CN116455278A - Self-tuning method for pid parameters of permanent magnet synchronous motor control system based on neural network self-adaptive control - Google Patents

Abstract

The invention provides a method for self-tuning of PID parameters of a permanent magnet synchronous motor control system based on neural network adaptive control, the method comprising the following steps: Step 1: designing the neural network structure of the adaptive controller of the speed loop; Step 2 : Using the gradient descent algorithm to adjust the neural network weights of the neural network adaptive controller online; Step 3: Feedback control of the motor speed. The invention can realize the self-tuning of the pid parameters of the permanent magnet synchronous motor speed loop control system, improve the dynamic performance of the motor operation, improve the self-adaptive ability of the motor control algorithm, and get rid of the dependence on the system model parameters.

Description

A permanent magnet synchronous motor control system pid based on neural network adaptive control Parameter self-tuning method

technical field

The invention belongs to the field of permanent magnet synchronous motor control, in particular to a method for self-tuning of PID parameters of a permanent magnet synchronous motor control system based on neural network adaptive control.

Background technique

With the development of power electronics technology, computer technology and automatic control technology, AC motor speed regulation technology has also been greatly developed, and the performance of AC motors has shown a strong advantage. In the AC speed control system, compared with the asynchronous motor speed control, the synchronous motor speed control has the advantages of high power factor, small frequency converter capacity and small moment of inertia. In the high-power AC drive system, the advantages of the synchronous motor speed control system obvious. With the development and application of permanent magnet materials, permanent magnet synchronous motor (Permanent Magnet Synchronous Motor, PMSM) has attracted more and more attention due to its high power density, large torque-to-inertia ratio and fast dynamic response.

The controller of the speed loop and current loop of the traditional permanent magnet synchronous motor is a PI controller, but the parameter setting of the PI controller can only be applied to a specific working range. When the working state of the motor changes, the control effect of the PI controller changes. Poor, unable to provide better performance, and the neural network can be adjusted by learning rules to optimize the system characteristics.

Contents of the invention

In order to overcome the problem that the PI controller of the traditional permanent magnet synchronous motor can only be applied to a specific working range, and the control effect of the PI controller becomes worse when the working state of the motor changes, and cannot provide better performance, the present invention provides A self-tuning method for pid parameters of permanent magnet synchronous motor control system based on neural network adaptive control.

Technical scheme of the present invention is as follows:

The present invention at first provides a kind of method for self-tuning of PID parameter of permanent magnet synchronous motor control system based on neural network self-adaptive control, and it comprises the following steps:

1) Design the neural network structure of the adaptive controller of the speed loop, determine the output of the input layer of the neural network according to the structure of the motor control system, determine the output of the hidden layer of the neural network according to the output of the input layer and the activation function of the hidden layer, and determine the output of the hidden layer of the neural network according to the output of the hidden layer of the neural network and The activation function determines the output of the neural network output layer;

2) Obtain the motor speed and rotor position information from the motor sensor, take the reference speed of the motor, the actual speed of the motor and the speed difference between the two as the input of the speed loop neural network adaptive controller, and use the gradient descent algorithm to adjust the neural network adaptive control online The neural network weights of the controller; the output value of the neural network output layer is the gain link parameter k _p , the integral link parameter k _i , and the differential link parameter k _d of the speed loop pid, and then the neural network self-adaptive Controller output:

u(k)=u(k-1)+k _p (e(k)-e(k-1))+k _i e(k)+k _d (e(k)-2e(k-1)+ e(k-2))

Among them, k represents a discrete amount of time, and the controller output u(k) is the q-axis reference current i _q ^* ,

3) Input the q-axis reference current and the actual q-axis current value into the current loop feed-forward pi controller to obtain the q-axis and d-axis reference voltages; through coordinate transformation according to the q-axis and d-axis reference voltages and rotor position information, The switch of the inverter bridge of the permanent magnet synchronous motor control system is adjusted by means of svpwm modulation to control the motor speed and realize the feedback control of the motor speed.

As a preferred solution of the present invention, the neural network structure of the adaptive controller of the design speed loop in the described step 1) is specifically:

Choose a three-layer neural network, including an input layer, a hidden layer and an output layer. The input node corresponds to the motor reference speed, the actual speed of the motor and the speed difference of the motor control system; the output node corresponds to the three parameters kp of the speed loop pid , ki, kd; the number j of input layer nodes is 3, the number i of output layer nodes is 3, and the number of hidden layer nodes is

As a preferred solution of the present invention, in the described step 1), the input layer output is determined according to the structure of the motor control system, specifically:

For a motor control system, the output of the input layer of the neural network is chosen to be:

Among them, x(j) represents the input of the neural network, that is, the reference speed of the motor, the actual speed of the motor and the speed difference, and j represents the neuron node; the subscript (1) represents the first layer of the neural network.

As a preferred solution of the present invention, in the described step 1), the hidden layer output is determined according to the input layer output and the hidden layer activation function, specifically:

The hidden layer input obtained from the input layer output is:

Among them, ω _ij represents the connection weight between corresponding neurons; O _j is the output of the input layer; the subscript (2) represents the hidden layer;

The output of the hidden layer is:

Among them, g(x) represents the activation function of the hidden layer,

As a preferred solution of the present invention, the output of the output layer is determined according to the output of the hidden layer and the activation function in the described step 1), specifically:

The output layer input is:

The output of the output layer is:

Among them, ω _li represents the connection weight between the corresponding neurons in the hidden layer and the output layer; O _i is the output of the hidden layer; the subscript (3) represents the output layer; f(x) represents the activation function of the hidden layer. Corresponding to pid parameters k _p , _ki , k _d , so f(x) is a non-negative activation function, f(x)=ReLU(x)=max(0,x);

As a preferred solution of the present invention, the online adjustment of the neural network weights of the neural network adaptive controller in the described step 2) is specifically:

Take the performance index function as:

Where r(k) represents the reference speed, y(k) represents the actual speed;

Use the gradient descent algorithm to modify the connection weight function of the neural network, and add an inertia term to make the search quickly converge to the global minimum, so we get:

Among them, η is the learning rate, α is the inertia coefficient, Solve for the first term of the expression for the amount of change in the connection weights from the hidden layer to the output layer:

The actual motor control pid algorithm uses a discrete incremental algorithm, for the incremental algorithm,

So you can get:

Then according to the input and output functions of the previous layers, we can get:

make Then the above formula can be written as:

In the same way, the formula for calculating the weight of the hidden layer is:

in

Then the weights can be updated according to the collected data.

As a preferred solution of the present invention, in the step 2), the learning rate η is updated, specifically:

Learning rate η(k)=σ*η(k-1)

Among them, the change factor σ=2 ^λ , λ=sign(g(k)*g(k-1));

where λ reflects the gradient direction; therefore, the learning rate gradient directions for the output layer and the hidden layer are:

The learning rate is updated according to the above.

Compared with the prior art, the present invention has the advantages that: the present invention replaces the traditional pid controller with a neural network adaptive controller to reduce the reduction and improve the system control accuracy; meanwhile, the gradient descent algorithm is used to adjust the neural network weights online to realize Adaptive function, improve the adaptive ability of the control system.

Description of drawings

Fig. 1 is a flow chart of the system of the present invention;

Fig. 2 is a schematic diagram of the system of the present invention;

Fig. 3 is the neural network schematic diagram that the present invention designs;

Fig. 4 is the subject circulation process of neural network algorithm of the present invention;

Fig. 5 is the speed curve diagram of the present invention and traditional pid algorithm when no-load;

Fig. 6 is the speed curve diagram of the present invention and traditional pid algorithm during direct load;

Fig. 7 is a speed curve diagram of the present invention and the traditional pid algorithm when the load is changed.

Detailed ways

The present invention will be further elaborated and described below in combination with specific embodiments. The embodiments are merely exemplary of the disclosure and do not delineate the scope of limitation. The technical features of the various implementations in the present invention can be combined accordingly on the premise that there is no conflict with each other.

Referring to Figure 2, the overall process of permanent magnet synchronous motor control mainly includes the following:

① Acquisition of the running state of the motor includes information such as rotor rotational angular velocity ω _r , position θr, etc. The position information is used as the basis for coordinate transformation, and the rotational speed information converts the angular velocity ω _r into the rotational speed per minute N and sends it into the rotational speed loop;

② According to the difference between the current speed information N and the expected reference speed N _ref , it is sent to the PI controller to make the actual speed track the reference speed, which is often called the speed loop;

③The output of the speed loop PI regulator can be used as the reference value i _qref of the q-axis current in the next link, and then compared with the actual q-axis current value i _q to be sent to the PI regulator for tracking, and then combined with the d-axis for PI adjustment , so that the q-axis and d-axis reference voltages v _q ^* and v _d ^* are obtained, which is the current loop;

④ According to the q-axis and d-axis reference voltage and the rotor position, v _α ^* and v _β ^* can be obtained through coordinate transformation, and then the switch of the inverter bridge is adjusted through svpwm modulation to control the motor speed; thereby realizing the feedback of the motor speed control.

The invention mainly aims at the problem that the PI controller of the traditional permanent magnet synchronous motor can only be applied to a specific working range, and the control effect of the PI controller becomes worse when the working state of the motor changes. Flow chart of the present invention as shown in Figure 1, comprises the steps:

1) Design the neural network structure of the adaptive controller of the speed loop, determine the output of the input layer of the neural network according to the structure of the motor control system, determine the output of the hidden layer of the neural network according to the output of the input layer and the activation function of the hidden layer, and determine the output of the hidden layer of the neural network according to the output of the hidden layer of the neural network and The activation function determines the output of the neural network output layer;

2) Obtain the motor speed and rotor position information from the motor sensor, take the reference speed of the motor, the actual speed of the motor and the speed difference between the two as the input of the speed loop neural network adaptive controller, and use the gradient descent algorithm to adjust the neural network adaptive control online The neural network weights of the controller; the output value of the neural network output layer is the gain link parameter k _p , the integral link parameter k _i , and the differential link parameter k _d of the speed loop pid, and then the neural network self-adaptive Controller output:

u(k)=u(k-1)+k _p (e(k)-e(k-1))+k _i e(k)+k _d (e(k)-2e(k-1)+ e(k-2))

Among them, k represents a discrete amount of time, and the controller output u(k) is the q-axis reference current i _qref ,

3) Input the q-axis reference current and the actual q-axis current value into the current loop feed-forward pi controller to obtain the q-axis and d-axis reference voltages; through coordinate transformation according to the q-axis and d-axis reference voltages and rotor position information, The switch of the inverter bridge of the permanent magnet synchronous motor control system is adjusted by means of svpwm modulation to control the motor speed and realize the feedback control of the motor speed.

The invention determines the input and output physical quantities of the neural network self-adaptive controller by establishing a permanent magnet synchronous motor mathematical model. First write the motion equation of the three-phase permanent magnet synchronous motor:

Among them, J is the moment of inertia, B is the damping coefficient, T _e is the electromagnetic torque, T _L is the load torque, ω _m is the mechanical angular velocity of the motor; p _n is the number of pole pairs of the motor, and i _d and i _q represent d, The current component of the q-axis, L _d , L _q represent the inductance of the d-axis and q-axis respectively, and ψ _f is the flux linkage of the permanent magnet;

Write the current equation in the dq coordinate system of the rotor field-oriented synchronous speed rotation as follows:

Among them, u _d and u _q represent the voltage components of the d and q axes respectively; ω _e is the electrical angular velocity of the motor;

It can be seen from the above formula that the stator currents _id and i _q generate cross-coupling electromotive force in the directions of the d-axis and q-axis respectively. If the stator currents _id and i _q are completely decoupled, the above formula will become:

u _d ^* and u _q ^* represent the voltage components of the d and q axes respectively after current decoupling; when using a conventional pi regulator, the voltages of the d and q axes can be obtained as:

Where v _d ^* and v _q ^* are reference voltages after pi adjustment. Therefore, for the current loop, the current error can be used as the input of the neural network adaptive controller, and the d and q voltages can be used as the output of the controller; for the speed loop, the speed error can be used as the input of the neural network adaptive controller, and the q-axis current can be used as the input of the neural network adaptive controller. Controller output. The invention mainly provides a neural network adaptive controller for the speed loop. For the current loop, a neural network adaptive controller can be designed with reference to the method of the present invention, or an existing traditional adaptive controller can be used.

As shown in Fig. 4, the present invention mainly designs a kind of rotational speed ring neural network adaptive controller, and learns neural network, online adjustment network parameter, realizes pid control parameter adaptive adjustment, the design of neural network adaptive controller And parameter adjustment specifically includes the following steps:

Step 1: Determine the neural network structure.

As shown in Figure 3, a three-layer neural network is selected, with an input layer, a hidden layer and an output layer. The input nodes correspond to the state quantities of the motor control system—the reference speed of the motor, the actual speed of the motor, and the speed difference; the output nodes correspond to Based on the three parameters kp, ki, kd of the pid controller. Therefore, the number of nodes in the input layer is j=3, the number of nodes in the output layer is i=3, and the number of nodes in the hidden layer is So take l=5.

Step 2: Determine the input layer output according to the motor control system structure.

For this control system, the input layer output of the neural network is chosen as:

Among them, x represents the input of the neural network, that is, the reference speed of the motor, the actual speed of the motor and the speed difference, and j represents the neuron node; the subscript (1) represents the first layer of the neural network;

Step 3: Determine the output of the hidden layer according to the output of the input layer and the activation function of the hidden layer.

The hidden layer input obtained from the input layer output is:

Among them, ω _ij represents the connection weight between corresponding neurons; O _j is the output of the input layer; the subscript (2) represents the hidden layer; the output of the output layer is:

Among them, g(x) represents the activation function of the hidden layer,

Step 4: Determine the output layer output according to the hidden layer output and activation function.

The output layer function is similar to the above formula:

Among them, ω _li represents the connection weight between the corresponding neurons in the hidden layer and the output layer; O _i is the output of the hidden layer; the subscript (3) represents the output layer; f(x) represents the activation function of the hidden layer. Corresponding to the pid parameter, so f(x) is a non-negative activation function, f(x)=ReLU(x)=max(0,x);

Step 5: Next, update the connection weights of each layer.

Take the performance index function as:

Use the gradient descent algorithm to modify the connection weight function of the network, and add an inertial term so that the search can quickly converge to the global minimum, so we can get:

Among them, η is the learning rate, α is the inertia coefficient, and solve the first term of the expression:

which for Since the relative change of y(k) and u(k) can be measured, that is

It can also be approximated by a symbolic function, that is, This can simplify the calculation. On the other hand, in actual operation, a small amount can be added to the denominator to prevent the denominator from being 0 at the initial moment; the actual motor control pid algorithm can use a discrete incremental algorithm. For the incremental algorithm, />

So you can get:

Then according to the input and output functions of the previous layers, we can get:

make Then the above formula can be written as:

In the same way, the formula for calculating the weight of the hidden layer is:

in

Then the weights can be updated according to the collected data.

Step 6: Update of learning rate.

The additional inertia item faces the difficulty of selecting the selection rate, which leads to the contradiction between the convergence speed and the convergence. Therefore, the introduction of learning rate adaptive design is also considered.

η(k)=σ*η(k-1)

Where σ is the variation factor σ=2 ^λ , λ=sign(g(k)*g(k-1));

where λ reflects the gradient direction. Therefore, the learning rate gradient directions for the output layer and the hidden layer are:

The learning rate can be updated according to the above.

Then the weights can be updated according to the collected data. The output value of the neural network output layer is the k _p , k _i , k _d of the pid parameters, and then the controller output is obtained by using the incremental pid algorithm:

u(k)=u(k-1)+k _p (e(k)-e(k-1))+k _i e(k)+k _d (e(k)-2e(k-1)+ e(k-2))

The controller output u(k) is the q-axis reference current i _q ^* , and the q-axis reference current is input into the current loop feedforward pi controller to control the operation of the motor.

Analysis of simulation results: In order to verify the feasibility of the proposed method, simulation verification was carried out in the Matlab/Simulink environment, and it was compared with the PI regulator.

Fig. 5 is the speed curve diagram of the present invention and traditional pid algorithm when no-load, as can be seen from Fig. 5 when motor is no-load, the speed curve overshoot of the present invention is far less than traditional pid algorithm, and the adjustment time to reach steady state is also far Less than the traditional pid algorithm; Fig. 6 is the speed curve diagram of the present invention and the traditional pid algorithm during direct load, as can be seen from Fig. 6 when the motor is directly applied with load operation, the speed curve overshoot of the present invention is far less than the traditional pid algorithm, and reaches The adjustment time of the steady state is also much shorter than the traditional pid algorithm; Fig. 7 is the speed curve diagram of the present invention and the traditional pid algorithm when the load is changed. -0.4s when the load runs, and then it can be seen that the time for the present invention to restore the stable speed after the sudden load is shorter than the traditional pid algorithm; from the above figure, it can be seen that the dynamic performance of the algorithm of the present invention such as overshoot and adjustment time is excellent In the traditional pid algorithm. In summary, the algorithm proposed in this paper can achieve the expected control accuracy and can effectively improve the performance of the motor operation.

Finally, it should be noted that: the above is only a preferred embodiment of the present invention, and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still The technical solutions recorded in the foregoing embodiments may be modified, or some technical features thereof may be equivalently replaced. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (7)

1. A method for self-tuning of permanent magnet synchronous motor control system pid parameters based on neural network adaptive control, is characterized in that, comprises the steps: 1) Design the neural network structure of the adaptive controller of the speed loop, determine the output of the input layer of the neural network according to the structure of the motor control system, determine the output of the hidden layer of the neural network according to the output of the input layer and the activation function of the hidden layer, and determine the output of the hidden layer of the neural network according to the output of the hidden layer of the neural network and The activation function determines the output of the neural network output layer; 2) Obtain the motor speed and rotor position information from the motor sensor, take the reference speed of the motor, the actual speed of the motor and the speed difference between the two as the input of the speed loop neural network adaptive controller, and use the gradient descent algorithm to adjust the neural network adaptive control online The neural network weights of the controller; the output value of the neural network output layer is the gain link parameter k _p , the integral link parameter k _i , and the differential link parameter k _d of the speed loop pid, and then the neural network self-adaptive Controller output: u(k)=u(k-1)+k _p (e(k)-e(k-1))+k _i e(k) +k _d (e(k)-2e(k-1)+e(k-2)) Among them, k represents a discrete amount of time, and the controller output u(k) is the q-axis reference current i _q ^* , 3) Input the q-axis reference current and the actual q-axis current value into the current loop feed-forward pi controller to obtain the q-axis and d-axis reference voltages; through coordinate transformation according to the q-axis and d-axis reference voltages and rotor position information, The switch of the inverter bridge of the permanent magnet synchronous motor control system is adjusted by means of svpwm modulation to control the motor speed and realize the feedback control of the motor speed. 2. method according to claim 1, is characterized in that, the neural network structure of the adaptive controller of the design speed loop in described step 1), specifically: Choose a three-layer neural network, including an input layer, a hidden layer and an output layer. The input node corresponds to the motor reference speed, the actual speed of the motor and the speed difference of the motor control system; the output node corresponds to the three parameters kp of the speed loop pid , ki, kd; the number j of input layer nodes is 3, the number i of output layer nodes is 3, and the number of hidden layer nodes is 3. The method according to claim 1, characterized in that, in said step 1), determining the input layer output according to the motor control system structure is specifically: For a motor control system, the output of the input layer of the neural network is chosen to be: Among them, x(j) represents the input of the neural network, that is, the reference speed of the motor, the actual speed of the motor and the speed difference, and j represents the neuron node; the subscript (1) represents the first layer of the neural network. 4. method according to claim 1, is characterized in that, in described step 1) according to input layer output and hidden layer activation function, determine hidden layer output, specifically: The hidden layer input obtained from the input layer output is: Among them, ω _ij represents the connection weight between corresponding neurons; O _j is the output of the input layer; the subscript (2) represents the hidden layer; The output of the hidden layer is: Among them, g(x) represents the activation function of the hidden layer, 5. method according to claim 1, is characterized in that, in described step 1) according to hidden layer output and activation function, determine output layer output, specifically: The output layer input is: The output of the output layer is: Among them, ω _li represents the connection weight between the corresponding neurons in the hidden layer and the output layer; O _i is the output of the hidden layer; the subscript (3) represents the output layer; f(x) represents the activation function of the hidden layer. Corresponding to pid parameters k _p , _ki , k _d , so f(x) is a non-negative activation function, f(x)=ReLU(x)=max(0,x); 6. method according to claim 1, is characterized in that, described step 2) in the neural network weights of online adjustment neural network adaptive controller, specifically: Take the performance index function as: Where r(k) represents the reference speed, y(k) represents the actual speed; Use the gradient descent algorithm to modify the connection weight function of the neural network, and add an inertia term to make the search quickly converge to the global minimum, so we get: Among them, η is the learning rate, α is the inertia coefficient, Solve for the first term of the expression for the amount of change in the connection weights from the hidden layer to the output layer: The actual motor control pid algorithm uses a discrete incremental algorithm, for the incremental algorithm, So we can get: Then according to the input and output functions of the previous layers, we can get: make Then the above formula can be written as: In the same way, the formula for calculating the weight of the hidden layer is: in Then the weights can be updated according to the collected data. 7. method according to claim 6, is characterized in that, in described step 2), learning rate n is updated, specifically: Learning rate η(k)=σ*η(k-1) Among them, the change factor σ=2 ^λ , λ=sign(g(k)*g(k-1)); where λ reflects the gradient direction; therefore, the learning rate gradient directions for the output layer and the hidden layer are: The learning rate is updated according to the above.