Multi-axis cooperative control method and system of brushless motor for industrial robot

Abstract

The invention provides a brushless motor multi-axis cooperative control method and system for an industrial robot, and relates to the technical field of control, and the method comprises the steps: processing the real-time parameters of a brushless motor of each joint axis through constructing an extended dynamic model, and obtaining a temperature drift coefficient and a load inertia parameter; a self-adaptive pre-compensation model, a neural network compensator and a recurrent neural network prediction model are combined, a torque control instruction is output, a three-phase current prediction value is calculated, a distributed model prediction controller is adopted to generate an optimal switching sequence, multi-axis cooperative control is achieved, and the system stability and the response speed can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of control technology, and particularly to a multi-axis collaborative control method and system for brushless motors used in industrial robots. Background Art

[0002] Due to advantages such as high efficiency, high power density, low noise, and long service life, brushless motors are widely used in the drive systems of industrial robots. In multi-axis collaborative control, the brushless motors of each joint must operate in coordination to ensure that the robot accurately completes the predetermined actions.

[0003] Traditional industrial robot control systems usually adopt classical control methods such as PID, and achieve the position, speed, and current control of each joint by constructing a single-axis independent control loop. With the continuous improvement of the requirements for robot motion accuracy, dynamic response, and adaptability in industrial production, more advanced control algorithms are needed to handle complex non-linear dynamic characteristics, multi-axis coupling, and environmental disturbances and other problems.

[0004] The existing technologies have the following defects and deficiencies in the multi-axis collaborative control of industrial robots:

[0005] Firstly, traditional control algorithms are difficult to effectively cope with the temperature change problem during the operation of industrial robots. The temperature rise of the brushless motor during long-term operation will cause the motor parameters to drift, and the traditional control algorithm lacks the ability to compensate for this temperature-related parameter change in real time, thus affecting the control accuracy.

[0006] Secondly, it is difficult to accurately identify and compensate the real-time changes of the load inertia parameters of the multi-axis system. Under different postures and load conditions of industrial robots, the equivalent inertia of each joint axis will change significantly. The existing control methods often adopt a fixed parameter model and cannot adapt to this dynamic change, resulting in a reduction in control performance. Summary of the Invention

[0007] The embodiments of the present invention provide a multi-axis collaborative control method and system for brushless motors used in industrial robots, which can solve the problems in the existing technologies.

[0008] In the first aspect of the embodiments of the present invention, a multi-axis collaborative control method for brushless motors used in industrial robots is provided, including:

[0009] Construct an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtain the temperature drift coefficient and load inertia parameter of the brushless motor of each joint axis;

[0010] Input the temperature drift coefficient and load inertia parameter into an adaptive pre-compensation model, and calculate the system disturbance amount and fractional-order sliding mode surface in combination with the extended dynamic model;

[0011] Input the system disturbance quantity and the fractional-order sliding mode surface into the neural network compensator. The neural network compensator processes the system disturbance quantity and the fractional-order sliding mode surface using the backpropagation algorithm and outputs a torque compensation signal;

[0012] Superimpose the torque compensation signal and the compensation torque to obtain a torque control command, and convert the torque control command into the three-phase current set values of the brushless motors of each joint axis;

[0013] Input the three-phase current set values and the phase current signals into the recurrent neural network prediction model, and the recurrent neural network prediction model outputs the three-phase current prediction values of the brushless motors of each joint axis;

[0014] Input the three-phase current prediction values into the distributed model predictive controller, and calculate the optimal switching sequence of each joint axis based on the coupling torque between the joint axes as the inverter drive signal of the brushless motor of each joint axis.

[0015] Construct an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtain the temperature drift coefficient and the load inertia parameters of the brushless motors of each joint axis, including:

[0016] Construct a winding resistance temperature coefficient model and a permanent magnet remanence density temperature coefficient model, and calculate a temperature-related torque constant model according to the winding resistance temperature coefficient model and the permanent magnet remanence density temperature coefficient model;

[0017] Construct an extended dynamic model based on the temperature-related torque constant model and the real-time operating parameters, input the real-time operating parameters into the extended dynamic model to calculate the parameter estimation matrix, and calculate the parameter identification error according to the parameter estimation matrix;

[0018] Construct an adaptive forgetting factor based on the parameter identification error, and update the parameter estimation matrix and the covariance matrix based on the adaptive forgetting factor;

[0019] Construct a state space model, substitute the parameter estimation matrix and the covariance matrix into the Kalman filter for parameter prediction, and obtain the parameter prediction value and the prediction covariance matrix;

[0020] Calculate the Kalman gain according to the parameter prediction value and the prediction covariance matrix, and use the Kalman gain and the temperature-related torque constant model for parameter correction to obtain the corrected state estimation value;

[0021] Establish parameter constraint conditions based on the corrected state estimation value, and calculate the temperature drift coefficient and the load inertia parameters of the brushless motors of each joint axis according to the parameter constraint conditions.

[0022] Input the system disturbance quantity and the fractional-order sliding mode surface into the neural network compensator. The neural network compensator processes the system disturbance quantity and the fractional-order sliding mode surface using the backpropagation algorithm and outputs a torque compensation signal, including:

[0023] Construct a composite fractional-order differential expression based on the system disturbance quantity and the fractional-order sliding mode surface, and input the composite fractional-order differential expression into a group of nonlinear disturbance observers. The group of nonlinear disturbance observers calculates the subsystem disturbance quantity based on the hybrid adaptive time-varying gain matrix and the multi-nested high-order observer state equation;

[0024] Construct a hierarchical recursive disturbance prediction model according to the subsystem disturbance quantity. The hierarchical recursive disturbance prediction model calculates the system comprehensive disturbance quantity using the weighted moving average method;

[0025] Calculate a multi-dimensional disturbance prediction error vector based on the system comprehensive disturbance quantity and the subsystem disturbance quantity, update the hybrid adaptive time-varying gain matrix according to the multi-dimensional disturbance prediction error vector, and feedback the updated hybrid adaptive time-varying gain matrix to the group of nonlinear disturbance observers;

[0026] Input the system disturbance quantity, the fractional-order sliding mode surface, the system comprehensive disturbance quantity, and the multi-dimensional disturbance prediction error vector into the neural network compensator. The neural network compensator performs multi-level signal processing on the input signals and generates multi-level output signals;

[0027] Construct a multi-objective composite error function based on the multi-dimensional disturbance prediction error vector and the multi-level output signals;

[0028] Construct a recursive network structure evaluation model based on the multi-objective composite error function, and generate a determined torque compensation signal according to the recursive network structure evaluation model.

[0029] Construct a recursive network structure evaluation model based on the multi-objective composite error function, and generate a determined torque compensation signal according to the recursive network structure evaluation model, including:

[0030] Construct a recursive network structure evaluation model based on the multi-objective composite error function. The recursive network structure evaluation model calculates the neuron importance index matrix and the network topology feature vector;

[0031] Construct a two-way propagation mechanism based on the network topology feature vector. The forward channel of the two-way propagation mechanism constructs an adaptive pruning criterion based on the neuron importance index matrix to generate optimized neuron connection weights. The reverse channel of the two-way propagation mechanism combines the optimized neuron connection weights with the neuron importance index matrix to construct a new constraint term and updates the multi-objective composite error function;

[0032] Input the neuron importance index matrix and the optimized neuron connection weights into the learning parameter generation module to generate an adaptive learning rate matrix;

[0033] Construct a dynamic momentum factor matrix based on the updated multi-objective composite error function and the network topology feature vector, and adjust the dynamic momentum factor matrix in combination with the neuron importance index matrix to generate optimized momentum update parameters;

[0034] Input the optimized momentum update parameters and the dynamic momentum factor matrix into the state calculation module to calculate the output state signals of each hidden layer, and determine the torque compensation signal based on the output state signals of each hidden layer.

[0035] Input the three-phase current given value and the phase current signal into the recurrent neural network prediction model, and the recurrent neural network prediction model outputs the three-phase current prediction values of the brushless motors of each joint axis, including:

[0036] Input the three-phase current given value and the phase current signal into the recurrent neural network prediction model, and the recurrent neural network prediction model performs multi-layer decomposition on the phase current signal by using wavelet packet decomposition to obtain the filtered signal;

[0037] Perform normalization processing on the three-phase current given value and the filtered signal to obtain normalized features, and construct a time-series sample matrix according to an adaptive time window for the normalized features;

[0038] Extract the principal component feature vector of the time-series sample matrix, process the principal component feature vector by using different types of recurrent neural network units, and output the corresponding hidden layer states;

[0039] Construct a hybrid attention mechanism based on the hidden layer states output by each recurrent neural network unit, and calculate the time-series correlation weight and the feature dimension weight of the hidden layer states based on the hybrid attention mechanism;

[0040] Input the time-series correlation weight and the feature dimension weight into a multi-layer perceptron to generate a fused attention vector, and perform weighted fusion on the fused attention vector and the hidden layer states to obtain an enhanced feature vector;

[0041] Construct a distributed gradient calculation matrix based on the enhanced feature vector, the hidden layer states and the principal component feature vector, and determine the three-phase current prediction values of the brushless motors of each joint axis in combination with a multi-head attention fusion network.

[0042] Construct a distributed gradient calculation matrix based on the enhanced feature vector, the hidden layer states and the principal component feature vector, and determine the three-phase current prediction values of the brushless motors of each joint axis in combination with a multi-head attention fusion network, including:

[0043] Input the enhanced feature vector, the hidden layer state, and the principal component feature vector into the multi-objective combined loss function construction module, and construct a distributed gradient calculation matrix based on the time series prediction loss and the feature reconstruction loss;

[0044] Output a multi-scale learning rate matrix based on the distributed gradient calculation matrix, and input the multi-scale learning rate matrix into a hybrid parameter optimizer, which updates the parameters of the recurrent neural network prediction model;

[0045] Process the enhanced feature vector based on the updated parameters of the recurrent neural network prediction model to obtain the current prediction tensors of each sub-network, and use a multi-head attention fusion network to weight the current prediction tensors of each sub-network to obtain the initial three-phase current prediction values of each joint-axis brushless motor;

[0046] Input the initial three-phase current prediction values into a hierarchical prediction result optimization system, output an evaluation index vector, and perform time-domain and frequency-domain correction on the initial three-phase current prediction values based on the evaluation index vector to obtain the corrected three-phase current prediction values;

[0047] Calculate the error matrix between the corrected three-phase current prediction values and the actual three-phase current values, and solve a multi-constraint optimization problem based on the error matrix to obtain an optimal parameter configuration vector;

[0048] Feed back the optimal parameter configuration vector to the corresponding module for parameter update, and re-perform prediction calculations based on the updated parameters to output the three-phase current prediction values of each joint-axis brushless motor.

[0049] Input the three-phase current prediction values into a distributed model predictive controller, and calculate the optimal switching sequence of each joint axis as the inverter drive signal of each joint-axis brushless motor based on the coupling torque between the joint axes, including:

[0050] Input the three-phase current prediction values into a distributed model predictive controller, which constructs a hierarchical recursive network structure based on the position, speed, and acceleration of each joint axis. The hierarchical recursive network structure models the dynamic coupling characteristics between joints and outputs a joint-interaction coupling torque model;

[0051] Construct a distributed state predictor based on the joint-interaction coupling torque model, and the distributed state predictor outputs a sequence of joint coupling torques within the prediction time domain;

[0052] Input the sequence of joint coupling torques and the three-phase current prediction values into a multi-objective optimizer, and the multi-objective optimizer constructs a loss evaluation model based on the conduction time of power devices;

[0053] Construct a current error constraint based on the predicted values of the three-phase current, construct a torque balance constraint based on the joint coupling torque sequence, and construct a rolling horizon optimization problem based on the loss evaluation model, the current error constraint, and the torque balance constraint;

[0054] Solve the rolling horizon optimization problem using the branch and bound algorithm to obtain the candidate switch sequences for each joint axis, construct a parallel evaluator based on the current error constraint and the torque balance constraint, and use the parallel evaluator to perform multi-objective evaluation on the candidate switch sequences, and select the optimal switch sequence for each joint axis as the inverter drive signal for the brushless motor of each joint axis.

[0055] In a second aspect of the embodiments of the present invention, there is provided a multi-axis cooperative control system for a brushless motor used in an industrial robot, including:

[0056] A first unit for constructing an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtaining the temperature drift coefficient and the load inertia parameter of the brushless motors of each joint axis;

[0057] A second unit for inputting the temperature drift coefficient and the load inertia parameter into an adaptive pre-compensation model, and calculating a system disturbance amount and a fractional-order sliding mode surface in combination with the extended dynamic model;

[0058] A third unit for inputting the system disturbance amount and the fractional-order sliding mode surface into a neural network compensator, and the neural network compensator processes the system disturbance amount and the fractional-order sliding mode surface using the backpropagation algorithm and outputs a torque compensation signal;

[0059] A fourth unit for superimposing the torque compensation signal and the compensation torque to obtain a torque control command, and converting the torque control command into a three-phase current given value for the brushless motors of each joint axis;

[0060] A fifth unit for inputting the three-phase current given value and the phase current signal into a recurrent neural network prediction model, and the recurrent neural network prediction model outputs the predicted values of the three-phase current of the brushless motors of each joint axis;

[0061] A sixth unit for inputting the predicted values of the three-phase current into a distributed model predictive controller, and calculating the optimal switch sequence for each joint axis as the inverter drive signal for the brushless motor of each joint axis based on the coupling torque between the joint axes.

[0062] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including:

[0063] A processor;

[0064] A memory for storing instructions executable by the processor;

[0065] Among them, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0066] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0067] The beneficial effects of this application are as follows:

[0068] In the present invention, the extended dynamics model is used to process the parameters of the brushless motors of each joint axis of the industrial robot in real time, obtain the temperature drift coefficient and the load inertia parameters, and combine the adaptive pre-compensation model to calculate the system disturbance amount and the sliding mode surface, effectively improving the operation accuracy and stability of the industrial robot.

[0069] The present invention uses a neural network compensator to process the system disturbance amount and the fractional-order sliding mode surface, outputs a torque compensation signal, and predicts the three-phase current values of the brushless motors of each joint axis through a recurrent neural network prediction model, enhancing the system's ability to suppress external disturbances and dynamic response performance.

[0070] The present invention uses a distributed model predictive controller to calculate the optimal switching sequence as the inverter drive signal based on the coupling torque between the joint axes, realizing multi-axis coordinated control, reducing the system energy consumption, improving the control accuracy, simplifying the system structure, and enhancing the overall performance and reliability of the industrial robot. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic flowchart of the multi-axis coordinated control method for brushless motors used in the industrial robot according to the embodiments of the present invention;

[0072] Figure 2 It is a schematic architecture diagram of the neural network compensator;

[0073] Figure 3 It is a schematic flowchart for determining the inverter drive signal. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0074] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0075] The technical solution of the present invention will be described in detail below with specific embodiments. The following several specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0076] Referring to Figures 1 to 3 , the multi-axis cooperative control method for a brushless motor of an industrial robot in an embodiment of the present invention includes:

[0077] Construct an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtain the temperature drift coefficient and load inertia parameter of the brushless motor of each joint axis;

[0078] Input the temperature drift coefficient and load inertia parameter into the adaptive pre-compensation model, and calculate the system disturbance amount and fractional-order sliding mode surface in combination with the extended dynamic model;

[0079] Input the system disturbance amount and fractional-order sliding mode surface into the neural network compensator. The neural network compensator processes the system disturbance amount and fractional-order sliding mode surface by using the backpropagation algorithm and outputs a torque compensation signal;

[0080] Superimpose the torque compensation signal and the compensation torque to obtain a torque control command, and convert the torque control command into a three-phase current set value of the brushless motor of each joint axis;

[0081] Input the three-phase current set value and the phase current signal into the recurrent neural network prediction model, and the recurrent neural network prediction model outputs the three-phase current prediction value of the brushless motor of each joint axis;

[0082] Input the three-phase current prediction value into the distributed model predictive controller, and calculate the optimal switching sequence of each joint axis based on the coupling torque between the joint axes as the inverter drive signal of the brushless motor of each joint axis.

[0083] In an optional implementation manner, constructing an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtaining the temperature drift coefficient and load inertia parameter of the brushless motor of each joint axis, includes:

[0084] Construct a winding resistance temperature coefficient model and a permanent magnet remanence density temperature coefficient model, and calculate a temperature-related torque constant model according to the winding resistance temperature coefficient model and the permanent magnet remanence density temperature coefficient model;

[0085] Based on the temperature-related torque constant model and the real-time operating parameters, construct an extended dynamic model, input the real-time operating parameters into the extended dynamic model to calculate a parameter estimation matrix, and calculate a parameter identification error according to the parameter estimation matrix;

[0086] Construct an adaptive forgetting factor based on the parameter identification error, and update the parameter estimation matrix and covariance matrix based on the adaptive forgetting factor;

[0087] Construct a state space model, substitute the parameter estimation matrix and the covariance matrix into a Kalman filter for parameter prediction, and obtain a parameter prediction value and a prediction covariance matrix;

[0088] Calculate a Kalman gain according to the parameter prediction value and the prediction covariance matrix, and use the Kalman gain and the temperature-related torque constant model to perform parameter correction to obtain a corrected state estimation value;

[0089] Establish parameter constraint conditions based on the corrected state estimation value, and calculate the temperature drift coefficient and load inertia parameters of the brushless motors of each joint axis according to the parameter constraint conditions.

[0090] The implementation manner of the present invention will be described in detail below in combination with an actual application scenario.

[0091] For a brushless motor, the characteristic that the winding resistance changes with temperature can be expressed as that the current resistance value is equal to the resistance value at the reference temperature multiplied by the functional relationship between the resistance temperature coefficient and the temperature difference. For example, in a test environment, when the motor temperature rises from 25 °C to 85 °C, the winding resistance increases from 1.5 ohms to about 1.8 ohms, and the resistance temperature coefficient is about 0.004 / °C.

[0092] At the same time, the temperature change of the remanent magnetic density of the permanent magnet can be expressed as that the current remanent magnetic density is equal to the remanent magnetic density at the reference temperature multiplied by the functional relationship between the magnetic flux temperature coefficient and the temperature difference. For example, for a motor using N35H neodymium iron boron permanent magnets, when the temperature rises from 25 °C to 85 °C, the remanent magnetic density decreases from 1.2 T to about 1.14 T, and the corresponding magnetic flux temperature coefficient is about -0.001 / °C.

[0093] Based on the above two temperature coefficient models, calculate a temperature-related torque constant model. The torque constant can be expressed as the torque constant at the reference temperature multiplied by the functional relationship between the comprehensive temperature coefficient and the temperature difference. In actual tests, when the temperature rises from 25 °C to 85 °C, the torque constant decreases from 0.12 N·m / A to about 0.114 N·m / A, and the comprehensive temperature coefficient is about -0.0007 / °C.

[0094] Construct an extended dynamics model based on the temperature-related torque constant model and real-time operating parameters. The real-time operating parameters include information such as motor current, voltage, speed, position, and temperature. For example, for the first joint of an industrial robot under a certain working condition, data such as the motor operating current of 5 A, voltage of 24 V, speed of 2000 rpm, and temperature of 65 °C are collected.

[0095] Input the real-time operating parameters into the extended dynamics model to calculate the parameter estimation matrix, which contains the estimated values of two key parameters: the temperature drift coefficient and the load inertia. Calculate the parameter identification error based on the difference between the parameter estimation matrix and the actual measured values. For example, in the initial estimation stage, the estimated value of the temperature drift coefficient is -0.0005 / °C, and the actual value is -0.0007 / °C. The estimated value of the load inertia is 0.05 kg·m 2 , and the actual value is 0.06 kg·m 2 , and the resulting identification errors are 28.6% and 16.7% respectively.

[0096] Construct an adaptive forgetting factor based on the parameter identification error. When the identification error is large, use a smaller forgetting factor (such as 0.85) to accelerate the convergence rate; when the identification error is small, use a larger forgetting factor (such as 0.98) to improve stability. Based on this adaptive forgetting factor, update the parameter estimation matrix and the covariance matrix to reduce the estimation bias.

[0097] Construct a state space model, which represents the motor dynamics characteristics in the form of state equations and observation equations. Substitute the parameter estimation matrix and the covariance matrix into the Kalman filter for parameter prediction to obtain the parameter prediction values and the prediction covariance matrix. For example, after the prediction stage, the predicted value of the temperature drift coefficient is adjusted to -0.00065 / °C, and the predicted value of the load inertia is adjusted to 0.055 kg·m 2 .

[0098] Calculate the Kalman gain based on the parameter prediction values and the prediction covariance matrix. For example, the Kalman gains of the temperature drift coefficient and the load inertia are calculated to be 0.45 and 0.38 respectively. Use the Kalman gain and the torque constant model related to temperature to correct the parameters and obtain the corrected state estimation values. After correction, the estimated value of the temperature drift coefficient is updated to -0.00068 / °C, and the estimated value of the load inertia is updated to 0.058 kg·m 2 .

[0099] Establish parameter constraint conditions based on the corrected state estimation values. For the temperature drift coefficient, its value range should be between -0.0010 / °C and -0.0005 / °C; for the load inertia parameter, its value should be positive and not exceed 120% of the maximum allowable load inertia of the motor. According to these constraint conditions, the finally calculated temperature drift coefficient is -0.00068 / °C, and the load inertia parameter is 0.058 kg·m 2 , and the identification errors are reduced to 2.9% and 3.3% respectively.

[0100] Through the above steps, the present invention can identify the temperature drift coefficient and load inertia parameters of the brushless motors of each joint axis in real time during the operation of the industrial robot, provide accurate parameter estimation for the robot control system, and thus improve the motion control accuracy. Practical applications show that after adopting this method, in a temperature fluctuation environment (20°C to 80°C), the end positioning accuracy of the robot is improved by about 28%, and the trajectory tracking error is reduced by about 35%, significantly improving the performance of the industrial robot in high-precision application scenarios.

[0101] In an alternative embodiment, the system disturbance quantity and the fractional-order sliding mode surface are input into a neural network compensator, and the neural network compensator processes the system disturbance quantity and the fractional-order sliding mode surface by using a backpropagation algorithm and outputs a torque compensation signal, including:

[0102] Construct a composite fractional-order differential expression based on the system disturbance quantity and the fractional-order sliding mode surface, and input the composite fractional-order differential expression into a group of nonlinear disturbance observers. The group of nonlinear disturbance observers calculates the subsystem disturbance quantity based on a hybrid adaptive time-varying gain matrix and a multi-nested high-order observer state equation;

[0103] Construct a hierarchical recursive disturbance prediction model according to the subsystem disturbance quantity, and the hierarchical recursive disturbance prediction model calculates the system comprehensive disturbance quantity by using a weighted moving average method;

[0104] Calculate a multi-dimensional disturbance prediction error vector based on the system comprehensive disturbance quantity and the subsystem disturbance quantity, update the hybrid adaptive time-varying gain matrix according to the multi-dimensional disturbance prediction error vector, and feedback the updated hybrid adaptive time-varying gain matrix to the group of nonlinear disturbance observers;

[0105] Input the system disturbance quantity, the fractional-order sliding mode surface, the system comprehensive disturbance quantity, and the multi-dimensional disturbance prediction error vector into the neural network compensator. The neural network compensator performs multi-level signal processing on the input signals and generates multi-level output signals;

[0106] Construct a multi-objective composite error function based on the multi-dimensional disturbance prediction error vector and the multi-level output signals;

[0107] Construct a recursive network structure evaluation model based on the multi-objective composite error function, and generate a determined torque compensation signal according to the recursive network structure evaluation model.

[0108] The following details the complete process of inputting the system disturbance quantity and the fractional-order sliding mode surface into the neural network compensator to obtain the torque compensation signal.

[0109] The system receives the system disturbance quantity and the fractional - order sliding mode surface as the initial input signals, and constructs a composite fractional - order differential expression using these two inputs. This composite fractional - order differential expression can be expressed as a linear combination of the fractional - order derivative of the system disturbance quantity and the fractional - order sliding mode surface. For example, when the value of the system disturbance quantity is 0.35, the value of the fractional - order sliding mode surface is 0.42, and the order of the fractional - order derivative is 0.5, the value of the composite fractional - order differential expression obtained through linear combination is approximately 0.57.

[0110] The constructed composite fractional - order differential expression is input into a group of nonlinear disturbance observers. This observer group contains three sub - observers working in parallel, and each sub - observer is responsible for estimating the disturbance components in different frequency ranges. The group of nonlinear disturbance observers calculates the subsystem disturbance quantity based on the hybrid adaptive time - varying gain matrix and the multiple nested high - order observer state equation. Specifically, the initial value of the hybrid adaptive time - varying gain matrix can be set as a diagonal matrix, and the diagonal elements are 4.5, 3.2, and 2.8 respectively. The multiple nested high - order observer state equation adopts a fifth - order structure, and the initial value of the state vector is set as a zero vector. After being processed by the observer, the obtained subsystem disturbance quantity may contain a low - frequency component of 0.22, a medium - frequency component of 0.13, and a high - frequency component of 0.08.

[0111] A hierarchical recursive disturbance prediction model is constructed based on the subsystem disturbance quantity. This model uses the weighted moving average method to calculate the system comprehensive disturbance quantity. In actual implementation, three different time windows can be set, namely the short - term window (5 sampling points), the medium - term window (15 sampling points), and the long - term window (30 sampling points). The corresponding weighting factors are 0.5, 0.3, and 0.2 respectively. Through weighted moving average calculation, if the average value of the short - term window is 0.45, the average value of the medium - term window is 0.39, and the average value of the long - term window is 0.36, then the calculated system comprehensive disturbance quantity is 0.416.

[0112] Based on this, the system calculates a multi - dimensional disturbance prediction error vector based on the system comprehensive disturbance quantity and the subsystem disturbance quantity. The calculation method of this error vector is the system comprehensive disturbance quantity minus the weighted sum of each subsystem disturbance quantity. For example, if the system comprehensive disturbance quantity is 0.416 and the weighted sum of the three subsystem disturbance quantities is 0.403, then the disturbance prediction error is 0.013. In practical applications, the multi - dimensional disturbance prediction error vector contains multiple components, corresponding to the prediction errors in different frequency ranges. According to the calculated multi - dimensional disturbance prediction error vector, the system updates the hybrid adaptive time - varying gain matrix. The update rule adopts the gradient descent method, and the learning rate is set to 0.05. For example, if the prediction error is positive, the gain value of the corresponding component is increased; otherwise, it is decreased. The updated hybrid adaptive time - varying gain matrix is fed back to the group of nonlinear disturbance observers for the next round of disturbance estimation.

[0113] The system disturbance quantity, fractional-order sliding mode surface, system comprehensive disturbance quantity, and multi-dimensional disturbance prediction error vector are input into the neural network compensator. The neural network compensator adopts a multi-layer structure, including an input layer (4 neurons), two hidden layers (12 neurons and 8 neurons respectively), and an output layer (1 neuron). The neural network compensator performs multi-level signal processing on the input signal and uses the ReLU activation function. For example, when the input signal is [0.35, 0.42, 0.416, 0.013], the output of the first hidden layer is a vector of length 12, and some values are [0.56, 0.23, 0.78, 0.44...]; the output of the second hidden layer is a vector of length 8, and some values are [0.67, 0.39, 0.51...]. These constitute the multi-level output signals of the neural network.

[0114] Based on the multi-dimensional disturbance prediction error vector and the multi-level output signals, the system constructs a multi-objective composite error function. This error function consists of three parts: disturbance compensation error, control smoothness error, and control energy consumption error. The weights of the three parts are set to 0.5, 0.3, and 0.2 respectively. For example, if the disturbance compensation error is 0.015, the control smoothness error is 0.022, and the control energy consumption error is 0.018, then the value of the multi-objective composite error function is 0.0179.

[0115] The system constructs a recursive network structure evaluation model based on the multi-objective composite error function. This model evaluates the performance of the neural network in a dynamically adjusted manner and adjusts the network parameters according to the evaluation results. The recursive network structure evaluation model searches for the optimal network parameters through an iterative method, and the evaluation threshold for each round of iteration is set to 0.01. When the evaluation value is lower than the threshold, the system considers that the current neural network structure has achieved the expected performance. Based on the finally optimized neural network structure, the system generates a determined torque compensation signal. For example, when the system disturbance quantity is 0.35 and the fractional-order sliding mode surface is 0.42, the finally generated torque compensation signal value is 0.384.

[0116] Through the above steps, the mapping from the system disturbance quantity and the fractional-order sliding mode surface to the torque compensation signal is realized, effectively improving the robustness and anti-disturbance ability of the control system.

[0117] In an alternative embodiment, constructing a recursive network structure evaluation model based on the multi-objective composite error function and generating a determined torque compensation signal according to the recursive network structure evaluation model includes:

[0118] Constructing a recursive network structure evaluation model based on the multi-objective composite error function, and the recursive network structure evaluation model calculates the neuron importance index matrix and the network topology feature vector;

[0119] Construct a two-way propagation mechanism based on the network topology feature vector. The forward channel of the two-way propagation mechanism constructs an adaptive pruning criterion based on the neuron importance index matrix to generate optimized neuron connection weights. The reverse channel of the two-way propagation mechanism combines the optimized neuron connection weights with the neuron importance index matrix to construct a new constraint term and updates the multi-objective composite error function.

[0120] Input the neuron importance index matrix and the optimized neuron connection weights into the learning parameter generation module to generate an adaptive learning rate matrix.

[0121] Construct a dynamic momentum factor matrix based on the updated multi-objective composite error function and the network topology feature vector, and adjust the dynamic momentum factor matrix in combination with the neuron importance index matrix to generate optimized momentum update parameters.

[0122] Input the optimized momentum update parameters and the dynamic momentum factor matrix into the state calculation module to calculate the output state signals of each hidden layer, and determine the torque compensation signal based on the output state signals of each hidden layer.

[0123] Construct a recursive network structure evaluation model based on the multi-objective composite error function. This evaluation model includes an input layer, multiple hidden layers, and an output layer, and the nodes between each layer are connected by connection weights. During the construction process, first initialize the network parameters, including the connection weights and bias values of the neurons in each layer. The initial connection weights can be random values within the range of [-0.1, 0.1]. Then, calculate the activation values of the neurons in each layer through forward calculation, and calculate the network output error using the multi-objective composite error function. This multi-objective composite error function includes a control accuracy term, a system stability term, and an energy efficiency term, and the weight ratio of the three is set to 4:3:3. Next, calculate the neuron importance index matrix, which reflects the contribution degree of each neuron to the overall performance of the network. Specifically, for the j-th neuron in the i-th layer, its importance index value Iij is comprehensively determined by the output value of this neuron, the connection weight, and the influence degree on the activation value of the next layer. At the same time, calculate the network topology feature vector, which includes parameters such as the number of network layers, the number of neurons in each layer, and the connection density, and is used to describe the overall structure characteristics of the network.

[0124] In the forward channel, an adaptive pruning criterion is constructed using the neuron importance index matrix. Specifically, the importance threshold θ = 0.15 is set, and the neuron connections with importance index values less than θ are pruned, that is, their connection weights are set to 0; for the connections with importance index values greater than θ, the connection weights are enhanced proportionally according to their importance values, thereby generating the optimized neuron connection weights. For example, if the original weight of a certain connection is 0.35 and the corresponding neuron importance index value is 0.8, the optimized weight is approximately 0.42. In the reverse channel, the optimized neuron connection weights and the neuron importance index matrix are combined to construct a new constraint term, which promotes the evolution of the network structure towards a more efficient direction and updates the multi-objective composite error function. The updated error function adds a structure optimization term on the basis of the original three terms, and the weight ratio is adjusted to 3:3:2:2.

[0125] The neuron importance index matrix and the optimized neuron connection weights are input into the learning parameter generation module to generate an adaptive learning rate matrix. This module first calculates the average importance values of each layer of the network, and then combines the connection weight distribution characteristics to assign different learning rates to different neurons. Neurons with high importance obtain smaller learning rates to maintain stability, while neurons with low importance obtain larger learning rates to promote rapid adaptation. For example, for a neuron with an importance index value of 0.9, its learning rate may be 0.001; while for a neuron with an importance index value of 0.2, its learning rate may be 0.01.

[0126] Based on the updated multi-objective composite error function and the network topology feature vector, a dynamic momentum factor matrix is constructed. This matrix takes into account the learning status and convergence of each part of the network and can adaptively adjust the momentum factors for updating the parameters of each neuron. Specifically, the dynamic momentum factor matrix is adjusted in combination with the neuron importance index matrix. Larger momentum factors are assigned to regions with high importance to accelerate convergence, and smaller momentum factors are assigned to regions with low importance to improve the exploration ability. For example, if the average importance of a certain region is 0.75 and its initial momentum factor is 0.7, it may increase to 0.85 after adjustment. After adjustment, the optimized momentum update parameters are generated for the network weight update process.

[0127] Input the optimized momentum update parameters and the dynamic momentum factor matrix into the state calculation module to calculate the output state signals of each hidden layer. First, calculate the activation states of neurons in each hidden layer based on the current input and historical state information, combined with the optimized connection weights and momentum parameters. Then, obtain the final output state through forward propagation. Based on the output state signals of each hidden layer, determine the torque compensation signal. In practical applications, for a torque control signal range of [-50, 50] N·m, when it is detected that an external disturbance causes the control error to exceed ±2 N·m, the recursive network evaluation model will generate a corresponding compensation signal. For example, when a -3.5 N·m error is detected, the model may generate a 4.2 N·m compensation torque, effectively suppressing the disturbance effect and controlling the system error within the range of ±0.8 N·m.

[0128] In actual tests, the torque control system using this method has a 35% shorter response time, a 50% reduced overshoot, a steady-state error reduced to 1 / 3 of the original, and shows stronger robustness in the face of sudden load changes compared with traditional PID control.

[0129] Through the above steps, this embodiment realizes a method for determining the torque compensation signal based on the recursive network structure evaluation model, effectively improving the system control accuracy and stability.

[0130] In an alternative embodiment, input the three-phase current setpoint and the phase current signal into the recurrent neural network prediction model, and the recurrent neural network prediction model outputs the predicted values of the three-phase currents of the brushless motors of each joint axis, including:

[0131] Input the three-phase current setpoint and the phase current signal into the recurrent neural network prediction model, and the recurrent neural network prediction model performs multi-layer decomposition on the phase current signal using wavelet packet decomposition to obtain the filtered signal;

[0132] Perform normalization processing on the three-phase current setpoint and the filtered signal to obtain the normalized features, and construct a time-series sample matrix with the normalized features according to an adaptive time window;

[0133] Extract the principal component feature vectors of the time-series sample matrix, and use different types of recurrent neural network units to process the principal component feature vectors and output the corresponding hidden layer states;

[0134] Construct a hybrid attention mechanism based on the hidden layer states output by each recurrent neural network unit, and calculate the time-series correlation weights and feature dimension weights of the hidden layer states based on the hybrid attention mechanism;

[0135] Input the temporal correlation weight and the feature dimension weight into a multi-layer perceptron to generate a fused attention vector, and perform weighted fusion of the fused attention vector and the hidden layer state to obtain an enhanced feature vector;

[0136] Construct a distributed gradient calculation matrix based on the enhanced feature vector, the hidden layer state, and the principal component feature vector, and determine the three-phase current prediction values of each joint axis brushless motor in combination with a multi-head attention fusion network.

[0137] Among them, the three-phase current given value is an instruction signal issued by the robot control system, and the phase current signal is the actual operating current signal of the motor collected in real time by the sensor. In this embodiment, the sampling frequency is 10 kHz, that is, a signal is collected every 0.1 ms.

[0138] Input the collected three-phase current given value and phase current signal into a recurrent neural network prediction model. This model first uses wavelet packet decomposition to perform multi-layer decomposition on the phase current signal to filter out noise. Specifically, select the Daubechies wavelet basis function for 4-level wavelet packet decomposition, decompose the phase current signal into 16 frequency bands, retain the medium and low frequency band (bands 1-8) signals according to the energy distribution characteristics, and discard the high-frequency noise band (bands 9-16) signals, and then reconstruct to obtain the filtered signal. In practical applications, for the motor of the 3rd axis of a certain six-axis robot, the peak value of the original phase A current signal is 5.2 A, containing about ±0.8 A of high-frequency noise. After wavelet packet decomposition filtering, the high-frequency noise is effectively suppressed to within ±0.15 A.

[0139] Adopt the maximum-minimum normalization method to map all signals to the [-1, 1] interval to obtain normalized features. Taking the collected phase A current as an example, the original current range is from -15 A to +15 A, and through normalization processing, it is mapped to the [-1, 1] interval.

[0140] The time window length is adaptively determined according to the dynamic response characteristics of the motor. In this embodiment, a basic window length of 50 ms is used. When it is detected that the motor speed change rate exceeds 100 rpm / s, it is automatically shortened to 25 ms to improve the response ability of the model to rapidly changing states. For standard working conditions, the dimension of the temporal sample matrix is 500×6 (50 ms×10 kHz×6 features), where the 6 features include the three-phase current given value and the three-phase actual current signal.

[0141] For the constructed temporal sample matrix, use the principal component analysis method to extract the principal component feature vector. Retain the principal components that can explain 85% of the data variance, usually the first 4 principal components. For example, for a certain temporal sample matrix with an original dimension of 500×6, a principal component feature vector with a dimension of 500×4 is obtained after principal component analysis.

[0142] The principal component feature vectors are input into different types of recurrent neural network units for processing. In this embodiment, three types of recurrent neural network units are used simultaneously: long short-term memory (LSTM) units, gated recurrent units (GRUs), and simple recurrent units (Simple RNNs). Each network unit contains two layers, with 64 neurons in each layer. Each unit processes the principal component feature vectors respectively and outputs the corresponding hidden layer states. For an input sequence of 500 time steps, each recurrent neural network unit outputs a hidden layer state with a dimension of 500×64.

[0143] A hybrid attention mechanism is constructed based on the hidden layer states output by each recurrent neural network unit. First, the temporal correlation weights are calculated to weight the time dimension of the hidden layer states; then the feature dimension weights are calculated to weight the feature dimension of the hidden layer states. In the specific implementation process, self-attention scores are calculated for the hidden layer states output by LSTM, GRU, and Simple RNN to generate a temporal weight vector with a dimension of 500×1; at the same time, the feature channel attention weights with a dimension of 1×64 are calculated.

[0144] The temporal correlation weights and the feature dimension weights are input into a multi-layer perceptron to generate a fused attention vector. The multi-layer perceptron consists of three fully connected networks, with the number of neurons being 128, 64, and 64 respectively, and the activation function is ReLU. The dimension of the fused attention vector is the same as that of the hidden layer state, which is 500×64.

[0145] The fused attention vector and the hidden layer state are weighted and fused to obtain an enhanced feature vector. Specifically, the hidden layer states of LSTM, GRU, and Simple RNN are respectively subjected to element-wise multiplication operations with the corresponding fused attention vectors to obtain three enhanced feature vectors, each with a dimension of 500×64.

[0146] A distributed gradient calculation matrix is constructed based on the enhanced feature vectors, the hidden layer states, and the principal component feature vectors. The enhanced feature vectors of the three networks are concatenated in the feature dimension to obtain a feature matrix with a dimension of 500×192; then the original hidden layer states (with a dimension of 500×192, concatenated by the three networks) and the principal component feature vectors (with a dimension of 500×4) are concatenated, and finally a distributed gradient calculation matrix with a dimension of 500×388 is obtained.

[0147] The distributed gradient calculation matrix is processed by a multi-head attention fusion network to determine the three-phase current prediction value of the brushless motor of each joint axis. The multi-head attention network contains 8 attention heads, each with an output dimension of 32. After fusion, the three-phase current prediction value is output through two fully connected layers (dimensions are 128 and 3 respectively). In practical applications, the model can predict the three-phase current of the third axis motor of a six-axis industrial robot for the next 20ms with an accuracy of more than 95%, and the average prediction error is less than 0.2A, which effectively supports the robot control system to perform feedforward compensation control.

[0148] In an optional implementation, a distributed gradient calculation matrix is constructed based on the enhanced feature vector, the hidden layer state and the principal component feature vector, and a three-phase current prediction value of each joint axis brushless motor is determined in combination with a multi-head attention fusion network, including:

[0149] Inputting the enhanced feature vector, the hidden layer state and the principal component feature vector into a multi-objective combined loss function construction module, and constructing a distributed gradient calculation matrix based on time series prediction loss and feature reconstruction loss;

[0150] Outputting a multi-scale learning rate matrix based on the distributed gradient calculation matrix, inputting the multi-scale learning rate matrix into a hybrid parameter optimizer, and the hybrid parameter optimizer updates the parameters of the recurrent neural network prediction model;

[0151] Processing the enhanced feature vector based on the parameters of the updated recurrent neural network prediction model to obtain the current prediction tensor of each sub-network, and using a multi-head attention fusion network to weight the current prediction tensor of each sub-network to obtain the initial three-phase current prediction value of the brushless motor of each joint axis;

[0152] Inputting the initial three-phase current prediction value into a hierarchical prediction result optimization system, outputting an evaluation index vector, and performing time domain and frequency domain correction on the initial three-phase current prediction value based on the evaluation index vector to obtain a corrected three-phase current prediction value;

[0153] Calculating an error matrix between the corrected three-phase current prediction value and the actual three-phase current value, and solving a multi-constraint optimization problem based on the error matrix to obtain an optimal parameter configuration vector;

[0154] The optimal parameter configuration vector is fed back to the corresponding module for parameter update, and the prediction calculation is re-performed based on the updated parameters to output the three-phase current prediction value of the brushless motor of each joint axis.

[0155] In this embodiment, a technical implementation process of constructing a distributed gradient calculation matrix based on enhanced eigenvectors, hidden layer states, and principal component eigenvectors, and determining the three-phase current prediction values of the brushless motors of each joint axis in combination with a multi-head attention fusion network is described in detail.

[0156] Specifically, the time series prediction loss uses the root mean square error calculation method to calculate the difference between the predicted value and the actual value; the feature reconstruction loss uses the structural similarity index to quantify the quality of the feature representation. By setting the weight coefficients α = 0.7 and β = 0.3 to control the contribution ratios of the time series prediction loss and the feature reconstruction loss respectively, the two losses are weighted and combined to form a comprehensive loss function. Based on this comprehensive loss function, the gradient vectors of each network parameter are calculated, and these gradient vectors are organized into a distributed gradient calculation matrix G with a dimension of p×q, where p represents the number of network parameters and q represents the batch size. For example, when the number of parameters is 1024 and the batch size is 32, the dimension of the gradient calculation matrix G is 1024×32.

[0157] Based on the distributed gradient calculation matrix, a multi-scale learning rate matrix is output, and the multi-scale learning rate matrix is input into the hybrid parameter optimizer, which updates the parameters of the recurrent neural network prediction model. The generation of the multi-scale learning rate matrix L adopts an adaptive scaling mechanism, and different learning rates are set for different parameter groups in the gradient calculation matrix G, with the learning rate ranging from 0.0001 to 0.01. For example, for fast-changing features, a larger learning rate of 0.01 is set; for slow-changing features, a smaller learning rate of 0.0001 is set. The hybrid parameter optimizer integrates the advantages of the Adam and RMSprop optimization algorithms, sets the momentum term parameter to 0.9, the second momentum term parameter to 0.999, and the epsilon value to 1e-8, and updates the network parameters through batch gradient descent.

[0158] The recurrent neural network adopts a bidirectional LSTM structure, with the number of hidden layer units set to 128 and the time step set to 20. After the input enhanced feature vector is processed, current prediction tensors of multiple sub-networks are obtained, and each sub-network corresponds to a joint axis. The multi-head attention fusion network weights the current prediction tensors of each sub-network to obtain the initial three-phase current prediction values of the brushless motors of each joint axis. The multi-head attention mechanism sets 8 attention heads, with the dimension of each attention head being 64. By calculating the correlation between the query (Query), key (Key), and value (Value), weights are assigned to the prediction results of different sub-networks. For example, for the current prediction of joint axis 1, the prediction results of the four sub-networks are weighted and fused with the attention weights [0.25, 0.15, 0.35, 0.25].

[0159] Input the initial three-phase current prediction values into the hierarchical prediction result optimization system, and output the evaluation index vector. Based on the evaluation index vector, perform time-domain and frequency-domain correction on the initial three-phase current prediction values to obtain the corrected three-phase current prediction values. The evaluation index vector includes three indicators: root mean square error (RMSE), mean absolute percentage error (MAPE), and peak signal-to-noise ratio (PSNR), corresponding to accuracy, relative error, and signal quality respectively. The time-domain correction uses the moving window average method with the window size set to 5; the frequency-domain correction removes high-frequency noise through wavelet transform with the threshold set to 0.05. For example, for the initial predicted current values [4.2A, 4.5A, 4.7A, 4.6A, 4.3A], the result after time-domain correction is 4.46A; after removing the high-frequency components less than the 0.05 threshold through frequency-domain correction, a smoother current prediction curve is obtained.

[0160] Calculate the error matrix between the corrected three-phase current prediction values and the actual three-phase current values, and solve the multi-constraint optimization problem based on the error matrix to obtain the optimal parameter configuration vector. The dimension of the error matrix E is m×n, where m represents the number of samples and n represents the three-phase current components. Set the optimization goal to minimize the Frobenius norm of the error matrix, and the constraint conditions include that the prediction delay does not exceed 10 milliseconds, the computational complexity does not exceed O(n2), and the storage space does not exceed 100MB. Search for the optimal configuration in the parameter space through the grid search algorithm, with the search step set to 0.05 and the upper limit of the number of iterations set to 200 times. The obtained optimal parameter configuration vector contains parameter values in multiple dimensions such as the learning rate, regularization coefficient, and attention weight.

[0161] Feed back the optimal parameter configuration vector to the corresponding module for parameter update, and re-perform the prediction calculation based on the updated parameters to output the three-phase current prediction values of the brushless motors of each joint axis. For example, for joint axis 2, before parameter update, the three-phase current prediction values are [5.2A, -2.8A, -2.4A], and the average prediction error is 0.31A; after parameter update, the three-phase current prediction values are adjusted to [5.3A, -2.9A, -2.4A], and the average prediction error is reduced to 0.18A, and the prediction accuracy is improved by 41.9%. Through iterative optimization, the system can finally achieve high-precision prediction of the three-phase current of the brushless motors of each joint axis, and the consistency between the predicted value and the actual value reaches more than 95%, meeting the requirements of precise control of the robot.

[0162] In an alternative embodiment, input the three-phase current prediction values into a distributed model predictive controller, and calculate the optimal switching sequence of each joint axis based on the coupling torque between the joint axes as the inverter drive signal of the brushless motor of each joint axis, including:

[0163] Input the predicted three-phase current values into the distributed model predictive controller. The distributed model predictive controller constructs a hierarchical recursive network structure based on the positions, velocities, and accelerations of each joint axis. The hierarchical recursive network structure models the dynamic coupling characteristics between joints and outputs the inter-joint coupling torque model.

[0164] Construct a distributed state predictor based on the inter-joint coupling torque model. The distributed state predictor outputs the joint coupling torque sequence within the prediction time domain.

[0165] Input the joint coupling torque sequence and the predicted three-phase current values into the multi-objective optimizer. The multi-objective optimizer constructs a loss evaluation model based on the conduction time of the power devices.

[0166] Construct a current error constraint based on the predicted three-phase current values, and construct a torque balance constraint based on the joint coupling torque sequence. Construct a receding horizon optimization problem based on the loss evaluation model, the current error constraint, and the torque balance constraint.

[0167] Use the branch and bound algorithm to solve the receding horizon optimization problem to obtain the candidate switching sequences for each joint axis. Construct a parallel evaluator based on the current error constraint and the torque balance constraint. Use the parallel evaluator to perform multi-objective evaluation on the candidate switching sequences, and select the optimal switching sequence for each joint axis as the inverter drive signal for the brushless motor of each joint axis.

[0168] In this embodiment, the predicted three-phase current values include the three-phase current values IA, IB, and IC of the brushless motor of each joint axis. After receiving these current prediction values, the distributed model predictive controller constructs a hierarchical recursive network structure based on the positions, velocities, and accelerations of each joint axis. The design of this network structure is as follows: The first layer contains 6 nodes, corresponding to the position data of 6 joint axes respectively; the second layer contains 6 nodes, corresponding to the velocity data of 6 joint axes respectively; the third layer contains 6 nodes, corresponding to the acceleration data of 6 joint axes respectively. The nodes in each layer are fully connected, and the connection weights are obtained through training with historical data. For example, the initial value of the connection weight between joint 1 and joint 2 is set to 0.85, and then it is adjusted according to the actual operation data.

[0169] When the hierarchical recursive network structure models the dynamic coupling characteristics between joints, it adopts two stages: forward propagation and backward propagation. In the forward propagation stage, each node receives the outputs of all nodes in the previous layer and sums them up with weights. In the backward propagation stage, the weights are adjusted according to the error between the actual torque and the predicted torque. Through repeated iteration, the network finally outputs the inter-joint coupling torque model. For example, when the position of joint 1 is 30 degrees, the velocity is 2 rad / s, and the acceleration is 0.5 rad / s 2 ², the model predicts that the coupling torque of joint 1 on joint 2 is 2.35 Nm.

[0170] A distributed state predictor is constructed based on the inter-joint coupling torque model. This predictor uses a recursive method to calculate the joint states and coupling torques at the next N time instants starting from the current time instant t. In this embodiment, the prediction horizon N is set to 10, and the sampling time interval is set to 2 ms. The state predictor takes the positions, velocities, accelerations, and torques of each joint axis at the current time as inputs and outputs the sequence of joint coupling torques within the prediction horizon. For example, the sequence of coupling torques of joint 1 on joint 2 at the next 10 time instants may be [2.35, 2.41, 2.46, 2.50, 2.53, 2.55, 2.56, 2.57, 2.57, 2.56] Nm.

[0171] The multi-objective optimizer first constructs a loss evaluation model based on the conduction time of the power device. This model takes into account two parts: the conduction loss and the switching loss of the power device. The conduction loss is proportional to the square of the current and the conduction time, and the switching loss is proportional to the switching frequency. In practical applications, taking the IGBT power module CM75DU-12F as an example, its on-resistance is 8.5 mΩ, and the switching energy loss is 0.35 mJ per switching. Based on these parameters, the energy loss in different switching states can be calculated. For example, when the three-phase currents are [15 A, -8 A, -7 A] and the switching frequency is 5 kHz, the energy loss per switching cycle is approximately 0.85 J.

[0172] The current error constraint ensures that the error between the actual current and the target current does not exceed a preset threshold. In this embodiment, the current error threshold is set to 5% of the rated current. That is, for a system with a rated current of 20 A, the maximum allowable error is 1 A. At the same time, a torque balance constraint is constructed based on the sequence of joint coupling torques to ensure that the torques between joints are balanced with each other, avoiding system vibration or instability caused by torque imbalance. The torque balance constraint is expressed as the sum of the coupling torques between joint axes should be equal to the sum of the external torques, and the allowable imbalance does not exceed 5% of the rated torque.

[0173] Based on the loss evaluation model, the current error constraint, and the torque balance constraint, a rolling horizon optimization problem is constructed. The objective of this optimization problem is to find the switching sequence with the minimum energy loss under the conditions of satisfying the current error constraint and the torque balance constraint. The optimization variables are the switching states at the next N time instants, the constraint conditions are the above two constraints, and the objective function is the weighted sum of the energy losses.

[0174] The algorithm starts from the root node, divides the search space into multiple sub-spaces, prunes by calculating the upper and lower bounds of each sub-space, and finally finds the optimal solution or approximate optimal solution. In practical applications, the maximum number of iterations is set to 100 and the convergence threshold is set to 0.01. The algorithm outputs the candidate switch sequences of each joint axis. For example, the candidate switch sequence of joint 1 is [101, 100, 100, 110, 110, 010, 010, 011, 001, 101], where every three digits represent the switch states of the three upper bridge arms at a moment.

[0175] A parallel evaluator is constructed based on the current error constraint and torque balance constraint, and this evaluator is used to perform multi-objective evaluation on the candidate switch sequences. The evaluation indicators include three aspects: current tracking error, torque balance degree, and energy loss, with weights of 0.4, 0.4, and 0.2 respectively. Through comprehensive scoring, the optimal switch sequence of each joint axis is selected as the inverter drive signal of the brushless motor of each joint axis. For example, after evaluation, the optimal switch sequence of joint 1 is [101, 100, 110, 110, 010, 010, 011, 001, 001, 101]. This sequence not only ensures the current tracking accuracy, but also realizes the torque balance between joints, and at the same time minimizes the energy loss.

[0176] Through the above method, the coordinated control of the multi-joint robot is realized, and the dynamic performance and energy efficiency of the system are improved.

[0177] The brushless motor multi-axis cooperative control system for an industrial robot in the embodiment of the present invention includes:

[0178] The first unit is used to construct an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtain the temperature drift coefficient and load inertia parameters of the brushless motors of each joint axis;

[0179] The second unit is used to input the temperature drift coefficient and load inertia parameters into the adaptive pre-compensation model, and calculate the system disturbance amount and fractional-order sliding mode surface in combination with the extended dynamic model;

[0180] The third unit is used to input the system disturbance amount and fractional-order sliding mode surface into the neural network compensator. The neural network compensator processes the system disturbance amount and fractional-order sliding mode surface using the backpropagation algorithm and outputs a torque compensation signal;

[0181] The fourth unit is used to superimpose the torque compensation signal and the compensation torque to obtain a torque control instruction, and convert the torque control instruction into the three-phase current given values of the brushless motors of each joint axis;

[0182] A fifth unit for inputting the three-phase current set value and the phase current signal into a recurrent neural network prediction model, and the recurrent neural network prediction model outputs the predicted three-phase current values of the brushless motors of each joint axis;

[0183] A sixth unit for inputting the predicted three-phase current values into a distributed model predictive controller, and calculating an optimal switching sequence for each joint axis as the inverter drive signal of the brushless motor of each joint axis based on the coupling torque between the joint axes.

[0184] In a third aspect of the embodiments of the present invention, an electronic device is provided, including:

[0185] A processor;

[0186] A memory for storing instructions executable by the processor;

[0187] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0188] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0189] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are loaded.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Brushless motor multi-axis collaborative control method for industrial robots, characterized in that Including: Construct an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtain the temperature drift coefficient and load inertia parameters of the brushless motors of each joint axis; Input the temperature drift coefficient and load inertia parameters into the adaptive pre-compensation model, and calculate the system disturbance quantity and fractional-order sliding mode surface in combination with the extended dynamic model; Input the system disturbance quantity and fractional-order sliding mode surface into the neural network compensator. The neural network compensator processes the system disturbance quantity and fractional-order sliding mode surface using the backpropagation algorithm and outputs a torque compensation signal; Superimpose the torque compensation signal and the compensation torque to obtain a torque control instruction, and convert the torque control instruction into the three-phase current given value of the brushless motors of each joint axis; Input the three-phase current given value and the phase current signal into the recurrent neural network prediction model, and the recurrent neural network prediction model outputs the three-phase current prediction values of the brushless motors of each joint axis; Input the three-phase current prediction values into the distributed model predictive controller, and calculate the optimal switching sequence of each joint axis based on the coupling torque between the joint axes as the inverter drive signal of the brushless motors of each joint axis.

2. The method according to claim 1, wherein Construct an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtain the temperature drift coefficient and load inertia parameters of the brushless motors of each joint axis, including: Construct a winding resistance temperature coefficient model and a permanent magnet remanence density temperature coefficient model, and calculate a temperature-related torque constant model according to the winding resistance temperature coefficient model and the permanent magnet remanence density temperature coefficient model; Based on the temperature-related torque constant model and the real-time operating parameters, construct an extended dynamic model. Input the real-time operating parameters into the extended dynamic model to calculate the parameter estimation matrix, and calculate the parameter identification error according to the parameter estimation matrix; Construct an adaptive forgetting factor based on the parameter identification error, and update the parameter estimation matrix and covariance matrix based on the adaptive forgetting factor; Construct a state space model, substitute the parameter estimation matrix and the covariance matrix into the Kalman filter for parameter prediction, and obtain the parameter prediction value and the predicted covariance matrix; Calculate the Kalman gain according to the parameter prediction value and the predicted covariance matrix, and use the Kalman gain to correct the parameters with the temperature-related torque constant model to obtain the corrected state estimation value; Establish parameter constraint conditions based on the corrected state estimation value, and calculate the temperature drift coefficient and load inertia parameters of the brushless motors of each joint axis according to the parameter constraint conditions.

3. The method according to claim 1, wherein Input the system disturbance quantity and fractional-order sliding mode surface into the neural network compensator. The neural network compensator processes the system disturbance quantity and fractional-order sliding mode surface using the backpropagation algorithm and outputs a torque compensation signal, including: Construct a composite fractional-order differential expression based on the system disturbance quantity and the fractional-order sliding mode surface, and input the composite fractional-order differential expression into a group of nonlinear disturbance observers. The group of nonlinear disturbance observers calculates the subsystem disturbance quantity based on a hybrid adaptive time-varying gain matrix and a multi-nested high-order observer state equation; Construct a hierarchical recursive disturbance prediction model according to the subsystem disturbance quantity. The hierarchical recursive disturbance prediction model calculates the system comprehensive disturbance quantity by using the weighted moving average method; Calculate a multi-dimensional disturbance prediction error vector based on the system comprehensive disturbance quantity and the subsystem disturbance quantity, update the hybrid adaptive time-varying gain matrix according to the multi-dimensional disturbance prediction error vector, and feedback the updated hybrid adaptive time-varying gain matrix to the group of nonlinear disturbance observers; Input the system disturbance quantity, the fractional-order sliding mode surface, the system comprehensive disturbance quantity, and the multi-dimensional disturbance prediction error vector into a neural network compensator. The neural network compensator performs multi-level signal processing on the input signals to generate multi-level output signals; Construct a multi-objective composite error function based on the multi-dimensional disturbance prediction error vector and the multi-level output signals; Construct a recursive network structure evaluation model based on the multi-objective composite error function, and generate a determined torque compensation signal according to the recursive network structure evaluation model.

4. The method according to claim 3, wherein Construct a recursive network structure evaluation model based on the multi-objective composite error function, and generate a determined torque compensation signal according to the recursive network structure evaluation model, including: Construct a recursive network structure evaluation model based on the multi-objective composite error function. The recursive network structure evaluation model calculates a neuron importance index matrix and a network topology feature vector; Construct a two-way propagation mechanism based on the network topology feature vector. The forward channel of the two-way propagation mechanism constructs an adaptive pruning criterion based on the neuron importance index matrix to generate optimized neuron connection weights. The reverse channel of the two-way propagation mechanism combines the optimized neuron connection weights with the neuron importance index matrix to construct a new constraint term and updates the multi-objective composite error function; Input the neuron importance index matrix and the optimized neuron connection weights into a learning parameter generation module to generate an adaptive learning rate matrix; Construct a dynamic momentum factor matrix based on the updated multi-objective composite error function and the network topology feature vector, and adjust the dynamic momentum factor matrix in combination with the neuron importance index matrix to generate optimized momentum update parameters; Input the optimized momentum update parameters and the dynamic momentum factor matrix into a state calculation module to calculate the output state signals of each hidden layer, and determine the torque compensation signal based on the output state signals of each hidden layer.

5. The method according to claim 1, characterized in that, Input the three-phase current given value and the phase current signal into a recurrent neural network prediction model. The recurrent neural network prediction model outputs the three-phase current prediction values of the brushless motors of each joint axis, including: Input the three-phase current given value and the phase current signal into the recurrent neural network prediction model. The recurrent neural network prediction model uses wavelet packet decomposition to perform multi-layer decomposition on the phase current signal to obtain a filtered signal; Perform normalization processing on the three-phase current given value and the filtered signal to obtain normalized features, and construct a time series sample matrix according to an adaptive time window for the normalized features; Extract the principal component feature vector of the time series sample matrix, and use different types of recurrent neural network units to process the principal component feature vector to output corresponding hidden layer states; Construct a hybrid attention mechanism based on the hidden layer states output by each recurrent neural network unit, and calculate the time series correlation weight and feature dimension weight of the hidden layer states based on the hybrid attention mechanism; Input the time series correlation weight and the feature dimension weight into a multi-layer perceptron to generate a fused attention vector, and perform weighted fusion of the fused attention vector and the hidden layer states to obtain an enhanced feature vector; Construct a distributed gradient calculation matrix based on the enhanced feature vector, the hidden layer states, and the principal component feature vector, and combine a multi-head attention fusion network to determine the three-phase current prediction values of the brushless motors of each joint axis.

6. The method according to claim 5, wherein Construct a distributed gradient calculation matrix based on the enhanced feature vector, the hidden layer states, and the principal component feature vector, and combine a multi-head attention fusion network to determine the three-phase current prediction values of the brushless motors of each joint axis, including: Input the enhanced feature vector, the hidden layer states, and the principal component feature vector into a multi-objective combined loss function construction module, and construct a distributed gradient calculation matrix based on time series prediction loss and feature reconstruction loss; Output a multi-scale learning rate matrix based on the distributed gradient calculation matrix, input the multi-scale learning rate matrix into a hybrid type parameter optimizer, and the hybrid type parameter optimizer updates the parameters of the recurrent neural network prediction model; Process the enhanced feature vector based on the parameters of the updated recurrent neural network prediction model to obtain the current prediction tensors of each sub-network, and use a multi-head attention fusion network to weight the current prediction tensors of each sub-network to obtain the initial three-phase current prediction values of the brushless motors of each joint axis; Input the initial three-phase current prediction values into a hierarchical prediction result optimization system to output an evaluation index vector, and perform time domain and frequency domain correction on the initial three-phase current prediction values based on the evaluation index vector to obtain the corrected three-phase current prediction values; Calculate the error matrix between the corrected three-phase current prediction values and the actual three-phase current values, and solve a multi-constraint optimization problem based on the error matrix to obtain an optimal parameter configuration vector; Feedback the optimal parameter configuration vector to the corresponding module for parameter update, and re-perform prediction calculation based on the updated parameters to output the three-phase current prediction values of the brushless motors of each joint axis.

7. The method according to claim 1, characterized in that, Input the three-phase current prediction values into a distributed model predictive controller, and calculate the optimal switching sequence of each joint axis based on the coupling torque between the joint axes as the inverter drive signal of the brushless motors of each joint axis, including: Input the predicted three-phase current values into the distributed model predictive controller. The distributed model predictive controller constructs a hierarchical recursive network structure based on the positions, speeds, and accelerations of each joint axis. The hierarchical recursive network structure models the dynamic coupling characteristics between joints and outputs the inter-joint coupling torque model. Construct a distributed state predictor based on the inter-joint coupling torque model. The distributed state predictor outputs the sequence of joint coupling torques within the prediction time domain. Input the sequence of joint coupling torques and the predicted three-phase current values into the multi-objective optimizer. The multi-objective optimizer constructs a loss evaluation model based on the conduction time of power devices. Construct a current error constraint based on the predicted three-phase current values, and construct a torque balance constraint based on the sequence of joint coupling torques. Construct a receding horizon optimization problem based on the loss evaluation model, the current error constraint, and the torque balance constraint. Use the branch and bound algorithm to solve the receding horizon optimization problem to obtain the candidate switching sequences for each joint axis. Construct a parallel evaluator based on the current error constraint and the torque balance constraint. Use the parallel evaluator to perform multi-objective evaluation on the candidate switching sequences, and select the optimal switching sequence for each joint axis as the inverter drive signal for the brushless motor of each joint axis.

8. A brushless motor multi-axis collaborative control system for industrial robots, which is used to implement the method described in any one of claims 1-7, characterized in that, Includes: The first unit is used to construct an extended dynamic model to process the real-time operating parameters of the brushless motors of each joint axis of the industrial robot, and obtain the temperature drift coefficient and load inertia parameters of the brushless motors of each joint axis. The second unit is used to input the temperature drift coefficient and load inertia parameters into the adaptive pre-compensation model, and calculate the system disturbance quantity and the fractional-order sliding mode surface in combination with the extended dynamic model. The third unit is used to input the system disturbance quantity and the fractional-order sliding mode surface into the neural network compensator. The neural network compensator uses the backpropagation algorithm to process the system disturbance quantity and the fractional-order sliding mode surface and outputs a torque compensation signal. The fourth unit is used to superimpose the torque compensation signal and the compensation torque to obtain a torque control command, and convert the torque control command into the given value of the three-phase current of the brushless motor of each joint axis. The fifth unit is used to input the given value of the three-phase current and the phase current signal into the recurrent neural network prediction model. The recurrent neural network prediction model outputs the predicted three-phase current values of the brushless motors of each joint axis. The sixth unit is used to input the predicted three-phase current values into the distributed model predictive controller, and calculate the optimal switching sequence for each joint axis based on the coupling torque between the joint axes as the inverter drive signal for the brushless motor of each joint axis.

9. An electronic device, characterized in that, Includes: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images