Five-phase permanent magnet motor multi-objective optimization method and control system thereof

Abstract

The invention discloses a five-phase permanent magnet motor multi-objective optimization method and a control system thereof, and belongs to the technical field of motor body design optimization and motor control. According to the multi-objective optimization method for the five-phase permanent magnet motor, the optimal optimization value of the motor parameters is obtained based on the genetic aggregation response surface and the multi-objective genetic algorithm, the design variables of the motor are comprehensively considered, and the optimization accuracy is improved; according to the control system of the five-phase permanent magnet fault-tolerant motor, iterative learning control serves as a current control technology and is used for recovering the performance of the five-phase permanent magnet motor under the fault condition, the advantages of iterative learning control and counter electromotive force control are combined, high-performance operation of the motor can be ensured under the fault condition, and the reliability of the five-phase permanent magnet fault-tolerant motor is improved. Ripple-free torque output is realized, and ohmic loss is minimized. The fault diagnosis module of the control system diagnoses short-circuit faults and open-circuit faults by using a time sequence neural network and a Kalman filter, and has real-time performance and strong adaptive ability.

Description

Technical Field

[0001] The present invention relates to the field of motor body design optimization and motor control technology, and in particular to a five-phase permanent magnet synchronous motor multi-objective optimization method and a control system thereof. Background Art

[0002] Five-phase permanent magnet motors play an important role in many industrial applications, but in their design process, multiple objectives need to be considered simultaneously, such as average torque, efficiency, and torque ripple, to improve the performance of the motor. Traditional optimization methods have the problems of high computational cost, low accuracy, or difficulty in finding the global optimal solution.

[0003] In addition to the optimization method of the motor body design, the safety and reliability of the electric drive system are also very critical to the motor. Short circuit or open circuit faults may occur during the operation of the five-phase permanent magnet motor, which will seriously affect the performance and reliability of the motor and even cause the system to shut down. Therefore, fault diagnosis and fault-tolerant control are crucial to ensure the safe and reliable operation of the five-phase permanent magnet motor.

[0004] Traditional fault diagnosis methods have problems such as being susceptible to noise interference, being insensitive to early fault diagnosis, and having difficulty in feature extraction. In response to this, various universities and research institutes have proposed some new fault diagnosis methods in recent years, but there are still more or less problems. For example, the document "A Five-Phase Permanent Magnet Motor Turn-to-Turn Short Circuit Fault Diagnosis Method" (Authorization Publication No.: CN116626491B) discloses a method for diagnosing five-phase permanent magnet motor turn-to-turn short circuit faults using a multi-scale convolutional residual network with an attention mechanism. Although it has a relatively high diagnostic accuracy, it relies on the Bayesian optimization algorithm to automatically select the optimal hyperparameter combination, and there is a risk of unstable algorithm convergence. In addition, the data enhancement method using a specific equally spaced sliding window to extract features has certain limitations and cannot adapt to all data characteristics and fault modes.

[0005] Therefore, a fault diagnosis method with real-time performance and high computational efficiency is needed to overcome existing defects, determine the fault type, and perform fault-tolerant control. Five-phase permanent magnet motors have advantages in fault-tolerant operation, but traditional fault-tolerant control schemes have limitations. They usually rely on accurate back-EMF models and complex fault detection, and need to consider torque and current measurement errors. Therefore, learning and repetitive control algorithms have gradually attracted attention. Among them, iterative learning control has great potential in dealing with periodic torque pulsation under motor fault conditions, providing a new idea for fault-tolerant control of motors. Summary of the invention

[0006] In view of the deficiencies of the prior art in the above-mentioned background technology, the purpose of the present invention is to provide a five-phase permanent magnet motor multi-objective optimization method and a control system thereof, so as to solve the problems of high computational cost, low precision and difficulty in finding a global optimal solution of the five-phase permanent magnet motor optimization method, and to improve the safety and reliability of the five-phase permanent magnet motor control system.

[0007] The technical solution to achieve the purpose of the present invention is:

[0008] A five-phase permanent magnet motor multi-objective optimization method, the five-phase permanent magnet motor multi-objective optimization method comprising the following steps:

[0009] Step S1-1, clarifying the optimization objectives of the five-phase permanent magnet motor, including average torque, efficiency, and torque pulsation performance parameters, and determining corresponding objective functions and constraints; the objective function is used to maximize efficiency, minimize torque pulsation, and improve average torque; the constraints include the number of stator winding turns, permanent magnet thickness, air gap length, rotor pole pair number, inductance, and resistance;

[0010] Step S1-2: Based on the optimization target, the central composite design method is used to generate sample points, determine the number and location of design points, and record the data of relevant parameters of each design point;

[0011] Step S1-3, using GARS to perform sensitivity analysis on the generated sample points, quantify the sensitivity of the motor response to each design feature, and evaluate the sensitivity index of the design variable to the optimization target; the GARS is a genetic aggregation response surface; the design features are various factors that affect the performance of the motor, including stator outer diameter, permanent magnet thickness, air gap length, centrifugal height, and notch depth; the design variables are specific parameters for quantifying these design features;

[0012] Step S1-4: Based on the results of the sensitivity analysis, a response surface model is constructed based on the weighted average of GARS, and the model includes multiple different meta-models to accurately describe the relationship between the design variables in the design space; through the response surface model, the influence of different combinations of design variables on the optimization target is characterized, providing a basis for multi-objective optimization; the design space is a set consisting of the value ranges of all design variables;

[0013] Step S1-5, based on the NSGA-II algorithm, multi-objective optimization is performed on the average torque, efficiency and torque ripple of the five-phase permanent magnet motor, a non-dominated solution set is generated under the constraints, the balance of the optimization objectives is achieved, and the optimization results are verified by finite element analysis; the NSGA-II is a non-dominated genetic algorithm.

[0014] Furthermore, the central composite design method of step S1-2 specifically includes the following steps:

[0015] Step S1-2-1, determining the design variables related to the performance of the five-phase permanent magnet motor and their value ranges;

[0016] Step S1-2-2, select the design points of the center point, the axial point and the star point; the center point is located at the center of the design space and is used to estimate the experimental error to determine the benchmark motor performance; the axial point is located at the boundary of the value range of the design variable to ensure that the boundary conditions of the motor performance are fully considered; the star point is determined by setting the position of the asterisk arm and is used to estimate the nonlinear response of the design variable and evaluate the influence of the curvature of the design variable on the motor performance;

[0017] Step S1-2-3, based on the design points of the center point, axial point and star point selected in step S1-2-2, generate corresponding sample points; the sample points can cover the design space and are used for experiments and simulations to obtain motor performance data, and then analyze the relationship between design variables and motor performance, providing data support for motor optimization.

[0018] Further, the step S1-3 uses a minimum-maximum search algorithm to search for the minimum and maximum values ​​of each output parameter and its corresponding input parameter to comprehensively evaluate the response values ​​of all variables, thereby assessing the impact of the design variables on the sensitivity index of the output parameters;

[0019] Furthermore, the functional relationship between the global objective function and the constraint function in the response surface design space in step S1-4 is:

[0020]

[0021] Among them, y(x) represents the relationship between the global objective function and the constraint function. represents the basic function, α i represents the weighting parameter, N m Indicates the number of metamodels;

[0022] The response surface model solves the complex nonlinear relationship between the design variables and the average efficiency, torque, and torque ripple by a quadratic approximation method. The prediction of the i-th response of the response surface model is expressed as:

[0023]

[0024] in, represents the predicted value of the ith response, a0 represents the baseline value of the response variable when all independent variables are zero, and a i Indicates the linear influence of the ith independent variable on the response, a ii Indicates the nonlinear influence of the square term of the ith independent variable on the response, a ijIt is expressed as the degree of influence of the interaction between the ith independent variable and the jth independent variable on the response, and all a are determined by least squares regression:

[0025] a=(X T X) -1 X T y

[0026] The GARS used in step S1-4 is represented as a set using the weighted average of several different meta-models:

[0027]

[0028] Among them, w i satisfy:

[0029]

[0030] Furthermore, the NSGA-II algorithm in step S1-5 includes the following steps:

[0031] Step S1-5-1, generate the parent population P0, whose size is N P ; Then, the offspring population Q0 is generated through crossover and mutation operations; the subsequent stage is denoted as the nth generation;

[0032] Step S1-5-2: Generate combined population R n =P n ∪Q n , population R n The scale is 2N P ;

[0033] Step S1-5-3: Combined population R n Decode and determine the fitness of the objective function; use the non-dominated solution sets L1, L2, and L3 to decode the population R n Sort them; the solution belonging to the best non-dominated solution set L1 is the best solution, followed by the solution in the set L2, and the solution selected from the set L3 is ranked last;

[0034] Step S1-5-4: When the total number of members of the selected set is equal to or exceeds N P When , select the initial set; calculate the crowding distance of each member in the selected set; for the next generation P n +1, select the first k1 sets according to the elitist principle; sort the kth set according to the crowding distance, select the member with the larger crowding distance and add it to P n In the +1 generation, until the population reaches N P; The elitism principle is a commonly used strategy in genetic algorithms, which aims to retain individuals with higher fitness in each generation to ensure that excellent genes can be passed on to the next generation, thereby accelerating the convergence speed of the algorithm and improving the performance of the algorithm;

[0035] The rule for selecting the initial set is given by the following formula:

[0036] g<h,if(g rank <h rank )or(g rank <h rank and(g dist <h dist ))

[0037] Among them, g and h represent two individuals used for comparison and selection in the algorithm, g rank and h rank denote the non-dominated ranks of members g and h, respectively. dist and h dist They represent the crowding distances of members g and h respectively;

[0038] Step S1-5-5, if the limit of the number of iterations is reached, the program ends and finds the fitness and non-dominated solution of the objective function; if not, go to step S1-5-6;

[0039] Step S1-5-6: For the new population P n +1 Perform crossover and mutation operations to generate a new population Q n +1, then enter step S1-5-2, and loop steps S1-5-2 to S1-5-5 until the fitness and non-dominated solution of the objective function are output.

[0040] A control system for controlling a five-phase permanent magnet motor optimized by the multi-objective optimization method comprises: a data acquisition module, a signal processing module, a rectifier, an inverter, a fault diagnosis module, a back electromotive force model module, an iterative learning controller module, a current controller module, and a five-phase permanent magnet motor;

[0041] The stator of the five-phase permanent magnet motor adopts a pole-slot combination of 10 slots and 8 poles, the winding is a double-layer fractional slot concentrated winding structure, the rotor is a surface permanent magnet rotor structure and the magnet adopts a centrifugal structure;

[0042] The control strategy of the five-phase permanent magnet motor control system includes the following steps:

[0043] Step S5-1, in the data acquisition module, the signal parameter values ​​of the five-phase permanent magnet motor are obtained in real time, including the current, voltage, rotor position, and torque signals of each phase, the stray magnetic flux around the motor is monitored by a fluxgate sensor, and the back electromotive force of the motor under different states is obtained by performing two-dimensional finite element operations on the motor, and finally the collected data is filtered and standardized for preprocessing;

[0044] Step S5-2, inputting the collected current signal parameters into the signal processing module, obtaining the alternating current on the αβ coordinate axis through coordinate transformation, and preprocessing the transformed current; the preprocessing includes standardization processing and window segmentation;

[0045] Step S5-3, input the pre-processed current into the fault diagnosis module, predict the current by using the transformer neural network and the model-based method respectively, and compare it with the actual current to determine the motor operation state; the operation state includes normal operation state, open circuit fault state and short circuit fault state;

[0046] Step S5-4: when the motor is running in a fault state, a back-electromotive force model module is established according to the back-electromotive force signal in step S5-1, and an optimal current value is calculated according to different objective functions of open circuit fault and short circuit fault to provide an initial optimal reference current;

[0047] Step S5-5, the iterative learning controller module updates the healthy phase current according to the optimal current in step S5-4 and the torque error calculated in combination with the torque signal in step S5-1;

[0048] Step S5-6, inputting the updated current into the current controller to generate a gate drive signal for the voltage source inverter so that the stator phase current can follow the optimal reference current;

[0049] Step S5-7: Perform two-dimensional finite element analysis on the motor to obtain torque pulsation, current distribution and ohmic loss data of the motor under different fault conditions, and use the data to evaluate the effect of the control scheme.

[0050] Further, the fault diagnosis module in step S5-3 includes a short circuit fault diagnosis module and an open circuit fault diagnosis module;

[0051] The overall process of the short-circuit fault diagnosis module is as follows: the pre-processed current signal is input into the TNN-based diagnosis model, and the diagnosis model includes multiple encoders, each encoder includes an MHA module and an FF module; the input signal is encoded by the MHA module and the FF, and an output tensor is obtained after hierarchical processing by multiple encoders; the output tensor is used to estimate the number of short-circuited coils and the short-circuit current, and then the third harmonic component of the stray magnetic flux is used to detect and locate the short-circuit fault, and finally the number of short-circuited coils, the short-circuit current amplitude and the stray magnetic flux analysis results are comprehensively estimated to determine the short-circuit fault and its degree; the TNN is a transformer neural network module, the MHA is a multi-head attention module, and the FF is a fully connected feedforward network module;

[0052] The MHA module calculation process is as follows:

[0053] Step S6-1-1: Perform a linear transformation on the input sequence to obtain a query matrix Q = XW Q , key matrix K = XW K Sum matrix V = XW V , where X = [x1, x2, …, x Ts ] T is the input vector matrix, t=1,2,…,Ts; is the trainable weight matrix, d feat is the number of input features, Ts is the time step;

[0054] Step S6-1-2, calculate self-attention according to the formula, and its function is used to normalize the attention score;

[0055] Step S6-1-3, perform calculations through multiple parallel self-attention layers, i.e., multiple heads, and concatenate and linearly transform the results to obtain the output of the MHA module;

[0056] The FF module specifically includes two linear transformation layers, and residual connection and normalization processing are performed before and after each linear layer respectively;

[0057] The FF module calculation process is as follows:

[0058] Step S6-2-1: The input of the first linear transformation layer is X e ', the output is GELU(X e 'W FF1 +b FF1 ),in are the trainable weights of the fully connected feedforward network, is the bias of the fully connected feedforward network, GELU is the Gaussian error linear unit activation function;

[0059] Step S6-2-2: The input of the second linear transformation layer is the output of the first linear transformation layer, and the output is GELU (X e 'W FF1 +b FF1 )W FF2 +b FF2 ,in are the trainable weights of the fully connected feedforward network, is the bias of the fully connected feed-forward network;

[0060] The overall process of the open circuit fault diagnosis module is as follows: using a model-based approach to perform fault detection, specifically including using an extended Kalman filter to estimate the current of the motor and establish a discretized state space model;

[0061] The open circuit fault diagnosis steps are as follows:

[0062] Step S6-3-1, using an extended Kalman filter to estimate the current of the motor;

[0063] Step S6-3-2: Establish a discretized state space model, which is divided into a prediction module and an update module:

[0064] In the prediction module: is the estimated value of the state at time k at time k-1; is the estimated value of the state at time k-1, u k-1 is the input at time k-1; is the state transfer function, P k|k-1 =F k P k-1|k-1 F k T +Q k-1 is the estimated error covariance matrix of time k-1 to time k, where I is the identity matrix, is the system matrix, Where L d , L q are the direct-axis inductance and quadrature-axis inductance, respectively, PM is the permanent magnet flux, T s is the sampling time, F k is the Jacobian matrix of A, P k-1|k-1 is the estimated error covariance matrix at time k-1, Q k-1 is the covariance matrix of process noise;

[0065] In the update module: is the Kalman gain, where H k is the measurement matrix, R k is the covariance matrix of the measurement noise; is the updated state estimate at time k, where h(x) is the measurement function and y k is the measured value at time k; P k|k =(IK k H k ) k|k-1 is the estimated error covariance matrix after updating at time k; where x = [i d i q ] T ,u=[V d V q ω e ] T ,h(x (k) )=Cx (k) ,i d 、i q are the direct-axis current and quadrature-axis current, V d 、V q is the direct-axis voltage and quadrature-axis voltage, ω e is the electrical angular velocity, is the transformation matrix; calculate the residual: It is applied to the CUSUM algorithm for fault detection. The CUSUM algorithm detects the occurrence of faults by accumulating the change of residuals. At the same time, the whiteness test is used to determine whether the sensor is faulty. The whiteness test calculates the autocorrelation function of the residuals. And the normalized ACF: Where N is the number of data samples, ε t is the residual at time t, is the mean of the residuals; if the residuals exceed the set threshold or the whiteness test shows an abnormality, it is determined that a fault exists; the CUSUM algorithm is a cumulative sum algorithm.

[0066] Furthermore, the optimal current in step S5-4 is obtained by first taking the electromagnetic torque without ripple and with minimum ohmic loss under fault conditions as the optimization target, and then using a closed-form equation to calculate the reference optimal current according to the back electromotive force model when the influence of the fault type is represented by a lookup table;

[0067] The optimal current calculation for the open circuit fault condition is as follows:

[0068] The objective function to minimize ohmic loss is:

[0069]

[0070] in represents the phase current vector, k represents the speed normalized back EMF vector, T* represents the desired torque, F represents the fault matrix whose elements are used to represent the constraints imposed by the stator winding connection and the fault location, and p1 and p2 are Lagrange multipliers;

[0071] Find the optimal current for open circuit fault as:

[0072]

[0073] Among them, F -1 is the left inverse of the matrix, defined as F -1 =(F T F) -1 F T , F' is given by F'=(I n -FF l -1 ) is given;

[0074] The optimal current calculation under the short-circuit fault condition is as follows:

[0075] The open-circuit fault-tolerant control is extended to short-circuit faults, and the modified objective function is defined as:

[0076]

[0077] Among them, i sc New parameters introduced to consider the effect of short-circuit current;

[0078] Find the optimal short-circuit fault current as:

[0079]

[0080] Among them, F τ -T is the right inverse of the matrix, defined as F τ -T =F(F T F) -1 ,i sc Indicates that the short-circuit related current is equal to zero in healthy operation and is equal to zero in single-phase fault. Given.

[0081] Furthermore, the specific method of updating the current in step S5-5 is to calculate based on the system equation, motor equation performance, error signal and learning algorithm, and suppress the torque pulsation with reference to the current update law;

[0082] The system equation is:

[0083] y j (m) = P(q)u j (m)+d(m)

[0084] Where m is the time index, j is the iteration index, q is the forward time shift operator qx(m)=x(m+1), y j is the output, u jis the control input, d is the repeated exogenous signal;

[0085] The motor equation is:

[0086] T m =k T i+T d

[0087] Among them, T m Represents the output torque, T d Indicates oscillating torque disturbance;

[0088] The error signal is:

[0089] e j (m) = y d (m)-y j (m)

[0090] Among them, y d is the expected output;

[0091] The learning algorithm is:

[0092] u j+1 (m) = Q(q)[u j (m)+L(q)e j (m+1)]

[0093] Among them, Q(q) and L(q) are defined as Q filter and learning function respectively, e j =T*-T m,j , T m,j represents the motor output torque of the jth iteration, T* represents the expected torque;

[0094] The iterative learning current update formula is:

[0095] i * ILC,j+1 =w+β(F'k(k T F'k) -1 )(T*-T m,j )

[0096] Among them, i * ILC,j represents the iterative learning current reference of the jth iteration, and β is a constant.

[0097] Furthermore, the current controller in step S5-6 implements the current following process as follows:

[0098] Step S5-6-1, the current controller monitors the actual value of the stator phase current in real time and compares it with the optimal current reference;

[0099] Step S5-6-2: Based on the comparison result, the current controller adjusts the gate drive signal to ensure that the stator phase current quickly and accurately tracks the optimal reference current.

[0100] Furthermore, the two-dimensional finite element analysis method in step S5-7 is as follows:

[0101] Step S5-7-1, establishing a two-dimensional geometric model according to the number of slots, inner diameter, outer diameter of the stator, the size of the rotor, the position and size parameters of the permanent magnet in the actual structure of the motor;

[0102] Step S5-7-2, defining the material properties of each part of the motor, wherein the material includes the material of the stator core, the material of the permanent magnet, and the material of the winding;

[0103] Step S5-7-3, meshing the model, discretizing the geometric model into 20,000 units according to accuracy requirements and calculation time limits;

[0104] Step S5-7-4, applying boundary conditions to the geometric model, defining the outer boundary of the motor as a magnetic insulation boundary condition, simulating the air environment around the motor, applying permanent magnet excitation to the motor and applying updated current excitation to achieve fault-tolerant control;

[0105] Step S5-7-5, using the two-dimensional finite element method to solve the partial differential equation of the motor's electromagnetic field, to obtain the distribution of the electromagnetic field quantity inside the motor, the electromagnetic field quantity including the magnetic field distribution and the magnetic flux density, thereby obtaining the torque pulsation, current distribution and ohmic loss data.

[0106] Compared with the prior art, the multi-objective optimization method for a five-phase permanent magnet motor and the control system thereof according to the present invention have the following beneficial effects:

[0107] 1. The multi-objective optimization method for a five-phase permanent magnet motor proposed in the present invention uses GARS for sensitivity analysis, which can more reliably determine important design factors and build an accurate response surface model, thereby providing more precise guidance for optimization. Compared with traditional methods, GARS takes into account the interaction between design variables and improves the reliability of the optimization design model through numerous response surfaces and cross-validation procedures;

[0108] 2. The NSGA-II algorithm is used for optimization in the multi-objective optimization method of a five-phase permanent magnet motor proposed in the present invention. The algorithm performs better in finding multiple solutions and converging to the exact Pareto optimal set, and can find a wider range of design solutions, providing more options for the design of five-phase permanent magnet motors.

[0109] 3. The fault diagnosis module of a five-phase permanent magnet motor control system proposed in the present invention performs short-circuit fault diagnosis based on TNN, which has advantages in accuracy and adaptability, and has the advantages of real-time performance and high computational efficiency; the model-based method performs open-circuit fault diagnosis, utilizes existing current sensors, does not require additional hardware, and is easy to apply to various occasions;

[0110] 4. The five-phase permanent magnet motor control system proposed in the present invention combines the advantages of back electromotive force control and iterative learning control, can ensure high-performance operation of the motor under fault conditions, achieve ripple-free torque output, and minimize ohmic losses;

[0111] 5. The five-phase permanent magnet motor control system proposed in the present invention does not require overly complex fault detection and diagnosis algorithms, can respond to faults quickly, and eliminate repeated torque pulsations caused by model mismatch, making the current close to the optimal solution, while ensuring the fault tolerance and stability of the system;

[0112] 6. The five-phase permanent magnet motor control system proposed in the present invention can be extended to multi-phase permanent magnet motors with any number of phases and any stator winding connections, and has wide applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 It is a flow chart of the multi-objective optimization method of a five-phase permanent magnet motor proposed by the present invention;

[0114] Figure 2 It is a control system topology diagram based on the combination of back-EMF control and iterative learning control;

[0115] Figure 3 It is a flowchart of a short-circuit fault diagnosis method for a control system of a five-phase permanent magnet motor proposed by the present invention;

[0116] Figure 4 It is a flowchart of an open circuit fault diagnosis method for a control system of a five-phase permanent magnet motor proposed by the present invention. DETAILED DESCRIPTION

[0117] The present invention is described below based on specific embodiments, but the present invention is not limited to the specific embodiments. In the detailed description of the present invention below, some specific details are described in detail. For those skilled in the art, the present invention can be fully understood without the description of these details.

[0118] A multi-objective optimization method for a five-phase permanent magnet motor, such as Figure 1 FIG. 1 is a flow chart of a multi-objective optimization method for a five-phase permanent magnet motor according to the present invention, wherein the method comprises the following steps:

[0119] Step S1-1, clarifying the optimization objectives of the five-phase permanent magnet motor, including average torque, efficiency, and torque pulsation performance parameters, and determining corresponding objective functions and constraints; the objective function is used to maximize efficiency, minimize torque pulsation, and improve average torque; the constraints include the number of stator winding turns, permanent magnet thickness, air gap length, rotor pole pair number, inductance, and resistance;

[0120] Step S1-2: Based on the optimization target, the central composite design method is used to generate sample points, determine the number and location of design points, and record the data of relevant parameters of each design point;

[0121] Step S1-3, using GARS to perform sensitivity analysis on the generated sample points, quantify the sensitivity of the motor response to each design feature, and evaluate the sensitivity index of the design variable to the optimization target; the GARS is a genetic aggregation response surface; the design features are various factors that affect the motor performance; the design variables are specific parameters for quantifying these design features;

[0122] Step S1-4: Based on the results of the sensitivity analysis, a response surface model is constructed based on the weighted average of GARS, and the model includes multiple different meta-models to accurately describe the relationship between the design variables in the design space; through the response surface model, the influence of different combinations of design variables on the optimization target is characterized, providing a basis for multi-objective optimization; the design space is a set consisting of the value ranges of all design variables;

[0123] Step S1-5, performing multi-objective optimization on the average torque, efficiency and torque ripple of the five-phase permanent magnet motor based on the NSGA-II algorithm, generating a non-dominated solution set under the constraint conditions, achieving a balance of the optimization objectives, and verifying the optimization results through finite element analysis; the NSGA-II is a non-dominated genetic algorithm;

[0124] Furthermore, the central composite design method of step S1-2 specifically includes the following steps:

[0125] Step S1-2-1, determining the design variables related to the performance of the five-phase permanent magnet motor and their value ranges;

[0126] Step S1-2-2, select the design points of the center point, the axial point and the star point; the center point is located at the center of the design space and is used to estimate the experimental error to determine the benchmark motor performance; the axial point is located at the boundary of the value range of the design variable to ensure that the boundary conditions of the motor performance are fully considered; the star point is determined by setting the position of the asterisk arm and is used to estimate the nonlinear response of the design variable and evaluate the influence of the curvature of the design variable on the motor performance;

[0127] Step S1-2-3, based on the design points of the center point, axial point and star point selected in step S1-2-2, generate corresponding sample points; the sample points can cover the design space and are used for experiments and simulations to obtain motor performance data, and then analyze the relationship between design variables and motor performance, providing data support for motor optimization.

[0128] Further, the step S1-3 uses a minimum-maximum search algorithm to search for the minimum and maximum values ​​of each output parameter and its corresponding input parameter to comprehensively evaluate the response values ​​of all variables, thereby assessing the impact of the design variables on the sensitivity index of the output parameters;

[0129] The functional relationship between the global objective function and the constraint function in the response surface design space in step S1-4 is:

[0130]

[0131] Among them, y(x) represents the relationship between the global objective function and the constraint function. represents the basic function, α i represents the weighting parameter, N m Indicates the number of metamodels;

[0132] When solving quadratic approximation problems, the second-order polynomial can be expressed as:

[0133]

[0134] The response surface model solves the complex nonlinear relationship between the design variables and the average efficiency, torque, and torque ripple by a quadratic approximation method. The prediction of the i-th response of the response surface model is expressed as:

[0135]

[0136] in, represents the predicted value of the ith response, a0 represents the baseline value of the response variable when all independent variables are zero, and a i Indicates the linear influence of the ith independent variable on the response, a ii Indicates the nonlinear influence of the square term of the ith independent variable on the response, a ij It is expressed as the degree of influence of the interaction between the ith independent variable and the jth independent variable on the response, and all a are determined by least squares regression:

[0137] a=(X T X) -1 X T y

[0138] The GARS used in step S1-4 is represented as a set using the weighted average of several different meta-models:

[0139]

[0140] Among them, w i satisfy:

[0141]

[0142] Furthermore, the NSGA-II algorithm in step S1-5 includes the following steps:

[0143] Step S1-5-1, generate the parent population P0, whose size is N P ; Then, the offspring population Q0 is generated through crossover and mutation operations; the subsequent stage is denoted as the nth generation;

[0144] Step S1-5-2: Generate combined population R n =P n ∪Q n , population R n The scale is 2N P ;

[0145] Step S1-5-3: Combined population R n Decode and determine the fitness of the objective function; use the non-dominated solution sets L1, L2, and L3 to decode the population R n Sort them; the solution belonging to the best non-dominated solution set L1 is the best solution, followed by the solution in the set L2, and the solution selected from the set L3 is ranked last;

[0146] Step S1-5-4: When the total number of members of the selected set is equal to or exceeds N P When , select the initial set; calculate the crowding distance of each member in the selected set; for the next generation P n +1, select the first k1 sets according to the elitist principle; sort the kth set according to the crowding distance, select the member with the larger crowding distance and add it to P n In the +1 generation, until the population reaches N P ; The elitism principle is a commonly used strategy in genetic algorithms, which aims to retain individuals with higher fitness in each generation to ensure that excellent genes can be passed on to the next generation, thereby accelerating the convergence speed of the algorithm and improving the performance of the algorithm;

[0147] The rule for selecting the initial set is given by the following formula:

[0148] g<h,if(g rank <h rank )or(g rank <h rank and(g dist <hdist ))

[0149] Among them, g and h represent two individuals used for comparison and selection in the algorithm, g rank and h rank denote the non-dominated ranks of members g and h, respectively. dist and h dist They represent the crowding distances of members g and h respectively;

[0150] Step S1-5-5, if the limit of the number of iterations is reached, the program ends and finds the fitness and non-dominated solution of the objective function; if not, go to step S1-5-6;

[0151] Step S1-5-6: For the new population P n +1 Perform crossover and mutation operations to generate a new population Q n +1, then enter step S1-5-2, and loop steps S1-5-2 to S1-5-5 until the fitness and non-dominated solution of the objective function are output.

[0152] Furthermore, the NSGA-II algorithm parameters in step S1-5 are configured to perform 10,000 estimation evaluations, 500 samples per iteration, and 500 initial samples. The maximum number of candidates is three, and the allowed Pareto percentage is 70%. As more generations are generated, the search direction of the objective function is generated under the set flux density target. The design variables and output variables and the corresponding performance are analyzed, and the candidate points Candidate A, Candidate B, and Candidate C represent the best design generated by the optimization program. In this embodiment, efficiency has the highest priority, followed by torque pulsation and average torque, so the candidate finally selected will have higher efficiency. Finally, a 2D finite element simulation of the optimal design is performed using Maxwell software based on finite element analysis;

[0153] like Figure 2 As shown, it is a control system topology diagram based on the combination of back electromotive force control and iterative learning control. The five-phase permanent magnet motor control system includes: a data acquisition module, a signal processing module, a rectifier, an inverter, a fault diagnosis module, a back electromotive force model module, an iterative learning controller module, a current controller module, and a five-phase permanent magnet motor;

[0154] The stator of the five-phase permanent magnet motor adopts a pole-slot combination of 10 slots and 8 poles, the winding is a double-layer fractional slot concentrated winding structure, the rotor is a surface permanent magnet rotor structure and the magnet adopts a centrifugal structure, the rated speed is 15000r / min, the rated power is 7kW, the rated torque is 4.5N·m, SmCo30 is used as the permanent magnet material, and T-100 is used as the core material;

[0155] The control strategy of the five-phase permanent magnet motor control system includes the following steps:

[0156] Step S5-1, in the data acquisition module, the signal parameter values ​​of the five-phase permanent magnet motor are obtained in real time, including the current, voltage, rotor position, and torque signals of each phase, the stray magnetic flux around the motor is monitored by a fluxgate sensor, and the back electromotive force of the motor under different states is obtained by performing two-dimensional finite element operations on the motor, and finally the collected data is filtered and standardized for preprocessing;

[0157] Step S5-2, inputting the collected current signal parameters into the signal processing module, obtaining the alternating current on the ɑβ coordinate axis through coordinate transformation, and preprocessing the transformed current; the preprocessing includes standardization processing and window segmentation;

[0158] Step S5-3, input the pre-processed current into the fault diagnosis module, predict the current by using the transformer neural network and the model-based method respectively, and compare it with the actual current to determine the motor operation state; the operation state includes normal operation state, open circuit fault state and short circuit fault state;

[0159] Step S5-4: when the motor is running in a fault state, a back-electromotive force model module is established according to the back-electromotive force signal in step S5-1, and an optimal current value is calculated according to different objective functions of open circuit fault and short circuit fault to provide an initial optimal reference current;

[0160] Step S5-5, the iterative learning controller module updates the healthy phase current according to the optimal current in step S5-4 and the torque error calculated in combination with the torque signal in step S5-1;

[0161] Step S5-6, inputting the updated current into the current controller to generate a gate drive signal for the voltage source inverter so that the stator phase current can follow the optimal reference current;

[0162] Step S5-7: Perform two-dimensional finite element analysis on the motor to obtain torque pulsation, current distribution and ohmic loss data of the motor under different fault conditions, and use the data to evaluate the effect of the control scheme.

[0163] Furthermore, the optimal current in step S5-4 is obtained by first taking the electromagnetic torque without ripple and with minimum ohmic loss under fault conditions as the optimization target, and then using a closed-form equation to calculate the reference optimal current according to the back electromotive force model when the influence of the fault type is represented by a lookup table;

[0164] The optimal current calculation for the open circuit fault condition is as follows:

[0165] The objective function to minimize ohmic loss is:

[0166]

[0167] in represents the phase current vector, k represents the speed normalized back EMF vector, T* represents the desired torque, F represents the fault matrix whose elements are used to represent the constraints imposed by the stator winding connection and the fault location, and p1 and p2 are Lagrange multipliers;

[0168] Find the optimal current for open circuit fault as:

[0169]

[0170] Among them, F -1 is the left inverse of the matrix, defined as F -1 =(F T F) -1 F T , F' is given by F'=(I n -FF l -1 ) is given;

[0171] The optimal current calculation under the short-circuit fault condition is as follows:

[0172] The open-circuit fault-tolerant control is extended to short-circuit faults, and the modified objective function is defined as:

[0173]

[0174] Among them, i sc New parameters introduced to consider the effect of short-circuit current;

[0175] Find the optimal short-circuit fault current as:

[0176]

[0177] Among them, F τ -T is the right inverse of the matrix, defined as F τ -T =F(F T F) -1 ,i sc Indicates that the short-circuit related current is equal to zero in healthy operation and is equal to zero in single-phase fault. Given.

[0178] Furthermore, the specific method of updating the current in step S5-5 is to calculate based on the system equation, motor equation performance, error signal and learning algorithm, and suppress the torque pulsation with reference to the current update law;

[0179] The system equation is:

[0180] y j (m) = P(q)u j (m)+d(m)

[0181] Where m is the time index, j is the iteration index, q is the forward time shift operator qx(m)=x(m+1), y j is the output, u j is the control input, d is the repeated exogenous signal;

[0182] The motor equation is:

[0183] T m =k T i+T d

[0184] Where T m Represents the output torque, T d Indicates oscillating torque disturbance;

[0185] The error signal is:

[0186] e j (m) = y d (m)-y j (m)

[0187] Among them, y d is the expected output;

[0188] The learning algorithm is:

[0189] u j+1 (m) = Q(q)[u j (m)+L(q)e j (m+1)]

[0190] Among them, Q(q) and L(q) are defined as Q filter and learning function respectively, e j =T*-T m,j , T m,j represents the motor output torque of the jth iteration, T* represents the expected torque;

[0191] The iterative learning current update formula is:

[0192] i * ILC,j+1 =w+β(F'k(k T F'k) -1 )(T*-T m,j )

[0193] Among them, i * ILC,jrepresents the iterative learning current reference of the jth iteration, and β is a constant.

[0194] Furthermore, the current controller in step S5-6 implements the current following process as follows:

[0195] Step S5-6-1, the current controller monitors the actual value of the stator phase current in real time and compares it with the optimal current reference;

[0196] Step S5-6-2: Based on the comparison result, the current controller adjusts the gate drive signal to ensure that the stator phase current quickly and accurately tracks the optimal reference current.

[0197] Furthermore, the two-dimensional finite element analysis method in step S5-7 is as follows:

[0198] Step S5-7-1, establishing a two-dimensional geometric model according to the number of slots, inner diameter, outer diameter of the stator, the size of the rotor, the position and size parameters of the permanent magnet in the actual structure of the motor;

[0199] Step S5-7-2, defining the material properties of each part of the motor, wherein the material includes the material of the stator core, the material of the permanent magnet, and the material of the winding;

[0200] Step S5-7-3, mesh the model and discretize the geometric model into 20,000 units according to the accuracy requirements and calculation time limit.

[0201] Step S5-7-4, applying boundary conditions to the geometric model, defining the outer boundary of the motor as a magnetic insulation boundary condition, simulating the air environment around the motor, applying permanent magnet excitation to the motor and applying updated current excitation to achieve fault-tolerant control;

[0202] Step S5-7-5, using a two-dimensional finite element method to solve the partial differential equation of the motor's electromagnetic field, to obtain the distribution of the electromagnetic field quantity inside the motor, the electromagnetic field quantity including the magnetic field distribution and the magnetic flux density, thereby obtaining the torque pulsation, current distribution and ohmic loss data;

[0203] Furthermore, after the fault diagnosis module determines the system fault, the fault-tolerant control system adjusts the healthy phase current through the speed outer loop and the current inner loop to control the motor to run with the fault. The current inner loop determines the motor fault type by analyzing the current in the stationary coordinate system, and passes each phase current into the back electromotive force model to define the objective function to solve the optimal current. The optimal current is compared with the actual current through the current controller to ensure current tracking; the speed outer loop analyzes and compares the rotor speed w with the given reference speed w * , the torque reference value is obtained through the speed PI controller;

[0204] Furthermore, the data acquisition module includes sensors for multiple signals such as current, voltage, speed, torque, etc., an optical encoder for sensing the rotor position angle, and a device for filtering and standardizing each signal; the torque sensor transmits the acquired torque information to the iterative learning controller module in real time, so that the iterative learning controller can make adjustments in time according to the latest torque information. After receiving the torque information, the iterative learning controller will compare it with the reference torque and adjust it according to the current update law. When the output torque is lower than the expected torque, the iterative learning module will increase the current accordingly, otherwise, it will reduce the current accordingly to form a closed-loop control loop. The reference torque is estimated by the PI controller based on the speed error; since the vector control based on SVPWM has excellent performance in normal operation, the vector control based on SVPWM is selected as the control method for normal operation of the motor.

[0205] The fault diagnosis module in step S5-3 includes a short circuit fault diagnosis module and an open circuit fault diagnosis module;

[0206] like Figure 3 As shown, the overall process of the short circuit fault diagnosis module is as follows:

[0207] The preprocessed current signal is input into a diagnostic model based on TNN, wherein the diagnostic model includes multiple encoders, each of which includes an MHA module and an FF module; the input signal is encoded by the MHA module and the FF, and an output tensor is obtained after hierarchical processing by multiple encoders; the output tensor is used to estimate the number of short-circuited coils and the short-circuit current, and then the third harmonic component of the stray magnetic flux is used to detect and locate the short-circuit fault, and finally the number of short-circuited coils, the short-circuit current amplitude and the stray magnetic flux analysis results are comprehensively estimated to determine the short-circuit fault and its degree; the TNN is a transformer neural network module, the MHA is a multi-head attention module, and the FF is a fully connected feedforward network module;

[0208] Furthermore, the calculation process of the MHA module is as follows:

[0209] Step S6-1-1: Perform a linear transformation on the input sequence to obtain a query matrix Q = XW Q , key matrix K = XW K Sum matrix V = XW V , where X = [x1, x2, …, x Ts ] T is the input vector matrix, t=1,2,…,Ts; is the trainable weight matrix, d feat is the number of input features, Ts is the time step;

[0210] Step S6-1-2, calculate self-attention according to the formula, and its function is used to normalize the attention score;

[0211] Step S6-1-3, perform calculations through multiple parallel self-attention layers, i.e., multiple heads, and concatenate and linearly transform the results to obtain the output of the MHA module;

[0212] Furthermore, the FF module specifically includes two linear transformation layers, and residual connection and normalization processing are performed before and after each linear layer respectively;

[0213] Furthermore, the FF module calculation process is as follows:

[0214] Step S6-2-1: The input of the first linear transformation layer is X e ', the output is GELU(X e 'W FF1 +b FF1 ),in are the trainable weights of the fully connected feedforward network, is the bias of the fully connected feedforward network, GELU is the Gaussian error linear unit activation function;

[0215] Step S62-2: The input of the second linear transformation layer is the output of the first linear transformation layer, and the output is GELU (X e 'W FF1 +b FF1 )W FF2 +b FF2 ,in are the trainable weights of the fully connected feedforward network, is the bias of the fully connected feed-forward network;

[0216] like Figure 4 As shown, the overall process of the open circuit fault diagnosis module is as follows: using a model-based method to perform fault detection, specifically including using an extended Kalman filter to estimate the current of the motor and establish a discretized state space model;

[0217] Furthermore, the open circuit fault diagnosis steps are as follows:

[0218] Step S6-3-1, using an extended Kalman filter to estimate the current of the motor;

[0219] Step S6-3-2: Establish a discretized state space model, which is divided into a prediction module and an update module:

[0220] In the prediction module: is the estimated value of the state at time k-1; is the estimated value of the state at time k-1, u k-1is the input at time k-1; is the state transfer function, P k|k-1 =F k P k-1|k-1 F k T +Q k-1 is the estimated error covariance matrix of time k-1 to time k, where I is the identity matrix, is the system matrix, Where L d , L q are the direct-axis inductance and quadrature-axis inductance, respectively, PM is the permanent magnet flux, T s is the sampling time, F k is the Jacobian matrix of A, P k-1|k-1 is the estimated error covariance matrix at time k-1, Q k-1 is the covariance matrix of process noise;

[0221] In the update module: is the Kalman gain, where H k is the measurement matrix, R k is the covariance matrix of the measurement noise; is the updated state estimate at time k, where h(x) is the measurement function and y k is the measured value at time k; P k|k =(IK k H k ) k|k-1 is the estimated error covariance matrix after updating at time k; where x = [i d i q ] T ,u=[V d V q ω e ] T ,h(x (k) )=Cx (k) ,i d 、i q are the direct-axis current and quadrature-axis current, V d 、V q is the direct-axis voltage and quadrature-axis voltage, ω e is the electrical angular velocity, is the transformation matrix; calculate the residual: It is applied to the CUSUM algorithm for fault detection. The CUSUM algorithm detects the occurrence of faults by accumulating the change of residuals. At the same time, the whiteness test is used to determine whether the sensor is faulty. The whiteness test calculates the autocorrelation function of the residuals. And the normalized ACF: Where N is the number of data samples, εt is the residual at time t, is the mean of the residuals; if the residuals exceed the set threshold or the whiteness test shows an abnormality, it is determined that a fault exists; the CUSUM algorithm is a cumulative sum calculation.

[0222] After the fault diagnosis module determines the system fault, the fault-tolerant control system adjusts the healthy phase current through the speed outer loop and the current inner loop to control the motor to run with the fault. The current inner loop determines the motor fault type by analyzing the current in the stationary coordinate system, and passes each phase current into the back electromotive force model to define the objective function to solve the optimal current. The optimal current is compared with the actual current through the current controller to ensure current tracking; the speed outer loop analyzes and compares the rotor speed w with the given reference speed w * , the torque reference value is obtained through the speed PI controller.

Claims

1. A multi-objective optimization method for a five-phase permanent magnet motor, characterized in that: The multi-objective optimization method for a five-phase permanent magnet motor comprises the following steps: Step S1-1, clarifying the optimization objectives of the five-phase permanent magnet motor, including average torque, efficiency, and torque pulsation performance parameters, and determining corresponding objective functions and constraints; the objective function is used to maximize efficiency, minimize torque pulsation, and improve average torque; the constraints include the number of stator winding turns, permanent magnet thickness, air gap length, rotor pole pair number, inductance, and resistance; Step S1-2: Based on the optimization target, the central composite design method is used to generate sample points, determine the number and location of design points, and record the data of relevant parameters of each design point; Step S1-3, using GARS to perform sensitivity analysis on the generated sample points, quantify the sensitivity of the motor response to each design feature, and evaluate the sensitivity index of the design variable to the optimization target; the GARS is a genetic aggregation response surface; the design features are various factors that affect the performance of the motor, including stator outer diameter, permanent magnet thickness, air gap length, centrifugal height, and notch depth; the design variables are specific parameters for quantifying these design features; Step S1-4: Based on the results of the sensitivity analysis, a response surface model is constructed based on the weighted average of GARS, and the model includes multiple different meta-models to accurately describe the relationship between the design variables in the design space; through the response surface model, the influence of different combinations of design variables on the optimization target is characterized, providing a basis for multi-objective optimization; the design space is a set consisting of the value ranges of all design variables; Step S1-5, based on the NSGA-II algorithm, multi-objective optimization is performed on the average torque, efficiency and torque ripple of the five-phase permanent magnet motor, a non-dominated solution set is generated under the constraints, the balance of the optimization objectives is achieved, and the optimization results are verified by finite element analysis; the NSGA-II is a non-dominated genetic algorithm.

2. A five-phase permanent magnet motor multi-objective optimization method according to claim 1, characterized in that: The central composite design method of step S1-2 specifically includes the following steps: Step S1-2-1, determining the design variables related to the performance of the five-phase permanent magnet motor and their value ranges; Step S1-2-2, select the design points of the center point, the axial point and the star point; the center point is located at the center of the design space and is used to estimate the experimental error to determine the benchmark motor performance; the axial point is located at the boundary of the value range of the design variable to ensure that the boundary conditions of the motor performance are fully considered; the star point is determined by setting the position of the asterisk arm and is used to estimate the nonlinear response of the design variable and evaluate the influence of the curvature of the design variable on the motor performance; Step S1-2-3, based on the design points of the center point, axial point and star point selected in step S1-2-2, generate corresponding sample points; the sample points can cover the design space and are used for experiments and simulations to obtain motor performance data, and then analyze the relationship between design variables and motor performance, providing data support for motor optimization.

3. A five-phase permanent magnet motor multi-objective optimization method according to claim 1, characterized in that: The step S1-3 uses a minimum-maximum search algorithm to search for the minimum and maximum values ​​of each output parameter and its corresponding input parameter to comprehensively evaluate the response values ​​of all variables, thereby evaluating the impact of the design variables on the sensitivity index of the output parameters; The functional relationship between the global objective function and the constraint function in the response surface design space in step S1-4 is: Among them, y(x) represents the relationship between the global objective function and the constraint function. represents the basic function, α i represents the weighting parameter, N m Indicates the number of metamodels; The response surface model solves the complex nonlinear relationship between the design variables and the average efficiency, torque, and torque ripple by a quadratic approximation method. The prediction of the i-th response of the response surface model is expressed as: in, represents the predicted value of the ith response, a0 represents the baseline value of the response variable when all independent variables are zero, and a i Indicates the linear influence of the ith independent variable on the response, a ii Indicates the nonlinear influence of the square term of the ith independent variable on the response, a ij It is expressed as the degree of influence of the interaction between the ith independent variable and the jth independent variable on the response, and all a are determined by least squares regression: a=(X T X) -1 X T y The GARS used in step S1-4 is represented as a set using the weighted average of several different meta-models: Among them, w i satisfy:

4. A five-phase permanent magnet motor multi-objective optimization method according to claim 1, characterized in that: The NSGA-II algorithm in step S1-5 includes the following steps: Step S1-5-1, generate the parent population P0, whose size is N P ; Then, the offspring population Q0 is generated through crossover and mutation operations; the subsequent stage is denoted as the nth generation; Step S1-5-2: Generate combined population R n =P n ∪Q n , population R n The scale is 2N P ; Step S1-5-3: Combined population R n Decode and determine the fitness of the objective function; use the non-dominated solution sets L1, L2, and L3 to decode the population R n Sort them; the solution belonging to the best non-dominated solution set L1 is the best solution, followed by the solution in the set L2, and the solution selected from the set L3 is ranked last; Step S1-5-4: When the total number of members of the selected set is equal to or exceeds N P When , select the initial set; calculate the crowding distance of each member in the selected set; for the next generation P n +1, select the first k1 sets according to the elitist principle; sort the kth set according to the crowding distance, select the member with the larger crowding distance and add it to P n In the +1 generation, until the population reaches N P ; The elitism principle is a commonly used strategy in genetic algorithms, which aims to retain individuals with higher fitness in each generation to ensure that excellent genes can be passed on to the next generation, thereby accelerating the convergence speed of the algorithm and improving the performance of the algorithm; The rule for selecting the initial set is given by the following formula: g<h,if(g rank <h rank )or(g rank <h rank and(g dist <h dist )) Among them, g and h represent two individuals used for comparison and selection in the algorithm, g rank and h rank denote the non-dominated ranks of members g and h, respectively. dist and h dist They represent the crowding distances of members g and h respectively; Step S1-5-5, if the limit of the number of iterations is reached, the program ends and finds the fitness and non-dominated solution of the objective function; if not, go to step S1-5-6; Step S1-5-6: For the new population P n +1 Perform crossover and mutation operations to generate a new population Q n +1, then enter step S1-5-2, and loop steps S1-5-2 to S1-5-5 until the fitness and non-dominated solution of the objective function are output.

5. A multi-objective optimization method for a five-phase permanent magnet motor according to claim 1, and a control system for controlling the five-phase permanent magnet motor optimized by the multi-objective optimization method, characterized in that: The control system includes: a data acquisition module, a signal processing module, a rectifier, an inverter, a fault diagnosis module, a back electromotive force model module, an iterative learning controller module, a current controller module, and a five-phase permanent magnet motor; The stator of the five-phase permanent magnet motor adopts a pole-slot combination of 10 slots and 8 poles, the winding is a double-layer fractional slot concentrated winding structure, the rotor is a surface permanent magnet rotor structure and the magnet adopts a centrifugal structure; The control strategy of the control system includes the following steps: Step S5-1, in the data acquisition module, the signal parameter values ​​of the five-phase permanent magnet motor are obtained in real time, including the current, voltage, rotor position, and torque signals of each phase, the stray magnetic flux around the motor is monitored by a fluxgate sensor, and the back electromotive force of the motor under different states is obtained by performing two-dimensional finite element operations on the motor, and finally the collected data is filtered and standardized for preprocessing; Step S5-2, inputting the collected current signal parameters into the signal processing module, obtaining the alternating current on the αβ coordinate axis through coordinate transformation, and preprocessing the transformed current; the preprocessing includes standardization processing and window segmentation; Step S5-3, input the pre-processed current into the fault diagnosis module, predict the current by using the transformer neural network and the model-based method respectively, and compare it with the actual current to determine the motor operation state; the operation state includes normal operation state, open circuit fault state and short circuit fault state; Step S5-4: when the motor is running in a fault state, a back-electromotive force model module is established according to the back-electromotive force signal in step S5-1, and an optimal current value is calculated according to different objective functions of open circuit fault and short circuit fault to provide an initial optimal reference current; Step S5-5, the iterative learning controller module updates the healthy phase current according to the optimal current in step S5-4 and the torque error calculated in combination with the torque signal in step S5-1; Step S5-6, inputting the updated current into the current controller to generate a gate drive signal for the voltage source inverter so that the stator phase current can follow the optimal reference current; Step S5-7: Perform two-dimensional finite element analysis on the motor to obtain torque pulsation, current distribution and ohmic loss data of the motor under different fault conditions, and use the data to evaluate the effect of the control scheme.

6. A five-phase permanent magnet motor control system according to claim 5, characterized in that: The fault diagnosis module in step S5-3 includes a short circuit fault diagnosis module and an open circuit fault diagnosis module; The overall process of the short-circuit fault diagnosis module is as follows: the preprocessed current signal is input into the TNN-based diagnosis model, the diagnosis model includes multiple encoders, each encoder includes an MHA module and an FF module; the input signal is encoded by the MHA module and the FF, and after hierarchical processing by multiple encoders, an output tensor is obtained; The output tensor is used to estimate the number of short-circuited coils and the short-circuit current, and the third harmonic component of the stray magnetic flux is used to detect and locate the short-circuit fault. Finally, the number of short-circuited coils, the short-circuit current amplitude and the stray magnetic flux analysis results are comprehensively estimated to determine the short-circuit fault and its degree; the TNN is a transformer neural network module, the MHA is a multi-head attention module, and the FF is a fully connected feedforward network module; The MHA module calculation process is as follows: Step S6-1-1: Perform a linear transformation on the input sequence to obtain a query matrix Q = XW Q , key matrix K = XW K Sum matrix V = XW V , where X = [x1, x2, …, x Ts ] T is the input vector matrix, is the trainable weight matrix, d feat is the number of input features, Ts is the time step; Step S6-1-2, calculate self-attention according to the formula, and its function is used to normalize the attention score; Step S6-1-3, perform calculations through multiple parallel self-attention layers, i.e., multiple heads, and concatenate and linearly transform the results to obtain the output of the MHA module; The FF module specifically includes two linear transformation layers, and residual connection and normalization processing are performed before and after each linear layer respectively; The FF module calculation process is as follows: Step S6-2-1: The input of the first linear transformation layer is X′ e , the output is GELU(X′ e W FF1 +b FF1 ),in are the trainable weights of the fully connected feedforward network, is the bias of the fully connected feedforward network, GELU is the Gaussian error linear unit activation function; Step S6-2-2: The input of the second linear transformation layer is the output of the first linear transformation layer, and the output is GELU(X′ e W FF1 +b FF1 )W FF2 +b FF2 ,in are the trainable weights of the fully connected feedforward network, is the bias of the fully connected feed-forward network; The overall process of the open circuit fault diagnosis module is as follows: using a model-based approach to perform fault detection, specifically including using an extended Kalman filter to estimate the current of the motor and establish a discretized state space model; The open circuit fault diagnosis steps are as follows: Step S6-3-1, using an extended Kalman filter to estimate the current of the motor; Step S6-3-2: Establish a discretized state space model, which is divided into a prediction module and an update module: In the prediction module: is the estimated value of the state at time k at time k-1; is the estimated value of the state at time k-1, u k-1 is the input at time k-1; is the state transfer function, P k|k-1 =F k P k-1|k-1 F k T +Q k-1 is the estimated error covariance matrix of time k-1 to time k, where I is the identity matrix, is the system matrix, Where L d , L q are the direct-axis inductance and quadrature-axis inductance, respectively, PM is the permanent magnet flux, T s is the sampling time, F k is the Jacobian matrix of A, P k-1|k-1 is the estimated error covariance matrix at time k-1, Q k-1 is the covariance matrix of process noise; In the update module: is the Kalman gain, where H k is the measurement matrix, R k is the covariance matrix of the measurement noise; is the updated state estimate at time k, where h(x) is the measurement function and y k is the measured value at time k; P k|k =(IK k H k ) k|k-1 is the estimated error covariance matrix after updating at time k; where x = [i d i q ] T ,u=[V d V q ω e ] T ,h(x (k) )=Cx (k) ,i d 、i q are the direct-axis current and quadrature-axis current, V d 、V q is the direct-axis voltage and quadrature-axis voltage, ω e is the electrical angular velocity, is the transformation matrix; calculate the residual: It is applied to the CUSUM algorithm for fault detection. The CUSUM algorithm detects the occurrence of faults by accumulating the change of residuals. At the same time, the whiteness test is used to determine whether the sensor is faulty. The whiteness test calculates the autocorrelation function of the residuals. And the normalized ACF: Where N is the number of data samples, ε t is the residual at time t, is the mean of the residuals; if the residuals exceed the set threshold or the whiteness test shows an abnormality, it is determined that a fault exists; the CUSUM algorithm is a cumulative sum algorithm.

7. A five-phase permanent magnet motor control system according to claim 5, characterized in that: The optimal current in step S5-4 is obtained by first taking the electromagnetic torque without ripple and with minimum ohmic loss under fault conditions as the optimization target, and then using a closed-form equation to calculate the reference optimal current according to the back electromotive force model when the fault type influence is represented by a lookup table; The optimal current calculation for the open circuit fault condition is as follows: The objective function to minimize ohmic loss is: in represents the phase current vector, k represents the speed normalized back EMF vector, T* represents the desired torque, F represents the fault matrix whose elements are used to represent the constraints imposed by the stator winding connection and the fault location, and p1 and p2 are Lagrange multipliers; Find the optimal current for open circuit fault as: Among them, F -1 is the left inverse of the matrix, defined as F -1 =(F T F) -1 F T , F' is given by F'=(I n -FF l -1 ) is given; The optimal current calculation under the short-circuit fault condition is as follows: The open-circuit fault-tolerant control is extended to short-circuit faults, and the modified objective function is defined as: Among them, i sc New parameters introduced to consider the effect of short-circuit current; Find the optimal short-circuit fault current as: Among them, F τ -T is the right inverse of the matrix, defined as F τ -T =F(F T F) -1 ,i sc Indicates that the short-circuit related current is equal to zero in healthy operation and equal to zero in single-phase fault. Given.

8. The control system of a five-phase permanent magnet motor according to claim 5, characterized in that: The specific method of updating the current in step S5-5 is to calculate based on the system equation, motor equation performance, error signal and learning algorithm, and suppress the torque pulsation with reference to the current update law; The system equation is: y j (m)=P(q)u j (m)+d(m) Where m is the time index, j is the iteration index, q is the forward time shift operator qx(m)=x(m+1), y j is the output, u j is the control input, d is the repeated exogenous signal; The motor equation is: T m =k T i+T d Where T m Represents the output torque, T d Indicates oscillating torque disturbance; The error signal is: e j (m)=y d (m)-y j (m) Among them, y d is the expected output; The learning algorithm is: u j+1 (m)=Q(q)[u j (m)+L(q)e j (m+1)] Among them, Q(q) and L(q) are defined as Q filter and learning function respectively, e j =T*-T m,j , T m,j represents the motor output torque of the jth iteration, T* represents the expected torque; The iterative learning current update formula is: i * ILC,j+1 =w+β(F'k(k T F'k) -1 )(T*-T m,j )j Among them, i * ILC,j represents the iterative learning current reference of the jth iteration, and β is a constant.

9. A control system for a five-phase permanent magnet motor according to claim 5, characterized in that: The steps for the current controller in step S5-6 to implement current following are as follows: Step S5-6-1, the current controller monitors the actual value of the stator phase current in real time and compares it with the optimal current reference; Step S5-6-2: Based on the comparison result, the current controller adjusts the gate drive signal to ensure that the stator phase current quickly and accurately tracks the optimal reference current.

10. A control system for a five-phase permanent magnet motor according to claim 5, characterized in that: The two-dimensional finite element analysis method steps in step S5-7 are: Step S5-7-1, establishing a two-dimensional geometric model according to the number of slots, inner diameter, outer diameter of the stator, the size of the rotor, the position and size parameters of the permanent magnet in the actual structure of the motor; Step S5-7-2, defining the material properties of each part of the motor, wherein the material includes the material of the stator core, the material of the permanent magnet, and the material of the winding; Step S5-7-3, meshing the model, discretizing the geometric model into 20,000 units according to accuracy requirements and calculation time limits; Step S5-7-4, applying boundary conditions to the geometric model, defining the outer boundary of the motor as a magnetic insulation boundary condition, simulating the air environment around the motor, applying permanent magnet excitation to the motor and applying updated current excitation to achieve fault-tolerant control; Step S5-7-5, using the two-dimensional finite element method to solve the partial differential equation of the motor's electromagnetic field, to obtain the distribution of the electromagnetic field quantity inside the motor, the electromagnetic field quantity including the magnetic field distribution and the magnetic flux density, thereby obtaining the torque pulsation, current distribution and ohmic loss data.

Images