Humanoid robot joint module reliability test method based on machine learning

Abstract

The application discloses a humanoid robot joint module reliability test method based on machine learning, and relates to the technical field of robot component testing, comprising: adopting a phase credible distance value determination method to perform phase credibility judgment on synchronous test data corresponding to phase damage stress sequences of the joint module, mark low credibility degradation states, and generate a phase credibility determination record; based on the phase credibility determination record, adopting a counterfactual phase diagnosis action to backfill loading characteristics of adjacent phases to a suspected degradation phase, confirming degradation attribution according to abnormal response differences before and after backfilling, and generating a joint module diagnosis action response record. The application realizes degradation attribution confirmation by adopting a counterfactual phase diagnosis action to backfill loading characteristics of adjacent phases to a suspected degradation phase, and combining thermal degradation recovery stripping to distinguish reversible thermal degradation and irreversible degradation, thereby improving the traceability of the joint module reliability test report.

Description

Technical Field

[0001] This invention relates to the field of robot component testing technology, and in particular to a reliability testing method for humanoid robot joint modules based on machine learning. Background Technology

[0002] Reliability testing of humanoid robot joint modules primarily focuses on load-bearing motion components such as the hip, knee, ankle, shoulder, and elbow. Through task motion acquisition, joint bench loading, multi-source response monitoring, and degradation state analysis, the operational stability of the joint modules under continuous motion, load variations, directional impacts, and temperature accumulation conditions is evaluated. Conventional methods typically establish test sequences based on preset motion conditions or rated load conditions, collecting response data such as torque, angle, speed, and temperature, and combining this data with the number of cycles, performance changes, and temperature rise to form a reliability evaluation result.

[0003] Conventional methods in joint module testing typically focus more on response changes under fixed load conditions, failing to adequately characterize phase differences, damage source attribution, and damage equivalence relationships between action phases during humanoid robot task movements. Furthermore, when test data anomalies occur, conventional methods often rely on response deviations to directly determine the degradation state, which is insufficient in distinguishing between reversible anomalies caused by temperature rise and irreversible degradation related to joint structures, thus affecting the relevance of reliability test conclusions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning-based reliability testing method for humanoid robot joint modules to solve the problems of difficulty in converting humanoid robot task motion damage into equivalent bench loading content and difficulty in distinguishing between abnormal temperature rise and irreversible degradation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a machine learning-based method for testing the reliability of humanoid robot joint modules, which includes: collecting the identity information of the joint module under test and the humanoid robot's task action data, performing phase segmentation and damage contribution merging, and generating a phase damage equivalent sequence; The phase damage stress closed-loop mapping method is adopted to transform the damage contribution in the phase damage equivalent sequence into bench loading content, and rearrange it according to the damage equivalence relationship and temperature continuity relationship to generate the phase damage stress sequence of the joint module. The phase reliability distance value determination method is used to determine the phase reliability of the synchronous test data corresponding to the phase damage stress sequence of the joint module and mark the low reliability degradation state to generate a phase reliability determination record. Based on the reliable phase determination record, the loading characteristics of adjacent phases are re-injected into the suspected degenerate phase using counterfactual phase diagnostic actions, and the degeneration attribution is confirmed based on the difference in abnormal response before and after re-injection, generating a joint module diagnostic action response record. Thermal degradation recovery stripping is performed on the diagnostic motion response records of the joint module. Based on the degree of abnormal residue after temperature recovery, reversible thermal degradation and irreversible degradation are distinguished, and a joint module reliability test report is generated.

[0007] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the identity information of the joint module under test includes module number, installation joint position, rated torque, rated speed, reduction ratio, drive motor model, encoder model, manufacturing batch and cumulative running time. The humanoid robot's task motion data includes motion name, motion start and end time, target joint angle, actual joint angle, actual joint speed, actual output torque, joint temperature, plantar pressure value, body posture angle, and end-effector load weight.

[0008] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the steps of performing phase segmentation and damage contribution merging to generate a phase damage equivalent sequence are as follows: The identity information of the joint module to be tested is bound to the action data of the humanoid robot task to generate joint module identity-associated action data; Based on the time stamps and joint origin relationships in the joint module identity-associated motion data, motion continuity correction is performed to generate joint module aligned motion data; Based on the force and motion direction changes in the joint module alignment motion data, the motion phase is divided, and a motion phase segmentation record is generated. The damage contribution of each motion phase is marked by torque change, temperature rise change and backlash change, and a motion phase damage contribution record is generated. Action phases with the same damage origin in the action phase damage contribution record are equivalently merged to generate an equivalent sequence of phase damage.

[0009] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the step of using the phase damage stress closed-loop mapping method to transform the damage contribution in the phase damage equivalent sequence into bench loading content is as follows: Read the action phase and damage contribution in the phase damage equivalent sequence, and use the phase damage stress closed-loop mapping method to establish the phase loading mapping relationship between the damage contribution and the bench executable loading content. Based on the phase loading mapping relationship, the damage contribution is transformed into torque loading trajectory and angle loading trajectory, and the loading duration condition is matched to generate bench loading content.

[0010] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the step of rearranging according to the damage equivalence relationship and temperature continuity relationship to generate the joint module phase damage stress sequence is as follows: By combining the loading content of the test bench with the temperature status of the joint module under test, the temperature rise connection relationship between adjacent test bench loading contents is verified, and a continuous temperature verification record is generated. The execution order and loading intensity of the bench loading content record are adjusted according to the continuous temperature verification record to ensure that the damage equivalence relationship is consistent with the continuous temperature relationship, thereby generating a phase damage stress sequence for the joint module.

[0011] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the step of using the phase reliability distance value determination method to determine the phase reliability of the synchronous test data corresponding to the phase damage stress sequence of the joint module and mark the low reliability degradation state to generate a phase reliability determination record is as follows: Multi-source response data of the joint module under test in each action phase were collected according to the joint module phase damage stress sequence and written into the same time axis to generate synchronous test data. A phase reliability distance value determination method is used to perform phase response consistency analysis on synchronous test data, and the degree of response deviation under the same action phase is converted into a phase reliability distance value. Based on the phase confidence distance value, the current action phase is judged for phase confidence, and the action phase with an abnormally decreasing phase confidence distance value is marked as a low confidence degradation state. The low-confidence degradation state is linked with the corresponding action phase, phase confidence distance value and synchronous test data to generate a phase confidence determination record.

[0012] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the step of using counterfactual phase diagnosis to re-feed the loading features of adjacent phases to the suspected degenerate phase includes the following steps: Read the low-confidence degradation state and action phase source in the phase confidence determination record, and determine the position where the phase confidence distance value drops abnormally as the suspected degradation phase; Based on the suspected degenerate phase, find the loading features of the adjacent phases before and after the suspected degenerate phase, and filter the loading features that have an action connection relationship with the suspected degenerate phase; Loaded features with action connection relationships are fed back to suspected degenerate phases through counterfactual phase diagnosis, and adjacent phase perturbation verification is performed to identify abnormal response changes before and after execution.

[0013] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the steps of confirming the degradation attribution based on the difference in abnormal response before and after reflow and generating a joint module diagnostic action response record are as follows: Degradation attribution is determined based on the migration and changes of abnormal responses generated by the recharge of loading features, and degradation attribution tags are generated. The degradation attribution marker is associated with the phase reliability determination record and written into the diagnostic actions and abnormal response differences of the suspected degradation phase and counterfactual phase to generate the joint module diagnostic action response record.

[0014] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the step of performing thermal degradation recovery stripping on the joint module diagnostic action response records, and distinguishing between reversible thermal degradation and irreversible degradation based on the degree of abnormal residue after temperature recovery, is as follows: Read the degradation attribution markers and abnormal response migration changes in the joint module diagnostic action response records, filter out diagnostic segments that have the same abnormal response migration changes before and after temperature drop, and generate temperature recovery response control data. Compare the changes in abnormal responses before and after temperature recovery based on temperature recovery response control data, and calculate the degree of abnormal residual in the migration and changes of the same abnormal response before and after temperature recovery. Based on the degree of abnormal residue, reversible thermal degradation and irreversible degradation are distinguished and written into the corresponding diagnostic fragments to generate thermal degradation stripping markers.

[0015] As a preferred embodiment of the machine learning-based humanoid robot joint module reliability testing method of the present invention, the joint module reliability test report is generated by associating thermal degradation peeling marks with joint module diagnostic action response records, summarizing degradation attribution marks, abnormal response migration changes, and abnormal residue levels.

[0016] The beneficial effects of this invention are as follows: By adopting the phase damage stress closed-loop mapping method, the damage contribution in the phase damage equivalent sequence is transformed into bench loading content, and rearranged according to the damage equivalence relationship and temperature continuity relationship, realizing the continuous transformation of task action damage into the joint module phase damage stress sequence, enabling bench testing to take into account action phase, damage source and temperature state; by adopting counterfactual phase diagnostic action to backfeed the loading characteristics of adjacent phases to the suspected degradation phase, and combining thermal degradation recovery stripping to distinguish between reversible thermal degradation and irreversible degradation, the degradation attribution is confirmed, and the traceability of the joint module reliability test report is improved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a machine learning-based reliability testing method for humanoid robot joint modules.

[0019] Figure 2 A flowchart for generating response records for counterfactual phase diagnostic actions and joint module diagnostic actions.

[0020] Figure 3 Flowchart for generating reliability test reports for thermal degradation recovery peeling and joint modules.

[0021] Figure 4 A flowchart for phase damage stress closed-loop mapping and joint module phase damage stress sequence generation.

[0022] Figure 5 A flowchart for determining the reliable distance value of the phase and generating the reliable determination record of the phase is provided. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a machine learning-based method for reliability testing of humanoid robot joint modules, comprising the following steps: S1. Collect the identity information of the joint module under test and the action data of the humanoid robot task, and perform phase segmentation and damage contribution merging to generate a phase damage equivalent sequence.

[0027] The identification information of the joint module under test includes the module number, installation joint position, rated torque, rated speed, reduction ratio, drive motor model, encoder model, batch number, and cumulative running time.

[0028] Humanoid robot task motion data includes motion name, motion start and end time, target joint angle, actual joint angle, actual joint speed, actual output torque, joint temperature, plantar pressure value, body posture angle, and end-effector load weight.

[0029] The identity information of the joint module under test is bound to the task action data of the humanoid robot to generate joint module identity-related action data.

[0030] Furthermore, the module number, installation joint position, rated load parameters, and cumulative runtime are extracted from the identity information of the joint module under test. The action name, action start and end time, joint position identifier, and joint operation data are extracted from the humanoid robot task action data. Identity matching is established according to the consistency between the installation joint position and the joint position identifier. Then, action segments belonging to the same joint module under test are extracted according to the action start and end time. The module number, installation joint position, and rated load parameters are written into the corresponding action segments, so that each action segment has joint source, load constraints, and time boundaries, generating joint module identity-related action data that can continue to perform action continuity correction and phase segmentation.

[0031] Based on the time stamps and joint origin relationships in the motion data associated with the joint module identity, motion continuity correction is performed to generate joint module aligned motion data.

[0032] Furthermore, the time stamps and joint origin relationships in the joint module identity-associated motion data are read. When there is a time gap between adjacent motion segments belonging to the same joint module under test, and the time gap is larger than the allowable interval of the sampling period, the time gap is marked as a motion breakpoint. When there is a time overlap between adjacent motion segments belonging to the same joint module under test, and the joint origin relationship of the time overlap part is consistent, the motion segment with higher time stamp continuity and smoother joint angle changes is retained, and duplicate overlapping motion segments are deleted. When the actual joint angle, actual joint velocity, and actual output torque on both sides of the motion breakpoint meet the continuous change condition, the gap data is filled in according to the boundary values ​​on both sides of the motion breakpoint. When the continuous change condition is not met on both sides of the motion breakpoint, the motion breakpoint is written as the motion boundary into the corrected time stamps and joint origin relationships to generate joint module aligned motion data that can continue to perform motion phase division.

[0033] It should be noted that the joint origin relationship refers to the correspondence between the joint position identifiers, data acquisition channels, and motion segment time markers in the humanoid robot's task motion data, and the module number and installation joint position in the identity information of the joint module under test. This is used to identify which joint module under test each segment of motion data belongs to. The continuous change condition means that the actual joint angle difference, actual joint velocity difference, and actual output torque difference on both sides of the motion breakpoint are all within the allowable fluctuation range of the sampling period, and the direction of change is consistent with the motion trend before and after the breakpoint.

[0034] Based on the force and motion direction changes in the joint module alignment motion data, motion phases are divided, and motion phase segmentation records are generated. The damage contribution of each motion phase is marked by torque changes, temperature rise changes, and backlash changes, and a motion phase damage contribution record is generated.

[0035] Furthermore, the actual output torque, actual joint angle, and actual joint velocity in the joint module alignment motion data are read. Increases, decreases, and sudden changes in the actual output torque are considered force changes, and positive, negative, and near-stationary changes in the actual joint velocity are considered motion direction changes. When the actual output torque continuously increases and the actual joint velocity direction remains consistent, it is classified as the load-bearing driving phase; when the actual output torque remains stable and the actual joint velocity is close to stationary, it is classified as the load-bearing holding phase; when the actual joint velocity direction reverses and the actual output torque shows peak fluctuations, it is classified as the reversing impact phase; when the actual output torque decreases and the actual joint velocity is close to stationary, it is classified as the reversing impact phase. As the speed gradually decreases, the process is divided into an unloading recovery phase, and the start and end times, phase type, and corresponding joint source of the action phase are written to generate an action phase segmentation record. When marking damage contributions, the torque change, temperature rise change, and hysteresis change corresponding to the action phase segmentation record are read. When the torque change is continuously under high load, fatigue damage contribution is marked. When the temperature rise change continues to accumulate and the temperature drop is insufficient, thermal accumulation damage contribution is marked. When the hysteresis change expands within the same action phase, transmission clearance damage contribution is marked. The damage contribution is then written to the action phase start and end times, phase type, and joint source to generate an action phase damage contribution record.

[0036] It should be noted that backlash variation refers to the deviation between the actual output torque and the actual joint angle when the forward and reverse movements of the tested joint module pass through the same joint angle position in the same action phase. It is used to reflect the increased state of transmission clearance and joint response lag.

[0037] Action phases with the same damage origin in the action phase damage contribution record are equivalently merged to generate an equivalent sequence of phase damage.

[0038] Furthermore, each action phase includes a phase type, start and end times, joint source, fatigue damage contribution, thermal accumulation damage contribution, and transmission clearance damage contribution. When multiple action phases share the same primary damage source and the joint source is consistent, these action phases are grouped into the same damage source set. When the torque change direction, temperature rise accumulation trend, and hysteresis expansion trend are consistent within the same damage source set, the fatigue damage contribution, thermal accumulation damage contribution, and transmission clearance damage contribution are accumulated according to the start and end times of the action phase, forming the corresponding equivalent duration and equivalent damage contribution. When action phases with opposite trends exist within the same damage source set, these opposite-trend action phases are separated to avoid different degradation mechanisms from being mixed into the same equivalent result. The phase type, joint source, equivalent duration, equivalent damage contribution, and damage source corresponding to each damage source set are arranged in chronological order to generate a phase damage equivalent sequence.

[0039] S2. Using the phase damage stress closed-loop mapping method, the damage contribution in the phase damage equivalent sequence is transformed into bench loading content, and rearranged according to the damage equivalence relationship and temperature continuity relationship to generate the joint module phase damage stress sequence.

[0040] The action phase and damage contribution in the phase damage equivalent sequence are read, and the phase loading mapping relationship between the damage contribution and the bench executable loading content is established by using the phase damage stress closed-loop mapping method.

[0041] Furthermore, the damage contribution corresponding to each action phase is extracted according to the action phase sequence in the phase damage equivalent sequence. The phase damage stress closed-loop mapping method uses the damage contribution as the basis for inference, mapping the continuous change of high torque load to the torque amplitude, loading frequency and cycle duration in the test bench executable loading content, mapping the cumulative change of temperature rise to the loading holding time, temperature rise target and heat dissipation interval in the test bench executable loading content, and mapping the hysteresis expansion change to the forward and reverse switching angle, reverse loading torque and reversal number in the test bench executable loading content. The expected damage contribution is calculated for the test bench executable loading content corresponding to each action phase. When the expected damage contribution is consistent with the damage contribution in the phase damage equivalent sequence, the action phase, damage contribution and test bench executable loading content are written into the same correspondence to generate the phase loading mapping relationship.

[0042] It should be noted that the phase damage stress closed-loop mapping method refers to mapping the continuous changes in high torque load, cumulative temperature rise, and hysteresis amplification into bench-executable loading content based on the action phase and damage contribution in the equivalent phase damage sequence. The loading parameters are then corrected through reverse verification of the predicted damage contribution and the actual damage contribution until a consistent closed-loop phase loading mapping relationship is formed. Bench-executable loading content refers to the loading parameters and actions that can be directly applied to the joint module under test by the joint module test bench, including torque amplitude, loading frequency, cycle duration, load holding time, temperature rise target, heat dissipation interval, forward / reverse switching angle, reverse loading torque, and number of reversals.

[0043] Based on the phase loading mapping relationship, the damage contribution is transformed into torque loading trajectory and angle loading trajectory, and the loading duration condition is matched to generate bench loading content.

[0044] Furthermore, based on the phase loading mapping relationship, the torque amplitude, loading frequency, cycle duration, forward / reverse switching angle, reverse loading torque, and commutation number corresponding to the damage contribution are determined. The continuous change of high torque load is divided into loading stage, stable load stage, and unloading stage according to the torque amplitude, and repeated according to the loading frequency and cycle duration to form a torque loading trajectory. The hysteresis expansion change is divided into forward rotation stage, reverse rotation stage, and commutation dwell stage according to the forward / reverse switching angle, and the reverse loading torque is superimposed in the reverse rotation stage to form an angle loading trajectory. The loading holding time, temperature rise target, heat dissipation interval, commutation number completion status, and equivalent duration are used as loading duration conditions. When the torque loading trajectory and angle loading trajectory reach the corresponding damage contribution and meet the loading duration conditions, the bench loading content is generated.

[0045] It should be noted that the torque loading trajectory refers to the time-varying torque application process within the bench loading content, including the loading stage, stabilization stage, unloading stage, and the corresponding loading frequency and cycle duration. The angle loading trajectory refers to the time-varying joint angle movement process within the bench loading content, including the forward rotation stage, reverse rotation stage, and reversing dwell stage, and reflects the forward / reverse switching angle, reverse loading torque, and number of reversals. The loading duration condition refers to the conditions for determining whether the bench loading content has reached the end requirements, including whether the equivalent duration is met, whether the cycle duration is completed, whether the temperature rise target is reached, whether the heat dissipation interval is met, and whether the number of reversals is completed.

[0046] By combining the loading content of the test bench with the temperature status of the joint module under test, the temperature rise connection relationship between adjacent test bench loading contents is verified, and a continuous temperature verification record is generated.

[0047] Furthermore, by combining the loading start and end times, torque loading trajectory, angle loading trajectory, loading holding time, and heat dissipation interval in the test bench loading content, as well as the pre-loading temperature, loading end temperature, and temperature drop rate in the temperature status of the joint module under test, temperature continuity verification is performed on adjacent test bench loading content. When the loading end temperature at the end of the current test bench loading content, after being corrected for the temperature drop rate according to the heat dissipation interval, is within the temperature continuity allowable range when compared with the loading end temperature at the beginning of the next test bench loading content, the temperature rise connection between adjacent test bench loading content is determined to be continuous. When the loading end temperature of the current test bench loading content is too large compared with the loading end temperature of the next test bench loading content, and the heat dissipation interval is insufficient to explain the temperature drop amplitude, it is marked as temperature discontinuity. When the loading end temperature of the current test bench loading content, when superimposed with the expected temperature rise of the next test bench loading content, exceeds the temperature rise target, it is marked as temperature overshoot. The adjacent test bench loading content number, loading end temperature, loading end temperature, temperature drop correction value, temperature rise connection status, and abnormal reason are written to generate a temperature continuity verification record.

[0048] It should be noted that the temperature status of the joint module under test refers to the temperature changes of the joint module before, during, and after the loading process on the test bench, including the temperature before loading, the temperature during loading, the temperature at the end of loading, the rate of temperature drop, and the temperature stability. The temperature rise continuity refers to whether the temperature changes between adjacent test bench loading processes can be continuously connected. Specifically, it refers to whether the temperature at the end of loading after the previous test bench loading process, after a heat dissipation interval, can continuously correspond to the temperature before loading at the beginning of the next test bench loading process.

[0049] The execution order and loading intensity of the bench loading content record are adjusted according to the continuous temperature verification record to ensure that the damage equivalence relationship is consistent with the continuous temperature relationship, thereby generating a phase damage stress sequence for the joint module.

[0050] Further adjustments to the execution order and loading intensity of the test bench loading content records are made based on the temperature discontinuity, temperature rise overshoot, and temperature rise continuity states in the continuous temperature verification records. When there is a temperature discontinuity between adjacent test bench loading content records, the test bench loading content records with insufficient temperature drop are moved backward, or a heat dissipation interval is inserted between adjacent test bench loading content records so that the pre-loading temperature of the subsequent test bench loading content record can withstand the loading end temperature of the previous test bench loading content record. When there is a temperature rise overshoot between adjacent test bench loading content records, the torque amplitude, loading frequency, or loading holding time in the subsequent test bench loading content record is reduced, and the temperature continuity is maintained by extending the heat dissipation interval. When the adjusted test bench loading content records result in the expected damage contribution being lower than the damage contribution in the phase loading mapping relationship, the damage equivalence relationship is supplemented by increasing the cycle duration or the number of reversals. When the damage equivalence relationship and the temperature continuity relationship are satisfied simultaneously, the adjusted execution order, loading intensity, loading duration conditions, and corresponding action phases are written into the same sequence to generate the joint module phase damage stress sequence.

[0051] It should be noted that the loading intensity is obtained based on the damage contribution, torque amplitude, loading frequency, loading holding time, reverse loading torque, and temperature rise target in the phase loading mapping relationship, characterizing the magnitude of the stress applied to the joint module under test by the bench loading content record. The damage equivalence relationship refers to the correspondence between the expected damage contribution formed by the adjusted bench loading content record and the damage contribution in the phase damage equivalence sequence. The temperature continuity relationship refers to the relationship where the loading end temperature, the temperature drop after the heat dissipation interval, and the pre-loading temperature can be continuously maintained between adjacent bench loading content records. Temperature rise overshoot refers to the state where the loading end temperature after the previous bench loading content ends has not fully dropped, and the temperature rise expected from the next bench loading content, after being superimposed, exceeds the corresponding temperature rise target or the upper limit of the allowable temperature of the joint module under test.

[0052] S3. Using the phase reliability distance value determination method, the synchronous test data corresponding to the phase damage stress sequence of the joint module are judged for phase reliability and marked as low reliability degradation state, and a phase reliability determination record is generated.

[0053] Multi-source response data of the joint module under test in each action phase were collected according to the phase damage stress sequence of the joint module and written into the same time axis to generate synchronous test data.

[0054] Furthermore, data acquisition is initiated according to the start and end times of the action phase in the joint module phase damage stress sequence. The actual output torque is obtained by the torque sensor at the output end of the joint module under test, the actual joint angle is obtained by the encoder, the actual joint speed is calculated by the change in actual joint angle between adjacent sampling times, and the joint temperature is obtained by the temperature sensor on the surface of the joint module under test. The actual output torque, actual joint angle, actual joint speed, and joint temperature are all written into the corresponding acquisition time, and the start time of the action phase is used as a unified time zero point. Different acquisition times are converted into relative times on the same time axis to generate synchronous test data.

[0055] A phase reliability distance value determination method is used to perform phase response consistency analysis on synchronous test data, and the degree of response deviation under the same action phase is converted into a phase reliability distance value.

[0056] The expression for converting the response deviation under the same action phase into a phase confidence distance value is as follows: ; in, Indicates the first The reliable distance value of the phase of each action phase. Index variable representing the phase of the action. This represents the full-scale value of the reliable phase distance. Indicates the first The influence weight of response deviation in the determination of phase reliability distance value for multi-source response data. Indicates the first The first action phase The current test value of multi-source response data. Indicates the first The first action phase Health baseline values ​​for multi-source response data. This represents the allowable fluctuation scale of the i-th type of multi-source response data under the p-th action phase.

[0057] Furthermore, the synchronous test data is aggregated into multi-source response data under the same action phase according to the source of the action phase. The current test value of each type of multi-source response data is compared with the corresponding health benchmark value, and converted into the response deviation degree according to the allowable fluctuation scale. The comprehensive response deviation degree under the same action phase is obtained by weighting and synthesizing the multi-source response data according to the influence of each type of multi-source response data on the degradation judgment. The comprehensive response deviation degree is converted through an exponential decay relationship, so that the larger the response deviation degree, the lower the confidence degree, and the smaller the response deviation degree, the higher the confidence degree, thus generating a phase confidence distance value.

[0058] It should be noted that the response deviation impact weight is based on the sensitivity of various multi-source response data to anomalies in the degradation test of similar joint modules under test. It is set by comparing the magnitude of the impact of deviations in actual output torque, actual joint angle, actual joint speed, and joint temperature on the decrease in the phase reliability distance value. The health benchmark value is based on repeated test data of the same action phase under the healthy state of similar joint modules under test. It is set by selecting the average value within the stable operating range. The allowable fluctuation scale is based on the normal deviation that can occur in the same action phase under the healthy state. It is set by statistically analyzing the maximum normal deviation between multiple health test values ​​and the health benchmark value and adding the sensor acquisition error.

[0059] The current action phase is judged based on the phase confidence distance value, and the action phase with an abnormally decreasing phase confidence distance value is marked as a low confidence degradation state.

[0060] Furthermore, the phase confidence distance value of the current action phase is compared with the confidence range formed by the same action phase under healthy conditions (e.g., when the phase confidence distance value is normalized to full scale and is between 0 and 1, the confidence range is between 0.75 and 1.00). When the phase confidence distance value falls within the confidence range, the current action phase response is determined to be consistent. When the phase confidence distance value is lower than the lower limit of the confidence range, and at least one of the multi-source response data (actual output torque, actual joint angle, actual joint speed, and joint temperature) under the same action phase continuously deviates from the healthy baseline value, the current action phase is determined to have an abnormal confidence decrease. When the abnormal confidence decrease occurs continuously in adjacent acquisition times, and the source of the action phase corresponding to the abnormal confidence decrease remains consistent, the action phase with the abnormally decreasing phase confidence distance value is marked as a low confidence degradation state.

[0061] The low-confidence degradation state is linked with the corresponding action phase, phase confidence distance value and synchronous test data to generate a phase confidence determination record.

[0062] Furthermore, after the low-confidence degradation state is formed, the corresponding action phase is determined according to the source of the action phase, and the falling time, falling amplitude, and number of consecutive falling of the phase confidence distance value are extracted. The start and end time of the corresponding action phase is used as the head of the chain, the phase confidence distance value is used as the judgment node, and the actual output torque, actual joint angle, actual joint speed, and joint temperature in the same time period in the synchronous test data are used as the response nodes. A chain correspondence relationship of "action phase - phase confidence distance value - synchronous test data - low-confidence degradation state" is established in chronological order to generate a phase confidence judgment record that can trace the source of the low-confidence degradation state.

[0063] S4. Based on the reliable phase determination record, the loading characteristics of adjacent phases are re-injected into the suspected degenerate phase using the counterfactual phase diagnosis action, and the degeneration attribution is confirmed according to the difference in abnormal response before and after re-injection, generating a joint module diagnostic action response record.

[0064] Read the low-confidence degradation state and action phase source in the phase confidence determination record, and determine the position where the phase confidence distance value drops abnormally as the suspected degradation phase; Furthermore, the source of the action phase, the phase confidence distance value, and the time period of the synchronous test data corresponding to the low confidence degradation state are extracted from the phase confidence determination record. When the phase confidence distance value of the current action phase is lower than the lower limit of the confidence range, and the decrease in the phase confidence distance value of the current action phase relative to the phase confidence distance value in the healthy state of the same action phase exceeds the decrease judgment threshold, and the state below the lower limit of the confidence range is repeatedly observed within the continuous acquisition time, it is determined that the phase confidence distance value has abnormally decreased. The source of the action phase and the start and end time of the action phase corresponding to the first occurrence of the abnormal decrease are determined as the suspected degradation phase.

[0065] It should be noted that the lower limit of the confidence range is based on the distribution of the phase confidence distance values ​​of the same action phase under the healthy state of similar joint modules under test, and is set by selecting the lowest phase confidence distance value that is still acceptable in the health test; the descent judgment threshold is based on the normal descent amplitude in the health test and the abnormal descent amplitude in the degradation test, and is set by taking the boundary value between the upper limit of the normal descent amplitude and the lower limit of the abnormal descent amplitude.

[0066] Based on the suspected degenerate phase, find the loading features of the adjacent phases before and after the suspected degenerate phase, and filter the loading features that have an action connection relationship with the suspected degenerate phase.

[0067] Furthermore, based on the start and end times of the action phase of the suspected degenerate phase and the joint origin, the preceding and following action phases with adjacent time boundaries and consistent joint origins are searched in the phase damage stress sequence of the joint module. The torque amplitude, loading frequency, angular loading direction, number of reversals, and loading duration conditions in the preceding and following action phases are extracted as loading features. When there is a relationship between the extracted loading features and the suspected degenerate phase that is time continuous, consistent in joint origin, can be inherited in motion direction, and can be transitioned in torque change, it is determined that there is an action connection relationship between the loading features and the suspected degenerate phase. When the temperature rise state corresponding to the loading feature will cause temperature rise overshoot, or the angular loading direction cannot form a continuous rotational relationship with the suspected degenerate phase, the loading feature is removed, and the loading features that can be fed back to the suspected degenerate phase for perturbation verification are retained, generating adjacent phase loading feature records.

[0068] It should be noted that the action connection relationship refers to the correspondence between the time boundary, joint source, motion direction, torque change and angle change of adjacent action phases and the suspected degenerate phase, so that the loading characteristics of adjacent action phases can be reasonably fed back to the suspected degenerate phase for perturbation verification.

[0069] Loaded features with action connection relationships are fed back to suspected degenerate phases through counterfactual phase diagnosis, and adjacent phase perturbation verification is performed to identify abnormal response changes before and after execution.

[0070] Furthermore, when loading features with action connection relationships are fed back to the suspected degradation phase, the torque amplitude change, angle loading direction change, and reversal number change in the previous or subsequent action phase are superimposed on the original torque loading trajectory and angle loading trajectory of the suspected degradation phase, while keeping them unchanged, to form the counterfactual phase diagnostic action. The adjacent phase disturbance verification specifically involves collecting the actual output torque, actual joint angle, actual joint speed, and joint temperature of the suspected degradation phase before and after feeding back the load under the same loading duration conditions, and comparing the response deviation position, deviation amplitude, and duration before and after the implementation of the counterfactual phase diagnostic action. When the abnormal response is enhanced, weakened, or transferred to the corresponding position of the adjacent loading feature as the feeding back load feature increases, decreases, or shifts, it is identified as an abnormal response change.

[0071] The degradation attribution is determined based on the migration and changes of abnormal responses generated by the loading feature recharge, and a degradation attribution tag is generated.

[0072] Furthermore, the position, magnitude, and duration of the response deviation before and after the implementation of the counterfactual phase diagnostic action are compared. When the abnormal response persists within the suspected degraded phase and the deviation magnitude continues to increase after the loaded feature is reinjected, the degradation is confirmed to be in the suspected degraded phase. When the abnormal response moves to the position corresponding to the loaded feature of the adjacent phase with the loading feature reinjection, the degradation is confirmed to be in the loaded feature of the adjacent phase. When the abnormal response weakens significantly with the loading feature reinjection and does not form a continuous deviation, the degradation is confirmed to be a phase connection perturbation anomaly, and the confirmation result is written into the degradation attribution flag.

[0073] The degradation attribution marker is associated with the phase reliability determination record and written into the diagnostic actions and abnormal response differences of the suspected degradation phase and counterfactual phase to generate the joint module diagnostic action response record.

[0074] S5. Perform thermal degradation recovery stripping on the joint module diagnostic action response record, distinguish between reversible thermal degradation and irreversible degradation based on the degree of abnormal residue after temperature recovery, and generate a joint module reliability test report.

[0075] Read the degradation attribution markers and abnormal response migration changes in the joint module diagnostic action response records, filter out diagnostic segments that have the same abnormal response migration changes before and after temperature drop, and generate temperature recovery response control data.

[0076] Furthermore, degeneration attribution markers, abnormal response migration changes, diagnostic action execution times, and joint temperature change times are extracted from the joint module diagnostic action response records. Diagnostic segments are divided into pre-temperature drop and post-temperature drop segments based on the time boundary of joint temperature decreasing from a high-temperature state to a stable temperature state. When the pre-temperature drop and post-temperature drop diagnostic segments correspond to the same degeneration attribution marker, and the abnormal response migration changes both exhibit the same response deviation position, the same deviation direction, and the same persistence characteristic, a before-and-after comparison relationship is established between the two diagnostic segments. When the abnormal response migration changes disappear or change to other deviation positions after temperature drop, they are not included in the same comparison relationship. Diagnostic segments that satisfy the same abnormal response migration change conditions, pre-temperature drop response values, post-temperature drop response values, and degeneration attribution markers are written accordingly to generate temperature recovery response comparison data.

[0077] By comparing the changes in abnormal responses before and after temperature recovery with the temperature recovery response control data, the degree of abnormal residual in the migration and changes of the same abnormal response before and after temperature recovery is calculated.

[0078] The expression for calculating the residual degree of the anomaly before and after temperature recovery, considering the migration changes of the same anomalous response, is as follows: ; in, Indicates the degree of abnormal residue. This indicates the abnormal response value after temperature recovery. Indicates the health baseline response value. This indicates the abnormal response value before the temperature recovered. This indicates the temperature before it returned to normal. This indicates the temperature after the temperature has recovered. This indicates the baseline temperature for health.

[0079] Furthermore, taking the same abnormal response migration and change as the calculation object, the abnormal response value before temperature recovery, the abnormal response value after temperature recovery, and the healthy baseline response value are extracted. The absolute difference between the abnormal response value after temperature recovery and the healthy baseline response value is taken as the abnormal deviation after recovery, and the absolute difference between the abnormal response value before temperature recovery and the healthy baseline response value is taken as the abnormal deviation before recovery. The difference between the two is used to obtain the abnormal retention ratio. The difference between the temperature before temperature recovery and the temperature after temperature recovery is divided by the difference between the temperature before temperature recovery and the healthy baseline temperature to obtain the temperature recovery correction ratio. The abnormal retention ratio and the temperature recovery correction ratio are multiplied to calculate the degree of abnormal residue of the same abnormal response migration and change before and after temperature recovery.

[0080] It should be noted that the extraction of abnormal response values ​​before temperature recovery, abnormal response values ​​after temperature recovery, and healthy baseline response values ​​is specifically as follows: in the temperature recovery response control data, the diagnostic segment corresponding to the same abnormal response migration change is identified; the response value that deviates the most from the healthy baseline in the diagnostic segment before temperature drop is taken as the abnormal response value before temperature recovery; the response value at the corresponding time and corresponding response position after temperature drop is taken as the abnormal response value after temperature recovery; and the stable response value of the same action phase and the same response position of the same joint module under test in the healthy state is taken as the healthy baseline response value.

[0081] Based on the degree of abnormal residue, reversible thermal degradation and irreversible degradation are distinguished and written into the corresponding diagnostic fragments to generate thermal degradation stripping markers.

[0082] Furthermore, the degree of abnormal residue corresponding to the migration and change of the same abnormal response in the temperature recovery response control data is compared with the residue judgment threshold. When the degree of abnormal residue is lower than the residue judgment threshold and the deviation of the abnormal response is significantly reduced after temperature recovery, the corresponding diagnostic segment is marked as reversible thermal degradation. When the degree of abnormal residue reaches the residue judgment threshold and the deviation of the abnormal response remains at the same response position and in the same deviation direction after temperature recovery, the corresponding diagnostic segment is marked as irreversible degradation. The degradation type, degree of abnormal residue, migration and change of abnormal response and the corresponding diagnostic segment number are associated and written to generate a thermal degradation stripping mark.

[0083] It should be noted that the residue determination threshold is based on the distribution of abnormal residue levels of the same type of joint module in reversible thermal degradation samples and irreversible degradation samples, and is set by selecting the boundary interval of abnormal residue levels of the two types of samples.

[0084] By associating thermal degradation stripping markers with joint module diagnostic action response records, and summarizing degradation attribution markers, abnormal response migration changes, and abnormal residue levels, a joint module reliability test report is generated.

[0085] Furthermore, based on the diagnostic segment number, action phase source, and counterfactual phase diagnostic action implementation time in the thermal degradation stripping mark, an association is established with the same diagnostic segment in the joint module diagnostic action response record. In the associated diagnostic segment, the degradation attribution mark is retained, and the abnormal response migration and change details and the degree of abnormal residue are written. A degradation source description is formed by combining the reversible thermal degradation and unreliable degradation types in the thermal degradation stripping mark. When multiple diagnostic segments exist for the same action phase, they are merged according to the degree of abnormal residue from high to low, retaining the diagnostic segment with the highest degree of abnormal residue as the primary degradation basis. The action phase source, degradation attribution mark, abnormal response migration and change details, degree of abnormal residue, and thermal degradation stripping mark are organized into test conclusion items to generate a joint module reliability test report.

[0086] In summary, this invention achieves continuous transformation of task action damage into the joint module phase damage stress sequence by employing a phase damage stress closed-loop mapping method, which converts the damage contribution in the phase damage equivalent sequence into bench loading content, and rearranges it according to the damage equivalence relationship and temperature continuity relationship. This enables bench testing to accommodate action phase, damage source, and temperature state. Furthermore, by employing counterfactual phase diagnostic actions to re-inject the loading characteristics of adjacent phases into suspected degradation phases, and combining thermal degradation recovery stripping to distinguish between reversible and irreversible thermal degradation, the degradation attribution is confirmed, improving the traceability of joint module reliability test reports.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A reliability testing method for humanoid robot joint modules based on machine learning, characterized in that, include: Collect the identity information of the joint module under test and the action data of the humanoid robot task, and perform phase segmentation and damage contribution merging to generate a phase damage equivalent sequence; The phase damage stress closed-loop mapping method is adopted to transform the damage contribution in the phase damage equivalent sequence into bench loading content, and rearrange it according to the damage equivalence relationship and temperature continuity relationship to generate the phase damage stress sequence of the joint module. The phase reliability distance value determination method is used to determine the phase reliability of the synchronous test data corresponding to the phase damage stress sequence of the joint module and mark the low reliability degradation state to generate a phase reliability determination record. Based on the reliable phase determination record, the loading characteristics of adjacent phases are re-injected into the suspected degenerate phase using counterfactual phase diagnostic actions, and the degeneration attribution is confirmed based on the difference in abnormal response before and after re-injection, generating a joint module diagnostic action response record. Thermal degradation recovery stripping is performed on the diagnostic motion response records of the joint module. Based on the degree of abnormal residue after temperature recovery, reversible thermal degradation and irreversible degradation are distinguished, and a joint module reliability test report is generated.

2. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 1, characterized in that: The identity information of the joint module under test includes the module number, installation joint position, rated torque, rated speed, reduction ratio, drive motor model, encoder model, manufacturing batch, and cumulative running time. The humanoid robot's task motion data includes motion name, motion start and end time, target joint angle, actual joint angle, actual joint speed, actual output torque, joint temperature, plantar pressure value, body posture angle, and end-effector load weight.

3. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 2, characterized in that, The steps for performing phase segmentation and damage contribution merging to generate an equivalent phase damage sequence are as follows: The identity information of the joint module under test is bound to the action data of the humanoid robot task to generate joint module identity-related action data; Based on the time stamps and joint origin relationships in the joint module identity-associated motion data, motion continuity correction is performed to generate joint module aligned motion data; Based on the force and motion direction changes in the joint module alignment motion data, the motion phase is divided, and a motion phase segmentation record is generated. The damage contribution of each motion phase is marked by torque change, temperature rise change and backlash change, and a motion phase damage contribution record is generated. Action phases with the same damage origin in the action phase damage contribution record are equivalently merged to generate an equivalent sequence of phase damage.

4. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 1, characterized in that, The phase damage stress closed-loop mapping method is used to transform the damage contribution in the equivalent phase damage sequence into bench loading content. The steps are as follows: Read the action phase and damage contribution in the phase damage equivalent sequence, and use the phase damage stress closed-loop mapping method to establish the phase loading mapping relationship between the damage contribution and the bench executable loading content. Based on the phase loading mapping relationship, the damage contribution is transformed into torque loading trajectory and angle loading trajectory, and the loading duration condition is matched to generate bench loading content.

5. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 4, characterized in that, The steps for rearranging the damage stress sequence of the joint module according to the damage equivalence relationship and the temperature continuity relationship are as follows: By combining the loading content of the test bench with the temperature status of the joint module under test, the temperature rise connection relationship between adjacent test bench loading contents is verified, and a continuous temperature verification record is generated. The execution order and loading intensity of the bench loading content record are adjusted according to the continuous temperature verification record to ensure that the damage equivalence relationship is consistent with the continuous temperature relationship, thereby generating a phase damage stress sequence for the joint module.

6. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 1, characterized in that, The phase reliability distance value determination method is used to determine the phase reliability of synchronous test data corresponding to the phase damage stress sequence of the joint module and mark the low reliability degradation state to generate a phase reliability determination record. The steps are as follows: Multi-source response data of the joint module under test in each action phase were collected according to the joint module phase damage stress sequence and written into the same time axis to generate synchronous test data. A phase reliability distance value determination method is used to perform phase response consistency analysis on synchronous test data, and the degree of response deviation under the same action phase is converted into a phase reliability distance value. Based on the phase confidence distance value, the current action phase is judged for phase confidence, and the action phase with an abnormally decreasing phase confidence distance value is marked as a low confidence degradation state. The low-confidence degradation state is linked with the corresponding action phase, phase confidence distance value and synchronous test data to generate a phase confidence determination record.

7. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 1, characterized in that, The steps for using counterfactual phase diagnostics to re-feed the loading features of adjacent phases to the suspected degenerate phase are as follows: Read the low-confidence degradation state and action phase source in the phase confidence determination record, and determine the position where the phase confidence distance value drops abnormally as the suspected degradation phase; Based on the suspected degenerate phase, find the loading features of the adjacent phases before and after the suspected degenerate phase, and filter the loading features that have an action connection relationship with the suspected degenerate phase; Loaded features with action connection relationships are fed back to suspected degenerate phases through counterfactual phase diagnosis, and adjacent phase perturbation verification is performed to identify abnormal response changes before and after execution.

8. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 7, characterized in that, The steps for determining the attribution of degeneration based on the difference in abnormal responses before and after re-irrigation, and generating a joint module diagnostic action response record, are as follows: Degradation attribution is determined based on the migration and changes of abnormal responses generated by the recharge of loading features, and degradation attribution tags are generated. The degradation attribution marker is associated with the phase reliability determination record and written into the diagnostic actions and abnormal response differences of the suspected degradation phase and counterfactual phase to generate the joint module diagnostic action response record.

9. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 1, characterized in that, The process of performing thermal degradation recovery stripping on the joint module diagnostic motion response records, and distinguishing between reversible and irreversible thermal degradation based on the degree of abnormal residue after temperature recovery, is as follows: Read the degradation attribution markers and abnormal response migration changes in the joint module diagnostic action response records, filter out diagnostic segments that have the same abnormal response migration changes before and after temperature drop, and generate temperature recovery response control data. Compare the changes in abnormal responses before and after temperature recovery based on temperature recovery response control data, and calculate the degree of abnormal residual in the migration and changes of the same abnormal response before and after temperature recovery. Based on the degree of abnormal residue, reversible thermal degradation and irreversible degradation are distinguished and written into the corresponding diagnostic fragments to generate thermal degradation stripping markers.

10. The reliability testing method for humanoid robot joint modules based on machine learning as described in claim 9, characterized in that, The joint module reliability test report is generated by associating thermal degradation peeling marks with joint module diagnostic action response records, summarizing degradation attribution marks, abnormal response migration changes, and abnormal residue levels.

Images