An industrial robot joint residual life prediction method based on transfer learning

Abstract

The application discloses a kind of industrial robot joint residual life prediction methods based on transfer learning, it is related to equipment failure prediction technical field, including the following steps: obtaining target object target servo data, target event data and source domain life data containing source domain commutation segment;The target servo data and the target event data are intercepted effective commutation segment and calculate target wear compensation and target micro hysteresis intensity, generate target hysteresis order, and then determine the target load type of the effective commutation segment according to the target event data.This scheme extracts the source domain load type and source domain hysteresis order of source domain commutation segment and is screened with target load type, target hysteresis order double consistent, so that the source domain sample participating in migration meets working condition load consistent and commutation hysteresis process consistent simultaneously, reduce the mismatch of similar amplitude but different mechanism.

Description

Technical Field

[0001] This invention relates to the field of equipment failure prediction technology, specifically to a method for predicting the remaining lifespan of industrial robot joints based on transfer learning. Background Technology

[0002] Industrial robots are widely used in automated production scenarios such as precision electronic assembly, automotive parts processing, metal processing, and packaging and handling. Their joint components are usually composed of servo motors, reduction mechanisms, transmission components, and position feedback components. The joint operating state directly affects the robot's end-effector trajectory following ability, repeatability, and continuous operation stability. As industrial robots operate in a high-frequency, repetitive operation state for a long time, the reducers, bearings, gear meshing pairs, and transmission connections inside the joints will gradually show degradation phenomena such as wear, increased clearance, changes in lubrication status, or increased elastic backlash.

[0003] In existing technologies, life prediction for industrial robot joints typically employs state monitoring methods based on vibration, current, or temperature signals. Some existing solutions collect joint motor current, joint position errors, or external vibration signals, extracting features such as root mean square (RMS), peak value, and error statistics. These features are then evaluated using threshold judgment, machine learning models, or transfer learning models. The advantage of these solutions is that the data sources are relatively clear, allowing the use of the robot control system or external sensors to acquire operational status information. Furthermore, they can identify abnormal states or estimate remaining life when component degradation has shown a clear trend or stable fault characteristics exist. However, industrial robot joints, in processes such as connector assembly, dispensing, and... In high-frequency reciprocating motions with small angles, such as screw tightening, the current compensation, angle lag, and recovery process at the moment of commutation are not only caused by joint wear, but are also affected by the end-effector contact resistance, the program segment action stage, the reverse start-up process, and servo control compensation. If the life stage is judged solely based on similar response amplitudes, similar error magnitudes, or similar overall trends, it is difficult to distinguish whether the increase in compensation is caused by external contact load or by internal joint wear and obstruction. It is also difficult to distinguish whether the order of compensation peaks, lag peaks, and recovery inflection points matches the actual commutation degradation process. This leads to a technical mismatch in the life stage mapping object, causing subsequent remaining life predictions to be based on inconsistent commutation response processes. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the remaining lifespan of industrial robot joints based on transfer learning, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, this invention discloses a method for predicting the remaining lifespan of industrial robot joints based on transfer learning, applied to the scenario of predicting the early degradation lifespan of joints in electronic precision assembly robots, comprising the following steps:

[0007] Acquire the target servo data, target event data, and source domain lifetime data containing source domain commutation segments of the target object;

[0008] The effective commutation segment is extracted from the target servo data and the target event data, and the target wear compensation amount and the target micro-hysteresis intensity are calculated to generate the target hysteresis sequence. Then, the target load type of the effective commutation segment is determined according to the target event data.

[0009] Extract the source domain load type and source domain hysteresis sequence corresponding to each source domain commutation segment in the source domain lifetime data, and retain only the source domain commutation segments whose source domain load type is consistent with the target load type and whose source domain hysteresis sequence is consistent with the target hysteresis sequence.

[0010] In the retained source domain commutation segment, extreme value statistics are performed on the source domain lifetime data to generate source domain migration anchor point data containing the source domain wear compensation range and source domain micro-hysteresis range for each stage, and the target wear compensation amount and the target micro-hysteresis intensity are mapped to the source domain migration anchor point data:

[0011] When the target wear compensation amount falls within the source domain wear compensation range and the target micro-hysteresis intensity falls within the source domain micro-hysteresis range, the corresponding source domain lifetime data is determined as a stable migration stage.

[0012] Statistical convergence is performed on the effective commutation segment and its associated target hysteresis sequence and the stable migration stage to generate lifetime prediction data.

[0013] Secondly, this invention discloses an industrial robot joint remaining life prediction system based on transfer learning, comprising:

[0014] The data acquisition module is used to acquire the target servo data, target event data, and source domain lifetime data containing source domain commutation segments of the target object.

[0015] The target load attribution determination module is used to extract effective commutation segments from the target servo data and the target event data, calculate the target wear compensation amount and the target micro-hysteresis intensity, generate the target hysteresis sequence, and determine the target load type of the effective commutation segment based on the target event data.

[0016] The source domain same load same sequence filtering module is used to extract the source domain load type and source domain hysteresis sequence corresponding to each source domain commutation segment in the source domain lifetime data, and retain only the source domain commutation segments whose source domain load type is consistent with the target load type and whose source domain hysteresis sequence is consistent with the target hysteresis sequence.

[0017] The stable migration stage determination module is used to perform extreme value statistics on the source domain lifetime data in the retained source domain commutation segment, generate source domain migration anchor point data containing the source domain wear compensation range and source domain micro-hysteresis range for each stage, and map the target wear compensation amount and the target micro-hysteresis intensity to the source domain migration anchor point data.

[0018] When the target wear compensation amount falls within the source domain wear compensation range and the target micro-hysteresis intensity falls within the source domain micro-hysteresis range, the corresponding source domain lifetime data is determined as a stable migration stage.

[0019] The lifetime prediction output module is used to perform statistical convergence on the effective commutation segment and its associated target hysteresis sequence and the stable migration stage to generate lifetime prediction data.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. This scheme extracts the source domain load type and source domain hysteresis sequence of the source domain commutation segment and filters them with dual consistency of target load type and target hysteresis sequence. This ensures that the source domain samples participating in the migration simultaneously meet the requirements of consistent operating load and consistent commutation hysteresis process, reducing mismatches with similar amplitudes but different mechanisms. The scheme also performs extreme value statistics of wear compensation amount and microhysteresis intensity according to the life stage of the retained segment to form comparable stage boundaries. This ensures that when the target wear compensation amount and target microhysteresis intensity fall within the corresponding range, they can uniquely constrain the stable migration stage, improving the traceability and consistency of stage mapping.

[0022] 2. This scheme extracts source domain lifetime data that simultaneously satisfy the conditions that the target wear compensation amount falls within the source domain wear compensation range and the target microhysteresis intensity falls within the source domain microhysteresis range as candidate lifetime data, thus excluding stages that do not simultaneously match both types of degradation characteristics. By calculating the smaller absolute differences between the target wear compensation amount and the target microhysteresis intensity and the first and last values ​​of the corresponding range, the boundary safety distance of the target data within each candidate stage range can be quantified. Then, the smaller value between the first inner distance and the second inner distance is taken to generate the stage retention margin, which makes the stage retention result subject to a weaker matching dimension constraint. Finally, the first candidate lifetime data is selected in descending order of stage retention margin, which can determine the stable migration stage with the largest boundary margin when multiple candidate stages overlap, reducing stage mismapping. Attached Figure Description

[0023] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0024] Figure 1 A flowchart illustrating the steps of a method for predicting the remaining lifespan of an industrial robot joint based on transfer learning, provided by this invention.

[0025] Figure 2 A schematic diagram of the process for generating the target hysteresis order provided by the present invention;

[0026] Figure 3 A schematic diagram of the process for generating a stable migration phase provided by the present invention;

[0027] Figure 4 A schematic diagram of the process for generating lifetime prediction data provided by the present invention;

[0028] Figure 5 This invention provides a schematic diagram of the module functions of an industrial robot joint remaining life prediction system based on transfer learning. Detailed Implementation

[0029] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0030] Industrial robot joints endure high-frequency, small-angle reciprocating motions for extended periods in repetitive tasks such as precision electronic assembly. This leads to gradual degradation in internal components like reducers, bearings, gear meshing, and transmission connections, manifesting as wear, increased clearance, changes in lubrication, and increased elastic backlash. Current methods for predicting the lifespan of industrial robot joints typically rely on vibration signals, current signals, temperature signals, joint motor current, joint position errors, or external vibration signals. These signals are then used to extract features such as root mean square (RMS), peak value, and error statistics before thresholding, machine learning evaluation, or transfer learning estimation. While this approach is applicable when degradation has established a clear trend or exhibits relatively stable fault characteristics, it is less effective in small-angle, high-frequency reciprocating motions such as connector assembly, dispensing, and screw tightening. The current compensation, angle hysteresis, and recovery processes during reversal are simultaneously affected by end-effector contact resistance, program segment actions, reverse startup processes, and servo control compensation. This results in reversal segments with similar response amplitudes, error magnitudes, or overall trends not necessarily corresponding to the same internal joint degradation state. If current lifespan stage determination relies solely on response amplitude, error magnitude, or overall trend similarity, it is difficult to distinguish whether the increase in compensation originates from external contact load or internal wear and tear on the joint. It is also difficult to distinguish whether the order of compensation peak, hysteresis peak, and recovery inflection point conforms to the commutation degradation process. This results in a mismatch between the mapping object technology and the source domain lifespan stage and the target effective commutation segment.

[0031] When an electronic precision assembly robot performs connector assembly tasks, the joints of the target object need to perform small-angle, high-frequency reciprocating movements within a work cycle, and at the moment of reversal, there are increases in current compensation, angle hysteresis, and changes in the recovery process. This reversal process may occur in the program segment before the end tool contacts the workpiece, or it may occur in the program segment after the connector has contacted and is experiencing contact resistance. If the existing processing determines the life stage only based on the similarity between the compensation amplitude, error magnitude, or overall trend in the target servo data and the source domain reversal segment in the source domain life data, then the increase in current compensation caused by contact resistance may be confused with the increase in current compensation caused by wear resistance inside the joint. At the same time, if the order of compensation peak, hysteresis peak, and recovery inflection point in the reversal process is not distinguished, source domain reversal segments with increased compensation but inconsistent recovery processes may also be used as the life stage mapping object of the target effective reversal segment. This problem is more easily exposed when the target load type, target hysteresis sequence, and source domain load type and source domain hysteresis sequence are not correspondingly constrained, specifically manifested as the target effective reversal segment being mapped to source domain life data that is inconsistent with its reversal response process.

[0032] If the above issues are not addressed, the target wear compensation amount and target micro-hysteresis intensity may fall into the source domain wear compensation range and source domain micro-hysteresis range that do not correspond to the actual commutation response process when mapping lifetime stages, causing a shift in the basis for determining the stable migration stage. This shift will continue to propagate along the data processing link: firstly, it will cause inconsistencies in the load state and hysteresis order between the target effective commutation segment and the source domain commutation segment; secondly, it will cause the stable migration stage on which statistical convergence is based to include technically mismatched objects, thus making the lifetime prediction data based on inconsistent commutation response processes. The ultimate adverse consequence is that the correspondence between the source domain lifetime stage on which the remaining lifetime prediction depends and the actual commutation degradation process of the target joint will be inaccurate, affecting the basis for subsequent predictions of the early degradation lifetime of the joints in electronic precision assembly robots.

[0033] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] Example 1:

[0035] This embodiment provides a method for predicting the remaining lifespan of industrial robot joints based on transfer learning. See [link to relevant documentation]. Figure 1 The method can be executed by a processor deployed in an industrial robot control cabinet, edge computing gateway, or production line health management server. For ease of understanding, the following explanation uses the early degradation life prediction of joints in the assembly, dispensing, and screw-fastening stations of an electronic precision assembly robot as an example. However, this does not constitute a limitation on the application scenario; the method is also applicable to joints in automotive parts processing, metal processing, and packaging handling scenarios that primarily involve small-angle, high-frequency reciprocating motions. The target object refers to a single joint of the industrial robot whose remaining life prediction is to be performed. This joint consists of a servo motor, a reduction mechanism, a transmission assembly, and a position feedback assembly.

[0036] The method first acquires the target servo data, target event data, and source domain lifetime data containing the source domain commutation segment of the target object. The target servo data refers to the timing data output by the servo drive system according to the control cycle during the operation of the target object, which includes at least feedback angular velocity, torque command, servo current, angle command, and feedback angle. It comes from the servo bus sampling of the robot control system or the readback of the internal register of the driver. The purpose of acquiring the target servo data is that the joint wear disturbance at the moment of commutation is concentrated in the servo current compensation and angle feedback hysteresis. Only by acquiring the above-mentioned servo quantities synchronized with the control cycle can the response process be reconstructed and the wear component removed before and after the commutation zero position time in subsequent steps.

[0037] Target event data refers to the action event records of the workstation where the target object is located within a work cycle. It includes at least the action start time, input / output trigger time, action end time, and return time range. It comes from the robot program running log or the workstation cycle event stream of the production line controller. The purpose of collecting target event data is to determine whether the robot end effector has made contact with the external workpiece at the moment of reversal and what stage of operation it is in. This directly determines whether an external contact load is superimposed in the reversal response. Only by obtaining the time boundary of the action event can the load contact state of each effective reversal segment be determined in subsequent steps.

[0038] Source domain lifetime data refers to a collection of historical operation records from historical robots of the same model, with known lifetime stage affiliations. This collection includes several source domain reversal segments, each carrying a lifetime stage identifier from the source domain lifetime data. The lifetime stage identifier is a category identifier marking the degradation stage of the source domain reversal segment, indicating its proximity to the failure boundary. The lifespan stage identifier numbering convention is that a larger number indicates proximity to the failure boundary, and a smaller number indicates less degradation. The reference sample source types for the source domain lifetime data include operation records from historical robots of the same model. There are three types of data: historical robot operation record library, manual maintenance and replacement record document, and full life cycle failure boundary offline test set. Among them, the historical robot operation record library of the same model and the full life cycle failure boundary offline test set both contain servo timing content of successive commutation, which is used to calculate the source domain wear compensation amount and source domain micro hysteresis intensity according to the same caliber as the target side. Manual maintenance and replacement record document usually only records the replacement time point and does not contain the servo timing of successive commutation. Therefore, it is only used to mark the life stage identifier and failure boundary of each source domain commutation segment, and is not used as a data source for calculating the source domain wear compensation amount and source domain micro hysteresis intensity.

[0039] After data acquisition, the method extracts effective commutation segments from the target servo data and target event data, calculates the target wear compensation amount and target micro-hysteresis intensity, generates the target hysteresis sequence, and then determines the target load type of the effective commutation segment based on the target event data. An effective commutation segment refers to a continuous servo sampling segment extracted from the target servo data, centered on a single actual commutation; the target wear compensation amount refers to the net servo current excess caused by wear resistance within the joint after stripping normal servo control compensation; the target micro-hysteresis intensity refers to the joint characterization of the amplitude and duration of the angle feedback hysteresis caused by joint wear; the target hysteresis sequence refers to the temporal category corresponding to the order in which the compensation response and hysteresis response appear during the recovery process; and the target load type refers to the load contact state category between the robot end effector and the external workpiece at the moment of commutation. These four quantities characterize the same effective commutation segment from four dimensions: wear amplitude, hysteresis process, temporal structure, and external load. The target wear compensation amount and target micro-hysteresis intensity are used for numerical determination of subsequent domain mapping, while the target hysteresis sequence and target load type are used for strong constraint alignment in subsequent source domain screening.

[0040] It should be noted that in calculating the target wear compensation amount, this embodiment does not directly use the servo current amplitude or current increment of the commutation recovery process as the degradation index. Instead, it first uses the commutation zero time corresponding to the effective commutation segment as the boundary, and decouples the effective commutation segment into a braking phase and a recovery phase according to the physical timing sequence. Then, it constructs a compensation benchmark based on the normal control compensation correspondence between the torque command and the servo current under the same torque command change direction within the braking phase. The actual servo current of the recovery phase is subtracted from the reference current of the compensation benchmark under the same torque command to obtain a difference sequence. The first digit of the absolute value of the difference sequence is used as the target wear compensation amount. Since the braking phase and the recovery phase correspond to the same commutation, share the same servo controller parameters and the same joint mechanical structure, the compensation benchmark can represent the normal servo control compensation level of this commutation. Therefore, after the above comparison and subtraction, the normal servo control compensation is deducted as a reference amount. The target wear compensation amount represents the net current excess amount exceeding the normal servo control compensation, that is, the component caused by wear resistance inside the joint.

[0041] After obtaining the four quantities from the target side, the method extracts the source domain load type and source domain hysteresis sequence corresponding to each source domain commutation segment in the source domain lifetime data, and retains only the source domain commutation segments whose source domain load type and target load type are consistent and whose source domain hysteresis sequence is consistent with the target hysteresis sequence. The source domain load type and source domain hysteresis sequence are category quantities obtained by performing load contact state determination and response timing structure determination on the source domain commutation segment in the same way as on the target side, and their determination criteria are exactly the same as on the target side. The retention operation here is the first locking of the double strong constraint alignment. Its practical basis is that only when the two commutation segments are consistent in both external load contact state and internal response timing structure can their wear responses be compared and allowed to enter the same numerical range for comparison. Otherwise, even if the wear compensation values ​​are close, they may belong to different physical causes and a stage correspondence should not be established.

[0042] In the preserved source domain commutation segments, the method performs extreme value statistics on the source domain lifetime data to generate source domain migration anchor point data containing the source domain wear compensation range and source domain microhysteresis range for each stage. The target wear compensation amount and target microhysteresis intensity are then mapped to the source domain migration anchor point data. The source domain migration anchor point data refers to a reference target indexed by lifetime stage identifiers and containing the value ranges of source domain wear compensation amount and source domain microhysteresis intensity for each stage. When the target wear compensation amount falls within the source domain wear compensation range and the target microhysteresis intensity falls within the source domain microhysteresis range, the corresponding source domain lifetime data is identified as a stable migration stage. A stable migration stage refers to the source domain degradation stage anchored under the premise of consistent physical mechanisms for the target commutation event.

[0043] Finally, the method performs statistical convergence on the effective commutation segments and their associated target hysteresis sequences and stable migration stages to generate lifetime prediction data. Statistical convergence refers to the convergence and competitive selection of discrete single-event stage mapping results within a continuous observation window to eliminate interference caused by discrete noise and random shutdowns. Lifetime prediction data refers to the predicted output encapsulating the remaining number of cycles from the target object's joint to the failure boundary. The above steps—from data acquisition, effective commutation segment processing, dual strong constraint screening, and domain mapping to statistical convergence—are interconnected and together constitute the logical framework of the method in this embodiment.

[0044] Through the above technical solution, this embodiment decouples the wear compensation amount within the commutation response process using the normal control compensation relationship, constructs the hysteresis sequence based on the order of the compensation response and the hysteresis response, and then performs dual strong constraint alignment on the source domain commutation segment using the load type and the hysteresis sequence. Unlike the conventional processing method that judges the life stage based on the similarity of response amplitudes or overall trends, this embodiment performs stage mapping at the level of consistency of the physical mechanism of the commutation event rather than the level of numerical similarity. This makes the stable migration stage produced by the data acquisition and domain mapping the direct basis for determining the convergence window and checking the remaining number of cycles in the subsequent statistical convergence stage.

[0045] Unlike existing technologies that directly collect joint current, position error, or vibration signals and extract statistical features such as root mean square and peak values, then use similar feature values ​​or trends as the basis for lifespan stage mapping, this scheme first uses the normal control compensation relationship of the braking phase as an anchor point within the commutation response process. Around this anchor point, it extracts the net wear compensation amount in the recovery phase and constructs a hysteresis sequence based on the order of the compensation response and the hysteresis response. Then, it uses the dual consistency of load contact state and response timing structure as a prerequisite to screen source domain commutation segments. The technical effect generated by this organizational relationship, which eliminates non-corresponding stage mappings caused by numerical coincidences, is difficult to obtain through threshold tuning of existing statistical features or simple weighted combinations of several features. This is because threshold tuning and feature combinations always remain at the level of numerical similarity, making it difficult to introduce the two constraint dimensions independent of amplitude: load contact state and timing causal structure.

[0046] See Figure 2 The process of extracting effective commutation segments from target servo data and target event data, calculating target wear compensation and target micro-hysteresis intensity, and generating target hysteresis sequence number is further detailed as follows. This process contains independent formulas and multiple steps with independent inputs and outputs, and is therefore expanded in the form of sub-steps.

[0047] Extract the sampling position where the sign of the feedback angular velocity in the target servo data reverses, and extract the target event data based on the sampling position to generate a valid commutation segment; divide the valid commutation segment into braking phase and recovery phase according to the time sequence.

[0048] It should be noted that the process of truncating and dividing the phase of the effective commutation segment is as follows:

[0049] Sub-step 2.1: Input the feedback angular velocity sequence from the target servo data, determine the sign of the feedback angular velocity at each sampling point, and identify the sampling positions where the sign changes from positive to negative or from negative to positive as candidate commutation points. Output the sampling time of the candidate commutation points; where the feedback angular velocity sequence is denoted as... , For the sampling sequence number, when and When there are different signs, determine the first one. Each sampling location is a candidate reversal point.

[0050] Sub-step 2.2: Input the sampling time of the candidate reversing point and the start and end time range of the action in the target event data. Using the sampling time of the candidate reversing point as the reversing zero time, extract a preset number of continuous servo sampling points before and after it, and output the continuous servo sampling segment as the valid reversing segment; where the reversing zero time is denoted as... Sampling time within the segment by Converting zero point to local time ,Right now .

[0051] Sub-step 2.3: Input the torque command sequence and angle deviation characteristics within the effective commutation segment. Using the position of torque polarity change as the boundary, divide the sampling interval where the local time is less than zero and corresponds to the braking to zero process into the braking phase, and divide the sampling interval where the local time is greater than zero and corresponds to the reverse following recovery process into the recovery phase. Output the braking phase and the recovery phase. The braking phase refers to the sampling interval where the joint brakes from the motion state to zero speed before commutation, and the recovery phase refers to the sampling interval where the joint reverses and recovers following the angle command after commutation. The commutation zero position time output in the previous sub-step is the timing boundary point for dividing the phases in this sub-step.

[0052] It should be noted that the feedback angular velocity sequence may exhibit slight sign jitter near the commutation zero-position time due to sensor floor noise. Without constraints, this could lead to misclassification of pseudo-commutation points in the stationary hold state as candidate commutation points. Therefore, for the angle command sequences in the target servo data that do not form candidate commutation points, adjacent control cycle differentials are performed, and the maximum absolute value of the difference is extracted as the boundary of the position hold jitter band. Only when the absolute value of the angle command change before and after the candidate reversal point is greater than When the position is maintained, it is determined to be a valid commutation and is truncated. Since the position maintaining jitter band boundary comes from the jitter amplitude of the non-commutation period itself, its value adapts to the changes in ambient noise and hardware base noise, avoiding misjudgment caused by a fixed threshold.

[0053] A compensation benchmark is established based on the braking phase, and the difference sequence between the recovery phase and the compensation benchmark is calculated to generate the target wear compensation amount.

[0054] It should be noted that the calculation of the target wear compensation amount relies on the comparison of the currents of the braking phase and the recovery phase under the same torque command. The specific execution process will be elaborated in the explanation of the compensation benchmark and the difference sequence later. Here, we only explain its position in this process: after establishing the compensation benchmark from the braking phase, the difference sequence is obtained by performing a difference calculation between the actual servo current of the recovery phase and the compensation benchmark. The first digit of the absolute value sequence of the difference is taken as the target wear compensation amount. .

[0055] The angle deviation sequence of the recovered phase is calculated, and the peak value of the angle deviation sequence is multiplied by the recovery time span to generate the target microhysteresis intensity. The recovery time span is calculated by the difference between the time point of the peak value of the angle deviation sequence and the time point where the absolute value of the angle deviation changes from increasing to decreasing.

[0056] It should be noted that the calculation process for the target micro-hysteresis intensity is as follows: within the recovered phase, the difference between the angle command and the feedback angle is calculated point by point to obtain the angle deviation sequence. ,Right now ;

[0057] in, Angle command, For the feedback angle, the maximum absolute value of the angle deviation sequence is taken as the peak angle deviation. The corresponding local time is denoted as the hysteresis peak time. The local time corresponding to the sampling position where the absolute value of the angle deviation first changes from increasing to decreasing after the hysteresis peak is determined as the hysteresis recovery inflection point. Recovery time span The difference between the inflection point of the hysteresis recovery and the peak of the hysteresis; the target microhysteresis intensity. It is the product of the peak angle deviation and the recovery time span.

[0058] For example, the expression for calculating the target microhysteresis intensity is:

[0059] ;

[0060] It should be further explained that the reason for using the product of the peak angle deviation and the recovery time span, rather than using the peak angle deviation alone, is that joint wear during the recovery process simultaneously manifests as increased hysteresis and a prolonged recovery process. Using only the peak value would miss the degenerative information of a prolonged recovery process; by using the product form, In either case of increased amplitude or prolonged recovery, it monotonically increases with the degree of degradation, which physically means the approximate cumulative intensity of the angle deviation within the recovery phase over time.

[0061] The target hysteresis sequence is generated based on the order in which the peak values ​​of the difference sequence and the peak values ​​of the angle deviation sequence appear in the time sequence within the recovered phase.

[0062] It should be noted that the generation process of the target hysteresis sequence is as follows: within the recovered phase, the local time corresponding to the sampling position where the absolute value of the difference sequence reaches its maximum value is recorded as the peak time of the difference sequence. The hysteresis peak time corresponding to the peak angle deviation. In comparison; when Earlier When the target hysteresis order is of type I, it indicates that the compensated response reaches its peak value before the hysteresis response; when Later When the target hysteresis order is at the same sampling position, the second type is adopted, indicating that the hysteresis response reaches its peak value before the compensation response; when the two are at the same sampling position, the third type of target hysteresis order is adopted, indicating that the compensation response and the hysteresis response reach their peak values ​​simultaneously. The target hysteresis order is determined solely by... and The category quantity is determined by the order of their occurrence and does not change with the specific values ​​of their peak values.

[0063] It should be noted that this process relies on continuous servo sampling data input. When sampling packet loss occurs within a valid commutation segment, resulting in the loss of consecutive sampling points, if the number of missing points does not exceed the preset tolerable number of points, the process continues after bridging with linear interpolation of adjacent valid sampling points. If the number of missing points exceeds the tolerable number of points, or if it is difficult to detect the hysteresis recovery inflection point where the absolute value of the angle deviation changes from increasing to decreasing within the recovered phase, the valid commutation segment is deemed invalid, and the processing of that segment is terminated, moving on to the next candidate commutation point. Since interpolation bridging will not change the timing structure of the commutation response only when the missing range is small, the tolerable number of points ranges from no more than 3 to 5 consecutive sampling points, with a typical value of 3 sampling points.

[0064] Through the above technical solution, this embodiment decouples the commutation response process into a braking phase and a recovery phase by using the commutation zero-position time as the boundary. Within the recovery phase, the target wear compensation amount is extracted by a contrast-stripping method, the target micro-hysteresis intensity is quantified by the product of amplitude and time, and the target hysteresis sequence is encoded by the order of the peak values ​​of the compensation response and the hysteresis response. Unlike the conventional processing method of directly extracting a single amplitude value or error statistics from the entire response, this embodiment extracts features at the phase structure and temporal causality level of the commutation response process, so that the target wear compensation amount, the target micro-hysteresis intensity, and the target hysteresis sequence can be used as input basis for subsequent source domain screening and domain mapping.

[0065] The core concept of this embodiment regarding the effective commutation segment phase decoupling and hysteresis sequence encoding scheme lies in reconstructing the continuous and unified commutation servo response process into two independent analysis units, the braking phase and the recovery phase, according to the physical timing and control causal relationship. Within the recovery phase, the order of compensation response and hysteresis response is introduced as a timing category independent of amplitude. The difference from the prior art is that the prior art only extracts features in the amplitude dimension and lacks the characterization of the causal structure of the response timing. It can replace the existing joint degradation characterization technology based on single amplitude statistical features such as current root mean square and position error peak.

[0066] The process of establishing a compensation benchmark based on the braking phase and calculating the difference sequence between the recovered phase and the compensation benchmark to generate the target wear compensation amount is further detailed as follows. This process includes independent formulas and four independent input-output steps: benchmark establishment, interpolation to obtain the benchmark current, difference calculation, and sorting of values. Therefore, it is unfolded in the form of sub-steps.

[0067] The torque command and servo current in the same torque command change direction within the braking phase are extracted and combined into a coordinate point set. Then, the coordinate point set is sorted in ascending order according to the value of the torque command to generate a compensation reference.

[0068] It should be noted that the process for establishing the compensation benchmark is as follows:

[0069] Sub-step 3.1: Input the torque command sequence within the braking phase. With servo current sequence The sampling points corresponding to the same torque command change direction are selected, and the torque command and servo current of each sampling point are combined into a coordinate point, and the coordinate point set consisting of several coordinate points is output. The torque command change direction refers to the trend direction of the torque command increasing or decreasing with the sampling sequence number. Only sampling points with the same change direction are selected to avoid the current compensation characteristics of the braking process and the reverse process being mixed.

[0070] Sub-step 3.2: Input the coordinate point set output from sub-step 3.1, and execute the torque command at each coordinate point. The coordinate point set is sorted in ascending order of numerical values, and the relationship between the torque command and servo current after the ascending order is output as the compensation benchmark. The compensation benchmark refers to the monotonic relationship between the torque command and the required normal servo current under normal commutation conditions, which is used to provide a reference for the normal control compensation current under the same torque command for phase recovery.

[0071] Within the recovery phase, the servo current sampling data of the target servo data is extracted as the actual current value, and the corresponding servo current is extracted from the compensation reference as the reference current value according to each torque command; the difference between the actual current value and the corresponding reference current value is calculated to generate a difference sequence; the difference sequence is sorted by absolute value, and the difference at the top of the sorted result is used as the target wear compensation amount.

[0072] It should be noted that the calculation of the difference sequence and the extraction of the target wear compensation amount are as follows: within the recovery phase, the actual servo current at each sampling position is... As the actual current value, the torque command at this sampling location The corresponding servo current is retrieved from the compensation reference or obtained through linear interpolation and used as the reference current value. Subtracting the actual current value from the reference current value yields the difference sequence. Sort the difference sequence by absolute value from largest to smallest, and take the first difference in the sorted sequence as the target wear compensation amount. Furthermore, it is required that the direction of current exceeding the first and second difference values ​​corresponds to the direction of feedback hysteresis in terms of hysteresis recovery; otherwise, the second absolute value is taken.

[0073] For example, the expression for calculating the difference sequence is:

[0074] ;

[0075] It should be further explained that, since the braking phase and the recovery phase correspond to the same commutation, share the same servo controller parameters and the same joint mechanical structure, the compensation benchmark established based on the braking phase can represent the normal control compensation level for that commutation; after subtracting the benchmark current under the same torque command from the actual current of the recovery phase, the normal servo control compensation is deducted as a control quantity. This reflects the portion exceeding normal control compensation. The requirement that the direction of the current exceedance aligns with the direction of the feedback hysteresis is to eliminate occasional reverse current fluctuations and ensure... This corresponds to wear disturbances that originate from the same source as hysteresis, rather than random noise.

[0076] It should be noted that the compensation benchmark established in this process relies on the existence of a sufficient number of coordinate points within the braking phase that cover the torque command range of the recovery phase. When the number of coordinate points in the same torque command change direction within the braking phase is insufficient to cover the torque command value range of the recovery phase, for recovery phase sampling points falling outside the coverage range of the compensation benchmark torque command, the benchmark current value is obtained by extending the compensation benchmark endpoint, and the difference result of this extended interval is marked as an extrapolation estimate. If the proportion of the extrapolation interval to the recovery phase exceeds a preset upper limit, the compensation benchmark for this effective commutation segment is determined to be unreliable, and the calculation of the target wear compensation amount for this segment is terminated. Since an excessively high proportion of the extrapolation interval will reduce the representativeness of the compensation benchmark, the preset upper limit is set between 20% and 30%, with a typical value of 20%.

[0077] Through the above technical solution, this embodiment constructs a monotonic compensation benchmark for torque command and servo current using the braking phase of the same commutation, and uses this compensation benchmark to compare and subtract the actual current of the recovery phase under the same torque command. Unlike the conventional processing method that directly uses the absolute amplitude or current increment of the recovery process current as the degradation index, this embodiment extracts the target wear compensation amount at the level of net current excess after deducting normal servo control compensation, so that the target wear compensation amount serves as the core input for the numerical attribution determination in the subsequent domain mapping.

[0078] The core concept of this embodiment regarding the scheme of separating the target wear compensation amount based on the braking phase compensation benchmark is to use the normal control compensation relationship of the system itself in the same commutation as a reference to achieve decoupling of wear disturbance and normal servo control compensation. The difference from the prior art is that the prior art is difficult to deduct the normal servo control compensation from the commutation current, which causes external and internal factors to be mixed. It can replace the existing technology of judging joint degradation based on current amplitude threshold or current increment trend.

[0079] The process of determining the target load type of valid commutation segments based on target event data is further detailed as follows. This process includes multiple steps with independent inputs and outputs, such as extracting time boundaries and determining four types of attribution; therefore, it is presented in the form of sub-steps.

[0080] Extract the action start time, input / output trigger time, action end time, and return time range from the target event data; extract the reversal occurrence time from the valid reversal segment, and compare the reversal occurrence time with the action start time, input / output trigger time, action end time, and return time range respectively.

[0081] It should be noted that the process for determining the target load type is as follows:

[0082] Sub-step 4.1: Input the station event records in the target event data and extract the action start time, input / output trigger time, action end time and return time range; where the input / output trigger time refers to the moment when the robot end effector and the external workpiece make contact, which marks the start of the contact operation on the time axis; the return time range refers to the travel time interval during which the robot end effector leaves the workpiece and returns in the current work cycle.

[0083] Sub-step 4.2: Input the commutation zero time corresponding to the valid commutation segment as the commutation occurrence time, compare the commutation occurrence time with each time boundary extracted in sub-step 4.1, and output the target load type based on the comparison results.

[0084] Specifically, when the reversal occurs later than the start time of the action but earlier than the input / output trigger time, the target load type is determined as pre-contact reversal, indicating that the reversal occurs in the stage where the robot has started the action but has not yet made contact with the workpiece; when the reversal occurs no earlier than the input / output trigger time and no later than the end time of the action, the target load type is determined as in-contact reversal, indicating that the reversal occurs in the stage where the robot end effector is in contact with the workpiece; when the reversal occurs within the return time range, the target load type is determined as post-contact return reversal, indicating that the reversal occurs in the stage where the robot detaches from the workpiece and returns; when the reversal occurs in any of the above conditions, the target load type is determined as idle stroke reversal, indicating that the reversal occurs in the idle stroke stage where no contact operation has been triggered.

[0085] It should be noted that the above four types of judgment conditions are divided in the order of action start time, input / output trigger time, action end time and return time range on the time axis. The commutation during contact adopts a closed interval judgment that is no earlier than the input / output trigger time and no later than the action end time. The commutation before contact adopts a judgment that is earlier than the input / output trigger time. The two are closely connected with the input / output trigger time as a common boundary. There is no commutation time that does not belong to either category and falls within the action interval, thus ensuring that each effective commutation segment obtains a definite target load type.

[0086] Through the above technical solution, this embodiment classifies commutation events into four target load types based on the relative positions of the commutation occurrence time with the action start time, input / output trigger time, action end time, and return time range. Unlike the conventional approach of treating all commutation events together without distinguishing the external load contact state, this embodiment assigns commutation events based on the external load contact state dimension at the moment of commutation. This allows the target load type to participate in the selection of source domain commutation segments as the first constraint in the dual strong constraint alignment that locks the external load conditions.

[0087] The core concept of this embodiment regarding the scheme for determining the target load type based on the time boundary of the workstation action lies in defining the force contact state between the robot and the external workpiece at the moment of reversal from the macroscopic workstation action dimension, and constraining the source domain and target domain to be equally aligned under external physical conditions. The difference from the prior art is that the prior art only judges the degradation from the servo signal itself and ignores the external load contact state at the moment of reversal. It can replace the existing joint state evaluation technology that does not distinguish load conditions.

[0088] The process of performing extreme value statistics on source domain lifetime data and generating source domain migration anchor point data that includes the source domain wear compensation range and source domain micro-hysteresis range for each stage is further detailed as follows. This process includes multiple steps with independent inputs and outputs, such as grouping, extraction, extreme value interval calculation, and correlation output, and is therefore presented as sub-steps.

[0089] Read the lifetime stage identifiers associated with the retained source domain commutation segments from the source domain lifetime data, and classify and group the retained source domain commutation segments according to the lifetime stage identifiers; extract the source domain wear compensation amount and source domain microhysteresis intensity corresponding to the source domain commutation segments in each classification group.

[0090] It should be noted that the process of generating source domain migration anchor data is as follows:

[0091] Sub-step 5.1: Input the source domain commutation segments and their lifetime stage identifiers after being retained under double strong constraints. Classify and group the retained source domain commutation segments according to their lifetime stage identifiers, and output several classification groups marked by their lifetime stage identifiers. Since the retention operation has ensured that each source domain commutation segment is consistent with the target side load type and hysteresis sequence, the grouping by lifetime stage identifiers here results in a set of source domain commutation segments for each degradation stage under the same load and sequence conditions.

[0092] Sub-step 5.2: Input each classification group, extract the source domain wear compensation amount and source domain micro-hysteresis intensity corresponding to the source domain commutation segment in each classification group, and output the source domain wear compensation amount set and source domain micro-hysteresis intensity set for each group; wherein the source domain wear compensation amount and source domain micro-hysteresis intensity are obtained by performing the same phase decoupling and feature calculation as the target side on the source domain commutation segment, and their calculation caliber is consistent with the target wear compensation amount and target micro-hysteresis intensity.

[0093] Sub-step 5.3: Input the source domain wear compensation set and source domain microhysteresis intensity set for each group. Within each categorized group, sort all source domain wear compensation amounts and all source domain microhysteresis in ascending order. Determine the range of the source domain wear compensation range for that group by the interval formed by the first and last values ​​of the sorted source domain wear compensation amounts. The range of the source domain microhysteresis is determined by the interval span formed by the first and last values ​​of the source domain microhysteresis intensity ranking results. Output the source domain wear compensation range and source domain microhysteresis range for each group.

[0094] Sub-step 5.4: Input the identifiers of each lifespan stage and their corresponding source domain wear compensation ranges and source domain micro-hysteresis ranges. Associate these three data points and output source domain migration anchor point data indexed by the lifespan stage identifiers and containing the source domain wear compensation ranges and source domain micro-hysteresis ranges. The two types of intervals output in the previous sub-step are the objects associated in this sub-step.

[0095] It should be further noted that the source domain lifetime data serves as a reference library, and the reference sample sources are historical robot operation records of the same model, manual maintenance and replacement records, or offline test sets of failure boundaries throughout the entire life cycle. Since the stability of the extreme value range depends on a sufficient number of samples, the number of source domain reversal segments under each lifetime stage identifier should not be too low. For example, the number of source domain reversal segments in each stage should not be less than several dozen. The source domain migration anchor point data uses the lifetime stage identifier as the key partitioning method, and its update condition is: when new historical robot operation records of the same model or new failure boundary offline test data are introduced, grouping and extreme value statistics are re-executed to refresh the source domain wear compensation range and source domain microhysteresis range of each stage.

[0096] Through the above technical solution, this embodiment groups source domain reversal segments under the same load and sequence position conditions according to the life stage identifier, and constructs intervals by arranging the first and last values ​​of the source domain wear compensation amount and source domain microhysteresis intensity in ascending order within each group. This is different from the conventional method of statistically distributing the characteristics of all source domain samples without distinguishing the load and sequence position. This embodiment establishes extreme value intervals for each stage under the same load and sequence position grouping conditions that are isomorphic to the target side, so that the source domain migration anchor point data can be used as a reference target for subsequent domain mapping comparison.

[0097] See Figure 3 The process of identifying the corresponding source domain lifetime data as the stable migration stage is further detailed as follows. This process includes independent formulas and multiple steps with independent inputs and outputs, such as candidate extraction, calculation of the two types of inner distances, residual calculation, and descending selection. Therefore, it is unfolded in the form of sub-steps.

[0098] Source domain lifetime data that satisfy the requirement that the target wear compensation amount falls within the source domain wear compensation range and the target microhysteresis intensity falls within the source domain microhysteresis range are extracted as candidate lifetime data.

[0099] It should be noted that the process of extracting candidate lifetime data is as follows: [The text abruptly shifts to a different topic] ...the target wear compensation amount... Source domain wear compensation range for each group Compare them one by one to determine the target microhysteresis strength. The source domain microhysteresis range of each group Compare them one by one, and simultaneously satisfy... and The source domain lifetime data corresponding to the grouping is extracted as candidate lifetime data; candidate lifetime data refers to the source domain degradation stage that simultaneously contains the target wear compensation amount and the target microhysteresis intensity in numerical terms, and there may be one or more of them.

[0100] For each candidate lifetime data, the absolute difference between the target wear compensation amount and the first and last values ​​of the source domain wear compensation range under the corresponding group is calculated, and the smaller value is extracted as the first inner distance; the absolute difference between the target microhysteresis intensity and the first and last values ​​of the source domain microhysteresis range under the corresponding group is calculated, and the smaller value is extracted as the second inner distance.

[0101] It should be noted that the calculation process for the first inner distance and the second inner distance is as follows: For each candidate lifetime data, calculate the absolute difference between the target wear compensation amount and the lower and upper bounds of the wear compensation range of the grouped source domain, take the smaller of the two values, and then divide it by the interval width of the wear compensation range of the grouped source domain for normalization to obtain the dimensionless first inner distance. Calculate the absolute difference between the target micro-hysteresis intensity and the lower and upper bounds of the micro-hysteresis range of the source domain in this group. Take the smaller of the two values, divide it by the interval width of the micro-hysteresis range of the source domain in this group, and normalize to obtain the dimensionless second inner distance. The first inner distance and the second inner distance respectively represent the proportion of the target wear compensation amount and the target microhysteresis intensity in the corresponding interval to the width of the interval. The values ​​are both between zero and half. The larger the value, the closer the landing point is to the center of the interval and the more robust the attribution.

[0102] It should be further explained that, since the target wear compensation is measured in amperes and the target microhysteresis intensity is measured in degrees per second, the absolute difference between the two has different dimensions and cannot be directly compared. Therefore, before taking the smaller value to form the stage retention margin, the smaller value of the absolute difference is normalized by the width of the source domain interval in each region, so that the first inner distance and the second inner distance are converted into dimensionless relative margins before comparison. This ensures that the physical meaning of the stage retention margin does not change with the selection of the units of current and angle.

[0103] For example, the expressions for calculating the first inner distance and the second inner distance are as follows:

[0104] ;

[0105] ;

[0106] in, This indicates taking the smaller value.

[0107] Compare the first inner distance and the second inner distance corresponding to the candidate lifetime data, and extract the smaller value to generate the stage retention margin; sort the stage retention margins corresponding to each candidate lifetime data in descending order, and take the candidate lifetime data corresponding to the stage retention margin at the top of the list as the stable migration stage.

[0108] It should be noted that the process of generating the stage retention margin and determining the stable migration stage is as follows: for each candidate lifetime data, the smaller value between the first inner distance and the second inner distance is taken as the stage retention margin of that candidate lifetime data. The candidate lifetime data are sorted in descending order of stage retention margin, and the candidate lifetime data corresponding to the first-ranked data is taken as the stable migration stage. The stage retention margin is the smaller of two dimensionless inner distances, which physically means the most conservative relative margin that satisfies the interval assignment in both the wear compensation and microhysteresis dimensions of the candidate lifetime data. The stage with the largest stage retention margin is taken as the stable migration stage, that is, the candidate with the most robust assignment in both dimensions is selected, thereby resolving the mapping conflict when the target wear compensation amount and the target microhysteresis intensity fall into multiple adjacent stage intervals at the same time.

[0109] For example, the calculation expression for the stage retention margin is as follows:

[0110] ;

[0111] It should be noted that the domain mapping in this process is based on the premise that the target reversal event and the source domain reversal segment are comparable under the same load type and the same hysteresis sequence. When there is no group that simultaneously accommodates the target wear compensation amount and the target micro-hysteresis intensity, i.e., the candidate lifetime data is an empty set, it is determined that the target reversal event is difficult to be mapped under the current source domain migration anchor point data, no stable migration stage is assigned to it, and the mapping of the event is terminated; the handling in this termination case is the same as the handling when the source domain retrieval fails, and no unfounded stage assignment is generated.

[0112] Through the above technical solution, this embodiment constructs a stage retention margin by using the inner distance from the target wear compensation amount and the target microhysteresis intensity to the boundary of the source domain interval, and selects the best among multiple candidate lifetime data based on the maximum stage retention margin. This is different from the conventional processing method of directly determining the stage assignment by the feature falling into a certain interval. This embodiment resolves the one-to-many mapping conflict caused by the overlap of adjacent stage intervals at the level of comparing the boundary margins of candidate stage intervals, so that the determined stable migration stage becomes the stage assignment result carried by each effective commutation segment in the subsequent statistical convergence.

[0113] The core concept of this embodiment regarding the scheme for determining the stable migration stage by retaining a stage margin lies in the dual extreme value determination by taking the smaller of the inner distances in two dimensions—wear compensation and microhysteresis—and then taking the larger one among the candidate stages. This ensures the uniqueness and robustness of the single-event migration mapping result. The difference from the prior art is that the prior art lacks a disambiguation criterion when the feature falls into multiple adjacent stage intervals, which easily leads to mapping ambiguity. This can replace the existing technology that directly determines the degradation stage by a single threshold or nearest neighbor.

[0114] The process involves sorting the stage retention margins corresponding to each candidate lifetime data in descending order and selecting the candidate lifetime data corresponding to the stage retention margin at the top of the list as the stable migration stage. When there are ties for the top position, the process is further refined as follows. This process includes multiple steps with independent inputs and outputs, such as extracting the maximum value, statistically analyzing the number of cases, and handling different scenarios; therefore, it is presented as sub-steps.

[0115] Extract the largest value from the stage retention margins after sorting in descending order; count the number of candidate lifetime data with the largest value; when the number is a single value, take the candidate lifetime data corresponding to that single value as the stable migration stage.

[0116] It should be noted that the process for determining the uniqueness of the maximum stage retention margin is as follows: extract the maximum value of the stage retention margin of each candidate lifetime data after sorting it in descending order, and count the number of candidate lifetime data whose stage retention margin is equal to the maximum value; when the number is a single value, directly take the corresponding candidate lifetime data as the stable migration stage. At this time, the mapping result is unique and no downgrading is required.

[0117] When there are two or more candidates, compare the lifetime stage identifiers associated with each candidate lifetime data with the largest value, select the candidate lifetime data corresponding to the item with the smaller lifetime stage identifier as the target degradation data, and generate an overlap conflict marker; attach the overlap conflict marker to the target degradation data, and use the target degradation data as the stable migration stage.

[0118] It should be noted that the degradation process for overlapping stage boundaries is as follows: When there are two or more candidate lifetime data corresponding to the largest stage retention margin, it indicates that the target reversal event has the same robustness to multiple adjacent stages in terms of wear compensation and microhysteresis dimensions, making it difficult to distinguish them based on the margin. In this case, the lifetime stage identifiers associated with these candidate lifetime data are compared, and the candidate lifetime data corresponding to the smaller lifetime stage identifier is selected as the target degradation data. The target degradation data refers to the stable migration stage conservatively selected in the case of overlapping stage boundaries. Selecting the stage with the smaller lifetime stage identifier is biased towards the side with a less severe degradation, in order to avoid overestimating the degradation. At the same time, an overlap conflict marker is generated and attached to the target degradation data, and the target degradation data with the overlap conflict marker is taken as the stable migration stage. The overlap conflict marker is a confidence degradation marker indicating that the stable migration stage of the event is determined under the low confidence case of overlapping stage boundaries.

[0119] It should be noted that the overlapping conflict marker, as a process status identifier, is passed along with the target degradation data and participates in the calculation of the low confidence ratio in the subsequent statistical convergence. After the overlapping conflict marker is added, the event still gives a stable transition stage to maintain the integrity of the event processing, but it is included in the low confidence ratio when the window is optimized, thereby suppressing the impact of stage boundary overlapping events on the final convergence result without discarding the event.

[0120] It should be noted that the output of this process for each target reversal event is an event mapping record. In the case of unique determination, the event mapping record includes the event identifier, target wear compensation amount, target micro-hysteresis intensity, target hysteresis sequence, target load type, and stable migration stage fields. In the case of overlapping stage boundaries, the event mapping record adds an overlap conflict marker field to the above fields.

[0121] Through the above technical solution, this embodiment conservatively selects the target degradation data with the smaller lifetime stage identifier when the stage retention margin is tied for first place, and isolates it for reliability degradation with overlapping conflict markers. This is different from the conventional processing method of forcibly specifying a single stage or directly discarding conflict events when there is mapping ambiguity. This embodiment isolates overlapping events at the stage boundary in the low reliability proportion statistical dimension while preserving the integrity of the event, so that the target degradation data and overlapping conflict markers serve as the source of low reliability proportion in the subsequent window selection.

[0122] The core concept of this embodiment regarding the degradation and isolation of overlapping stage boundary events lies in isolating matching events that are difficult to disambiguate with margin as reliability degradation indicators, rather than forcibly specifying or discarding them. This prevents the propagation of local errors caused by forced specification, which would then skew the overall prediction trend. The difference from existing technologies is that existing technologies lack a reliability classification and isolation mechanism for mapping ambiguity events. This can replace the existing technology that forcibly classifies or completely eliminates uncertain matches.

[0123] See Figure 4 The process of statistically converging effective commutation segments and their associated target hysteresis order positions and stable migration stages to generate lifetime prediction data is further detailed as follows. This process includes independent formulas and multiple steps with independent inputs and outputs, such as window partitioning, two-class ratio calculation, sorting and filtering, and reverse lookup output. Therefore, it is unfolded in the form of sub-steps.

[0124] All valid commutation segments are arranged in chronological order of occurrence, and consecutive valid commutation segments appearing on the timeline are categorized into consecutive commutation candidate windows.

[0125] It should be noted that the process of dividing the continuous commutation candidate window is as follows: all valid commutation segments are arranged in chronological order of their commutation zero-position time, and combined with the job cycle number, adjacent segments with consecutive cycle numbers on the timeline that all contain valid commutation segments are merged into the same window. When the cycle number is interrupted or a valid commutation segment under a certain number is missing, it is truncated at the interruption position, thereby classifying the continuously occurring valid commutation segments into several continuous commutation candidate windows; the continuous commutation candidate window refers to the observation window system that reflects the continuity of system wear, used to exclude time discontinuities caused by discrete noise events and random shutdowns.

[0126] Within each consecutive candidate window for reversal, the ratio of the number of segments with the same target hysteresis order to the total number of segments in the window is calculated to generate the order consistency ratio; the ratio of the number of segments with overlapping conflict markers to the total number of segments in the window is calculated to generate the low confidence ratio.

[0127] It should be noted that the calculation process for the ordinal consistency ratio and the low confidence ratio is as follows: For each consecutive candidate window of reversal, the mode of the number of segments with the same target hysteresis ordinal position and that have formed a stable migration stage is counted. The ratio of the number of segments corresponding to this mode to the total number of segments in the window is used as the ordinal consistency ratio. The number of segments with overlapping conflict markers within a statistical window is used as the ratio of this number to the total number of segments within the window, and this ratio is taken as the low-confidence proportion. The sequence consistency ratio characterizes the degree of consistency in the response timing structure within the window, while the low confidence ratio characterizes the proportion of overlapping events at stage boundaries within the window.

[0128] For example, the formulas for calculating the ordinal consistency ratio and the low reliability ratio are as follows:

[0129] ;

[0130] in, The mode is the number of segments within the window that share the same target hysteresis order and have formed a stable transition phase. The number of segments within the window that have overlapping conflict markers. This represents the total number of segments within the window.

[0131] Based on the order consistency ratio and low confidence ratio of each consecutive commutation candidate window, all consecutive commutation candidate windows are sorted and filtered to extract the convergence window; the stable migration stage at the end of the time series in the convergence window is extracted, and the remaining cycle number associated with the stable migration stage is extracted from the source domain lifetime data to generate lifetime prediction data.

[0132] It should be noted that the process of extracting the convergence window and generating lifetime prediction data is as follows: Based on the order consistency ratio and low confidence ratio, combined with stage monotonicity, each consecutive commutation candidate window is sorted and filtered. The specific rules for sorting and filtering are elaborated in the following section on the sorting and filtering of consecutive commutation candidate windows, ultimately yielding a unique convergence window. The stable transition stage at the end of the time series within the convergence window is extracted, and the remaining cycle count from the failure boundary associated with this stable transition stage is determined in the source domain lifetime data according to the following reverse lookup mapping rules. It is then packaged into lifetime prediction data and output.

[0133] It should be noted that the reverse lookup mapping rule for determining the remaining number of cycles from the last stable migration stage is as follows: First, count the number of consecutive occurrences of the last stable migration stage in the tail sequence of the convergence window. This number of consecutive occurrences is determined as the target event advancement sequence number. The target event advancement sequence number is used to characterize the relative position of the target object within the last stable migration stage. The more consecutive occurrences, the deeper the advancement within this stage. Next, retrieve source domain reference points from the source domain lifetime data that have the same baseline attributes as the convergence window, i.e., the source domain load type is the same as the target load type, the source domain hysteresis sequence is the same as the target hysteresis sequence, and the lifetime stage identifier is the same as the last stable migration stage. Read the intra-stage sequence number of each source domain reference point within this stage. The source domain reference point whose intra-stage sequence number is equal to or closest to the target event advancement sequence number is determined as the matching reference point. Finally, read the remaining effective cycle quantity from the matching reference point to the failure boundary, which is recorded in advance, as the remaining number of cycles. The intra-stage sequence number refers to the position sequence number of the source domain reference point within its lifetime stage, arranged according to the order of operation. It adopts the same intra-stage counting caliber as the target event advancement sequence number, thereby further refining the stage attribution to the advancement position within the stage and avoiding the problem of not distinguishing the remaining cycle count of all reference points within the same stage due to reverse lookup based solely on stage attribution.

[0134] It should be noted that when there are two or more matching reference points, the difference between the target event progression sequence number and the sequence number within each stage of each matching reference point is calculated. The remaining effective loop quantity corresponding to the matching reference point with the smallest difference is uniquely retained and used as the remaining loop count. Output: When there are still more than one matching reference point with the smallest difference, take the median of their remaining effective loop counts as the remaining loop count. Output to address source domain redundancy caused by repeated sampling under the same advancement sequence number in the same stage.

[0135] It should be noted that the output of this process has two scenarios. The first is the normal output scenario, in which the lifetime prediction data includes at least the target object joint identifier, convergence window baseline attributes, final stable migration stage, target event progression sequence number, and remaining cycle count fields. The second is the non-updated output scenario, i.e., when the source domain lifetime data reading is interrupted, no matching segment matching the same type of strong constraint is found, or the remaining cycle count obtained through reverse lookup is an empty set, this process does not output the remaining cycle count, and the status alarm indicator is not updated.

[0136] It should be noted that this process relies on continuous and valid commutation segments within the window. When the total number of segments in a continuous commutation candidate window is less than the preset minimum number of window segments, the window is deemed statistically insignificant and removed from the candidates. When all continuous commutation candidate windows are removed due to insufficient segments or no window meets the filtering criteria after sorting and filtering, the aforementioned no-update output scenario is entered, suspending the current prediction update cycle and not outputting lifetime prediction data to avoid misleading updates when data is insufficient. Since a small total number of window segments can render the statistics on ordinal consistency and low confidence unrepresentative, the minimum number of window segments ranges from 10 to 40, with a typical value of 20.

[0137] Through the above technical solution, this embodiment divides the continuous reversal candidate window by the continuity of the work cycle, and then performs quantitative evaluation of the window by the order consistency ratio and low confidence ratio, and then checks the remaining cycle number. Unlike the conventional processing method of directly averaging or taking the latest value of the prediction results of discrete single events, this embodiment converges at the quality and consistency evaluation level of the continuous observation window and then maps it to the remaining lifetime, so that the lifetime prediction data becomes the final remaining cycle number output for industrial use.

[0138] The core concept of this embodiment regarding the scheme of generating lifetime prediction data by statistical convergence of continuous commutation candidate windows lies in constructing an observation window system that reflects the continuity of wear, and then using the two quantitative dimensions of ordinal consistency and low confidence to screen out the most representative convergence window and then back-check the quantification of remaining lifetime. The difference from the prior art is that the prior art lacks quality screening and monotonicity verification of the prediction results at the continuous window level, which can replace the existing technology of directly outputting remaining lifetime from discrete sample points.

[0139] Based on the ordinal consistency ratio and low confidence ratio of each consecutive commutation candidate window, the process of sorting and filtering all consecutive commutation candidate windows to extract the convergence window is further refined as follows. This process includes multiple steps with independent inputs and outputs, such as initial screening based on ordinal consistency ratio, secondary screening based on low confidence ratio, and final selection based on monotonic evolution characteristics. Therefore, it is unfolded in the form of sub-steps.

[0140] All consecutive switching candidate windows are sorted in descending order of their sequence consistency ratio, and the first-ranked ratio in the sorted results is extracted as the lower consistency limit. Continuous switching candidate windows whose sequence consistency ratio reaches the lower consistency limit are selected to construct a preliminary screening window set.

[0141] It should be noted that the process of constructing the initial screening window set is as follows: all consecutively switching candidate windows are sorted according to their order consistency ratio. Arrange the windows in descending order, and take the first position of the consistency ratio as the lower limit of consistency. Include the consecutive candidate windows whose consistency ratio reaches the lower limit of consistency into the initial screening window set. The lower limit of consistency refers to the consistency ratio threshold used to screen out windows with insufficient consistency in the response timing structure. The initial screening window set refers to the set of windows retained after being screened by the consistency ratio of the position.

[0142] The windows in the initial screening window set are sorted in ascending order of their low reliability proportions, and the proportion at the top of the sorted results is taken as the upper limit of reliability. Windows whose low reliability proportions do not exceed the upper limit of reliability are selected to construct the secondary screening window set.

[0143] It should be noted that the process of constructing the secondary screening window set is as follows: each window in the initial screening window set is sorted according to a low confidence ratio. The windows are sorted in ascending order, and the lowest reliability percentage is taken as the upper limit of reliability. Windows with a low reliability percentage not exceeding the upper limit of reliability are included in the secondary screening window set. The upper limit of reliability refers to the low reliability percentage threshold used to screen out windows with an excessively high proportion of overlapping events at the stage boundaries. The secondary screening window set refers to the set of windows retained after low reliability percentage screening based on the initial screening window set. The initial screening window set output from the previous sub-step is the input of this sub-step.

[0144] By comparing the stable migration stages contained in each window within the screening window set in chronological order, it is determined whether each stable migration stage exhibits monotonic evolution characteristics. If so, the window exhibiting monotonic evolution characteristics is selected as the termination window.

[0145] It should be noted that the final selection process for the convergence window is as follows: For each window in the rescreening window set, its stable migration stage sequence is read in chronological order of the effective commutation segments. It is determined whether the subsequent stable migration stage in the sequence is not lower than the preceding stable migration stage, i.e., whether there is an anomalous rebound in the degree of degradation. Monotonic evolution characteristics refer to the stable migration stage progressing in a non-regressive manner over time. Window exhibiting monotonic evolution characteristics is selected as the convergence window. Since joint wear is an irreversible cumulative process, the stable migration stage sequence corresponding to the actual degradation should progress monotonically over time. Using monotonic evolution characteristics as the final selection condition can exclude windows where stage rebound is caused by occasional mismapping.

[0146] Additionally, regarding the uniqueness of multiple windows exhibiting monotonic evolution characteristics within the rescreening window set: when two or more windows exhibiting monotonic evolution characteristics remain in the rescreening window set, a uniqueness selection is made according to a preset priority order of indicators. The preset priority order of indicators is: ordinal consistency ratio takes precedence over low confidence ratio, and low confidence ratio takes precedence over the total number of window segments. That is, the higher ordinal consistency ratio, the lower low confidence ratio, and the longer total number of window segments are compared step by step until a unique closing window is selected. If it is still difficult to uniquely determine the window after comparing all priorities, the latest window in the time series is selected as the closing window to reflect the most recent degradation state.

[0147] Through the above technical solution, this embodiment uses a multi-level competitive filtering method, which involves initial screening based on ordinal consistency ratio, secondary screening based on low reliability ratio, and final selection based on monotonic evolution characteristics of stable migration stage. This method differs from the conventional processing method of ranking windows based on a single indicator or simple weighting. This embodiment couples macroscopic degradation monotonicity and microscopic reliability consistency into the step-by-step filtering, so that the selected convergence window becomes the only reference window for subsequent reverse lookup of the remaining loop count.

[0148] The core concept of this embodiment regarding the multi-level competitive filtering scheme for extracting convergence windows lies in sequentially coupling and filtering response timing consistency, stage boundary overlap reliability, and degradation process monotonicity to converge a unique reference window with the highest prediction reliability and physical representativeness. The difference from existing technologies is that existing technologies lack joint constraints on candidate windows in terms of consistency, reliability, and monotonicity, and can replace existing technologies that use a single quality index to screen sample windows.

[0149] Example 2:

[0150] Please see Figure 5 A system for predicting the remaining lifespan of industrial robot joints based on transfer learning, comprising:

[0151] The data acquisition module is used to acquire the target servo data, target event data, and source domain lifetime data containing source domain commutation segments of the target object.

[0152] The target load attribution determination module is used to extract valid commutation segments from target servo data and target event data, calculate the target wear compensation amount and target micro-hysteresis intensity, generate the target hysteresis sequence, and determine the target load type of the valid commutation segment based on the target event data.

[0153] The source domain same load same sequence filtering module is used to extract the source domain load type and source domain hysteresis sequence corresponding to each source domain commutation segment in the source domain lifetime data, and retain only the source domain commutation segments whose source domain load type is consistent with the target load type and whose source domain hysteresis sequence is consistent with the target hysteresis sequence.

[0154] The stable migration stage determination module is used to perform extreme value statistics on the source domain lifetime data in the retained source domain commutation segments, generate source domain migration anchor point data containing the source domain wear compensation range and source domain micro-hysteresis range for each stage, and map the target wear compensation amount and target micro-hysteresis intensity to the source domain migration anchor point data.

[0155] When the target wear compensation amount falls within the source domain wear compensation range and the target micro-hysteresis intensity falls within the source domain micro-hysteresis range, the corresponding source domain lifetime data is determined as the stable migration stage.

[0156] The lifetime prediction output module is used to perform statistical convergence on effective commutation segments and their associated target hysteresis sequences and stable migration stages to generate lifetime prediction data.

[0157] This embodiment has the same technical effects as Embodiment 1.

[0158] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The data mentioned in this application have undergone normalization and other preprocessing to unify dimensions during formula calculations.

[0159] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the remaining lifespan of industrial robot joints based on transfer learning, applied to the scenario of predicting the early degradation lifespan of joints in electronic precision assembly robots, characterized in that... Includes the following steps: Acquire the target servo data, target event data, and source domain lifetime data containing source domain commutation segments of the target object; The effective commutation segment is extracted from the target servo data and the target event data, and the target wear compensation amount and the target micro-hysteresis intensity are calculated to generate the target hysteresis sequence. Then, the target load type of the effective commutation segment is determined according to the target event data. Extract the source domain load type and source domain hysteresis sequence corresponding to each source domain commutation segment in the source domain lifetime data, and retain only the source domain commutation segments whose source domain load type is consistent with the target load type and whose source domain hysteresis sequence is consistent with the target hysteresis sequence. In the retained source domain commutation segment, extreme value statistics are performed on the source domain lifetime data to generate source domain migration anchor point data containing the source domain wear compensation range and source domain micro-hysteresis range for each stage, and the target wear compensation amount and the target micro-hysteresis intensity are mapped to the source domain migration anchor point data: When the target wear compensation amount falls within the source domain wear compensation range and the target micro-hysteresis intensity falls within the source domain micro-hysteresis range, the corresponding source domain lifetime data is determined as a stable migration stage. Statistical convergence is performed on the effective commutation segment and its associated target hysteresis sequence and the stable migration stage to generate lifetime prediction data.

2. The method for predicting the remaining lifespan of industrial robot joints based on transfer learning according to claim 1, characterized in that: Extracting effective commutation segments from the target servo data and the target event data, calculating the target wear compensation amount and the target micro-hysteresis intensity, and generating the target hysteresis sequence specifically includes: Extract the sampling position where the sign of the feedback angular velocity in the target servo data reverses, and extract the target event data based on the sampling position to generate a valid reversal segment; The effective commutation segment is divided into braking phase and recovery phase according to the timing sequence; A compensation benchmark is established based on the braking phase, and the difference sequence between the recovery phase and the compensation benchmark is calculated to generate the target wear compensation amount. The angle deviation sequence of the recovered phase is calculated, and the peak value of the angle deviation sequence is multiplied by the recovery time span to generate the target microhysteresis intensity; wherein, the recovery time span is calculated by the difference between the time point of the peak value of the angle deviation sequence and the time point where the absolute value of the angle deviation changes from increasing to decreasing; The target hysteresis sequence is generated based on the order in which the peak values ​​of the difference sequence and the peak values ​​of the angle deviation sequence appear in the time sequence within the recovered phase.

3. The method for predicting the remaining lifespan of industrial robot joints based on transfer learning according to claim 2, characterized in that: Based on the braking phase, a compensation benchmark is established, and the difference sequence between the recovery phase and the compensation benchmark is calculated to generate the target wear compensation amount, specifically including: The torque command and servo current in the same torque command change direction within the braking phase are extracted and combined into a coordinate point set. Then, the coordinate point set is sorted in ascending order according to the value of the torque command to generate a compensation reference. Within the recovery phase, the servo current sampling data of the target servo data is extracted as the actual current value, and the corresponding servo current is extracted from the compensation reference as the reference current value according to each torque command. The difference between the actual current value and the corresponding reference current value is calculated to generate a difference sequence; The difference sequence is sorted by absolute value, and the difference that is first in the sorting result is taken as the target wear compensation amount.

4. The method for predicting the remaining lifespan of industrial robot joints based on transfer learning according to claim 1, characterized in that: Determining the target load type of the effective commutation segment based on the target event data specifically includes: Extract the action start time, input / output trigger time, action end time, and return time range from the target event data; Extract the reversal occurrence time from the valid reversal segment, and compare the reversal occurrence time with the action start time, the input / output trigger time, the action end time, and the return time range respectively: When the commutation occurs later than the action start time but earlier than the input / output trigger time, the target load type is determined to be pre-contact commutation. When the commutation occurs no earlier than the input / output trigger time and no later than the action end time, the target load type is determined to be commutation in contact. When the reversal time is within the return time range, the target load type is determined to be a post-contact return reversal. If the reversal time does not meet any of the above conditions, the target load type is determined to be an idle reversal.

5. The method for predicting the remaining lifespan of industrial robot joints based on transfer learning according to claim 1, characterized in that: Extreme value statistics are performed on the source domain lifetime data to generate source domain migration anchor point data that includes the source domain wear compensation range and source domain micro-hysteresis range at each stage. Specifically, this includes: Read the lifetime stage identifier associated with the retained source domain commutation segment from the source domain lifetime data, and classify and group the retained source domain commutation segment according to the lifetime stage identifier; Extract the source domain wear compensation and source domain microhysteresis intensity corresponding to the source domain commutation segment in each classification group; Within each classification group, all the extracted source domain wear compensation amounts and all the source domain microhysteresis intensities are arranged in ascending order, and the range between the first and last values ​​of each arrangement result is determined as the corresponding source domain wear compensation range and source domain microhysteresis range, respectively. The identifiers of each lifespan stage are associated with their corresponding source domain wear compensation ranges and source domain micro-hysteresis ranges to generate source domain migration anchor point data.

6. The method for predicting the remaining lifespan of industrial robot joints based on transfer learning according to claim 5, characterized in that: Determining the corresponding source domain lifetime data as a stable migration phase specifically includes: The source domain lifetime data that satisfy the target wear compensation amount falling within the source domain wear compensation range and the target microhysteresis intensity falling within the source domain microhysteresis range are extracted as candidate lifetime data. For each candidate lifetime data, the absolute difference between the target wear compensation amount and the first and last values ​​of the source domain wear compensation range under the corresponding group is calculated, and the smaller value is extracted as the first inner distance. Calculate the absolute difference between the first and last values ​​of the target microhysteresis intensity and the source domain microhysteresis range under the corresponding group, and extract the smaller value as the second inner distance. By comparing the first inner distance and the second inner distance corresponding to the candidate lifetime data, the smaller value is extracted to retain a margin during the generation stage; The stage retention margins corresponding to each candidate lifetime data are sorted in descending order, and the candidate lifetime data corresponding to the stage retention margin at the top of the list is taken as the stable migration stage.

7. The method for predicting the remaining lifespan of an industrial robot joint based on transfer learning according to claim 6, characterized in that: The candidate lifetime data are sorted in descending order of the stage retention margins corresponding to each candidate lifetime data, and the candidate lifetime data corresponding to the first-ranked stage retention margin is taken as the stable migration stage. Specifically, this includes: Extract the largest value from the remaining balance of the stages after sorting in descending order; Count the number of candidate lifetime data that have the maximum value; When the quantity is a single value, the candidate lifetime data corresponding to that single value is taken as a stable migration stage; When the number is two or more, compare the lifetime stage identifiers associated with each candidate lifetime data having the largest value, select the candidate lifetime data corresponding to the item with the smaller lifetime stage identifier as the target degradation data, and generate an overlap conflict marker. The overlapping conflict markers are appended to the target degradation data, and the target degradation data is used as a stable migration phase.

8. The method for predicting the remaining lifespan of an industrial robot joint based on transfer learning according to claim 7, characterized in that: Statistical convergence of the effective commutation segment and its associated target hysteresis sequence and stable migration phase to generate lifetime prediction data specifically includes: All the effective commutation segments are arranged in chronological order of occurrence, and the effective commutation segments that appear consecutively on the timeline are classified into consecutive commutation candidate windows. Within each of the consecutive commutation candidate windows, the ratio of the number of segments with the same target hysteresis order to the total number of segments in the window is calculated to generate the order consistency ratio. Calculate the ratio of the number of segments with the aforementioned overlapping conflict markers to the total number of segments within the window, and generate a low confidence ratio; Based on the order consistency ratio and low confidence ratio corresponding to each of the consecutive commutation candidate windows, all the consecutive commutation candidate windows are sorted and filtered to extract the convergence window. Extract the stable migration phase at the end of the time series within the convergence window, and extract the remaining cycle count associated with the stable migration phase from the source domain lifetime data to generate lifetime prediction data.

9. The method for predicting the remaining lifespan of industrial robot joints based on transfer learning according to claim 8, characterized in that: Based on the order consistency ratio and low confidence ratio corresponding to each of the consecutive reversal candidate windows, all the consecutive reversal candidate windows are sorted and filtered to extract the convergence window, specifically including: All the continuous reversal candidate windows are sorted in descending order of numerical values ​​according to the order consistency ratio, and the proportion value of the first one in the sorting result is extracted as the lower limit of consistency. Filter the consecutive reversal candidate windows whose sequence consistency ratio reaches the lower limit of consistency, and construct a preliminary screening window set; The windows in the initial screening window set are arranged in ascending order of their low confidence ratios, and the first-ranked ratio in the arrangement is extracted as the upper confidence limit. Filter the windows whose low confidence ratio does not exceed the upper confidence limit to construct a set of secondary screening windows; By comparing the stable migration stages contained in each window within the set of screening windows in chronological order, it is determined whether each stable migration stage exhibits monotonic evolution characteristics. If so, the window exhibiting the monotonic evolution characteristics is selected as the closing window.

10. A system for predicting the remaining lifespan of industrial robot joints based on transfer learning, characterized in that, include: The data acquisition module is used to acquire the target servo data, target event data, and source domain lifetime data containing source domain commutation segments of the target object. The target load attribution determination module is used to extract effective commutation segments from the target servo data and the target event data, calculate the target wear compensation amount and the target micro-hysteresis intensity, generate the target hysteresis sequence, and determine the target load type of the effective commutation segment based on the target event data. The source domain same load same sequence filtering module is used to extract the source domain load type and source domain hysteresis sequence corresponding to each source domain commutation segment in the source domain lifetime data, and retain only the source domain commutation segments whose source domain load type is consistent with the target load type and whose source domain hysteresis sequence is consistent with the target hysteresis sequence. The stable migration stage determination module is used to perform extreme value statistics on the source domain lifetime data in the retained source domain commutation segment, generate source domain migration anchor point data containing the source domain wear compensation range and source domain micro-hysteresis range for each stage, and map the target wear compensation amount and the target micro-hysteresis intensity to the source domain migration anchor point data. When the target wear compensation amount falls within the source domain wear compensation range and the target micro-hysteresis intensity falls within the source domain micro-hysteresis range, the corresponding source domain lifetime data is determined as a stable migration stage. The lifetime prediction output module is used to perform statistical convergence on the effective commutation segment and its associated target hysteresis sequence and the stable migration stage to generate lifetime prediction data.

Images