Segmented robot joint nonlinear friction compensation method and system

Abstract

The invention discloses a sectional type robot joint non-linear friction compensation method and system, and relates to the technical field of non-linear friction compensation, the sectional type robot joint non-linear friction compensation method comprises the following steps: S1, obtaining data, and carrying out sectional classification labeling to form a robot sectional labeling data set; s2, extracting a segmented friction feature parameter set, and performing cross-modal data matching to obtain a segmented friction feature matching result; s3, performing weighted fusion on segmented friction feature matching results, and performing screening to form a segmented friction compensation feature parameter set; s4, constructing a sectional type multi-mode friction compensation model, and performing iterative training to obtain a sectional type friction compensation result; and S5, verifying the result, performing actual motion control on the result, and collecting parameters in real time to generate a report. According to the method, the multi-modal feature extraction network and the cross-modal data matching mechanism are combined, so that accurate integration of the friction characteristic data and the mechanical response data is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to, in particular, a segmented robot joint nonlinear friction compensation method and system. BACKGROUND

[0002] As a key support for improving the motion precision and stability of high-end equipment, the robot joint nonlinear friction compensation technology has become a key research direction in the fields of precision manufacturing, automated assembly, etc. Friction, as a complex nonlinear phenomenon inherent to robot joints, combines multiple characteristics such as static friction, Coulomb friction, and viscous friction, and is dynamically affected by factors such as motion speed, load change, and temperature fluctuation. Its existence not only causes joint positioning errors to increase and motion to crawl, but also exacerbates component wear and increases energy consumption, severely restricting the application value of robots in high-precision scenarios such as semiconductor processing and precision assembly.

[0003] Joint nonlinear friction compensation is not only a key means to address the lack of robot motion precision, directly improving product pass rate and production efficiency, providing data support for equipment operation and maintenance, and extending joint service life by predicting friction wear state, but also a core grip to upgrade industrial robots to high precision and high reliability. By optimizing friction suppression effect, reducing motor drive load and energy consumption, and providing friction characteristic data for next-generation robot joint structure design, it avoids equipment failure and production stagnation caused by compensation strategy failure.

[0004] However, existing segmented robot joint nonlinear friction compensation methods rely on single-dimensional motion parameters, ignore the diagnostic value of friction characteristic data and mechanical response data, and cannot cover the coupling logic of motion state, friction characteristics, and mechanical feedback, resulting in one-sidedness of the compensation strategy with heavy parameters and light association. Moreover, existing segmented robot joint nonlinear friction compensation methods use simple splicing or static weight fusion, which cannot solve the heterogeneity problem of multi-modal data such as motion state and friction characteristics, resulting in ineffective alignment of data semantics, loose feature association, and susceptibility to noise interference. No effective solutions have been proposed to address the problems in related technologies. SUMMARY

[0005] To address the problems in related technologies, the present application proposes a segmented robot joint nonlinear friction compensation method and system to overcome the above technical problems existing in the prior art.

[0006] To achieve the above purpose, the specific technical solutions adopted by the present application are as follows: According to one aspect of the present application, a segmented robot joint nonlinear friction compensation method is provided, comprising the following steps: S1, acquire multi-source robot motion data and robot joint nonlinear friction data under segmented robot joint motion conditions, and segment and classify the multi-source robot motion data according to the robot joint motion state to form a robot segmented labeled data set; S2, pre-process the robot segmented labeled data set, extract the segmented friction feature parameter set in the robot segmented labeled data set by using a deep learning framework, and perform cross-modal data matching between the segmented friction feature parameter set and the robot joint nonlinear friction data to obtain a segmented friction feature matching result; As a preferred solution, the pre-processing of the robot segmented labeled data set, the extraction of the segmented friction feature parameter set in the robot segmented labeled data set by using a deep learning framework, and the cross-modal data matching between the segmented friction feature parameter set and the robot joint nonlinear friction data to obtain a segmented friction feature matching result include the following steps: S21, according to the segmented classification labeling of the robot segmented labeled data set, remove noise data, abnormal acquisition data and redundant data, and perform time axis alignment correction; S22, build a segmented multi-modal feature extraction network model based on a deep learning framework, and extract a standardized motion state data set, a friction characteristic data set and a mechanical response data set in the robot segmented labeled data set by using the segmented multi-modal feature extraction network model; S23, integrate the standardized motion state data set, the friction characteristic data set and the mechanical response data set to obtain a segmented friction feature parameter set; S24, calculate the sample matching degree of the segmented friction feature parameter set and the robot joint nonlinear friction data by using a cosine similarity algorithm, and establish a cross-modal correlation mapping table based on the sample matching degree for cross-modal data matching.

[0007] S3, preset a segmented matching screening rule and a segmented feature weight distribution rule, weight fuse the segmented friction feature matching result based on the segmented feature weight distribution rule, and screen the weighted fused segmented friction feature matching result by using the segmented matching screening rule to form a segmented friction compensation feature parameter set; As a preferred solution, the pre-processing of the robot segmented labeled data set, the extraction of the segmented friction feature parameter set in the robot segmented labeled data set by using a deep learning framework, and the cross-modal data matching between the segmented friction feature parameter set and the robot joint nonlinear friction data to obtain a segmented friction feature matching result include the following steps: S31, extract the robot joint friction characteristic parameters of each segment in the robot joint nonlinear friction data, and preset a segmented matching screening rule and a segmented feature weight distribution rule; S32, extract the segmented friction feature parameters of each segment from the segmented friction feature matching result, and calculate the single-modal feature weight coefficient of the segmented friction feature parameters based on the segmented feature weight distribution rule; As a preferred solution, the step of extracting the segmented friction feature parameters of each segment from the segmented friction feature matching result, and calculating the single-modal feature weight coefficient of the segmented friction feature parameters based on the segmented feature weight distribution rule comprises the following steps: S321, extract the segmented friction feature parameters from the segmented friction feature matching result according to the segment classification label in the robot segmented labeling data set; S322, based on the segmented feature weight distribution rule, determine the weight calculation constraint condition of the segmented friction feature parameters; S323, construct a single-modal weight calculation model according to the weight calculation constraint condition, input the segmented friction feature parameters into the model, and calculate the segmented single-modal feature weight coefficient; S324, reasonable check the segmented single-modal feature weight coefficient, eliminate the abnormal coefficient beyond the constraint range, and form the single-modal feature weight coefficient set.

[0008] S33, weight fusion of the segmented friction feature parameters of each segment and the single-modal feature weight coefficient, and modal feature alignment processing to obtain the fusion feature parameter set; As a preferred solution, the step of weight fusion of the segmented friction feature parameters of each segment and the single-modal feature weight coefficient, and modal feature alignment processing to obtain the fusion feature parameter set comprises the following steps: S331, perform dimension product operation on the segmented friction feature parameters and the single-modal feature weight coefficient; S332, accumulate and sum the segmented friction feature parameters after the dimension product operation to obtain the segmented fusion feature set, and construct a segmented fusion feature set alignment mapping model to map the segmented fusion features in the segmented fusion feature set to a unified feature space; S333, perform dimension normalization processing on the segmented fusion features in the mapped segmented fusion feature set, and integrate to form the fusion feature parameter set.

[0009] S34, remove invalid feature data in the fusion feature parameter set using the segmented matching screening rule, and screen the core feature parameters to form the segmented friction compensation feature parameter set.

[0010] S4, construct a segmented multi-modal friction compensation model based on the attention mechanism, input the segmented friction compensation feature parameter set into the segmented multi-modal friction compensation model for iterative training, and obtain the segmented friction compensation result; As a preferred solution, the segmented multi-modal friction compensation model is constructed based on the attention mechanism, the segmented friction compensation feature parameter set is input into the segmented multi-modal friction compensation model for iterative training, and a segmented friction compensation result is obtained, including the following steps: S41, a segmented multi-modal friction compensation model is constructed based on an attention mechanism, and a modal difference penalty term is introduced to optimize and adjust the segmented multi-modal friction compensation model; S42, the segmented friction compensation feature parameter set is divided into a training set and a test set, and a modal difference penalty rule is preset; S43, the training set is sequentially input into the segmented multi-modal friction compensation model after optimization and adjustment, the segmented core friction feature weight in the segmented friction compensation feature parameter set is strengthened through the attention mechanism, and the segmented multi-modal friction compensation model is iteratively updated in combination with the modal difference penalty rule; As a preferred solution, the training set is sequentially input into the segmented multi-modal friction compensation model after optimization and adjustment, the segmented core friction feature weight in the segmented friction compensation feature parameter set is strengthened through the attention mechanism, and the segmented multi-modal friction compensation model is iteratively updated in combination with the modal difference penalty rule, including the following steps: S431, according to the segmented classification label in the robot segmented labeling data set, the segmented friction compensation feature parameter set corresponding to the training set is sequentially input into the segmented multi-modal friction compensation model after optimization and adjustment, and a segmented multi-modal friction compensation parameter is obtained; S432, the importance score of the segmented core friction feature weight is calculated through the attention mechanism, the core weight proportion of the segmented core friction feature weight is adjusted based on the importance score, and the segmented multi-modal friction compensation parameter is adjusted according to the core weight proportion; S433, the modal difference loss between the segmented multi-modal friction compensation parameter after adjustment and the measured friction compensation reference value in the robot joint nonlinear friction data is calculated based on the modal difference penalty rule, and the segmented multi-modal friction compensation model is iteratively updated in combination with the segmented multi-modal friction compensation parameter after adjustment.

[0011] As a preferred solution, the importance score of the segmented core friction feature weight is calculated through the attention mechanism, the core weight proportion of the segmented core friction feature weight is adjusted based on the importance score, and the segmented multi-modal friction compensation parameter is adjusted according to the core weight proportion, including the following steps: S4321, the core friction feature vector in the segmented friction compensation feature parameter set is extracted, and a feature importance scoring model is constructed based on the attention mechanism; S4322, the core friction feature vector is input into the feature importance scoring model to calculate the core importance score of the core friction feature vector, and the core importance score is normalized. S4323, ranking the normalized core importance scores, and adjusting the core weight proportion of the core friction feature vector based on the ranking result to generate a segmented core friction feature weight; S4324, weighting and correcting the segmented multi-modal friction compensation parameters based on the segmented core friction feature weight.

[0012] S44, verifying the performance of the segmented multi-modal friction compensation model updated iteratively by using the test set, outputting the friction compensation amount prediction value, and integrating to form a segmented friction compensation result.

[0013] S5, verifying the segmented friction compensation result, applying the verified segmented friction compensation result to the actual motion control of the robot joint, and collecting the actual motion control parameters of the robot joint in real time to generate a robot joint nonlinear friction compensation report.

[0014] As a preferred scheme, the verification of the segmented friction compensation result, the application of the verified segmented friction compensation result to the actual motion control of the robot joint, and the collection of the actual motion control parameters of the robot joint in real time to generate a robot joint nonlinear friction compensation report include the following steps: S51, presetting a result verification data set, and setting a compensation error threshold and a control accuracy threshold as a verification evaluation standard; S52, comparing and checking the segmented friction compensation result with the verification data set, and calculating the friction compensation error value and the joint motion control accuracy value of each segment; S53, fine-tuning and optimizing the friction compensation error value and the joint motion control accuracy value according to the verification evaluation standard to form a segmented friction compensation scheme; S54, deploying the segmented friction compensation scheme to the robot joint motion control system, and collecting the joint motion control parameters in real time; S55, combining the real-time collected control parameters with the segmented friction compensation scheme to generate a robot joint nonlinear friction compensation report.

[0015] According to another aspect of the present application, a segmented robot joint nonlinear friction compensation system is provided, which comprises a data acquisition labeling module, a data processing matching module, a feature fusion screening module, a model construction calculation module, and a verification application report module. The data acquisition labeling module is used to acquire multi-source robot motion data and robot joint nonlinear friction data under segmented robot joint motion conditions, and to segment and classify the multi-source robot motion data according to the robot joint motion state to form a robot segmented labeled data set. The data processing matching module is used for preprocessing the robot segmented annotation data set, extracting segmented friction feature parameter sets in the robot segmented annotation data set by using a deep learning framework, and performing cross-modal data matching on the segmented friction feature parameter sets and the robot joint nonlinear friction data to obtain segmented friction feature matching results. The feature fusion screening module is used for presetting segmented matching screening rules and segmented feature weight distribution rules, weighting and fusing the segmented friction feature matching results based on the segmented feature weight distribution rules, and screening the weighted and fused segmented friction feature matching results by using the segmented matching screening rules to form segmented friction compensation feature parameter sets. The model construction calculation module is used for constructing a segmented multi-modal friction compensation model based on an attention mechanism, inputting the segmented friction compensation feature parameter sets into the segmented multi-modal friction compensation model for iterative training, and obtaining segmented friction compensation results. The verification application report module is used for verifying the segmented friction compensation results, applying the verified segmented friction compensation results to robot joint actual motion control, and collecting robot joint actual motion control parameters in real time to generate a robot joint nonlinear friction compensation report.

[0016] The beneficial effects of the present application are as follows: 1. The present application classifies and annotates robot joint motion data according to motion states, combines a multi-modal feature extraction network and a cross-modal data matching mechanism, systematically eliminates noise, abnormal and redundant data, realizes accurate integration of standardized motion state data, friction characteristic data and mechanical response data, solves the problem of strong heterogeneity and loose feature correlation of multi-source motion data, improves the integrity and accuracy of the segmented friction feature parameter set, and dynamically generates a feature weight coefficient set by presetting a segmented feature weight distribution rule, replacing the mode of artificially setting fixed weights.

[0017] 2. The present application dynamically strengthens the core friction feature weight by feature importance scoring, calculates the modal difference loss by combining the modal difference penalty rule and iteratively updates the model, suppresses potential conflicts and invalid information interference between modes, improves the robustness of the model in a multi-segmented and multi-modal data environment, ensures that the core friction feature is not covered by redundant information, improves the accuracy of compensation prediction, and maps the segmented fusion features to a unified feature space and completes dimension normalization by modal feature alignment processing, ensures that the weight distribution of different segmented and different modal features is accurately matched with the actual friction influence law, avoids the problem of poor adaptability of a single weight standard, and improves the adaptability of the compensation model to different motion conditions. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description only need to be some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative work on the premise of the accompanying drawings.

[0019] Figure 1 It is a method flow chart of a segmented robot joint nonlinear friction compensation method according to an embodiment of the present application. Figure 2 It is a system block diagram of a segmented robot joint nonlinear friction compensation system according to an embodiment of the present application.

[0020] In the figure: 1, data acquisition and labeling module; 2, data processing and matching module; 3, feature fusion and screening module; 4, model construction and calculation module; 5, verification and application report module. DETAILED DESCRIPTION

[0021] The specific embodiments of the present application will be further described in detail below in combination with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application.

[0022] Therefore, the detailed description of the embodiments of the present application provided in the accompanying drawings below is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those skilled in the art on the basis of the embodiments in the present application without any creative work are within the scope of protection of the present application.

[0023] According to an embodiment of the present application, a segmented robot joint nonlinear friction compensation method and system are provided.

[0024] The present application will be further described in combination with the accompanying drawings and specific embodiments. According to an embodiment of the present application, as shown in Figure 1 The segmented robot joint nonlinear friction compensation method according to an embodiment of the present application includes the following steps: S1, acquiring multi-source robot motion data and robot joint nonlinear friction data under segmented robot joint motion conditions, and segmenting and classifying the multi-source robot motion data according to the robot joint motion state to form a robot segmented labeled data set; Specifically, by robot joint sensor, motion controller and force feedback device, synchronous acquisition of multi-source motion data (including joint angular velocity, displacement, driving torque, etc.) and nonlinear friction data (including static friction threshold, coulomb friction coefficient, viscous friction coefficient, etc.), based on the robot motion state (such as low-speed crawling, uniform speed running, acceleration / deceleration, start / stop, etc.) Set the classification standard of the segment, clear the motion parameter threshold and time interval of each segment, after the preliminary screening of the collected multi-source data, remove the obvious noise and abnormal collection points, according to the set motion state standard, divide and classify the data by time segment, and the annotation information includes segment type, time range, core motion parameter characteristics; Finally, integrate the annotated multi-source data and friction data, establish the corresponding association of the motion state label-multi-modal motion data-nonlinear friction data, and form the robot segment annotation data set.

[0025] S2, preprocessing the robot segment annotation data set, using a deep learning framework to extract the segment friction feature parameter set in the robot segment annotation data set, and performing cross-modal data matching between the segment friction feature parameter set and the robot joint nonlinear friction data to obtain a segment friction feature matching result; As a preferred solution, the preprocessing of the robot segment annotation data set, using a deep learning framework to extract the segment friction feature parameter set in the robot segment annotation data set, and performing cross-modal data matching between the segment friction feature parameter set and the robot joint nonlinear friction data to obtain a segment friction feature matching result includes the following steps: S21, according to the segment classification annotation of the robot segment annotation data set, remove noise data, abnormal collection data and redundant data, and perform time axis alignment correction; Specifically, according to the motion state characteristics of each segment (such as speed threshold, torque range, etc.), use statistical filtering method (such as 3σ criterion) to identify and remove noise data outside the reasonable interval, use outlier detection algorithm (such as isolation forest) to filter abnormal collection data caused by sensor failure and transmission interruption, and based on the semantic association of segment classification annotation, use redundant data removal strategy (such as Pearson correlation coefficient method) to remove repeated data with high correlation in the same segment without additional information gain. According to the time sequence difference of multi-source data (motion, friction, mechanical response data), use the system clock of the robot motion controller as the reference, align the time axis of each modal data through timestamp synchronization algorithm, correct the time sequence deviation caused by collection delay, ensure the accurate matching of multi-source data at the same time node, and form the preprocessed data set.

[0026] S22, based on the deep learning framework, build a segment multi-modal feature extraction network model, and use the segment multi-modal feature extraction network model to extract the standardized motion state data set, friction characteristic data set and mechanical response data set in the robot segment annotation data set; Specifically, first, a three-branch parallel network architecture is built to adapt to the modal characteristics of three types of data, motion state, friction characteristics, and mechanical response. Each branch contains convolutional layers, pooling layers, and batch normalization layers to extract single-modal local and global features. Second, a segmented adaptation module is embedded to optimize the cross-segment feature adaptation by configuring a dedicated feature extraction subnetwork for different motion state segments (e.g., low speed, constant speed, acceleration) based on the segmented classification labels of the dataset. Then, the preprocessed segmented labeled dataset is standardized (e.g., normalized, standardized scaling), and input into the network according to the segmented labels. Finally, the single-modal key features are extracted and fused layer by layer through each branch network, and the structured feature vector is output through the fully connected layer. The standardized motion state dataset (containing core parameter features such as angular velocity and displacement), friction characteristics dataset (containing features such as friction coefficient and threshold), and mechanical response dataset (containing features such as driving torque and contact force) are formed to ensure the feature integrity and modal specificity of the three types of datasets.

[0027] S23, the standardized motion state dataset, friction characteristics dataset, and mechanical response dataset are integrated to obtain a segmented friction feature parameter set. Specifically, the standardized motion state dataset, friction characteristics dataset, and mechanical response dataset are processed for feature dimension unification. The feature scaling (e.g., Min-Max normalization) is used to map the features of different dimensions to the same numerical interval, eliminating the influence of dimension differences. Then, the three types of datasets are grouped according to the motion state segments (e.g., low speed, constant speed, acceleration and deceleration) based on the segmented classification labels to ensure accurate correspondence of multi-modal data within the same segment. Feature concatenation and cross-fusion strategies are used to stack the motion state features (e.g., angular velocity, displacement), friction characteristics features (e.g., friction coefficient, threshold), and mechanical response features (e.g., driving torque, contact force) in the same segment. The attention mechanism is used to strengthen the weight proportion of key features. The fused feature data is structured and stored according to the segmented labels to form a segmented friction feature parameter set containing multi-modal core features, consistent dimensions, and close associations.

[0028] S24, the cosine similarity algorithm is used to calculate the sample matching degree of the segmented friction feature parameter set and the robot joint nonlinear friction data, and a cross-modal correlation mapping table is established based on the sample matching degree for cross-modal data matching.

[0029] Specifically, first, the segmented friction feature parameter set and the robot joint nonlinear friction data are standardized and preprocessed to map both types of data to the same feature space, eliminating the differences in dimensions and numerical ranges. Secondly, the cosine similarity algorithm is used to calculate the similarity value (value range 0-1) of each sample in the segmented friction characteristic parameter set and the nonlinear friction data sample one by one. The higher the similarity, the stronger the sample feature fit degree; then, a similarity threshold (such as 0.7) is set to filter out high matching sample pairs higher than the threshold and eliminate low similarity invalid matches. Finally, a cross-modal correlation mapping table is established based on the filtered effective sample pairs, which clearly records the correspondence between the segmented friction characteristic parameters and the nonlinear friction data, the similarity score, and the feature correlation dimension in each segment. Through this mapping table, accurate correlation matching of different modal data is realized.

[0030] S3, preset segment matching screening rules and segment feature weight allocation rules, based on the segment feature weight allocation rules, the segment friction feature matching results are weighted and fused, and the segment matching screening rules are used to screen the weighted and fused segment friction feature matching results to form a segmented friction compensation feature parameter set; As a preferred scheme, the preset segment matching screening rules and segment feature weight allocation rules, based on the segment feature weight allocation rules, the segment friction feature matching results are weighted and fused, and the segment matching screening rules are used to screen the weighted and fused segment friction feature matching results to form a segmented friction compensation feature parameter set, including the following steps: S31, extract the robot joint friction characteristic parameters of each segment in the robot joint nonlinear friction data, and preset segment matching screening rules and segment feature weight allocation rules; Specifically, for the robot joint nonlinear friction data, according to the divided motion state segments (such as low speed, uniform speed, acceleration and deceleration, etc.), the core friction characteristic parameters of each segment are extracted through feature extraction algorithms (such as statistical feature method, peak detection method), including static friction threshold, coulomb friction coefficient, viscous friction coefficient and friction stiffness, etc. Key indicators, form a segmented friction characteristic parameter subset, and then combine the friction compensation accuracy requirement and the data characteristics, preset segment matching screening rules: clearly set the similarity threshold of high matching samples (such as cosine similarity ≥ 0.75), feature integrity requirement (core parameters have no missing) and data stability standard (eliminate abnormal samples with fluctuation exceeding threshold), and preset segment feature weight allocation rules: according to the influence degree of each modal feature on friction compensation, set weight calculation constraints (such as the sum of weights is 1, the proportion of core feature weight is not less than 50%), and clearly define the weight adjustment logic of different motion segments (such as low speed segment focuses on static friction related feature weight, high speed segment strengthens viscous friction feature weight), providing unified specification for subsequent dynamic weight calculation and feature fusion.

[0031] S32, extract the segmented friction characteristic parameters of each segment from the segmented friction characteristic matching result, and calculate the single-modal characteristic weight coefficient of the segmented friction characteristic parameters based on the segmented characteristic weight distribution rule; As a preferred solution, the extracting the segmented friction characteristic parameters of each segment from the segmented friction characteristic matching result, and calculating the single-modal characteristic weight coefficient of the segmented friction characteristic parameters based on the segmented characteristic weight distribution rule comprises the following steps: S321, extracting the segmented friction characteristic parameters from the segmented friction characteristic matching result according to the segmented classification annotation in the robot segmented annotation dataset; Specifically, the core information of the segmented classification annotation in the robot segmented annotation dataset is determined, including the motion state type of each segment (such as low-speed crawling, uniform-speed running, acceleration and deceleration, start and stop, etc.), time interval boundary and corresponding motion parameter threshold (such as angular velocity range, displacement interval, etc.), which is used as the extraction basis. According to the segmented classification annotation time axis and motion state label, the matching result subsets corresponding to each segment are located one by one by correlating and mapping the segmented friction characteristic matching result. Then, in each segment subset, the characteristic parameters strongly related to the motion state of the segment are selected by combining the pre-set friction characteristic core indicators (such as static friction threshold, Coulomb friction coefficient, viscous friction coefficient, etc.), the redundant matching characteristics and invalid associated data across segments are removed, and the extracted characteristic parameters are classified and arranged according to the segment type, so that the segmented friction characteristic parameters of each segment are accurately corresponding to the corresponding motion state annotation, forming a structured segmented segmented friction characteristic parameter set, and laying a foundation for subsequent weight calculation and feature fusion.

[0032] S322, based on the segmented characteristic weight distribution rule, the weight calculation constraint condition of the segmented friction characteristic parameters is determined; Specifically, first, the pre-set segmented characteristic weight distribution rule is taken as the core basis, and the friction characteristic differences and compensation accuracy requirements of different motion segments (such as low-speed, uniform-speed, acceleration and deceleration) are combined to determine the core constraint direction of weight calculation.

[0033] Secondly, the basic constraint condition is set: the weight sum of all segmented friction characteristic parameters in each segment is 1, which ensures the normalization of weight distribution; the weight proportion of core friction characteristics (such as low-speed segment static friction threshold, high-speed segment viscous friction coefficient) is not less than 50%, which guarantees the influence of key features, and reasonable constraints are added: the weight value range of a single feature is limited to 0.01-0.4, which avoids the fusion deviation caused by too high or too low weight of a certain feature; the weight adjustment range of the same modal feature in different segments does not exceed 30%, which maintains the consistency of weight distribution.

[0034] Finally, the adaptability constraint is supplemented: the weight distribution needs to fit the correlation logic of the motion state and the friction characteristic of each segment, which ensures that the constraint condition meets the rule requirements.

[0035] S323, construct a single modal weight calculation model according to the weight calculation constraint condition, and input the segmented friction characteristic parameters into the model to calculate the segmented single modal characteristic weight coefficient; Specifically, taking the weight calculation constraint condition (the total weight sum is 1, the core feature proportion is greater than or equal to 50%, the single feature weight is 0.01-0.4, etc.) as the core criterion, a single modal weight calculation model is constructed: the analytic hierarchy process is used to determine the feature influence priority, the entropy weight method is used to quantify the feature information gain, and a hybrid weight calculation framework is formed by fusion, the constraint condition is embedded as a model regularization term to ensure that the output weight meets the preset rules, the segmented friction characteristic parameters of each segment are standardized to eliminate the dimensional difference, and then the batch is input into the model according to the segment classification. The model first calculates the subjective weight of the feature (based on the importance ranking of friction compensation) by the analytic hierarchy process, then calculates the objective weight (based on the feature data dispersion) by the entropy weight method, then fuses the subjective and objective weights in a ratio of 6:4, corrects the weight values that exceed the constraint range by regularization processing, and outputs the weight coefficient of each single modal feature in each segment to form a segmented single modal feature weight coefficient set that meets the constraint condition and fits the friction characteristic.

[0036] S324, reasonable verification is performed on the segmented single modal characteristic weight coefficient, and abnormal coefficients exceeding the constraint range are removed to form a single modal characteristic weight coefficient set.

[0037] Specifically, the core basis for weight rationality verification is the preset constraint condition (the total weight sum is 1, the core feature proportion is greater than or equal to 50%, the single feature weight is 0.01-0.4, etc.), a double verification mechanism is built, the first step is numerical boundary verification: each segmented single modal characteristic weight coefficient is checked one by one, and abnormal values less than 0.01 or greater than 0.4 are directly removed, and the core feature weight proportion of less than 50% is marked as a to-be-corrected subset, the second step is logical consistency verification: the total sum of all feature weights in each segment is calculated, and if the error deviates from 1 by more than ±0.05, it is adjusted by normalization algorithm; at the same time, the weight fluctuation amplitude of the same modal feature in different segments is verified, and if it exceeds 30%, it is logically corrected combined with the friction characteristic, and then the weight coefficient after verification and correction is rechecked to ensure that all coefficients meet the constraint condition, and a single modal feature weight coefficient set with standard structure and reliable data is formed according to the segment classification.

[0038] S33, the segmented friction characteristic parameters of each segment are weighted and fused with the single modal characteristic weight coefficient, and modal characteristic alignment processing is performed to obtain a fused feature parameter set; As a preferred scheme, the step of weighting and fusing the segmented friction characteristic parameters of each segment with the single modal characteristic weight coefficient, and performing modal characteristic alignment processing to obtain a fused feature parameter set comprises the following steps: S331, dimension product operation is performed on the segmented friction feature parameters and the single-modal feature weight coefficients; Specifically, the core logic of the dimension product operation is as follows: ensure the dimension accurate matching of the segmented friction feature parameters and the single-modal feature weight coefficients, avoid operation deviation, dimensionally regularize the segmented friction feature parameters of each segment, convert them into feature vectors of a uniform dimension (such as M rows and N columns, M is the sample quantity, and N is the feature dimension), at the same time, arrange the single-modal feature weight coefficients of the corresponding segment into a weight vector (1 row and N columns) consistent with the dimension of the feature vector, and ensure that each feature parameter corresponds to a unique weight coefficient.

[0039] Then, the element-wise product operation is adopted to multiply each element of the separated friction feature vector and the weight vector, that is, the feature value at each position in the feature vector is multiplied by the weight coefficient at the corresponding position in the weight vector, to obtain the weighted feature vector. During the operation, the dimension matching is verified synchronously, and if the dimension does not match, the feature dimension is supplemented or the weight coefficient is reshaped for adjustment, to ensure the operation compliance, and output the weighted separated friction feature vectors of each segment.

[0040] S332, the segmented friction feature parameters after the dimension product operation are accumulated and summed to obtain a segmented fusion feature set, and a segmented fusion feature set alignment mapping model is constructed to map the segmented fusion features in the segmented fusion feature set to a unified feature space; Specifically, first, the segmented friction feature parameters after the dimension product operation (which have been converted into weighted feature vectors of a uniform dimension) are accumulated and summed: according to the segment classification label, the weighted feature vectors of all samples in the same segment are accumulated and summed dimension by dimension to obtain the aggregated feature vector of the segment, and the aggregated feature vectors of all segments are integrated to form the segmented fusion feature set.

[0041] Then, the segmented fusion feature set alignment mapping model is constructed: based on the dimension and distribution characteristics of the segmented fusion features, an adaptive feature mapping algorithm (such as kernel principal component analysis or domain adaptive network) is adopted, the dimension standard and distribution constraint of the unified feature space are introduced, each feature vector in the segmented fusion feature set is input into the model, the mapping relationship of different segmented features is learned, the feature distribution deviation caused by the motion state difference of each segment is eliminated, all the segmented fusion features are uniformly mapped to the preset standardized feature space, and the model parameters are optimized synchronously during the mapping process to minimize the cross-segment feature difference, to ensure that the mapped features not only retain the original core information, but also have dimension consistency and comparability, and form the standardized segmented fusion feature set in the unified feature space.

[0042] S333, the segmented fusion features in the mapped segmented fusion feature set are subjected to dimension normalization processing, and are integrated to form a fusion feature parameter set.

[0043] Specifically, for the segmented fusion features mapped to the unified feature space, a Min-Max normalization algorithm is used for dimension normalization processing: taking the global maximum and minimum values of all segmented fusion features as the reference, the numerical values of each feature dimension are mapped to the [0, 1] interval, the formula is: normalized value = (original feature value-global minimum value) / (global maximum value-global minimum value), which eliminates the numerical range difference between different feature dimensions and ensures the balance of each feature weight.

[0044] And during the processing, abnormal feature points with abnormal fluctuations in normalized values (such as exceeding the range of [0, 1]) are simultaneously removed to ensure data consistency. Then, the segmented fusion features after normalization are structured and integrated according to the segmented classification labels, taking the segmented identification-feature dimension-normalized feature value as the core structure, establishing a feature index association, ensuring that each segmented feature accurately corresponds to the original motion state label, and forming a fusion feature parameter set with unified dimensions, standardized values, and clear associations.

[0045] S34, using a segmented matching screening rule to remove invalid feature data in the fusion feature parameter set, and screening core feature parameters to form a segmented friction compensation feature parameter set.

[0046] Specifically, based on the pre-set segmented matching screening rule, hierarchical screening is carried out on the fusion feature parameter set: in the first step, invalid feature data is removed through similarity threshold checking (retaining features with cosine similarity ≥0.75), integrity checking (removing samples with missing core parameters), and stability checking (removing abnormal data with fluctuations exceeding the pre-set threshold), filtering out invalid information without actual friction compensation value; in the second step, core feature parameters are screened, and the friction characteristics of each segmented motion state (such as low speed, constant speed, acceleration and deceleration) are combined to use feature importance evaluation algorithms (such as random forest and mutual information method) to quantify the contribution of features to compensation accuracy, and the top 60% core features (such as low-speed static friction threshold and high-speed viscous friction coefficient) are preferentially retained.

[0047] And during the screening process, the redundancy between features is simultaneously checked, and highly redundant features with a correlation coefficient ≥0.85 are removed through the Pearson correlation coefficient to ensure the simplicity and effectiveness of the feature set. The screened core features are integrated according to the segmented classification labels to form a segmented friction compensation feature parameter set that accurately adapts to each motion state, has no invalid redundancy, and highlights core information.

[0048] S4, based on the attention mechanism, a segmented multi-modal friction compensation model is constructed, the segmented friction compensation feature parameter set is input into the segmented multi-modal friction compensation model for iterative training, and a segmented friction compensation result is obtained; As a preferred solution, the segmented multi-modal friction compensation model is constructed based on the attention mechanism, the segmented friction compensation feature parameter set is input into the segmented multi-modal friction compensation model for iterative training, and a segmented friction compensation result is obtained, including the following steps: S41, a segmented multi-modal friction compensation model is constructed based on an attention mechanism, and a modal difference penalty term is introduced to optimize and adjust the segmented multi-modal friction compensation model; Specifically, the segmented multi-modal friction compensation model core architecture is built based on the attention mechanism: a three-branch encoder is used to process motion, friction, and mechanical response multi-modal features, a segmented adaptation module is embedded to match different motion state characteristics, a multi-head self-attention layer is integrated at the decoder end to dynamically capture the correlation weight of each modal core feature, and a modal difference penalty term is introduced to optimize the model: based on the distribution distance and semantic consistency of each modal feature, a penalty function is constructed and integrated into the model loss function (such as mean square error loss), when there is a conflict or distribution deviation between different modal features, the loss value is increased through the penalty term, which forces the model to adjust the parameters, and during the training process, the penalty term dynamically quantifies the modal difference loss, which is optimized together with the feature matching loss, and the model's adaptation ability to modal heterogeneity is strengthened, while the attention mechanism is used to strengthen the core feature weight and suppress redundant interference, forming an optimized model that takes into account the multi-modal fusion accuracy and segmented adaptation.

[0049] S42, the segmented friction compensation feature parameter set is divided into a training set and a test set, and a modal difference penalty rule is preset; Specifically, the segmented friction compensation feature parameter set is divided according to the stratified sampling principle: taking each motion segment (such as low speed, constant speed, acceleration and deceleration) as the stratification basis to ensure the consistency of the segmented distribution and feature distribution of the training set and the test set, and the data is split in a 7:3 or 8:2 ratio to avoid data bias, and the data labels are labeled after splitting to clearly indicate the corresponding friction compensation reference value of each sample, ensuring the integrity and usability of the data set.

[0050] The preset modal difference penalty rule is based on the semantic correlation and distribution characteristics of multi-modal features (motion, friction, and mechanical response): when the cosine similarity of different modal features is less than 0.65 or the distribution distance exceeds the preset threshold, the penalty mechanism is triggered; the penalty gradient is clearly defined, the larger the modal difference, the higher the penalty weight (value range 0.1-0.5), and at the same time, the proportion of the penalty term in the total loss is limited to not more than 30% to avoid excessive punishment affecting core feature learning; the rule needs to adapt to the friction characteristics of different segments to ensure that the modal conflicts under different working conditions can be effectively constrained.

[0051] S43, sequentially input the segmented multi-modal friction compensation model adjusted and optimized according to the training set, strengthen the segmented core friction feature weight in the segmented friction compensation feature parameter set through the attention mechanism, and iteratively update the segmented multi-modal friction compensation model combined with the modal difference penalty rule; As a preferred solution, the step of sequentially inputting the segmented multi-modal friction compensation model adjusted and optimized according to the training set, strengthening the segmented core friction feature weight in the segmented friction compensation feature parameter set through the attention mechanism, and iteratively updating the segmented multi-modal friction compensation model combined with the modal difference penalty rule comprises the following steps: S431, according to the segmented classification label in the robot segmented labeled data set, sequentially input the segmented friction compensation feature parameter set corresponding to the training set into the segmented multi-modal friction compensation model adjusted and optimized, and obtain the segmented multi-modal friction compensation parameter; Specifically, the segmented classification label core information in the robot segmented labeled data set is extracted, including the motion state type, time interval and feature identifier of each segment, and the segmented friction compensation feature parameter set corresponding to the training set is grouped and arranged according to this, to ensure that each group of parameters is accurately matched with the corresponding segment label, and each group of feature parameters is sequentially input into the segmented multi-modal friction compensation model adjusted and optimized according to the logical order of the segment type (such as low speed, uniform speed, acceleration and deceleration, etc.), while the model calls the exclusive feature processing subnetwork of the corresponding motion state through the segmented adaptation module, strengthens the core feature weight combined with the attention mechanism, and suppresses invalid interference by using the modal difference penalty term. After the model is operated layer by layer, the feature is fused and the parameter is iterated, the friction compensation amount, feature correlation weight and other key information corresponding to each segment are output, and the segmented multi-modal friction compensation parameter corresponding to the segmented classification label is integrated and formed.

[0052] S432, calculate the importance score of the segmented core friction feature weight through the attention mechanism, adjust the core weight proportion of the segmented core friction feature weight based on the importance score, and adjust the segmented multi-modal friction compensation parameter according to the core weight proportion; Specifically, first, extract the segmented core friction feature vector from the segmented friction compensation feature parameter set, construct a feature importance scoring model based on the attention mechanism, and calculate the attention weight of each core feature through the model as the importance score and perform normalization processing.

[0053] Then, sort the core friction feature weight according to the normalized importance score, clarify the core position of the high-score feature (such as the low-speed static friction threshold and the high-speed viscous friction coefficient), adjust the weight proportion, and ensure that the total proportion of the top 30% high-importance features is not less than 60%, while reducing the weight proportion of low-score redundant features.

[0054] Finally, the adjusted segmented core friction feature weight is weighted and fused with the segmented multi-modal friction compensation parameter, the compensation parameter is corrected according to the new weight proportion, the positive influence of the core feature on the compensation result is strengthened, the invalid feature interference is weakened, the adjusted compensation parameter is more suitable for the friction characteristics of each segment, and the precision of the compensation model is improved.

[0055] S433, based on the modal difference penalty rule, the modal difference loss between the adjusted segmented multi-modal friction compensation parameter and the measured friction compensation reference value in the robot joint nonlinear friction data is calculated, and the segmented multi-modal friction compensation model is iteratively updated combined with the adjusted segmented multi-modal friction compensation parameter.

[0056] As a preferred solution, the importance score of the segmented core friction feature weight is calculated by the attention mechanism, the core weight proportion of the segmented core friction feature weight is adjusted based on the importance score, and the segmented multi-modal friction compensation parameter is adjusted according to the core weight proportion, including the following steps: S4321, extract the core friction feature vector in the segmented friction compensation feature parameter set, and construct a feature importance scoring model based on the attention mechanism; Specifically, from the segmented friction compensation feature parameter set, combined with the friction characteristic logic of each motion segment (such as low speed, constant speed, acceleration and deceleration), through feature contribution degree sorting (based on random forest or mutual information method) and redundancy screening (Pears correlation coefficient < 0.85), the core indicators such as static friction threshold, Coulomb friction coefficient and viscous friction coefficient are extracted, and the segmented core friction feature vector with unified dimension is integrated to ensure that the vector retains key information and has no redundancy. Based on the attention mechanism, a feature importance scoring model is constructed: based on the Transformer architecture, a segmented adaptation layer is embedded to adapt to different motion state feature differences, a multi-head self-attention mechanism is used to capture the dynamic association between features, the attention weight of each core feature is calculated, a fully connected layer is introduced to map the weight to a standardized importance score (value 0-1), and a regularization term is embedded to avoid overfitting. During the model training process, the parameters are optimized to minimize the friction compensation error.

[0057] S4322, input the core friction feature vector into the feature importance scoring model to calculate the core importance score of the core friction feature vector, and normalize the core importance score; Specifically, the extracted segmented core friction feature vectors are preprocessed, including dimension regularization (to ensure matching with the model input dimension) and numerical standardization (to eliminate dimensional differences), to avoid affecting the scoring accuracy. The preprocessed core friction feature vectors are input into the trained feature importance scoring model one by one according to the segmented classification label. The multi-head self-attention mechanism is used to capture the correlation between features, and the original importance score of each feature is calculated. Then, the preliminary scoring results are output through the fully connected layer.

[0058] The Min-Max normalization algorithm is used to process the core importance score. The maximum and minimum values of the original scores of all segmented core features are used as the reference to map each score to the [0, 1] interval, ensuring that the importance scores of different segments and different dimensional features are comparable. After normalization, the score distribution is checked for reasonableness, and abnormal deviating values are removed to form a standardized core importance score set.

[0059] S4323, sort the normalized core importance scores, and adjust the core weight proportion of the core friction feature vector based on the sorting results to generate a segmented core friction feature weight; Specifically, the normalized importance scores (0-1 interval) of the segmented core friction feature vectors are sorted in descending order to determine the division boundary between high importance features (such as low-speed static friction threshold and high-speed viscous friction coefficient) and low-score redundant features. Then, based on the sorting results, the core weight proportion is adjusted: the weight proportion of high importance features is uniformly increased to 60%-70% of the total weight, with the TOP10% feature weight set to 0.3-0.4 and the 10%-30% feature weight set to 0.2-0.3. The low-score feature weight is adjusted in a stepwise manner according to the score proportion, ensuring that the single feature weight is not less than 0.01 and not more than 0.1, and the total weight of all features in each segment is 1. During the adjustment process, the adaptability of weight distribution is checked in combination with the friction characteristic logic of each segment's motion state to avoid conflicts with the actual friction law. The adjusted weight is sorted and arranged according to the segmented classification to generate a segmented core friction feature weight set.

[0060] S4324, based on the segmented core friction feature weight, the segmented multi-modal friction compensation parameters are weighted and adjusted.

[0061] Specifically, first, the correspondence between the segmented core friction feature weight and the segmented multi-modal friction compensation parameter is determined to ensure that the weight and the compensation parameter in the same segment are consistent in dimension and feature matching (such as the static friction weight corresponding to the static friction compensation amount).

[0062] Subsequently, the weighted product correction strategy is adopted, the core friction characteristic weight of each segment is multiplied with the corresponding friction compensation parameter element by element to obtain the weighted segment compensation parameter intermediate value, the influence of the compensation parameter corresponding to the high importance feature is strengthened, the weight normalization constraint is introduced to ensure that the sum of the modified segment compensation parameters remains adaptive to the original compensation parameter total amount, and the numerical imbalance is avoided. During the correction process, the rationality of the correction result is checked in combination with the motion state characteristics of each segment (such as focusing on static friction parameter correction in the low-speed segment and strengthening viscous friction parameter adjustment in the high-speed segment), and the abnormal value deviating from the friction law is fine-tuned again.

[0063] Finally, the modified compensation parameters of all segments are integrated to form an optimized segmented multi-modal friction compensation parameter set that accurately matches the friction characteristics of each segment and highlights the contribution of core features.

[0064] S44, the performance of the iteratively updated segmented multi-modal friction compensation model is verified using the test set, the friction compensation amount prediction value is output, and the segmented friction compensation result is integrated.

[0065] Specifically, the segmented test set is structured and arranged according to the segmented classification labels to ensure that the feature distribution and motion state type of each segment in the test set are consistent with those in the training set, ensuring the fairness of the verification. Then, the segmented friction compensation feature parameters of the test set are input into the iteratively updated segmented multi-modal friction compensation model one by one. The model accurately matches the friction characteristics of each segment through the segmented adaptation module and the attention mechanism, and outputs the friction compensation amount prediction value of each sample. The prediction results are verified using multi-dimensional performance indicators such as mean square error and mean absolute error to evaluate the prediction accuracy and stability of the model. Then, the friction compensation amount prediction values of all test samples are integrated according to the segmented classification labels, and the abnormal results exceeding the reasonable error range are removed to form a segmented friction compensation result set containing the motion state, feature parameters, predicted compensation amount, and performance evaluation indicators of each segment.

[0066] S5, the segmented friction compensation result is verified, and the verified segmented friction compensation result is applied to the actual motion control of the robot joint, and the actual motion control parameters of the robot joint are collected in real time to generate a robot joint nonlinear friction compensation report.

[0067] As a preferred solution, the verification of the segmented friction compensation result, and the application of the verified segmented friction compensation result to the actual motion control of the robot joint, and the real-time collection of the actual motion control parameters of the robot joint to generate a robot joint nonlinear friction compensation report include the following steps: S51, a result verification data set is preset, and a compensation error threshold and a control accuracy threshold are set as verification evaluation criteria; Specifically, the preset result verification dataset: from the actual operation of the robot full working condition data, according to the stratified sampling principle to select the sample covering all motion segments (low speed, uniform speed, acceleration and deceleration, etc.), Ensure that the sample proportion of each segment is consistent with the actual working condition, while containing friction data under different loads and temperatures to avoid single data; The size of the dataset is set to 10%-15% of the total data, independent of the training set and test set, to ensure the objectivity of the verification.

[0068] Set the verification evaluation standard: combined with the demand of friction compensation engineering, preset compensation error threshold, wherein the absolute error threshold ≤0.05N・m (torque compensation scene) or ≤0.1rad / s (speed compensation scene), the relative error threshold ≤5%; Set the control accuracy threshold, require joint positioning error ≤±0.02mm after compensation, speed fluctuation coefficient ≤3%.

[0069] Clear threshold judgment rule: if more than 85% of the samples in the verification data meet the error threshold and control accuracy threshold, the model is qualified; Otherwise, return to optimize the model parameters to ensure that the evaluation standard has practicality and strictness, and provide clear basis for the final verification of the model.

[0070] S52, compare the segmented friction compensation result with the verification dataset, calculate the friction compensation error value and joint motion control accuracy value of each segment; Specifically, first of all, clear the matching dimension of segmented friction compensation result and verification dataset, according to the segmented classification label, the predicted compensation amount and feature parameters in the compensation result are corresponding to the actual friction value and motion parameters in the verification dataset one by one, to ensure accurate comparison of the same segment and the same working condition sample.

[0071] Secondly, calculate the friction compensation error value of each segment: use the absolute error formula (|predicted compensation amount-actual friction value|) and the relative error formula (|predicted compensation amount-actual friction value| / actual friction value×100%), respectively. Calculate the error of each sample, take the average error of the samples in the segment as the overall compensation error of the segment, and calculate the joint motion control accuracy value: based on the joint positioning data and speed data after compensation, calculate the positioning deviation average (≤±0.02mm is qualified) and speed fluctuation coefficient (≤3% is qualified) in the segment, Quantify the control accuracy performance, and finally arrange the error value and control accuracy value according to the segment to form the performance verification report of each segment.

[0072] S53, according to the verification evaluation standard, fine-tune the friction compensation error value and joint motion control accuracy value, and form a segmented friction compensation scheme; Specifically, with the preset compensation error threshold (absolute error ≤0.05 N·m, relative error ≤5%) and control accuracy threshold (positioning error ≤±0.02 mm, speed fluctuation coefficient ≤3%) as the judgment reference, the friction compensation error value and the joint motion control accuracy value of each segment are compared one by one. For the segments that do not meet the standard, if the error is caused by the imbalance of the core feature weight, the proportion of the segment core friction feature weight is adjusted to strengthen the influence of the high importance feature, and if the accuracy is insufficient, the model attention mechanism parameters are optimized to improve the core feature capture ability.

[0073] Meanwhile, in combination with the motion characteristics of each segment (such as fine-tuning of the static friction parameter in the low-speed segment and optimization of the viscous friction compensation logic in the high-speed segment), the friction compensation parameters are corrected in a targeted manner, and after iterative fine-tuning, the indicators are checked again until more than 85% of the samples meet the evaluation standard. The compensation parameters, weight configuration and verification results of each optimized segment are integrated to form a segmented friction compensation scheme that is adapted to all working conditions and accurately matches the friction characteristics of each motion segment, ensuring the practicality and reliability of the scheme.

[0074] S54, deploying the segmented friction compensation scheme to the robot joint motion control system and collecting joint motion control parameters in real time; S55, generating a robot joint nonlinear friction compensation report by combining the real-time collected control parameters with the segmented friction compensation scheme.

[0075] According to another aspect of the present application, as Figure 2 shown, a segmented robot joint nonlinear friction compensation system is provided, which comprises a data acquisition labeling module 1, a data processing matching module 2, a feature fusion screening module 3, a model construction calculation module 4 and a verification application report module 5; The data acquisition labeling module 1 is used to acquire multi-source robot motion data and robot joint nonlinear friction data under segmented robot joint motion working conditions, and to segment and classify the multi-source robot motion data according to the robot joint motion state to form a robot segmented labeled data set. The data processing matching module 2 is used to pre-process the robot segmented labeled data set, extract the segmented friction feature parameter set in the robot segmented labeled data set using a deep learning framework, and perform cross-modal data matching between the segmented friction feature parameter set and the robot joint nonlinear friction data to obtain a segmented friction feature matching result. The feature fusion screening module 3 is used to preset a segmented matching screening rule and a segmented feature weight allocation rule, weight fuse the segmented friction feature matching result based on the segmented feature weight allocation rule, and screen the weighted fused segmented friction feature matching result using the segmented matching screening rule to form a segmented friction compensation feature parameter set. The model construction calculation module 4 is configured to construct a segmented multi-modal friction compensation model based on an attention mechanism, input a segmented friction compensation feature parameter set into the segmented multi-modal friction compensation model for iterative training, and obtain a segmented friction compensation result. The verification application report module 5 is configured to verify the segmented friction compensation result, apply the verified segmented friction compensation result to actual motion control of a robot joint, and collect actual motion control parameters of the robot joint in real time to generate a robot joint nonlinear friction compensation report.

[0076] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A segmented robot joint nonlinear friction compensation method, characterized by, The method comprises the following steps: S1, obtaining multi-source robot motion data and robot joint nonlinear friction data under segmented robot joint motion conditions, and segmenting and classifying the multi-source robot motion data according to the robot joint motion state to form a robot segmented labeled data set; S2, preprocessing the robot segmented labeled data set, extracting segmented friction feature parameter sets in the robot segmented labeled data set by using a deep learning framework, and performing cross-modal data matching on the segmented friction feature parameter sets and the robot joint nonlinear friction data to obtain segmented friction feature matching results; S3, presetting a segmented matching screening rule and a segmented feature weight distribution rule, weighting and fusing the segmented friction feature matching results based on the segmented feature weight distribution rule, and screening the weighted and fused segmented friction feature matching results based on the segmented matching screening rule to form a segmented friction compensation feature parameter set; S4, constructing a segmented multi-modal friction compensation model based on an attention mechanism, inputting the segmented friction compensation feature parameter set into the segmented multi-modal friction compensation model for iterative training, and obtaining a segmented friction compensation result; S5, verifying the segmented friction compensation result, applying the verified segmented friction compensation result to robot joint actual motion control, and collecting robot joint actual motion control parameters in real time to generate a robot joint nonlinear friction compensation report.

2. The method of claim 1, wherein, The preprocessing of the robot segmented labeled data set, the extraction of the segmented friction feature parameter sets in the robot segmented labeled data set by using the deep learning framework, and the cross-modal data matching of the segmented friction feature parameter sets and the robot joint nonlinear friction data to obtain the segmented friction feature matching results comprise the following steps: S21, according to the segmented classification labeling of the robot segmented labeled data set, the noise data, the abnormal acquisition data and the redundant data are removed, and the time axis is aligned and corrected; S22, a segmented multi-modal feature extraction network model is constructed based on a deep learning framework, and the segmented multi-modal feature extraction network model is used to extract a standardized motion state data set, a friction characteristic data set and a mechanical response data set in the robot segmented labeled data set; S23, the standardized motion state data set, the friction characteristic data set and the mechanical response data set are integrated to obtain a segmented friction feature parameter set; S24, the cosine similarity algorithm is used to calculate the sample matching degree of the segmented friction feature parameter set and the robot joint nonlinear friction data, and a cross-modal correlation mapping table is established based on the sample matching degree for cross-modal data matching.

3. The method of claim 1, wherein, The preprocessing of the robot segmented labeled data set, the extraction of the segmented friction feature parameter sets in the robot segmented labeled data set by using the deep learning framework, and the cross-modal data matching of the segmented friction feature parameter sets and the robot joint nonlinear friction data to obtain the segmented friction feature matching results comprise the following steps: S31, extracting the robot joint friction characteristic parameters of each segment in the robot joint nonlinear friction data, and presetting a segmented matching screening rule and a segmented feature weight distribution rule; S32, extract the segmented friction feature parameters of each segment from the segmented friction feature matching result, and calculate the single-modal feature weight coefficient of the segmented friction feature parameters based on the segmented feature weight distribution rule; S33, weight and fuse the segmented friction feature parameters of each segment with the single-modal feature weight coefficient, and perform modal feature alignment processing to obtain a fused feature parameter set; S34, remove invalid feature data in the fused feature parameter set using a segmented matching screening rule, and screen core feature parameters to form a segmented friction compensation feature parameter set.

4. The method of claim 1, wherein, The segmented multi-modal friction compensation model is constructed based on the attention mechanism, and the segmented friction compensation feature parameter set is input into the segmented multi-modal friction compensation model for iterative training to obtain a segmented friction compensation result, including the following steps: S41, construct a segmented multi-modal friction compensation model based on the attention mechanism, and introduce a modal difference penalty term to optimize and adjust the segmented multi-modal friction compensation model; S42, divide the segmented friction compensation feature parameter set into a training set and a test set, and predefine a modal difference penalty rule; S43, input the training set into the segmented multi-modal friction compensation model in turn according to the segmentation, strengthen the segmented core friction feature weight in the segmented friction compensation feature parameter set through the attention mechanism, and iteratively update the segmented multi-modal friction compensation model in combination with the modal difference penalty rule; S44, verify the performance of the iteratively updated segmented multi-modal friction compensation model using the test set, output a friction compensation amount prediction value, and integrate to form a segmented friction compensation result.

5. The method of claim 1, wherein, The segmented friction compensation result is verified, and the verified segmented friction compensation result is applied to actual motion control of a robot joint, and real-time acquisition of robot joint actual motion control parameters generates a robot joint nonlinear friction compensation report, including the following steps: S51, predefine a result verification data set, and set a compensation error threshold and a control accuracy threshold as a verification evaluation standard; S52, compare and verify the segmented friction compensation result with the verification data set, calculate the friction compensation error value and the joint motion control accuracy value of each segment; S53, fine-tune and optimize the friction compensation error value and the joint motion control accuracy value according to the verification evaluation standard to form a segmented friction compensation scheme; S54, deploy the segmented friction compensation scheme to the robot joint motion control system, and real-time acquisition of joint motion control parameters; S55, generate a robot joint nonlinear friction compensation report by combining the real-time acquisition control parameters with the segmented friction compensation scheme.

6. The method of claim 3, wherein, The segmented friction feature parameters of each segment are extracted from the segmented friction feature matching result, and the single-modal feature weight coefficient of the segmented friction feature parameters is calculated based on the segmented feature weight distribution rule, including the following steps: S321, extract the segmented friction feature parameters from the segmented friction feature matching result according to the segment classification label in the robot segmented label data set; S322, based on the segmented feature weight distribution rule, clearly define the weight calculation constraint conditions of the segmented friction feature parameters; S323, construct a single modal weight calculation model according to the weight calculation constraint condition, input the segmented friction characteristic parameters into the model, and calculate the segmented single modal characteristic weight coefficient; S324, reasonable check is carried out on the segmented single modal characteristic weight coefficient, the abnormal coefficient exceeding the constraint range is eliminated, and the single modal characteristic weight coefficient set is formed.

7. The method of claim 3, wherein, The step of weighting and fusing the segmented friction characteristic parameters of each segment with the single modal characteristic weight coefficient, and performing modal characteristic alignment processing to obtain the fused characteristic parameter set comprises the following steps: S331, dimension product operation is performed on the segmented friction characteristic parameters and the single modal characteristic weight coefficient; S332, the segmented friction characteristic parameters after the dimension product operation are accumulated and summed to obtain the segmented fusion characteristic set, and a segmented fusion characteristic set alignment mapping model is constructed to map the segmented fusion characteristics in the segmented fusion characteristic set to a unified characteristic space; S333, the segmented fusion characteristics in the mapped segmented fusion characteristic set are subjected to dimension normalization processing, and are integrated to form the fused characteristic parameter set.

8. The method of claim 4, wherein, The step of inputting the training set into the segmented multi-modal friction compensation model adjusted and optimized in sequence according to the segmentation, strengthening the segmented core friction characteristic weight in the segmented friction compensation feature parameter set through the attention mechanism, and iteratively updating the segmented multi-modal friction compensation model in combination with the modal difference penalty rule comprises the following steps: S431, according to the segmented classification label in the robot segmented labeling data set, the segmented friction compensation feature parameter set corresponding to the training set is input into the segmented multi-modal friction compensation model adjusted and optimized in sequence to obtain the segmented multi-modal friction compensation parameter; S432, the importance score of the segmented core friction characteristic weight is calculated through the attention mechanism, the core weight proportion of the segmented core friction characteristic weight is adjusted based on the importance score, and the segmented multi-modal friction compensation parameter is adjusted according to the core weight proportion; S433, the modal difference loss between the adjusted segmented multi-modal friction compensation parameter and the measured friction compensation benchmark value in the robot joint nonlinear friction data is calculated based on the modal difference penalty rule, and the segmented multi-modal friction compensation model is iteratively updated in combination with the adjusted segmented multi-modal friction compensation parameter.

9. The method of claim 8, wherein, The step of calculating the importance score of the segmented core friction characteristic weight through the attention mechanism, adjusting the core weight proportion of the segmented core friction characteristic weight based on the importance score, and adjusting the segmented multi-modal friction compensation parameter according to the core weight proportion comprises the following steps: S4321, the core friction feature vector in the segmented friction compensation feature parameter set is extracted, and a feature importance scoring model is constructed based on the attention mechanism; S4322, the core friction feature vector is input into the feature importance scoring model to calculate the core importance score of the core friction feature vector, and the core importance score is normalized; S4323, the normalized core importance score is sorted, and the core weight proportion of the core friction feature vector is adjusted based on the sorting result to generate the segmented core friction characteristic weight; S4324, the segmented multi-modal friction compensation parameter is weighted and corrected based on the segmented core friction characteristic weight.

10. A segmented robot joint nonlinear friction compensation system for implementing the segmented robot joint nonlinear friction compensation method of any one of claims 1-9, characterized by, The system comprises a data acquisition and labeling module, a data processing and matching module, a feature fusion and screening module, a model construction and calculation module, and a verification and application report module; The data acquisition and labeling module is used to acquire multi-source robot motion data and robot joint nonlinear friction data under segmented robot joint motion conditions, and to segment and classify label the multi-source robot motion data according to the robot joint motion state to form a robot segmented labeled data set; The data processing and matching module is used to preprocess the robot segmented labeled data set, extract segmented friction feature parameter sets in the robot segmented labeled data set using a deep learning framework, and perform cross-modal data matching between the segmented friction feature parameter sets and the robot joint nonlinear friction data to obtain segmented friction feature matching results; The feature fusion and screening module is used to preset segmented matching and screening rules and segmented feature weight distribution rules, weight fuse the segmented friction feature matching results based on the segmented feature weight distribution rules, and screen the weighted fused segmented friction feature matching results using the segmented matching and screening rules to form segmented friction compensation feature parameter sets; The model construction and calculation module is used to construct a segmented multi-modal friction compensation model based on an attention mechanism, input the segmented friction compensation feature parameter sets into the segmented multi-modal friction compensation model for iterative training, and obtain segmented friction compensation results; The verification and application report module is used to verify the segmented friction compensation results, apply the verified segmented friction compensation results to robot joint actual motion control, and real-time collect robot joint actual motion control parameters to generate a robot joint nonlinear friction compensation report.

Images