Machine learning based robot control system parameter optimization training system

Abstract

The application discloses a robot control system parameter optimization training system based on machine learning, and relates to the technical field of intelligent control.The system comprises a data acquisition and structured labeling module, a coupling influence strength evaluation module, a deviation mapping model construction module, a neural network training module, a real-time prediction and trajectory correction module and a closed-loop control execution module, real-time joint data is acquired, a lightweight machine learning model is constructed and trained to predict trajectory deviation caused by dynamic coupling, the target trajectory is dynamically corrected according to the trajectory deviation, and closed-loop online compensation is realized by a bottom controller, the whole process from data acquisition, prediction to trajectory correction and bottom control execution is continuously executed, so that closed-loop online compensation of joint dynamic coupling effect is realized, and the problems of trajectory tracking and dynamic response precision caused by dynamic coupling of a multi-joint robot are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a robot control system parameter optimization training system based on machine learning. Background Technology

[0002] In the fields of modern industrial automation and advanced manufacturing, multi-joint robot systems have become core productivity tools, with applications spanning multiple key areas such as industrial robotic arms, high-precision bionic robots, automated assembly lines, and human-robot collaboration. These systems, with their multiple degrees of freedom, high flexibility, and powerful operational capabilities, significantly improve production efficiency and operational accuracy.

[0003] In multi-joint robot systems, when one joint suddenly accelerates or decelerates, the resulting inertial torque immediately acts on the next joint connected to it, causing unpredictable positional shifts or vibrations. Traditional feedback controllers can only correct for deviations after they are detected, thus failing to provide proactive and predictive compensation for such instantaneous disturbances caused by dynamic coupling. This limits the performance ceiling and robustness of robot systems in high-speed, high-precision tasks. This inherent bottleneck—the difficulty of balancing real-time performance, accuracy, and robustness in traditional control methods, especially in handling the complex, time-varying, and difficult-to-model dynamic coupling effects of multi-joint systems—has become a key constraint on improving robot performance. Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based parameter optimization training system for robot control systems to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a robot control system parameter optimization training system based on machine learning.

[0006] The system includes: a data acquisition and structured labeling module, a coupling effect strength assessment module, a deviation mapping model construction module, a neural network training module, a real-time prediction and trajectory correction module, and a closed-loop control execution module.

[0007] Furthermore, the data acquisition and structured labeling module is connected to the coupling influence intensity assessment module and the deviation mapping model construction module; the coupling influence intensity assessment module is connected to the deviation mapping model construction module; the deviation mapping model construction module is connected to the neural network training module; the neural network training module is connected to the real-time prediction and trajectory correction module; and the real-time prediction and trajectory correction module is connected to the closed-loop control execution module.

[0008] Furthermore, the coupling influence strength assessment module includes: a joint pair management unit, a preset deviation management unit, a deviation value assessment unit, and a key coupling screening unit;

[0009] The joint pair management unit is used to collect the motion state parameters of the i-th joint of the robot at time t, record the motion state parameters of the joint according to the coupling path of the joint, and combine the motion state parameters of adjacent joints into adjacent joint data pairs. The motion state parameters include angular acceleration and driving current.

[0010] A preset deviation management unit is used to calculate the deviation δi+1 between the actual position of the (i+1)th joint and the preset trajectory position;

[0011] The deviation evaluation unit is used to calculate the absolute value of the ratio of the deviation Δδi+1 between the actual position and the preset trajectory position of the (i+1)th joint per unit time to the angular acceleration Δai of the i-th joint per unit time. The absolute value is recorded as the evaluation index of the joint coupling influence intensity of the node pair (i, i+1) composed of the i-th joint and the (i+1)-th joint. , ;

[0012] Before formal model training, the evaluation metrics For all adjacent joint pairs of the robot Calculations are performed. The calculation process involves systematically stimulating the first... The joint generates a preset angular acceleration profile, while simultaneously monitoring the first... The deviation between the actual position and the ideal position of a joint is due to the dynamic coupling between adjacent joints of a robot. Essentially, it is the change in the position deviation of the (i+1)th joint when the angular acceleration of the i-th joint changes by 1 unit. This directly reflects the sensitivity of the coupling effect. When the angular acceleration changes slightly, the coupling effect strength index can be calculated using the finite difference method.

[0013] The key coupling screening unit is used to obtain the joint coupling influence intensity evaluation index of all node pairs formed by adjacent joints, sort them from high to low, and take the top K% of node pairs as key coupling paths, where K∈(0,100).

[0014] Furthermore, the deviation mapping model construction module includes: an angle management unit, an angle deviation management unit, a data smoothing unit, and a label input unit;

[0015] The angle management unit is used to collect angle information at the robot joints, where the actual angle of the (i+1)th joint at time t+Δt is denoted as... The target angle of the (i+1)th joint at the corresponding time in the preset trajectory is denoted as . ;

[0016] The angle deviation management unit is used to output a label as the deviation between the actual position of the joint and the preset trajectory position in the next sampling period, where the deviation of the (i+1)th joint is... The calculation formula is: , where Δt represents the length of time per unit time;

[0017] The data smoothing unit is used to smooth the joint deviation by introducing a sliding window filtering mechanism. A window of W unit time is set, and the smoothed deviation value is: , The calculation formula is: Where wm represents the weight, the initial weight value is set to ω1, when m=0, wm=ω1, and the weight is decreased in the form of the (m+1)th weight being half of the mth weight value, ω1∈(0,1).

[0018] The label input unit is used to take the smoothed deviation value as the true label of the neural network.

[0019] Furthermore, in the neural network training module, the neural network structure includes an input layer with 2 nodes, a first hidden layer with 16 nodes, a second hidden layer with 8 nodes, and an output layer with 1 node.

[0020] The input layer nodes take the angular acceleration ai(t) at time t of the i-th node as the first input feature, and the driving current I at time t of the i-th node as the second input feature. i (t) is used as the second input feature;

[0021] The output layer output is used as the prediction bias for the i-th joint at time t+Δt. ;

[0022] During training, the mean squared error is used as the loss function L. The formula for calculating the loss function L is: Where N represents the number of samples in the batch, and These represent the true deviation and the predicted deviation of the k-th sample in the batch, respectively.

[0023] Furthermore, the real-time prediction and trajectory correction module inputs the angular acceleration and driving current of the i-th joint at the current moment into the trained neural network model to predict the position deviation of the (i+1)-th joint after one unit of time.

[0024] Get the current time ct, and the time ct+Δt one unit time later than the current time. Get the position deviation of the (i+1)th joint at time ct+Δt. The set value of the angle of the (i+1)th joint at time ct+Δt. ;

[0025] The target trajectory of the (i+1)th joint is dynamically corrected based on the predicted value. The dynamic correction includes calculating the dynamic correction value of the angle of the (i+1)th joint at time ct+Δt. , The calculation formula is:

[0026] , where k represents the compensation gain coefficient, k∈(0.8,1.2).

[0027] Compensation gain coefficient It can be fine-tuned according to the needs of actual application scenarios to balance the intensity of compensation and the stability of the system. In order to avoid ineffective compensation for minor disturbances and thus avoid introducing additional vibrations or instabilities, the target trajectory correction mechanism can be set to trigger only when the absolute value of the prediction deviation is greater than a preset threshold.

[0028] Furthermore, the closed-loop control execution module continuously and cyclically executes the data acquisition and trajectory correction process to achieve closed-loop online compensation;

[0029] Closed-loop online compensation includes: obtaining the correction error e(t) of the (i+1)th joint at time t, where e(t) is calculated using the following formula: ;

[0030] The control output ui+1(t) applied to the (i+1)th joint at time t is established using a PID control structure. The formula for calculating ui+1(t) is as follows: , where K p K represents the proportional gain. i K represents the integral gain. d This represents the differential gain.

[0031] Indicates at time The control output applied to the (i+1)th joint, such as a torque or current command. This represents the control error at the current moment. To further improve the adaptive capability and dynamic response performance of the control system, the proportional gain... Integral gain and differential gain The value is not fixed, but adaptively adjusted based on the instantaneous second derivative amplitude of the corrected trajectory, i.e., the corrected target angular acceleration. The specific adjustment strategy is as follows: when the corrected target angular acceleration exceeds a threshold, it indicates that the robot is in a high-speed or high-acceleration motion state; at this time, the proportional gain is increased. Increasing by one unit percentage, for example, by 20%, and increasing the differential gain Increasing the integral gain by one unit percentage, such as 15%, enhances the system's rapid response to trajectory deviations and its damping characteristics. This remains unchanged to ensure that excessive integral saturation is not introduced during high-speed motion.

[0032] Furthermore, the closed-loop control execution module, during the execution of the PID control structure, further includes: a dynamic gain adjustment unit;

[0033] The dynamic gain adjustment unit is used to obtain the corrected angular acceleration of the (i+1)th joint at time t. ;

[0034] when When the absolute value of the acceleration is greater than the preset acceleration threshold, the dynamic gain adjustment unit adjusts the proportional gain K. p The value increases by one unity gain, and the differential gain K is also increased. d The value also increases by one unit gain.

[0035] The entire process, from data acquisition and prediction to trajectory correction and low-level control execution, is continuously cyclically executed, thereby achieving closed-loop online compensation for the dynamic coupling effect of joints. The closed-loop control cycle is consistent with the data sampling cycle Δt, ensuring that the system can respond and adjust in real time at the same frequency as the physical process.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. This invention effectively addresses the limitations of traditional robot control systems in handling complex, time-varying, and difficult-to-accurately model dynamic coupling effects in multi-joint systems by introducing a deviation prediction model based on machine learning. By predicting trajectory deviations caused by inter-joint coupling, real-time and accurate compensation for dynamic disturbances is achieved, significantly improving the trajectory tracking accuracy of the robot under high-speed and high-precision conditions.

[0038] 2. The lightweight neural network model constructed in this invention is quantized and compressed, enabling it to run on embedded real-time control hardware with extremely low inference latency. This ensures the feasibility of real-time deployment of the entire optimization training method in actual industrial robot systems and avoids computational bottlenecks caused by excessive model complexity.

[0039] 3. This invention employs a low-level controller that adaptively adjusts the PID gain, combined with a machine learning prediction model, to form a multi-layered, adaptive control architecture. When the robot is in high-speed, high-acceleration motion, the dynamic adjustment of the PID gain further enhances the system's responsiveness and stability to complex dynamics, avoiding the performance bottleneck of a single control strategy. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the structure of the robot control system parameter optimization training system based on machine learning according to the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example: Figure 1 As shown, the present invention provides a technical solution: a robot control system parameter optimization training system based on machine learning.

[0043] The system includes: a data acquisition and structured labeling module, a coupling effect strength assessment module, a deviation mapping model construction module, a neural network training module, a real-time prediction and trajectory correction module, and a closed-loop control execution module.

[0044] Specifically, the data acquisition and structured labeling module is connected to the coupling influence intensity assessment module and the deviation mapping model construction module; the coupling influence intensity assessment module is connected to the deviation mapping model construction module; the deviation mapping model construction module is connected to the neural network training module; the neural network training module is connected to the real-time prediction and trajectory correction module; and the real-time prediction and trajectory correction module is connected to the closed-loop control execution module.

[0045] The data acquisition and structured labeling module is used to collect motion state data of each joint in real time during the robot's execution of a preset trajectory task. The motion state data includes joint angle, angular velocity, angular acceleration and driving current, and the dynamic influence between adjacent joints is structured and labeled.

[0046] The coupling influence strength assessment module is used to evaluate the influence relationship between adjacent joints based on the coupling path of robot joints, select joint combinations whose coupling influence strength ranking is higher than a threshold as key coupling paths, and establish a compensation model for key coupling paths.

[0047] The coupling influence intensity assessment module includes: a joint pair management unit, a preset deviation management unit, a deviation value assessment unit, and a key coupling screening unit.

[0048] The joint pair management unit is used to collect the motion state parameters of the i-th joint of the robot at time t, record the motion state parameters of the joint according to the coupling path of the joint, and combine the motion state parameters of adjacent joints into adjacent joint data pairs. The motion state parameters include angular acceleration and driving current.

[0049] The preset deviation management unit is used to calculate the deviation δ between the actual position of the (i+1)th joint and the preset trajectory position.i+1 ;

[0050] The deviation evaluation unit is used to calculate the deviation Δδ between the actual position of the (i+1)th joint and the preset trajectory position per unit time. i+1 , and the angular acceleration Δa of the i-th joint per unit time i The absolute value of the ratio is denoted as the joint coupling influence strength evaluation index of the node pair (i, i+1) consisting of the i-th joint and the (i+1)-th joint. , ;

[0051] The key coupling screening unit is used to obtain the joint coupling influence intensity evaluation index of all node pairs formed by adjacent joints, sort them from high to low, and take the top K% of node pairs as key coupling paths, where K∈(0,100).

[0052] The deviation mapping model construction module is used to construct a deviation mapping model of the joint path coupling effect based on structured labeled data. The deviation mapping model takes the angular acceleration and driving current of the joint as input features and the deviation between the actual position of the joint in the joint group and the preset trajectory position in the next sampling period as the output label.

[0053] The deviation mapping model construction module includes: an angle management unit, an angle deviation management unit, a data smoothing unit, and a label input unit;

[0054] The angle management unit is used to collect angle information at the robot joints, where the actual angle of the (i+1)th joint at time t+Δt is denoted as... The target angle of the (i+1)th joint at the corresponding time in the preset trajectory is denoted as . ;

[0055] The angle deviation management unit is used to output a label as the deviation between the actual position of the joint and the preset trajectory position in the next sampling period, where the deviation of the (i+1)th joint is... The calculation formula is: , where Δt represents the length of time per unit time;

[0056] The data smoothing unit is used to smooth the joint deviation by introducing a sliding window filtering mechanism. A window of W unit time is set, and the smoothed deviation value is: , The calculation formula is: , where w m This represents the weight, with an initial weight value ω1. When m=0, w m =ω1, which decreases in the form of the (m+1)th weight being half of the mth weight, where ω1∈(0,1).

[0057] The label input unit is used to take the smoothed deviation value as the true label of the neural network.

[0058] The neural network training module is used to train the bias mapping model using a neural network.

[0059] In the neural network training module, the neural network structure includes an input layer with 2 nodes, a first hidden layer with 16 nodes, a second hidden layer with 8 nodes, and an output layer with 1 node.

[0060] The input layer nodes take the angular acceleration ai(t) at time t of the i-th node as the first input feature, and the driving current I at time t of the i-th node as the second input feature. i (t) is used as the second input feature;

[0061] The output layer output is used as the prediction bias for the i-th joint at time t+Δt. ;

[0062] During training, the mean squared error is used as the loss function L. The formula for calculating the loss function L is: Where N represents the number of samples in the batch, and These represent the true deviation and the predicted deviation of the k-th sample in the batch, respectively.

[0063] The real-time prediction and trajectory correction module is used to monitor the motion state data of the robot joints in real time. It inputs the angular acceleration and drive current of the i-th joint at the current moment into the trained neural network model, predicts the position deviation of the joint after one unit of time, and dynamically corrects the target trajectory of the joint based on the prediction value.

[0064] Among them, the real-time prediction and trajectory correction module inputs the angular acceleration and driving current of the i-th joint at the current moment into the trained neural network model to predict the position deviation of the (i+1)-th joint after one unit of time.

[0065] Get the current time ct, and the time ct+Δt one unit time later than the current time. Get the position deviation of the (i+1)th joint at time ct+Δt. The set value of the angle of the (i+1)th joint at time ct+Δt. ;

[0066] The target trajectory of the (i+1)th joint is dynamically corrected based on the predicted value. The dynamic correction includes calculating the dynamic correction value of the angle of the (i+1)th joint at time ct+Δt. , The calculation formula is:

[0067] , where k represents the compensation gain coefficient, k∈(0.8,1.2).

[0068] The closed-loop control execution module is used to input the corrected target angle into the underlying position controller, drive the robot's joints to execute the optimized motion commands, and continuously cycle through the data acquisition and trajectory correction process to achieve closed-loop online compensation.

[0069] Among them, the closed-loop control execution module continuously and cyclically executes the data acquisition to trajectory correction process to achieve closed-loop online compensation;

[0070] Closed-loop online compensation includes: obtaining the correction error e(t) of the (i+1)th joint at time t, where e(t) is calculated using the following formula: ;

[0071] The control output u applied to the (i+1)th joint at time t is established using a PID control structure. i+1 (t), u i+1 The formula for calculating (t) is: , where K p K represents the proportional gain. i K represents the integral gain. d This represents the differential gain.

[0072] Furthermore, the closed-loop control execution module, during the execution of the PID control structure, further includes: a dynamic gain adjustment unit;

[0073] The dynamic gain adjustment unit is used to obtain the corrected angular acceleration of the (i+1)th joint at time t. ;

[0074] when When the absolute value of the acceleration is greater than the preset acceleration threshold, the dynamic gain adjustment unit adjusts the proportional gain K. p The value increases by one unity gain, and the differential gain K is also increased. d The value also increases by one unit gain.

[0075] The following will provide a detailed description of the machine learning-based robot control system parameter optimization training method using a specific embodiment.

[0076] An industrial robot with six tandem joints has five adjacent joint pairs: (J1, J2), (J2, J3), (J3, J4), (J4, J5), (J5, J6). If, in an actual test, the measured value of S (rad / (rad / s^2)) is as follows:

[0077] J1→J2: 0.0012, J2→J3: 0.0028, J3→J4: 0.0035, J4→J5: 0.0018, J5→J6: 0.0009;

[0078] J3→J4 was selected as the key coupling path, and a machine learning compensation model was built only for this joint pair.

[0079] 1. Data Acquisition and Structured Tagging:

[0080] Suppose a robot is performing a high-speed trajectory task. In this task, the J3 joint, as the preceding joint, has a significant dynamic coupling effect on the J4 joint.

[0081] At any moment Data collection begins at fixed time intervals, starting from a second.

[0082] At a specific moment Motion state data of the J3 joint were collected:

[0083] Instantaneous angle of J3 joint Instantaneous angular velocity of joint J3 Instantaneous angular acceleration of joint J3 J3 joint drive current ;

[0084] At the same time, the J4 joint was acquired at the current moment. and the next moment Data:

[0085] J4 joint The actual angle of the moment J4 joint Setting target angle at any time J4 joint The actual angle of the moment J4 joint Setting target angle at any time ;

[0086] Perform structured tagging: Marked as affecting the J4 joint Disturbance sources of time and position deviation.

[0087] 2. Assessment of the impact of joint coupling on strength:

[0088] Assume that before training, the coupling strength of the J3-J4 joint pair is evaluated through controlled experiments, and the actual position deviation of J4 is recorded when J3 generates a series of preset angular acceleration profiles.

[0089] Numerical methods were used to measure small changes in J3 angular acceleration. and the resulting change in J4 position deviation ,calculate Calculations yielded .

[0090] 3. Deviation mapping model construction and sliding window filtering:

[0091] Based on the collected data, the J4 joint was calculated. Original deviation of time :

[0092] .

[0093] To perform sliding window filtering, historical time-series bias data is required. Assume that... The four sampling periods prior to time (i.e.) Within this range, the original deviations of joint J4 are as follows:

[0094] , , , ;

[0095] According to the sliding window filtering formula ,in Weight .

[0096] Filtered deviation value The calculation is as follows:

[0097] ;

[0098] ;

[0099] ;

[0100] This filtered deviation value This will serve as the true label for the neural network.

[0101] 4. Neural Network Training:

[0102] Assume that sample data has been collected covering different working conditions, with load mass ranging from 0 kg to 10 kg, joint angular velocity ranging from 0.1 rad / s to 2.0 rad / s, and trajectory curvature radius greater than 2 m, and each sample contains input features. and the corresponding filtered deviation label .

[0103] Neural network structure:

[0104] Input layer: 2 nodes ( );

[0105] First hidden layer: 16 nodes, activated via ReLU;

[0106] Second hidden layer: 8 nodes, activated via ReLU;

[0107] Output layer: 1 node ( );

[0108] Training parameters:

[0109] Loss function: mean squared error; optimizer: Adam; learning rate: η=0.001;

[0110] Training rounds: 500 rounds, using an early stopping strategy, terminating when the validation set loss does not decrease for 20 consecutive rounds.

[0111] Training process example:

[0112] Suppose that in a certain training batch, the input of the k-th training sample is... The corresponding true filtered bias label is .

[0113] The prediction bias is obtained through forward propagation calculation using a neural network. .

[0114] If the initial prediction is Then the loss contribution of this sample is:

[0115] .

[0116] The Adam optimizer will adjust the weights and biases of the neural network through backpropagation based on this loss. After 500 training iterations or early stopping, the model converges.

[0117] 5. Real-time trajectory correction and low-level control:

[0118] During the robot's real-time operation phase, the system in each... Perform the following operations during the sampling period:

[0119] Current moment The input features of the J3 joint are: and .

[0120] These two values ​​are input into a trained and quantized neural network model, assuming the model predicts the J4 joint at the next time step. The deviation is Check and correct the trigger conditions: Set the threshold to 0.001 rad;

[0121] because The prediction deviation did not trigger target trajectory correction.

[0122] Therefore, the target angle for the correction of joint J4 remains the same. .

[0123] Another scenario in the embodiment:

[0124] If at another moment The more intense movement of the J3 joint leads to a deviation in the neural network's prediction of the J4 joint at the next moment. .

[0125] at this time Check and correct the trigger conditions: ;

[0126] because This triggers a target trajectory correction.

[0127] Using compensated gain The revised new target angle The calculation is as follows:

[0128] ;

[0129] .

[0130] This corrected target angle The input will be sent to the underlying position controller of the J4 joint.

[0131] The underlying PID controller executes:

[0132] Assuming in At that moment, the actual measured angle of joint J4 .

[0133] Controller calculation error .

[0134] But here The goal for the next moment, the current moment. The corrected target angle should be ;

[0135] Assuming in The time correction target is .

[0136] but .

[0137] PID parameter adaptive adjustment: Assuming the J4 joint is in Corrected target angular acceleration at any time The calculation result is .

[0138] because Therefore, the PID gain will be adjusted.

[0139] Assuming the original gain is , , .

[0140] The proportional gain is set to a single-step increase of 20%, and the derivative gain is set to a single-step increase of 15%. After adjustment: , , ;

[0141] The PID controller will use these adjusted gains based on the current error. Calculate the control output based on historical integral error and error change rate. A drive current command is calculated and sent to the servo driver of the J4 joint. After receiving the command, the servo driver drives the J4 joint to move towards the corrected target angle.

[0142] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A machine learning-based parameter optimization training system for robot control systems, characterized in that: The system includes: The module includes a data acquisition and structured labeling module, a coupling effect strength assessment module, a deviation mapping model construction module, a neural network training module, a real-time prediction and trajectory correction module, and a closed-loop control execution module. The data acquisition and structured labeling module is used to collect motion state data of each joint in real time during the robot's execution of a preset trajectory task. The motion state data includes joint angle, angular velocity, angular acceleration and driving current, and the dynamic influence between adjacent joints is structured and labeled. The coupling influence strength assessment module is used to assess the influence relationship between adjacent joints based on the coupling path of the robot joints, select joint combinations whose joint coupling influence strength ranking is higher than a threshold as key coupling paths, and establish a compensation model for the key coupling paths. The deviation mapping model construction module is used to construct a deviation mapping model of joint path coupling effect based on structured labeled data. The deviation mapping model takes the angular acceleration and driving current of the joint as input features and the deviation between the actual position of the joint in the joint group and the preset trajectory position in the next sampling period as the output label. A neural network training module is used to train the deviation mapping model using a neural network; The real-time prediction and trajectory correction module is used to monitor the motion state data of the robot joints in real time. It inputs the angular acceleration and drive current of the i-th joint at the current moment into the trained neural network model, predicts the position deviation of the joint after one unit of time, and dynamically corrects the target trajectory of the joint based on the prediction value. The closed-loop control execution module is used to input the corrected target angle into the underlying position controller, drive the robot's joints to execute the optimized motion commands, and continuously cycle through the data acquisition and trajectory correction process to achieve closed-loop online compensation. The coupling influence intensity assessment module includes: a joint pair management unit, a preset deviation management unit, a deviation value assessment unit, and a key coupling screening unit; The joint pair management unit is used to collect the motion state parameters of the i-th joint of the robot at time t, record the motion state parameters of the joint according to the coupling path of the joint, and combine the motion state parameters of adjacent joints into adjacent joint data pairs. The motion state parameters include angular acceleration and driving current. The preset deviation management unit is used to calculate the deviation δ between the actual position of the (i+1)th joint and the preset trajectory position. i+1 ; The deviation evaluation unit is used to calculate the deviation Δδ between the actual position of the (i+1)th joint and the preset trajectory position per unit time. i+1 , and the angular acceleration Δa of the i-th joint per unit time i The absolute value of the ratio is denoted as the joint coupling influence strength evaluation index of the node pair (i, i+1) formed by the i-th joint and the (i+1)-th joint. , ; The key coupling screening unit is used to obtain the joint coupling influence intensity evaluation index of all node pairs formed by adjacent joints, sort them from high to low, and take the top K% of node pairs as key coupling paths, where K∈(0,100).

2. The robot control system parameter optimization training system based on machine learning according to claim 1, characterized in that: The deviation mapping model construction module includes: an angle management unit, an angle deviation management unit, a data smoothing unit, and a label input unit; Angle management unit is used to collect angle information at the robot joints, wherein the actual angle of the (i+1)th joint at time t+Δt is denoted as... The target angle of the (i+1)th joint at the corresponding time in the preset trajectory is denoted as . ; An angle deviation management unit is used to output a label as the deviation between the actual position of the joint and the preset trajectory position in the next sampling period, wherein the deviation of the (i+1)th joint... The calculation formula is: , where Δt represents the length of time per unit time; The data smoothing unit is used to smooth the deviation of the joint by introducing a sliding window filtering mechanism. A window of W unit time is set, and the smoothed deviation value is: The The calculation formula is: , where w m This represents the weight, with an initial weight value ω1. When m=0, w m =ω1, which decreases in the form of the (m+1)th weight being half of the mth weight, where ω1∈(0,1). The label input unit is used to take the smoothed deviation value as the true label of the neural network.

3. The robot control system parameter optimization training system based on machine learning according to claim 2, characterized in that: In the neural network training module, the neural network structure includes an input layer with 2 nodes, a first hidden layer with 16 nodes, a second hidden layer with 8 nodes, and an output layer with 1 node. The input layer nodes respectively input the angular acceleration a at time t of the i-th node. i (t) is the first input feature, and the driving current I at time t of the i-th node is... i (t) is used as the second input feature; The output result of the output layer is used as the prediction deviation of the i-th joint at time t+Δt. ; During training, mean squared error is used as the loss function L, and the formula for calculating the loss function L is as follows: Where N represents the number of samples in the batch, and These represent the true deviation and the predicted deviation of the k-th sample in the batch, respectively.

4. The robot control system parameter optimization training system based on machine learning according to claim 3, characterized in that: The real-time prediction and trajectory correction module inputs the angular acceleration and driving current of the i-th joint at the current moment into the trained neural network model to predict the position deviation of the (i+1)-th joint after one unit of time. Obtain the current time ct, and the time ct+Δt one unit time later than the current time, and obtain the position deviation of the (i+1)th joint at time ct+Δt. and the set value of the angle of the (i+1)th joint at time ct+Δt. ; The target trajectory of the (i+1)th joint is dynamically corrected based on the predicted value. The dynamic correction includes calculating the dynamic correction value of the angle of the (i+1)th joint at time ct+Δt. The The calculation formula is: , where k represents the compensation gain coefficient, k∈(0.8,1.2).

5. The robot control system parameter optimization training system based on machine learning according to claim 4, characterized in that: The closed-loop control execution module continuously and cyclically executes the data acquisition to trajectory correction process to achieve closed-loop online compensation. The closed-loop online compensation includes: obtaining the correction error e(t) of the (i+1)th joint at time t, wherein the formula for calculating e(t) is: ; The control output u applied to the (i+1)th joint at time t is established using a PID control structure. i+1 (t), the u i+1 The formula for calculating (t) is: , where K p K represents the proportional gain. i K represents the integral gain. d This represents the differential gain.

6. The robot control system parameter optimization training system based on machine learning according to claim 5, characterized in that: The closed-loop control execution module, during the execution of the PID control structure, further includes: a dynamic gain adjustment unit; The dynamic gain adjustment unit is used to obtain the corrected angular acceleration of the (i+1)th joint at time t. ; When the When the absolute value of the acceleration is greater than a preset acceleration threshold, the dynamic gain adjustment unit adjusts the proportional gain K. p The value is increased by one unity gain, and the differential gain K is increased by one unity gain. d The value also increases by one unit gain.

7. The robot control system parameter optimization training system based on machine learning according to claim 1, characterized in that: The data acquisition and structured labeling module is connected to the coupling influence intensity assessment module and the deviation mapping model construction module; the coupling influence intensity assessment module is connected to the deviation mapping model construction module; the deviation mapping model construction module is connected to the neural network training module; the neural network training module is connected to the real-time prediction and trajectory correction module; and the real-time prediction and trajectory correction module is connected to the closed-loop control execution module.

8. The robot control system parameter optimization training system based on machine learning according to claim 1, characterized in that: The machine learning-based robot control system achieves online optimization training and compensation of robot control system parameters by executing a continuous cyclic process. The cyclic process includes: The motion state data is collected in real time, and the dynamic influence between adjacent joints is structurally labeled. Based on the structured labeled data and the coupling path, the strength of the joint coupling effect is evaluated, and the key coupling paths are selected. Based on the key coupling path and structured labeled data, the deviation mapping model is constructed and trained to obtain a trained neural network model capable of predicting the joint position deviation. During the actual operation of the robot, the trained neural network model is used to predict the motion state data of the robot joints in real time, and the prediction deviation is obtained. The target trajectory of the robot joint at the next moment is dynamically corrected based on the prediction deviation; The corrected target trajectory is input into the underlying position controller to drive the robot joints to execute optimized motion commands, thereby achieving continuous online compensation and optimization of the robot's motion trajectory.

Images