A method for predicting a magnetic field of an electrical equipment

Abstract

The application provides an electric equipment magnetic field prediction method, comprising the following steps: S1, combining the actual operation condition of the electric equipment, a finite element model of the electric equipment is established; S2, according to the working condition data to be predicted, several known working conditions are selected for finite element analysis, and the real magnetic field distribution diagram of the electric equipment under the corresponding several known working conditions is obtained; S3, several factors influencing the magnetic field distribution of the electric equipment are selected as input variables of a deep learning model, and an input information matrix is formed in combination with the geometric structure of the electric equipment; S4, the real magnetic field distribution diagram obtained in step S2 is normalized; S5, the deep learning model is trained, and the magnetic field of the electric equipment is predicted based on the trained deep learning model. Compared with other methods, the prediction accuracy of the application is higher, the prediction workload is smaller, and the application can be flexibly applied to magnetic field prediction under different working conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of equipment magnetic field prediction technology, and particularly relates to an electrical equipment magnetic field prediction method. BACKGROUND

[0002] The most common and important one in electrical equipment is electromagnetic mechanical device, which is applied to various fields such as daily production and life and military use. Magnetic field is a medium for realizing energy conversion. Understanding the magnetic field distribution of electrical equipment is not only beneficial to analyzing and mastering the electromagnetic performance of electrical equipment, realizing the design and optimization of equipment structure, but also beneficial to diagnosing and preventing operation failure and improving the service life of electrical equipment.

[0003] At present, there are some magnetic field prediction technologies for electrical equipment on the market. For example, patent CN112016172 uses an analytical method to predict the magnetic field, applies a sub-domain analytical model based on the separation of variables method to a two-dimensional magnetic quasi-static field, and thus predicts the air gap magnetic field of a solid rotor induction motor. Although this patent can accurately predict the air gap magnetic field by dividing the solution region into sub-regions, the application of this model mainly reflects in the analytical calculation of constant magnetic field. This patent applies it to a two-dimensional magnetic quasi-static field, and the magnetic field of other properties remains to be verified. Moreover, at present, the traditional magnetic circuit analytical calculation method cannot meet the accuracy requirements of magnetic field calculation.

[0004] In comparison, the finite element method reduces various assumptions in analytical calculation and can consider the saturation degree problem of nonlinear materials, and can model various complex structures of electrical equipment. However, the finite element method requires a large number of grid divisions, which leads to large amount of calculation and low efficiency. The working state of electrical equipment is different, and its magnetic field distribution will also be affected to a certain extent. Once the working condition changes, the magnetic field distribution needs to be recalculated, which is a huge task.

[0005] With the increasing development of deep learning, it can extract features from known data learning to realize the prediction and generation of unknown data. Patent CN108372026 proposes a magnetic field effect prediction method in electrostatic precipitator based on BP neural network, which predicts the magnetic field effect in the precipitator under different voltages and temperatures through BP neural network. Although this patent can make a high-precision prediction of the magnetic field effect, it does not explain the selection method of different voltages and temperatures of the electrostatic precipitator. SUMMARY

[0006] Therefore, in order to overcome the above defects, the application aims to provide an electrical equipment magnetic field prediction method.

[0007] To achieve the above purpose, the technical scheme of the application is as follows:

[0008] A method for predicting the magnetic field of electrical equipment, comprising the following steps:

[0009] S1. Establishing a finite element model of the electrical equipment in combination with the actual operating conditions of the electrical equipment;

[0010] S2. Selecting several known conditions for finite element analysis according to the operating condition data to be predicted, to obtain the real magnetic field distribution map of the electrical equipment under the corresponding several known conditions;

[0011] S3. Selecting several factors that have an impact on the magnetic field distribution of the electrical equipment as input variables of the deep learning model, and combining the geometric structure of the electrical equipment to form an input information matrix;

[0012] S4. Normalizing the real magnetic field distribution map obtained in step S2;

[0013] S5. Taking the input information matrix obtained in step S3 and the real magnetic field distribution map processed in step S4 as training sample data and test sample data, training the deep learning model, and realizing the magnetic field prediction of the electrical equipment based on the trained deep learning model.

[0014] Further, in step S1, the actual operating conditions of the electrical equipment include working voltage, working current, torque, frequency, ambient temperature, and humidity.

[0015] Further, in step S2, according to the variational principle of the finite element method, the problem of solving the magnetic field of the motor is converted into the problem of solving the extreme value of the energy functional, and then the electromagnetic field simulation software is used to obtain the real magnetic field distribution map.

[0016] Further, in step S3, the input information matrix formed includes the structural parameters of the electrical equipment and the variables involved in the actual operating conditions;

[0017] The structural parameters of the electrical equipment include the dimensions and materials of each part, and excitation.

[0018] Further, in step S4, the real magnetic field distribution map obtained by finite element analysis is an RGB three-channel color image, which needs to be converted to a single channel first, and then the pixel value is normalized to the [0, 1] interval.

[0019] Further, in step S5, the sample data is selected based on the interpolation priority principle;

[0020] When selecting sample data, the computational workload should also be considered, and the specific method is as follows:

[0021] According to the results of numerical simulation and historical experience, whether the magnetic field change is linear or nonlinear is observed, and an appropriate sample quantity is selected; if it is linear, a small number of samples are selected for training to reduce the amount of operation; if it is nonlinear, the sample quantity is increased as much as possible while reducing the amount of operation to improve the accuracy of the model.

[0022] Further, the deep learning model in step S5 includes a convolutional neural network, U-net, U-net+residual, linknet, and a local attention mechanism.

[0023] Compared with the prior art, the electric equipment magnetic field prediction method has the following advantages:

[0024] 1. Although the working conditions of the electric equipment change, for the case of changing the working conditions, only the corresponding data in the input information matrix needs to be changed, and the magnetic field distribution map of the electric equipment under this working condition can be directly generated through the prediction of the deep learning model. Compared with other methods, the prediction accuracy is higher, the prediction workload is smaller, and it can be flexibly applied to magnetic field prediction under different working conditions.

[0025] 2. For the selection of design parameters under different working conditions, the application has specific instructions, that is, the magnetic field distribution of the electric equipment under different working conditions can be predicted, and two methods of interpolation and extrapolation can be used. If interpolation prediction is performed, two end values of the specified influencing factors need to be selected, and the intermediate data is determined by using a specific sampling method, such as systematic sampling. If extrapolation prediction is performed, one end value of the specified influencing factor needs to be selected, and the data in the expanded range is determined by using a specific sampling method, such as simple random sampling, systematic sampling, stratified sampling, etc.

[0026] 3. The application proposes that according to the results of numerical simulation and historical experience, whether the magnetic field change is linear or nonlinear is observed, and an appropriate sample quantity is selected. If it is linear, a small number of samples can be selected for training to reduce the amount of operation; if it is nonlinear, the sample quantity is increased as much as possible while reducing the amount of operation to the greatest extent to improve the accuracy of the model. BRIEF DESCRIPTION OF DRAWINGS

[0027] The accompanying drawings, which form a part of the present application, are used to provide a further understanding of the present application, and the illustrative embodiments thereof and their descriptions are used to explain the present application, and do not constitute an improper limitation of the present application. In the drawings:

[0028] Figure 1 The electric equipment performance prediction method based on the pre-trained model described in the embodiments of the present application is shown in the flowchart;

[0029] Figure 2 The convolutional neural network structure diagram described in the embodiments of the present application is shown in the structure diagram;

[0030] Figure 3 A U-net model structure diagram according to an embodiment of the present application;

[0031] Figure 4 A linknet model structure diagram according to an embodiment of the present application;

[0032] Figure 5 A partial structure diagram of a permanent magnet synchronous motor according to an embodiment of the present application;

[0033] Figure 6 A real magnetic field distribution diagram of a permanent magnet synchronous motor obtained by finite element analysis according to an embodiment of the present application;

[0034] Figure 7 A magnetic field distribution diagram predicted by a convolutional neural network according to an embodiment of the present application;

[0035] Figure 8 A magnetic field distribution diagram predicted by a U-net model according to an embodiment of the present application;

[0036] Figure 9 A magnetic field distribution diagram predicted by a linknet model according to an embodiment of the present application;

[0037] Figure 10 A magnetic field distribution diagram predicted by a linknet+attention model according to an embodiment of the present application. DETAILED DESCRIPTION

[0038] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0039] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application. In addition, the terms "first", "second" and the like are only for the purpose of description and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features limited by "first", "second" and the like can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise stated, the meaning of "a plurality of" is two or more.

[0040] In the description of the present application, it should be noted that unless otherwise expressly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0041] As Figure 1 shown, a kind of electric equipment magnetic field prediction method, S1, the finite element model of specific structure electric equipment is established, which considers the actual operation condition of electric equipment;

[0042] S2, according to the data of the working condition to be predicted, select several known working conditions for finite element analysis, and obtain the real magnetic field distribution diagram of the electric equipment under several known working conditions through finite element analysis.

[0043] S3, select several factors that have influence on the magnetic field distribution of electric equipment as input variables of deep learning model, and form input information matrix combined with the geometric structure of the electric equipment;

[0044] S4, the real magnetic field distribution diagram obtained by finite element analysis is normalized;

[0045] S5, the input information matrix of several working conditions of the real magnetic field distribution diagram obtained in S3 and the preprocessed real magnetic field distribution diagram in S4 are used as training sample data and test sample data, and the deep learning model is trained. The deep learning model can be selected according to the actual situation, and the optional model is, for example, convolutional neural network, U-net, U-net+residual, linknet and local attention mechanism. The test sample data is tested, and the mean square error is used as the standard of whether the model is successfully trained or not;

[0046] S6, for the electric equipment under other working conditions, as long as the corresponding data in the input information matrix is changed, the magnetic field distribution diagram of the electric equipment under the working condition can be directly generated by the prediction of the deep learning model.

[0047] In S1, the structural parameters of electric equipment need to be mastered, for example, the number of poles, the number of slots, the air gap, the inner diameter and the outer diameter of the stator and rotor core of the motor. According to Maxwell differential equation, the mathematical model of each region of electric equipment is obtained. The actual working conditions of electric equipment include working voltage, working current, torque, frequency, surrounding temperature, humidity, etc.

[0048] In S2, according to the variational principle of finite element method, the problem of solving motor magnetic field is transformed into the problem of solving the extreme value of energy functional, and then the real magnetic field distribution diagram is obtained by using electromagnetic field simulation software.

[0049] Step S2, when performing finite element analysis on selected known working conditions, needs to combine the relevant parameters of the known working conditions for finite element analysis. By selecting different working condition parameters, the magnetic field distribution of the electrical equipment under different working conditions can be predicted. When selecting parameters, interpolation and extrapolation methods can be used. If interpolation prediction is used, two end values of the specified influencing factors need to be selected, and the intermediate data is determined by using a specific sampling method, such as system sampling. If extrapolation prediction is used, one end value of the specified influencing factors needs to be selected, and the data in the expanded range is determined by using a specific sampling method, such as simple random sampling, system sampling, stratified sampling, etc. It should be emphasized that,

[0050] Since both ends have constraint values when interpolating, the interpolation prediction method has higher accuracy than extrapolation. Under the allowed conditions, interpolation prediction is selected. According to this principle, select which working conditions to perform finite element analysis on as training sample data for the neural network.

[0051] The input information matrix in S3 is composed of two parts. One is the structural parameters of the electrical equipment, mainly including the sizes and materials of each part, excitation, etc. The other part is the variables involved in the actual working conditions, which are consistent with the working condition factors considered in the finite element analysis in S1.

[0052] The normalization process in S4 is because most deep learning models learn through gradient descent. The real magnetic field cloud map obtained by finite element simulation analysis is an RGB three-channel color image, which needs to be converted to a single channel and normalized to the [0, 1] interval.

[0053] In S5, in addition to following the interpolation priority principle, the problem of computational workload needs to be considered, which puts forward requirements for the number of sample selection. The present invention proposes to observe whether the magnetic field changes linearly or nonlinearly according to the results of numerical simulation and historical experience, and select an appropriate number of samples. If it is linear, a small number of samples can be selected for training to reduce the amount of computation; if it is nonlinear, the number of samples is increased as much as possible while reducing the amount of computation to the greatest extent to improve the accuracy of the model. The deep learning model involved in S5 can be varied. The following are several common models.

[0054] For example, Figure 2As shown, a Convolutional Neural Network (CNN) is a type of feedforward neural network that incorporates convolutional computations and has a deep structure. Its structure consists of three layers: an input layer, hidden layers, and an output layer. The hidden layers include convolutional layers, pooling layers, and fully connected layers. The input layer of a CNN can handle multi-dimensional data. Because it uses gradient descent for learning, its input features need to be standardized; that is, the input data needs to be normalized in the channel or time / frequency dimension before being input into the CNN. The convolutional layer extracts features from the input data. It contains multiple convolutional kernels, and the elements within the kernels represent weight coefficients. After feature extraction by the convolutional layer, the output feature map is passed to the pooling layer for feature selection and information filtering. The fully connected layer performs non-linear combinations of the extracted features to produce the output. The layer above the output layer is the fully connected layer, which outputs classification results, etc.

[0055] like Figure 3 As shown, the U-net model is a neural network with an encoder-decoder structure. It is an improvement and extension of the fully convolutional neural network model. This model consists of five layers. The right-pointing arrow represents a 3×3 convolution, the narrow right-pointing arrow represents a 1×1 convolution, the down-pointing arrow represents a 2×2 pooling, the up-pointing arrow represents a 2×2 upsampling, and the black arrow indicates that the feature map of that layer is copied and cropped. The horizontal numbers represent the number of channels, and the vertical numbers represent the resolution. The learning process of this model is as follows: First, in the encoder stage, two convolution operations are performed on the input image to extract features, increasing the number of channels from 1 to 64. Then, four pooling operations are performed, followed by two 3×3 convolution operations and a correction of the linear unit activation function after each pooling operation, ultimately increasing the number of channels to 1024 and reducing the resolution to 32. After two 1×1 convolution operations, the resolution becomes 28. The decoder stage follows. Four upsampling operations are performed using a 2×2 convolutional kernel. After each upsampling, two 3×3 convolution operations and a modified linear unit activation function are applied, resulting in 64 channels and a resolution of 388. During each upsampling process, the transformed feature map is concatenated and fused with the feature map extracted from the corresponding encoder convolutional layer via skip connections to ensure that the number of channels is consistent with the encoder stage. Finally, a 1×1 convolution operation maps the 64-dimensional feature vector to the output layer, predicting each pixel to obtain a 2-channel feature map.

[0056] The U-net+residual model is composed of a U-net model and a residual convolutional neural network, and is also an encoder-decoder structure. The advantage of the model is that it can use inter-layer skip connection to make the gradient directly return to the previous layer in the back propagation process, without increasing the operation complexity, thereby avoiding the performance degradation problem caused by deepening the level. In addition, a Dropout layer is added to the model to prevent network overfitting. The skip connection can effectively extract and fuse features of different levels, repair the lost image detail information in the pooling operation, and make the magnetic field prediction result more accurate.

[0057] As shown in Figure 4 The linknet structure is a fully convolutional neural network, that is, a convolutional layer is used instead of a fully connected layer at the end of the network, and an end-to-end model with an image as input and an image as output is obtained by training. It is different from the traditional convolutional neural network with an image as input and a class label as output. The advantage of this network is that it uses skip connection to cascade the encoder and the decoder, which can fully extract and fuse features of different levels. The improved linknet model structure is a convolutional neural network formed by combining attention mechanism and linknet. This network adds attention mechanism after the skip connection, which can focus attention on the target area to more accurately capture features, effectively solve the problem of loss of image detail information in the decoder stage, and the prediction result is more accurate

[0058] Taking a permanent magnet synchronous motor as an example, the scheme of the application is further described.

[0059] 1. First, master the structural parameters of the electrical equipment, and select the factors affecting the magnetic field distribution. Taking a permanent magnet synchronous motor as an example, as shown in Figure 5 The permanent magnet width is 31 mm, the permanent magnet thickness is 5 mm, the stator tooth height is 29 mm, and the stator tooth width is 5 mm. The working conditions are selected as current 220 A, frequency 50 Hz, and temperature 30°C. According to the Maxwell differential equation set, the mathematical model of each region of the permanent magnet synchronous motor can be obtained, and the regions can be divided into stator winding, air gap and permanent magnet. The magnetic field is solved by solving the extreme value problem of the energy functional using the variational principle of the finite element method.

[0060] 2. Use, but not limited to, Ansys simulation software to simulate the magnetic field of the electrical equipment. Set appropriate boundary conditions and mesh division modes to obtain the real magnetic field distribution diagram of the electrical equipment, as shown in Figure 6As shown, it is the real magnetic field distribution diagram of permanent magnet synchronous motor obtained by finite element analysis. For the selection of design parameters under different working conditions, the application has specific instructions, that is, the magnetic field distribution of electrical equipment under different working conditions can be predicted, and two methods of interpolation and extrapolation can be used. If interpolation prediction is made, two end values of the specified influencing factors need to be selected, and the intermediate data is determined by using a specific sampling method, such as system sampling; if extrapolation prediction is made, one end value of the specified influencing factor needs to be selected, and the data of the extended range is determined by using a specific sampling method, such as simple random sampling, system sampling, stratified sampling, etc. It needs to be emphasized that since there are constraint values at both ends when interpolation prediction is made, the prediction method of interpolation has higher accuracy compared with extrapolation, and interpolation prediction is selected under the permitted conditions. According to this principle, select which working conditions to perform finite element analysis, and use the training sample data as input to the neural network. For example, if the magnetic field of a permanent magnet synchronous motor under 200V voltage is to be predicted, the actual magnetic field distribution of the motor under 100V and 300V can be selected as the sample during model training, and the magnetic field prediction made by the interpolation method is more accurate.

[0061] 3. Convert the structure of the electrical equipment into an input structure matrix, convert the corresponding working condition into an input working condition matrix, and combine the two into an input information matrix. Normalize the real magnetic field distribution diagram obtained by finite element simulation analysis to obtain a magnetic field distribution diagram normalization matrix.

[0062] 4. Input the input information matrix and the magnetic field distribution diagram normalization matrix into the deep learning model. The selection of the number of samples should be based on the results of numerical simulation and historical experience to observe whether the magnetic field changes linearly or nonlinearly. If it is linear, a small number of samples can be selected for training to reduce the amount of calculation; if it is nonlinear, the number of samples is increased as much as possible while reducing the amount of calculation to the greatest extent to improve the accuracy of the model. Taking a permanent magnet synchronous motor as an example, 2300 training sample data and 200 test data can be selected, and a U-net+residual model can be selected, which can accurately predict the magnetic field distribution diagram of the electrical equipment. Take the mean square error as the standard for whether the model is successfully trained. If the test mean square error is less than the set value, it is considered that the training is successful, and if it is greater than the set value, the number of training samples is increased and the training is continued.

[0063] As Figures 7-10 shown, the effect diagram of the permanent magnet synchronous motor under different deep learning models for magnetic field distribution prediction.

[0064] 5. Predict the magnetic field distribution diagram of the electrical equipment under other working conditions. Change the input information matrix according to the working condition to be predicted, and input it into the trained deep learning model to predict the magnetic field distribution diagram of the electrical equipment.

[0065] The above merely provides the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method of magnetic field prediction for electrical equipment, characterized in that, It comprises the following steps: S1, combining the actual operation condition of the electrical equipment, a finite element model of the electrical equipment is established; S2, according to the working condition data to be predicted, several known working conditions are selected for finite element analysis, and the real magnetic field distribution diagram of the electrical equipment under the corresponding several known working conditions is obtained; S3, several factors affecting the magnetic field distribution of the electrical equipment are selected as input variables of the deep learning model, and the input information matrix is formed in combination with the geometric structure of the electrical equipment; S4, the real magnetic field distribution diagram obtained in step S2 is normalized; S5, the input information matrix obtained in step S3 and the real magnetic field distribution diagram processed in step S4 are used as training sample data and test sample data, the deep learning model is trained, and the magnetic field of the electrical equipment is predicted based on the trained deep learning model; Wherein, The input information matrix formed in step S3 includes the structural parameters of the electrical equipment and the variables involved in the actual operation condition; Wherein, the structural parameters of the electrical equipment include the size and material of each part, excitation; In step S5, sample data is selected based on the interpolation priority principle; When selecting sample data, the operation workload should also be considered, and the specific method is as follows: According to the results of numerical simulation and historical experience, observe whether the magnetic field changes linearly or nonlinearly, and select an appropriate number of samples; if it is linear, a small number of samples are selected for training to reduce the amount of calculation; if it is nonlinear, the number of samples is increased as much as possible while reducing the operation workload to improve the accuracy of the model; In step S1, the actual working condition of the electrical equipment includes working voltage, working current, torque, frequency, surrounding temperature and humidity; In step S2, according to the variational principle of finite element method, the problem of solving motor magnetic field is transformed into solving the extreme value problem of energy functional, and then electromagnetic field simulation software is used to obtain the real magnetic field distribution diagram; In step S2, when selecting sample data under different working conditions, the selection method includes interpolation and extrapolation; If interpolation prediction is made, two end values of the specified influencing factors need to be selected, and the intermediate data is determined by using a specific sampling method; if extrapolation prediction is made, one end value of the specified influencing factors needs to be selected, and the data in the extended range is determined by using a specific sampling method; Because there are constraint values at both ends when interpolation prediction is made, the prediction method of interpolation has higher accuracy than that of extrapolation, so interpolation prediction is selected under the allowable conditions, and according to this principle, it is determined which working conditions are selected for finite element analysis as training sample data to be input into the neural network.

2. The method of claim 1, wherein, In step S4, the real magnetic field distribution diagram obtained by finite element analysis is an RGB three-channel color image, which needs to be converted to a single channel first, and then the pixel value is normalized to the [0, 1] interval.

3. The method of claim 1, wherein: The deep learning model in step S5 includes convolutional neural network, U-net, U-net+residual, linknet and local attention mechanism.

Images