Self-adaptive joint control method and system for humanoid robot

Abstract

The invention relates to the technical field of robot control, in particular to a self-adaptive joint control method and system for a humanoid robot. Comprising the following steps: acquiring relative motion information, motion state information and internal mechanical state information to form comprehensive sensing data; performing contact prediction based on the comprehensive perception data, completing contact risk judgment and determining a potential contact area; and in response to the contact prediction result, a softening control instruction is sent to a joint driver in the potential contact area through a rapid instruction channel independent of a conventional trajectory position control process, so that the joint is adjusted to a low-rigidity state from a high-rigidity state within a preset response time. The problems of safety risks and equipment damage possibly caused by accidental contact between the humanoid robot and the outside when the humanoid robot executes a high-precision assembly task in a high-rigidity control state are solved, and the technical problem that an existing safety protection mechanism lags in response is solved.

Description

Technical Field

[0001] This application relates to the field of robot control technology, and more specifically, to an adaptive joint control method and system for humanoid robots. Background Technology

[0002] In the field of intelligent manufacturing equipment, humanoid robots are playing an increasingly important role. They need to be able to flexibly cope with various complex production environments, especially when performing sub-millimeter-level high-precision assembly tasks, where extremely high accuracy in positioning is required. To achieve this high precision, the drive mechanisms of robot joints are usually set to a high-rigidity control state to maximize the positional stability of the end effector and resist external interference.

[0003] However, under such high-rigidity control, existing technologies face serious challenges when humanoid robots come into unexpected physical contact with their surroundings or personnel. For example, during precision assembly, if a technician accidentally touches the robot arm, which is in a high-rigidity state, the robot's controller, which prioritizes eliminating positional errors, will immediately issue a strong corrective command, attempting to pull the arm back to its predetermined path. Due to the high rigidity of the joints, this forceful corrective action cannot smoothly absorb external impacts; instead, it may cause the arm to rebound violently, posing a serious safety risk to the technician and even damaging the precision workpiece held by the robot's end effector.

[0004] While existing robotic systems typically incorporate safety mechanisms based on force sensing or motor current monitoring to detect collisions and initiate emergency stops or enter compliant modes, these mechanisms inherently suffer from time lag. From the moment a sensor detects an abnormal force to information processing, logical judgment, and finally issuing a control mode switching command, the entire process takes tens or even hundreds of milliseconds. In contrast, high-stiffness position controllers react at the microsecond or millisecond level. This time difference means that before the safety mechanisms can intervene, a harsh and dangerous rebound action, driven by the high stiffness characteristics, has already occurred.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] This application discloses an adaptive joint control method and system for humanoid robots, aiming to solve the safety risks and equipment damage that may occur when humanoid robots perform high-precision assembly tasks under high-rigidity control conditions and come into accidental contact with the outside, as well as the technical problem of the lag in response of existing safety protection mechanisms.

[0007] The technical solution of this application is as follows: In a first aspect, this application discloses an adaptive joint control method for humanoid robots, applied to scenarios where the humanoid robot performs sub-millimeter-level high-precision assembly tasks and at least some joints are in a high-stiffness control state. The method includes: The system acquires relative motion information between external objects and the humanoid robot, motion state information of the humanoid robot itself, and internal mechanical state information of the joint drive device to form comprehensive perception data. Among them, the relative motion information includes at least the distance information and relative velocity information between the external object and the humanoid robot, the motion state information includes at least the joint velocity information and joint acceleration information, and the internal mechanical state information includes at least the joint motor current information and joint torque information. Contact prediction is performed based on comprehensive perception data to obtain contact prediction results. The contact prediction includes: assessing the possibility of contact between external objects and the humanoid robot, and determining potential contact areas. The contact risk assessment is determined based on one or more of the following conditions: the distance information is less than a preset safe distance threshold, and the relative speed information is greater than a preset approach speed threshold; within a preset time window, the joint motor current information and / or joint torque information exhibit abnormally rapid fluctuations exceeding a preset change threshold. The potential contact area is determined based on the joint that meets the conditions for the existence of contact risk and / or the joint closest to the external object. In response to the contact prediction results, a compliance control command is sent to the joint actuator in the potential contact area through a fast command channel independent of the conventional trajectory position control process. This causes the joint in the potential contact area to adjust from a high stiffness state to a low stiffness state within a preset response time. The compliance control command includes: switching the joint control mode from position control to torque control or impedance control, and / or limiting the maximum output torque of the joint to a preset safe torque limit.

[0008] Secondly, this application also discloses an adaptive joint control system for humanoid robots, applied to scenarios where the humanoid robot performs sub-millimeter-level high-precision assembly tasks and at least some joints are in a high-stiffness control state. The system includes: The perception data acquisition module is used to acquire relative motion information between external objects and humanoid robots, motion state information of humanoid robots themselves, and internal mechanical state information of joint drive devices to form comprehensive perception data. Among them, the relative motion information includes at least the distance information and relative velocity information between external objects and humanoid robots, the motion state information includes at least the joint velocity information and joint acceleration information, and the internal mechanical state information includes at least the joint motor current information and joint torque information. The contact prediction module is used to perform contact prediction based on comprehensive perception data and obtain contact prediction results. Contact prediction includes: assessing the possibility of contact between an external object and the humanoid robot, and determining potential contact areas. The contact risk assessment is based on one or more of the following conditions: distance information is less than a preset safe distance threshold, and relative speed information is greater than a preset approach speed threshold; within a preset time window, joint motor current information and / or joint torque information exhibit abnormally rapid fluctuations exceeding a preset change threshold. The potential contact area is determined based on the joint that meets the conditions for contact risk and / or the joint closest to the external object. The joint compliance control module is used to respond to the contact prediction results and send compliance control commands to the joint actuators in the potential contact area through a fast command channel independent of the conventional trajectory position control process. This causes the joints in the potential contact area to adjust from a high stiffness state to a low stiffness state within a preset response time. The compliance control commands include: switching the joint control mode from position control to torque control or impedance control, and / or limiting the maximum output torque of the joint to a preset safe torque limit.

[0009] Beneficial Effects: The adaptive joint control method for humanoid robots disclosed in this application solves the safety risks and equipment damage that may occur when humanoid robots perform high-precision assembly tasks under high-stiffness control conditions and come into accidental contact with external objects, as well as the technical problem of the lag in the response of existing safety protection mechanisms. This method can adjust the relevant joints from a high-stiffness state to a low-stiffness state at an extremely fast speed (within a preset response time) before or at the initial stage of a collision, thereby smoothly mitigating external impacts and avoiding the violent rebound that may be caused by the harsh corrective actions of high-stiffness controllers. This significantly improves the safety of humanoid robots in complex, high-precision working environments, protects operators and precision workpieces, and effectively reduces the risk of damage to the robot itself. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an adaptive joint control method for humanoid robots provided in this application.

[0011] Figure 2 A flowchart of an adaptive joint control system for humanoid robots provided in this application.

[0012] In the diagram: 1. Sensor data acquisition module; 2. Contact prediction module; 3. Joint compliance control module. Detailed Implementation

[0013] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0014] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0015] Reference Figure 1 This application proposes an adaptive joint control method for humanoid robots, applicable to scenarios where the humanoid robot performs sub-millimeter-level high-precision assembly tasks and at least some joints are in a high-stiffness control state. The method includes: S1000: Acquires relative motion information between external objects and humanoid robots, motion state information of the humanoid robot itself, and internal mechanical state information of joint drive devices to form comprehensive perception data; Among them, the relative motion information includes at least the distance information and relative velocity information between the external object and the humanoid robot, the motion state information includes at least the joint velocity information and joint acceleration information, and the internal mechanical state information includes at least the joint motor current information and joint torque information. S2000: Contact prediction is performed based on comprehensive sensing data to obtain contact prediction results; The contact prediction includes: assessing the possibility of contact between an external object and the humanoid robot, and determining the potential contact area. The contact risk assessment is based on one or more of the following conditions: the distance information is less than a preset safe distance threshold, and the relative speed information is greater than a preset approach speed threshold; within a preset time window, the joint motor current information and / or joint torque information exhibit abnormally rapid fluctuations exceeding a preset change threshold. The potential contact area is determined based on the joint that meets the conditions for the existence of contact risk and / or the joint closest to the external object. S3000: In response to the contact prediction result, it sends a compliance control command to the joint actuator in the potential contact area through a fast command channel independent of the conventional trajectory position control process, so that the joint in the potential contact area is adjusted from a high stiffness state to a low stiffness state within a preset response time. The compliance control commands include: switching the joint control mode from position control to torque control or impedance control, and / or limiting the maximum output torque of the joint to a preset safe torque limit.

[0016] Specifically, the "humanoid robot" referred to in this application refers to a robot system with multiple joints and degrees of freedom, capable of simulating human limb movements and performing complex tasks. Its joint drive mechanism typically employs servo motors in conjunction with reducers to achieve precise position and torque control. "Sub-millimeter-level high-precision assembly tasks" refer to tasks requiring extremely high assembly accuracy, with an error range in the sub-millimeter level (e.g., 0.1 mm to 0.9 mm), such as the insertion of microelectronic components or the alignment of precision mechanical parts. In such tasks, "high-stiffness control state" means that the joint controller is configured to prioritize maintaining a preset position or trajectory, exhibiting strong resistance to external disturbances to ensure the positional stability of the end effector. "Compliant control commands" refer to commands used to adjust joint control modes or limit output torque, aiming to reduce joint stiffness, allowing it to respond more "softly" to external contact, thereby reducing impact forces.

[0017] The adaptive joint control method proposed in this application is based on the fact that it effectively solves the contradiction between high stiffness control and collision safety in high-precision assembly tasks of humanoid robots through multi-source perception, contact prediction and rapid compliance control.

[0018] Specifically, various methods can be employed to acquire comprehensive sensing data. For example, the relative motion information between external objects and the humanoid robot can be obtained through visual sensors (such as depth cameras and LiDAR) installed on the robot itself or in the working environment. These sensors can measure the distance and relative velocity between objects in real time. Another method is to use an ultrasonic sensor array to calculate the distance by measuring the round-trip time of the sound waves and analyze the relative velocity through the Doppler effect. The humanoid robot's own motion state information, such as joint velocity and joint acceleration information, can usually be obtained directly from joint encoders and inertial measurement units (IMUs). Joint encoders provide precise joint positions, and velocity and acceleration can be obtained by differentiating the position data. IMUs can provide more comprehensive attitude and motion information. The internal mechanical state information of the joint drive mechanism, such as joint motor current and joint torque information, can be directly measured through current and torque sensors inside the motor driver. There is a certain proportional relationship between motor current and output torque, so current information can indirectly reflect torque state. Multi-source data is aggregated to form comprehensive sensing data, providing input for subsequent contact prediction.

[0019] In terms of contact prediction based on comprehensive sensing data, the following methods can be adopted. Contact risk assessment can be based on preset safe distance thresholds and approach speed thresholds: for example, when the vision sensor detects that the distance between an external object and the robot is less than a preset safe distance threshold (e.g., 50 mm) and the relative speed is greater than a preset approach speed threshold (e.g., 0.1 m / s), the system determines that there is a contact risk. In addition, it can also be assessed by monitoring whether joint motor current information and / or joint torque information exhibit abnormally rapid fluctuations exceeding a preset change threshold (e.g., 20% of the normal fluctuation range) within a preset time window (e.g., 10 ms); abnormally rapid fluctuations can characterize early signals of external impact or contact. Potential contact areas can be determined based on joints that meet the contact risk assessment conditions: for example, when the motor current of a certain joint exhibits abnormal fluctuations, the joint is identified as a potential contact area; or, when the vision sensor identifies the joint closest to an external object, the joint is identified as a potential contact area.

[0020] In responding to contact prediction results and sending compliance control commands, the following approach can be adopted. Once the contact prediction result indicates a contact risk, the system sends compliance control commands to the joint actuators within the potential contact area via a fast command channel independent of the conventional trajectory position control process. This fast command channel can be a separate hardware interrupt or a higher-priority software thread to ensure that the commands reach the joint actuators with low latency and are executed within a preset response time (e.g., 10 milliseconds). Compliance control commands may include switching the joint control mode from position control to torque control or impedance control: in position control mode, the joint maintains its preset position or trajectory, while in torque or impedance control mode, the joint adjusts its response according to external forces or preset impedance characteristics, thus exhibiting lower stiffness. Furthermore, compliance control commands may also include limiting the maximum output torque of the joint to a preset safe torque limit (e.g., limiting the maximum output torque to 30% of the normal operating torque) to further reduce the impact force during contact.

[0021] The adaptive joint control method proposed in this application works by performing multi-dimensional, real-time perception of the humanoid robot and its working environment to form comprehensive perception data. Based on this comprehensive perception data, contact prediction is performed to assess contact risk and determine potential contact areas. When contact risk exists, a compliant control command is sent to the joint actuators within the potential contact area via a fast command channel independent of the conventional trajectory position control process. This causes the joints within the potential contact area to adjust from a high-stiffness control state to a low-stiffness state within a preset response time. It should be noted that contact prediction can occur either before actual contact or in the early stages of contact (e.g., triggered by abnormally rapid fluctuations in motor current and / or joint torque information) to ensure that compliant adjustment is completed before the impact energy is fully transmitted and causes a violent rebound. Through joint control mode switching and / or output torque limiting, the joints can absorb impact energy and suppress violent rebound when actual contact occurs, thereby improving the safety and robustness of the humanoid robot in high-precision assembly tasks.

[0022] In another embodiment of this application, a compliance control command is further proposed to be sent to the joint actuator within the potential contact area, so that the joint within the potential contact area adjusts from a high stiffness state to a low stiffness state within a preset response time, and the method further includes: S4000: Performs recovery control of the joint within the potential contact area from a low-stiffness state to a high-stiffness state, wherein the recovery control includes: S4100: Apply a preset weak stimulus to the joint and collect dynamic response data of the joint; S4200: Analyzes dynamic response data to quantify the joint's internal resistance to weak stimuli and its rapid recovery ability, obtaining quantitative results; S4300: Based on the quantification results, adjust the recovery speed and recovery path of the joint from a low stiffness state to a high stiffness state; S4400: During the joint recovery process, the actual motion state of the humanoid robot's end effector is continuously monitored to verify the effectiveness of the recovery speed and recovery path in ensuring the stability of the end effector. When the actual motion state of the end effector does not meet the preset stability conditions, the recovery speed and / or recovery path are adjusted. The end effector is used to perform sub-millimeter-level high-precision assembly tasks.

[0023] Specifically, after the robot joints have undergone compliance processing and potential contact risks have been eliminated, the joints need to be gradually restored from a low-stiffness state to the high-stiffness state required for normal operation to ensure that the robot can continue to perform high-precision tasks. The process of "applying a preset weak stimulus to the joint and collecting dynamic response data" aims to detect the current mechanical properties and health status of the joint in a non-invasive manner. A weak stimulus can be understood as a small-amplitude, controlled disturbance signal, such as a tiny torque pulse, frequency scan signal, or vibration signal, with an amplitude far smaller than the threshold that might cause joint instability or affect assembly accuracy. By applying such a stimulus, a small dynamic response can be induced in the joint. This response data (such as changes in joint position, velocity, acceleration, torque, or current) is precisely collected by sensors for subsequent analysis. The purpose is to obtain real-time "pulse" information of the joint without interfering with the normal operation of the robot.

[0024] Furthermore, the collected dynamic response data is processed and interpreted to evaluate the mechanical performance of the joint in its current state. "Resistance" can be understood as the joint's ability to suppress external disturbances, such as its damping characteristics and instantaneous stiffness; while "rapid recovery capability" refers to the speed at which the joint's motion state or mechanical parameters recover to a stable state after being disturbed. These capabilities can be quantified by analyzing parameters such as amplitude, frequency, attenuation rate, and phase hysteresis of the dynamic response data, for example, by calculating the joint's equivalent stiffness, damping coefficient, or response time. The aim is to provide objective and quantitative evidence for the joint's recovery process.

[0025] Based on this, a dynamic stiffness recovery strategy for the joint is planned according to a quantitative assessment of its resistance and rapid recovery capabilities. For example, if the quantitative results show that the joint's resistance is weak or its recovery capability is slow, the stiffness recovery speed can be reduced accordingly, a gentler stiffness increase curve can be adopted, or a more conservative recovery path can be selected to avoid instability during the recovery process. Conversely, if the joint performance is good, the recovery speed can be appropriately accelerated to improve efficiency. The aim is to achieve adaptability in the joint stiffness recovery process and ensure a smooth transition.

[0026] Furthermore, throughout the joint stiffness recovery process, the end effector of the humanoid robot undergoes real-time, high-precision motion state monitoring. The end effector is the component that directly performs sub-millimeter-level high-precision assembly tasks, and its stability directly affects the success of the task. "Preset stability conditions" can include indicators such as the end effector's positional deviation, velocity fluctuations, and vibration amplitude. For example, the positional deviation must be less than a certain micrometer-level threshold, or the vibration frequency and amplitude must be within acceptable ranges. If the actual motion state of the end effector does not meet these conditions, it indicates a potential problem with the current recovery strategy. The system will immediately provide feedback and adjust the current recovery speed and / or recovery path, such as pausing recovery, slowing down the recovery speed, or reverting to a lower stiffness state, until the end effector stabilizes. The purpose is to provide a real-time safety mechanism to ensure that the stability of the high-precision assembly task is not affected under any circumstances.

[0027] The solution proposed in this application effectively solves the stability problems that may be caused by traditional recovery strategies by introducing an active detection and feedback mechanism during the process of joint recovery from a low stiffness state to a high stiffness state.

[0028] In another embodiment of this application, S4200 further includes: S4210: Perform time-frequency analysis on dynamic response data to obtain the energy spectral density within a specific frequency range; S4220: Compare the energy spectral density with a pre-established baseline energy spectral density under healthy conditions to identify micro-vibration signals with preset time-frequency characteristics; S4230: Calculate the signal-to-noise ratio of the micro-vibration signal; S4240: Performs adaptive filtering on micro-vibration signals with a signal-to-noise ratio lower than a preset signal-to-noise ratio threshold to enhance the characteristics of the micro-vibration signals and obtain the enhanced micro-vibration signals; S4250: Based on the intensity and duration of the enhanced micro-vibration signal, the resistance and rapid recovery capability of the joint's internal system to weak excitation are quantified to obtain the quantification results.

[0029] Specifically, time-frequency analysis of dynamic response data refers to using signal processing techniques such as Fourier transform, wavelet transform, or Hilbert-Huang transform to convert the acquired time-domain dynamic response data to the frequency domain or time-frequency domain, thereby revealing the energy distribution of the signal at different frequencies. Through this analysis, the energy spectral density within a specific frequency range can be obtained. The energy spectral density characterizes the distribution of vibrational energy of the internal structure of a joint at a specific frequency when subjected to weak excitation.

[0030] The comparison of the energy spectral density with a pre-established baseline energy spectral density under healthy conditions aims to identify micro-vibration signals that differ from those under healthy conditions. The baseline energy spectral density is obtained using the same excitation and acquisition methods when the joint is in a normal, undamaged state, and serves as a reference standard. Through comparison, micro-vibration signals with preset time-frequency characteristics can be identified; these signals are used to characterize potential abnormalities or damage within the joint.

[0031] In practical applications, calculating the signal-to-noise ratio (SNR) of a micro-vibration signal is used to assess the quality of the identified micro-vibration signal. The SNR is the ratio of signal power to noise power; a high SNR indicates a clear signal, while a low SNR indicates severe noise interference. The calculation result is used to determine whether subsequent adaptive filtering processing of the micro-vibration signal is necessary.

[0032] Furthermore, adaptive filtering is applied to micro-vibration signals with a signal-to-noise ratio (SNR) below a preset threshold. The purpose of this is to remove noise and enhance the characteristics of the micro-vibration signal. Adaptive filtering can dynamically adjust filter parameters based on the statistical characteristics of the signal and noise, thereby effectively suppressing noise and obtaining an enhanced micro-vibration signal without losing key signal information.

[0033] Ultimately, based on the intensity and duration of the enhanced micro-vibration signal, the joint's internal resistance to weak excitations and its rapid recovery capability are quantified. For example, signal intensity can reflect the stiffness or damping characteristics of the joint's internal structure, while duration may be related to energy dissipation or recovery speed. These quantitative indicators provide a quantitative assessment of the joint, offering a precise basis for subsequent recovery control.

[0034] The solution proposed in this application achieves accurate quantification of the internal state of the joint by performing multi-dimensional and refined analysis of the dynamic response data of the joint.

[0035] In another embodiment of this application, S4100 further includes: S4110: Records timing information and excitation parameters of weak excitation; S4120: Within a preset acquisition period, apply weak excitation to the joint multiple times under the same excitation parameters, and acquire the corresponding dynamic response data for each joint. S4130: Based on the timing information, the dynamic response data collected multiple times are synchronously aligned and synchronously superimposed and averaged to obtain the enhanced micro-vibration signal. S4140: Acquire sensor data of other unexcited joints of the humanoid robot. Other unexcited joints are those that have a mechanical coupling relationship with the joint and for which no weak excitation is applied during the weak excitation period. S4150: Based on the correlation between sensor data from other unexcited joints and the enhanced micro-vibration signal, an adaptive noise cancellation filter is constructed, and the sensor data from other unexcited joints is used as a reference noise signal to input into the adaptive noise cancellation filter to obtain the crosstalk component in the enhanced micro-vibration signal. S4160: Subtract crosstalk components from the enhanced micro-vibration signal to obtain a crosstalk-free micro-vibration signal, which serves as input for subsequent analysis of dynamic response data; S4170: During the process of collecting dynamic response data multiple times within a preset acquisition period, the environmental noise in the humanoid robot's workspace is monitored to obtain information on the frequency and intensity of the environmental noise. S4180: Based on the frequency and intensity of the ambient noise, adjust the frequency of the weak excitation in subsequent acquisition cycles, and use the adjusted weak excitation frequency in subsequent weak excitation application and dynamic response data acquisition, so that the dynamic response data corresponds to the adjusted weak excitation frequency.

[0036] Weak excitation can be understood as a low-energy, short-duration disturbance signal, the purpose of which is to stimulate the joint's inherent vibration modes without significantly affecting its normal operation. Timing information can include the excitation's start time and duration, while excitation parameters can include the amplitude, frequency, and waveform. Recording this information provides a precise time reference and excitation condition reference for subsequent data processing.

[0037] Furthermore, within a preset acquisition period, weak excitations are applied to the joint multiple times, and dynamic response data is collected separately. The purpose is to improve the signal-to-noise ratio of the signal through repeated measurements. The preset acquisition period can be set according to the actual application scenario and the required measurement accuracy. By synchronously aligning and synchronously superimposing and averaging the dynamically response data collected multiple times, random noise can be effectively suppressed, and the micro-vibration signal caused by weak excitation can be enhanced, thereby obtaining a clearer and more reliable enhanced micro-vibration signal. Synchronous alignment can be achieved through timestamps or feature point matching, while superimposing and averaging is accomplished by summing the aligned signals point by point and taking the average value.

[0038] Furthermore, sensor data from other unexcited joints of the humanoid robot are acquired. These unexcited joints are those mechanically coupled to the excited joints but without any applied weak excitation during this period. The sensor data from these joints can reflect structural vibrations or conducted noise within the robot body. Based on the correlation between the sensor data from these other unexcited joints and the enhanced micro-vibration signal, an adaptive noise cancellation filter can be constructed. This filter can identify and separate crosstalk components caused by mechanical coupling in the enhanced micro-vibration signal. The sensor data from these other unexcited joints is used as a reference noise signal input to the filter. The filter outputs an estimate of the crosstalk component in the enhanced micro-vibration signal, thereby accurately estimating and removing crosstalk to obtain a crosstalk-free micro-vibration signal that more purely reflects the true dynamic response of the excited joint.

[0039] Specifically, during the multiple acquisitions of dynamic response data, the environmental noise in the humanoid robot's workspace is continuously monitored to obtain its frequency and intensity information. For example, this can be done using an additional microphone or vibration sensor. Based on the monitored frequency and intensity of the environmental noise, the frequency of the weak excitation in subsequent acquisition cycles can be dynamically adjusted. The purpose is to ensure that the weak excitation frequency avoids the strong interference frequency band of the environmental noise, or to select a frequency band with relatively weak environmental noise for excitation, thereby further improving the signal-to-noise ratio and effectiveness of the dynamic response data. It should be noted that, to ensure the effectiveness of synchronous superposition and averaging processing, the frequency adjustment is applied to subsequent acquisition cycles, so that multiple excitations used for superposition and averaging within the same preset acquisition cycle maintain the same excitation parameters. The adjusted weak excitation frequency will be used in subsequent excitation application and data acquisition.

[0040] The solution proposed in this application effectively solves the problem that dynamic response data is susceptible to noise and crosstalk interference in traditional methods through a series of refined data acquisition and processing steps.

[0041] In some preferred embodiments, it is assumed that when a humanoid robot is performing a high-precision assembly task, its elbow joint enters a low-stiffness state due to accidental contact. To assess the elbow joint's recovery capability and perform recovery control, the system applies a series of preset weak sinusoidal excitations to the elbow joint. For example, within a 10-second acquisition period, a torque excitation lasting 50 milliseconds, with a frequency of 200 Hz and an amplitude of 0.1 Nm is applied every second. After each excitation, the elbow joint's accelerometer and torque sensor acquire its dynamic response data, and the system records the precise timestamp and excitation parameters for each excitation. After the acquisition period ends, all acquired dynamic response data are precisely synchronized and aligned according to the timestamps, and then superimposed and averaged to eliminate random noise, resulting in an enhanced micro-vibration signal.

[0042] Simultaneously, acceleration sensor data from the shoulder joint (as an unexcited joint), which is mechanically coupled to the elbow joint, is continuously collected. The system analyzes the correlation between the sensor data from the shoulder joint and the enhanced micro-vibration signal from the elbow joint, constructing an adaptive least mean square (LMS) filter. The sensor data from the shoulder joint is input into the LMS filter as reference noise to estimate and subtract the crosstalk component caused by the robot's structure from the enhanced micro-vibration signal from the elbow joint, thus obtaining a clean micro-vibration signal free of crosstalk.

[0043] Furthermore, throughout the data acquisition process, ambient noise in the robot's workspace is monitored in real time via an independent microphone. If strong ambient noise interference is detected around 200Hz, the system intelligently adjusts the weak excitation frequency of subsequent acquisition cycles, for example, to 250Hz, to avoid the peak frequency of ambient noise and ensure that the acquired dynamic response data has a higher signal-to-noise ratio at the new excitation frequency. Ultimately, this highly purified micro-vibration signal, processed through multiple steps and boasting a high signal-to-noise ratio, will be used for subsequent joint state quantification analysis, providing a reliable basis for precise recovery control.

[0044] In another embodiment of this application, S4210 further includes: S4211: Preprocess the dynamic response data to obtain preprocessed dynamic response data, wherein the preprocessing includes removing DC bias and power frequency interference; S4212: Based on the recovery phase of the joint from a low-stiffness state to a high-stiffness state and the historical damage assessment data of the joint, the dynamically adjusted characteristic frequency range is obtained. S4213: Set the analysis frequency band for time-frequency analysis based on the dynamically adjusted characteristic frequency range, perform time-frequency analysis on the preprocessed dynamic response data, and calculate the energy spectral density based on the dynamically adjusted characteristic frequency range.

[0045] Specifically, preprocessing refers to the initial signal processing of the raw acquired dynamic response data to eliminate or reduce unwanted components in the data. DC bias removal aims to eliminate constant offsets in the signal, ensuring its zero-mean characteristic, which is crucial for the accuracy of subsequent time-frequency analysis. Power frequency interference removal refers to filtering out noise of specific frequencies (e.g., 50Hz or 60Hz) introduced by the power supply system or other electrical equipment, which may mask the true micro-vibration signals of the joint.

[0046] Furthermore, the recovery phase of a joint from a low-stiffness state to a high-stiffness state refers to different time points or process segments in the gradual recovery of the joint from a low-stiffness state after compliance control to a normal high-stiffness state, such as the initial recovery stage, the middle recovery stage, and the final recovery stage. Historical damage assessment data for a joint refers to information accumulated during its past operation regarding its health status, damage type, damage severity, and corresponding vibration characteristics. This data can come from regular maintenance inspections, fault diagnosis records, or trend data obtained through long-term monitoring.

[0047] The dynamically adjusted characteristic frequency range refers to a frequency interval that adaptively reflects the joint's health status or potential damage, determined based on the joint's current recovery stage and historical damage assessment data. For example, in the early stages of recovery, it may be necessary to focus on low-frequency signals related to friction or initial wear; while in the later stages of recovery, it may be necessary to focus on mid-to-high-frequency signals related to structural loosening or fatigue. The analysis band is the frequency range of interest during time-frequency analysis, and its setting should match the dynamically adjusted characteristic frequency range to ensure the analysis's relevance and effectiveness. Energy spectral density is a function describing the distribution of signal energy at different frequencies. By calculating its value within a specific frequency range, the vibration intensity of the joint within that frequency range can be quantified. It should be noted that the dynamically adjusted characteristic frequency range is used to define the center frequency and bandwidth of the analysis band, thereby focusing the energy spectral density calculation on the frequency components within the characteristic frequency range.

[0048] The solution proposed in this application can effectively remove DC bias and power frequency interference by preprocessing the dynamic response data, thereby purifying the original signal and ensuring the accuracy and reliability of subsequent time-frequency analysis.

[0049] In another embodiment of this application, after calculating the energy spectral density based on the dynamically adjusted characteristic frequency range, the method further includes: S4214: Acquire the time-frequency characteristics of environmental noise in the workspace of the humanoid robot, and during the application of weak excitation to the joint, collect the sensor time-varying data of other unexcited joints of the humanoid robot, and extract the vibration time-frequency characteristics from the sensor time-varying data of other unexcited joints. Other unexcited joints are those that have a mechanical coupling relationship with the joint and are not subjected to weak excitation during the weak excitation period. S4215: Monitor the time-varying characteristics of the energy spectral density of a joint within a dynamically adjusted characteristic frequency range; S4216: Compare the time-varying characteristics with the time-frequency characteristics of the environmental noise to identify the first time-varying component that satisfies the preset first similarity rule with the environmental noise; S4217: Compare the time-varying characteristics with the vibration time-frequency characteristics of other unexcited joints to identify the second time-varying component that satisfies the preset second similarity rule with the vibration of other unexcited joints. S4218: Subtract the first time-varying component and the second time-varying component from the time-varying characteristics to obtain the energy spectral density time-varying characteristics after removing similar components; S4220: Compare the energy spectral density with a pre-established baseline energy spectral density under healthy conditions to identify micro-vibration signals with preset time-frequency characteristics, including: comparing the time-varying characteristics of the energy spectral density with the time-varying characteristics of the baseline energy spectral density to identify micro-vibration signals with preset time-frequency characteristics.

[0050] Specifically, obtaining the time-frequency characteristics of environmental noise can be achieved by deploying environmental noise sensors (such as microphones or vibration sensors) in the humanoid robot's workspace. These sensors can collect acoustic or vibration signals from the environment in real time and extract their frequency, amplitude, phase, and other characteristics using time-frequency analysis methods such as Fourier transform. Time-varying data from sensors of other unexcited joints refers to the simultaneous acquisition of sensor data from other joints that are mechanically coupled to the target joint but not directly excited, such as accelerometers and force sensors, when a weak excitation is applied to the target joint. This data reflects the vibration transmission and crosstalk within the mechanical structure; extracting time-frequency characteristics from this data yields its vibration modes and energy distribution.

[0051] Monitoring the time-varying characteristics of the energy spectral density of a joint within its dynamically adjusted characteristic frequency range involves continuously tracking changes in the energy spectral density over time to capture its dynamic evolution. Comparing the monitored time-varying characteristics with the time-frequency characteristics of environmental noise aims to identify interference components caused by environmental noise. A pre-defined first similarity rule, based on indicators such as correlation coefficient, spectral similarity, or energy distribution overlap, is used to determine which parts of the time-varying characteristics are highly similar to environmental noise. Similarly, comparing the time-varying characteristics with the vibration time-frequency characteristics of other unexcited joints identifies crosstalk components caused by mechanical coupling; a pre-defined second similarity rule can also employ similar metric standards.

[0052] In practical applications, subtracting the identified first and second time-varying components from the time-varying characteristics means estimating the contributions of the first and second time-varying components to the time-varying characteristics and removing this estimate from the time-varying characteristics. This processing can employ signal processing techniques, such as adaptive filtering, noise cancellation algorithms, or spectral subtraction, to effectively remove these interfering components, thereby obtaining a purer energy spectral density time-varying characteristic that reflects the true internal state of the joint. Therefore, the subsequent micro-vibration signal identification process will be based on the energy spectral density time-varying characteristic after removing similar components, and compared with the time-varying characteristic of the baseline energy spectral density established in the healthy state, thereby more accurately identifying micro-vibration signals with preset time-frequency characteristics.

[0053] The solution proposed in this application effectively solves the problem of insufficient accuracy in identifying micro-vibration signals under complex working conditions by introducing a mechanism for identifying and removing environmental noise and other unexcited joint vibration crosstalk.

[0054] In another embodiment of this application, an adaptive joint control method is further proposed, which includes: S4219: Extract multidimensional features of the time-varying characteristics of energy spectral density to obtain a multidimensional feature vector, wherein the multidimensional features include at least frequency features, amplitude features, phase features, bandwidth features, energy centroid features, and statistical features characterizing the time-varying distribution pattern. S42110: Compare the multidimensional feature vector with the pre-established damage pattern feature library, and calculate the similarity between the multidimensional feature vector and the feature vector of each damage pattern in the damage pattern feature library. The damage pattern feature library contains multiple damage types and their corresponding feature vectors under different damage degrees. S42111: Based on similarity, identify the most matching damage pattern and output preliminary damage assessment results, wherein the preliminary damage assessment results include at least the most matching damage pattern and its corresponding similarity. S42112: Based on the preliminary damage assessment results, when the similarity is lower than the preset similarity threshold, the micro-vibration signal corresponding to the multidimensional feature vector is marked as an unknown damage mode. S42113: Continuously collect unclassified damage event information from the historical operation data of humanoid robots, and extract and classify the damage event information to dynamically update the damage pattern feature library. S42114: Adjust the feature vectors in the damage mode feature library according to the recovery stage of the joint from a low stiffness state to a high stiffness state, so as to adapt to the changes in damage characteristics under different working conditions.

[0055] Specifically, multidimensional feature vectors refer to a set of numerical values ​​extracted from the time-varying characteristics of energy spectral density that comprehensively characterize the properties of joint micro-vibration signals. These features are not limited to single frequency or amplitude information, but encompass the signal's performance across different dimensions. For example, frequency features can refer to the dominant frequency, harmonic frequencies, or frequency distribution within a specific frequency band of the micro-vibration signal; amplitude features can refer to the signal's peak value, root mean square value, or energy magnitude; phase features can reveal the signal's relative position or synchronicity over time; bandwidth features reflect the frequency distribution range of the signal; energy centroid features can indicate the concentration trend of energy along the frequency axis; and statistical features characterizing the time-varying distribution, such as kurtosis, skewness, and entropy, can describe the non-Gaussianity or complexity of the signal's changes over time. By extracting these multidimensional features, the essence of micro-vibration signals can be more comprehensively and precisely characterized, providing a rich data foundation for subsequent damage pattern recognition.

[0056] The damage pattern feature library is a pre-built database that stores various known damage types (such as bearing wear, gear fracture, poor lubrication, and loosening) and their corresponding multidimensional feature vectors at different damage levels. The feature library can be constructed based on experimental data, simulation models, or historical fault data. The feature vector for each damage pattern represents the typical manifestation of damage in micro-vibration signals. By comparing the currently extracted multidimensional feature vector with the feature vectors in the library, the similarity between the current joint state and various known damage patterns can be quantified.

[0057] In practical applications, similarity is an indicator used to measure the degree of matching between the currently extracted multidimensional feature vector and the feature vectors of various damage patterns in the damage pattern feature library. Similarity can be calculated using various mathematical methods, such as Euclidean distance, cosine similarity, Mahalanobis distance, or classification confidence based on machine learning models. Higher similarity indicates that the current joint's micro-vibration signal is closer to the typical characteristics of the damage pattern. Based on the calculated similarity, the system identifies the damage pattern with the highest similarity to the current micro-vibration signal and uses it as the best-matching damage pattern. The preliminary damage assessment result includes the name of this best-matching damage pattern and its corresponding similarity value, providing a preliminary basis for subsequent decision-making.

[0058] Furthermore, when the similarity of the calculated best-matching damage pattern is lower than a preset similarity threshold, it indicates that the current micro-vibration signal's characteristics do not match any known damage pattern in the database. In this case, the signal will be marked as an unknown damage pattern. This helps the system identify new, unpredictable anomalies and trigger further analysis or human intervention. To improve the accuracy and adaptability of damage identification, this application continuously collects damage event information marked as unknown damage patterns from the humanoid robot's historical operational data. After feature extraction and expert analysis, this unclassified damage event information can be classified as new damage patterns or used as supplementary samples to existing damage patterns, thereby dynamically updating and expanding the damage pattern feature database. This allows the system to continuously learn and adapt to new damage types or damage manifestations.

[0059] Furthermore, during the recovery process from a low-stiffness state to a high-stiffness state, the internal mechanical properties and vibration response of a joint may change with the recovery stage (e.g., gradual increase in stiffness, load changes, etc.). Therefore, to ensure the accuracy of damage assessment, the feature vectors in the damage mode feature library need to be dynamically adjusted according to the current recovery stage. For example, in the early stages of recovery, the joint may still be in a low-stiffness state, and its vibration characteristics may differ from the high-stiffness state after full recovery. By adjusting the feature vectors according to the recovery stage, the identification of damage modes can more accurately reflect the actual damage under the current working conditions.

[0060] The solution proposed in this application overcomes the limitation of existing technologies, which can only identify the presence of micro-vibration signals but cannot perform fine damage diagnosis, by introducing a multi-dimensional feature extraction and damage pattern comparison mechanism.

[0061] In some preferred embodiments, this application is implemented as follows: Assuming that when a humanoid robot is performing an assembly task, its elbow joint enters a low-stiffness state after contact prediction and begins recovery control. During the recovery process, the system continuously applies weak excitation and collects dynamic response data of the elbow joint. After time-frequency analysis and noise cancellation, the time-varying characteristics of the elbow joint's energy spectral density are obtained. The system extracts multidimensional features from the time-varying characteristics, such as the main vibration frequency, its corresponding amplitude, phase lag angle, vibration energy bandwidth, energy centroid, and statistical features like the kurtosis coefficient and skewness coefficient of the vibration signal within a specific frequency range (e.g., 500Hz-1500Hz), forming a multidimensional feature vector. Subsequently, the multidimensional feature vector is input into a pre-established damage mode feature library for comparison. The feature library may contain various damage modes such as "initial bearing wear," "minor gear pitting," and "lubricating oil deterioration," and their corresponding feature vectors. For example, by calculating cosine similarity, if the current feature vector has a similarity of 0.92 with the feature vector of the "early bearing wear" pattern, while its similarity with other patterns is below 0.85, the system will identify "early bearing wear" as the best-matching damage pattern and output a preliminary damage assessment result, which includes "early bearing wear" and a similarity of 0.92. If the similarity between the extracted feature vector and all known damage patterns at a certain moment is below a preset similarity threshold (e.g., 0.7), the micro-vibration signal will be marked as an "unknown damage pattern." The system will record the feature vector and related operational data of the unknown damage pattern and continuously collect similar events. When enough unknown damage pattern data has been accumulated, engineers can analyze it and classify it as a new damage pattern (e.g., "loose connecting bolts"), and then update it to the damage pattern feature library, thereby enhancing the system's recognition capability. Furthermore, at different stages of elbow joint recovery from low to high stiffness, such as when stiffness is adjusted to 50% and 80%, the system dynamically adjusts the weights or ranges of feature vectors in the damage mode feature library. For example, in the early stages of recovery with lower stiffness, the vibration characteristics of certain damage modes may not be obvious or may exhibit different forms. In this case, the corresponding feature vectors in the feature library will be adjusted to better match the damage performance under the current stiffness, ensuring the accuracy of the assessment. In this way, the system can provide more accurate diagnostic information based on the actual type of joint injury and the recovery stage, thereby guiding the recovery control module to adjust the recovery speed and path of the elbow joint. For example, for the initial wear of bearings, a slower stiffness recovery speed can be adopted and high-load operations can be avoided to prevent the damage from aggravating.

[0062] In another embodiment of this application, it is further proposed that, based on similarity, the most matching damage pattern be identified and a preliminary damage assessment result be output, including: S42111-1: When the similarity calculation results show that the similarity of multiple damage patterns is higher than the preset similarity threshold, and the similarity difference between multiple damage patterns is less than the preset difference threshold, the feature vectors of multiple damage patterns are analyzed to obtain the difference information of multiple damage patterns in frequency features, amplitude features, duration features and statistical features representing time-varying distribution patterns. The difference information is then weighted and evaluated in combination with the current operating conditions of the joint to determine the most matching damage pattern. The current operating conditions include at least one or more of the following: operating load, joint speed, joint output torque and ambient temperature. S42111-2: When the similarity calculation result cannot meet the preset discrimination conditions to distinguish different damage modes, the verification process is triggered, a set of probing stimuli are applied to the joint, and transient response data of the joint under the probing stimuli are collected; high-resolution time-frequency analysis is performed on the transient response data to extract unique response features for distinguishing different damage modes, and the matching results of the damage modes are updated based on the unique response features. S42111-3: Based on the determined damage mode type and the current operating state of the joint, select the preset damage assessment rule, and combine it with the actual operating parameters of the joint to perform preliminary quantification of the damage degree, and output the preliminary damage assessment result. The actual operating parameters include at least one or more of the following: joint torque fluctuation, joint position deviation, and temperature change trend. The preliminary damage assessment result includes at least the best matching damage mode and its corresponding similarity, as well as the preliminary quantification result of the damage degree. The preliminary quantification result of the damage degree is used to determine the adjustment direction of the recovery speed and / or recovery path for the recovery control of the joint from a low stiffness state to a high stiffness state.

[0063] Specifically, when similarity calculations show that multiple damage patterns have similarities higher than a preset similarity threshold, and the differences in similarity between these patterns are less than a preset difference threshold, it indicates that a clear judgment cannot be made based solely on similarity. In this case, a deeper analysis of the feature vectors of these candidate damage patterns is needed to identify subtle differences in frequency characteristics, amplitude characteristics, duration characteristics, and statistical characteristics representing time-varying distribution patterns. For example, these differences can be obtained by comparing the energy distribution, peak amplitude, signal duration, and signal envelope trends of different patterns in specific frequency bands. Furthermore, this difference information is weighted and evaluated in conjunction with the current operating conditions of the joint. Current operating conditions can be understood as external or internal conditions affecting joint performance and damage manifestations. For example, operational load refers to the magnitude of the external load borne by the joint, joint speed refers to the speed of joint movement, joint output torque refers to the torque generated by the joint actuator, and ambient temperature refers to the temperature of the environment in which the joint is located. By using this operating condition information as a weighting factor, the probability of different damage modes under the current actual operating conditions can be assessed more accurately, thereby determining the most matching damage mode. For example, under heavy load conditions, the characteristics of some damage modes may be more significant, so they will be given higher weights in the evaluation.

[0064] When the similarity calculation results fail to meet the preset discrimination conditions to distinguish different damage modes—for example, when weighted evaluation still fails to clearly distinguish between them or the confidence level is insufficient—the system triggers a verification process. This verification process aims to obtain more discriminative information through active probing. Specifically, a set of probing stimuli is applied to the joint. These stimuli can be preset short-duration pulses or swept-frequency signals with specific frequencies or amplitudes. The purpose is to elicit unique responses that the joint may exhibit under different damage modes. Simultaneously, transient response data of the joint under these probing stimuli is collected. This transient response data includes the joint's rapid dynamic feedback to the stimuli within a short period. Subsequently, high-resolution time-frequency analysis is performed on the transient response data, using methods such as wavelet transform and Hilbert-Huang transform, to extract unique response features for distinguishing different damage modes. These unique response features may include transient vibration modes, resonant frequency shifts, and changes in damping characteristics. Finally, the matching results of the damage modes are updated based on these unique response features, thereby improving the accuracy of damage identification.

[0065] In practical applications, after determining the type of damage mode and the current operating state of the joint, preset damage assessment rules are selected. These rules are assessment models or algorithms pre-established for different damage types and operating conditions. For example, for bearing wear, an assessment rule based on vibration signal harmonic analysis might be selected; for gear pitting, an assessment rule based on meshing frequency modulation characteristics might be selected. Subsequently, the degree of damage is initially quantified by combining the actual operating parameters of the joint. These actual operating parameters refer to measurable physical quantities of the joint during actual operation, such as joint torque fluctuations, joint position deviations, and temperature change trends. Changes in these parameters are often directly related to the degree of damage. For example, increased joint torque fluctuations may indicate internal friction or jamming, position deviations may indicate increased clearance, and increased temperature may indicate abnormal friction. By inputting these parameters into the selected assessment rules, a preliminary quantitative result of the degree of damage can be obtained, such as a damage index or percentage. The preliminary injury assessment results include not only the best matching injury pattern and its corresponding similarity, but also the preliminary quantitative results of the injury severity. The preliminary quantitative results of the injury severity are used to determine the direction of adjustment of the recovery speed and / or recovery path for the joint to recover from a low stiffness state to a high stiffness state. For example, a faster recovery speed can be used when the injury is mild, and a slower and more cautious recovery path can be used when the injury is severe to avoid secondary injury.

[0066] The solution proposed in this application effectively solves the problems of ambiguity in damage pattern recognition and inaccuracy in damage degree quantification under complex working conditions by introducing multi-dimensional feature difference analysis, active probing stimulus verification, and adaptive damage assessment rule selection.

[0067] In some preferred embodiments, assuming that after a humanoid robot's elbow joint performs an assembly task, its micro-vibration signal is processed and its features extracted. After comparison with a damage mode feature library, it is found that the similarity between the two damage modes, "early bearing wear" and "minor gear pitting," is higher than a preset threshold (e.g., 0.8), and the difference in similarity between the two is very small (e.g., 0.02), making it difficult for the system to directly determine which type of damage it is. At this point, the solution of this application will initiate further analysis: First, the system deeply analyzes the feature vectors of the two modes, "early bearing wear" and "minor gear pitting," and identifies subtle differences in their energy distribution, harmonic characteristics, and vibration signal envelope morphology within a specific frequency range. For example, bearing wear may exhibit more obvious random impact characteristics at higher frequency bands, while gear pitting may exhibit modulation phenomena at the meshing frequency and its harmonics. At the same time, the system combines the current operating conditions of the elbow joint, such as a high current operating load and a slight increase in joint output torque fluctuation, thus favoring the characteristics of gear pitting in the weighted evaluation, and determining the most matching damage mode accordingly.

[0068] If a clear distinction cannot be made after weighted evaluation, for example, if the confidence level is still below a certain preset value, the system will trigger a verification process: apply a set of probing stimuli to the elbow joint, such as a short-duration, low-amplitude frequency sweep signal, and collect transient response data of the joint; then perform high-resolution time-frequency analysis on the transient response data, such as using short-time Fourier transform or wavelet transform, to extract more distinctive and unique response features. For example, bearing wear may exhibit more obvious nonlinear response or resonant frequency drift under specific stimuli, while gear pitting may exhibit a specific impact response mode under transient impact; based on these unique response features, the system updates the damage mode matching results, thereby accurately identifying the best-matching damage mode, such as finally determining it as "minor gear pitting".

[0069] After identifying the damage as "minor pitting of the gear," the system selects a preset damage assessment rule based on the type of damage mode and the joint's current operating state (e.g., moderate load, normal temperature). For example, it selects a damage quantification model based on gear meshing vibration characteristics. Simultaneously, it incorporates actual joint operating parameters, such as joint torque fluctuations (e.g., fluctuation amplitude of 15% of normal value), joint position deviations (e.g., 0.05 mm), and temperature change trends (e.g., a slow increase of 0.5 degrees Celsius / hour). These parameters are input into the selected assessment rule to calculate a preliminary damage quantification result, such as a damage index of 30% (indicating minor damage). This quantification result will guide the joint's recovery control. For example, based on a 30% damage level, the system adjusts the speed at which the joint recovers from a low-stiffness state to a high-stiffness state to 70% of the normal recovery speed and selects a smoother recovery path that avoids high impacts to ensure that the damage is not aggravated during recovery while maintaining stable robot operation.

[0070] In another embodiment of this application, a preset damage assessment rule is selected based on the determined type of damage pattern and the current operating state of the joint, specifically including: S42111-31: Based on the determined damage mode type, select the corresponding initial evaluation rule from the preset damage type-evaluation model mapping table; S42111-32: Based on the current operating state of the joint, the parameters of the initial evaluation rule are dynamically adjusted to obtain the adjusted evaluation rule; S42111-33: Input the actual operating parameters of the joint into the adjusted evaluation rules to obtain the damage degree index that characterizes the degree of damage; S42111-34: When the damage severity index exceeds the applicable range of the preset damage assessment rule, an adaptive correction process is triggered to collect real-time running data of the joint and signal features used to characterize the damage. The signal features include time-varying characteristics of energy spectral density and / or unique response features. Combined with historical similar damage case data, the parameters and structure of the adjusted assessment rule are corrected to obtain the damage assessment rule. S42111-35: During the injury assessment process, continuously monitor the evolution trend of injury signal characteristics and adjust the injury severity index in combination with joint operation load and environmental conditions.

[0071] Specifically, firstly, based on the determined damage pattern type, an initial assessment rule corresponding to the damage pattern type is selected from a pre-established damage type-assessment model mapping table. The mapping table pre-stores the correspondence between different damage types and their respective applicable assessment models, ensuring the initial matching of the assessment rules. Furthermore, considering the current operating state of the joint, the parameters of the selected initial assessment rule are dynamically adjusted to obtain an adjusted assessment rule. This dynamic adjustment mechanism allows the assessment rule to better adapt to current operating loads, joint speeds, ambient temperatures, and other specific working conditions, improving the real-time performance and accuracy of the assessment. Subsequently, the actual operating parameters of the joint, such as joint torque fluctuations, joint position deviations, and temperature change trends, are input into the adjusted assessment rule to calculate a damage severity index characterizing the degree of joint damage. This damage severity index is used to characterize the health status of the joint.

[0072] In a preferred implementation, when the damage severity index exceeds the applicable range of the preset damage assessment rules, the system triggers an adaptive correction process. During this process, real-time joint operation data and signal characteristics characterizing the damage are collected. These signal characteristics may include time-varying energy spectral density properties and / or unique response features. Simultaneously, historical data on similar damage cases are combined to modify the parameters and structure of the adjusted assessment rules, resulting in more accurate and robust damage assessment rules. This correction mechanism ensures that the assessment rules can continuously learn and adapt to new damage conditions, improving their generalization ability. It should be noted that the damage assessment rules obtained from the adaptive correction process are used to update the damage severity index, allowing assessment results exceeding the applicable range to be recalculated or corrected based on the modified assessment rules. Furthermore, throughout the damage assessment process, the system continuously monitors the evolution trend of the damage signal characteristics and, in conjunction with the joint's operational load and environmental conditions, corrects the damage severity index in real time. This continuous monitoring and correction mechanism further enhances the dynamism and accuracy of the damage assessment, ensuring continuous and accurate tracking of the joint's health status.

[0073] The solution proposed in this application effectively addresses the problem of insufficient accuracy and robustness of traditional static damage assessment rules under complex and variable working conditions by introducing dynamic adjustment, adaptive correction, and continuous monitoring mechanisms.

[0074] In some preferred embodiments, a specific example is given below. Suppose a humanoid robot is performing a high-precision assembly task, and a damage pattern, such as bearing wear, is identified in one of its elbow joints. First, the system selects an initial bearing wear assessment rule from a preset damage type-assessment model mapping table based on the type of damage pattern, "bearing wear." This rule can be a machine learning model based on vibration signal characteristics. Next, considering the current operating state of the elbow joint, such as the assembly load it is bearing, the current joint speed, and the ambient temperature, the system dynamically adjusts the parameters in the initial assessment rule. For example, when the current load is high, the sensitivity of the assessment model to vibration amplitude is increased to identify signs of accelerated wear earlier. Then, the actual operating parameters of the elbow joint, such as minor fluctuations in joint torque, position tracking errors, and temperature change trends in the bearing area, are input into the adjusted assessment rule to calculate a damage severity index. If the damage severity index suddenly exceeds the preset "slight wear" range, for example, reaching the "moderate wear" threshold, the system will immediately trigger an adaptive correction process. At this point, the system will collect real-time vibration signals of the elbow joint (time-varying energy spectral density characteristics) and transient response data under specific probing excitations (unique response characteristics), and combine this with case data similar to "moderate bearing wear" from the historical database. Based on this information, the system will correct the parameters of the currently used evaluation rules (such as weights and thresholds) and even the model structure (such as adding new feature dimensions or adjusting model levels), and update the damage severity index based on the corrected evaluation rules to more accurately reflect the actual damage status of the elbow joint. Throughout the evaluation process, the system will continuously monitor the evolution of the elbow joint's vibration signal characteristics, such as the energy change trend of specific frequency components. Simultaneously, it will combine changes in the operational load of the assembly task and fluctuations in ambient temperature to correct the calculated damage severity index in real time. For example, if the ambient temperature rises and the bearing lubrication performance decreases, even if the vibration signal change is not significant, the damage severity index will be appropriately increased to reflect the increased potential risk. In this way, continuous, accurate, and adaptive evaluation of the elbow joint's health status is ensured.

[0075] Reference Figure 2 This application proposes an adaptive joint control system for humanoid robots, applicable to scenarios where humanoid robots perform sub-millimeter-level high-precision assembly tasks and at least some joints are in a high-stiffness control state. The system includes: The perception data acquisition module 1 is used to acquire relative motion information between external objects and humanoid robots, motion state information of humanoid robots themselves, and internal mechanical state information of joint drive devices to form comprehensive perception data; wherein, the relative motion information includes at least the distance information and relative velocity information between external objects and humanoid robots, the motion state information includes at least the joint velocity information and joint acceleration information, and the internal mechanical state information includes at least the joint motor current information and joint torque information. The contact prediction module 2 is used to perform contact prediction based on comprehensive perception data to obtain contact prediction results. The contact prediction includes: judging the possibility of contact between external objects and humanoid robots, and determining potential contact areas. The contact risk judgment is based on one or more of the following conditions to determine the existence of contact risk: the distance information is less than a preset safe distance threshold, and the relative speed information is greater than a preset approach speed threshold; within a preset time window, the joint motor current information and / or joint torque information show abnormal rapid fluctuations exceeding a preset change threshold. The potential contact area is determined based on the joint that meets the conditions for the existence of contact risk and / or the joint closest to the external object. The joint compliance control module 3 is used to respond to the contact prediction result and send compliance control commands to the joint actuators in the potential contact area through a fast command channel independent of the conventional trajectory position control process. This causes the joints in the potential contact area to adjust from a high stiffness state to a low stiffness state within a preset response time. The compliance control commands include: switching the joint control mode from position control to torque control or impedance control, and / or limiting the maximum output torque of the joint to a preset safe torque limit.

[0076] The adaptive joint control system proposed in this application is based on the modular design that enables multi-source perception, intelligent prediction and rapid compliant control, effectively solving the contradiction between high stiffness control and collision safety in high-precision assembly tasks of humanoid robots.

[0077] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An adaptive joint control method for humanoid robots, characterized in that, The method, applicable to scenarios where humanoid robots perform sub-millimeter-level high-precision assembly tasks and at least some joints are in a high-stiffness control state, includes: The relative motion information between the external object and the humanoid robot, the motion state information of the humanoid robot itself, and the internal mechanical state information of the joint drive device are acquired to form comprehensive perception data; wherein, the relative motion information includes at least the distance information and relative velocity information between the external object and the humanoid robot, the motion state information includes at least the joint velocity information and joint acceleration information, and the internal mechanical state information includes at least the joint motor current information and joint torque information. Contact prediction is performed based on the comprehensive sensing data to obtain contact prediction results; wherein, the contact prediction includes: assessing the possibility of contact between the external object and the humanoid robot, and determining the potential contact area. The contact risk assessment is determined based on one or more of the following conditions: the distance information is less than a preset safe distance threshold, and the relative speed information is greater than a preset approach speed threshold; within a preset time window, the joint motor current information and / or joint torque information exhibit abnormally rapid fluctuations exceeding a preset change threshold, wherein the potential contact area is determined based on the joint that satisfies the condition of having the contact risk and / or the joint closest to the external object; In response to the contact prediction result, a compliance control command is sent to the joint actuator in the potential contact area through a fast command channel independent of the conventional trajectory position control process. This causes the joint in the potential contact area to adjust from a high stiffness state to a low stiffness state within a preset response time. The compliance control command includes: switching the joint control mode from position control to torque control or impedance control, and / or limiting the maximum output torque of the joint to a preset safe torque upper limit.

2. The adaptive joint control method according to claim 1, characterized in that, Sending a compliance control command to the joint actuator within the potential contact area, causing the joint within the potential contact area to adjust from a high-stiffness state to a low-stiffness state within a preset response time, further includes: Perform recovery control of the joint within the potential contact area from a low-stiffness state to a high-stiffness state, wherein the recovery control includes: A preset weak stimulus is applied to the joint, and the dynamic response data of the joint is collected; The dynamic response data is analyzed to quantify the joint's internal resistance to weak stimuli and its rapid recovery capability, resulting in quantitative analysis. Based on the quantification results, adjust the recovery speed and recovery path of the joint from the low stiffness state to the high stiffness state; During the joint recovery process, the actual motion state of the end effector of the humanoid robot is continuously monitored to verify the effectiveness of the recovery speed and recovery path in ensuring the stability of the end effector. When the actual motion state of the end effector does not meet the preset stability conditions, the recovery speed and / or recovery path are adjusted, wherein the end effector is used to perform sub-millimeter-level high-precision assembly tasks.

3. The adaptive joint control method according to claim 2, characterized in that, The dynamic response data is analyzed to quantify the joint's internal resistance to weak stimuli and its rapid recovery capability, yielding quantitative results including: Time-frequency analysis is performed on the dynamic response data to obtain the energy spectral density within a specific frequency range; The energy spectral density is compared with a baseline energy spectral density established in the healthy state to identify micro-vibration signals with preset time-frequency characteristics. Calculate the signal-to-noise ratio of the micro-vibration signal; Adaptive filtering is performed on micro-vibration signals with a signal-to-noise ratio lower than a preset signal-to-noise ratio threshold to enhance the characteristics of the micro-vibration signals, thereby obtaining enhanced micro-vibration signals. Based on the intensity and duration of the enhanced micro-vibration signal, the resistance and rapid recovery capability of the joint to weak excitation are quantified to obtain the quantification result.

4. The adaptive joint control method according to claim 3, characterized in that, Apply a preset weak stimulus to the joint and collect dynamic response data of the joint, including: Record the timing information and excitation parameters of the weak excitation; Within a preset acquisition period, the weak excitation is applied to the joint multiple times under the same excitation parameters, and the corresponding dynamic response data is acquired respectively. Based on the timing information, the dynamic response data collected multiple times are synchronously aligned and synchronously superimposed and averaged to obtain an enhanced micro-vibration signal. Acquire sensor data of other unexcited joints of the humanoid robot, wherein the other unexcited joints are those that have a mechanical coupling relationship with the joint and for which the weak excitation was not applied during the weak excitation; Based on the correlation between the sensor data of the other unexcited joints and the enhanced micro-vibration signal, an adaptive noise cancellation filter is constructed, and the sensor data of the other unexcited joints is used as a reference noise signal to input the adaptive noise cancellation filter to obtain the crosstalk component in the enhanced micro-vibration signal. The crosstalk component is subtracted from the enhanced micro-vibration signal to obtain the crosstalk-free micro-vibration signal, which is used as input for subsequent analysis of the dynamic response data; During the process of collecting the dynamic response data multiple times within a preset collection period, the environmental noise in the workspace of the humanoid robot is monitored to obtain the frequency and intensity information of the environmental noise. Based on the frequency and intensity of the environmental noise, the frequency of the weak excitation in subsequent acquisition cycles is adjusted, and the adjusted weak excitation frequency is used in subsequent weak excitation application and dynamic response data acquisition, so that the dynamic response data corresponds to the adjusted weak excitation frequency.

5. The adaptive joint control method according to claim 3, characterized in that, Time-frequency analysis is performed on the dynamic response data to obtain the energy spectral density within a specific frequency range, including: The dynamic response data is preprocessed to obtain preprocessed dynamic response data, wherein the preprocessing includes removing DC bias and power frequency interference; Based on the recovery phase of the joint from the low stiffness state to the high stiffness state and the historical damage assessment data of the joint, the dynamically adjusted characteristic frequency range is obtained. The analysis frequency band for the time-frequency analysis is set based on the dynamically adjusted characteristic frequency range, the time-frequency analysis is performed on the preprocessed dynamic response data, and the energy spectral density is calculated based on the dynamically adjusted characteristic frequency range.

6. The adaptive joint control method according to claim 5, characterized in that, After calculating the energy spectral density based on the dynamically adjusted characteristic frequency range, the method further includes: The time-frequency characteristics of environmental noise in the workspace of the humanoid robot are obtained, and during the application of the weak excitation to the joint, the sensor time-varying data of other unexcited joints of the humanoid robot are collected, and the vibration time-frequency characteristics are extracted from the sensor time-varying data of the other unexcited joints. The other unexcited joints are those that have a mechanical coupling relationship with the joint and are not subjected to the weak excitation during the weak excitation. Monitor the time-varying characteristics of the energy spectral density of the joint within the dynamically adjusted characteristic frequency range; The time-varying characteristics are compared with the time-frequency characteristics of the environmental noise to identify the first time-varying component that satisfies the preset first similarity rule with the environmental noise; The time-varying characteristics are compared with the vibration time-frequency characteristics of the other unexcited joints to identify the second time-varying component that satisfies the preset second similarity rule with the vibration of the other unexcited joints. Subtracting the first time-varying component and the second time-varying component from the time-varying characteristics yields the energy spectral density time-varying characteristics after removing similar components; The step of comparing the energy spectral density with a pre-established baseline energy spectral density under healthy conditions to identify micro-vibration signals with preset time-frequency characteristics includes: comparing the time-varying characteristics of the energy spectral density with the time-varying characteristics of the baseline energy spectral density to identify micro-vibration signals with preset time-frequency characteristics.

7. The adaptive joint control method according to claim 6, characterized in that, Also includes: Extract the multidimensional features of the time-varying characteristics of the energy spectral density to obtain a multidimensional feature vector, wherein the multidimensional features include at least frequency features, amplitude features, phase features, bandwidth features, energy centroid features, and statistical features characterizing the time-varying distribution pattern. The multidimensional feature vector is compared with a pre-established damage pattern feature library, and the similarity between the multidimensional feature vector and the feature vector of each damage pattern in the damage pattern feature library is calculated. The damage pattern feature library contains multiple damage types and their corresponding feature vectors under different damage degrees. Based on the similarity, the most matching damage pattern is identified and a preliminary damage assessment result is output, wherein the preliminary damage assessment result includes at least the most matching damage pattern and its corresponding similarity. Based on the preliminary damage assessment results, when the similarity is lower than a preset similarity threshold, the micro-vibration signal corresponding to the multidimensional feature vector is marked as an unknown damage mode. Unclassified damage event information is continuously collected from the historical operation data of the humanoid robot, and the damage event information is feature extracted and classified to dynamically update the damage pattern feature library. Based on the recovery stage of the joint from the low stiffness state to the high stiffness state, the feature vectors in the damage mode feature library are adjusted to adapt to the changes in damage characteristics under different working conditions.

8. The adaptive joint control method according to claim 7, characterized in that, The step of identifying the most matching damage pattern based on the similarity and outputting a preliminary damage assessment result includes: When the similarity calculation results show that the similarity of multiple damage patterns is higher than the preset similarity threshold, and the similarity difference between multiple damage patterns is less than the preset difference threshold, the feature vectors of multiple damage patterns are analyzed to obtain the difference information of multiple damage patterns in frequency features, amplitude features, duration features and statistical features representing time-varying distribution patterns. The difference information is then weighted and evaluated in conjunction with the current operating conditions of the joint to determine the most matching damage pattern. The current operating conditions include at least one or more of the following: operating load, joint speed, joint output torque and ambient temperature. When the similarity calculation result fails to meet the preset discrimination conditions to distinguish different damage modes, the verification process is triggered, a set of probing stimuli are applied to the joint, and transient response data of the joint under the probing stimuli are collected; high-resolution time-frequency analysis is performed on the transient response data to extract unique response features for distinguishing different damage modes, and the matching result of the damage mode is updated based on the unique response features; Based on the determined damage pattern type and the current operating state of the joint, a preset damage assessment rule is selected, and the damage degree is preliminarily quantified in combination with the actual operating parameters of the joint, and a preliminary damage assessment result is output. The actual operating parameters include at least one or more of the following: joint torque fluctuation, joint position deviation, and temperature change trend. The preliminary damage assessment result includes at least the best matching damage pattern and its corresponding similarity, as well as the preliminary quantification result of the damage degree. The preliminary quantification result of the damage degree is used to determine the adjustment direction of the recovery speed and / or recovery path for the recovery control of the joint from the low stiffness state to the high stiffness state.

9. The adaptive joint control method according to claim 8, characterized in that, Based on the determined damage pattern type and the current operating state of the joint, a preset damage assessment rule is selected, including: Based on the determined damage mode type, select the corresponding initial assessment rule from the preset damage type-assessment model mapping table; Based on the current operating state of the joint, the parameters of the initial evaluation rule are dynamically adjusted to obtain the adjusted evaluation rule; The actual operating parameters of the joint are input into the adjusted evaluation rules to obtain a damage degree index that characterizes the degree of damage. When the damage severity index exceeds the applicable range of the preset damage assessment rule, an adaptive correction process is triggered to collect the real-time running data of the joint and the signal features used to characterize the damage. The signal features include the time-varying characteristics of the energy spectral density and / or the unique response features. Combined with historical similar damage case data, the parameters and structure of the adjusted assessment rule are corrected to obtain the damage assessment rule. During the damage assessment process, the evolution trend of damage signal characteristics is continuously monitored, and the damage severity index is corrected in combination with the operating load of the joint and environmental conditions.

10. An adaptive joint control system for humanoid robots, characterized in that, The system, applicable to scenarios where humanoid robots perform sub-millimeter-level high-precision assembly tasks and at least some joints are in a high-stiffness control state, includes: The perception data acquisition module is used to acquire relative motion information between an external object and the humanoid robot, motion state information of the humanoid robot itself, and internal mechanical state information of the joint drive device to form comprehensive perception data; wherein, the relative motion information includes at least distance information and relative velocity information between the external object and the humanoid robot, the motion state information includes at least joint velocity information and joint acceleration information, and the internal mechanical state information includes at least joint motor current information and joint torque information; A contact prediction module is used to perform contact prediction based on the comprehensive sensing data to obtain a contact prediction result. The contact prediction includes: assessing the possibility of contact between the external object and the humanoid robot, and determining a potential contact area. The contact risk assessment is based on one or more of the following conditions: the distance information is less than a preset safe distance threshold, and the relative speed information is greater than a preset approach speed threshold; within a preset time window, the joint motor current information and / or joint torque information exhibit abnormally rapid fluctuations exceeding a preset change threshold. The potential contact area is determined based on the joint that satisfies the condition of having the contact risk and / or the joint closest to the external object. The joint compliance control module is used to respond to the contact prediction result by sending compliance control commands to the joint actuators in the potential contact area through a fast command channel independent of the conventional trajectory position control process. This causes the joints in the potential contact area to adjust from a high stiffness state to a low stiffness state within a preset response time. The compliance control commands include: switching the joint control mode from position control to torque control or impedance control, and / or limiting the maximum output torque of the joint to a preset safe torque upper limit.

Images