Method for constructing decoupling controller of permanent-magnet-assisted bearingless synchronous reluctance electric motor

Abstract

Disclosed in the present invention is a method for constructing a decoupling controller of a permanent-magnet-assisted bearingless synchronous reluctance electric motor. A decoupling controller formed by an IDBO-FNN dynamic prediction module and an IDBO-FNN system is connected to a composite controlled object, wherein the IDBO-FNN dynamic prediction module is formed by three parallel IDBO-FNN prediction modules, three parallel control increment calculation modules, three parallel prediction value calculation modules and a composite signal calculation module; and the IDBO-FNN system is formed by means of connecting a second-order difference processor and first-order difference processor of a fourth IDBO parameter optimization module before a fourth fuzzy neural network model, each IDBO parameter optimization module using an improved dung beetle optimizer to optimize the membership function width, membership function center value and weight of a fuzzy neural network. The structure of a fuzzy neural network is optimized, such that the degree of dependence of the fuzzy neural network on expert experience is reduced, the convergence speed of the weight and membership function of the fuzzy neural network is improved, high robustness is achieved, and various good static and dynamic performances such as rotor radial position control and electric motor speed control can be obtained.

Description

A method for constructing a decoupling controller for a permanent magnet-assisted bearingless synchronous reluctance motor Technical Field

[0001] The present invention belongs to the permanent magnet bearingless synchronous magnetic group motor control technology in the field of electric transmission control equipment. It is a motor reverse decoupling controller that provides stable decoupling control for the high-speed operation of the permanent magnet assisted bearingless synchronous reluctance motor and is widely used in the field of high-speed electric transmission. Background Art

[0002] Compared with traditional bearingless motors, permanent magnet assisted bearingless synchronous reluctance motors have a simple structure, reliable operation and low cost. Yiwei Industry's design of adding permanent magnets to the rotor greatly improves its torque density and power factor, which perfectly meets the requirements of modern high-performance drive motors such as smoothness and easy maintenance, and is more suitable for high-speed and ultra-high-speed applications.

[0003] The permanent magnet-assisted bearingless synchronous reluctance motor (PRM) is a complex, strongly coupled, nonlinear, multi-input, multi-output (MIMO) system. Stable rotor suspension and operation require dynamic decoupling control between the electromagnetic torque and the mirror suspension force, as well as between the radial suspension force components in two perpendicular directions. Traditional decoupling control methods include feedback decoupling control and inverse system decoupling control. However, decoupling control using an α-order inverse system requires a precise mathematical model of the controlled object, making it difficult to achieve decoupling during actual motor operation. Neural network inverse control can address these shortcomings. Specifically, it achieves linearized decoupling without relying on a precise mathematical model and avoids control errors caused by system instability. Although neural network inverse control does not require a precise mathematical model, it does have inherent limitations, such as significant influence of training examples on weight adjustment, slow learning speed, and unclear operating principles.

[0004] Chinese patent number CN201710511782.9, titled "Fuzzy Neural Network Decoupling Controller for Five-Degree-of-Freedom Bearingless Permanent Magnet Synchronous Motor," uses a fuzzy neural network to decouple permanent magnet synchronous motors. However, this method suffers from certain drawbacks, as it can easily render all nodes in the fuzzy rule layer invalid without expert experience. Chinese patent number CN201110021862.9, titled "Construction Method for Generalized Inverse Decoupling Controller for Bearingless Synchronous Reluctance Motor," uses a neural network to decouple bearingless synchronous reluctance motors. However, this method requires high data sample accuracy and a slow learning rate, requiring weight optimization based on the neural network algorithm.

[0005] Summary of the Invention

[0006] The purpose of the present invention is to solve the problems existing in the existing decoupling control algorithm of permanent magnet assisted bearingless synchronous reluctance motor, and propose a method for constructing a decoupling controller of permanent magnet assisted bearingless synchronous reluctance motor based on a fuzzy neural network (FNN) algorithm optimized by an improved dung beetle optimizer (IDBO), which has strong robustness, low sample requirements and fast learning speed, so as to achieve precise control of the bearingless synchronous reluctance motor on the basis of the original decoupling control.

[0007] To achieve the above-mentioned object, the technical solution adopted by the method for constructing a decoupling controller of a permanent magnet assisted bearingless synchronous reluctance motor of the present invention includes the following steps:

[0008] Step A): The IDBO-FNN dynamic prediction module is composed of three parallel IDBO-FNN prediction modules, three parallel control increment calculation modules, three parallel prediction value calculation modules and a composite signal calculation module; an IDBO-FNN prediction module, a control increment calculation module and a prediction value calculation module are connected in series to form a series branch, which are two rotor radial displacement branches and a speed branch. The speed branch obtains the speed control value n c , the outputs of the two rotor radial displacement branches are used to obtain the rotor eccentric displacement r through the composite signal calculation module;

[0009] Step B): The IDBO-FNN system is formed by connecting the second-order differential processor and the first-order differential processor of the fourth IDBO parameter optimization module before the fourth fuzzy neural network model. The rotor eccentric displacement r is processed by the second-order differential processor to obtain the first-order and second-order displacement differential signals, and the speed control variable n c The first-order speed differential signal is obtained through the first-order differential processor; the rotor eccentric displacement r, the speed control value n c The membership function width, center value, weight, first-order and second-order displacement differential signals of the fourth fuzzy neural network model and the first-order speed differential signal are used as inputs of the fourth fuzzy neural network model, and the fourth fuzzy neural network model outputs a control current signal;

[0010] Each of the IDBO-FNN prediction modules is composed of a corresponding fuzzy neural network model connected to a corresponding IDBO parameter optimization module; each of the IDBO parameter optimization modules uses the following improved dung beetle optimization algorithm to optimize the membership function width, center value and weight of the fuzzy neural network:

[0011] Step 1: Set four populations: rolling dung beetles, breeding dung beetles, small dung beetles, and stealing dung beetles, with the optimization dimension D set to 3, the maximum number of iterations T, the initialization of the judgment parameter α, the number of each dung beetle population, and the optimal fitness;

[0012] Step 2: Determine whether the number of iterations has reached the maximum value or whether the individual has reached the optimal fitness. If not, proceed to step 3. If so, terminate the iteration.

[0013] Step 3: Determine the population of the current individual. If the individual is a dung beetle, proceed to step 4. If the individual is not a dung beetle, proceed to step 5.

[0014] Step 4: Update the position according to the rolling ball dung beetle position update algorithm;

[0015] Step 5: If the individual is a juvenile dung beetle, proceed to step 6; if the individual is not a juvenile dung beetle, proceed to step 7;

[0016] Step 6: Update the position of the young dung beetle, and then jump to step 10;

[0017] Step 7: If the individual is a small dung beetle, proceed to step 8; if the individual is not a small dung beetle, proceed to step 9;

[0018] Step 8: Update the position of the dung beetle, then jump to step 10;

[0019] Step 9: If the individual is a thieving dung beetle, update the location of the thieving dung beetle and then jump to step 10;

[0020] Step 10: Generate a random number r1∈(0,1), set the perturbation probability pv, and determine whether r1 is less than pv. If it is less than or equal to pv, proceed to step 11; if it is greater than pv, proceed to step 15.

[0021] Step 11: Generate a random number r2∈(0,1), set a constant S, and determine whether r2 is less than ST. If so, proceed to step 12; if greater than ST, proceed to step 13.

[0022] Step 12: Update the position of the follower according to the sparrow search algorithm; if the obtained position remains unchanged for five consecutive iterations, mutate the individual according to the Cauchy-Gauss formula;

[0023] Step B): constructing a composite controlled object including a permanent magnet assisted bearingless synchronous reluctance motor;

[0024] Step C): The IDBO-FNN dynamic prediction module and the IDBO-FNN system constitute a decoupling controller connected to the composite controlled object.

[0025] The advantages of the present invention using the above technical solution are:

[0026] 1. The present invention combines the advantages of fuzzy logic control's low sample requirements and neural network's good system learning ability and predictive control dynamic performance, further optimizes the fuzzy neural network structure, reduces the fuzzy neural network's dependence on expert experience, and proposes an optimization algorithm to improve the convergence speed of the fuzzy neural network weights and membership functions. It has great advantages in processing the complex systems of permanent magnet-assisted bearingless synchronous reluctance motors with nonlinearity, strong coupling and multiple variables, has strong robustness, and can obtain good static and dynamic performance such as rotor radial position and motor speed control.

[0027] 2. The IDBO-FNN decoupling control module employed in this invention uses a composite signal of radial displacement, rotational speed, and rotor eccentricity angle θ as the control signal, and a current signal as the output signal. A feedback loop is added after the output current. Compared to using radial displacement alone as the input signal, the composite signal not only better reflects the overall operating state of the motor, but also allows for targeted control of radial displacement in either direction by adjusting weight parameters, improving control flexibility. Compared to directly converting the output current for control, the added feedback provides more precise control and allows for real-time adjustment of the motor by adjusting the PI parameters.

[0028] 3. The IDBO-FNN dynamic prediction module employed in this invention offers superior fitting capabilities, faster convergence, and a relatively simple algorithm compared to neural network data prediction. By training the IDBO-FNN to generate a system model and using the current value to predict the next moment, the IDBO-FNN dynamic prediction module exhibits greater robustness and tracking performance compared to PID control of the current error. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] FIG1 is a block diagram of a decoupling controller for a permanent magnet-assisted bearingless synchronous reluctance motor constructed using the method of the present invention;

[0030] FIG2 is a structural block diagram of the composite controlled object in FIG1 ;

[0031] FIG3 is a block diagram of the internal structure of the IDBO-FNN dynamic prediction module in FIG1 ;

[0032] FIG4 is an optimization flow chart of the improved dung beetle algorithm;

[0033] FIG5 is a block diagram of the IDBO-FNN system in FIG1 ;

[0034] In the figure: 1. Permanent magnet assisted bearingless synchronous reluctance motor composite controlled object; 2. IDBO-FNN system; 3. IDBO-FNN dynamic prediction module; 4. Angle calculation module; 5, 6, 7, 8. PI regulator; 9, 11. Park inverse transform; 10, 12. Park transform; 13, 15. Space vector PWM; 14, 16. Clark transform; 17, 18. Voltage-type inverter; 19. Suspension force subsystem; 20. Torque subsystem; 21, 52, 54, 56. IDBO parameter optimization model; 22. Second-order difference processor; 23. First-order difference processor; 24. IDBO-FNN; 31, 34, 38. IDBO-FNN prediction module; 24, 51, 53, 55 fuzzy neural network model; 32, 35, 39. Control increment calculation module; 33, 36, 3A. Prediction value calculation module; 37. Composite signal calculation module. DETAILED DESCRIPTION

[0035] As shown in Figure 1, the decoupling controller of the permanent magnet assisted bearingless synchronous reluctance motor consists of an IDBO-FNN dynamic prediction module 3 and an IDBO-FNN system 2, which are connected in series before the permanent magnet assisted bearingless synchronous reluctance motor composite controlled object 1. The IDBO-FNN dynamic prediction module 3, the IDBO-FNN system 2 and the permanent magnet assisted bearingless synchronous reluctance motor composite controlled object 1 are connected in phase. The permanent magnet assisted bearingless synchronous reluctance motor composite controlled object 1 outputs the actual rotor radial displacement x, y and speed n. The actual rotor radial displacement x, y and speed n at time t acquired by the sensor are equal to the given rotor radial displacement x at time t. * ,y * And the given speed n * The output of IDBO-FNN dynamic prediction module 3 is the eccentric displacement r of the rotor at time t+1 and the speed control value n at time t+1. c The input of the angle calculation module 4 is the actual speed n, and the output is the actual eccentric angle of the rotor θ=nt. The eccentric displacement r at time t+1 output by the IDBO-FNN dynamic prediction module 3 is combined with the speed control value n at time t+1. c The eccentric angle θ obtained by the angle calculation module 4 is used as the input of the IDBO-FNN system 2. After being processed by the IDBO-FNN system 2, the reference current component of the suspension winding at time t+1 is obtained. and the torque winding reference current component The output of the IDBO-FNN system 2 is used as the input of the permanent magnet assisted bearingless synchronous reluctance motor composite controlled object 1, and the decoupling control of the permanent magnet assisted bearingless synchronous reluctance motor is finally completed after being processed by the permanent magnet assisted bearingless synchronous reluctance motor composite controlled object 1.

[0036] As shown in Figure 2, a composite controlled object 1 is constructed, including a permanent magnet-assisted bearingless synchronous reluctance motor. The permanent magnet-assisted bearingless synchronous reluctance motor includes a suspension force subsystem 19 and a rotor subsystem 20. A three-phase inverter circuit is connected between the suspension force subsystem 19 and the rotor subsystem 20. The three-phase inverter circuit consists of a first PI controller 15, a second PI controller 16, a third PI controller 17, a fourth PI controller 18, a first Park inverse transform 9, a second Park inverse transform 11, a first Park transform 10, a second Park transform 12, a first space vector PWM 13, a second space vector PWM 15, a first Clark transform 14, a second Clark transform 16, a first voltage-source inverter 17, and a second voltage-source inverter 18. In the control of the suspension force subsystem, the d- and q-axis suspension force winding reference current components at time t+1, processed and output by the IDBO-FNN system 2, are used as the output. and The winding current i at time t fed back by the first Park transformation 10 Bd and i Bq The difference is made respectively and converted into the expected value of the suspension winding voltage at time t+1 by the first PI controller 5 and the second PI controller 6. and After inputting into the first PARK inverse transformation 9, the transformed output is the flux tracking control input signal at time t+1 and The flux tracking control input signal at time t+1 and The input is converted into six PWM waves in the first space vector PWM13 and input into the first voltage source inverter 17. After transformation, the three-phase suspension winding current i at time t+1 is obtained. BU 、i BV and i BW , controls the rotor suspension and outputs the rotor radial displacement x, y through the sensor. Three-phase suspension force winding current i BU 、i BV and i BW After the first Clark transformation 14, it is transformed into the winding current i Bα and i Bβ The input is fed into the first Park transformation 10 to form a current negative feedback closed loop. The control process of the torque subsystem is similar to that of the suspension force subsystem. Therefore, the input of the composite controlled object is the suspension force winding reference current component at time t+1 obtained by processing the output of the IDBO-FNN system 2. And the torque winding reference current component and, The output is the rotor radial displacement x, y and speed n.

[0037] As shown in Figure 3, the IDBO-FNN dynamic prediction module 3 is constructed. The IDBO-FNN dynamic prediction module 3 consists of three parallel IDBO-FNN prediction modules 31, 34, and 38, three parallel control increment calculation modules 32, 35, and 39, three parallel prediction value calculation modules 33, 36, and 3A, and a composite signal calculation module 37. An IDBO-FNN prediction module 31, a control increment calculation module 32, and a prediction value calculation module 33 are connected in series to form the first series branch, which predicts the rotor radial displacement x. An IDBO-FNN prediction module 34, a control increment calculation module 35, and a prediction value calculation module 36 are connected in series to form the second series branch, which predicts the rotor radial displacement y. An IDBO-FNN prediction module 38, a control increment calculation module 39, and a prediction value calculation module 3A are connected in series to form the third series branch, which predicts the rotor speed n. This totals two rotor radial displacement branches and one speed branch. The outputs of prediction value calculation module 33 and prediction value calculation module 36 are connected to a composite signal calculation module 37, and the predicted rotor radial displacement x and y values ​​are output as composite signals. The three parallel IDBO-FNN prediction modules 31, 34, and 38 have identical structures, respectively consisting of a fuzzy neural network model 51, 53, and 55 connected to an IDBO parameter optimization module 52, 54, and 56. Each IDBO-FNN prediction module correspondingly predicts a radial displacement x and y and a speed n. The outputs of the two series branches that predict the rotor radial displacement x and y are calculated by composite signal calculation module 37 to form a composite signal.

[0038] Given the rotor radial displacement x * and y * And the given speed n * The difference between the actual radial displacement x and y of the rotor and the speed n is input into the corresponding IDBO-FNN prediction module 31, 34, 38, that is, the given rotor radial displacement x * The difference between the actual rotor radial displacement x and the given rotor radial displacement y is input into the first IDBO-FNN prediction module 31. * The difference between the actual radial displacement y and the rotor is input into the second IDBO-FNN prediction module 34, and the given rotor speed n is * The difference between the actual rotor speed n and the result is input into the third IDBO-FNN prediction module 38.

[0039] Each IDBO-FNN prediction module 31, 34, 38 is composed of a corresponding fuzzy neural network model 51, 53, 55 connected to a corresponding IDBO parameter optimization module 52, 54, 56. That is, the IDBO-FNN prediction module 31 is composed of a fuzzy neural network model 51 and an IDBO parameter optimization module 52 connected, the IDBO-FNN prediction module 34 is composed of a fuzzy neural network model 53 and an IDBO parameter optimization module 54 connected, and the IDBO-FNN prediction module 38 is composed of a fuzzy neural network model 55 and an IDBO parameter optimization module 56 connected.

[0040] The three IDBO-FNN prediction modules 31, 34, and 38 process the input signal in the same manner. The specific processing process is described below using the IDBO-FNN prediction module 31 as an example:

[0041] The IDBO-FNN prediction module 31 consists of a fuzzy neural network model 51 and an IDBO parameter optimization module 52. The fuzzy neural network model 51 predicts the radial displacement given at input time t and the difference △x between the radial displacements. The fuzzy neural network does not require complex expressions, but the prediction accuracy depends on the appropriate membership function width. and the center value and weights The IDBO parameter optimization module is used to optimize the width of the membership function in the fuzzy neural network. and the center value and weights The simulation model is established based on the permanent magnet assisted bearingless synchronous reluctance motor composite controlled object 1 given in Figure 2 and the IDBO-FNN dynamic prediction module 3 given in Figure 3. As a step excitation signal, it is added to the control end of the permanent magnet assisted bearingless synchronous reluctance motor. At the same time, the actual radial displacement signal x of the rotor of the permanent magnet assisted bearingless synchronous reluctance motor at time t is measured by the sensor. The actual radial displacement signal x of the rotor at time t is sampled, and a total of 1000 sets of data are sampled. Since the data collected by the actual sensor will lag behind the measured object by △t time, 1000 sets of radial displacement signals x″ are sampled again at an interval of △t as the output value of the fuzzy neural network model 51 at time t+1. Then, the first set of 1000 sets of sampled data {x} are compared with the displacement given value {x *}, a set of data {△x} is obtained by subtraction as the sample set of the input signal of the fuzzy neural network model 51, denoted as X1, and {x″} obtained by sampling again is used as the sample set of the output signal of the fuzzy neural network model 51, denoted as Y1. The samples are normalized and all variable values ​​are limited to [-1, 1] to avoid the influence of the order of magnitude on the calculation. 80% of the sample data are used as the training sample X1train 、Y 1train , as training set data, and the remaining 20% ​​as test samples X 1test 、Y 1test , as the test set data. The training set is input into the fuzzy neural network prediction model 51 for parameter training. The fuzzy segmentation number d=2, the fuzzy language number m=3, and the membership function uses the Gaussian function. The test set data is sent to the IDBO parameter optimization module 52 for parameter training. Optimization. Taking the population fitness performance index as the parameter After being processed by the IDBO parameter optimization module 52, a set of parameters with the best fitness value is The fuzzy neural network model 51 is sent to the fuzzy neural network model 51 for back propagation training and then used for displacement prediction. The back propagation training of the fuzzy neural network model 51 is mainly for weights. To update, first make the difference between the output value predicted by the fuzzy neural network and the actual output value in the training set to get the error e=y(X 1train )-Y 1train , then in the back propagation the weights are updated by

[0042] Where z is the number of iterations, z is 1, H is the total number of training samples, H = 800, and k is the learning rate of the fuzzy neural network.

[0043] The fuzzy neural network model 51 is divided into five layers. The first layer is the input layer, which is used to input external input into the fuzzy neural network. Let the input be a vector X, and the dimension of X is h, that is, X=[X1,X2........,X h ] T , the input of the fuzzy neural network model 51 is the displacement difference △x, X=[△x] T , the number of nodes in the first layer is N=1.

[0044] The second layer is the separation layer, which is used to divide the input vector X into components of dimension d = 2, and divide it into p subsets in total, where If N is not divisible by d, add 0 to the front and back of the input vector so that the dimension of the input vector is divisible by d. At this time, p = 1, and the dimension of the second layer becomes N' = 1. The number of nodes in the second layer is N', and

[0045] The third layer is the fuzzification layer, which is used to map each node in the segmentation layer to a fuzzy language value. At this time, for the only input △x in the segmentation layer, there is

[0046] Where i = 1, 2, 3...., h, N′ = 1, 2, 3..., m, s i =1,2,3...,m, m is the number of fuzzy languages ​​for each input, m=3, For input △x at its sth i The membership degree on the fuzzy linguistic value, is the sth input △x i Membership functions, all fuzzy neural network membership functions in the present invention use Gaussian functions, For the s i The central value of the membership function, For the s i The width of the membership function.

[0047] The number of nodes in the third layer is N′×m.

[0048] The fourth layer is the fuzzy rule layer, which is used to superimpose the fuzzy rule strength. Each rule strength node has

[0049] Where, j=1,2,3......,p,k=1,2,3,...,m, For each rule strength node in the fourth layer.

[0050] The number of nodes in the fourth layer is p×m.

[0051] The fifth layer is the output layer, which is used to associate the fuzzy rule strength of each subset p with the output number. The predicted output of the fuzzy neural network model 51 for the input △x is the rotor radial displacement x″ at time t+1.

[0052] Where, is the weight of the output layer, r is the output dimension, and the output dimension is 1.

[0053] The fuzzy neural network model 51 outputs the radial displacement x″ at time t+1 as the output of the IDBO-FNN prediction module 31, which is input into the control increment calculation module 32 of the same series branch to calculate the control increment △u. The size of △u is related to the historical error at time t. The calculation number l is set, and the value range of l is between 1 and t+1. The radial displacement output by the fuzzy neural network model 51 at t=1 is recorded as x″(l). The calculation formula of △u is:

[0054] Where d(l) is the error weight parameter, which ranges from 0 to 2 and can be set according to the actual control effect.

[0055] The control increment △u(t) output by the control increment module 32 is input to the prediction value calculation module 33 of the same series branch. The prediction value calculation module calculates the final radial prediction value x at time t+1 according to the formula c (t+1). The formula is: x c (t+1)=x″ c (t+1)+λ△u(t) (7)

[0056] Where λ is the incremental weight parameter value, ranging from 0 to 2, and is set according to actual conditions.

[0057] The other two IDBO-FNN prediction modules 34 and 38 process the input values ​​in the same way as the IDBO-FNN prediction module 31, and the corresponding prediction value calculation module 36 outputs the final radial prediction value y c (t+1). The prediction value calculation module 3A outputs the final speed prediction value n c (t+1). The radial prediction value x c (t+1),y c (t+1) is used as the input of the composite signal calculation module 37, which processes the two input prediction values ​​to obtain the rotor eccentric displacement signal at time t+1. Where a is a weight parameter, and the value range of a is 0 to 1. The composite signal calculation module 37 and the prediction value calculation module 3A respectively output the rotor eccentric displacement signal r and the speed control signal n. c To IDBO-FNN system 2.

[0058] As shown in FIG4 , the IDBO parameter optimization modules 21, 51, 53, and 55 of the present invention all use the improved dung beetle optimization algorithm to optimize the fuzzy neural network membership function width. and the center value and weights The improved dung beetle algorithm is optimized as follows:

[0059] Based on the original dung beetle algorithm, an adaptive weight factor is introduced into the position update formula of the chicks and stealing dung beetles. The algorithm reduces the probability of falling into a local optimum. Secondly, it integrates the follower position update mechanism in the sparrow search algorithm to perturb the algorithm and uses a greedy strategy to update the position, which improves the convergence accuracy of the algorithm. Finally, when the algorithm stagnates, it introduces the Cauchy-Gauss mutation strategy to improve the algorithm's ability to escape the local optimal solution. The specific implementation method is as follows:

[0060] Step 1: Initialize the population and set the population size to 100. There are four populations, namely, rolling dung beetles, breeding dung beetles, small dung beetles and thieving dung beetles. There are three parameters to be optimized in the fuzzy neural network, so the optimization dimension D is set to 3, that is, D=3. The initial positions of each population are random, and the position coordinates are the parameters to be optimized in the fuzzy neural network: the width of the membership function and the center value and weights The maximum number of iterations T is set to 500, the judgment parameter α is initialized, the population size of each dung beetle species is set to pop = 25, and the optimal fitness value is set to 9.5.

[0061] Step 2: Determine whether the number of iterations has reached the maximum value or whether the individual has reached the optimal fitness of 9.5. If not, proceed to step 3. If so, terminate the iteration. The population fitness formula is as follows, taking the IDBO parameter optimization module 52 as an example:

[0062] Where, y(X i 1test ) is the fuzzy neural network model 51 when the input is the test set input signal X i 1test The predicted output value when X i 1test is the output value in the test set, m1 is the number of test set data, and m=200 in the present invention.

[0063] Step 3: Determine the population of the current individual. If the individual is a ball-rolling dung beetle, proceed to step 4. If the individual is not a ball-rolling dung beetle, proceed to step 5.

[0064] Step 4: Update the position according to the rolling ball dung beetle position update algorithm. After the position update is completed, jump to step 10. The rolling ball dung beetle position update formula is as follows:

[0065] Where σ is a random number between (0, 1), ξ represents the probability of a dung beetle encountering an obstacle, ξ∈(0,1), k is the deviation coefficient, k∈(0,0.02], b is a constant, b∈(0,1), △X represents the light intensity, △X=|X i -X w |, where X w represents the worst position in the population, represents the deviation coefficient, when hour, Meaningless, the position is not updated.

[0066] Step 5: If the individual is a juvenile dung beetle, proceed to step 6; if the individual is not a juvenile dung beetle, proceed to step 7.

[0067] Step 6: Update the position according to the position update algorithm of the young dung beetle. After completing this operation, jump to step 10. The breeding dung beetles will move within the safe area. The safe area is defined as follows: Lb * =max(X * ×(1-R),Lb) (11) Ub * =min(X * ×(1+R),Ub) (12)

[0068] Where, X * Represents the optimal position of the group in this iteration, R = t / T, Lb * , Ub * They represent the lower and upper bounds of the safe area, Lb and Ub represent the lower and upper bounds of the feasible area, respectively.

[0069] The position update formula of the chicks laid by the breeding dung beetle is improved by introducing an adaptive weight factor Constructing adaptive weight factors

[0070] Where υ = exp(cos(πt / T) × rand(-1, 1)), t is the current iteration number, which is consistent with the current time t, T is the maximum number of iterations, and rand(-1, 1) is a random number.

[0071] The improved position update formula of young dung beetles is:

[0072] Where b1 and b2 are independent random vectors of 1×D, and D is the dimension of the optimization parameter. In this case, D is 3.

[0073] Step 7: If the individual is a small dung beetle, proceed to step 8; if the individual is not a small dung beetle, proceed to step 9.

[0074] Step 8: Update the position according to the position update formula of the dung beetle, and then jump to step 10. The position update of the dung beetle when foraging is as follows: Lb b =max(X b ×(1-R),Lb) (16) Ub b =max(X b ×(1-R),Ub) (17) X i (t+1)=X i (t)+C1×(X i(t)-Lb b )+C2×(X i (t)-Ub b ) (18)

[0075] Where, X b is the global optimal position, Lb b , Ub b are the lower and upper bounds of the optimal foraging, C1 is a random vector obeying the normal distribution, and C2 is a random vector with a value range of (0, 1).

[0076] Step 9: The individual is a thieving dung beetle. It updates its position according to the thieving dung beetle position update formula. After the update is completed, jump to step 10. The thieving dung beetle will steal near the global optimal position. Add an adaptive weight factor to the thieving dung beetle position update formula. The improved position update formula is as follows:

[0077] Where g is a random vector that obeys the normal distribution, and S represents a constant.

[0078] Step 10: Generate a random number r1∈(0,1), set the perturbation probability pv, t is the current iteration number. Determine whether r1 is less than or equal to pv. If so, proceed to step 11. If so, proceed to step 15.

[0079] Step 11: Generate a random number r2∈(0,1), set a constant S, and determine whether r2 is less than ST, where T is the maximum number of iterations. If so, proceed to step 12; if so, proceed to step 13. In the present invention, ST takes a value of 0.65.

[0080] Step 12: Update the first formula in the follower position formula based on the sparrow search algorithm: Update the location and proceed to step 14 after the update.

[0081] After the dung beetle population position is updated, the follower position update mechanism in the sparrow search algorithm is introduced into the dung beetle optimization algorithm as a perturbation strategy, allowing the group to more fully traverse the solution space and improve population quality. Specifically, based on the perturbation probability pv, when perturbation is required, the sparrow algorithm follows the best-positioned individual in the four populations. The optimal position of the finder in the sparrow search algorithm is defined to correspond to the global optimal position in the dung beetle optimization algorithm. The followers are responsible for following the finder in search of food. When perturbation is not required, the optimal position of the population after iteration is directly updated using the Cauchy-Gauss formula mutation and the next round of iteration. The follower position is updated as follows:

[0082] Where Q is a random number that obeys the normal distribution, Xp (t+1) represents the optimal position of the current discoverer, X w Indicates the global worst position, A is a 1×d matrix, the elements in the matrix are 1 or -1, A + =A T (AA T ) -1 , L represents a 1×d matrix with all elements set to 1. When i>pop / 2, it indicates that the i-th follower has not obtained food, has a low fitness, and is in a hungry state. At this time, it needs to fly to other places to find food.

[0083] Step 13: Update the second formula X in the formula based on the follower position in the sparrow search algorithm p (t+1)+|X i (t)-X p (t+1)|×A + ×L performs location update, and then proceeds to step 14.

[0084] Step 14: Determine whether the obtained position remains unchanged for five consecutive iterations. If so, proceed to step 15; if not, proceed to step 16.

[0085] Step 15: Mutate the individual according to the Cauchy-Gauss formula, and then proceed to step 16.

[0086] In order to solve the problem that the dung beetle algorithm is prone to falling into the local optimal solution when there are multiple peaks, the Cauchy-Gauss mutation is introduced to increase the probability of the algorithm jumping out of the local optimal solution. Specifically, if the optimal position of the population remains unchanged after five iterations, Gauss mutation is performed, otherwise the next round of iteration is performed. The Cauchy-Gauss mutation method is as follows: Xnew i (t) = X i (t)×[1+β1Cauchy(0,1)+β2Causs(0,1)] (21)

[0087] Where, Cauchy(0,1) is a random number that obeys the Cauchy distribution, Causs(0,1) is a random number that obeys the Gaussian distribution; Xnew i (t) is the position after mutation, β1=1-t 2 / T 2 ,β2=t 2 / T 2 .

[0088] Step 16: Repeat steps 2 to 15.

[0089] When the population fitness value requirement is met after 31 iterations, the fitness values ​​of the 100 populations are sorted and the population position coordinates of the population with the best fitness value are That is, the optimal membership function width and the center value and weights The data is input into the fuzzy neural network and trained offline using the gradient descent method to form a fuzzy neural network model.

[0090] As shown in Figure 5, the IDBO-FNN system 2 is constructed. Before designing the IDBO-FNN system 2, the reversibility analysis of the permanent magnet assisted bearingless synchronous reluctance motor must be performed first, as follows:

[0091] Establish the state equation of the permanent magnet assisted bearingless synchronous reluctance motor:

[0092] Where y Md 、y Mq are the magnetic flux of the stator torque winding on the d-axis and q-axis respectively; i Md 、i Mq are the torque currents of the stator torque winding on the d-axis and q-axis respectively; L Md 、L Mq are the synchronous inductances of the stator torque winding on the d-axis and q-axis respectively; f x 、f y are the disturbances received by the rotor in the x and y directions respectively; k x =k M -k L ; coefficient m is the rotor weight; ω m is the mechanical angular frequency of the permanent magnet assisted bearingless synchronous reluctance motor rotor; J is the moment of inertia of the rotor; T L represents the torque load on the rotor shaft; T e is the electromagnetic torque.

[0093] Select input variable u=[u1,u2,u3,u4] T =[i Md ,i Mq ,i Bd ,i Bq ] T , state variables And the output variable y = [y1,y2,y3] T =[x,y,n] T .

[0094] By taking multiple derivatives of the output variable until the input variable is explicitly included:

[0095] Will The Jacobian matrix is ​​obtained by differentiating the input variables:

[0096] The full rank of the matrix shows that the motor control system is reversible.

[0097] The IDBO-FNN system 2 is composed of a fourth IDBO parameter optimization module 21, a second-order difference processor 22, and a first-order difference processor 23 connected before the fourth fuzzy neural network model 24. The second-order difference processor 22 processes the input rotor eccentricity displacement signal r. The signal processing method is as follows: the first-order difference signal of the rotor eccentricity displacement signal r at time t The rotor eccentric displacement signal r(t-4) at time t-4, the rotor eccentric displacement signal r(t-3) at time t-3, the rotor eccentric displacement signal r(t-1) at time t-1 and the rotor eccentric displacement signal r(t) at time t are calculated using the following formula:

[0098] The second-order differential signal of the rotor eccentric displacement signal r at time t is The rotor eccentric displacement signal r(t-4) at time t-4, the rotor eccentric displacement signal r(t-3) at time t-3, the rotor eccentric displacement signal r(t-2) at time t-2, the rotor eccentric displacement signal r(t-1) at time t-1 and the rotor eccentric displacement signal r(t) at time t are calculated using the following formula:

[0099] The first-order difference processing of its 23 pairs of input speed control signals n c For processing, the signal processing method is as follows: the first-order difference signal at time t The speed control signal n at time t-4 c (t-4) and speed control signal n at time t-3 c (t-3) Speed ​​control signal n at time t-1 c (t-1) and the speed control signal n at time t c (t) is calculated, and the calculation formula is:

[0100] The fuzzy neural network model 24 is used to calculate the output signals of the IDBO-FNN dynamic prediction module 3, the second-order difference processor 22, the first-order difference processor 23 and the angle calculation module 4. The specific process is as follows:

[0101] The simulation model is established by using the permanent magnet assisted bearingless synchronous reluctance motor composite controlled object 1 given in Figure 2 and the IDBO-FNN dynamic prediction module 3 given in Figure 3. By inputting random current signals The motor rotor eccentric displacement r and speed n at time t+1 are obtained by sampling and processed by IDBO-FNN dynamic prediction module 3. c and the eccentric angle θ calculated by the angle calculation module 4, and at the same time, the rotor eccentric displacement signal r at time t+1 is numerically differentiated to obtain its first-order differential signal With the second-order differential signal The speed control signal n at time t+1 c Find the first-order difference signal composition As the input signal sample set of the fuzzy neural network model 24, denoted as X2, the random current signal As the output signal sample set of the fuzzy neural network model 24, denoted as Y2, a total of 1000 groups of sample data are sampled, of which 80% are used as the training set, denoted as X 2train 、Y 2train , 20% as the test set, denoted as X 2test 、Y 2test , the training set sample data is input into the fuzzy neural network model 24 for training, the input dimension h = 6, the fuzzy segmentation number d = 3, the fuzzy language number is set to 8, the membership function uses the Gaussian function, and its parameter training process is similar to the fuzzy neural network model 51 in IDBO-FNN31. The test set sample data is input into the IDBO parameter optimization module 21 for the initial weights in the fuzzy neural network model 24 and the central value in the initial membership function and width Perform parameter optimization. When the optimization is completed, a set of parameters with the best fitness value will be The fuzzy neural network model 24 is sent to perform a back propagation iteration, and then the trained fuzzy neural network model 24 is connected with the IDBO parameter optimization model 21, the second-order difference processor 22, and the first-order difference processor 23 to convert the input signal Converted into control current signal

[0102] The present invention first optimizes the predecessor's fuzzy neural network algorithm, and uses the IDBO algorithm to optimize the weights and initial values ​​of the membership function of the new fuzzy neural network, establishes the controlled object 1 shown in Figure 2, and then establishes the IDBO-FNN dynamic prediction module 3 shown in Figure 3, so that the model prediction is almost accurate in line with the actual controlled process, and finally establishes the IDBO-FNN system 2 shown in Figure 5. Finally, the IDBO-FNN dynamic prediction module 3, the IDBO-FNN inverse system 2 angle calculation module 4 and the three-phase inverter circuit are connected in series, and a closed-loop control is designed to form a complete permanent magnet assisted bearingless synchronous reluctance motor IDBO-FNN inverse decoupling controller.

Claims

1. A method for constructing a decoupling controller for a permanent magnet assisted bearingless synchronous reluctance motor, characterized in that The following steps are involved: Step A): an IDBO-FNN dynamic prediction module (3) is formed by three parallel IDBO-FNN prediction modules, three parallel control increment calculation modules, three parallel prediction value calculation modules and a composite signal calculation module; an IDBO-FNN prediction module, a control increment calculation module and a prediction value calculation module are sequentially connected in series to form a series branch, which are two rotor radial displacement branches and a speed branch. The speed branch obtains the speed control value n c , the outputs of the two rotor radial displacement branches are used to obtain the rotor eccentric displacement r through the composite signal calculation module; Step B): The IDBO-FNN system (2) is formed by connecting the second-order differential processor and the first-order differential processor of the fourth IDBO parameter optimization module before the fourth fuzzy neural network model. The rotor eccentric displacement r is processed by the second-order differential processor to obtain the first-order and second-order displacement differential signals, and the speed control variable n is c Obtain the first-order speed differential signal through the first-order differential processor; Rotor eccentric displacement r, speed control value n c The membership function width, center value, weight, first-order and second-order displacement differential signals of the fourth fuzzy neural network model and the first-order speed differential signal are used as inputs of the fourth fuzzy neural network model, and the fourth fuzzy neural network model outputs a control current signal; Each of the IDBO-FNN prediction modules is composed of a corresponding fuzzy neural network model connected to a corresponding IDBO parameter optimization module; each of the IDBO parameter optimization modules uses the following improved dung beetle optimization algorithm to optimize the membership function width, center value and weight of the fuzzy neural network: Step 1: Set four populations: rolling dung beetles, breeding dung beetles, small dung beetles, and stealing dung beetles, with the optimization dimension D set to 3, the maximum number of iterations T, the initialization of the judgment parameter α, the number of each dung beetle population, and the optimal fitness; Step 2: Determine whether the number of iterations has reached the maximum value or whether the individual has reached the optimal fitness. If not, proceed to step 3. If so, terminate the iteration. Step 3: Determine the population of the current individual. If the individual is a dung beetle, proceed to step 4. If the individual is not a dung beetle, proceed to step 5. Step 4: Update the position according to the rolling ball dung beetle position update algorithm; Step 5: If the individual is a juvenile dung beetle, proceed to step 6; if the individual is not a juvenile dung beetle, proceed to step 7; Step 6: Update the position of the young dung beetle, and then jump to step 10; Step 7: If the individual is a small dung beetle, proceed to step 8; if the individual is not a small dung beetle, proceed to step 9; Step 8: Update the position of the dung beetle, then jump to step 10; Step 9: If the individual is a thieving dung beetle, update the location of the thieving dung beetle and then jump to step 10; Step 10: Generate a random number r1∈(0,1), set the perturbation probability pv, and determine whether r1 is less than pv. If it is less than or equal to pv, proceed to step 11; if it is greater than pv, proceed to step 15. Step 11: Generate a random number r2∈(0,1), set a constant S, and determine whether r2 is less than ST. If so, proceed Step 12: If it is greater than ST, proceed to step 13; Step 12: Update the position of the follower according to the sparrow search algorithm; if the obtained position remains unchanged for five consecutive iterations, mutate the individual according to the Cauchy-Gauss formula; Step B): constructing a composite controlled object including a permanent magnet assisted bearingless synchronous reluctance motor; Step C): The IDBO-FNN dynamic prediction module (3) and the IDBO-FNN system (2) form a decoupling controller connected to the composite controlled object.

2. The construction method according to claim 1, wherein: The fuzzy neural network model is divided into five layers. The first layer is the input layer. The input is a vector X. The dimension of X is h. X=[X1,X2........,X h ] T , the input of the fuzzy neural network model is the displacement difference △x, X=[△x] T , the number of nodes in the first layer is N = 1; The second layer is the separation layer, which is divided into p subsets, and the input vector X is equally divided into components with a dimension of d = 2. The number of nodes in the second layer is The third layer is the fuzzification layer. The only input △x in the segmentation layer is the input △x in its sth i Membership degree on fuzzy linguistic value Membership function m is the number of fuzzy languages ​​for each input, For the s i The central value of the membership function, For the s i The width of the membership function; The number of nodes in the third layer is N′×m; The fourth layer is the fuzzy rule layer, each rule strength node The number of nodes in the fourth layer is p×m; The fifth layer is the output layer, which predicts the radial displacement of the rotor r is 1, is the weight.

3. The construction method according to claim 2, wherein: Optimal fitness y(X i 1test ) is the fuzzy neural network model when the input is the test set input signal X i 1test The predicted output value when X i 1test is the output value in the test set, and m1 is the number of test set data.

4. The construction method according to claim 3, wherein: The position update formula of the rolling ball dung beetle is: σ is a random number between (0, 1), ξ is the probability of the dung beetle encountering an obstacle, ξ∈(0,1), k is the deviation coefficient, k∈(0,0.02], b is a constant, b∈(0,1), △X is the light intensity, △X=|X i -X w |, X w represents the worst position in the population, represents the deviation coefficient, when π / 2,π time, Meaningless, the position is not updated.

5. The construction method according to claim 4, characterized in that: The breeding dung beetles move in a safe area, which is defined as: Lb * =max(X * ×(1-R),Lb),Ub * =min(X * ×(1+R),Ub),X * Represents the optimal position of the group in this iteration, R = t / T, Lb * , Ub * They represent the lower and upper bounds of the safe area, Lb and Ub represent the lower and upper bounds of the feasible area, respectively.

6. The construction method according to claim 5, characterized in that: The position update formula of the young dung beetle is: Adaptive weighting factor υ=exp(cos(πt / T)×rand(-1,1)), where t is the current iteration number, rand(-1,1) is a random number, b1 and b2 are independent random vectors of 1×D, and D is 3.

7. The construction method according to claim 6, characterized in that: The position update formula of the dung beetle when foraging is: Lb b =max(X b ×(1-R),Lb), Ub b =max(X b ×(1-R),Ub), X i (t+1)=X i (t)+C1×(X i (t)-Lb b )+C2×(X i (t)-Ub b ), X b is the global optimal position, Lb b , Ub b are the lower and upper bounds of the optimal foraging, C1 is a random vector obeying the normal distribution, and C2 is a random vector with a value range of (0, 1).

8. The construction method according to claim 7, characterized in that: Perturbation probability 9. The construction method according to claim 8, characterized in that: The follower position update formula of the sparrow search algorithm is: Q is a random number that follows a normal distribution, X p (t+1) represents the optimal position of the current discoverer, X w Indicates the global worst position, A is a 1×d matrix, the elements in the matrix are 1 or -1, A + =A T (AA T ) -1 , L represents a 1×d matrix with all elements set to 1. When i>pop / 2, it indicates that the i-th follower has not obtained food, has a low fitness, and is in a hungry state. At this time, it needs to fly to other places to find food.

10. The construction method according to claim 9, characterized in that: Cauchy-Gauss variant is Xnew i (t) = X i (t)×[1+β1Cauchy(0,1)+β2Causs(0,1)] Cauchy(0,1) is a random number that follows the Cauchy distribution, and Causs(0,1) is a random number that follows the Gaussian distribution; Xnew i (t) is the position after mutation, β1=1-t 2 / T 2 , β2=t 2 / T 2 .

Images