Adaptive integral terminal sliding mode permanent magnet synchronous motor control method based on RBF neural network

Abstract

The invention provides a self-adaptive integral terminal sliding mode permanent magnet synchronous motor control method based on an RBF neural network, and the method comprises the steps: firstly constructing a permanent magnet synchronous motor mathematical model and a self-adaptive integral terminal sliding mode rotating speed loop control model, and carrying out the automatic calibration of controller parameters under different working conditions through a gradient descent method on this basis; then a radial basis function (RBF) neural network model is constructed and trained to enable controller parameters to be automatically adjusted according to actual operation conditions, and finally the neural network is combined with a self-adaptive integral terminal sliding mode, so that rotating speed ring controller parameters can be automatically adjusted according to the target rotating speed and the load torque, the dynamic response performance is obviously improved, and the method is suitable for large-scale popularization and application. And meanwhile, overshoot is effectively reduced, and the steady-state performance is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of permanent magnet synchronous motor control, and in particular relates to an adaptive integral terminal sliding mode permanent magnet synchronous motor control method based on RBF neural network. Background Art

[0002] In the control of permanent magnet synchronous motor (PMSM), the traditional method is to use a double closed-loop double PI vector control system composed of multiple speed loops and current loops. However, since PMSM is a nonlinear time-varying system, the traditional PI control is insufficient in dynamic response and steady-state performance. In addition, when facing different load torques and target speeds, it is often necessary to manually calibrate the controller parameters to adapt to various working conditions. Some recent PMSM control technologies have introduced sliding mode variable structure control, which has the advantages of simple structure, fast response, and strong robustness. It is more suitable for nonlinear control than traditional methods. However, the performance of sliding mode control often depends on reasonable controller parameter settings. If the parameters are not set properly, it may cause system jitter and reduce control performance. Due to the different requirements for controller parameters under different working conditions, the parameter calibration process is often complicated and time-consuming. Therefore, there is an urgent need in this field for a nonlinear control method with parameter adaptive adjustment capability, which can significantly improve the PMSM control performance under multiple working conditions while simplifying the debugging process. Summary of the invention

[0003] In view of this, in order to solve the technical problems existing in the art, the present invention provides an adaptive integral terminal sliding mode permanent magnet synchronous motor control method based on RBF neural network, which specifically includes the following steps:

[0004] Step 1, construct an equivalent mathematical model of the permanent magnet synchronous motor in a synchronous rotating coordinate system, including voltage equations, electromagnetic torque equations and motor dynamics; adopt deadbeat predictive control (DPCC) for the current loop of the permanent magnet synchronous motor, and use the current at time k in combination with the voltage equation to predict the control voltage at time k+1; use an adaptive integral terminal sliding mode surface to control the speed loop; combine the equations to obtain an adaptive integral terminal sliding mode speed loop control model;

[0005] Step 2: Based on the difference between the actual speed and the ideal speed, a loss function and its related controller parameters to be optimized are defined, and the controller parameters under various working conditions are automatically calibrated using the gradient descent method;

[0006] Step 3, establish a radial basis function (RBF) neural network model, and train the neural network using the corresponding relationship between different operating conditions and their corresponding calibrated controller parameters, so that the trained neural network can automatically adjust to obtain the optimal controller parameters according to the operating conditions;

[0007] Step 4: Combine the trained RBF neural network with the adaptive integral terminal sliding mode speed loop control model. The RBF neural network uses the load torque and target speed under different working conditions to calculate the optimal controller parameters, and obtains the control voltage reference value through the speed loop and current loop in turn.

[0008] Furthermore, in step 1, an equivalent mathematical model is established specifically for the surface-mounted permanent magnet synchronous motor, including: establishing the following dq axis voltage equations:

[0009]

[0010] In the formula, u d and u q are the d-axis and q-axis stator voltage components respectively; i d and i q are the d-axis and q-axis stator current components respectively; R s is the stator resistance; L s is the stator inductance; ψ f is the rotor permanent magnet flux, ω e is the rotor electrical angular velocity, t is the time;

[0011] The following permanent magnet synchronous motor torque equation is established:

[0012]

[0013] Where, T e is the electromagnetic torque; p is the number of pole pairs;

[0014] And establish the following permanent magnet synchronous motor dynamic equation:

[0015]

[0016] Where, J is the rotor moment of inertia; ω m is the rotor mechanical angular velocity; T l is the load torque; B is the viscous friction coefficient.

[0017] Furthermore, in step 1, deadbeat predictive control is specifically adopted and the following prediction model is established to predict the control voltage at time k+1:

[0018]

[0019] where u d (k+1),u q (k+1) are the predicted voltage vectors of d and q axes at time k+1 respectively; are the d-axis and q-axis command current vectors at time k respectively; T s is the sampling time; i d(k), i q (k) is the sampling vector of d and q axes at time k; ω e (k) is the motor sampled electrical angular velocity at time k;

[0020] The adaptive integral terminal sliding surface of the designed speed loop is as follows:

[0021] s=e+s I

[0022] Where s is the sliding surface, e is the mechanical angular velocity error, and s I is the sliding mode integral term;

[0023] Adaptive terminal integral sliding mode term s I for:

[0024] s I =max[-I max ,min(I,I max )]

[0025] In the formula, I max is the adaptive integral term limit; I is the integral terminal sliding mode term;

[0026] The integral terminal sliding mode term I is:

[0027]

[0028] In the formula, η1 and η2 are adaptive adjustment parameters respectively; tanh is the hyperbolic cosine function;

[0029] Adaptive integral term limit I max for:

[0030] I max =i qmax (1-e -0.5e )

[0031] In the formula, i qmax is the maximum value of the motor q-axis current;

[0032] The following adaptive reaching law is adopted:

[0033]

[0034] In the formula, η3 is also an adaptive adjustment parameter;

[0035] Taking the derivative of the sliding surface, we get:

[0036]

[0037] The superscript · indicates the derivative of the corresponding parameter;

[0038] Since the speed setting is a constant, Will and Lianlide:

[0039]

[0040] Finally, the following adaptive integral terminal sliding mode speed loop control model is obtained by combining the permanent magnet synchronous motor dynamics equation and the electromagnetic torque equation:

[0041]

[0042] Among them, η1, η2, and η3 are the parameters of the adaptive integral terminal synovial speed loop controller, and their values ​​are all positive numbers.

[0043] Furthermore, the process of automatically calibrating the controller parameters under various working conditions in step 2 includes:

[0044] Consider the load torque T corresponding to different working conditions l ref and target speed n ref And define the following loss function J:

[0045]

[0046] Where N is the total number of sampling points, which is determined by the sampling frequency and sampling time of the controller; n is the sampling speed of the controller; n ideal The ideal speed curve is determined according to the current working conditions, which is specifically defined as:

[0047]

[0048] Where, T emax is the maximum value of the motor electromagnetic torque; t r The time required for the motor to reach the target speed.

[0049] Set the initial value of the controller parameter to be optimized to Then, the iteration begins, the startup experiment is carried out, and the initial loss function J is calculated. 0 ;

[0050] Specify the learning rate δ, and let the parameters to be optimized increase automatically Conduct a startup experiment and calculate the loss function J 1

[0051] Increment the optimized parameters Conduct a startup experiment and calculate the loss function J 2 ;

[0052] Increment the optimized parameters Conduct a startup experiment and calculate the loss function J3 ;

[0053] Calculate the gradient of the objective function with respect to the controller parameters

[0054]

[0055] In the formula, J i =[J 1 J 2 J 3 ],

[0056] The initial parameter values ​​are updated as follows:

[0057]

[0058] Iterate until the result converges or reaches the set maximum number of iterations, and then stop to complete the automatic calibration of the controller parameters to be optimized.

[0059] Furthermore, the construction and training process of the RBF neural network in step 3 includes the following steps:

[0060] a. Define input data Output data Y = [η1η2η3], normalize the input and output data, and construct training and test sets;

[0061] b. Build the RBF neural network model, set the network parameters including the number of training times, learning rate, minimum error of training target, and set the initial values ​​of weights and biases;

[0062] c. Use the training set to train the RBF neural network model;

[0063] d. Use the test set to test the RBF neural network model;

[0064] e. Determine whether the training results meet expectations by denormalizing the predicted results and comparing them with the true values;

[0065] f. If the prediction result meets the expected error, the training is terminated, the weights and biases of the neural network are output, and the trained optimal controller parameter prediction model is obtained; otherwise, the weights and biases are updated, and steps be are repeated until the training is completed.

[0066] Furthermore, the process of calculating the voltage reference value in step 4 includes:

[0067] a. Input the load torque and target speed into the trained RBF neural network and calculate the corresponding optimal controller parameters η1, η2, η3;

[0068] b. Input the optimal controller parameters η1, η2, η3 into the adaptive integral terminal speed loop sliding mode controller model to calculate the current reference value i q ref ;

[0069] c. Set the current reference value i s ref Input the current loop based on deadbeat predictive control and calculate the voltage reference value u d ref and u q ref .

[0070] The adaptive integral terminal sliding mode permanent magnet synchronous motor control method based on RBF neural network provided by the present invention first constructs a mathematical model of the permanent magnet synchronous motor and an adaptive integral terminal sliding mode speed loop control model, and on this basis uses the gradient descent method to automatically calibrate the controller parameters under different working conditions, and then constructs and trains a radial basis function (RBF) neural network model so that the controller parameters can be automatically adjusted according to the actual operating conditions. Finally, the neural network is combined with the adaptive integral terminal sliding mode, which can realize automatic adjustment of the speed loop controller parameters according to the target speed and load torque, and the dynamic response performance is significantly improved, while effectively reducing overshoot and improving steady-state performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A control block diagram of the method provided by the present invention;

[0072] Figure 2 Flowchart for calibrating controller parameters for the gradient descent method;

[0073] Figure 3 This is the training and testing flow chart for RBF neural network;

[0074] Figure 4 This is a comparison diagram of the effects of the present invention and the traditional PI control method;

[0075] Figure 5 This is a comparison chart of the effects of the present invention and another example using the traditional PI control method. DETAILED DESCRIPTION

[0076] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] like Figure 1As shown, the control system constructed based on the present invention mainly includes a DPCC current loop controller, a parameter-adjustable adaptive integral terminal sliding mode speed loop controller, an RBF neural network module, a Clarke / Park transformation module, an SVPWM modulation module, a three-phase voltage inverter and a permanent magnet synchronous motor model. The current sensor detects the three-phase current of the motor in real time, and obtains the real-time d and q axis currents of the motor through coordinate transformation; the speed sensor detects the real-time speed signal and rotor position signal of the motor. The target speed and load torque are calculated by the RBF neural network module to obtain the speed loop controller parameters, and the control parameters and the speed error are input into the speed loop controller to obtain the q axis reference current. Adopting i d =0 control strategy, the d, q axis reference current and actual current are input into the current loop controller to obtain the d, q axis reference voltage, and the α, β axis voltage is obtained through coordinate transformation. The space vector modulation is performed through the SVPWM module, and the modulation signal is output to the inverter module, which finally drives the permanent magnet synchronous motor to operate.

[0078] The corresponding control method specifically includes the following steps:

[0079] Step 1, construct an equivalent mathematical model of the permanent magnet synchronous motor in a synchronous rotating coordinate system, including voltage equations, electromagnetic torque equations and motor dynamics; adopt deadbeat predictive control (DPCC) for the current loop of the permanent magnet synchronous motor, and use the current at time k in combination with the voltage equation to predict the control voltage at time k+1; use an adaptive integral terminal sliding mode surface to control the speed loop; combine the equations to obtain an adaptive integral terminal sliding mode speed loop control model;

[0080] Step 2: Based on the difference between the actual speed and the ideal speed, a loss function and its related controller parameters to be optimized are defined, and the controller parameters under various working conditions are automatically calibrated using the gradient descent method;

[0081] Step 3, establish a radial basis function (RBF) neural network model, and train the neural network using the corresponding relationship between different operating conditions and their corresponding calibrated controller parameters, so that the trained neural network can automatically adjust to obtain the optimal controller parameters according to the operating conditions;

[0082] Step 4: Combine the trained RBF neural network with the adaptive integral terminal sliding mode speed loop control model. The RBF neural network uses the load torque and target speed under different working conditions to calculate the optimal controller parameters, and obtains the control voltage reference value through the speed loop and current loop in turn.

[0083] In a preferred embodiment of the present invention, in step 1, an equivalent mathematical model is established specifically for the surface-mounted permanent magnet synchronous motor, including: establishing the following dq axis voltage equations:

[0084]

[0085] In the formula, u d and u q are the d-axis and q-axis stator voltage components respectively; i d and i q are the d-axis and q-axis stator current components respectively; R s is the stator resistance; L s is the stator inductance; ψ f is the rotor permanent magnet flux, ω e is the rotor electrical angular velocity, t is the time;

[0086] The following permanent magnet synchronous motor torque equation is established:

[0087]

[0088] Where, T e is the electromagnetic torque; p is the number of pole pairs;

[0089] And establish the following permanent magnet synchronous motor dynamic equation:

[0090]

[0091] Where, J is the rotor moment of inertia; ω m is the rotor mechanical angular velocity; T l is the load torque; B is the viscous friction coefficient.

[0092] In a preferred embodiment of the present invention, deadbeat predictive control is specifically adopted in step 1 and the following prediction model is established to predict the control voltage at time k+1:

[0093]

[0094] where u d (k+1),u q (k+1) are the predicted voltage vectors of d and q axes at time k+1 respectively; are the d-axis and q-axis command current vectors at time k respectively; T s is the sampling time; i d (k), i q (k) is the sampling vector of d and q axes at time k; ω e (k) is the motor sampled electrical angular velocity at time k;

[0095] Designing the adaptive integral terminal sliding surface of the speed loop can improve the system robustness, improve the system control performance, and reduce the chattering phenomenon, but there is a problem of difficulty in parameter adjustment. The sliding surface form is as follows:

[0096] s=e+s I

[0097] Where s is the sliding surface, e is the mechanical angular velocity error, and s I is the sliding mode integral term;

[0098] Adaptive terminal integral sliding mode term s I for:

[0099] s I =max[-I max ,min(I,I max )]

[0100] In the formula, I max is the adaptive integral term limit; I is the integral terminal sliding mode term;

[0101] The integral terminal sliding mode term I is:

[0102]

[0103] In the formula, η1 and η2 are adaptive adjustment parameters respectively; tanh is the hyperbolic cosine function;

[0104] When the speed error is large, the limit of the adaptive integral term is not affected. When the speed error is close to 0, the limit is also close to 0, thereby reducing the chattering phenomenon of the sliding surface. The adaptive integral term limit I max for:

[0105] I max =i qmax (1-e -0.5e )

[0106] In the formula, i qmax is the maximum value of the motor q-axis current;

[0107] The design idea of ​​the adaptive reaching law is similar to the above. When the speed error is large, its limit value is not affected. When the speed error is close to 0, its limit value is also close to 0, thereby reducing the chattering phenomenon of the sliding surface. Therefore, the following adaptive reaching law is adopted:

[0108]

[0109] In the formula, η3 is also an adaptive adjustment parameter;

[0110] Taking the derivative of the sliding surface, we get:

[0111]

[0112] The superscript · indicates the derivative of the corresponding parameter;

[0113] Since the speed setting is a constant, Will and Lianlide:

[0114]

[0115] Finally, the following adaptive integral terminal sliding mode speed loop control model is obtained by combining the permanent magnet synchronous motor dynamics equation and the electromagnetic torque equation:

[0116]

[0117] Among them, η1, η2, and η3 are the parameters of the adaptive integral terminal synovial speed loop controller, and their values ​​are all positive numbers.

[0118] Select the Lyapunov function:

[0119]

[0120] The stability of the system is analyzed by calculating the derivative of the Lyapunov function. If the derivative is less than 0, the system is stable.

[0121]

[0122] From the above formula, we can see that the derivative of the Lyapunov function is always less than 0, and the system is stable. The disadvantage of the adaptive integral terminal sliding mode is that the parameter adjustment is complicated, and each working condition corresponds to different controller parameters. In order to simplify the difficulty of parameter adjustment, the gradient descent method is used to automatically calibrate the controller parameters of each working condition.

[0123] Multi-condition load torque T l ref and target speed n ref The motor load torque range is 0-10Nm, and the target speed range is 0-1000rpm. Figure 2 As shown in Figure 1, it is a flow chart of calibrating controller parameters using the gradient descent method.

[0124] In a preferred embodiment of the present invention, the process of automatically calibrating the controller parameters under various working conditions in step 2 includes:

[0125] Consider the load torque T corresponding to different working conditions l ref and target speed n ref And define the following loss function J:

[0126]

[0127] Where N is the total number of sampling points, which is determined by the sampling frequency and sampling time of the controller; n is the sampling speed of the controller; n ideal The ideal speed curve is determined according to the current working conditions, which is specifically defined as:

[0128]

[0129] Where, T emax is the maximum value of the motor electromagnetic torque; t r The time required for the motor to reach the target speed.

[0130] Set the initial value of the controller parameter to be optimized to Then, the iteration begins, the startup experiment is carried out, and the initial loss function J is calculated. 0 ;

[0131] Specify the learning rate δ, and let the parameters to be optimized increase automatically Conduct a startup experiment and calculate the loss function J 1

[0132] Increment the optimized parameters Conduct a startup experiment and calculate the loss function J 2 ;

[0133] Increment the optimized parameters Conduct a startup experiment and calculate the loss function J 3 ;

[0134] Calculate the gradient of the objective function with respect to the controller parameters

[0135]

[0136] In the formula, J i =[J 1 J 2 J 3 ],

[0137] The initial parameter values ​​are updated as follows:

[0138]

[0139] Iterate until the result converges or reaches the set maximum number of iterations, and then stop to complete the automatic calibration of the controller parameters to be optimized.

[0140] In a preferred embodiment of the present invention, step 3 includes the following steps in sequence:

[0141] a. Define input data Output data Y = [η1η2η3], normalize the input and output data, and construct training and test sets;

[0142] b. Build the RBF neural network model, set the network parameters including the number of training times, learning rate, minimum error of training target, and set the initial values ​​of weights and biases;

[0143] c. Use the training set to train the RBF neural network model;

[0144] d. Use the test set to test the RBF neural network model;

[0145] e. Determine whether the training results meet expectations by denormalizing the predicted results and comparing them with the true values;

[0146] f. If the prediction result meets the expected error, the training is terminated, the weights and biases of the neural network are output, and the trained optimal controller parameter prediction model is obtained; otherwise, the weights and biases are updated, and steps be are repeated until the training is completed.

[0147] In a preferred embodiment of the present invention, the process of calculating the voltage reference value in step 4 includes:

[0148] a. Input the load torque and target speed into the trained RBF neural network and calculate the corresponding optimal controller parameters η1, η2, η3;

[0149] b. Input the optimal controller parameters η1, η2, η3 into the adaptive integral terminal speed loop sliding mode controller model to calculate the current reference value i q ref ;

[0150] c. Set the current reference value i s ref Input the current loop based on deadbeat predictive control and calculate the voltage reference value u d ref and u q ref .

[0151] Figure 4 The experimental diagrams of speed and d and q axis current of the PI speed loop DPCC current loop (left) and the RBF neural network-based adaptive integral terminal sliding mode speed loop DPCC current loop (right) under 300RPM / 9Nm operating conditions are shown.

[0152] Figure 5 The experimental diagrams of the 500RPM / 7Nm working condition PI speed loop DPCC current loop (left) and the RBF neural network adaptive integral terminal sliding mode speed loop DPCC current loop (right) speed and d and q axis current are shown. The speed loop control period is 4e-4s, and the current loop control period is 5e-5s. The experimental results show that the traditional method has problems such as overshoot and slow response speed, and different working conditions are affected differently. The improved control strategy is not limited by the operating conditions and can automatically adjust to the optimal control parameters, which enhances the robustness of the controller and ensures the dynamic response performance and steady-state performance of the motor under various operating conditions.

[0153] It should be understood that the size of the serial numbers of the steps in the embodiment of the present invention does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0154] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An adaptive integral terminal sliding mode permanent magnet synchronous motor control method based on RBF neural network, characterized in that: The specific steps include: Step 1, construct an equivalent mathematical model of the permanent magnet synchronous motor in a synchronous rotating coordinate system, including voltage equations, electromagnetic torque equations and motor dynamics; adopt deadbeat predictive control for the current loop of the permanent magnet synchronous motor, and use the current at time k in combination with the voltage equation to predict the control voltage at time k+1; The speed loop is controlled by an adaptive integral terminal sliding mode surface; the adaptive integral terminal sliding mode speed loop control model is obtained by combining the equations; Step 2: Based on the difference between the actual speed and the ideal speed, a loss function and its related controller parameters to be optimized are defined, and the controller parameters under various working conditions are automatically calibrated using the gradient descent method; Step 3: Establish a radial basis function (RBF) neural network model, and train the neural network using the corresponding relationship between different operating conditions and their corresponding calibrated controller parameters, so that the trained neural network can automatically adjust to obtain the optimal controller parameters according to the operating conditions; Step 4: Combine the trained RBF neural network with the adaptive integral terminal sliding mode speed loop control model. The RBF neural network uses the load torque and target speed under different working conditions to calculate the optimal controller parameters, and obtains the control voltage reference value through the speed loop and current loop in turn.

2. The method according to claim 1, characterized in that: In step 1, an equivalent mathematical model is established specifically for the surface-mounted permanent magnet synchronous motor, including: establishing the following dq axis voltage equations: In the formula, u d and u q are the d-axis and q-axis stator voltage components respectively; i d and i q are the d-axis and q-axis stator current components respectively; R s is the stator resistance; L s is the stator inductance; ψ f is the rotor permanent magnet flux, ω e is the rotor electrical angular velocity, t is the time; The following permanent magnet synchronous motor torque equation is established: Where, T e is the electromagnetic torque; p is the number of pole pairs; And establish the following permanent magnet synchronous motor dynamic equation: Where, J is the rotor moment of inertia; ω m is the rotor mechanical angular velocity; T l is the load torque; B is the viscous friction coefficient.

3. The method according to claim 2, characterized in that: In step 1, deadbeat predictive control is specifically adopted and the following prediction model is established to predict the control voltage at time k+1: where u d (k+1),u q (k+1) are the predicted voltage vectors of d and q axes at time k+1 respectively; are the d-axis and q-axis command current vectors at time k respectively; T s is the sampling time; i d (k), i q (k) is the sampling vector of d and q axes at time k; ω e (k) is the motor sampled electrical angular velocity at time k; The adaptive integral terminal sliding surface of the designed speed loop is as follows: s=e+s I Where s is the sliding surface, e is the mechanical angular velocity error, and s I is the sliding mode integral term; Adaptive terminal integral sliding mode term s I for: s I =max[-I max ,min(I,I max )] In the formula, I max is the adaptive integral term limit; I is the integral terminal sliding mode term; The integral terminal sliding mode term I is: In the formula, η1 and η2 are adaptive adjustment parameters respectively; tanh is the hyperbolic cosine function; Adaptive integral term limit I max for: I max =i qmax (1-e -0.5e ) In the formula, i qmax is the maximum value of the motor q-axis current; The following adaptive reaching law is adopted: In the formula, η3 is also an adaptive adjustment parameter; Taking the derivative of the sliding surface, we get: The superscript · indicates the derivative of the corresponding parameter; Since the speed setting is a constant, Will and Lianlide: Finally, the following adaptive integral terminal sliding mode speed loop control model is obtained by combining the permanent magnet synchronous motor dynamics equation and the electromagnetic torque equation: Among them, η1, η2, and η3 are the parameters of the adaptive integral terminal synovial speed loop controller, and their values ​​are all positive numbers.

4. The method according to claim 3, characterized in that: Step 2: The process of automatically calibrating the controller parameters under various working conditions includes: Consider the load torque T corresponding to different working conditions l ref and target speed n ref And define the following loss function J: Where N is the total number of sampling points, which is determined by the sampling frequency and sampling time of the controller; n is the sampling speed of the controller; n ideal The ideal speed curve is determined according to the current working conditions, which is specifically defined as: Where, T emax is the maximum value of the motor electromagnetic torque; t r The time required for the motor to reach the target speed. Set the initial value of the controller parameter to be optimized to Then, the iteration begins, the startup experiment is carried out, and the initial loss function J is calculated. 0 ; Specify the learning rate δ, and let the parameters to be optimized increase automatically Conduct a startup experiment and calculate the loss function J 1 Increment the optimized parameters Conduct a startup experiment and calculate the loss function J 2 ; Increment the optimized parameters Conduct a startup experiment and calculate the loss function J 3 ; Calculate the gradient of the objective function with respect to the controller parameters In the formula, J i =[J 1 J 2 J 3 ], The initial parameter values ​​are updated as follows: Iterate until the result converges or reaches the set maximum number of iterations, and then stop to complete the automatic calibration of the controller parameters to be optimized.

5. The method according to claim 4, characterized in that: Step 3: The construction and training process of the RBF neural network includes: a. Define input data Output data Y = [η1η2η3], normalize the input and output data, and construct training and test sets; b. Build the RBF neural network model, set the network parameters including the number of training times, learning rate, minimum error of the training target, and set the initial values ​​of weights and biases; c. Use the training set to train the RBF neural network model; d. Use the test set to test the RBF neural network model; e. Determine whether the training results meet expectations by denormalizing the predicted results and comparing them with the true values; f. If the prediction result meets the expected error, the training is terminated, the weights and biases of the neural network are output, and the trained optimal controller parameter prediction model is obtained; otherwise, the weights and biases are updated, and steps be are repeated until the training is completed.

6. The method according to claim 5, characterized in that: The process of calculating the voltage reference value in step 4 includes: a. Input the load torque and target speed into the trained RBF neural network and calculate the corresponding optimal controller parameters η1, η2, η3; b. Input the optimal controller parameters η1, η2, η3 into the adaptive integral terminal speed loop sliding mode controller model to calculate the current reference value i q ref ; c. Set the current reference value i s ref Input the current loop based on deadbeat predictive control and calculate the voltage reference value u d ref and u q ref .

Images