A control method of permanent magnet synchronous motor based on K-means optimization fuzzy RBF neural network

Abstract

The application relates to the technical field of motor control, and discloses a permanent magnet synchronous motor control method based on K-Means optimized fuzzy RBF neural network, which initializes fuzzy RBF neural network parameters through K-Means clustering; actual output values V of a permanent magnet motor vector control system are obtained through sampling, and the actual output values V are subtracted from given values V ★ to calculate the deviation and the deviation change of the system; a fuzzy RBF neural network controller is constructed, e(k) and e c (k) of the system are taken as inputs of the fuzzy RBF neural network controller, k p , k i , k d are obtained through a four-layer network, and the outputs are input into a PID controller to obtain an output control amount iq, and current loop PID control is continuously carried out; the motor control method has obviously improved steady-state and dynamic performance, has strong robustness, and can effectively enhance the anti-interference capability of the motor.

Description

Technical Field

[0001] This invention belongs to the field of motor control technology, and in particular relates to a control method for permanent magnet synchronous motors based on K-Means optimized fuzzy RBF neural networks. Background Technology

[0002] With the development of rare-earth permanent magnet materials and power electronics technology, the modern control theory of permanent magnet synchronous motors (PMSMs) has gradually become complete. Due to their advantages in power, size, precision, and fast response performance, PMSMs have become one of the mainstream motors widely used in high-precision servo systems. Traditional proportional-integral-derivative (PID) controllers and other linear theory algorithms have been widely applied in engineering, but they cannot overcome the system problems caused by the nonlinearity and time-varying parameters of PMSMs. Fortunately, intelligent control theory has proposed a series of ideas and methods to solve this problem, mainly including fuzzy control technology, chaotic control technology, expert control technology, and integrated intelligent control. However, besides being relatively complex, these methods have also encountered a series of problems in practical applications. To date, many traditional control experts have not fully accepted intelligent control, thus limiting the promotion of intelligent controllers in the field of motor control. Fuzzy RBF neural networks, after good training, have shown stable and fast control effects.

[0003] Research on Permanent Magnet Synchronous Motor Servo System Based on Fuzzy Neural Network [1] published a control method based on fuzzy RBF neural network controller. The paper proposes a fuzzy RBF neural network PID control strategy to replace the speed loop PID control in the vector control system. The neural network and fuzzy control are combined, and the PID parameters in the speed loop are dynamically adjusted by using the gradient descent optimization algorithm based on the incremental PID control method. However, in the above method, the initial parameters of the fuzzy RBF neural network: the number of neurons in the fuzzification layer, the base width of the RBF function and the center point do not have a reasonable range. Whether the fuzzy RBF neural network converges and the convergence speed during training are very important to the selection of the initial value of the parameters. If the initial value is 1 or a random number, the neural network needs to learn or acquire through many times during training. Such initial value selection will make it difficult to achieve the ideal effect of training. Since the number of neurons in the fuzzification layer will affect the performance of the network, if an inappropriate number of neurons in the fuzzification layer is selected, the constructed fuzzy RBF neural network model will not be able to match the controlled object, thus making it difficult to achieve the ideal control effect.

[0004] Therefore, a control method for permanent magnet synchronous motors based on K-Means-optimized fuzzy RBF neural networks is of great significance for selecting appropriate initialization parameters for training the fuzzy RBF neural network.

[0005] [1] Wang Peiyu, Ma Lixin. Research on servo system of permanent magnet synchronous motor based on fuzzy neural network [J]. Electronic Science and Technology, 2022, 35(06):83-88. DOI:10.16180 / j.cnki.issn1007-7820.2022.06.013. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a permanent magnet synchronous motor control method based on K-Means optimized fuzzy RBF neural network, so as to effectively improve the dynamic performance of permanent magnet synchronous motor control and the training speed of the network used.

[0007] This invention discloses a control method for a permanent magnet synchronous motor based on a K-Means optimized fuzzy RBF neural network, comprising the following steps:

[0008] S1. Initialize the parameters of the fuzzy RBF neural network by K-Means clustering, including the center point of the RBF function, the base width and the number of neurons in the fuzzification layer, and set the initial values ​​of the weights, learning rate η and momentum factor α of each layer in the fuzzy RBF neural network, including the input layer, fuzzification layer, fuzzy inference layer and output layer.

[0009] S2. Obtain the actual output value V of the permanent magnet motor vector control system through sampling, and compare it with the given value V. ★ By taking the difference, the system deviation e(k) = V is calculated. ★ -V and deviation change e c (k) = e(k) - e(k-1);

[0010] S3. Construct a fuzzy RBF neural network controller

[0011] The system's e(k) and e c (k) is used as the input to the fuzzy RBF neural network controller, and is obtained by passing it through four layers: the input layer, the fuzzification layer, the fuzzy inference layer, and the output layer. p k i k d The input is fed into the PID controller to obtain the output control quantity iq. The control quantity iq is then added to the controlled object to generate the actual output value V(k+1) at the next sampling time, where k p k i k d These are the proportional adjustment coefficient, integral adjustment coefficient, and derivative adjustment coefficient, respectively.

[0012] S4. Update the base width b of the fuzzy RBF neural network using the gradient descent method. ij and weight ω;

[0013] S5. Let k = k + 1, return to steps S2-S3 and recalculate until k... p k i k d No further iterations; where k p k i k d These are the proportional adjustment coefficient, integral adjustment coefficient, and derivative adjustment coefficient, respectively.

[0014] The K-Means clustering in step S1 above divides Q(1,2,3…q) samples into K(1,2,3…k) categories by iteratively updating the cluster centers, ensuring that the sum of squared errors (distances from each data point to its corresponding cluster center) is minimized. When the data error remains unchanged, it indicates that the clustering function has converged. In each iteration, the classification of each sample must be considered. If it is incorrect, adjustments must be made to ensure that the data groups are as similar as possible within each group and the similarity between groups is as small as possible. The Euclidean distance method and the proximity principle are used to determine the cluster centers and the categories to which the samples belong during classification.

[0015] In step S1 above, the parameters of the fuzzy RBF neural network are initialized using K-Means clustering as follows:

[0016] Samples X of the same cluster i (i = 1, 2, 3…n) with respect to cluster center X c The Euclidean distance is:

[0017]

[0018] In equation (1), X i Let Q be the number of samples (i = 1, 2, 3…n). k is the number of clusters; w is the number of samples X. i Number of attribute metrics, X c X is the cluster center of the sample (c = 1, 2, 3…k). ki For sample X i The k-attribute index value, X kc The cluster centers of the index attribute k (c = 1, 2, 3…k); d(X) i ,X c ) is the sample X i Regarding cluster center X c The Euclidean distance, d(X) i ,X c This is used as the initial value for the basis width of the RBF neural network;

[0019] Sample X i The samples are divided into K classes through calculation, and the mean value of samples within each class is calculated as the new cluster center. Then, the sample X is recalculated.i The distance from the cluster centers is determined by multiple iterations until the cluster centers no longer move, thus obtaining the center point of the RBF function.

[0020]

[0021] In equation (2), m is the number of iterations, and X kc (m) D represents the cluster center of attribute index k after the m-th iteration; t (m-1) The classification result after the (m-1)th iteration (t = 1, 2, 3…K); h t (m-1) For category D after the (m-1)th iteration t (m-1) The number of samples in X kc (m) As the center point of the RBF neural network's base width, the number of neurons in the fuzzy layer is B = S, where K ≤ S ≤ Q.

[0022] In step S2 above, the permanent magnet motor vector control system adopts a vector control method with id=0 for both speed and current closed loops.

[0023] The above input layer is the first layer. Each node in this layer is directly connected to each component of the input, passing the input to the next layer. The input and output of each node i in this layer is represented as: f1(i) = x i ;

[0024] The second layer: the fuzzification layer, which uses a Gaussian function as the membership function, c ij and b j These are the center point and weight of the membership function of the j-th fuzzy set of the i-th input variable:

[0025] f2(i,j)=exp(net) j 2 (3)

[0026]

[0027] The third layer is the fuzzy inference layer. This layer completes the matching of fuzzy rules through connection with the fuzzification layer. Fuzzy operations are performed between the nodes, and the output of each node j is the product of all the input signals of that node.

[0028]

[0029] In equation (5), N i It is the number of the i-th input membership function in the input layer, i.e., the number of nodes in the fuzzification layer;

[0030] Fourth layer: Output layer, the output layer is f4.

[0031]

[0032] In equation (6), l is the number of output layer nodes, and w is the connection weight matrix between the output layer nodes and each node in the third layer;

[0033] The final output of f4 is k p k i k d Three parameters are input into the PID controller to obtain the output control quantity iq. This control quantity iq is then input into the current loop for further processing. d Vector control with =0 generates the actual output value V(k+1) of the motor at the next sampling time.

[0034] In step S4 above, the basis width b of the fuzzy RBF neural network is updated. ij The specific steps for determining the weight ω are as follows:

[0035] The performance index function of the permanent magnet motor vector control system is:

[0036]

[0037] Gradient descent is used to adjust network parameters, meaning that E adjusts the parameters that need adjustment in the direction of the negative gradient. Each parameter is adjusted according to the following formula.

[0038]

[0039] In equation (8), the learning efficiency η ranges from 0 to 1. The smaller η is, the slower the learning. If η is too large, it will cause oscillations in the control system. The momentum factor α is taken as 0. x is the parameter that needs to be adjusted, including the output weight ω and the base width b. ij ,in:

[0040] Weight ω adjustment:

[0041] Base width b ij Adjustment:

[0042] Where p, i, and d are the proportional, integral, and derivative control parameters, respectively.

[0043] In equation (10), Use the sig function instead of calculation.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] This invention proposes a control method for permanent magnet synchronous motors (PMSMs) based on a K-Means optimized fuzzy RBF neural network. The PMSM employs a speed-current dual-closed-loop vector control method with id=0. Specifically, the K-Means clustering method is used to initialize the parameters of the fuzzy RBF neural network controller, including the number of hidden layers, the base width of the RBF function, and its center point. The fuzzy RBF neural network controller is applied to the current-loop PID controller and contains four layers: an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. The k-value is obtained through the fuzzy RBF neural network controller. p k i k d The input is fed into the PID controller to obtain iq, and then current loop PID control is performed. This method combines classical control theory and intelligent control methods. The design process is simple and efficient. Compared with conventional PID control, the steady-state and dynamic performance is significantly improved, and it has strong robustness, which can effectively enhance the motor's anti-interference ability. Attached Figure Description

[0046] Figure 1 The block diagram shows a permanent magnet synchronous motor control system based on a K-Means optimized fuzzy RBF neural network.

[0047] Figure 2 This is a diagram of the fuzzy RBF neural network structure.

[0048] Figure 3 The flowchart shows the control method for permanent magnet synchronous motors based on K-Means optimized fuzzy RBF neural networks. Detailed Implementation

[0049] Now combined with the appendix Figure 1-3 The present invention will be further described as follows:

[0050] This invention proposes a control method for permanent magnet synchronous motors based on a K-Means optimized fuzzy RBF neural network, wherein: this method combines the characteristics of K-Means clustering parameter initialization and fuzzy RBF neural network, and the design steps are as follows:

[0051] Step 1: Initialize the fuzzy RBF neural network using K-Means clustering, including the center point of the RBF function, the base width, the number of neurons in the fuzzification layer, setting the initial values ​​of the weights of each layer of the network, the learning rate η, and the momentum factor α.

[0052] Step 2: Obtain the actual system output value V through sampling, and compare it with the given value V. ★ By taking the difference, the system deviation e(k) = V is calculated. ★ -V and deviation change e c (k) = e(k) - e(k-1);

[0053] Step 3: Combine e(k) and e c (k) is used as the input of the fuzzy RBF neural network controller. kp, ki, and kd are obtained through a four-layer network and then input into the PID controller to obtain the output control quantity iq. The control quantity iq is added to the controlled object to generate the actual output value V(k+1) at the next sampling time.

[0054] Step 4: Update the base width b of the fuzzy RBF neural network ij and weight ω;

[0055] Step 5: Let k = k + 1, return to step 2 and recalculate until kp, ki, and kd no longer iterate.

[0056] In the implementation process, K-Means clustering in step 1 divides N samples into K categories by iteratively updating the cluster centers, ensuring that the sum of the squared distances (errors) from each data point to its corresponding cluster center is minimized. When the data error remains unchanged, it indicates that the clustering function has converged. During each iteration, the classification of each sample must be considered for accuracy. If incorrect, adjustments are made to ensure that data groups are as similar as possible within each group and as similar between groups as possible. The Euclidean distance method and the proximity principle are used to determine the cluster centers and the categories to which the samples belong.

[0057] During implementation, the permanent magnet motor vector control system in step 2 adopts a vector control method with id=0 for both speed and current closed loops.

[0058] In the implementation process, the four-layer network in step 3 consists of an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. kp, ki, and kd are obtained through the fuzzy RBF neural network controller, and then input into the PID controller to obtain iq, which is then used for current loop PID control.

[0059] In the implementation process, step 4, updating the basis width and weights of the fuzzy RBF neural network, employs the gradient descent method. The system performance index function is defined as follows: Next, gradient descent is used to adjust the network parameters so that E adjusts the parameters that need to be adjusted in the direction of the negative gradient.

[0060] This invention proposes a control method for permanent magnet synchronous motors (PMSMs) based on a K-Means optimized fuzzy RBF neural network. The PMSM employs a speed-current dual-loop vector control method with id=0. The K-Means clustering method is used to initialize the parameters of the fuzzy RBF neural network controller, including the number of hidden layers, the base width of the RBF function, and its center point. The fuzzy RBF neural network controller, applied to the current loop PID controller, contains four layers: an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer. kp, ki, and kd are obtained through the fuzzy RBF neural network controller and input into the PID controller to obtain iq, which is then used for further current loop PID control. The control system block diagram is shown below. Figure 1 As shown in the figure. This method combines classical control theory and intelligent control methods. The design process is simple and efficient. Compared with conventional PID control, it significantly improves steady-state and dynamic performance, has strong robustness, and can effectively enhance the anti-interference ability of the motor.

[0061] The specific implementation process of the permanent magnet synchronous motor control method based on K-Means optimized fuzzy RBF neural network is as follows:

[0062] Step 1: Initialize the fuzzy RBF neural network using K-Means clustering, including the center point of the RBF function, the base width, the number of neurons in the fuzzification layer, the initial values ​​of the weights of each layer of the network, the learning rate η, and the momentum factor α.

[0063] K-Means clustering divides Q(1,2,3…q) samples into K(1,2,3…k) categories by iteratively updating cluster centers, ensuring that the sum of squared distances (errors) from each data point to its corresponding cluster center is minimized. When the data error remains unchanged, the clustering function has converged. During each iteration, the classification of each sample is considered; if incorrect, adjustments are made to ensure that data groups are as similar as possible within each group and as similar between groups as possible. Euclidean distance and the proximity principle are used to determine cluster centers and the category to which a sample belongs. Based on the K-Means clustering method, samples X belonging to the same cluster are obtained. i (i = 1, 2, 3…n) with respect to cluster center X c Euclidean distance In the formula, X i Let Q be the number of samples (i = 1, 2, 3…n). X c The cluster centers of the samples are (c = 1, 2, 3…k), where k is the number of clusters; w is the number of samples X. i Number of attribute metrics; X ki For sample X i The k-attribute index value, X kcThe cluster centers of the index attribute k (c = 1, 2, 3…k); d(X) i ,X c ) is the sample X i Regarding cluster center X c The Euclidean distance, d(X) i ,X c This is used as the initial value for the basis width of the RBF neural network. Sample X i The samples are divided into K classes through calculation, and the average value of the samples within each group is used as the new cluster center. Then, the distance between the samples and the cluster centers is recalculated. After multiple iterations, the classification is completed when the cluster centers no longer move. In the formula, m is the number of iterations, and X kc (m) D represents the cluster center of attribute index k after the m-th iteration; t (m-1) The classification result after the (m-1)th iteration (t = 1, 2, 3…K); h t (m-1) For category D after the (m-1)th iteration t (m-1) The number of samples in X kc (m) The center point of the RBF neural network's basis width is used. The number of neurons in the fuzzy layer is B = S, (K ≤ S ≤ Q).

[0064] Step 2: Obtain the actual system output value V through sampling, and compare it with the given value V. ★ By taking the difference, the system deviation e(k) = V is calculated. ★ -V and deviation change e c (k) = e(k) - e(k-1); The actual speed V of the motor is obtained by sampling through a rotary transformer, and compared with the given speed V ★ By subtracting, we get e(k) and e c The value of (k).

[0065] Step 3: Constructing a fuzzy RBF neural network controller

[0066] Let e(k) and e c (k) serves as the input to the fuzzy RBF neural network controller, and kp, ki, and kd are obtained through a four-layer network. These four layers are the input layer, fuzzification layer, fuzzy inference layer, and output layer, respectively. Figure 2 As shown.

[0067] The first layer is the input layer. Each node in this layer is directly connected to a component of the input layer, passing the input to the next layer. The input and output of each node i in this layer is represented as: f1(i) = x i ;

[0068] The second layer: the fuzzification layer, which uses a Gaussian function as the membership function, c ij and b j These are the center point and weight of the membership function of the j-th fuzzy set of the i-th input variable, respectively. That is, f2(i,j) = exp(net). j 2 ),

[0069] The third layer is the fuzzy inference layer. This layer completes fuzzy rule matching through connections with the fuzzification layer, and fuzzy operations are performed between the nodes. The output of each node j is the product of all its input signals, i.e. In the formula, N i It is the number of the i-th input membership function in the input layer, i.e., the number of fuzzy layer nodes.

[0070] Fourth layer: Output layer, the output layer is f4, that is... In the formula, l is the number of output layer nodes, and w is the connection weight matrix between the output layer nodes and each node in the third layer;

[0071] The final output of f4 consists of three parameters: kp, ki, and kd. These three parameters are input into the PID controller to obtain the output control quantity iq. The control quantity iq is then input into the current loop for i... d Vector control with =0 generates the actual output value V(k+1) of the motor at the next sampling time;

[0072] Step 4: Update the base width b of the fuzzy RBF neural network ij and weight ω

[0073] Define the system performance index function as follows: Next, gradient descent is used to adjust the network parameters, specifically adjusting E in the direction of the negative gradient for the parameters that need adjustment. The parameters are configured according to the formula... In the adjustment formula, the learning efficiency η ranges from 0 to 1. A smaller η results in slower learning, while a larger η can cause oscillations in the control system. The momentum factor α is set to 0; x represents the parameters to be adjusted, including the output weight ω and the base width b. ij Weight ω adjustment:

[0074] Base width b ij Adjustment:

[0075] Where p, i, and d are the proportional, integral, and derivative control parameters, respectively.

[0076] In equation (10), Use the sig function to replace the calculation;

[0077] Step 5: Let k = k + 1, return to step 2 and recalculate until kp, ki, and kd no longer iterate.

[0078] The design of a permanent magnet synchronous motor control system based on a K-Means optimized fuzzy RBF neural network was completed.

[0079] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A control method of a permanent magnet synchronous motor based on K-Means optimization fuzzy RBF neural network, characterized in that, Includes the following steps: S1. Initialize the parameters of the fuzzy RBF neural network using K-Means clustering, including the center point of the RBF function, the base width, and the number of neurons in the fuzzification layer. Set the initial values ​​of the weights and the learning rate for each layer in the fuzzy RBF neural network, including the input layer, fuzzification layer, fuzzy inference layer, and output layer. and momentum factor ; S2, obtain the actual output value of the permanent magnet motor vector control system through sampling , and the given value , and the deviation , and the change of the deviation ; S3. Construct a fuzzy RBF neural network controller The system and As the input to the fuzzy RBF neural network controller, k is obtained by sequentially passing through four layers: the input layer, the fuzzification layer, the fuzzy inference layer, and the output layer. p k i k d The input is fed into the PID controller to obtain the output control quantity iq. The control quantity iq is then added to the controlled object to generate the actual output value at the next sampling time. ; where k p k i k d These are the proportional adjustment coefficient, integral adjustment coefficient, and derivative adjustment coefficient, respectively. S4, updating the base width of the fuzzy RBF neural network by using gradient descent method and weights ; S5, Order Return to steps S2-S3 and recalculate until k is reached. p k i k d No further iterations, where k p k i k d These are the proportional adjustment coefficient, integral adjustment coefficient, and derivative adjustment coefficient, respectively. The parameters for initializing the fuzzy RBF neural network using K-Means clustering are specifically as follows: Samples of the same cluster The Euclidean distance of (i=1, 2, 3…n) about the cluster center is: （1） In equation (1), Let Q be the number of samples (i=1,2,3…n). k is the number of clusters; w is the number of samples. The cluster centers of the samples are (c=1,2,3…k). For the sample The k-attribute index value, The cluster centers for the index attribute k (c=1,2,3…k); For the sample About cluster centers The Euclidean distance will As the initial value for the basis width of the RBF neural network; sample The samples are divided into K classes through calculation, and the mean value of samples within each class is used as the new cluster center. Then the sample is recalculated. The distance from the cluster centers is determined by multiple iterations until the cluster centers no longer move, thus obtaining the center point of the RBF function. （2） In equation (2), m is the number of iterations. Let be the cluster center of attribute index k after the m-th iteration; The classification result after the (m-1)th iteration (t=1,2,3…K); The category after the (m-1)th iteration The number of samples in the sample will As the center point of the basis width of the RBF neural network, the number of neurons in the fuzzy layer is B=S. .

2. The K-Means optimization based fuzzy RBF neural network-based permanent magnet synchronous motor control method according to claim 1, wherein, The K-Means clustering in step S1 involves iteratively updating the cluster centers to divide Q (1,2,3…q) samples into K (1,2,3…k) categories, ensuring that the sum of squared errors (distances from each data point to its corresponding cluster center) is minimized. When the data error remains unchanged, it indicates that the clustering function has converged. In each iteration, the classification of each sample must be considered. If incorrect, adjustments must be made to ensure that data groups are as similar as possible within each group and as similar as possible between groups. The Euclidean distance method and the proximity principle are used to determine the cluster centers and the categories to which the samples belong during classification.

3. The permanent magnet synchronous motor control method based on K-Means optimized fuzzy RBF neural network as described in claim 1, characterized in that, In step S2, the permanent magnet motor vector control system adopts a vector control method with id = 0 in the speed-current dual closed loop.

4. The control method of permanent magnet synchronous motor based on K-Means optimized fuzzy RBF neural network according to claim 3, wherein, The input layer is the first layer, each node of which is directly connected to each component of the input, transmitting the input quantity to the next layer. The input and output of each node i of the layer is expressed as: ; The second layer: the fuzzy layer, using Gaussian function as membership function, and are the center point and weight value of membership function of the jth fuzzy set of the ith input variable, respectively. （3） （4） Third layer: fuzzy inference layer, which completes the matching of fuzzy rules through the connection with the fuzzy layer, and realizes fuzzy operation between nodes, and the output of each node j is the product of all input signals of the node, (5) In equation (5), , It is the number of the i-th input membership function in the input layer, i.e., the number of nodes in the fuzzification layer; Fourth layer: output layer, the output layer is , （6） In formula (6), is the number of output layer nodes, is a connection weight matrix of the output layer nodes and nodes of the third layer. final The output is k p k i k d Three parameters are input into the PID controller to obtain the output control quantity iq. This control quantity iq is then input into the current loop for further processing. Vector control generates the actual output value of the motor at the next sampling time. .

5. The control method of permanent magnet synchronous motor based on K-Means optimized fuzzy RBF neural network according to claim 1, wherein, In step S4, the base width of the fuzzy RBF neural network is updated. and weight The specific steps are as follows: The performance index function of the permanent magnet motor vector control system is: （7） Gradient descent is used to adjust network parameters, meaning that E adjusts the parameters that need adjustment in the direction of the negative gradient. Each parameter is adjusted according to the following formula: （8） In equation (8), learning efficiency The value range is 0-1. The younger they are, the slower they learn. If the value is too large, it will cause oscillations in the control system; therefore, a momentum factor should be used. ; The parameters that need to be adjusted include the output weights. and base width ,in: weight adjustment (9) Base width Adjustment: (10) wherein, P, I, D are proportional, integral, derivative tuning parameters, respectively; In equation (10), , Use the sig function instead of calculation.

Images