Robot joint torque control method based on long short-term memory network

Abstract

This invention relates to the field of robot control technology, and particularly to a robot joint torque control method based on long short-term memory networks (LSM). The method includes acquiring multi-source data and extracting features, forward propagation and thermal drift prediction using a dual-channel LSM network, real-time feedforward torque compensation based on the predicted values, model learning and optimization, and performance monitoring and diagnosis. In this scheme, a dual-channel LSM network prediction model is used to integrate temperature time-series information and joint motion state features to achieve dynamic prediction of thermally induced torque drift. The model employs a dual-channel, separate processing structure for temperature and motion, enabling feature extraction and pattern learning in both the thermal accumulation fundamental effect and motion modulation effect dimensions. This significantly improves the modeling capability for slow-time-varying, nonlinear thermal drift processes. During online operation, the model can correct torque commands in real-time using feedforward compensation, effectively suppressing zero-point torque drift caused by temperature rise.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a robot joint torque control method based on long short-term memory networks. Background Technology

[0002] As industrial robots evolve towards high precision and long-term continuous operation, the stability and accuracy of joint torque control have become key factors determining robot performance. In actual operation, heat is generated within robot joints due to friction and motor losses, causing the temperature of transmission components to gradually rise. This leads to slow, nonlinear drift in the joint output torque, a phenomenon particularly pronounced during long-term operation, severely impacting the robot's positioning accuracy, trajectory tracking performance, and force control stability. Traditional joint torque control methods are mostly based on static or quasi-static models, which struggle to accurately characterize this highly time-varying and nonlinear thermal drift process. Especially under complex and variable workloads and motion conditions, model mismatch becomes a significant issue, hindering the practical application of robots in high-precision work scenarios.

[0003] In existing technologies, compensation for joint thermal drift typically relies on simplified empirical models or filtering methods based on fixed parameters. These methods often struggle to adapt to dynamic thermal processes under different operating conditions and fail to effectively integrate the coupling relationship between temperature evolution over time and real-time motion state. Furthermore, traditional methods lack consideration for the long-term adaptability of the model, failing to keep pace with performance changes in the robot body due to wear, aging, etc., resulting in a gradual decrease in compensation effectiveness over extended operating time. Simultaneously, existing solutions generally lack safety monitoring and closed-loop evaluation mechanisms for the compensation process, making it difficult to guarantee the robustness and maintainability of the system under abnormal operating conditions. Therefore, there is an urgent need for an intelligent joint torque control method capable of predicting and compensating for thermally induced torque drift online in real time, and possessing self-learning and self-monitoring capabilities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a robot joint torque control method based on long short-term memory networks, thereby solving the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Robot joint torque control methods based on long short-term memory networks include:

[0007] S1: Collect multi-source data and extract features: The robot control system starts the multi-source data acquisition process, synchronously reads the raw data of each sensor unit in each control cycle, extracts physical features after preprocessing and calculation, constructs a fused spatiotemporal feature sequence and performs standardization processing to generate a feature matrix;

[0008] S2: Forward propagation and thermal drift prediction of the dual-channel long short-term memory network: The feature matrix is ​​input into the pre-deployed dual-channel long short-term memory network model for forward propagation calculation to obtain the original thermally induced drift torque prediction value, and a first-order low-pass filter is applied to the original thermally induced drift torque prediction value to obtain the smoothed drift prediction value.

[0009] S3: Real-time feedforward torque compensation based on predicted values: The upper-level controller calculates the theoretical expected torque command, subtracts the smoothed drift prediction value from the command to generate the compensated torque command, verifies the compensated torque command by combining the compensation confidence management and safety restriction mechanism, and sends the verified torque command to the joint servo driver.

[0010] S4: Model learning and optimization: Incremental learning process is triggered by periodic and event-driven strategies. High-confidence data samples are extracted to form an incremental dataset. The dual-channel long short-term memory network model is fine-tuned and trained using the elastic weight consolidation algorithm. The performance of the updated model is evaluated on the validation dataset. If the preset conditions are met, the original model parameters are replaced with the updated model parameters.

[0011] S5: Performance Monitoring and Diagnosis: Through an independently running monitoring thread, the torque tracking error, thermal drift prediction value and related sensor data generated during the operation of the robot control system are aggregated, analyzed and tracked over a long period of time. The compensation function switch comparison test is performed regularly to calculate the compensation effectiveness index, system health check is performed, and corresponding response measures are taken and relevant data are stored according to abnormal or performance degradation signs.

[0012] In one possible implementation, in step S1, the sensor unit includes a high-precision digital temperature sensor mounted on the surface of the harmonic reducer housing, a strain gauge joint torque sensor installed between the motor output shaft and the reducer input shaft, and a dual encoder system located at the motor end and the load end respectively; the raw data includes joint housing temperature readings, joint output torque readings, motor-side encoder values, load-side encoder values, and three-phase current values ​​of the motor driver.

[0013] In one possible implementation, in step S1, the preprocessing calculation specifically involves: calculating the load-side joint angle and joint angular velocity based on the motor-side encoder value and the load-side encoder value, and calculating the equivalent amplitude of the motor current based on the three-phase current value.

[0014] In one possible implementation, in step S1, the fused spatiotemporal feature sequence is constructed as follows: the robot control system maintains a fixed-length sequence. A sliding time window, for each historical point within the window. Constructing multidimensional feature vectors The vector is composed of temperature-related feature sub-vectors and motion-related feature sub-vectors. The temperature-related feature sub-vectors include the filtered temperature value at the current moment, the temperature change rate over the past several cycles, and the accumulated running time since the start of this power-on operation. The motion-related feature sub-vectors include the joint angular velocity at the current moment, the equivalent amplitude of the motor current, and the variance of the torque signal.

[0015] In one possible implementation, in step S2, the dual-channel long short-term memory network model includes a first channel and a second channel. The first channel receives and processes a subsequence composed of the temperature-related feature sub-vectors in chronological order, and the second channel receives and processes a subsequence composed of the motion-related feature sub-vectors in chronological order. The deep output features of the two channels interact and integrate in the fusion layer of the network, and finally output the original thermally induced drift torque prediction value through a fully connected regression layer.

[0016] In one possible implementation, in step S3, the compensation confidence management and safety constraint mechanism specifically involves: the robot control system calculating the current prediction confidence index in real time. ,like If the value exceeds the preset safety threshold, a full compensation strategy will be adopted; if... If the temperature drops below the safety threshold, a degradation strategy is initiated, introducing a time-varying compensation gain. This makes the actual compensation amount become ,in The smoothed drift prediction value is used to record the event and alert the operator.

[0017] In one possible implementation, in step S4, the periodic strategy refers to automatically starting the incremental learning process after the robot has run continuously for 24 hours; the event-driven strategy refers to immediately starting the incremental learning process when the system continuously monitors that the prediction confidence is lower than the threshold and the actual torque tracking error is significantly higher than the historical baseline; each sample in the incremental dataset includes a historical feature matrix and its corresponding target value label, and the target value label is estimated by a background calibration algorithm within a specific time window when the robot performs low-speed, uniform, and unloaded motion.

[0018] In one possible implementation, in step S5, the compensation effectiveness index is calculated by comparing the torque tracking error variance between the compensation activation phase and the compensation deactivation phase; the response measures include recording warning logs, displaying warning information on the operation interface, or switching the control system to a safe mode when a serious fault is diagnosed.

[0019] Beneficial effects compared to existing technologies:

[0020] 1. In this scheme, a dual-channel long short-term memory network prediction model is used to integrate temperature time-series information and joint motion state characteristics to achieve dynamic prediction of thermally induced torque drift. The model employs a dual-channel structure for separate processing of temperature and motion, enabling feature extraction and pattern learning on both the fundamental effect of thermal accumulation and the motion modulation effect, significantly improving its modeling capability for slow-time-varying, nonlinear thermal drift processes. During online operation, the model can correct torque commands in real time using feedforward compensation, effectively suppressing zero-point torque drift caused by temperature rise, thus maintaining high accuracy and long-term stability of torque output throughout the entire process from a cold state to thermal equilibrium.

[0021] 2. In this scheme, an incremental learning mechanism based on an elastic weight consolidation algorithm is introduced, enabling the thermal drift prediction model to continuously self-optimize and adapt during the robot's long-term operation. This mechanism uses recent high-confidence data to fine-tune the model, while regularization constraints protect important historical knowledge from being forgotten. Furthermore, the model undergoes performance validation before updates to ensure that accuracy improves on new data without significant degradation in historical performance. This significantly improves the long-term reliability and environmental adaptability of the compensation method.

[0022] 3. In this solution, a self-contained performance monitoring and diagnostic thread is constructed to achieve real-time evaluation and health management of the thermal drift compensation system's performance. This thread periodically performs comparative tests on and off of the compensation function, quantifies and calculates the compensation effectiveness index, and continuously checks the rationality of sensor data and detects anomalies in the prediction curve. This enables early identification of model failure or hardware malfunction signs, and allows for tiered responses based on preset strategies to ensure system operational safety. Attached Figure Description

[0023] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0024] Figure 1 This is a schematic diagram of the operation process of the robot joint torque control method of the present invention. Detailed Implementation

[0025] Preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention can also be implemented in various different forms, and therefore the present invention is not limited to the embodiments described below. In addition, for the purpose of more clearly describing the present invention, parts not connected to the invention will be omitted from the drawings.

[0026] The technical solutions in this application are designed to address the problems described in the background, and are generally as follows:

[0027] Example

[0028] This embodiment introduces a robot joint torque control method based on a long short-term memory network. Furthermore, it provides an online real-time compensation method for the physical problem of slow, nonlinear zero-point drift in joint output torque caused by the increased temperature of transmission components due to internal frictional heat generation and motor losses during long-term continuous operation of industrial robot joints. The specific implementation steps are as follows:

[0029] S1: Collect multi-source data and extract features

[0030] The robot control system first initiates a high-precision, highly synchronized multi-source data acquisition process. On each robot joint requiring thermal drift compensation, the following sensor units are integrated: a high-precision digital temperature sensor mounted on the surface of the harmonic reducer housing to directly measure real-time temperature changes of key components in the joint transmission system; a strain gauge joint torque sensor installed between the motor output shaft and the reducer input shaft to measure the actual output torque of the joint; and a dual encoder system located at the motor end and the load end respectively to accurately calculate the absolute position, velocity, and elastic deformation angle of the joint due to the limited torsional stiffness of the transmission chain.

[0031] In each control cycle, typically 1 ms, the system synchronously reads raw data from each sensor. This data includes joint housing temperature readings. Joint output torque reading Motor-side encoder value Load-side encoder value and the three-phase current value of the motor driver. , , .

[0032] Subsequently, the system performs preprocessing calculations to extract physical features that can be used as model input. Specifically, based on the motor-side encoder values, the load-side joint angle is calculated. and joint angular velocity Joint angular velocity It is obtained by calculating the angle using the first-order difference, i.e. ,in, For the current control cycle (the 1st cycle) The joint angle value (per cycle). For the previous control cycle (the first) The joint angle value (per cycle). The sampling period is [value]. Simultaneously, the equivalent amplitude of the motor current is calculated. This value can indirectly reflect the heating status of the motor windings. Next, the original temperature signal... and torque signal A low-pass filter with a cutoff frequency of 10Hz is applied to suppress high-frequency measurement noise while fully preserving the low-frequency components related to the thermal dynamics process.

[0033] After preprocessing, the system constructs a fused spatiotemporal feature sequence for input to the Long Short-Term Memory (LSTM) network model. The system maintains a fixed-length sequence. The sliding time window continuously covers the timeline from the current moment backwards. Historical data for each control cycle. For each historical point in this window. The system constructs a multidimensional feature vector. This vector is composed of two physically related but distinct feature sub-vectors. The first sub-vector focuses on temperature-related features, containing the filtered temperature value at the current moment. Rate of temperature change over the past several periods And the cumulative operating time since this power-on operation. The second sub-vector focuses on motion-related features, including the joint angular velocity at the current moment. Motor current equivalent amplitude And a statistic reflecting the intensity of motion within that short time window, namely the variance of the torque signal. Before inputting the feature vectors into the model, all features undergo standardization based on long-term historical statistical data to ensure they conform to a distribution with a mean of zero and a standard deviation of one. Ultimately, the system generates a feature vector with dimensions of [missing value] in each control cycle. Feature matrix ,in, It is the total dimension of the feature vector. This matrix not only contains the recent temperature history of the joint, but also incorporates the corresponding motion state context.

[0034] S2: Forward Propagation and Thermal Drift Prediction in Dual-Channel Long Short-Term Memory Networks

[0035] When the latest feature matrix Once ready, the system immediately feeds the data into a pre-deployed and loaded dual-channel long short-term memory network model for forward propagation calculations. The goal is to predict the joint torque drift caused purely by thermal effects within the current control cycle. This predictive model network structure is specifically designed to model the slow time-varying, highly nonlinear, and motion history-sensitive characteristics of thermally induced drift.

[0036] The model employs a unique dual-channel input architecture. The first channel specifically receives and processes the temperature-related feature subsequence extracted in step S1, while the second channel specifically processes the motion-related feature subsequence. This physically meaningful channel separation design allows the network to more effectively learn patterns in two different feature spaces: the temperature channel primarily learns the fundamental and gradual impact of the joint temperature rise process itself and its historical cumulative effects on the drift amount; the motion channel learns the dynamic modulation effects of different speeds, loads, and motion patterns on the heat generation rate and heat conduction process. The deep output features of the two channels interact and integrate in the network's fusion layer, ultimately outputting a single predicted thermally induced drift torque value through a fully connected regression layer. .

[0037] Furthermore, each channel is composed of two stacked layers of long short-term memory units. The final output layer of the network is a fully connected layer that fuses the high-level feature representations extracted from the two channels at the last time step and maps them to a scalar output. Specifically, let the final hidden state vector of the temperature channel output be... The final hidden state vector output by the motion channel is Concatenate these two vectors to obtain the fused feature vector. Subsequently, the predicted thermally induced drift torque value was calculated using a fully connected layer. The calculation process is expressed as follows:

[0038] ;

[0039] in, It is the weight matrix of the output layer. This is the bias term. The model's parameters were trained offline using a large-scale dataset collected over extended periods of operation by the robot under various typical load and speed curves. The training objective was to minimize the mean square error between the model's predicted values ​​and the actual thermally induced drift separated through precise calibration experiments and signal processing algorithms. The trained model parameters were permanently stored in the controller's non-volatile memory for real-time access during the online phase.

[0040] During online operation, the forward propagation calculation is efficiently completed within a 1 ms control cycle. To suppress high-frequency jitter in the predicted values ​​that may be caused by input noise and to ensure the smoothness of the compensation signal, the system applies a first-order low-pass filter to the network's raw output. The filtered drift predicted value... The calculation formula is as follows:

[0041] ;

[0042] in, Index for the current control cycle, This is the filtered drift estimate for the current control cycle. The current control cycle is represented by the raw drift prediction value directly output by the Long Short-Term Memory network model. This is the drift estimate after filtering from the previous control cycle. These are the filter coefficients, with values ​​between 0 and 1. A small value, such as 0.1, is typically chosen to achieve effective smoothing without excessively introducing phase lag. This is the filtered predicted value. It will be used for subsequent real-time torque compensation without introducing high-frequency interference to the control system.

[0043] S3: Real-time feedforward torque compensation based on predicted values

[0044] This step applies the predicted thermal drift to the joint's torque control loop for real-time compensation. The compensation is performed in a feedforward manner. First, the robot's upper-level controller calculates the theoretically expected torque command for the joint in the current cycle based on trajectory planning or force control algorithms. .

[0045] Subsequently, the compensation module performs the core operation: subtracting the predicted thermal drift from the theoretically expected torque command to generate the final compensated torque command sent to the joint servo driver. This operation can be represented as Its physical significance lies in actively offsetting the systematic deviation component introduced into the output torque due to changes in the internal temperature field of the joint. This makes the compensated torque command sent to the servo driver... It can more accurately correspond to the expected net output torque, thereby suppressing the decrease in torque control accuracy caused by temperature rise from the source.

[0046] To ensure the robustness and safety of the system under various operating conditions, this embodiment also integrates a set of compensated confidence management and safety constraint mechanisms. The system calculates the current prediction confidence index in real time. This metric comprehensively evaluates the similarity between the current input features and the distribution of the model training data, as well as whether the rate of change of the predicted value over a short period of time is reasonable. If the confidence level... If the value exceeds the preset safety threshold, the system will fully employ the aforementioned compensation strategy. If the confidence level... If the temperature drops below the safety threshold, the system determines that the current operating condition may exceed the model's training range or that the sensor data is abnormal. At this point, the system will initiate a degradation strategy, gradually reducing the weight of the compensation effect; for example, introducing a time-varying compensation gain. (Gradually decreasing from 1 to 0), making the actual compensation amount become Simultaneously, the event is recorded and the operator is alerted. This mechanism effectively avoids injecting erroneous commands into the control system due to occasional misjudgments by the model, ensuring the absolute stability of the entire robot system. Finally, the torque command, after thermal drift compensation and safety verification, is... The command is sent to the joint servo driver, which operates in current loop mode to precisely track the command, thereby achieving high-precision torque output.

[0047] S4: Model Learning and Optimization

[0048] To enable the thermal drift prediction model to adapt to slow changes in robot performance over the long term, such as alterations in frictional characteristics due to mechanical wear, this embodiment also includes a background-running incremental learning and parameter update mechanism. This mechanism runs periodically with a low priority, without affecting the determinism of the real-time control thread.

[0049] The incremental learning process combines periodic and event-driven strategies. Periodic triggering, for example, automatically initiates a learning cycle after the robot has run continuously for 24 hours. Event-driven triggering is linked to the confidence management mechanism in step S3. When the system continuously detects low prediction confidence and a significantly higher actual torque tracking error than the historical baseline, it determines that the current model may have mismatched with the actual physical process, thus immediately triggering the learning cycle. After triggering, the system extracts a batch of recently collected high-confidence data samples from the cache, forming an incremental dataset. Each sample contains a historical feature matrix. and their corresponding target value labels This label value, the "true" thermal drift, is not directly measured but estimated through a background calibration algorithm. Specifically, within a specific time window of the robot performing low-speed, uniform, unloaded motion, its output torque mainly comprises frictional torque and thermal drift torque. Using a basic frictional model, an approximate thermal drift is separated from the measured torque and used as a label for supervised learning.

[0050] Subsequently, the system uses the parameters of the currently running LSTM model as the initialization starting point, and then initializes the incremental dataset. Fine-tuning training is performed with a limited number of iterations. To prevent the model from rapidly forgetting previously learned knowledge while learning new knowledge, an elastic weight consolidation algorithm is used during training. This algorithm penalizes excessive modifications to important network parameters by adding a smart regularization term to the loss function. The importance of parameters is determined by approximately diagonalizing the Fisher information matrix of each parameter from historical training data. To measure. The overall loss function of incremental learning. The structure is as follows:

[0051] ;

[0052] in, This represents the mean squared error loss of the model on incremental new data; These are the current parameters to be optimized in the model; This is a snapshot of the parameters saved before the start of this incremental learning process. This is a key hyperparameter used to balance learning new knowledge with retaining old knowledge. By optimizing this composite loss function, the model can absorb new experiences and adapt to new changes while maximally maintaining its original, generalized predictive ability.

[0053] After fine-tuning the training, the system rigorously evaluates the overall performance of the updated model on an independent and representative validation dataset. This validation set includes some historical typical data and recent data. The evaluation criteria are: the updated model should show a significant improvement in prediction accuracy on the new data, while its performance degradation on historical data must be controlled within a pre-defined, extremely small acceptable range (e.g., the mean squared error increase does not exceed 5%). Only when both conditions are met will the system safely replace the original parameters with the updated model parameters, completing a safe online self-evolution. If the evaluation fails, the update is automatically abandoned, the system reverts to the original model, and hyperparameters such as the learning rate may be adjusted for future attempts. This closed-loop mechanism endows the compensation system with continuous self-optimization and adaptation capabilities throughout the entire product lifecycle.

[0054] S5: Performance Monitoring and Diagnostics

[0055] An independent system monitoring and evaluation thread operates continuously throughout the robot's operation. This thread runs at a low frequency (e.g., 10 Hz) to aggregate, analyze, and track massive amounts of torque tracking errors, thermal drift predictions, and related sensor data generated during system operation. This allows for a macroscopic assessment of the overall effectiveness of the thermal drift compensation system and monitoring of the health status of its various components.

[0056] The core of performance evaluation lies in objectively quantifying the actual engineering benefits brought about by compensation measures. The system periodically (e.g., hourly) or under operator instruction, under absolutely safe operating conditions, performs a brief, automated on / off comparison test of the compensation function. During the short time window when the compensation function is temporarily disabled, the system records the tracking error sequence of joint torques. During the period when the compensation function is normally enabled, the system synchronously records the corresponding torque tracking error sequence. By calculating and comparing the variance of the torque tracking error in these two stages, an intuitive index of compensation effectiveness can be defined. :

[0057] ;

[0058] in, To compensate for the variance of torque tracking error during the start-up phase; To compensate for the variance of torque tracking error during the closing phase. This quantitatively reflects the contribution of the thermally induced drift compensation system to the accuracy of the lifting torque control; the closer the value is to 1, the better the compensation effect. The system is continuously plotted and tracked. The curve showing how the value changes over time, temperature, or cumulative operating time is a direct indicator of whether the performance of the compensation system has degraded over time or requires maintenance.

[0059] In addition, the monitoring thread performs multi-dimensional health checks. It checks the reasonableness and consistency of data from various sensors, such as verifying whether temperature readings are within a physically possible range and whether torque sensor readings and motor current commands show a reasonable proportional relationship. Simultaneously, it analyzes thermal drift predictions. The time series curves can be used to detect whether there are unexpected sharp jumps, long-term saturation, or significant disconnection from temperature changes. These may be early signs of model failure or sensor malfunction.

[0060] Once any anomalies or signs of performance degradation are detected, the monitoring thread responds according to a preset policy level. Response measures include logging warnings, displaying alerts on the user interface, or proactively switching the control system to a safe mode when a serious fault is diagnosed. All operational data, performance metrics, and event logs are persistently stored, forming a complete digital archive. These archives can be used to trace the root cause of problems, optimize system parameters, and provide crucial data for predictive maintenance of the robot, such as inferring the wear condition of transmission components by analyzing slow changes in the long-term baseline of thermal drift.

[0061] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A robot joint torque control method based on long short-term memory networks, characterized in that, include: S1: Collect multi-source data and extract features: The robot control system starts the multi-source data acquisition process, synchronously reads the raw data of each sensor unit in each control cycle, extracts physical features after preprocessing and calculation, constructs a fused spatiotemporal feature sequence and performs standardization processing to generate a feature matrix; S2: Forward propagation and thermal drift prediction of the dual-channel long short-term memory network: The feature matrix is ​​input into the pre-deployed dual-channel long short-term memory network model for forward propagation calculation to obtain the original thermally induced drift torque prediction value, and a first-order low-pass filter is applied to the original thermally induced drift torque prediction value to obtain the smoothed drift prediction value. S3: Real-time feedforward torque compensation based on drift prediction: The upper-level controller calculates the theoretical expected torque command, subtracts the smoothed drift prediction value from the command to generate the compensated torque command, verifies the compensated torque command by combining the compensation confidence management and safety restriction mechanism, and sends the verified torque command to the joint servo driver. The compensation confidence management and safety constraint mechanism specifically involves the robot control system calculating the current prediction confidence index in real time. ,like If the value exceeds the preset safety threshold, a full compensation strategy will be adopted; if... If the temperature drops below the safety threshold, a degradation strategy is initiated, introducing a time-varying compensation gain. This makes the actual compensation amount become ,in The smoothed drift prediction value is used as the basis for recording the event and alerting the operator. S4: Model learning and optimization: Incremental learning process is triggered by periodic and event-driven strategies. High-confidence data samples are extracted to form an incremental dataset. The dual-channel long short-term memory network model is fine-tuned and trained using the elastic weight consolidation algorithm. The performance of the updated model is evaluated on the validation dataset. If the preset conditions are met, the original model parameters are replaced with the updated model parameters. The periodic strategy refers to automatically starting the incremental learning process after the robot has run continuously for 24 hours; the event-driven strategy refers to immediately starting the incremental learning process when the system continuously monitors that the prediction confidence is lower than the threshold and the actual torque tracking error is significantly higher than the historical baseline; each sample in the incremental dataset includes a historical feature matrix and its corresponding target value label, and the target value label is estimated by a background calibration algorithm within a specific time window when the robot performs low-speed, uniform, and unloaded motion. S5: Performance Monitoring and Diagnosis: Through an independently running monitoring thread, the torque tracking error, thermal drift prediction value and related sensor data generated during the operation of the robot control system are aggregated, analyzed and tracked over a long period of time. The compensation function switch comparison test is performed regularly to calculate the compensation effectiveness index, system health check is performed, and corresponding response measures are taken and relevant data are stored according to abnormal or performance degradation signs.

2. The robot joint torque control method based on long short-term memory network as described in claim 1, characterized in that, In step S1, the sensor unit includes a high-precision digital temperature sensor mounted on the surface of the harmonic reducer housing, a strain gauge joint torque sensor installed between the motor output shaft and the reducer input shaft, and a dual encoder system located at the motor end and the load end respectively; the raw data includes joint housing temperature readings, joint output torque readings, motor-side encoder values, load-side encoder values, and three-phase current values ​​of the motor driver.

3. The robot joint torque control method based on long short-term memory network as described in claim 2, characterized in that, In step S1, the preprocessing calculation specifically involves: calculating the load-side joint angle and joint angular velocity based on the motor-side encoder value and the load-side encoder value, and calculating the equivalent amplitude of the motor current based on the three-phase current value.

4. The robot joint torque control method based on long short-term memory network as described in claim 3, characterized in that, In step S1, the fused spatiotemporal feature sequence is constructed as follows: the robot control system maintains a fixed length of... A sliding time window, for each historical point within the window. Constructing multidimensional feature vectors The vector is composed of temperature-related feature sub-vectors and motion-related feature sub-vectors. The temperature-related feature sub-vectors include the filtered temperature value at the current moment, the temperature change rate over the past several cycles, and the accumulated running time since the start of this power-on operation. The motion-related feature sub-vectors include the joint angular velocity at the current moment, the equivalent amplitude of the motor current, and the variance of the torque signal.

5. The robot joint torque control method based on long short-term memory network as described in claim 4, characterized in that, In step S2, the dual-channel long short-term memory network model includes a first channel and a second channel. The first channel receives and processes a subsequence composed of temperature-related feature sub-vectors in chronological order, and the second channel receives and processes a subsequence composed of motion-related feature sub-vectors in chronological order. The deep output features of the two channels interact and integrate in the fusion layer of the network, and finally output the original thermally induced drift torque prediction value through a fully connected regression layer.

6. The robot joint torque control method based on long short-term memory network as described in claim 5, characterized in that, In step S5, the compensation effectiveness index is calculated by comparing the torque tracking error variance between the compensation activation phase and the compensation deactivation phase; the response measures include recording warning logs, displaying warning information on the operation interface, or switching the control system to a safe mode when a serious fault is diagnosed.

Images