruct each trajectory into a high-dimensional phase point sequence. Alternatively, linear or nonlinear dimensionality reduction techniques such as principal component analysis (PCA) or independent component analysis (ICA) can be used to extract features from the original trajectory data within a sliding window and project them into a low-dimensional space, thereby forming a high-dimensional phase point sequence.

[0216] Step S3113: Perform local linear embedding on all phase point sequences in the high-dimensional phase space, extract low-dimensional manifold coordinates that can characterize the intrinsic structure of fault feature evolution, and obtain the initial evolutionary manifold.

[0217] Locally Linear Embedding (LLE) is a nonlinear dimensionality reduction algorithm. Its core idea is that each data point in high-dimensional data can be approximated by a linear combination of its nearest neighbors, and this local linear relationship is preserved after dimensionality reduction to a lower-dimensional manifold. LLE can effectively discover the low-dimensional nonlinear manifold structure embedded in high-dimensional space, thereby extracting low-dimensional coordinates that more fundamentally reflect the evolution of fault characteristics. Besides LLE, other manifold learning algorithms such as isomaps, Laplacian eigenmaps, or t-distributed random neighborhood embeddings (t-SNE) can also be used. These algorithms can similarly extract low-dimensional nonlinear manifold structures from high-dimensional data to characterize the intrinsic laws governing the evolution of fault characteristics.

[0218] Step S3114: Perform principal curve fitting on the initial evolving manifold to obtain a one-dimensional continuous curve passing through the central region of the initial evolving manifold coordinates, which serves as the evolutionary reference manifold.

[0219] Master curve fitting is a non-parametric curve fitting method designed to find the "skeleton" or "central path" of a data point cloud. After obtaining the initial evolving manifold, master curve fitting can extract a smooth, continuous one-dimensional curve that best represents the main trend and path of fault feature evolution under normal conditions, while also exhibiting good robustness to noise and outliers. For example, master curve fitting can be achieved through iterative convergence: an initial curve is set, data points are iteratively projected onto the curve, and the curve's shape is updated based on the projected points until convergence. Alternatively, a curve fitting method based on spline functions (such as B-splines or cubic splines) can be used. A series of control points are selected on the initial evolving manifold, and the positions of these control points are optimized so that the fitted spline curve optimally passes through the central region of the initial evolving manifold coordinates, satisfying the requirements of smoothness and continuity.

[0220] Through steps S3111-S3114, the problem of inaccurate fitting of the evolutionary baseline manifold can be effectively solved, thereby improving the reliability of fault early warning. Specifically, by extracting multiple historical evolutionary trajectories from multiple historical fault-free periods and performing time axis alignment and amplitude normalization, the comprehensiveness, representativeness, and comparability of the data used for constructing the evolutionary baseline manifold are ensured. On this basis, phase space reconstruction is performed on each trajectory in the standardized trajectory set, revealing the inherent dynamic structure of the system and providing richer geometric information for subsequent manifold learning. Furthermore, local linear embedding is performed on all phase point sequences in the high-dimensional phase space to extract low-dimensional manifold coordinates that can characterize the inherent structure of fault feature evolution, effectively filtering out noise, highlighting core evolutionary features, and obtaining the initial evolutionary manifold. The initial evolutionary manifold is fitted with a principal curve to obtain a one-dimensional continuous curve passing through the central region of the initial evolutionary manifold coordinates, which serves as the evolutionary baseline manifold. This evolutionary baseline manifold is smooth and continuous, accurately capturing the main trend of fault feature evolution under normal conditions, and has good robustness to noise and anomalies.

[0221] Step S312: Based on the historical fault characteristics and current fault characteristics of the fourth historical time period, perform time-series evolution trend analysis to obtain the real-time evolution trajectory.

[0222] Step S313: Compare the evolutionary reference manifold with the real-time evolutionary trajectory, and calculate the instantaneous offset angle and offset rate of the real-time evolutionary trajectory relative to the evolutionary reference manifold.

[0223] This step aims to quantify the difference between the real-time evolution trajectory and the baseline manifold. The comparison can be achieved through various geometric or statistical methods. For example, the real-time evolution trajectory can be projected onto the manifold space containing the baseline manifold, and then the distance between the projected point and the nearest point on the baseline manifold, as well as the projection direction of the real-time evolution trajectory onto the manifold space, can be calculated. The instantaneous offset angle can be expressed as the angle between the tangent vector of the real-time evolution trajectory in the manifold space and the tangent vector of the corresponding point on the baseline manifold, reflecting the degree to which the fault feature evolution direction deviates from the baseline. The offset rate can be expressed as the speed at which the real-time evolution trajectory moves in the manifold space, or the rate of change of the distance from the trajectory point of the real-time evolution trajectory to the corresponding point on the baseline manifold over time, reflecting the speed at which the fault feature deviates from the baseline.

[0224] Step S314: Based on the instantaneous offset angle and the offset rate, the dynamic evolution direction vector in the fourth historical time period is fitted.

[0225] This step aims to synthesize the instantaneous offset angle and offset rate to generate a high-precision time-series steering signal. Fitting the dynamically evolving direction vector can be achieved by using the instantaneous offset angle and offset rate as inputs, and training and predicting using vector synthesis or machine learning models (such as support vector machines, neural networks, etc.). For example, the offset angle and offset rate can be used as features to train a regression model to output a multi-dimensional vector, where the magnitude represents the intensity of the offset and the direction represents the trend of the offset. Alternatively, the instantaneous offset angle can be converted into a unit direction vector and weighted by the offset rate to obtain a dynamically evolving direction vector that integrates direction and velocity information.

[0226] Through steps S311-S314, the evolutionary baseline manifold of fault characteristics under normal conditions can be fitted, thus establishing a robust reference baseline unaffected by current environmental noise. Based on this, the real-time evolution trajectory is determined by comparing historical fault characteristics from the fourth historical time period with current fault characteristics, accurately reflecting the dynamic changes of fault characteristics under the current system operating state. Furthermore, comparing the real-time evolution trajectory with the evolutionary baseline manifold and calculating the instantaneous offset angle and offset rate allows for precise quantification of the difference between the real-time evolution trajectory and the normal historical pattern. Based on the instantaneous offset angle and offset rate, a dynamic evolution direction vector within the fourth historical time period is fitted, integrating direction and velocity information to generate a high-precision temporal guidance signal. This enables subsequent cross-physics semantic fusion to more accurately identify the evolution trend and pattern of faults, effectively distinguish between real faults and environmental interference, improve the accuracy and reliability of fault early warning, and provide a solid foundation for early warning of quadruped robots in complex environments such as cable tunnels.

[0227] Step S32: Using the time-series guidance signal as a constraint, extract the fault evolution path associated with the dynamic evolution direction from the target knowledge graph. The fault evolution path includes sequentially associated physical field feature nodes, intermediate state nodes, and fault mode nodes.

[0228] Using time-series guiding signals as constraints means that when extracting fault evolution paths from the target knowledge graph, the dynamic evolution direction of fault features must be considered. Fault evolution paths that align with the actual fault development trend must be selected, avoiding the extraction of irrelevant or erroneous paths. The target knowledge graph is a structured knowledge base containing physical field feature nodes, intermediate state nodes, and fault mode nodes related to quadruped robots, along with their interrelationships. It is typically represented as a graph, where nodes represent entities (e.g., "joint jamming," "motor overheating," "high-frequency vibration"), and edges represent relationships between entities (e.g., "leads to," "manifests as," "related to"). Its function is to provide domain expert knowledge to guide fault diagnosis and reasoning. A fault evolution path refers to a sequence starting from one or more physical field feature nodes, passing through a series of intermediate state nodes, and arriving at one or more fault mode nodes. This path describes the gradual evolution of a fault from its initial physical manifestation to its fault type. Its function is to provide a logical framework for understanding and interpreting how multiple physical field sensor data collectively point to a specific fault. In the target knowledge graph, physical field feature nodes represent abnormal phenomena or characteristics observable by various physical field sensors, such as acoustic signature data, vibration data, electromagnetic field data, and energy consumption data. Examples include "high-frequency vibration," "abnormal current fluctuations," and "abnormal acoustic signature noise." Intermediate state nodes represent transitional stages or internal mechanisms in the fault evolution process. These states may not be directly observable by a single physical field sensor but can be inferred through reasoning or the fusion of multiple physical field sensor data. Examples include "accelerated bearing wear," "partial discharge in windings," "poor lubrication," and "motor overheating." Fault mode nodes represent fault types that need to be identified and addressed. Examples include "joint jamming," "motor overload," and "reducer damage."

[0229] Step S32 above proposes to generate fault semantic descriptions by guiding cross-physical field semantic fusion through target knowledge graph. However, in its implementation, due to the lack of constraints on the dynamic evolution direction, the fault evolution path extracted from the target knowledge graph may deviate from the actual fault evolution trend, thereby affecting the accuracy and efficiency of fault identification.

[0230] To address this, a method is further proposed that uses time-series guidance signals as constraints to extract fault evolution paths associated with the dynamic evolution direction from the target knowledge graph. Specifically, this includes:

[0231] Step S321: Based on the dynamic evolution direction contained in the time-series guidance signal, generate the corresponding direction quantization representation, and map the direction quantization representation to the embedding space where the physical field feature nodes in the target knowledge graph are located to obtain the direction query embedding.

[0232] This step aims to transform the abstract dynamic evolution direction of a fault into a computable and comparable numerical form and associate it with semantic information in the target knowledge graph. Direction quantization is the numerical representation of the dynamic evolution direction, while direction query embedding projects this numerical representation onto the semantic space of the target knowledge graph for similarity calculation and node matching. Specifically, the dynamic evolution direction (e.g., a multi-dimensional vector) can be normalized and then encoded into a fixed-dimensional vector using a pre-trained neural network (such as a multilayer perceptron or autoencoder) as the direction quantization representation. This representation is then input into another mapping layer aligned with the embedding space of physical field feature nodes in the target knowledge graph to obtain the direction query embedding. Alternatively, rule-based or statistical methods can be used to decompose the dynamic evolution direction into multiple predefined base directions (e.g., "increased torque," "increased vibration frequency," etc.) and calculate its projection intensity in each base direction, forming a sparse vector as the direction quantization representation. Then, this sparse vector is directly mapped to the embedding space of the target knowledge graph through a lookup table or linear transformation to obtain the directional query embedding.

[0233] Step S322: Based on the similarity between the directional query embedding and the embedding of each physical field feature node in the target knowledge graph, recall the set of initial physical field feature nodes that match the dynamic evolution direction from the target knowledge graph.

[0234] The purpose of this step is to identify the physical field feature nodes most likely related to the current fault initiation stage in the target knowledge graph based on the dynamic evolution direction of the fault. By calculating the similarity between the query embedding and the embeddings of all physical field feature nodes in the target knowledge graph, the semantic association between them can be quantified, thereby filtering out highly relevant physical field feature nodes. In specific implementation, similarity metrics such as cosine similarity, the reciprocal of Euclidean distance, or the reciprocal of Manhattan distance can be used to calculate the similarity between the directional query embedding and each physical field feature node embedding. Then, a similarity threshold is set, and all physical field feature nodes with similarity exceeding the threshold are added to the initial physical field feature node set. Alternatively, a hard threshold can be omitted, and a Top-K strategy can be adopted, that is, selecting the K most similar physical field feature nodes as the initial physical field feature node set, where K is a preset integer.

[0235] Step S323: Using each physical field feature node in the initial physical field feature node set as the starting point of the path, search along the preset association type in the target knowledge graph to obtain multiple candidate propagation paths from each path starting point to the fault mode node.

[0236] This step aims to explore all possible fault evolution paths from the target knowledge graph. Starting with the initial physical field feature node in the identified set of initial physical field feature nodes, the search proceeds recursively or iteratively along various relationships (such as "cause," "accompany," "influence," etc.) defined in the target knowledge graph until a fault mode node representing the fault type is reached. Specifically, breadth-first search (BFS) or depth-first search (DFS) algorithms can be used. During the search, the search depth and path length are limited, and traversal is only performed along preset relationship types related to the fault propagation logic. For example, a relationship path of "physical field feature node -> intermediate state node -> fault mode node" can be defined. Alternatively, methods based on random walks or graph neural networks (GNNs) can be used. By performing multiple random walks on the target knowledge graph and recording the path from the initial physical field feature node to the fault mode node, or by utilizing GNNs to learn complex relationships between nodes, multiple candidate propagation paths can be generated.

[0237] Step S324: Based on the evolution rate contained in the timing guidance signal, multiple candidate propagation paths are screened for path time consistency, and the candidate propagation path with the highest overall semantic similarity to the dynamic evolution direction among the remaining candidate propagation paths is taken as the fault evolution path associated with the dynamic evolution direction.

[0238] This step is crucial for path refinement. It utilizes the evolution rate information of the fault to verify candidate propagation paths over time and combines this with overall semantic similarity to select the path that best matches the actual fault evolution trend. The evolution rate reflects the time or speed required for a fault to develop from one state to another, which is essential for eliminating candidate propagation paths that do not conform to the actual timeline. In practice, an average evolution time or time range can be preset for each type of association in the target knowledge graph. During screening, the total time of all associations on each candidate propagation path is calculated and compared with the evolution rate contained in the time-series guidance signal. For example, if the evolution rate indicates rapid fault development, candidate propagation paths with shorter total times are selected; conversely, those with longer total times are selected. Simultaneously, the average semantic similarity between the embedding vectors of all nodes on each of the remaining candidate propagation paths and the dynamic evolution direction is calculated, and the path with the highest similarity is selected. Alternatively, a time consistency model can be constructed, which takes the candidate propagation path and evolution rate as input and outputs the matching degree between the candidate propagation path and the rate. For example, recurrent neural networks (RNNs) or Transformer models can be used to learn the correspondence between node sequences in a path and time series. Then, paths with high matching scores are retained. For semantic similarity, the cosine similarity between the weighted average vector of all node embeddings in each of the remaining candidate propagation paths and the query embedding in the dynamic evolution direction can be calculated to select the path with the highest similarity.

[0239] Through steps S321-S324, a temporal guidance signal is introduced as a constraint to extract paths associated with the dynamic evolution direction of the fault. This solves the problem that fault evolution paths extracted from the target knowledge graph may deviate from the actual fault evolution trend, thereby improving the accuracy and efficiency of fault identification. Specifically, by transforming the dynamic evolution direction contained in the temporal guidance signal into a directional quantization representation and mapping it to the embedding space of the target knowledge graph, a directional query embedding is obtained. This allows the abstract dynamic evolution direction to be semantically aligned and quantitatively compared with the physical field feature nodes in the target knowledge graph. Based on the similarity between this directional query embedding and the embeddings of each physical field feature node, a set of initial physical field feature nodes that highly match the current fault dynamic evolution direction can be recalled, avoiding interference from irrelevant nodes and ensuring the relevance of the path starting point. Using these starting nodes as path starting points, searching along preset association types in the target knowledge graph can comprehensively cover multiple possible fault propagation paths, ensuring the diversity of path exploration. Furthermore, by utilizing the evolution rate contained in the temporal guidance signal to screen candidate propagation paths for temporal consistency, candidate propagation paths with mismatched temporal attributes are effectively eliminated, enhancing the practical feasibility of the selected paths. By selecting the path with the highest overall semantic similarity to the dynamic evolution direction from the remaining candidate propagation paths after screening, the fault evolution path associated with the dynamic evolution direction is chosen, ensuring that the extracted path can maximally reflect the actual dynamic evolution process of the fault. This path extraction mechanism, which combines temporal dynamic information with semantic association with the target knowledge graph, enables subsequent fault semantic descriptions to more accurately capture the evolutionary patterns of faults, providing a solid foundation for early warning and adaptive fault-tolerant control of quadruped robots.

[0240] Step S33: Perform multi-resolution temporal decomposition on multiple physical field sensing data to obtain the evolution components of acoustic data, vibration data, electromagnetic field data, and energy consumption data at different time scales (i.e., feature extraction process for the current multi-dimensional physical field data), thereby determining the initial multi-physical field synchronization feature sequence, and reordering and aligning the initial multi-physical field synchronization feature sequence according to the node association order of the fault evolution path to generate the target multi-physical field synchronization feature sequence that matches the fault evolution path.

[0241] Multiple physical field sensing data refer to the acoustic signature data, vibration data, electromagnetic field data, and energy consumption data synchronously collected by the quadruped robot during its operation in cable tunnels. These data reflect the operating status of the quadruped robot from different dimensions. Multi-resolution temporal decomposition refers to decomposing the original physical field sensing data into components with different frequency and amplitude characteristics at different time scales. Its function is to capture the manifestation of faults at different time scales. For example, high-frequency impacts in vibration data may indicate bearing damage, while low-frequency fluctuations may indicate structural loosening. For instance, wavelet basis functions at different scales can be used to decompose the signal to obtain detailed and approximate components in different frequency bands, thus achieving multi-resolution analysis. Alternatively, complex signals can be adaptively decomposed into a series of Intrinsic Mode Functions (IMFs), each IMF representing the oscillation mode of the signal at a specific time scale. Or, the signal can be converted to the frequency domain, and then different bandpass filters can be designed to extract signal components in different frequency ranges, followed by inverse transformation back to the time domain. Evolutionary components (i.e., the corresponding physical field synchronization features) refer to the subsequences or features extracted from the original physical field sensing data after multi-resolution temporal decomposition, representing their dynamic changes at a specific time scale. Reordering and phase alignment refer to rearranging and synchronizing the evolutionary components of different physical fields according to the logical order and temporal relationship of nodes in the extracted fault evolution path. This ensures that the physical field data from different physical fields can be fused according to the logical order of the fault evolution path and maintain temporal consistency for subsequent semantic matching. For example, if the nodes in the target knowledge graph contain temporal information or relative temporal relationships, the starting points or key event points of different physical field data can be aligned based on this information. Alternatively, the similarity between the evolutionary components of different physical fields can be calculated, and the optimal time offset can be determined by maximizing the similarity, thereby achieving phase alignment. Furthermore, event markers related to specific nodes in the fault evolution path can be identified in different physical field data, and these events can be used as synchronization points for alignment. The target multiphysics synchronization feature sequence refers to a sequence that combines evolutionary components obtained from different physics and decomposed at different time scales after reordering and phase alignment. This sequence is consistent with the fault evolution path in both time and semantics, providing a structured input for subsequent path attention network processing.

[0242] Step S34: Input the target multiphysics synchronization feature sequence into the path attention network sequentially along the fault evolution path. Use the node embedding of each node in the fault evolution path as the path guide, and calculate the semantic matching degree between the target multiphysics synchronization feature sequence and the current node node by node.

[0243] Path attention networks are deep learning models whose core idea is to use attention mechanisms to process sequential data, dynamically focusing on the parts of the sequence most relevant to the current task. Here, it's used to semantically fuse target multi-physics synchronization feature sequences under the guidance of a fault evolution path. Its role is to selectively weight and integrate different physics synchronization features based on the contextual information of the fault evolution path, thereby more accurately identifying faults. For example, a Transformer encoder-decoder structure can be constructed, where the encoder processes the target multi-physics synchronization feature sequence, and the decoder uses node embeddings as queries to calculate attention weights for each part of the sequence. Alternatively, an attention module can be introduced at each time step of a recurrent neural network (RNN, such as LSTM or GRU), enabling it to assign attention weights to different parts of the input sequence based on the current node embedding and historical hidden states. Furthermore, the fault evolution path can be viewed as a graph structure, and a Graph Attention Network (GAT) can be used to learn attention weights between nodes, thereby propagating and aggregating information along the path. Node embedding refers to the low-dimensional vector representation of each node (including physics feature nodes, intermediate state nodes, and failure mode nodes) in the target knowledge graph. These vectors are learned through knowledge graph embedding techniques such as TransE (Translational Embedding), RotatE (Rotational Embedding), and GraphSAGE (Graph Sample and AggregatE). They capture the semantic information and relationships of nodes, serving as "path guidance" to provide semantic context for the path attention network, guiding it to find the information that best matches the current path node in the target multiphysics synchronization feature sequence. Semantic matching degree refers to the similarity between a physics synchronization feature in the target multiphysics synchronization feature sequence and the semantic concept represented by the current node in the failure evolution path. It quantifies the correlation strength between the currently observed physical phenomenon and the failure evolution steps defined in the target knowledge graph, and its role is to evaluate the fit between data and knowledge.

[0244] The above step S34 proposes to calculate the semantic matching degree node by node to evaluate the matching degree between the target multiphysics synchronization feature sequence and the fault evolution path nodes. However, due to the lack of consideration for the path context, the matching degree calculation may be inaccurate, affecting the accuracy of the fault semantic description.

[0245] To address this, a method is further proposed that uses the node embeddings of each node in the fault evolution path as path guidance to calculate the semantic matching degree between the target multiphysics synchronization feature sequence and the current node node by node. Specifically, this includes:

[0246] Step S341: Extract the node embedding of the current node sequentially from the fault evolution path, and extract the physical field synchronization features corresponding to the node type from the target multiphysics synchronization feature sequence according to the node type of the current node.

[0247] Extracting the node embedding of the current node sequentially from the fault evolution path refers to obtaining the vector representation of the currently processed node in order while traversing the fault evolution path. Node embeddings are semantic representations of each node in the target knowledge graph, typically obtained through pre-training using graph embedding algorithms (such as DeepWalk, Node2Vec, TransE, etc.). These algorithms map discrete nodes into a low-dimensional continuous vector space to capture the semantic relationships between nodes. Their role is to provide a structured and computable semantic benchmark for subsequent semantic matching. For example, the corresponding embedding vector can be directly retrieved from a pre-built node embedding lookup table based on the unique identifier of the current node. Alternatively, if the target knowledge graph is dynamically updated, a lightweight graph neural network (GNN) layer can be used to generate the node embedding in real time based on the neighbor information of the current node.

[0248] Extracting physical field synchronization features corresponding to the current node's type from the target multiphysics synchronization feature sequence means selecting and extracting semantically relevant physical field synchronization features from the already aligned and synchronized target multiphysics synchronization feature sequence based on the physical field type represented by the current node (e.g., acoustic signature, vibration, electromagnetic field, or energy consumption). For example, if the current node is "high-frequency vibration anomaly," then high-frequency vibration synchronization features are extracted. If the current node is "motor current fluctuation," then energy consumption synchronization features are extracted. This extraction can be achieved through predefined indexes, time window sliding, or feature filtering based on node type.

[0249] Step S342: Project the physical field synchronization features onto the semantic space where the node embedding is located to obtain the feature vector to be matched. Calculate the cosine similarity between the feature vector to be matched and the node embedding as the initial semantic matching degree of the current node.

[0250] Projecting physical field synchronization features onto the semantic space of the node embedding to obtain the feature vector to be matched involves using a mapping function or transformation matrix to convert the physical field synchronization features into a vector representation with the same dimension and semantic characteristics as the node embedding. This projection process aims to eliminate heterogeneity between different data sources, enabling direct comparison between the physical field synchronization features and the node embeddings in the target knowledge graph. For example, a fully connected neural network layer can be used, taking the physical field synchronization features as input and outputting a vector with the same dimension as the node embedding. Alternatively, an attention-based module can be designed to interact with the physical field synchronization features and the node embeddings, generating a more semantically aligned feature vector to be matched.

[0251] Calculating the cosine similarity between the feature vector to be matched and the embedding of the current node, as the initial semantic matching degree of the current node, refers to using cosine similarity, a commonly used vector similarity metric, to quantify the correlation between the physical field synchronization features and the semantics of the target knowledge graph node. Cosine similarity measures the degree of similarity between two vectors by calculating the cosine of the angle between them; its value ranges from -1 to 1, with a higher value indicating higher similarity. For example, one can directly calculate the dot product of the two vectors and divide by their respective magnitudes. Alternatively, one can first normalize the two vectors and then calculate their dot product.

[0252] Step S343: Obtain the historical matching degree (i.e., the path guidance semantic matching degree corresponding to the predecessor node of the current node in the fault evolution path, and use the historical matching degree as the path context weight to perform weighted correction on the initial semantic matching degree of the current node to obtain the path guidance semantic matching degree of the current node.

[0253] Obtaining the historical matching degree of the predecessor node of the current node in the fault evolution path refers to retrieving the semantic matching degree (i.e., historical matching degree) of the directly preceding node on the fault evolution path from the storage medium before processing the current node. This historical matching degree contains the matching information between the predecessor node and the corresponding physical field synchronization features, representing the strength of semantic associations that have occurred on the path. For example, the historical matching degree of the predecessor node can be retrieved from a stack structure or queue. Alternatively, if a recursive or iterative approach is used to process the path, the historical matching degree of the predecessor node is passed as a parameter to the calculation process of the current node.

[0254] Using historical matching scores as path context weights to weight and adjust the initial semantic matching score of the current node, we obtain the path-guided semantic matching score of the current node. This means adjusting the initial semantic matching score of the synchronization features between the current node and the physical field using the matching score information of predecessor nodes, reflecting the influence of the path context on the initial semantic matching score. For example, the historical matching score can be used as a multiplicative factor, directly multiplied by the initial semantic matching score. Alternatively, a gating mechanism can be designed to fuse the historical matching score and the initial semantic matching score, dynamically adjusting the weight allocation through learning, thereby generating a path-guided semantic matching score that better conforms to the path logic.

[0255] Step S344: Associate and store the path-guided semantic matching degree of the current node with the feature vector to be matched, and use it as the historical matching degree referenced in the calculation of subsequent nodes. Repeat the above steps along the fault evolution path until the path-guided semantic matching degree calculation of all nodes is completed.

[0256] Associating and storing the path-guided semantic matching degree of the current node with the feature vector to be matched, and using this as the historical matching degree referenced in the calculation of subsequent nodes, means saving the path-guided semantic matching degree calculated by the current node and corrected by the path context, along with its corresponding feature vector to be matched. This allows it to be referenced as the historical matching degree of the "predecessor node" when processing the next node on the fault evolution path. For example, this information can be stored in an ordered list or a hash table, where the key is the node identifier and the value is the combination of the path-guided semantic matching degree and the feature vector to be matched.

[0257] Repeating the above steps sequentially along the fault evolution path until the path-guided semantic matching degree calculation for all nodes is completed means performing the aforementioned feature extraction, projection, similarity calculation, context correction, and result storage operations one by one, starting from the initial node of the path, according to the order defined in the fault evolution path, until all nodes on the path have completed the calculation of their path-guided semantic matching degree. This process ensures that the path-guided semantic matching degree of each node on the path takes into account the contextual information of its predecessor node, thus forming a coherent and path-dependent matching degree sequence.

[0258] By introducing path context information through steps S341-S344, the problem of inaccurate matching degree calculation caused by the lack of consideration for path context when calculating semantic matching degree node by node in the fault evolution path is solved. Specifically, by obtaining the historical matching degree of the predecessor node and using it as the path context weight, the initial semantic matching degree of the current node is weighted and corrected. This ensures that the path-guided semantic matching degree of the current node not only reflects its own features and the similarity of the node embedding, but also incorporates the overall logic of the fault evolution path and the matching state of the predecessor node. This context-based correction mechanism avoids the bias caused by isolated calculations, making the semantic matching results more consistent with the actual situation of the fault evolution process. In addition, the path-guided semantic matching degree of the current node is associated and stored with the feature vector to be matched, and used as the historical matching degree referenced in the calculation of subsequent nodes, ensuring the effective transmission and accumulation of context information, thereby forming a coherent and self-consistent semantic matching chain throughout the fault evolution path. This method can generate more interpretable fault semantic descriptions, providing a more reliable basis for subsequent fault diagnosis and fault-tolerant decision-making, and improving the accuracy and robustness of fault warning and adaptive fault-tolerant control of quadruped robots in complex cable tunnel environments.

[0259] Step S35: Accumulate the semantic matching degree of each node along the path to generate a fault semantic description that integrates path evolution information and multi-physics synchronization features.

[0260] Path accumulation refers to integrating the semantic matching degrees calculated node by node along the fault evolution path to form a comprehensive metric. This is not just a simple summation, but may involve operations such as weighting and sequence aggregation to reflect the accumulated evidence of the fault throughout the entire fault evolution path. Its role is to integrate all information along the fault evolution path to form a comprehensive judgment of the fault. The fault semantic description is a generated, detailed, and interpretable textual or structured representation of the current fault state of the quadruped robot. It integrates multi-source verification information (from real-time sensor data) and historical fault features, and combines fault evolution path information from the target knowledge graph. Its role is to provide rich semantic input for subsequent fault-tolerant decisions, making the decision-making process more targeted and interpretable.

[0261] Step S35 above proposes to accumulate the semantic matching degree of each node to generate a fault semantic description that integrates path evolution information and target multi-physics synchronization feature sequence. However, in its implementation, the accumulation of semantic matching degree may lack an effective weighting mechanism and fusion strategy, resulting in the generated fault semantic description being inaccurate or unable to fully integrate path evolution information and target multi-physics synchronization feature sequence, thereby affecting the identification of fault type and location.

[0262] To address this, a method is further proposed that accumulates the semantic matching degree of each node along the path to generate a fault semantic description that integrates path evolution information and multi-physics synchronization features. Specifically, this includes:

[0263] Step S351: Sequentially obtain the path guidance semantic matching degree of each node along the fault evolution path, and extract the physical field synchronization features corresponding to each node from the target multi-physics field synchronization feature sequence.

[0264] The fault evolution path indicates the process of a fault evolving from its initial symptoms to its current failure mode. It consists of a series of interconnected nodes (e.g., physical field feature nodes, intermediate state nodes, and failure mode nodes). The fault evolution path provides a structured framework for understanding and tracking the occurrence and development of faults. The path-guided semantic matching degree represents the degree of semantic relevance or agreement between the target multiphysics synchronization feature sequence and a specific node in the fault evolution path, taking into account path context information. Its function is to quantify the degree of agreement between the currently observed physical field synchronization features and a specific node on the fault evolution path. The target multiphysics synchronization feature sequence is a set of time-series features, including acoustic signatures, vibrations, electromagnetic fields, and energy consumption, that are reordered and phase-aligned to match the structure of the fault evolution path. Its function is to provide preprocessed, multi-dimensional, and time-synchronized physical field observation data corresponding to each stage of the fault evolution path. The implementation of the target multiphysics synchronization feature sequence can include: extracting features (e.g., frequency domain features, time domain statistical features) from the raw time-series data collected by different physical field sensors, and then aligning the time using dynamic time warping (DTW) or a cross-correlation-based method. Alternatively, a multimodal fusion network can be used to map different physical field data to a unified feature space and synchronize them. Physical field synchronization features are local feature subsets extracted from the target multiphysics synchronization feature sequence, corresponding to specific node types or stages in the fault evolution path. Their function is to provide each path node with the most relevant physical field information, avoiding interference from irrelevant information. The implementation of physical field synchronization features can include: selective extraction based on node type (e.g., extracting vibration synchronization features if the node is "vibration anomaly," and energy consumption synchronization features if the node is "motor overheating"). Alternatively, a sliding window mechanism can be used to truncate feature segments overlapping the current node's time window from the target multiphysics synchronization feature sequence.

[0265] Step S352: Arrange the physical field synchronization features corresponding to each node according to the order of the nodes in the fault evolution path to obtain the path alignment feature sequence.

[0266] Path-aligned feature sequences are feature sequences formed by rearranging the physical field synchronization features corresponding to each node in a fault evolution path according to the logical or temporal order of these nodes in the fault evolution path. Their purpose is to ensure that the organizational structure of the feature sequence is consistent with the logical order of the fault evolution path, thereby accurately reflecting the fault evolution process. Path-aligned feature sequences can be implemented by: simply concatenating the corresponding physical field synchronization features according to the index or topological order of the nodes in the fault evolution path; or, during the concatenation process, inserting specific delimiters or padding values ​​based on the correlation strength or time interval between nodes to maintain the integrity and interpretability of the sequence.

[0267] Step S353: Using the path guidance semantic matching degree corresponding to each node as a weight coefficient, the physical field synchronization features at the corresponding positions in the path alignment feature sequence are weighted and summed to obtain the path cumulative vector.

[0268] The weighting coefficient here refers to the path-guided semantic matching degree corresponding to each node, used to measure the importance or contribution of the corresponding physical field synchronization feature in the fault semantic description. Its function is to dynamically adjust the influence of the physical field synchronization feature in the fusion result based on the semantic matching degree between the physical field synchronization feature and the fault evolution path nodes, highlighting the physical field synchronization features highly correlated with the current fault state. The implementation of the weighting coefficient can include: directly using the original value of the path-guided semantic matching degree as the weight; or normalizing the path-guided semantic matching degree (e.g., using the Softmax function) to make its sum equal to 1, ensuring the stability of the weighted summation. Weighted summation is a method for fusing multiple physical field synchronization features, where each physical field synchronization feature is multiplied by a weighting coefficient, and then all weighted physical field synchronization features are summed. Its function is to fuse the physical field synchronization features corresponding to different nodes according to their semantic matching degree with the fault evolution path, generating a path accumulation vector that comprehensively reflects fault evolution information. The weighted summation can be implemented by multiplying each dimension of each physical field synchronization feature element-wise with its corresponding weight coefficient, and then summing all the weighted physical field synchronization feature vectors. Alternatively, a more complex weighted fusion strategy can be used, such as attention-based weighting, which allows the model to learn the dynamic weights of different physical field synchronization features. The path accumulation vector is a vector representation obtained after weighted summation, which integrates fault evolution path information and the target multi-physics synchronization feature sequence. Its role is to serve as a preliminary representation of fault semantic description, containing key physical field synchronization features during fault evolution and their semantic matching degree with path nodes. The path accumulation vector can be implemented as a fixed-dimensional vector, whose dimension is determined by the dimension and number of physical field synchronization features, or a variable-dimensional vector, whose dimension is dynamically adjusted according to the length and complexity of the fault evolution path.

[0269] Step S354: The path accumulation vector and the path guidance semantic matching degree sequence are concatenated, and the concatenated vector is input into the fully connected layer for feature fusion, and the fault semantic description that integrates path evolution information and target multi-physics field synchronization feature sequence is output.

[0270] The path-guided semantic matching degree sequence is a sequence of all path-guided semantic matching degrees arranged in the node order of the fault evolution path. Its function is to provide semantic matching degree information for each node on the fault evolution path, serving as an independent feature, and is input into the subsequent fusion module along with the path accumulation vector. The path-guided semantic matching degree sequence can be implemented as a one-dimensional vector with a length equal to the number of nodes in the fault evolution path, or as a two-dimensional matrix containing path-guided semantic matching degree values ​​and other relevant path context information. Concatenation is the operation of joining two or more vectors or sequences in a dimensional dimension. Its function is to combine the path accumulation vector (containing weighted fusion of physical field synchronization features) with the path-guided semantic matching degree sequence (containing matching degree information of the path context) to form a more comprehensive feature representation. Concatenation can be implemented as a simple vector concatenation in the feature dimension, or as a sequence concatenation in the time dimension to form a longer temporal feature. A fully connected layer is a basic layer in a neural network, where each input neuron is connected to every output neuron. Its function is to perform nonlinear transformation and deep fusion on the concatenated vectors, thereby learning a higher-level and more abstract semantic representation of faults from these original features. Fully connected layers can be implemented using single-layer or multi-layer perceptrons (MLPs) with activation functions (e.g., ReLU (Rectified Linear Unit), Sigmoid) to introduce nonlinearity. Alternatively, they can be combined with regularization techniques such as Dropout to prevent overfitting. Feature fusion is the process of integrating features from different sources or representations to generate a more informative and discriminative unified feature representation. Its function is to fully utilize the information in the path accumulation vector and the path-guided semantic matching degree sequence, and through the learning capability of fully connected layers, generate a vector that can comprehensively and accurately describe the semantics of faults. Feature fusion can be implemented by automatically discovering and fusing complex relationships between different features through neural network weight learning, or by employing more complex fusion architectures, such as attention mechanisms or gating mechanisms, to dynamically adjust the contributions of different features. The fault semantic description is an output vector representation that integrates fault evolution path information, target multiphysics synchronization feature sequences, and semantic matching degree information. Its role is to serve as a comprehensive, high-level semantic description of the current fault state, which can be used for subsequent fault diagnosis, classification, and fault-tolerant decision-making. The implementation of the fault semantic description can include: a fixed-dimensional vector, the dimensions of which are designed according to the needs of downstream tasks; or an interpretable symbolic representation that can be directly mapped to specific fault types and locations.

[0271] Steps S351-S354 introduce ordered path accumulation, dynamic weighting based on semantic matching degree, and a deep feature fusion mechanism, overcoming the limitations of existing technologies that lack effective weighting and fusion strategies for semantic matching degree accumulation. Specifically, the path-guided semantic matching degree corresponding to each node is sequentially obtained along the fault evolution path, and physical field synchronization features are extracted. These features are then arranged in path order, ensuring a strict correspondence and temporal consistency between multi-source physical field synchronization features and the fault evolution path. This avoids the randomness of feature extraction and provides a structured and ordered foundation for subsequent fusion, thereby enabling more accurate capture of the fault evolution process. Simultaneously, the path-guided semantic matching degree is used as a weighting coefficient to weight and sum the physical field synchronization features, ensuring that physical field synchronization features with high semantic matching degrees to fault evolution path nodes occupy a larger proportion in the path accumulation vector. This dynamic weighting mechanism effectively highlights key physical field synchronization features highly correlated with the current fault state while suppressing the influence of noise and irrelevant information. It solves the problem of lacking effective weighting in the semantic matching degree accumulation process, allowing the path accumulation vector to better reflect the core features of the fault. Furthermore, by concatenating the path accumulation vector with the path-guided semantic matching degree sequence and inputting it into a fully connected layer for feature fusion, a deep integration of path evolution information and target multi-physics synchronization feature sequences is achieved. The fully connected layer can learn the complex nonlinear relationships between these heterogeneous information, thereby generating a highly abstract and information-rich fault semantic description. This deep fusion strategy solves the problem of insufficient multi-source information fusion in traditional methods, enabling the fault semantic description to not only include physical field synchronization features but also incorporate contextual information and semantic matching degree of the fault evolution path, greatly improving the ability to identify fault types and locations. In summary, the generated fault semantic description is not only more accurate and information-rich but also fully integrates fault evolution path information and target multi-physics synchronization feature sequences, providing a solid foundation for subsequent fault diagnosis and fault-tolerant decision-making, and improving the accuracy and reliability of fault warning for quadruped robots in complex environments.

[0272] Through steps S31-S35, a temporal evolution trend analysis of fault features is introduced, and their dynamic evolution direction is used as a temporal guidance signal for cross-physics field fusion. This effectively solves the problem that traditional methods lack effective utilization of temporal evolution trends during semantic fusion. By capturing the dynamic evolution direction changes of fault features, the nascent stage and development trend of faults can be identified earlier, avoiding false alarms or missed alarms that may be caused by judging based solely on instantaneous states, thus improving the timeliness and accuracy of fault early warning. Furthermore, the temporal guidance signal is used as a constraint condition to extract fault evolution paths that are highly correlated with the current dynamic evolution direction from the target knowledge graph. This path extraction mechanism based on real-time trends overcomes the blindness of traditional knowledge graph search, ensuring the consistency between the selected path and the actual fault development process, thereby providing an accurate semantic framework for subsequent multi-physics field data fusion. The physical field feature nodes, intermediate state nodes, and fault mode nodes included in the fault evolution path provide clear logical clues for the step-by-step diagnosis and reasoning of faults. Furthermore, multi-resolution temporal decomposition of various physical field sensor data comprehensively captures the evolutionary characteristics of acoustic signature data, vibration data, electromagnetic field data, and energy consumption data at different time scales, avoiding the limitations of single-time-scale analysis. By reordering and phase-aligning the initial multi-physics synchronization feature sequence according to the node association order of the fault evolution path, a target multi-physics synchronization feature sequence that highly matches the fault evolution path in terms of time and semantics is successfully generated, laying a solid data foundation for subsequent deep semantic fusion and ensuring the temporal and logical synergy of different physical field information. The target multi-physics synchronization feature sequence is input into the path attention network along the fault evolution path, and the node embeddings of each node in the fault evolution path are used as path guides to calculate the semantic matching degree node by node and accumulate the path. This mechanism enables the system to selectively focus on and integrate the most relevant parts of various physical field sensor data to the fault evolution path based on the semantic context of the knowledge graph, thereby generating a fault semantic description that integrates path evolution information and target multi-physics synchronization features. This fault semantic description not only includes multi-source verification information and historical fault features, but also provides interpretable semantic representations of fault types and locations, greatly improving the reliability and precision of fault identification. It provides high-quality input for subsequent unified fault-tolerant decision-making models, thereby realizing early warning and adaptive fault-tolerant control of quadruped robot faults.

[0273] Step S4: Based on the fault semantic description, motion adjustment instructions are generated through a unified fault-tolerant decision model obtained by knowledge fusion of multiple fault-specific control strategies.

[0274] The unified fault-tolerant decision model can be a simple rule engine that directly matches keywords in the fault semantic description (e.g., "left front leg motor overload") to a preset fault-specific control strategy (i.e., candidate fault-specific control strategies, such as "left front leg gait adjustment strategy") and extracts its core control parameters (e.g., gait cycle, joint swing amplitude) to construct a customized control parameter combination for the quadruped robot. The motion adaptation correction module of the unified fault-tolerant decision model can perform simple linear scaling on the gait cycle parameter and joint swing amplitude in the customized control parameter combination based on the quadruped robot's current joint motion data (e.g., current speed, posture) and joint torque margin assessment values. For example, if the joint torque margin assessment value is small, the swing amplitude is reduced. The corrected control parameter combination (i.e., the corrected control parameter combination) is then matched with a preset gait primitive library (e.g., containing several fixed gaits such as walking, trotting, and slow walking), selecting the closest gait primitive and generating motion adjustment instructions.

[0275] The Unified Fault-Tolerant Decision Model is an intelligent decision-making system that integrates multiple fault-specific control strategies. Through knowledge fusion, it integrates and optimizes response strategies for different fault scenarios, forming a general framework capable of adaptively generating motion adjustment commands based on the current fault semantic description.

[0276] Motion adjustment commands are instructions generated by a unified fault-tolerant decision model based on fault semantic descriptions, used to modify the motion parameters of a quadruped robot in real time. These commands are designed to enable the quadruped robot to adjust its gait, posture, or joint movements when a fault occurs, in order to maintain the stability and functionality of the quadruped robot and prevent the fault from worsening or causing mission failure.

[0277] Step S4 above proposes generating motion adjustment instructions based on fault semantic description to achieve fault-tolerant control of the quadruped robot. However, in its implementation, the existing method cannot adaptively generate customized control parameter combinations based on fault semantics, resulting in motion adjustment instructions that lack specificity and adaptability, thus affecting the control effect.

[0278] To address this, a method is further proposed that generates motion adjustment commands based on fault semantic description and through a unified fault-tolerant decision model obtained by knowledge fusion of multiple fault-specific control strategies. Specifically, this method includes:

[0279] Step S41: Input the fault semantic description into the policy routing network in the unified fault-tolerant decision model. The policy routing network maps the fault semantic description to the policy embedding space composed of multiple fault-specific control policies and outputs the semantic similarity distribution between the fault semantic description and each fault-specific control policy.

[0280] Fault semantic description is a high-level, abstract vector representation of the current fault state of a quadruped robot. It is not merely a compilation of raw sensor data, but a comprehensive and structured description of the fault type, location, severity, and evolution trend. For example, a fault semantic description can be a high-dimensional vector, where each dimension represents a specific attribute or feature of the fault, such as the probability distribution of the fault mode, the type of physical field affected, or the direction of evolution of model prediction bias. The policy routing network is a key component in the unified fault-tolerant decision model, responsible for matching the input fault semantic description with various fault-specific control strategies, evaluating their semantic relevance, and outputting a semantic similarity distribution. The policy routing network can be an attention-based neural network that represents semantic similarity by calculating the attention weights between the fault semantic description embedding and each strategy embedding (i.e., the fault-specific control strategy vector). Alternatively, it can be a multilayer perceptron that takes the fault semantic description and strategy embeddings as input and outputs a similarity score. The policy routing network maps the fault semantic description to a policy embedding space composed of multiple fault-specific control strategies and outputs the semantic similarity distribution between the fault semantic description and each fault-specific control strategy. The policy embedding space is a high-dimensional vector space where each fault-specific control policy is represented as a unique vector. These vectors capture the semantic information and functional characteristics of the fault-specific control policies, ensuring that semantically similar fault-specific control policies are close together in this space. The semantic similarity distribution represents the probability or weight distribution of the semantic similarity between the current fault semantic description and all available fault-specific control policies, indicating which fault-specific control policies are most relevant to the current fault and to what extent.

[0281] Step S42: Based on semantic similarity distribution, extract the core control parameters (i.e., control parameters) of the top N strategies (i.e., multiple candidate fault-specific control strategies) that are closest to the current fault semantic description from multiple fault-specific control strategies. Then, perform semantic similarity-weighted interpolation reconstruction on the core control parameters of the top N strategies in the strategy embedding space to generate a customized control parameter combination corresponding to the current fault semantic description, where N is a positive integer.

[0282] Multiple fault-specific control strategies refer to a series of control rules, parameter sets, or behavioral patterns pre-designed or learned to address various specific fault modes that a quadruped robot may encounter. For example, a fault-specific control strategy can be a set of parameterized controllers with different parameter sets preset for different fault types. Alternatively, it can be a series of rule-based expert systems that trigger corresponding rule sets to adjust the quadruped robot's behavior when a specific fault mode is detected. Core control parameters refer to the set of parameters most critical to each fault-specific control strategy and most influential on the quadruped robot's motion behavior and performance. For example, for a gait control strategy, core control parameters might include gait cycle parameters, swing leg height, support leg phase, and joint target angles. Semantic similarity-weighted interpolation reconstruction is a parameter generation method that, based on the semantic similarity between the fault semantic description and the top N strategies, performs weighted averaging or more complex interpolation operations on the core control parameters of the top N strategies, thereby generating a customized parameter combination that integrates the advantages of the top N strategies and highly fits the current fault semantics. For example, linear interpolation can be used, with semantic similarity as weight, to perform a weighted summation of the core control parameters of the top N strategies. Alternatively, Gaussian process regression can be used, employing the policy embedding and core control parameters as training data to predict the parameters corresponding to the query point using a Gaussian process model. Customized control parameter combinations are a set of control parameters tailored to the current fault semantic description; they are dynamically generated rather than simply selected from preset strategies.

[0283] Step S42 above proposes to interpolate and reconstruct the core control parameters of the first N strategies in the strategy embedding space to generate customized control parameter combinations. However, in this process, if the weighted interpolation is not performed considering the semantic similarity between the fault-specific control strategy and the current fault semantic description, the generated customized control parameter combinations may not accurately match the specific characteristics of the current fault, thus affecting the adaptability and accuracy of fault-tolerant control.

[0284] To address this, a method is further proposed that involves performing semantic similarity-weighted interpolation reconstruction on the core control parameters of the first N policies within the policy embedding space to generate customized control parameter combinations corresponding to the current fault semantic description. Specifically, this includes:

[0285] Step S421: Extract the core control parameter vectors corresponding to each fault-specific control strategy from the first N strategies, and obtain the strategy embedding vectors of the first N strategies in the strategy embedding space.

[0286] The core control parameter vector is a vector composed of key parameters of a pre-defined fault-specific control strategy designed for a specific fault mode or type. These parameters directly determine the motion adjustment behavior of the quadruped robot when facing a specific fault, such as gait period, joint swing amplitude, joint stiffness, damping coefficient, and torque limitation. It can be implemented by pre-setting and storing a set of optimal or suboptimal control parameters for each fault-specific control strategy through expert experience, offline simulation optimization, or reinforcement learning. Alternatively, it can be achieved by analyzing historical fault data and combining machine learning models (such as decision trees and neural networks) to learn and extract control parameter patterns associated with fault modes of specific fault-specific control strategies from a large amount of operational data. The strategy embedding vector is a low-dimensional dense vector used to characterize the semantic position of each fault-specific control strategy in the strategy embedding space. These vectors are mapped from discrete fault-specific control strategies or strategy description information to a continuous vector space using embedding techniques (such as Word2Vec, Doc2Vec, and Transformer encoders), ensuring that semantically similar fault-specific control strategies are closer together in the strategy embedding space, while semantically dissimilar fault-specific control strategies are farther apart. This can be achieved by encoding the textual description of fault-specific control strategies using pre-trained natural language processing models (such as BERT (Bidirectional Encoder Representations from Transformers) or GPT (Generative Pre-trained Transformer)) to obtain their policy embedding vectors. Alternatively, graph neural networks (GNNs) can be used to learn the relationships between fault-specific control strategies, generating policy embedding vectors that reflect the structural and semantic relationships between these strategies.

[0287] Step S422: Map the fault semantic description to the policy embedding space to obtain the query embedding point corresponding to the current fault semantic description, and calculate the distance between the query embedding point and the policy embedding vectors corresponding to the first N policies to obtain the spatial distance between each fault-specific control policy and the query embedding point.

[0288] The fault semantic description is a vector that provides a high-level abstraction and semantic representation of the current fault state of the quadruped robot. It contains key information such as the fault type, location, severity, evolution trend, and deviation from historical behavior. This can be achieved as a high-dimensional feature vector, obtained through semantic fusion guided by a target knowledge graph, derived from the current fault features generated in the preceding steps and the current multi-dimensional physical field data. The query embedding point is the vector representation of the current fault semantic description in the policy embedding space. By mapping the fault semantic description to the same space as the policy embedding vector, semantic alignment and comparison between the fault state and the preset control policy can be achieved. This can be achieved by encoding the fault semantic description using the same embedding model as the generated policy embedding vector (such as a Transformer encoder) to obtain its position in the policy embedding space. Alternatively, a specially trained mapping network (such as a multilayer perceptron) can be used to convert the fault semantic description into a query embedding point in the policy embedding space. Spatial distance is a measure of the similarity or correlation between the query embedding point and the policy embedding vector. The smaller the distance, the closer the semantic description of the current fault is to the semantics of the corresponding fault-specific control strategy, and the higher the applicability of the fault-specific control strategy to the current fault. This can be achieved by using Euclidean distance to calculate the straight-line distance between two vectors in multidimensional space. Alternatively, cosine similarity can be used, which measures the directional consistency of two vectors by calculating the cosine of the angle between them, and then converting it into a distance metric (e.g., 1-cosine similarity).

[0289] Step S423: Using the reciprocal of the spatial distance as the basis function, radial basis function interpolation is performed on the core control parameter vectors corresponding to the first N strategies to construct a parameter interpolation function with the query embedding point as the independent variable.

[0290] Using the reciprocal of spatial distance as the basis function means that fault-specific control strategies closer to the query embedding point (i.e., with higher semantic similarity) contribute more to the interpolation result. This non-linear weighting method allows the fault-specific control strategy most relevant to the current fault semantic description to play a dominant role in generating customized control parameter combinations. This can be implemented by directly using 1 / distance as the basis function, or by using the Gaussian radial basis function exp(-ε×distance). 2), where ε is a shape parameter used to adjust the decay rate. Radial basis function interpolation is a method to construct a continuous function using known data points (here referring to the core control parameter vectors of the first N policies and their positions in the policy embedding space). This function can output an interpolation result based on any input point (here referring to the query embedding point), characterized by the interpolation result being mainly influenced by neighboring data points. It can be implemented using a combination of polynomial interpolation and radial basis functions, such as thin plate splines. Alternatively, a pure radial basis function network (RBFN) can be used, where each radial basis function corresponds to a policy embedding vector, and its output is obtained by weighted summation. The parametric interpolation function is a mathematical function that takes the position in the policy embedding space (i.e., the query embedding point) as the independent variable and the core control parameter vector as the dependent variable. Once constructed, this function can accept any valid query embedding point as input and output a customized control parameter vector (i.e., a customized combination of control parameters) that semantically matches it. The parametric interpolation function can be expressed as F(x)=Σ(wi× (||x-ci||))+P(x), where x is the query embedding point and ci is the policy embedding vector. Here, is the radial basis function, wi is the weight, and P(x) is a low-order polynomial. Alternatively, an interpolation function can be achieved by training a neural network (such as a multilayer perceptron) with the policy embedding vector and the corresponding core control parameter vector as training data, enabling it to learn the mapping relationship from the embedding spatial location to the control parameters.

[0291] Step S424: Substitute the query embedding point into the parameter interpolation function to calculate the parameter vector corresponding to the position of the current fault semantic description in the policy embedding space, and use the parameter vector as a customized control parameter combination.

[0292] Substituting the query embedding point into the parameter interpolation function is a crucial step in obtaining the customized control parameter combination, ensuring that the generated customized control parameter combination is tailored to the specific fault state. This can be achieved by directly inputting the query embedding point into the radial basis function interpolation formula. Alternatively, if the parameter interpolation function is implemented using a neural network, the query embedding point is used as the input to the neural network, and the output is obtained through forward propagation. The calculation of the parameter interpolation function yields a parameter vector (i.e., the vector representation of the customized control parameter combination) that highly matches the position of the current fault semantic description in the policy embedding space. This parameter vector is the result of a weighted fusion of the core control parameter vectors corresponding to the first N policies in the semantic space, combining the advantages of the fault-specific control policies most relevant to the current fault. This can be implemented as a list or vector containing all core control parameters (such as gait period, joint swing amplitude, etc.), or as a structured data object where each field corresponds to a specific control parameter and its value. Using the calculated parameter vector as the customized control parameter combination means that this combination will be directly used in the subsequent motion adaptation correction module to guide the quadruped robot in precise fault-tolerant motion adjustments. This customization ensures the adaptability and effectiveness of the control strategy. This can be achieved by directly passing the customized control parameter combination to the motion adaptation correction module for further adjustment and optimization, or by storing it in a temporary variable for later use by subsequent modules.

[0293] Through steps S421-S424, a customized control parameter combination is accurately generated using a semantic similarity-weighted radial basis function interpolation method, effectively solving the problem of inaccurate parameter matching in traditional methods. Specifically, by extracting the core control parameter vector and the policy embedding vector, a structured raw data and embedding representation foundation are provided for interpolation. The fault semantic description is mapped to the policy embedding space to obtain the query embedding point, and its spatial distance with the policy embedding vector is calculated, thereby quantifying the semantic similarity between the current fault and each fault-specific control policy, providing an objective basis for weighted interpolation. Using the reciprocal of the spatial distance as the basis function for radial basis function interpolation ensures that fault-specific control policies with high semantic similarity contribute more to the parameters, avoiding the bias caused by simple averaging and improving the adaptability of the parameter combination. Substituting the query embedding point into the parameter interpolation function to calculate the parameter vector generates a parameter combination matching the current semantic position, ensuring that the control parameters closely fit the fault characteristics, thereby enhancing the accuracy of fault-tolerant decisions. This method can not only intelligently generate new, unpreset, customized control parameter combinations from existing fault-specific control strategies, but also make these combinations highly consistent with the current fault semantic description, greatly improving the robustness and effectiveness of quadruped robots in fault early warning and adaptive fault-tolerant control in complex cable tunnel environments.

[0294] Step S43: Input the customized control parameter combination into the motion adaptation correction module in the unified fault-tolerant decision model. The motion adaptation correction module adaptively scales the original gait cycle parameter and the original joint swing amplitude in the customized control parameter combination based on the current joint motion data and joint torque margin evaluation value of the quadruped robot to obtain the corrected control parameter combination (corrected control parameter combination).

[0295] The motion adaptation correction module is a sub-module within this unified fault-tolerant decision model. Its role is to dynamically adjust and optimize the customized control parameter combination after it has been generated, by combining the current joint motion data and joint torque margin assessment values ​​of the quadruped robot. This ensures that the generated motion adjustment commands are physically feasible and maximize the quadruped robot's remaining motion capabilities. For example, this module could be a rule-based expert system that adjusts parameters based on the current joint motion data and joint torque margin assessment values. Alternatively, it could be a model predictive control-based module that optimizes the customized control parameter combination online to meet physical constraints. Current joint motion data refers to various kinematic and dynamic information of the quadruped robot at the current moment, such as current joint angles, current joint angular velocities, and current joint linear velocities. This data can be directly measured using encoders, inertial measurement units, foot force sensors, etc., or estimated by fusing multi-sensor data through state estimation algorithms. The joint torque margin assessment value refers to the difference between the rated maximum torque that the motor or actuator can provide for each joint of a quadruped robot in the current motion state and the actual output torque (i.e., the current joint torque measurement). It reflects the additional dynamic margin that the joint can provide when dealing with unexpected situations. For example, it can be calculated by real-time monitoring of joint current and motor temperature, combined with motor characteristic curves and safety margins. Alternatively, the actual output torque can be directly measured by a torque sensor and compared with the rated maximum torque of the motor. Gait cycle parameters refer to the time required for a quadruped robot to complete a full gait cycle, which is a key parameter affecting the quadruped robot's movement speed, stability, and energy consumption. Joint swing amplitude refers to the maximum range or distance that the joint or foot tip moves in space during the swing phase, affecting the quadruped robot's stride length, obstacle-crossing ability, and movement flexibility. The corrected control parameter combination is a parameter set that has been further adjusted and optimized by the motion adaptation correction module based on the current joint motion data and joint torque margin assessment value, based on the customized control parameter combination, making it more robust and physically feasible.

[0296] Step S43 above proposes to use a motion adaptation correction module to correct the customized control parameter combination. However, in its implementation, because the current joint motion data and joint torque margin assessment value are not fully considered, it is unable to dynamically adapt to the actual load capacity and motion stability requirements of the quadruped robot, resulting in an inaccurate corrected control parameter combination.

[0297] To address this, a further proposed method is used whereby the motion adaptation correction module adaptively scales the gait period parameter and joint swing amplitude in the customized control parameter combination based on the current joint motion data and joint torque margin assessment value of the quadruped robot, thereby obtaining a corrected control parameter combination. Specifically, this method includes:

[0298] Step S431: Calculate the current torque margin of each joint based on the difference between the actual torque output value of each joint at the current moment (i.e., the current joint torque measurement) and the torque limit value, and aggregate the current torque margins of each joint to obtain the global torque margin evaluation value (i.e., the joint torque margin evaluation value).

[0299] The actual torque output value refers to the actual torque generated by the motors of each joint of the quadruped robot at the current moment. It is usually estimated by combining the current sensor inside the motor driver with the motor parameters, or directly measured by a dedicated torque sensor. The torque limit value refers to the maximum output torque allowed by each joint motor under design or operational safety specifications. This value is usually determined by factors such as motor model, reduction ratio, driver performance, and robot structural strength, and can be read from the configuration file during system initialization or obtained through online calibration. Calculating the current torque margin of each joint involves comparing the actual torque output value with the torque limit value to quantify the remaining load capacity of each joint. For example, the absolute torque margin can be obtained by subtracting the actual torque output value from the torque limit value, or the normalized relative torque margin can be obtained by calculating (torque limit value - actual torque output value) / torque limit value. Aggregation refers to integrating the current torque margin information of each joint into a comprehensive index to reflect the overall torque margin of the entire quadruped robot. Aggregation methods may include: taking the minimum of all joint torque margins to reflect the limitation of the weakest link; calculating the average of all joint torque margins; or performing a weighted average based on the importance of each joint in motion (e.g., supporting joints have a higher weight than swinging joints).

[0300] Step S432: Based on the global torque margin assessment value, determine the adjustable range of the gait cycle parameter and the adjustable range of the joint swing amplitude, and use the upper limit of the adjustable range of the gait cycle parameter as the target gait cycle and the lower limit of the adjustable range of the joint swing amplitude as the target swing amplitude to construct an initial scaling target vector.

[0301] Determining the adjustable range refers to dynamically setting the allowable adjustment range of gait cycle parameters and joint swing amplitude parameters based on the global torque margin assessment value. For example, a mapping function can be preset to map the global torque margin assessment value to the safe range of gait cycle and swing amplitude. Alternatively, the corresponding adjustable range can be directly obtained according to different torque margin levels by consulting a pre-calibrated lookup table. When the global torque margin assessment value is low, it indicates that the quadruped robot's load is close to its limit. At this time, the adjustable range of the gait cycle parameter may be limited to a longer cycle range (i.e., slowing down the speed), while the adjustable range of the joint swing amplitude may be limited to a smaller amplitude range (i.e., reducing the stride length) to reduce torque demand. Using the upper limit of the adjustable range of the gait cycle parameter as the target gait cycle and the lower limit of the adjustable range of the joint swing amplitude as the target swing amplitude is based on safety and robustness considerations. In fault conditions, a more conservative motion strategy is usually adopted, that is, reducing the demand on joint torque by extending the gait cycle (reducing the speed) and reducing the joint swing amplitude (reducing the stride length), thereby avoiding overload and instability. This choice ensures that the quadruped robot can continue to move in the safest and most stable way under torque constraints. Constructing the initial scaling target vector means combining the target gait period and target swing amplitude determined above into a vector. This vector represents the ideal adjustment direction and target value of the quadruped robot's motion parameters, considering only the torque margin constraint. For example, this vector can be represented as [target gait period, target swing amplitude].

[0302] Step S433: Determine the tracking error between the actual motion trajectory and the expected motion trajectory of the quadruped robot at the current moment based on the current joint motion data, calculate the stability margin of the current motion state of the quadruped robot based on the tracking error, and use the stability margin as the scaling factor to dynamically adjust the initial scaling target vector to obtain the target scaling target vector.

[0303] The actual motion trajectory refers to the path actually executed by the quadruped robot's feet or body in space. It is determined through current joint motion data and can also be obtained through foot position sensors, IMUs, or visual odometry. The desired motion trajectory refers to the ideal path that the quadruped robot should follow according to preset gait planning or task instructions. Tracking error is the deviation between the actual trajectory and the desired trajectory, usually quantified by metrics such as Euclidean distance, mean squared error (MSE), or dynamic time warping (DTW). Calculating the stability margin of the quadruped robot's current motion state means evaluating the stability of the quadruped robot's current motion based on the trajectory tracking error. For example, a small tracking error indicates stable quadruped robot motion with a high stability margin. A large tracking error indicates that the quadruped robot motion may have an unstable trend, with a low stability margin. The stability margin can be a normalized scalar that reflects the robustness of the motion. The calculation method can be to convert the tracking error into a stability margin value between 0 and 1 through an inverse mapping function (such as 1 / (1+k×error)), or to evaluate it based on the error magnitude and rate of change using a fuzzy logic system, where k is the error sensitivity coefficient and error is the tracking error. Using the stability margin as a scaling factor to dynamically adjust the initial scaling target vector means using the stability margin as a weight or scaling factor to correct the initial scaling target vector. For example, when the stability margin is low, the scaling factor will further extend the target gait period and further reduce the target swing amplitude to enhance stability. When the stability margin is high, the scaling factor allows the target parameters to be closer to the parameters in the initial scaling target vector, even slightly improving performance within a safe range. This adjustment can be achieved by multiplying the initial scaling target vector with the stability margin, or by mapping it through a nonlinear function. The target scaling target vector, after being corrected by the stability margin, is a vector used to guide the adaptive scaling of parameters. It comprehensively considers constraints on both joint torque margin and motion stability. For example, this vector can be represented as [corrected target gait period, corrected target swing amplitude].

[0304] Step S434: The original gait cycle parameters and original joint swing amplitude in the customized control parameter combination are linearly interpolated to the target scaling target vector to generate the target gait cycle parameters and target joint swing amplitude after adaptive scaling. Based on the target gait cycle parameters and target joint swing amplitude after adaptive scaling, the corrected control parameter combination is determined.

[0305] Linear interpolation approximation refers to a mathematical method for smoothly transitioning between the original gait period parameters and original joint swing amplitude in a customized control parameter combination, as well as the target scaling target vector. For example, the formula P_new = P_original + alpha1 × (P_target - P_original) can be used, where alpha1 is an interpolation coefficient between 0 and 1, which can gradually change according to time or the number of iterations to achieve incremental parameter adjustment and avoid abrupt parameter changes. P_new is the adaptively scaled target gait period parameter or target joint swing amplitude, P_original is the original gait period parameter or original joint swing amplitude in the customized control parameter combination, and P_target is the corrected target gait period or corrected target swing amplitude in the target scaling target vector. Another approach is to use cubic spline interpolation to obtain a smoother transition curve. Generating adaptively scaled gait period parameters and joint swing amplitude refers to obtaining adaptively scaled target gait period parameters or target joint swing amplitude for motion control through a linear interpolation approximation process. These parameters consider both the initial intent of the fault semantic description and the current physical limitations and motion stability of the quadruped robot. Determining the corrected control parameter combination refers to reintegrating the adaptively scaled gait cycle parameters and joint swing amplitude, along with other unscaled parameters from the customized control parameter combination (such as foot landing point and fuselage height), into a complete set of control parameters that can be directly used to generate motion adjustment commands. This combination ensures that all necessary control parameters have been optimized based on the current state.

[0306] Through steps S431-S434, a motion adaptation correction module is introduced. Based on the current joint motion data and joint torque margin assessment value of the quadruped robot, the gait cycle parameter and joint swing amplitude in the customized control parameter combination are adaptively scaled to obtain a corrected control parameter combination. This process effectively solves the problem that directly applying the customized control parameter combination may lead to instability in the quadruped robot's motion or joint torque exceeding the safe range. Specifically, based on the actual torque output value and torque limit value of each joint, a global torque margin assessment value is calculated and aggregated, which can accurately quantify the current load capacity and potential overload risk of the quadruped robot. Based on the global torque margin assessment value, the adjustable range of the gait cycle parameter and joint swing amplitude is dynamically determined, and an initial scaling target vector is constructed using a conservative strategy (upper limit of gait cycle, lower limit of swing amplitude), ensuring that parameter adjustments are always made within the safe margin of joint torque and avoiding joint overload. In addition, the actual motion trajectory is determined based on the current joint motion data, and the stability margin of the quadruped robot's current motion state is calculated by combining the tracking error between the actual motion trajectory and the expected motion trajectory. Using this stability margin as a scaling factor, the initial scaling target vector is dynamically adjusted to obtain the target scaling target vector. This mechanism ensures that parameter adjustment not only considers torque constraints but also fully takes into account the real-time motion stability of the quadruped robot. When the quadruped robot's motion stability decreases, the stability margin guides the parameters to adjust in a more conservative direction, thereby enhancing motion robustness and avoiding trajectory deviations or oscillations caused by improper parameters. By linearly interpolating the original gait period parameters and original joint swing amplitudes in the customized control parameter combination to the target scaling target vector, adaptively scaled parameters are generated and combined into a corrected control parameter combination. This smooth interpolation method reduces abrupt changes in control commands and improves the continuity and reliability of system response. In summary, through the above process, when the quadruped robot encounters faults in complex environments such as cable tunnels, it can dynamically adjust its motion strategy according to its real-time physical constraints and motion stability requirements, thereby maximizing its fault tolerance and task execution efficiency while ensuring safety, and improving the quadruped robot's autonomous adaptability and reliability.

[0307] Step S44: Dynamically match the modified control parameter combination with the preset gait primitive library, select the corresponding gait primitive according to the matching result, and then generate a motion adjustment command that adapts to the current fault state and remaining motion capability.

[0308] The preset gait primitive library is a collection storing various basic gait patterns (such as walking, trotting, jogging, crawling, etc.). Each gait primitive contains a series of predefined joint trajectories, foot landing points, phase relationships, etc. Dynamic matching refers to the process of finding the gait primitive in the preset gait primitive library that best matches the current parameter characteristics based on the modified control parameter combination. For example, matching can be performed by calculating the distance or similarity between the modified control parameter combination and the parameters of each gait primitive in the library. Alternatively, a decision tree or classifier can be used to map the modified control parameter combination to the most suitable gait primitive. Gait primitives are used to construct a complete motion adjustment command. The motion adjustment command is a series of instructions generated to control the joint motors or actuators of the quadruped robot, including target joint angles, angular velocities, torques, etc., to achieve real-time adjustment of the quadruped robot's motion parameters, thereby responding to faults.

[0309] Through steps S41-S44, a unified fault-tolerant decision model is introduced, solving the problem in existing methods where control strategies cannot adaptively generate customized control parameter combinations based on fault semantic descriptions, leading to a lack of specificity and adaptability in the control strategies and affecting fault tolerance. Specifically, the policy routing network maps fault semantic descriptions to the policy embedding space and outputs the semantic similarity distribution between the fault semantic description and each fault-specific control strategy. This mechanism enables the system to assess the correlation betwee...

Claims

1. A method for generating a robot motion adjustment instruction based on a fault, characterized by, include: The robot acquires its current joint motion data, current joint torque measurement value, and current multidimensional physical field data at the current moment, wherein the physical field data includes at least voiceprint data, vibration data, electromagnetic field data, and energy consumption data. Based on the current joint motion data, determine the current joint torque prediction set of the robot; Based on the current joint torque prediction set and the current joint torque measurement value, the current fault characteristics of the robot are determined; Based on the current multidimensional physical field data, the current fault features, and the target knowledge graph, a fault semantic description of the robot is determined. The target knowledge graph includes fault mode nodes, intermediate state nodes, physical field feature nodes, and the relationships between nodes. The intermediate state nodes represent the internal mechanism degradation or transitional damage of the robot during the fault evolution process, and the physical field feature nodes represent abnormal phenomena that can be directly mapped according to the current multidimensional physical field data. Based on the fault semantic description, the joint drive command sequence of the robot in a predetermined time period is determined, wherein the predetermined time period is the time period after the current moment.

2. The method of claim 1, wherein, The step of determining the current joint torque prediction set of the robot based on the current joint motion data includes: Based on the current joint motion data, the robot's current joint angle, current joint angular velocity, and current joint current are determined. Based on the current joint angle, the current joint angular velocity, and the current joint current, a first heterogeneous motor model is used to obtain a predicted value of the first joint torque, wherein the first heterogeneous motor model is a physical mechanism model constructed based on the motor dynamics equations; Based on the current joint angle, the current joint angular velocity, and the current joint current, a second heterogeneous motor model is used to obtain a second joint torque prediction value. The second heterogeneous motor model has pre-learned the correlation between the current joint angle, the current joint angular velocity, the current joint current, and the second joint torque prediction value. Based on the current joint current, a third heterogeneous motor model is used to obtain the predicted value of the third joint torque. The third heterogeneous motor model is a pre-constructed lookup table model that characterizes the mapping relationship between the predicted joint current and the predicted joint torque. Based on the first joint torque prediction value, the second joint torque prediction value, and the third joint torque prediction value, the current joint torque prediction value set is obtained.

3. The method of claim 1, wherein, The step of determining the current fault characteristics of the robot based on the current joint torque prediction set and the current joint torque measurement value includes: The historical prediction consistency distribution characteristics and historical fluctuation patterns of the robot are obtained in a first historical time period. During this first historical time period, the robot is fault-free, and the historical environmental conditions corresponding to the first historical time period are the same as the current environmental conditions corresponding to the current moment. The historical prediction consistency distribution characteristics characterize the statistical distribution and statistical characteristics of the dispersion of the historical joint torque prediction value set within the first historical time period. The historical fluctuation patterns are obtained based on statistical analysis of the historical deviation value set within the first historical time period, which is determined based on the historical joint torque prediction value set and the historical joint torque measurement values ​​within the first historical time period. Based on the current predicted joint torque value set and the current measured joint torque value, determine the current deviation value set; Based on the current set of deviation values ​​and the historical fluctuation patterns, a set of deviation characteristics is determined; Based on the current set of predicted joint torque values, a current prediction consistency coefficient is determined, wherein the current prediction consistency coefficient characterizes the consistency between the first joint torque prediction value, the second joint torque prediction value, and the third joint torque prediction value in the current set of predicted joint torque values. Based on the current prediction consistency coefficient and the historical prediction consistency distribution characteristics, a consistency deviation characteristic is determined, wherein the consistency deviation characteristic characterizes the degree of deviation of the current prediction consistency coefficient from the historical prediction consistency distribution characteristics; Based on the deviation feature set and the consistency deviation feature, the current fault feature is determined.

4. The method according to claim 3, characterized in that, The process of determining the deviation feature set based on the current deviation value set and the historical fluctuation pattern includes: Based on the historical fluctuation patterns, determine the historical deviation mean and historical deviation standard deviation of the historical deviation value set; The current set of deviations is determined based on the current set of deviations and the historical average deviation. Based on the current set of differences and the historical standard deviations, determine the set of deviation multiples; Based on the aforementioned deviation multiple set, a dynamic confidence set is determined; The deviation feature set is determined based on the dynamic confidence set and the current deviation value set.

5. The method according to claim 3, characterized in that, The step of determining the current prediction consistency coefficient based on the current set of joint torque prediction values ​​includes: Determine the historical consensus center value, wherein the historical consensus center value is determined based on the set of historical joint torque prediction values ​​for the second historical time period; Based on the current set of predicted joint torque values ​​and the historical consensus center value, a set of deviation distances is determined, wherein the set of deviation distances includes a first deviation distance between a first predicted joint torque value and the historical consensus center value, a second deviation distance between a second predicted joint torque value and the historical consensus center value, and a third deviation distance between a third predicted joint torque value and the historical consensus center value; The current prediction consistency coefficient is determined based on the deviation distance set.

6. The method according to claim 3, characterized in that, The determination of consistency deviation characteristics based on the current prediction consistency coefficient and the historical prediction consistency distribution characteristics includes: Based on the historical prediction consistency distribution characteristics, a benchmark membership function is constructed; Based on the current joint angle in the current joint motion data and the historical joint angle in the third historical time period, the joint angular velocity fluctuation amplitude is determined. The target membership function is obtained by correcting the reference membership function based on the joint angular velocity fluctuation amplitude. Based on the current prediction consistency coefficient and the target membership function, the initial membership degree is obtained; The initial membership degree is corrected based on the deviation feature set to obtain the consistency deviation feature.

7. The method according to claim 1, characterized in that, The step of determining the robot's fault semantic description based on the current multidimensional physical field data, the current fault features, and the target knowledge graph includes: Based on the current fault characteristics and the historical fault characteristics of the fourth historical time period, a timing guidance signal is determined, wherein the timing guidance signal characterizes the evolution direction and evolution rate of the fault characteristics. Feature extraction is performed on the current multi-dimensional physical field data to obtain an initial multi-physics field synchronization feature sequence; Based on the time-series guidance signal and the target knowledge graph, a fault evolution path is determined, wherein the fault evolution path starts from the physical field feature node and ends at the fault mode node; Based on the node association order of the fault evolution path, the initial multiphysics synchronization feature sequence is reordered and phase-aligned to obtain the target multiphysics synchronization feature sequence. The fault semantic description is determined based on the target multiphysics synchronization feature sequence.

8. The method according to any one of claims 1 to 7, characterized in that, The step of determining the joint drive command sequence of the robot within a predetermined time period based on the fault semantic description includes: Based on the semantic similarity between the fault semantic description and multiple fault-specific control strategy vectors, multiple candidate fault-specific control strategies are selected from multiple fault-specific control strategies. Based on the control parameters corresponding to the multiple candidate fault-specific control strategies, a customized combination of control parameters for the robot is determined. Based on the current joint motion data and joint torque margin assessment value, the gait cycle parameter and joint swing amplitude in the customized control parameter combination are corrected to obtain the corrected control parameter combination. The joint torque margin assessment value is determined based on the current joint torque measurement value, and the gait cycle parameter refers to the time required for the robot to complete a complete gait cycle. Based on the modified control parameter combination, the motion adjustment command of the robot is determined; Based on the motion adjustment command, the joint drive command sequence is determined.

9. A fault-based robot motion adjustment instruction generation device, characterized in that, include: The data acquisition module is used to acquire the robot's current joint motion data, current joint torque measurement value, and current multidimensional physical field data at the current moment. The physical field data includes at least voiceprint data, vibration data, electromagnetic field data, and energy consumption data. The first determining module is used to determine the current joint torque prediction value set of the robot based on the current joint motion data; The second determining module is used to determine the current fault characteristics of the robot based on the current joint torque prediction value set and the current joint torque measurement value; The third determining module is used to determine the fault semantic description of the robot based on the current multidimensional physical field data, the current fault features, and the target knowledge graph. The target knowledge graph includes fault mode nodes, intermediate state nodes, physical field feature nodes, and the relationships between nodes. The intermediate state nodes represent the internal mechanism degradation or transitional damage of the robot during the fault evolution process, and the physical field feature nodes represent abnormal phenomena that can be directly mapped according to the current multidimensional physical field data. The fourth determining module is used to determine the joint driving command sequence of the robot within a predetermined time period based on the fault semantic description, wherein the predetermined time period is the time period after the current moment.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores multiple instructions, which are adapted to be loaded by a processor and executed by the fault-based robot motion adjustment instruction generation method according to any one of claims 1 to 8.

Images