A dual-mode self-calibration method and system for lower limb assistive exoskeleton sensors

Abstract

The application discloses a kind of lower limb power-assisted exoskeleton sensor's dual-mode self-calibration method and system, belongs to intelligent control system and robot technical field, it includes acquisition real-time multi-source sensing data flow and lower limb movement physical constraint relationship dynamic library;To data stream is weighted constraint verification calculation, obtains consistency index set;Through exception combination search and correlation dependency analysis, generate consistency exception mode;To the mode is abnormal source attribution analysis in time and space dimension, obtains calibration execution strategy;To real-time multi-source sensing data flow executes local parameter perturbation correction or global parameter optimization solution, updates sensor parameter set.The application adopts the data consistency verification means based on physical constraint to carry out online monitoring to sensor network, and according to the isolated or systemic abnormal mode diagnosed, selects to execute local parameter perturbation correction or global parameter optimization solution, can realize the self-diagnosis and adaptive calibration of control system sensor fault.

Description

Technical Field

[0001] This invention relates to the field of intelligent control systems and robotics, and in particular to a dual-mode self-calibration method and system for a lower limb assistive exoskeleton sensor. Background Technology

[0002] A lower limb assistive exoskeleton is a mechatronic device worn on the lower limbs. It uses an integrated sensor network to sense the user's movement intentions and status in real time, and a controller drives actuators to provide corresponding assistive torque to help the user perform actions such as walking and standing. This sensor network is the foundation of the entire control system and typically includes joint angle sensors, inertial measurement units, and plantar pressure sensors. The accuracy of its measurement data directly determines the exoskeleton's performance, safety, and user experience.

[0003] In related technologies, Chinese invention patent application CN121222037A discloses a lower limb weight-bearing gait rehabilitation training system, including a data acquisition module, a data processing and analysis module, a patient individualization modeling module, an intelligent decision-making and control module, a rehabilitation execution module, and a human-computer interaction and medical information interface module. These modules are connected via network communication. The data acquisition module is used to acquire multimodal data of the patient in real time, including static vital signs data, dynamic physiological parameters, kinematic and dynamic parameters, and non-motor physiological and psychological state data. The data processing and analysis module is used to preprocess, extract features, and perform deep analysis on the raw data, outputting structured patient individualization feature vectors and evaluation results.

[0004] However, while the aforementioned solution integrates multiple high-precision sensors such as inertial measurement units, rotary encoders, and plantar pressure sensors to acquire patients' multimodal motion parameters, its data processing logic is mainly limited to basic filtering, denoising, and timing alignment of single-modal signals, lacking a multi-source data consistency verification mechanism based on human biomechanical physical constraints. During long-term weight-bearing training, the system's preset static calibration parameters are prone to failure due to factors such as sensor temperature drift, mechanical vibration causing loose installation, or slippage of the wearing position. Existing technologies cannot distinguish between isolated anomalies of a single sensor and systemic correlation deviations, nor do they possess the ability to dynamically self-correct parameters online based on physical constraints during operation. This leads to a decrease in the accuracy of the digital twin model over time, thereby affecting the precision and safety of rehabilitation control. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a dual-mode self-calibration method and system for lower limb assistive exoskeleton sensors. It employs a data consistency verification method based on physical constraints to monitor the sensor network online, and selects to perform local parameter perturbation correction or global parameter optimization solution based on the diagnosed isolated or systematic abnormal modes. This enables autonomous diagnosis and adaptive calibration of sensor faults in the control system.

[0006] The above objectives can be achieved through the following approach: A dual-mode self-calibration method for a lower limb assistive exoskeleton sensor includes: acquiring real-time multi-source sensor data streams during the operation of the lower limb assistive exoskeleton and a dynamic library of physical constraint relationships describing lower limb motion constraint relationships; performing weighted constraint verification calculations on the real-time multi-source sensor data streams using the dynamic library of physical constraint relationships and user motion state recognition results to obtain a consistency index set; performing violation combination retrieval and correlation dependency analysis on the consistency index set to generate consistency violation patterns characterizing systematic anomalies or isolated anomalies; performing spatiotemporal anomaly source attribution analysis on the consistency violation patterns to obtain a calibration execution strategy containing a calibration target index and correction algorithm identifiers; and, in response to the calibration execution strategy, performing local parameter perturbation correction or global parameter optimization on the real-time multi-source sensor data streams to update the sensor parameter set.

[0007] Optionally, acquiring the real-time multi-source sensor data stream during the operation of the lower limb-assisted exoskeleton and the dynamic library of physical constraint relationships used to describe the lower limb motion constraint relationships includes: synchronously acquiring joint angle sensor data, inertial measurement unit data, and plantar pressure sensor data to obtain a real-time multi-source sensor data stream; performing a deviation statistical analysis on the foot position calculated from the joint angle sensor data and the foot position estimated from the inertial measurement unit data to generate a first geometric constraint interval; performing a mechanical balance calculation on the total ground reaction force measured from the plantar pressure sensor data and the real-time body weight support component obtained through human kinematics calculation to generate a second mechanical constraint interval; and integrating the first geometric constraint interval and the second mechanical constraint interval to obtain the dynamic library of physical constraint relationships.

[0008] Optionally, obtaining the consistency index set includes: acquiring the current user motion state recognition result, assigning weights to the first geometric constraint interval and the second mechanical constraint interval to generate a confidence weight set; and using the confidence weight set to perform weighted statistics on the real-time multi-source sensor data stream to generate the consistency index set.

[0009] Optionally, generating consistency violation patterns representing systematic or isolated anomalies includes: retrieving index items from the consistency index set that simultaneously exceed the first geometric constraint interval and the second mechanical constraint interval, generating a violation constraint combination; performing correlation dependency analysis on the violation constraint combination to generate sensor anomaly source pointing information; determining the number of sources in the sensor anomaly source pointing information to generate an isolated sensor anomaly pattern representing a single-source anomaly or a systematic associated anomaly pattern representing multiple-source anomalies, as consistency violation patterns.

[0010] Optionally, obtaining the calibration execution strategy containing the calibration target index and the correction algorithm identifier includes: extracting the duration characteristics of the consistency violation pattern in the time dimension and the sensor topology distribution characteristics in the spatial dimension; performing cluster analysis on the consistency violation pattern using the duration characteristics and the sensor topology distribution characteristics to obtain the abnormal influence domain; matching the abnormal influence domain with a preset sensor fault mapping table to generate a calibration target index pointing to a specific sensor and a corresponding parameter correction algorithm identifier; and integrating the calibration target index and the parameter correction algorithm identifier to generate the calibration execution strategy.

[0011] Optionally, the execution of local parameter perturbation correction includes: in response to the calibration execution strategy containing an identifier of an isolated sensor anomalous mode, extracting the corresponding target sensor data according to the calibration target index; performing micro-perturbation calculation on the target sensor data and the current sensor parameters to generate candidate parameter update amounts; and using the candidate parameter update amounts to perform tentative correction on the current sensor parameters to generate a temporary parameter set.

[0012] Optionally, the method further includes: re-performing the constraint verification calculation on the real-time multi-source sensing data stream using the temporary parameter set to obtain a verified consistency index set; comparing the deviation magnitude between the verified consistency index set and the consistency index set to generate a consistency improvement judgment result; if the consistency improvement judgment result is true, then adding the candidate parameter update amount to the current sensor parameters to generate an updated sensor parameter set.

[0013] Optionally, the global parameter optimization solution includes: in response to the identification of a systematically associated abnormal pattern in the calibration execution strategy, collecting the calibration sensor data stream during the user's execution of a preset calibration action sequence; using the physical constraint relationship dynamic library as the convergence target, substituting the calibration sensor data stream into the optimizer for global solution, and generating a reconstructed sensor parameter set.

[0014] Optionally, substituting the calibration sensor data stream into the optimizer for global solution includes: performing state prediction calculations using the current sensor parameters and the calibration sensor data stream to obtain predicted output data; calculating the sum of deviations between the predicted output data and all associated constraints in the physical constraint relationship dynamic library to obtain an overall consistency deviation value; performing a minimization iterative operation on the overall consistency deviation value until the deviation converges to determine the reconstructed sensor parameter set.

[0015] Based on the same inventive concept, this invention also provides a dual-mode self-calibration system for a lower limb assistive exoskeleton sensor, comprising: a data and constraint acquisition module, used to acquire real-time multi-source sensor data streams during the operation of the lower limb assistive exoskeleton and a dynamic library of physical constraint relationships describing lower limb movement constraint relationships; a weighted constraint verification module, used to perform weighted constraint verification calculations on the real-time multi-source sensor data streams using the dynamic library of physical constraint relationships and user movement state recognition results to obtain a consistency index set; a violation pattern analysis module, used to perform violation combination retrieval and correlation dependency analysis on the consistency index set to generate consistency violation patterns characterizing systematic anomalies or isolated anomalies; a calibration strategy generation module, used to perform spatiotemporal dimension anomaly source attribution analysis on the consistency violation patterns to obtain a calibration execution strategy containing a calibration target index and a correction algorithm identifier; and a parameter correction execution module, used to respond to the calibration execution strategy by performing local parameter perturbation correction or global parameter optimization on the real-time multi-source sensor data streams to update the sensor parameter set.

[0016] Compared with the prior art, the present invention has the following advantages: This invention achieves continuous online monitoring of sensor data streams by constructing a dynamic reference frame based on kinematic and dynamic physical constraints. This method requires no external calibration equipment or system shutdown, and can assess data consistency in real time during the normal operation of the exoskeleton. It can promptly detect sensor parameter drift caused by factors such as component aging, environmental changes, or physical impacts, thereby ensuring the long-term accuracy and reliability of the control system's input signals.

[0017] The consistency violation pattern analysis mechanism proposed in this invention can perform in-depth diagnosis and attribution of monitored abnormal data. Through the retrieval of violation combinations and correlation dependency analysis, the system can not only determine the existence of anomalies, but also distinguish between isolated anomalies caused by a single sensor failure and systemic anomalies caused by the coupling effect of multiple sensors, and locate the specific source of the anomaly. This precise fault diagnosis capability lays the foundation for subsequent targeted correction strategies, avoids blindly adjusting parameters, and improves calibration efficiency and accuracy.

[0018] The dual-mode calibration execution strategy employed in this invention enhances the system's adaptive repair capabilities and operational efficiency. For simple, isolated anomalies, the system uses rapid local parameter perturbation correction to quickly restore sensor performance with minimal computational overhead. For complex systemic anomalies, a comprehensive global parameter optimization solution is initiated to reconstruct the parameter system. This hierarchical processing and on-demand calibration approach achieves an optimal balance between correction speed and correction depth, ensuring that the exoskeleton control system exhibits strong robustness and adaptability when facing sensor failures of varying degrees.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a dual-mode self-calibration method for a lower limb assistive exoskeleton sensor according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the dynamic generation of constraint intervals according to an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the convergence process of global parameter optimization solution in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of a dual-mode self-calibration system for a lower limb assistive exoskeleton sensor according to an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Reference Figure 1One embodiment of the present invention proposes a dual-mode self-calibration method and system for a lower limb assistive exoskeleton sensor. It employs a data consistency verification method based on physical constraints to monitor the sensor network online, and selects to perform local parameter perturbation correction or global parameter optimization solution based on the diagnosed isolated or systematic abnormal modes, thereby enabling autonomous diagnosis and adaptive calibration of sensor faults in the control system.

[0027] The method described in this embodiment specifically includes: S1. Acquire real-time multi-source sensor data streams during the operation of the lower limb assistive exoskeleton and a dynamic library of physical constraint relationships used to describe the lower limb motion constraint relationships; S2. Using the physical constraint relationship dynamic library and the user motion state recognition results, perform weighted constraint verification calculation on the real-time multi-source sensor data stream to obtain a consistency index set; S3. Perform violation combination retrieval and correlation dependency analysis on the consistency index set to generate consistency violation patterns that characterize systematic anomalies or isolated anomalies. S4. Perform spatiotemporal anomaly source attribution analysis on the consistency violation pattern to obtain a calibration execution strategy that includes the calibration target index and the correction algorithm identifier. S5. In response to the calibration execution strategy, perform local parameter perturbation correction or global parameter optimization on the real-time multi-source sensor data stream, and update the sensor parameter set.

[0028] Optionally, the acquisition of real-time multi-source sensor data streams during the operation of the lower limb-assisted exoskeleton and a dynamic library of physical constraint relationships used to describe the lower limb motion constraint relationships includes: Simultaneously acquire joint angle sensor data, inertial measurement unit data, and plantar pressure sensor data to obtain a real-time multi-source sensor data stream; A first geometric constraint interval is generated by performing a statistical analysis of the deviation between the foot position calculated from the joint angle sensor data and the foot position estimated from the inertial measurement unit data. A second mechanical constraint interval is generated by performing a mechanical balance calculation on the total ground reaction force measured by the foot pressure sensor data and the real-time body weight support component obtained by human kinematics calculation. By integrating the first geometric constraint interval and the second mechanical constraint interval, a dynamic library of physical constraint relationships is obtained.

[0029] Specifically, the central processing unit simultaneously acquires data from multiple sensors at a sampling frequency of no less than 100 Hz. These sensors include rotary encoders mounted at the hip, knee, and ankle joints, providing real-time joint angle sensor data; a nine-axis inertial measurement unit mounted on the foot or lower leg linkage, providing inertial measurement unit data including acceleration, angular velocity, and magnetic field strength; and a flexible pressure sensor array integrated under the insole or sole plate, providing plantar pressure sensor data. Each acquired data stream is accompanied by a high-precision timestamp and integrated into a unified format real-time multi-source sensor data stream as input for subsequent processing.

[0030] Based on the known length of the exoskeleton links and real-time joint angles obtained from joint angle sensor data The relative position from the hip joint base to the foot is calculated. Gravity component compensation and double integration are performed on the acceleration signal from the foot inertial measurement unit data to estimate the displacement change at the foot end, thus obtaining the foot position. Since inertial integration has inherent drift, while forward kinematics calculation is relatively accurate, constraints are generated by calculating the Euclidean distance deviation between the two position vectors. This deviation... It can be represented as: , in It is the three-dimensional position vector of the foot calculated using a positive kinematics model. The three-dimensional position vector of the foot is obtained by integrating the data from the inertial measurement unit. Within a sliding time window of 1 to 2 seconds, the deviation is... Perform statistical analysis and calculate its moving average. and moving standard deviation First geometrically constrained interval This is generated from the statistical results and confidence coefficients. The coefficient is determined. The value is typically between 2 and 3, corresponding to a confidence level of approximately 95% to 99.7%.

[0031] Total ground reaction force It is the vertical ground reaction force obtained directly by integrating and summing the data from the plantar pressure sensor. Real-time body weight support component. This is obtained through an inverse dynamics model of the human body and exoskeleton system. Under simplified quasi-static or slow motion conditions, this component can be approximated as the user's body weight. With gravitational acceleration The product of these two is dynamically adjusted based on the current motion state. The mechanical deviation between them is calculated. : , Here It is a scalar value of the total ground reaction force obtained from data measured by plantar pressure sensors. This is the real-time scalar value of the body weight support component required to maintain the current motion state, calculated through an inverse dynamics model. Similar to generating the first geometric constraint interval, the deviation is calculated within a sliding time window. moving average and moving standard deviation And combined with confidence coefficient Determine the second mechanical constraint interval The boundary. For example... Figure 2 As shown, this illustrates the process of constructing and dynamically generating geometric and mechanical constraint intervals in the dynamic library of physical constraint relationships.

[0032] The generated geometric and mechanical constraints are integrated to form a dynamic library of physical constraint relationships. This dynamic library of physical constraint relationships contains the first geometric constraint interval. Second mechanical constraint interval The data structure is dynamic, and the statistical parameters constituting these two intervals, namely the moving average and standard deviation, are constantly updated as the system continues to run without anomalies, allowing the constraint library to adapt to different user movement patterns or environmental changes.

[0033] For example, a user walks on flat ground at a speed of 0.8 m / s, and the central processing unit synchronously collects data at a frequency of 100 Hz via a rotary encoder, inertial measurement unit, and plantar pressure sensor. At the sampling time... The system uses joint angle data and a 0.5-meter link length to calculate the foot position vector using forward kinematics. Meters; simultaneously, gravity compensation and quadratic integration are performed on the inertial measurement data to estimate the foot position vector. Meters. The system calculates the Euclidean distance deviation between the two. Meters. Simultaneously, the total ground reaction force was measured. Newton, while the weight support component calculated by the inverse dynamics model. Newton, calculating mechanical deviations Newton. The system statistically analyzes data within a sliding time window of 2 seconds to obtain the moving average of the geometric deviation. Meters, moving standard deviation Meters. Set confidence coefficient. The first geometric constraint interval is determined based on the statistical results. Boundary: Lower limit is meters, with an upper limit of meters, that is, the generated interval is Meters. Similarly, for mechanical deviations, a moving average is statistically obtained. Newton, moving standard deviation Newton, setting the confidence coefficient The lower limit of calculation is Newton, because the physical definition of deviation is non-negative, the actual lower limit is 0, and the upper limit is... Newton, that is, determining the second mechanical constraint interval for Newton. This method achieves spatiotemporal alignment and fusion of multi-source heterogeneous data. By calculating the upper and lower boundaries of the statistical distribution, a strict physical constraint interval is constructed, ensuring that the constraint conditions can both adaptively follow changes in the user's motion pattern and effectively contain normal sensor noise fluctuations and accurately identify abnormal deviations through the double boundaries.

[0034] Optionally, the obtained consistency index set includes: Obtain the current user motion state recognition result, assign weights to the first geometric constraint interval and the second mechanical constraint interval, and generate a confidence weight set; The confidence weight set is used to perform weighted statistics on the real-time multi-source sensor data stream to generate a consistency index set.

[0035] Specifically, finite state machines or machine learning models such as Hidden Markov Models are used to analyze the real-time multi-source sensor data stream, especially the temporal characteristics of plantar pressure sensor data and joint angle sensor data, to determine the user's current motion state, such as the foot-on-ground phase, swing phase, or standing balance phase. After obtaining the user's motion state identification result, the corresponding weights are retrieved from a preset weight mapping table to generate a confidence weight set. For example, during the swing phase when the foot is completely off the ground, the plantar pressure sensor data is theoretically zero, and the reliability of the second mechanical constraint interval is extremely low at this time, so a weight value close to zero is assigned to it. For example, values ​​from 0.0 to 0.2, while assigning weights close to one to the first geometric constraint interval. For example, 0.8 to 1.0; conversely, during the single-leg stable support period, the foot is stationary relative to the ground, and the drift error of the inertial measurement unit data integration will accumulate, leading to a decrease in the reliability of the first geometric constraint interval, thus requiring a lower allocation. and higher .

[0036] Calculate the normalized geometric deviation at the current time. and normalized mechanical deviation These two normalized biases are obtained by calculating the geometric biases in real time. and mechanical deviation The original deviations, which have units of length and force, are converted into dimensionless ratios by dividing each deviation by the upper tolerance limit of its corresponding constraint interval. The final consistency index is then calculated using the following formula. .

[0037] , in, The overall consistency index value represents the current moment; and These are the weight coefficients for the first geometric constraint interval and the second mechanical constraint interval, obtained from the set of confidence weights. It is the normalized geometric bias, which is achieved by adjusting the real-time geometric bias. Divide by the width of the first geometric constraint interval to obtain; It is the normalized mechanical deviation, which is achieved by adjusting the real-time mechanical deviation. This is obtained by dividing by the width of the second mechanical constraint interval. This calculation is performed continuously at a frequency of at least 100 Hz to generate a consistency index that varies over time. The sequence of values ​​in this sequence constitutes a set of consistency indicators. The higher the value in this set of indicators, the greater the deviation of the real-time multi-source sensor data stream from physical constraints, and the higher the probability of anomalies.

[0038] For example, the system identifies that the user is currently in a "single-foot stable support phase," at which point the foot contact is stable but the inertial components are prone to integral drift. Therefore, a table is consulted to assign weights. Assign weights to the first geometrically constrained interval. Given the second mechanical constraint interval. The real-time geometric deviation measured at the current moment. Meters, exceeding the constraint limit of 0.030 meters; real-time mechanical deviation. Newtons, exceeding the upper limit of the 11-Newton constraint. The system calculates the normalized geometric deviation. Normalized mechanical deviation Then, the overall consistency index was calculated using a weighted formula. .because The system determines that the current data stream has an inconsistency anomaly. This method introduces a dynamic weight allocation mechanism that depends on the motion state, which can intelligently adjust the focus of constraint verification according to the differences in the physical characteristics of sensors at different motion stages, and transform physical deviations of different dimensions into a unified dimensionless consistency index, thereby achieving a precise quantitative assessment of the system's operating status.

[0039] Optionally, the generation of consistency violation patterns representing systematic or isolated anomalies includes: Retrieve index items from the consistency index set that simultaneously exceed the first geometric constraint interval and the second mechanical constraint interval, and generate a violation constraint combination; Correlation dependency analysis is performed on the combination of violation constraints to generate sensor anomaly source pointing information; The number of sources is determined for the sensor anomaly source pointing information, and an isolated sensor anomaly pattern representing a single source anomaly or a systematic associated anomaly pattern representing multiple source anomalies is generated as a consistency violation pattern.

[0040] Specifically, at frequencies no lower than 50 Hz, the normalized geometric deviation is continuously monitored. and normalized mechanical deviation When detection occurs within the same sampling moment or a very short time window, such as 10 to 20 milliseconds, and When the values ​​are all greater than 1.0, that is, the real-time geometric deviation... Exceeding the first geometric constraint interval The range, and real-time mechanical deviation It also exceeds the second mechanical constraint range. If the range is specified, a double constraint violation is determined. The timestamp of the event and the relevant deviation value will be recorded to generate an instance of the violation constraint combination.

[0041] For each combination of violation constraints, the process traces back to the original input that caused the violation, i.e., the real-time multi-source sensor data stream. A sensitivity analysis-based approach is used to quantify the dependency of the violation deviation on each sensor input by calculating its partial derivatives. For example, for geometric deviations... Calculate its data related to the angle sensors of each joint. and inertial measurement unit data The sensitivity coefficient. For mechanical deviations. Then calculate its data related to the plantar pressure sensor. And the sensitivity coefficients used to solve the joint state quantities of the dynamic model. These sensitivity coefficients... It can be represented as: , Here It is a sensor The sensitivity coefficient, Represents geometric deviation or mechanical deviation , Representative from sensor The raw data readings are used. Sensors with higher sensitivity coefficients mean that even small fluctuations in their data will have a greater impact on the final deviation results, and are therefore considered more suspicious sources of anomalies. The sensitivity coefficients of all relevant sensors are sorted to generate a list ranked in descending order of suspicion, serving as information pointing to sensor anomalies.

[0042] Analyze the list of sensor anomaly sources. If the sensitivity coefficient of the sensor ranked first in this list... If a sensor's sensitivity coefficient is significantly higher than all other sensors—for example, more than an order of magnitude higher than the second-highest sensitivity coefficient—then the anomaly is determined to be from a single source. In this case, an isolated sensor anomaly pattern representing a single-source anomaly is generated. Conversely, if two or more sensors in the list have similar sensitivity coefficients that are all significantly higher than other sensors, it indicates that the anomaly is caused by multiple data sources. This usually points to a deeper systemic problem, such as physical offset of the sensor's installation location affecting the measurement of multiple joint angles. In this case, a systemic correlated anomaly pattern representing a multi-source anomaly is generated. These two patterns together constitute the output of the consistency violation pattern, providing a decision-making basis for subsequent calibration strategy selection.

[0043] For example, the monitoring system continuously scans the sensor data stream at a frequency of 50 Hz. Within a specific 15-millisecond time window, the normalized geometric deviation reaches 1.5 and the normalized mechanical deviation... If both values ​​reach 1.2, exceeding the threshold of 1.0, the system determines a double constraint violation and records the violation event with a timestamp. For this combination of violations, the system traces back to the corresponding real-time multi-source sensor data stream, initiates the sensitivity analysis process, and applies the formula... Calculate the sensitivity coefficients of each sensor input. The system reads the knee joint angle sensor values. Applying a numerical differential increment of 0.01 radians results in a geometric deviation. A change of 0.05 meters yields the following calculations: Simultaneously calculate inertial measurement unit data. The sensitivity coefficient is and plantar pressure sensor data Regarding mechanical deviation The sensitivity coefficient is The system generates a list of sensor anomaly source information, sorted in descending order, based on the above coefficients. Analysis shows that the number one ranked The second highest The value is an order of magnitude larger than the expected value, which does not conform to the characteristic of similar values ​​from multiple sources. Therefore, the system determines that the anomaly source is a single source and generates a "single sensor anomaly pattern" violation combination pointing to the knee joint encoder. This method, through quantitative sensitivity analysis and ranking comparison based on partial derivatives, can accurately quantify the contribution of each sensor to the system failure at a deeper data level. It can also intelligently distinguish between single device failures and complex systemic deviations based on sensitivity distribution characteristics, thus providing an objective and interpretable decision-making basis for subsequent accurate matching and calibration strategies.

[0044] Optionally, the calibration execution strategy that includes the calibration target index and the correction algorithm identifier includes: Extract the duration characteristics of the consistency violation pattern in the time dimension and the sensor topology distribution characteristics in the spatial dimension; Cluster analysis is performed on the consistency violation patterns using the duration characteristics and sensor topology distribution characteristics to obtain the anomaly impact domain; Based on the abnormal influence domain, a preset sensor fault mapping table is matched to generate a calibration target index pointing to a specific sensor and a corresponding parameter correction algorithm identifier. Integrate the calibration target index and the parameter correction algorithm identifier to generate a calibration execution strategy.

[0045] Specifically, for each consistency violation pattern, a two-dimensional analysis is performed. In the temporal dimension, its duration is extracted, calculating the time from the onset of the violation pattern to its disappearance. A threshold of 200 milliseconds is set; patterns shorter than this value are marked as transient anomalies, while those longer are marked as persistent anomalies. In the spatial dimension, its sensor topology distribution features are extracted. Here, sensor topology refers to a pre-stored data structure describing the relative positions of all sensors on the exoskeleton's physical structure, such as which sensors belong to the same leg or which sensors are on adjacent joints. The sensor anomaly source information associated with the violation pattern is mapped onto this topology, and its distribution patterns are analyzed to generate topology distribution descriptors such as "single joint isolation," "ipsilateral limb chain," or "bilateral symmetrical joints."

[0046] Cluster analysis is performed on consistency violation patterns using extracted spatiotemporal features to determine their anomalous influence domains. Feature vectors composed of duration and sensor topology distribution features are used as input, and a pre-trained classifier, such as a decision tree or support vector machine model, is applied for cluster analysis. This analysis classifies different violation patterns into several predefined categories, each category representing an anomalous influence domain. For example, feature vectors of "transient anomalies" and "single joint isolation" are categorized into the "isolated sensor noise spike" influence domain; while feature vectors of "persistent anomalies" and "ipsilateral limb chain-like" are categorized into the "limb segment link parameter drift" influence domain.

[0047] Based on the identified anomaly impact domain, a pre-defined sensor fault mapping table is used to generate specific calibration instructions. This sensor fault mapping table is a built-in expert knowledge base that stores the correspondence between anomaly impact domains and calibration strategies in key-value pairs. For example, the mapping table contains the following rules: if the anomaly impact domain is "isolated sensor noise spikes," then its corresponding calibration target index points to that isolated sensor, and the correction algorithm identifier is specified as the zero-bias fast calibration algorithm in "local parameter perturbation correction"; if the anomaly impact domain is "limb segment link parameter drift," then the calibration target index will include all relevant sensors on that limb chain, such as angle sensors for the hip, knee, and ankle joints, and the correction algorithm identifier is specified as "global parameter optimization solution." Using the anomaly impact domain as the query index, the corresponding calibration target index and parameter correction algorithm identifier are retrieved from this table.

[0048] The acquired calibration target index, i.e., a single sensor ID or a list of sensor IDs, is combined with the parameter correction algorithm identifier, i.e., an enumerated value or string pointing to a specific algorithm module, into a unified data structure. This structure is the final generated calibration execution strategy, which is passed to the parameter correction execution module to trigger subsequent local parameter perturbation correction or global parameter optimization solution processes.

[0049] For example, a multi-dimensional feature analysis is initiated for the violation pattern pointing to the encoder of the right knee joint. In the temporal dimension, the system extracts the duration of the anomaly as 50 milliseconds and compares it to a 200-millisecond threshold. Since it is less than the threshold, it is marked as a "transient anomaly." In the spatial dimension, the system maps the anomaly source to a pre-stored sensor topology. Analysis reveals that the anomaly involves only the single node of the right knee joint and does not affect adjacent joints, thus generating a "single-joint isolated" topology distribution descriptor. The system then constructs a feature vector from the extracted "transient anomaly" temporal features and "single-joint isolated" spatial features, inputs it into a pre-trained decision tree classifier model for cluster analysis, and the analysis results clearly classify the violation pattern into the "isolated sensor noise spike" anomaly influence domain. Based on this determined anomaly influence domain, the system performs index matching in the built-in sensor fault mapping table. Based on the key-value pair rules stored in the mapping table, for the influence domain of "isolated sensor noise spikes," the system retrieves the corresponding calibration target index as the single sensor ID "ID_Right_Knee_Encoder" and assigns the parameter correction algorithm identifier as "Zero_Bias_Rapid_Calib," corresponding to the zero-bias fast calibration algorithm in local parameter perturbation correction. Finally, the system combines the obtained calibration target index and parameter correction algorithm identifier to generate a unified data structure containing complete instruction information—the calibration execution strategy. This method, through refined extraction of spatiotemporal features and machine learning-based clustering inference, combined with rule mapping from an expert knowledge base, achieves automated and intelligent transformation from abstract violation patterns to specific executable calibration strategies, ensuring that the system can accurately match effective repair solutions for faults of different natures.

[0050] Optionally, the process of performing local parameter perturbation correction includes: In response to the calibration execution strategy including an identifier of an isolated sensor anomaly mode, the corresponding target sensor data is extracted according to the calibration target index; Perform micro-perturbation calculations on the target sensor data and current sensor parameters to generate candidate parameter update values; The current sensor parameters are tentatively corrected using the candidate parameter update amount to generate a temporary parameter set.

[0051] Specifically, once a calibration execution strategy containing an isolated sensor anomaly pattern identifier is received, the system immediately extracts the existing parameters of the target sensor from the current sensor parameter set based on the calibration target index in the strategy, such as the zero-point offset or gain coefficient of the joint angle sensor. Simultaneously, it retrieves the small segment of historical data associated with the target sensor that led to the anomaly determination from the buffer of the real-time multi-source sensor data stream.

[0052] A small perturbation calculation is performed on the current parameters of the target sensor to generate candidate parameter updates. The core of this calculation is based on the idea of ​​gradient descent, that is, fine-tuning the parameters along the direction that causes the consistency index to decrease most rapidly. The consistency index is calculated online using either the sensitivity of the sensor parameters to the deviations calculated during the violation analysis phase, or through numerical methods. For target sensor parameters The partial derivatives. Candidate parameter update amount. The calculation formula can be expressed as: , in, Update amount for generated candidate parameters; It is the preset perturbation step size or learning rate, which is a very small positive number, usually ranging from 0.01 to 0.1, to ensure the smallness and stability of the correction; Indicators of consistency Relative to target parameters The gradient, which indicates the direction of parameter adjustment, is obtained by making small positive and negative perturbations near the current parameter value and observing... The gradient is numerically estimated by changing the parameters; the negative sign ensures that parameter updates are made in the direction of reducing the consistency index.

[0053] Update the calculated candidate parameters Compared with the current parameter value extracted from the sensor parameter set The values ​​are superimposed to generate temporary parameter values. Copy the current set of sensor parameters and replace the parameters of the target sensor with this temporary parameter value. This allows us to obtain a complete set of temporary parameters.

[0054] For example, in response to the "zero bias quick calibration" command, the system adjusts the zero bias parameters of the right knee joint encoder. The current value is Perform local perturbation. Set the learning rate. The system calculates consistency indicators online. Relative to parameters Gradient: Numerical experiments show that increasing the parameter makes As the parameter increases, it decreases, making Descent, Estimating Gradient Values According to the formula Calculate candidate update quantity The system generates temporary parameter values. Store the parameters in a temporary set. This method employs a gradient-based local parameter perturbation approach, which can quickly calculate the correction direction and minute step size of the parameters while the system is running online without interrupting service. This allows for timely compensation for minor errors caused by environmental changes or sensor drift, maintaining the long-term operational accuracy of the system.

[0055] Optionally, the method further includes: The constraint verification calculation is re-executed on the real-time multi-source sensor data stream using the temporary parameter set to obtain a set of verified consistency indicators. The deviation magnitude of the verified consistency index set is compared with that of the consistency index set to generate a consistency improvement judgment result. If the consistency improvement determination result is true, the candidate parameter update amount is added to the current sensor parameters to generate an updated sensor parameter set.

[0056] Specifically, a temporary parameter set is used, and the original real-time multi-source sensor data stream that caused the anomaly is retrieved again. The weighted constraint verification calculation is then performed again using the exact same algorithm, i.e., the geometric deviation is recalculated. and mechanical deviation This, combined with the same set of confidence weights, ultimately generates a new consistency index based on temporary parameters. This new sequence of indicators is called the post-validation consistency index set.

[0057] Extract the original consistency index value at the time the anomaly occurred. And the post-validation consistency index value calculated using a temporary parameter set on the same data points. The consistency improvement judgment result is generated by calculating the relative improvement rate of the consistency index. This judgment logic can be expressed as: , In this expression, the result is a Boolean value; It is the original consistency index value before the correction; It is a post-verification consistency index value calculated using a temporary parameter set; This is a preset consistency improvement threshold, typically set as a positive number between 0.1 and 0.3. It requires that the validated metric show at least a 10% to 30% improvement to be considered a valid improvement. This threshold setting prevents negligible metric changes caused by computational noise and other factors from being incorrectly accepted as valid corrections. If this inequality holds, the consistency improvement determination result is true.

[0058] If the consistency improvement determination result is true, the local parameter perturbation correction is considered effective. At this point, the previously calculated candidate parameter update is formally added to the corresponding current sensor parameter in the sensor parameter set, generating an updated sensor parameter set, which will immediately take effect in subsequent operations. Conversely, if the consistency improvement determination result is false, indicating that this tentative correction has not brought significant improvement or may even worsen the situation, the candidate parameter update and the entire temporary parameter set are discarded, the original sensor parameters remain unchanged, and a failed correction attempt is recorded in the internal log.

[0059] For example, the system utilizes including A temporary parameter set for right knee zero bias is used to re-perform forward and inverse kinematics calculations on historical outlier data segments. The corrected normalized geometric bias is reduced to Normalized mechanical deviation decreased to Calculate the consistency index after verification. Compare with the original indicators. The relative improvement rate was calculated. Since the improvement rate exceeded the preset 20% threshold, the system determined the correction was effective and officially added the candidate update values ​​to the sensor parameters, updating the right knee zero bias to... This method introduces a virtual verification mechanism based on historical data playback to quantitatively evaluate the effect of parameter correction before the parameters officially take effect. This prevents parameter deterioration caused by computational noise or incorrect estimation, and ensures the safety and robustness of the online calibration process.

[0060] Optionally, the global parameter optimization solution includes: In response to the calibration execution strategy including an identifier of a systematically associated anomaly pattern, the calibration sensor data stream is collected during the user's execution of a preset calibration action sequence; Using the dynamic library of physical constraint relationships as the convergence target, the calibration sensor data stream is substituted into the optimizer for global solution to generate the reconstructed sensor parameter set.

[0061] Specifically, once a calibration execution strategy containing a systematically correlated abnormal pattern is received, the system immediately guides the user to execute a preset sequence of calibration actions through a human-computer interaction interface, such as voice prompts or display screen information. This sequence typically includes a series of movements that fully stimulate all degrees of freedom of the lower limbs, such as slow squats, single-leg standing, and small-amplitude stepping, with a duration generally between 30 and 60 seconds. During the user's execution of these actions, data from all relevant sensors is continuously collected at a high sampling rate, forming a dedicated, information-rich calibration sensor data stream, which will be used as the sole input for global optimization.

[0062] This method uses physical constraints as the optimization objective and performs a global solution using the calibration data stream. Unlike local corrections that focus on a single bias, it uses all relevant constraints defined in the dynamic library of physical constraint relationships—namely, the first geometric constraint interval and the second mechanical constraint interval—as the convergence objective of the optimizer. The entire calibration sensor data stream is taken as input and fed into a nonlinear optimizer, such as the Levenberg-Marquardt algorithm or a sequential quadratic programming algorithm. The optimizer's task is to search the entire sensor parameter space for a set of optimal parameters that minimizes the weighted sum of the geometric and mechanical biases calculated from these parameters throughout the entire calibration sequence. Finally, the parameter combination output by the optimizer after convergence is the reconstructed set of sensor parameters.

[0063] For example, when the system detects a systematic correlation anomaly involving multiple joints in both legs, it guides the user via voice to complete a 30-second preset calibration action, such as squatting in place. The system collects a calibration sensor data stream covering the entire action cycle and uses the geometric and mechanical constraints in the dynamic library of physical constraint relationships as the global optimization objective. It then calls the Levenberg-Marquardt nonlinear optimizer, substituting the calibration data stream, to solve for the combination of link parameters and installation error parameters that minimizes the sum of all constraint violations, generating a reconstructed set of sensor parameters. Through global optimization, the system can utilize rich data information under specific excitation actions to solve complex, highly coupled, multi-parameter systematic deviation problems, achieving high-precision reconstruction of global parameters such as link length and sensor installation posture, restoring the physical consistency of the system.

[0064] Optionally, substituting the calibration sensor data stream into the optimizer for global solution includes: Using the current sensor parameters and the calibration sensor data stream, state prediction calculations are performed to obtain predicted output data; The total deviation between the predicted output data and all associated constraints in the physical constraint dynamic library is calculated to obtain the overall consistency deviation value. Perform a minimization iterative operation on the overall consistency deviation value until the deviation converges, and determine the reconstructed sensor parameter set.

[0065] Specifically, in each iteration of the optimization process, the sensor parameter set of the current iteration is used. This set contains the current values ​​of all parameters to be optimized, such as the zero bias and gain of each joint sensor, and the mounting attitude matrix of the inertial measurement unit. The entire calibration sensor data stream is used as input, and this set of parameters is used to perform forward kinematics calculation, inertial navigation calculation, and inverse dynamics calculation on each frame of data in the data stream. This results in a predicted output data sequence of the same length as the calibration data stream, containing information such as the foot position and ground reaction force frame by frame.

[0066] The overall deviation between the predicted output data and the physical constraints is calculated to quantify the quality of the current parameter set. This involves iterating through each time point of the predicted output data sequence and calculating the geometric deviation of the prediction at each time point. and mechanical deviation To eliminate the influence of different physical dimensions, these deviations are normalized. Finally, the overall consistency deviation is obtained by weighted summation of the normalized deviations over the entire time series, and its calculation formula can be expressed as: , In this formula, This represents the overall consistency deviation value; The symbol represents all time steps in the calibration sensor data stream. Perform summation; and It is a preset weighting coefficient used to balance the contribution of geometric constraints and mechanical constraints to the total deviation, and its value range is usually between 0.1 and 1.0; At a certain point in time The normalized geometric bias, by using real-time geometric bias Divide by the tolerance of the first geometric constraint interval to obtain; At a certain point in time The normalized mechanical deviation is obtained by adjusting the real-time mechanical deviation. The result is obtained by dividing by the tolerance of the second mechanical constraint interval. A squared term is used to penalize larger deviations and ensure the differentiability of the objective function.

[0067] The optimal set of parameters is determined by minimizing the overall consistency deviation through iterative computation until convergence. The objective function of the optimization algorithm uses all sensor parameters to be calibrated as independent variables. In each iteration, the optimizer calculates the direction and step size of parameter updates based on the gradient of the objective function with respect to each parameter, generating new iteration parameters. Then, the state prediction and deviation calculation steps described above are repeated to obtain a new overall consistency deviation value. This iterative process continues until... The value no longer decreases significantly, meaning the difference between two iterations is less than the preset convergence threshold, for example... Or, the maximum number of iterations has been reached, typically between 100 and 500. Upon convergence, the last set of sensor parameters output by the optimizer is determined as the reconstructed set of sensor parameters. For example... Figure 3 As shown, the point-line trajectory depicts the iterative process of the optimizer: starting from the initial parameter set with high deviation values, the parameters are continuously corrected along the gradient direction where the deviation decreases the fastest, and finally converge to the valley position where the overall consistency deviation is the smallest. The corresponding coordinate point at this time is the reconstructed sensor parameter set.

[0068] For example, in a certain iteration of the optimization process, the system extracts the sensor parameter set of the current iteration version, including specific parameters such as the zero bias and gain of each joint sensor, and the mounting attitude matrix of the inertial measurement unit. The system uses this set of parameters to perform forward kinematics calculation, inertial navigation calculation, and inverse kinematics calculation on each frame of data in the entire calibration sensor data stream, thereby generating a predicted output data sequence of the same length as the original stream. For the first frame of this sequence... The normalized geometric bias was calculated at each time point. and normalized mechanical deviation Set the preset weighting coefficients. and According to the formula The deviation contribution value of this frame is calculated. The system iterates through the entire time series, performs a weighted sum of the deviations at all time steps, and obtains the overall consistency deviation value. Then, the optimizer will As the objective function, all parameters to be calibrated are treated as independent variables, and the calculation is performed. Based on the gradients of each parameter, the direction and step size of the parameter updates are determined, generating new iteration parameters. This iterative process continues until the convergence condition is met: that is, the interval between two iterations... The decrease in value is less than the preset convergence threshold. The iteration continues until the maximum number of iterations is reached. Finally, the optimizer outputs the last set of parameters at the point where iteration stops, which is then used as the reconstructed set of sensor parameters. This method transforms the parameter calibration problem into a mathematical least-squares optimization problem by constructing an overall objective function that includes all time steps and multi-dimensional constraints. This allows for full utilization of redundant information in massive datasets to suppress random noise, thereby obtaining statistically optimal system parameter estimates.

[0069] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides a dual-mode self-calibration system for a lower limb assistive exoskeleton sensor, comprising: The data and constraint acquisition module is used to acquire real-time multi-source sensor data streams during the operation of the lower limb assistive exoskeleton and a dynamic library of physical constraint relationships used to describe the constraint relationships of lower limb movement. The weighted constraint verification module is used to perform weighted constraint verification calculations on the real-time multi-source sensor data stream using the physical constraint relationship dynamic library and the user motion state recognition results, so as to obtain a set of consistency indicators. The violation pattern analysis module is used to perform violation combination retrieval and correlation dependency analysis on the consistency index set to generate consistency violation patterns that characterize systematic anomalies or isolated anomalies. The calibration strategy generation module is used to perform spatiotemporal dimension anomaly source attribution analysis on the consistency violation pattern to obtain a calibration execution strategy containing the calibration target index and the correction algorithm identifier. The parameter correction execution module is used to respond to the calibration execution strategy by performing local parameter perturbation correction or global parameter optimization on the real-time multi-source sensor data stream and updating the sensor parameter set.

[0070] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0071] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A dual-mode self-calibration method for a lower limb assistive exoskeleton sensor, characterized in that, The method includes: The process involves acquiring real-time multi-source sensor data streams during the operation of a lower limb-assisted exoskeleton and a dynamic library of physical constraint relationships describing lower limb motion constraints. This acquisition includes: simultaneously collecting data from joint angle sensors, inertial measurement units (IMUs), and plantar pressure sensors to obtain a real-time multi-source sensor data stream; performing a statistical analysis of the deviation between the foot position calculated from the joint angle sensor data and the foot position estimated from the IMU data to generate a first geometric constraint interval; performing a mechanical balance calculation between the total ground reaction force measured from the plantar pressure sensor data and the real-time body weight support component obtained through human kinematics to generate a second mechanical constraint interval; and integrating the first geometric constraint interval and the second mechanical constraint interval to obtain the dynamic library of physical constraint relationships. The real-time multi-source sensor data stream is weighted and constrained to perform a consistency index set by using the physical constraint relationship dynamic library and the user motion state recognition results. The process of obtaining the consistency index set includes: acquiring the current user motion state recognition results, assigning weights to the first geometric constraint interval and the second mechanical constraint interval to generate a confidence weight set; and using the confidence weight set to perform weighted statistics on the real-time multi-source sensor data stream to generate the consistency index set. The consistency index set is subjected to violation combination retrieval and correlation dependency analysis to generate consistency violation patterns representing systemic or isolated anomalies. The generation of consistency violation patterns representing systemic or isolated anomalies includes: retrieving index items from the consistency index set that simultaneously exceed the first geometric constraint interval and the second mechanical constraint interval, generating violation constraint combinations; performing correlation dependency analysis on the violation constraint combinations to generate sensor anomaly source pointing information; determining the number of sources in the sensor anomaly source pointing information to generate isolated sensor anomaly patterns representing single-source anomalies or systemic associated anomaly patterns representing multi-source anomalies, as consistency violation patterns. The consistency violation pattern is analyzed for spatiotemporal anomaly source attribution to obtain a calibration execution strategy including a calibration target index and a correction algorithm identifier. This calibration execution strategy includes: extracting the duration features and sensor topology distribution features of the consistency violation pattern; marking transient or persistent anomalies with the duration features; generating descriptors for isolated single joints, ipsilateral limb chains, or bilateral symmetrical joints with the sensor topology distribution features; inputting the duration features and sensor topology distribution features into a feature vector into a pre-trained classifier for clustering to obtain anomaly influence domains; matching the anomaly influence domains to a preset sensor fault mapping table; indexing the corresponding sensor and identifying zero-bias fast calibration if it is an isolated sensor noise spike; indexing the corresponding limb chain sensor and identifying global parameter optimization if it is a limb segment link parameter drift; generating a calibration target index and a corresponding parameter correction algorithm identifier; and integrating the calibration target index and the parameter correction algorithm identifier to generate the calibration execution strategy. In response to the calibration execution strategy, local parameter perturbation correction or global parameter optimization is performed on the real-time multi-source sensor data stream to update the sensor parameter set.

2. The dual-mode self-calibration method for a lower limb assistive exoskeleton sensor according to claim 1, characterized in that, The execution of local parameter perturbation correction includes: In response to the calibration execution strategy including an identifier of an isolated sensor anomaly mode, the corresponding target sensor data is extracted according to the calibration target index; Perform micro-perturbation calculations on the target sensor data and current sensor parameters to generate candidate parameter update values; The current sensor parameters are tentatively corrected using the candidate parameter update amount to generate a temporary parameter set.

3. The dual-mode self-calibration method for a lower limb assistive exoskeleton sensor according to claim 2, characterized in that, The method further includes: The constraint verification calculation is re-executed on the real-time multi-source sensor data stream using the temporary parameter set to obtain a set of verified consistency indicators. The deviation magnitude of the verified consistency index set is compared with that of the consistency index set to generate a consistency improvement judgment result. If the consistency improvement determination result is true, the candidate parameter update amount is added to the current sensor parameters to generate an updated sensor parameter set.

4. The dual-mode self-calibration method for a lower limb assistive exoskeleton sensor according to claim 1, characterized in that, The global parameter optimization solution includes: In response to the calibration execution strategy including an identifier of a systematically associated anomaly pattern, the calibration sensor data stream is collected during the user's execution of a preset calibration action sequence; Using the dynamic library of physical constraint relationships as the convergence target, the calibration sensor data stream is substituted into the optimizer for global solution to generate the reconstructed sensor parameter set.

5. The dual-mode self-calibration method for a lower limb assistive exoskeleton sensor according to claim 4, characterized in that, The step of substituting the calibration sensor data stream into the optimizer for global solution includes: State prediction calculations are performed using the current sensor parameters and the calibration sensor data stream to obtain predicted output data; Calculate the sum of the deviations between the predicted output data and all associated constraints in the physical constraint dynamic library to obtain the overall consistency deviation value; Perform a minimization iterative operation on the overall consistency deviation value until the deviation converges, and determine the reconstructed sensor parameter set.

6. A dual-mode self-calibration system for a lower limb assistive exoskeleton sensor, characterized in that, The system includes: The data and constraint acquisition module is used to acquire real-time multi-source sensor data streams during the operation of the lower limb-assisted exoskeleton and a dynamic library of physical constraint relationships describing the lower limb motion constraint relationships. Acquiring these data streams includes: simultaneously acquiring joint angle sensor data, inertial measurement unit data, and plantar pressure sensor data to obtain a real-time multi-source sensor data stream; performing a deviation statistical analysis on the foot position calculated from the joint angle sensor data and the foot position estimated from the inertial measurement unit data to generate a first geometric constraint interval; performing a mechanical balance calculation on the total ground reaction force measured from the plantar pressure sensor data and the real-time body weight support component obtained through human kinematics calculations to generate a second mechanical constraint interval; and integrating the first geometric constraint interval and the second mechanical constraint interval to obtain the dynamic library of physical constraint relationships. The weighted constraint verification module is used to perform weighted constraint verification calculations on the real-time multi-source sensor data stream using the physical constraint relationship dynamic library and the user motion state recognition results to obtain a consistency index set. The process of obtaining the consistency index set includes: acquiring the current user motion state recognition results; assigning weights to the first geometric constraint interval and the second mechanical constraint interval to generate a confidence weight set; and using the confidence weight set to perform weighted statistics on the real-time multi-source sensor data stream to generate the consistency index set. The violation pattern analysis module is used to perform violation combination retrieval and correlation dependency analysis on the consistency index set to generate consistency violation patterns that characterize systemic anomalies or isolated anomalies. The generation of consistency violation patterns that characterize systemic anomalies or isolated anomalies includes: retrieving index items from the consistency index set that simultaneously exceed the first geometric constraint interval and the second mechanical constraint interval, generating violation constraint combinations; performing correlation dependency analysis on the violation constraint combinations to generate sensor anomaly source pointing information; determining the number of sources in the sensor anomaly source pointing information, and generating isolated sensor anomaly patterns characterizing single-source anomalies or systemic associated anomaly patterns characterizing multi-source anomalies, as consistency violation patterns. A calibration strategy generation module is used to perform spatiotemporal anomaly source attribution analysis on the consistency violation patterns to obtain a calibration execution strategy containing a calibration target index and a correction algorithm identifier. The process of obtaining the calibration execution strategy includes: extracting the duration features and sensor topology distribution features of the consistency violation patterns; marking transient or persistent anomalies with the duration features; generating descriptors for isolated single joints, ipsilateral limb chains, or bilateral symmetrical joints with the sensor topology distribution features; inputting the duration features and sensor topology distribution features into a feature vector to a pre-trained classifier for clustering to divide the anomaly influence domain; matching the anomaly influence domain to a preset sensor fault mapping table; indexing the corresponding sensor and identifying zero-bias fast calibration if it is an isolated sensor noise spike; indexing the corresponding limb chain sensor and identifying global parameter optimization if it is a limb segment link parameter drift; generating a calibration target index and a corresponding parameter correction algorithm identifier; and integrating the calibration target index and the parameter correction algorithm identifier to generate the calibration execution strategy. The parameter correction execution module is used to respond to the calibration execution strategy by performing local parameter perturbation correction or global parameter optimization on the real-time multi-source sensor data stream and updating the sensor parameter set.

Images