Construction method of radial six-pole hybrid magnetic bearing without displacement sensor control system

Abstract

The application discloses a construction method of a radial six-pole hybrid magnetic bearing without displacement sensor control system, determines the structure of a BP neural network, initializes genetic parameters and coding length of the BP neural network, each individual in a population contains all weights and thresholds of the BP neural network, and determines a fitness function; the individual with the highest fitness value in each generation population is reserved as an elite individual and directly genetically passed to the next generation; the roulette method is adopted to select and operate the remaining individuals; adaptive crossover probability and adaptive mutation probability are calculated; optimal weights and thresholds are output after iteration; the BP neural network is trained; and a displacement prediction model is obtained; genetic algorithm is adopted to optimize initial weights and thresholds of the BP neural network, improve prediction accuracy and convergence speed, the selection operator is redesigned to avoid loss and damage of the best genes in the current population, and the convergence and global optimization ability are improved through adaptive adjustment of the crossover probability and the mutation probability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of high-speed and super-high-speed electric drive control, and relates to the control technology of magnetic suspension bearing (referred to as magnetic bearing), in particular to a construction method of a radial six-pole hybrid magnetic bearing without displacement sensor control system, which is suitable for the fields of industrial manufacturing, life science, energy transportation and aerospace. BACKGROUND

[0002] The magnetic suspension bearing is a high-performance bearing which realizes the mechanical contact between the rotor and the stator, and uses the electromagnetic force of the permanent magnet or the energized coil to make the rotor stably suspend at the geometric center position of the stator. Due to the unique mechanical structure and working principle of the magnetic bearing, it has the advantages of no friction, no wear, no lubrication, high speed, high precision, low power consumption and the like. The measurement of the rotor displacement of the magnetic bearing is the key to realize the stable operation of the system, and at present, the displacement sensor is often used to directly measure the displacement of the rotor of the magnetic bearing. However, the displacement sensor has the problems of high cost, large size, complicated installation and complex control system, and is not suitable for occasions with high requirements on size and speed. Therefore, the displacement sensorless measurement of the magnetic bearing is adopted.

[0003] At present, the common displacement sensorless technology implementation methods mainly include the following categories: the first category is the state observation method, such as Kalman filter method, which needs to obtain the accurate mathematical model of the research object. However, the accurate mathematical model of the alternating current magnetic bearing is difficult to establish due to its strong coupling and nonlinearity. The second category is the parameter estimation method, such as high-frequency signal injection method, which needs to add a circuit to realize the processing of special signals, and the disadvantage is that it makes the control system more complex. The support vector machine is also used for indirect detection of the rotor displacement of the magnetic bearing. This method does not need additional current and signal processing circuit, and the parameters of the prediction model are easy to determine. However, the support vector machine can only solve the single output function prediction problem, and for the multi-output displacement prediction model of the magnetic bearing, multiple support vector machines need to be designed for prediction. The BP neural network has strong non-linear mapping ability, does not need to rely on the mathematical model of the magnetic bearing, and its multi-input multi-output structure characteristics can predict multiple parameters at the same time, which provides a new method for accurate and efficient prediction of the rotor displacement of the magnetic bearing.

[0004] On the other hand, although the BP neural network has strong non-linear mapping ability, the selection of initial weights and thresholds is based on experience and has certain randomness, which may cause slow convergence speed in the calculation process and easily fall into local extremum, resulting in large error between the predicted value and the actual value. SUMMARY

[0005] The application aims at the problems of the existing magnetic bearing displacement sensor, and provides a construction method of a radial six-pole hybrid magnetic bearing displacement sensor control system based on an improved genetic algorithm optimized BP neural network, so as to realize magnetic bearing rotor displacement self-detection and improve control accuracy.

[0006] The construction method of the radial six-pole hybrid magnetic bearing displacement sensor control system comprises the following steps:

[0007] Step 1): Collecting multiple groups of radial control currents i a , i b , i c of the radial six-pole hybrid magnetic bearing, and preprocessing the initial sample set to obtain a training sample set and a test sample set;

[0008] Step 2): Determining the structure of the BP neural network, initializing the genetic parameters and the coding length of the BP neural network, randomly setting N population individuals, each individual in the population containing all the weights w and thresholds b of the BP neural network, and determining the fitness function, and calculating the fitness of all individuals in the population according to the fitness function;

[0009] Step 3): Reserving the individual with the highest fitness value in each generation population as an elite individual to be directly inherited to the next generation, and selecting the remaining N-1 individuals by using the roulette method, and selecting the parent chromosomes according to the fitness values of the individuals in the population, and the selected probability f m is the fitness of the mth individual, N is the number of individuals in the population, and 1≤m≤N-1;

[0010] Step 4): Calculating the adaptive crossover probability P c and the adaptive mutation probability P b , and outputting the optimal weights w and thresholds b after iteration:

[0011]

[0012]

[0013] n is the current iteration number, n max is the maximum iteration number, P cmax is the maximum crossover probability 0.8, P cmin is the minimum crossover probability 0.3, F mean is the average fitness value of the current population, F max is the larger fitness value of the two individuals to be crossed, F is the fitness value of the parent chromosome, P bmax is the maximum mutation probability 0.1, and P bmin is the minimum mutation probability 0.001.

[0014] Step 5) based on the optimal weight w and threshold b, using the training sample set to train the BP neural network to obtain a displacement prediction model;

[0015] Step 6) connecting the output end of the displacement prediction module in sequence with a linear closed-loop controller, a force / current conversion module, a Clark inverse conversion module, a current hysteresis three-phase power inverter and a radial six-pole hybrid magnetic bearing, using a current sensor to measure the radial control current i a , i b , i c and inputting the displacement prediction module to form a radial six-pole hybrid magnetic bearing displacement sensorless control system.

[0016] Further, the method for training the BP neural network comprises: using a Sigmiod function as an activation function of a neuron, respectively calculating an implicit layer output and an output layer output, determining an error function, updating the weight w and the threshold b, constantly adjusting the weight w and the threshold b between each layer of the input layer, the implicit layer and the output layer, stopping iteration when the error meets the accuracy requirement or reaches the maximum iteration number, establishing a nonlinear mapping relationship of the radial control current i a , i b , i c and the displacement x a , y a in the x and y directions, and obtaining the trained BP neural network.

[0017] Further, the test sample set is input into the trained BP neural network model, and the error function is used to test the accuracy of the model.

[0018] Further, the Sigmiod function expression is: x is a function independent variable, and e is a natural base.

[0019] Further, the implicit layer output the output layer output the error function w ij is the weight from the input layer to the implicit layer, x i is the input value of the input layer, b ij is the threshold from the input layer to the implicit layer, i is the number of input layer nodes, j is the number of implicit layer nodes, w jk is the weight from the implicit layer to the output layer, b jk is the threshold from the implicit layer to the output layer.

[0020] The present application has the following advantages:

[0021] 1. The control system constructed by the present invention reduces the size and production cost of the magnetic bearing compared to using sensors to detect rotor displacement, eliminates the potential for sensors to become fault points, and improves the stability of the magnetic bearing system.

[0022] 2. The control system constructed by this invention uses a BP neural network to predict the displacement of the magnetic bearing rotor. It eliminates the need for precise mathematical models and design parameters. Utilizing the nonlinear mapping between the input and output of the BP neural network, a magnetic bearing current-displacement prediction model is established, simplifying the magnetic bearing control system. The BP neural network can achieve the mapping between input and output. Theoretical research shows that this three-layer neural network has high computational accuracy and, under conditions of reasonable network model design and a sufficiently large training sample size, can accurately approximate any nonlinear continuous function.

[0023] 3. To address the issue of randomly setting the initial weights w and threshold b in a BP neural network, this invention employs a genetic algorithm to optimize the initial weights w and threshold b. By seeking the global optimal solution for weights w and threshold b, the accuracy of the model's prediction results is improved. This overcomes the shortcomings of the prediction model being prone to getting trapped in local optima and having too many iterations, thereby improving the prediction accuracy and convergence speed of the BP neural network.

[0024] 4. Compared with the traditional genetic algorithm, the improved genetic algorithm redesigns the selection operator to avoid the loss and destruction of the best genes in the current population. Furthermore, by adaptively adjusting the crossover probability and mutation probability, it improves its convergence and global optimization capabilities, enabling the BP neural network to obtain better initial weights w and thresholds b. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the axial structure and magnetic flux of a radial six-stage hybrid magnetic bearing;

[0026] Figure 2 This is a block diagram of the control system for a radial six-pole hybrid magnetic bearing without displacement sensors;

[0027] Figure 3 yes Figure 2 The structural block diagram of the improved genetic algorithm and optimized BP neural network prediction module;

[0028] Figure 4 yes Figure 2 The flowchart of the construction of the improved genetic algorithm to optimize the BP neural network;

[0029] In the diagram, 1. Radial six-pole hybrid magnetic bearing; 11. Ring permanent magnet; 12. Left radial stator lamination; 13. Right radial stator lamination; 14. Rotor; 15. Radial control coil;

[0030] 2. Improved genetic algorithm optimization BP neural network displacement prediction module; 21. Sample data acquisition module; 22. Sample data preprocessing module; 23. BP neural network structure determination module; 24. Improved genetic algorithm optimization module; 25. BP neural network training module;

[0031] 3. Linear closed-loop controller; 31. First PID controller; 32. Second PID controller;

[0032] 4. Force / current conversion module; 5. Clark inverse transformation module; 6. Current hysteresis three-phase inverter; 7. Current sensor. DETAILED DESCRIPTION

[0033] As Figure 1 shown, the radial six-pole hybrid magnetic bearing is composed of annular permanent magnet 11, left radial stator sheet 12, right radial stator sheet 13, rotor 14 and radial control coil 15, annular permanent magnet 11 is located in the middle of left radial stator sheet 12 and right radial stator sheet 13, along the axial magnetization, provides radial bias magnetic flux, bias magnetic flux from the N pole of annular permanent magnet 11, flow through the left side of the radial stator 12, left side of the air gap, rotor 14, right side of the air gap, right side of the radial stator 13 back to the S pole of annular permanent magnet 11. Rotor 14 is located at the geometric center of the stator sheet, left radial stator sheet 12 and right radial stator sheet 13 are uniformly distributed along the circumference of 6 magnetic poles, respectively, radial control coil 15 is wound on each magnetic pole, the radial control coil 15 on the diametrically opposite two magnetic poles of left radial stator sheet 12 and right radial stator sheet 13 is wound in the same direction and connected in series, the two stator sheets are connected in series in axial opposite direction, 12 radial control coils 15 constitute a three-phase coil, which is connected in star shape and driven by a three-phase inverter. When working, the radial control coil 15 provides radial control magnetic flux, which realizes the radial control of the rotor by changing the size and direction of the three-phase current, and after superposition with the bias magnetic flux, the rotor 14 is stably suspended in the balance position.

[0034] As Figure 2As shown, the sensorless control system for a radial six-pole hybrid magnetic bearing according to the present invention includes a displacement prediction module 2 optimized by an improved genetic algorithm BP neural network 1, a linear closed-loop controller 3, a force / current conversion module 4, a Clark inverse conversion module 5, and a current hysteresis three-phase power inverter 6 connected in series. The output of the current hysteresis three-phase power inverter 6 is connected to the input of the radial six-pole hybrid magnetic bearing 1. The output of the radial six-pole hybrid magnetic bearing 1 is connected to the displacement prediction module 2 optimized by the improved genetic algorithm BP neural network 1. The output of the BP neural network displacement prediction module is then connected to the input of the linear closed-loop controller 3, thus forming a closed-loop control system. The linear closed-loop controller 3 consists of a first PID controller 31 and a second PID controller 32 connected in parallel, corresponding to radial suspension in the x and y directions, respectively.

[0035] The radial control current i of the input radial six-pole hybrid magnetic bearing 1 is measured using current sensor 7. a i b i c Radial control current i a i b i c As the three inputs to the improved genetic algorithm-optimized BP neural network displacement prediction module 2, the output of the improved genetic algorithm-optimized BP neural network displacement prediction module 2 is the displacement prediction value x in the x and y directions. a y a The predicted displacement values ​​x in the x and y directions. a y a Displacement reference value x in the x and y directions a * y a * By comparing and subtracting the values, we obtain the displacement difference e in the x-direction. x This serves as the input to the first PID controller 31. Similarly, the displacement difference e in the y-direction is obtained. y The displacement difference e in the y direction y The first PID controller 31 outputs the radial levitation force F in the x-direction, which is the input to the second PID controller 32. x * The output of the second PID controller 32 is the radial levitation force F in the y direction. y * The outputs of the first PID controller 31 and the second PID controller 32 are both connected to the input of the force / current conversion module 4, where the radial levitation force F... x * F y * After force / current transformation, it is converted into a radial control reference current i. x * iy * , radial control reference current i x * , i y * is converted into three-phase current expected value i a * , i b * , i c * , three-phase current expected value i a * , i b * , i c * as instruction current into current hysteresis three-phase power inverter 6, current hysteresis three-phase power inverter 6 tracks the radial control current i a , i b , i c , at the same time, the current sensor 7 measures the radial control current i a , i b , i c value feedback to the input end of current hysteresis three-phase power inverter 6.

[0036] As shown in Figure 3 , the improved genetic algorithm optimization BP neural network displacement prediction module 2 is composed of sample data acquisition module 21, sample data preprocessing module 22, BP neural network structure determination module 23, improved genetic algorithm optimization module 24 and BP neural network training module 25. Among them, the sample data acquisition module 21 and the sample data preprocessing module 22 are connected in series and then connected to the input end of the BP neural network training module 25, and the BP neural network structure determination module 23 and the improved genetic algorithm optimization module 24 are connected in series and then connected to another input end of the BP neural network training module 25. The input of the sample data acquisition module 21 is the radial control current i a , i b , i c , the output is the initial sample set I, which is used as the input of the sample data preprocessing module 22. After preprocessing by the sample data preprocessing module 22, the output is the training sample set I1 and the test sample set I2. The training sample set I1 is input into the improved genetic algorithm optimization module 24 and the BP neural network training module 25 at the same time, and the test sample set I2 is input into the BP neural network training module 25 after the BP neural network training is completed, in order to verify its prediction accuracy. The output of the improved genetic algorithm optimization module 24 is the optimal weight w and the threshold value b, and the output of the BP neural network training module 25 is the displacement prediction value x a , ya .

[0037] As shown in Figure 4 , the improved genetic algorithm optimizing BP neural network displacement prediction module 2 randomly generates an initial population containing all the weights w and threshold values b of the BP neural network. All individuals in the initial population are formed into multiple chromosomes through real number coding. The chromosomes are formed into a new population through improved selection, crossover and mutation operations. The iteration is stopped after meeting the termination condition. The optimal chromosome in the population formed in the last iteration is found. The optimal weights w and threshold values b are obtained by decoding the chromosome and input into the BP neural network training module. The specific construction process is as follows:

[0038] Step 1: The sample data acquisition module 21 acquires sample data. Multiple groups of radial control currents i a , i b , i c of the radial hybrid magnetic bearing 1 are continuously acquired as initial input sample data to form an initial sample set I.

[0039] Step 2: The sample data preprocessing module 22 preprocesses the data. The sample data preprocessing module 22 preprocesses the original data of the initial sample set I collected to remove invalid data and obtain a preprocessed data set. The sample set data is normalized to limit the numerical value to [-1, 1], which is more conducive to the training of the BP neural network. 1000 groups of processed sample data are randomly selected, of which 700 groups are used as a training sample set I1 for training the BP neural network training module 23; and the other 300 groups are used as a test sample set I2 for verifying the prediction accuracy of the displacement prediction model.

[0040] Step 3: The structure of the BP neural network is determined by the structure determination module 23. Since the input of the improved genetic algorithm optimizing BP neural network displacement prediction module 2 is the radial control current i a , i b , i c , and the output is the x, y direction displacement prediction value x a , y a , it can be determined that the input layer of the BP neural network has 3 nodes, the output layer has 2 nodes, and the number of nodes in the hidden layer can be determined as 8 according to the empirical formula, and the expression of the empirical formula is:

[0041]

[0042] Therefore, a BP neural network model with a topological structure of 3-8-2 is established, wherein N i represents the number of nodes in the input layer, N o represents the number of nodes in the output layer, and δ is a constant between 1 and 10.

[0043] Step 4: The improved genetic algorithm optimization module 24 optimizes the weights w and the threshold b of the BP neural network. Specifically:

[0044] Step 4.1: Initialize genetic parameters. Set the genetic parameters, including the population size of 20, the iteration number of 100, and the fitness accuracy of 0.001.

[0045] Step 4.2: Initial value coding. In genetic algorithms, the coding method for population individuals mainly includes binary coding and real number coding. The binary coding method is used to optimize the weights w and the threshold b of the BP neural network. The binary coding method is simple to encode and decode, and easy to implement crossover and mutation operations.

[0046] Step 4.3: Population initialization. Randomly set N population individuals, each individual in the population contains all the weights w and the threshold b of the BP neural network, that is, one individual corresponds to a determined BP neural network model. The value range of the weights is [-1, 1], and the value range of the threshold is [0, 1].

[0047] Step 4.4: Determine the fitness function. The sum of the squares of the differences between the actual values and the expected values of different output nodes of the BP neural network is used as the error function of the genetic algorithm, and the reciprocal of the error function is used to define the fitness function, and the fitness function expression is obtained as:

[0048]

[0049] In the formula, y ok and y k respectively represent the actual value and the expected value of the output node of the BP neural network, and k represents the number of output nodes.

[0050] Step 4.5: Calculate the fitness value. A group of sample data is randomly selected from the training sample set I1, and the fitness values of all individuals in the population are calculated according to the fitness function in step 4.4, combined with all the weights w and the threshold b of the BP neural network contained in the population individuals.

[0051] Step 4.6: Improved selection operation. The selection operation used in the traditional genetic algorithm may eliminate the best fitness individual of a population, so an improved selection operation is needed to reduce the error caused by the traditional selection operation. The improved selection operation described in the present application combines the elite reservation method with the roulette method. First, the individual with the highest fitness value in each population is reserved as an elite individual to be directly inherited to the next generation. Then, the roulette method is used to select the remaining N-1 individuals. The individual with a higher fitness value is more likely to be selected. Let the fitness of the mth individual be f mP = m / N m is:

[0052]

[0053] where N represents the number of individuals in the population, and 1≤m≤N-1.

[0054] Step 4.7: Adaptive crossover operation. In the traditional genetic algorithm, the crossover probability P c is a constant value, which is in the range of [0.3, 0.8]. The larger the value of the crossover probability, the stronger the global search ability of the genetic algorithm, but too large a crossover probability will destroy the original high fitness of the chromosome. Conversely, the smaller the value of the crossover probability, the smaller the global search ability and convergence speed of the genetic algorithm, so an adaptive crossover operation is designed to continuously adjust the crossover probability of the genetic algorithm according to the fitness value. In the early stage of population iteration, the individual fitness value is less than the average fitness value, and the crossover probability needs to be increased to expand the global search ability. In the later stage of population iteration, the individual fitness value is greater than the average fitness value, and the global search ability is reduced by reducing the crossover probability. The adaptive crossover probability expression is as follows:

[0055]

[0056] where n is the current iteration number, n max is the maximum iteration number, P cmax is the maximum crossover probability 0.8, P cmin is the minimum crossover probability 0.3, F mean is the average fitness value of the population, F max is the larger fitness value of the two individuals to be crossed.

[0057] Step 4.8: Adaptive mutation operation. In the traditional genetic algorithm, the default value of the mutation probability P b is in the range of [0.001, 0.1]. The reason for improving the mutation operation is the same as step 4.7, which will not be described in detail here. The adaptive mutation operation uses a smaller mutation probability in the early stage of population iteration to preserve good genes in the chromosome; in the later stage of population iteration, the mutation probability is increased to enhance the local search ability. The adaptive mutation probability expression is as follows:

[0058]

[0059] where F is the fitness value of the parent chromosome, P bmax is the maximum mutation probability 0.1, P bmin is the minimum mutation probability 0.001.

[0060] Step 4.9: Iteration operation. After the genetic operation of step 4.6, step 4.7 and step 4.8, the individuals in the population are updated, that is, a new population is obtained, at this time, it is necessary to return to step 4.5 to calculate the fitness value of the individuals in the new population.

[0061] Step 4.10: Output the optimal weight w and threshold b. There are two conditions for the end of the iteration operation, the first is to reach the set number of iterations, and the second is that the individual fitness value error reaches the requirement, and one of them can stop iteration. After iteration, the best individual in the population is decoded, and the optimal BP neural network weight w and threshold b are output.

[0062] Step 5: Based on the optimal weight w and threshold b and the training sample set, the BP neural network is trained to obtain the displacement prediction model. Specifically:

[0063] Step 5.1: Network parameter initialization. Set the BP neural training parameters, including the learning efficiency is 0.05, the number of iterations is 100, and the target accuracy is set to 0.0001.

[0064] Step 5.2: Determine the activation function of the neuron. The Sigmiod function is used as the activation function of the neuron, which is continuous itself and its derivative, which is convenient for data calculation and processing. The expression of the Sigmiod function is:

[0065]

[0066] In the formula, x is the function argument, and e is the natural base.

[0067] Step 5.3: Calculate the hidden layer output. The expression for calculating the hidden layer output value is:

[0068]

[0069] In the formula, w ij is the weight from the input layer to the hidden layer, x i is the input value of the input layer, b ij is the threshold value from the input layer to the hidden layer, i is the number of input layer nodes, and j is the number of hidden layer nodes.

[0070] Step 5.4: Calculate the output layer output. The expression for calculating the output layer output value is:

[0071]

[0072] In the formula, w jk is the weight from the hidden layer to the output layer, b jk is the threshold value from the hidden layer to the output layer.

[0073] Step 5.5: Determine the error function. The error function of the output node is determined based on the sum of squared errors between the actual and expected outputs of the network, and its expression is:

[0074]

[0075] Step 5.6: Update the weights w and the threshold b. The update expression is:

[0076]

[0077]

[0078]

[0079]

[0080] In the formula, w jk 'and b jk ' represents the updated weights and thresholds from the hidden layer to the output layer, w ij 'and b ij ' represents the updated weights and thresholds from the input layer to the hidden layer, respectively, and η is the learning efficiency, which is 0.05 in this invention.

[0081] The BP neural network is trained using the input training sample set I1 and the weights w and thresholds b obtained from the genetic algorithm in step 4. The weights w and thresholds b between the input, hidden, and output layers are continuously adjusted using the forward propagation and back feedback principles of the BP neural network. Iteration stops when the error meets the accuracy requirements or the maximum number of iterations is reached, ultimately establishing the radial control current i. a i b i c and displacements in the x and y directions x a y a The nonlinear mapping relationship is obtained, which leads to the improved genetic algorithm optimized BP neural network displacement prediction module 2.

[0082] Step 6: Verify the prediction accuracy of the BP neural network. Input the test sample set I2 into the trained BP neural network model, and use the error function expression determined in step 5.3 to verify whether the accuracy of the prediction model meets the requirements.

[0083] The output of the improved genetic algorithm-optimized BP neural network displacement prediction module 2 is connected in series with a linear closed-loop controller 3, a force / current conversion module 4, a Clark inverse conversion module 5, a current hysteresis three-phase power inverter 6, and a radial six-pole hybrid magnetic bearing 1. The radial control current i is measured by a current sensor 7. a i b i c Then the radial control current ia i b i c The input is into the improved genetic algorithm optimization BP neural network displacement prediction module 2, and a displacement sensorless control system of the radial six-pole hybrid magnetic bearing is formed.

[0084] The application replaces the traditional displacement sensor for measuring the rotor displacement of the magnetic bearing, greatly saves the cost and space, and has strong applicability; the BP neural network algorithm is adopted to train the sample data, the current-displacement prediction model is established, and the stability of the magnetic bearing system is improved; the improved genetic algorithm is adopted to initialize the weight w and the threshold b of the BP neural network, the convergence speed of the BP neural network is accelerated, and the local optimal solution is avoided, so that the prediction model of the BP neural network is more efficient and accurate.

Claims

1. A configuration method of a radial six-pole hybrid magnetic bearing non-displacement sensor control system, characterized in that The method comprises the following steps: Step 1): Collecting multiple groups of radial control currents i of a radial hexapole hybrid magnetic bearing a , i b , i c composing an initial sample set, preprocessing the initial sample set, obtaining a training sample set and a test sample set; Step 2): determining the structure of the BP neural network, initializing the genetic parameters and the coding length of the BP neural network, randomly setting N population individuals, each individual in the population containing all the weights w and thresholds b of the BP neural network, and determining a fitness function, and calculating the fitness of all individuals in the population according to the fitness function; Step 3): The individual with the highest fitness value in each generation population is reserved as the elite individual to be directly inherited to the next generation, and the roulette method is used for selection operation on the remaining N-1 individuals, and the selected probability is proportional to the fitness value of the individual in the population f m is the fitness value of the mth individual, and N is the number of individuals in the population, 1≤m≤N-1. Step 4) Calculate the adaptive crossover probability P c and the adaptive mutation probability P b and the threshold b after iteration. n is the current iteration number, n max P is the maximum iteration number cmax P is the maximum crossover probability 0.8 cmin F is the minimum crossover probability 0.3 mean F is the average fitness value of the current population max F is the larger fitness value of the two individuals to be crossed, F is the fitness value of the parent chromosome bmax P is the maximum mutation probability 0.1 bmin F is the minimum mutation probability 0.001 Step 5): based on the optimal weights w and thresholds b, training the BP neural network using the training sample set to obtain a displacement prediction model; Step 6): the output of the displacement prediction module is connected in series with a linear closed-loop controller, a force / current conversion module, a Clark inverse conversion module, a current hysteresis three-phase power inverter and a radial six-pole hybrid magnetic bearing, respectively, and a current sensor is used to measure the radial control current i a , i b , i c of the radial six-pole hybrid magnetic bearing and input into the displacement prediction module to form a radial six-pole hybrid magnetic bearing displacement sensorless control system.

2. The construction method of the radial hybrid six-pole magnetic bearing non-displacement sensor control system according to claim 1, characterized in that: The method for training the BP neural network comprises the following steps: adopting a Sigmiod function as an activation function of a neuron, respectively calculating an output of a hidden layer and an output of an output layer, determining an error function, updating a weight value w and a threshold value b, constantly adjusting the weight value w and the threshold value b between each layer of an input layer, a hidden layer and an output layer, stopping iteration when the error meets a precision requirement or reaches a maximum iteration number, and establishing a radial control current i a , i b , i c and a nonlinear mapping relationship of displacement x a , y a in x and y directions, so as to obtain the trained BP neural network.

3. The construction method of the radial hybrid six-pole magnetic bearing non-displacement sensor control system according to claim 2, characterized in that: The test sample set is input into the trained BP neural network model, and the error function is used to test the accuracy of the model.

4. The construction method of claim 2, wherein the radial hybrid six-pole magnetic bearing non-displacement sensor control system is characterized in that: The Sigmiod function expression is: x is a function argument, and e is a natural base.

5. The construction method of claim 4, wherein the radial hybrid six-pole magnetic bearing non-displacement sensor control system is characterized in that: The output of the hidden layer The output of the output layer The error function w ij is the weight from the input layer to the hidden layer, x i is the input value of the input layer, b ij is the threshold value from the input layer to the hidden layer, i is the number of input layer nodes, j is the number of hidden layer nodes, w jk is the weight from the hidden layer to the output layer, b jk is the threshold value from the hidden layer to the output layer.

6. The construction method of the radial hybrid six-pole magnetic bearing non-displacement sensor control system according to claim 5, characterized in that: The weights w and thresholds b are updated according to the following expression: w jk 'and b jk 'are the updated weights and thresholds from the hidden layer to the output layer, respectively, w ij 'and b ij 'are the updated weights and thresholds from the input layer to the hidden layer, respectively, and η is the learning efficiency.

7. The method of claim 1, wherein the method further comprises: providing a radial 6-pole hybrid magnetic bearing non-displacement sensor control system comprising: a radial 6-pole hybrid magnetic bearing; a radial 6-pole hybrid magnetic bearing controller; a radial 6-pole hybrid magnetic bearing sensor; and a radial 6-pole hybrid magnetic bearing sensor controller. The structure of the BP neural network is that the input layer has 3 nodes, the output layer has 2 nodes, and the hidden layer has 8 nodes.

8. The method of claim 1, wherein the method further comprises: providing a radial 6-pole hybrid magnetic bearing non-displacement sensor control system comprising: a radial 6-pole hybrid magnetic bearing; a radial 6-pole hybrid magnetic bearing controller; a radial 6-pole hybrid magnetic bearing sensor; and a radial 6-pole hybrid magnetic bearing sensor controller. The coding length L is 50.

9. The method of claim 1, wherein the method further comprises: providing a radial 6-pole hybrid magnetic bearing non-displacement sensor control system comprising: a radial 6-pole hybrid magnetic bearing; a radial 6-pole hybrid magnetic bearing controller; a radial 6-pole hybrid magnetic bearing sensor; and a radial 6-pole hybrid magnetic bearing sensor controller. The value range of the weights w is [-1, 1], and the value range of the threshold b is [0, 1].

Images