A cnn-ga-based active disturbance rejection controller parameter optimization method

Abstract

The application relates to a CNN-GA-based active disturbance rejection controller parameter optimization method, which comprises the following steps: an active disturbance rejection controller simulation model is built, a convolutional neural network model and a genetic algorithm model are constructed, the response speed of the active disturbance rejection control to a target and the fitness of the active disturbance rejection control to different research objects are taken as optimization targets, a deep deterministic policy gradient method is designed to make the parameter optimization process converge, the training of the active disturbance rejection controller parameter optimization model based on the convolutional neural network is completed, and the optimal parameters are obtained. By adopting the parameter optimization scheme, the problems of the active disturbance rejection controller, such as many parameters, strong coupling and difficult debugging, can be solved at the minimum artificial debugging cost.

Description

Technical Field

[0001] This invention relates to the field of control technology, and specifically to a method for optimizing the parameters of an active disturbance rejection controller based on CNN-GA. Background Technology

[0002] In the traditional field of control, PID control technology remains the mainstream. However, in recent years, the emergence of various new control methods, such as active disturbance rejection control, fuzzy control, and sliding mode control, has effectively compensated for the shortcomings of PID control. Active disturbance rejection controllers, as a popular motor control technology, are widely researched and applied in various fields of automation control and industrial production.

[0003] Active disturbance rejection controllers (ADRCs) have a large number of internal parameters, strong coupling, low sensitivity, and nonlinear characteristics. As a result, the parameters of ADRCs are generally adjusted manually and based on experience. This manual method of parameter optimization is time-consuming, inefficient, and cannot bring out the controller's optimal performance.

[0004] In recent years, some scholars have proposed using particle swarm optimization or genetic algorithm (GA) to optimize the parameters of active disturbance rejection controllers, replacing the manual optimization method. However, existing optimization methods are mainly based on a single algorithm, and there is still room for further optimization in terms of robustness and accuracy. In view of this, the inventors proposed an active disturbance rejection controller parameter optimization method based on CNN-GA (CNN stands for convolutional neural network), which can further improve prediction accuracy, save optimization time, and have stronger overall robustness. This case is based on this. Summary of the Invention

[0005] This invention discloses a method for optimizing the parameters of an active disturbance rejection controller based on CNN-GA, comprising the following steps:

[0006] Step 1: Collect data on the controlled object's response to the target and data from each part of the active disturbance rejection controller, which will be used as input and output data for the optimization model, respectively;

[0007] Step 2: Preprocess the collected raw data to remove invalid data and obtain a preprocessed dataset;

[0008] Step 3: Construct a convolutional neural network model by dividing the data in the preprocessed dataset into a training set and a test set;

[0009] Step 4: Normalize the training set data and output the normalized training set input transformed data and training set output transformed data;

[0010] Step 5: Construct a parameter optimization model for the active disturbance rejection controller based on a convolutional neural network, i.e., model net;

[0011] Step 6: Construct a genetic algorithm model, i.e., GA, and use the GA model to optimize the initial weight values ​​and thresholds of the model net;

[0012] Step 7: Assign the optimized weight values ​​and threshold values ​​from the genetic algorithm to the model net;

[0013] Step 8: Input the normalized training set input transformation data and the training set output transformation data to train the model net;

[0014] Step 9: Normalize the test set data and output the normalized test set input transformation data and test set output transformation data;

[0015] Step 10: Input the normalized test set into the transformed data and test the trained model net to obtain the encoded test results. Decode the obtained test results to obtain the test results of model net.

[0016] Step 11: Subtract the test results of model net from the normalized test set output data to obtain the test error of model net. Adjust the weight values ​​and thresholds of model net based on the test error to obtain the parameter optimization model of the active disturbance rejection controller of the convolutional neural network optimized by the genetic algorithm.

[0017] Furthermore, in step 1, it is necessary to establish an active disturbance rejection controller model. The active disturbance rejection controller consists of three parts: a tracking differentiator, an extended state observer, and a nonlinear state error feedback. The extended state observer is used as the optimization objective, and the optimization parameters are determined to be the three output quantities z1, z2, and z3 of the extended state observer. Among them, z1 and z2 are the state variable estimates of the active disturbance rejection controller model, and z3 is the global variable estimate of the active disturbance rejection controller model.

[0018] Furthermore, the model of the tracking differentiator is as follows:

[0019]

[0020] The model for the extended state observer is as follows:

[0021]

[0022] The model for the nonlinear state error feedback is as follows:

[0023] e1 = v1 - z1, e2 = v2 - z2

[0024] u=-fhan(e1, ce2, r, n).

[0025] Furthermore, in steps 4 and 9, the mapminmax function is used for normalization to normalize the input and output data of the training set and the test set to between -1 and 1.

[0026] Furthermore, the specific steps for constructing the model net in step 5 are as follows:

[0027] Step 5.1: Construct a feature extraction module, which includes convolutional layers, unfolded layers, and fully connected layers. The convolutional layers use one-dimensional convolution for computation.

[0028] The calculation expression for a single sample convolutional layer is Y = f((m,k)*W1 + b1), where (m,k) is the input matrix, m represents the number of samples, k represents the number of channels, W1 is an a×b×k×s weight matrix, a×b is defined as the kernel size, s is defined as the number of kernels, b1 is the bias term, f is the activation function relu:max(0,x), and Y is the output of the convolutional layer.

[0029] The expression for calculating the expanded layer is Y. flat =Reshape(Y,size), where Y flat The output of the unfolded layer is Reshape, where Reshape is the unfolding function and size represents the dimension of the transformation result.

[0030] The calculation expression for a fully connected layer is Z = f(W² * Y). flat +b2), where W2 is the weight matrix, b2 is the bias term, and Z is the output of the fully connected layer;

[0031] Step 5.2: Construct a regression prediction module. The regression prediction module includes a Drop layer and a regression layer. The Drop layer adds a regularization term to the loss function and randomly discards some data and fills it with 0.

[0032] The Drop layer expression is Z. drop =Drop(Z,keep_prop), where Z drop For the output of the Drop layer, keep_prop is the proportion of data to be retained out of the total data, and its value ranges from 0 to 1;

[0033] The regression layer expression is O = g(W3*Z) drop +b3), where W3 is the weight matrix, b3 is the bias term, g is the activation function sigmoid:1 / (1+exp(x)), and O is the output of the regression layer.

[0034] Furthermore, the specific steps for constructing the genetic algorithm model in step 6 are as follows:

[0035] Step 6.1: Set the parameters of the genetic algorithm based on the preprocessed dataset. The parameters include the number of iterations maxgen, the population size sizepop, the crossover probability pcross, and the mutation probability pmutation.

[0036] Step 6.2: Determine the total number of nodes numsum in the model net, set the chromosome length lenchrom and the boundary value bound, and make the chromosome length lenchrom equal to the total number of nodes numsum;

[0037] Step 6.3: Input the chromosome length lenchrom and boundary value bound from step 6.2 into the function code, and randomly generate chromosome individuals to form the population individuals;

[0038] Step 6.4: Input the information of individual chromosomes into the model net, and then input the encoded chromosome information and the information of the model net into the fitness calculation function fun to obtain the corresponding fitness individuals.fitness; find the chromosome with the best fitness in each generation and record the chromosome information, fitness value bestfitness, and average fitness avgfitness.

[0039] Step 6.5: Input the population information individuals and the population size sizepop into the selection function select, and output the new population individuals after selection;

[0040] Step 6.6: Input the crossover probability pcross, chromosome length lenchrom, chromosome information chrom, and population size sizepop into the crossover function cross. Select the chromosomes to be crossovered from the new population individuals obtained from the selection operation according to the crossover operation probability pcross. Randomly pair them up and exchange some genes to obtain the crossover chromosome individual chrom1.

[0041] Step 6.7: Input the mutation probability pmutation, chromosome length lenchrom, chromosome information chrom, population size sizepop, current iteration count num, maximum iteration count maxgen, and individual boundary bound into the mutation function Mutation. The mutation function Mutation selects the chromosomes to be mutated from the crossover chromosome chrom1 according to the mutation operation probability pmutation, and then mutates a certain gene on these chromosomes, outputting the mutated chromosome chrom2. The chromosomes to be mutated are randomly selected, and the mutation position is also randomly selected.

[0042] Step 6.8: Replacement Operation: Input the chromosome population after mutation operation into the fitness function fun one by one, recalculate the fitness, select the best individual, output and record the fitness value of the chromosome individuals.fitness, find the chromosomes with the minimum and maximum fitness and their positions in the population, and replace the chromosome with the maximum fitness with the chromosome with the minimum fitness value; record the best fitness, average fitness and fitness of all chromosomes in the population in each generation;

[0043] Step 6.9: Return to step 6.5 and proceed to the next iteration; once the maximum number of iterations maxgen is reached, end the iterative training and obtain the chromosome x with the highest fitness.

[0044] Step 6.10: Decode the best chromosome x and output the optimal initial weights and thresholds.

[0045] Furthermore, in step 6.5, the selection method in the select function adopts the roulette wheel method.

[0046] This invention utilizes neural network operations to construct a parameter optimization model for an active disturbance rejection controller based on convolutional neural networks and genetic algorithms. This model can adapt to the nonlinear correlation and coupling characteristics between input and output data, resulting in a model with small deviations and high prediction accuracy. This avoids spending a lot of time adjusting parameters, saving parameter optimization time, improving parameter optimization efficiency, and enhancing overall robustness. Attached Figure Description

[0047] Figure 1 This is a flowchart of the active disturbance rejection controller parameter optimization method based on CNN-GA in the embodiment;

[0048] Figure 2 This is a block diagram of the active disturbance rejection controller mentioned in this invention;

[0049] Figure 3 This is a flowchart of the genetic algorithm in the optimization method of this invention;

[0050] Figure 4 The graph shows the comparative effects of using traditional PID control, active disturbance rejection control without parameter optimization, and control according to the optimization method of this invention for a specific research object.

[0051] Figure 5 For replacement and Figure 4 After different research objects, the comparison results of traditional PID control, active disturbance rejection control without parameter optimization, and control according to the optimization method of this invention are shown in the figure. Detailed Implementation

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0053] This embodiment discloses a method for optimizing the parameters of an active disturbance rejection controller based on CNN-GA, such as... Figure 1 As shown, this parameter optimization method mainly includes establishing an active disturbance rejection controller (ADRC) model, building a convolutional neural network (CNN) model and a genetic algorithm model, designing a deep deterministic policy gradient method to converge the parameter optimization process, and completing the training of the ADRC parameter optimization model based on the CNN and genetic algorithm. Specifically, the parameter optimization method includes the following steps:

[0054] Step 1: Establish an active disturbance rejection controller model, select the parameters to be optimized based on the actual model, and set the optimization target.

[0055] Specifically, the active disturbance rejection controller (ADRC) comprises three parts: a tracking differentiator (TD), an extended state observer (ESO), and a nonlinear state error feedback. The tracking differentiator (TD) tracks the input signal and manages transient states, ensuring fast control without overshoot. The extended state observer (ESO) observes the output and its derivatives, treating both internal and external disturbances as a total disturbance and estimating it. The nonlinear state error feedback (NLSEF) primarily controls the nonlinear combination of the reference input and the extended state error signal, while also compensating for the observed total disturbance.

[0056] The model of the tracking differentiator is as follows:

[0057]

[0058] When the input signal is disturbed by noise, denoted by n(t), the following relationship holds:

[0059]

[0060] It is also known that the smaller the constant T, the greater the noise interference. Since the accuracy of the differential signal is related to the time constant T, traditional methods struggle to resolve this contradiction. Therefore, we attempt to rewrite the above equation in the following alternative form:

[0061]

[0062] Verification has shown that this type of differentiator is superior to traditional differentiators in suppressing noise interference because it not only provides a differential signal when processing the input signal but also tracks the signal, thus obtaining a highly accurate differential signal.

[0063] The model for the extended state observer is as follows:

[0064]

[0065] The fal function is a special power function y = |x| α The modified sgn(x) has the following specific form:

[0066]

[0067] The nonlinear state error feedback model is as follows:

[0068] e1 = v1 - z1, e2 = v2 - z2

[0069] u = -fhan(e1, ce2, r, n)

[0070] Combination Figure 2 The principle of the active disturbance rejection controller is explained as shown below: Figure 2 The left side ν represents the external input signal to the system. After passing through the tracking differentiator (TD), the input signal becomes two parts: ν1 represents the tracking signal of v, and ν2 represents the differential signal of v. TD can not only quickly track the input signal v, but also reduce the influence of noise signal (denoted by n(t)) on the differential signal ν2, thus obtaining a low-impurity, high-quality differential signal. The extended state observer (ESO) obtains the input signal from the system's output signal and performs state estimation by analyzing the output signal, giving three output quantities: z1, z2, and z3. z1 and z2 are estimates of the model's state variables, and z3 is an estimate of the model's overall variables. The ESO compares the tracking signal v1 with z1 and the differential signal v2 with z2 to obtain two sets of error signals e1 and e2. The nonlinear state error feedback (NLSEF) nonlinearly combines these two sets of error signals e1 and e2 to obtain the initial control value U0. The final control value U is obtained by multiplying the total disturbance variable obtained by the ESO by the coefficient 1 / b0 and subtracting the initial control value U0. t It can be seen that the most important aspect of active disturbance rejection control is the estimation of the system state, so the main focus is on optimizing the three parameters of the extended state observer.

[0071] Step 2: Collect data on the controlled object's response to the target and data from each part of the active disturbance rejection controller, which will be used as input and output data for the optimization model, respectively.

[0072] Step 3: Preprocess the collected raw data to remove invalid data, resulting in a preprocessed dataset of 100 sets. Invalid data refers to abnormal data caused by errors during system operation, resulting from accidental factors during operation.

[0073] Step 4: Construct a convolutional neural network model by randomly dividing the data in the preprocessed dataset into training and test sets.

[0074] Step 5: Use the mapminmax function to normalize the training set data, normalizing both the input and output data of the training set to between -1 and 1, and output the normalized training set input data (input train) and output data (output train).

[0075] Step 6: Construct a parameter optimization model for an active disturbance rejection controller based on a convolutional neural network, referred to as ModelNet.

[0076] Specifically, the steps to construct the model net are as follows:

[0077] (1) Construct a feature extraction module, which includes a convolutional layer, an unfolded layer and a fully connected layer. The convolutional layer uses one-dimensional convolution for calculation.

[0078] The expression for calculating the convolutional layer for a single sample is Y = f((m,k)*W1 + b1), where (m,k) is the input matrix, m represents the number of samples, k represents the number of channels, W1 is the a×b×k×s weight matrix, a×b is defined as the kernel size, s is defined as the number of kernels, b1 is the bias term, f is the activation function relu:max(0,x), and Y is the output of the convolutional layer.

[0079] The expression for calculating the expanded layer is Y. flat =Reshape(Y,size), where Y flat The output of the unfolded layer is Reshape, which is the unfolding function, and size represents the dimension of the transformation result.

[0080] The calculation expression for a fully connected layer is Z = f(W² * Y). flat +b2), where W2 is the weight matrix, b2 is the bias term, and Z is the output of the fully connected layer.

[0081] (2) Construct a regression prediction module. The regression prediction module includes a Drop layer and a regression layer. The Drop layer adds a regularization term to the loss function and randomly discards some data and fills it with 0.

[0082] The Drop layer expression is Z. drop =Drop(Z,keep_prop), where Z dropFor the Drop layer output, keep_prop is the proportion of data to be retained out of the total data, and its value ranges from 0 to 1.

[0083] The regression layer expression is O = g(W3*Z) drop +b3), where W3 is the weight matrix, b3 is the bias term, g is the activation function sigmoid:1 / (1+exp(x)), and O is the output of the regression layer.

[0084] Step 7: Build as follows Figure 3 The genetic algorithm model shown, abbreviated as GA, is used to optimize the initial weight values ​​and thresholds of the model net.

[0085] Specifically, the steps for constructing a genetic algorithm model are as follows:

[0086] (1) Set the parameters of the genetic algorithm according to the preprocessed dataset. The parameters include the number of iterations maxgen, the population size sizepop, the crossover probability pcross, and the mutation probability pmutation.

[0087] (2) Determine the total number of nodes numsum of the model net, set the chromosome length lenchrom and the boundary value bound, and the chromosome length lenchrom is equal to the total number of nodes numsum;

[0088] (3) Input the chromosome length lenchrom and boundary value bound from step (2) into the function code, and randomly generated chromosome individuals form the population individuals;

[0089] (4) Input the information of individual chromosomes into the model net, and then input the encoded chromosome information and the information of the model net into the fitness calculation function fun to obtain the corresponding fitness individuals.fitness; find the chromosome with the best fitness in each generation and record the chromosome information, fitness value bestfitness, and average fitness avgfitness.

[0090] (5) Input the population information individualss and the population size sizepop into the selection function select, and output the new population individuals after selection; the selection method in the selection function select adopts the roulette wheel selection method;

[0091] (6) Input the crossover probability pcross, chromosome length lenchrom, chromosome information chrom, and population size sizepop into the crossover function cross. Select the chromosomes to be crossovered from the new population individuals obtained from the selection operation according to the crossover operation probability pcross. Randomly pair them up and exchange some genes to obtain the crossover chromosome individual chrom1.

[0092] (7) Input the mutation probability pmutation, chromosome length lenchrom, chromosome information chrom, population size sizepop, current iteration number num, maximum iteration number maxgen, and individual boundary bound into the mutation function Mutation. Mutation selects the chromosomes that need to be mutated from the crossover chromosome chrom1 according to the mutation operation probability pmutation, and then mutates a certain gene on these chromosomes, outputting the mutated chromosome chrom2. The chromosomes to be mutated are randomly selected, and the mutation position is also randomly selected.

[0093] (8) Replacement operation: Input the chromosome population after mutation operation into the fitness function fun one by one, recalculate the fitness, select the best individual, output and record the fitness value of the chromosome individuals.fitness, find the chromosome with the smallest and largest fitness and their position in the population, replace the chromosome with the largest fitness with the chromosome with the smallest fitness value; record the best fitness, average fitness and fitness of all chromosomes in the population in each generation;

[0094] (9) Return to step (5) and proceed to the next iteration; once the maximum number of iterations maxgen is reached, end the iterative training and obtain the chromosome x with the highest fitness.

[0095] (10) Decode the best chromosome x and output the optimal initial weights and thresholds.

[0096] Step 8: Assign the optimized weight values ​​and threshold values ​​from the genetic algorithm to the model net.

[0097] Step 9: Train the model net: Set the training parameters, number of iterations 100, learning efficiency 0.05, and target accuracy 0.0001; use the train function to input the normalized input data input train and the output data output train to complete the training of the neural network.

[0098] Step 10: Use the mapminmax function to normalize the test set data, normalizing both the input and output data of the test set to between -1 and 1, and output the normalized test set input data input test and output data output test.

[0099] Step 11: Test the model net. Use the sim function to input the normalized test set into the transformed data input test and substitute it into the trained model net to test the model and obtain the encoded test result, denoted as an. Decode the obtained test result and record the decoded result as test_simu, which is the test result of the model net.

[0100] Step 12: Subtract the test result test_simu from the output transformed data output test of the test set to obtain the test error of model net. Adjust the weight values ​​and thresholds of model net based on the test error to obtain the parameter optimization model of the active disturbance rejection controller of the convolutional neural network optimized by the genetic algorithm.

[0101] To verify the effectiveness of the parameter optimization method of this invention, for a specific research object, the effects of traditional PID control, active disturbance rejection control without parameter optimization, and parameter-optimized control according to the method of this invention on the target response were compared. The comparison results are as follows: Figure 4 As shown in the figure, the PID curve represents the effect curve of the traditional PID control method, the ADRC curve represents the effect curve of the active disturbance rejection control without parameter optimization, and the CNN+GA+ADRC curve represents the effect curve of the optimization method of this invention. Figure 4 The results show that, for the same research object model, the optimized method of this invention significantly outperforms PID control and unoptimized active disturbance rejection control.

[0102] To further verify the generalizability of the parameter optimization method provided by this invention, the method was modified by replacing... Figure 4 After studying different but similar research objects, the effects of traditional PID control, active disturbance rejection control without parameter optimization, and parameter-optimized control according to the method of this invention on the target response were compared again. The comparison results are as follows: Figure 5 As shown, the results still demonstrate that, even after changing different research objects, the optimized performance of the method of the present invention is still significantly better than PID control and unoptimized active disturbance rejection control, verifying the generalization of the model of the present invention.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A CNN-GA based active disturbance rejection controller parameter optimization method, characterized in that, Includes the following steps: Step 1: Collect data on the controlled object's response to the target and data from each part of the active disturbance rejection controller, which will be used as input and output data for the optimization model, respectively; In step 1, it is necessary to establish an active disturbance rejection controller model. The active disturbance rejection controller consists of three parts: a tracking differentiator, an extended state observer, and a nonlinear state error feedback. The extended state observer is used as the optimization objective, and the optimization parameters are determined to be the three output quantities z1, z2, and z3 of the extended state observer. Among them, z1 and z2 are the state variable estimates of the active disturbance rejection controller model, and z3 is the global variable estimate of the active disturbance rejection controller model. Step 2: Preprocess the collected raw data to remove invalid data and obtain a preprocessed dataset; Step 3: Construct a convolutional neural network model by dividing the data in the preprocessed dataset into a training set and a test set; Step 4: Normalize the training set data and output the normalized training set input transformed data and training set output transformed data; Step 5: Construct a parameter optimization model for the active disturbance rejection controller based on a convolutional neural network, i.e., model net; Step 6: Construct a genetic algorithm model, i.e., GA, and use the GA model to optimize the initial weight values ​​and thresholds of the model net; Step 7: Assign the optimized weight values ​​and threshold values ​​from the genetic algorithm to the model net; Step 8: Input the normalized training set input transformation data and the training set output transformation data to train the model net; Step 9: Normalize the test set data and output the normalized test set input transformation data and test set output transformation data; Step 10: Input the normalized test set into the transformed data and test the trained model net to obtain the encoded test results. Decode the obtained test results to obtain the test results of model net. Step 11: Subtract the test results of model net from the normalized test set output data to obtain the test error of model net. Adjust the weight values ​​and thresholds of model net based on the test error to obtain the parameter optimization model of the active disturbance rejection controller of the convolutional neural network optimized by the genetic algorithm.

2. The CNN-GA based active disturbance rejection controller parameter optimization method according to claim 1, characterized in that: The model of the tracking differentiator is as follows: The model for the extended state observer is as follows: The model for the nonlinear state error feedback is as follows: 。 3. The CNN-GA based active disturbance rejection controller parameter optimization method of claim 1, wherein: In steps 4 and 9, the mapminmax function is used for normalization to normalize the input and output data of the training and test sets to between -1 and 1.

4. The CNN-GA based active disturbance rejection controller parameter optimization method of claim 1, wherein: The specific steps for constructing the model net in step 5 are as follows: Step 5.1: Construct a feature extraction module, which includes convolutional layers, unfolded layers, and fully connected layers. The convolutional layers use one-dimensional convolution for computation. The calculation expression for a single sample convolutional layer is Y=f((m,k)*W1+b1), where (m,k) is the input matrix, m represents the number of samples, k represents the number of channels, W1 is an a×b×k×s weight matrix, a×b is defined as the kernel size, s is defined as the number of kernels, b1 is the bias term, f is the activation function relu:max(0,x), and Y is the output of the convolutional layer. The unfolded layer calculates an expression as Y flat = Reshape(Y, size), where Y flat is the output of the unfolded layer, Reshape is the unfolding function, and size indicates the dimension of the converted result; The calculation expression for a fully connected layer is Z = f(W² * Y). flat +b2), where W2 is the weight matrix, b2 is the bias term, and Z is the output of the fully connected layer; Step 5.2: Construct a regression prediction module. The regression prediction module includes a Drop layer and a regression layer. The Drop layer adds a regularization term to the loss function and randomly discards some data and fills it with 0. Drop layer expression is Z drop = Drop(Z, keep_prop), where Z drop is the output of the Drop layer, and keep_prop is the proportion of data retained in the total data, which ranges from 0 to 1. The regression layer expression is O = g(W3*Z) drop +b3), where W3 is the weight matrix, b3 is the bias term, g is the activation function sigmoid:1 / (1+ exp( x)), and O is the output of the regression layer.

5. The CNN-GA based active disturbance rejection controller parameter optimization method of claim 1, wherein: The specific steps for constructing the genetic algorithm model in step 6 are as follows: Step 6.1: Set the parameters of the genetic algorithm based on the preprocessed dataset. The parameters include the number of iterations maxgen, the population size sizepop, the crossover probability pcross, and the mutation probability pmutation. Step 6.2: Determine the total number of nodes numsum in the model net, set the chromosome length lenchrom and the boundary value bound, and make the chromosome length lenchrom equal to the total number of nodes numsum; Step 6.3: Input the chromosome length lenchrom and boundary value bound from step 6.2 into the function code, and randomly generate chromosome individuals to form the population individuals; Step 6.4: Input the information of individual chromosomes into the model net, and then input the encoded chromosome information and the information of the model net into the fitness calculation function fun to obtain the corresponding fitness individuals.fitness; find the chromosome with the best fitness in each generation and record the chromosome information, fitness value bestfitness, and average fitness avgfitness. Step 6.5: Input the population information individuals and the population size sizepop into the selection function select, and output the new population individuals after selection; Step 6.6: Input the crossover probability pcross, chromosome length lenchrom, chromosome information chrom, and population size sizepop into the crossover function cross. Select the chromosomes to be crossovered from the new population individuals obtained from the selection operation according to the crossover operation probability pcross. Randomly pair them up and exchange some genes to obtain the crossover chromosome individual chrom1. Step 6.7: Input the mutation probability pmutation, chromosome length lenchrom, chromosome information chrom, population size sizepop, current iteration count num, maximum iteration count maxgen, and individual boundary bound into the mutation function Mutation. The mutation function Mutation selects the chromosomes to be mutated from the crossover chromosome chrom1 according to the mutation operation probability pmutation, and then mutates a certain gene on these chromosomes, outputting the mutated chromosome chrom2. The chromosomes to be mutated are randomly selected, and the mutation position is also randomly selected. Step 6.8: Replacement Operation: Input the chromosome population after mutation operation into the fitness function fun one by one, recalculate the fitness, select the best individual, output and record the fitness value of the chromosome individuals.fitness, find the chromosomes with the minimum and maximum fitness and their positions in the population, and replace the chromosome with the maximum fitness with the chromosome with the minimum fitness value; record the best fitness, average fitness and fitness of all chromosomes in the population in each generation; Step 6.9: Return to step 6.5 and proceed to the next iteration; once the maximum number of iterations maxgen is reached, end the iterative training and obtain the chromosome x with the highest fitness. Step 6.10: Decode the best chromosome x and output the optimal initial weights and thresholds.

6. The CNN-GA based active disturbance rejection controller parameter optimization method of claim 5, wherein: In step 6.5, the selection method in the select function adopts the roulette wheel method.

Images