Iron loss modeling method for mechanism-data hybrid driving of permanent magnet synchronous motor

Abstract

The invention belongs to the technical field of motor iron loss modeling, and particularly relates to a permanent magnet synchronous motor mechanism-data hybrid driven iron loss modeling method, which comprises the following steps of: generating a PWM (Pulse Width Modulation) induction magnetic flux density curve, a temperature curve and a mechanical stress distribution curve under different load conditions and frequencies; obtaining an improved iron loss analysis model considering the space harmonic wave and the PWM harmonic wave at the same time; a multi-physical mechanism coupling iron loss analysis model considering mechanical stress, temperature, space harmonic waves, load current and frequency changes at the same time is obtained, and iron loss data under different working conditions are obtained through calculation by means of the model; based on the trained CNN, multi-layer feature map output is obtained for input features, and various feature mappings extracted by a convolutional layer are integrated in a full connection layer; and converting an integration result of the full connection layer into probability distribution through Softmax logistic regression, and performing classification and final output. The method is high in prediction precision and can be suitable for various application scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of motor iron loss modeling, and specifically relates to a permanent magnet synchronous motor mechanism-data hybrid driven iron loss modeling method. Background Art

[0002] The accurate calculation of iron loss in permanent magnet synchronous motors remains challenging due to the interaction between different fields such as electromagnetics, magnetics, and thermal stress coupled dynamics. Before the rise of big data, methods such as classical loss models, hysteresis models, finite element models, frequency domain models, temperature dependent models, and empirical models were developed to calculate iron loss. However, due to the complexity of the iron loss mechanism involving multiple disciplines, each technique has its advantages and limitations in iron loss modeling.

[0003] Classical loss models are suitable for sinusoidal excitation, but lack accuracy for complex waveforms and non-sinusoidal conditions. Although hysteresis models and temperature-dependent modeling methods can simulate material properties, they increase the complexity of determining model parameters. Finite element models and frequency domain models can effectively alleviate these problems, but they require a lot of computing resources and time.

[0004] In recent years, machine learning technology has made significant progress. The flexibility, scalability, and deep abstract representation capabilities of neural networks enable them to be seamlessly applied in fields with sufficient data, inspiring researchers to solve scientific problems such as physical system modeling. Although statistical machine learning models have made significant progress in reducing modeling complexity and improving prediction accuracy with the help of big data, there are still many limitations when deploying these purely data-driven models in practical applications. For example, predictions may lack robustness, lack interpretability, and may violate physical constraints or common sense. Therefore, it is an urgent problem to find a modeling method that considers multi-physics field effects, has high prediction accuracy, can reduce modeling complexity, and has a wide range of practical applications. Summary of the invention

[0005] In view of the above deficiencies in the prior art, the object of the present invention is to provide a permanent magnet synchronous motor mechanism-data hybrid driven iron loss modeling method, which has high prediction accuracy and can be applied to various application scenarios.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is a permanent magnet synchronous motor mechanism-data hybrid driven iron loss modeling method, comprising the following steps: S1. Using Simplorer and Ansys software, generate three curves: pulse width modulation carrier PWM induced magnetic flux density curve, temperature curve and mechanical stress distribution curve under different load conditions and frequencies; S2. The finite element method is used to convert the magnetic flux density into a trapezoidal waveform to obtain an iron loss analysis model that takes into account spatial harmonics. The hysteresis loss is adjusted by defining a correction coefficient, and the Bertotti method is used to replace the eddy current loss with the cumulative effect of each component caused by harmonics to obtain an iron loss analysis model that takes into account PWM harmonics. The iron loss analysis model that takes into account spatial harmonics and the iron loss analysis model that takes into account PWM harmonics are integrated to obtain an improved iron loss analysis model that takes into account both spatial harmonics and PWM harmonics. S3. Based on the improved iron loss analysis model, the Pearson correlation coefficient is used to analyze the interaction between the iron loss coefficient and multiple physical factors, so as to calibrate the iron loss coefficient and obtain the calibrated iron loss analysis model; the iron loss coefficient is further defined as a function of the operating temperature and mechanical stress, and a multi-physical mechanism coupled iron loss analysis model that simultaneously considers mechanical stress, temperature, spatial harmonics, load current and frequency changes is obtained, and the iron loss data under different working conditions are calculated using the model; S4, based on the iron loss data under different working conditions, select features as the input of the convolutional neural network CNN; S5. Train CNN and set the loss function to quantify the error between the predicted value and the true value of iron consumption during forward reasoning. Use the error to guide the CNN back-propagation algorithm and gradient adjustment. Then use the Adam optimizer to update the network parameters including the connection weights and biases of the convolutional layer and the fully connected layer, and refine the mapping relationship between the CNN input and output. S6. Based on the trained CNN, for the input features, multiple one-dimensional convolution kernels of two convolutional layers are used to achieve multi-feature fusion, and the multi-layer feature map output is obtained through the ReLU activation function, and various feature maps extracted by the convolutional layer are integrated in the fully connected layer; S7. Convert the integrated results of the fully connected layer into probability distribution through Softmax logistic regression for classification and final output.

[0007] As a preferred embodiment of the present invention, in S2, the process of constructing an iron loss analysis model taking into account spatial harmonics is as follows: The magnetic flux density is converted into a trapezoidal waveform by the finite element method, and the iron loss density in the tooth and yoke regions is expressed as: (1); In the formula, , are the iron loss densities in the tooth and yoke regions, respectively; , are the empirical coefficients of the magnetic flux density amplitude in the tooth and yoke regions, respectively; , are the magnetic flux density amplitudes in the tooth and yoke regions, respectively; , T are the motor angular frequency and excitation period respectively; , The number of phases and slots per pole and phase respectively; are the derived fitting coefficients; , are the hysteresis loss coefficient and eddy current loss coefficient respectively; β is the magnet width coverage, expressed as: (2); In the formula, p is the number of poles; is the effective width of the magnet; r is the outer radius of the rotor; Determine the total iron losses in the stator core considering space harmonics , expressed as: (3); In the formula, , are the volumes of the tooth and yoke respectively.

[0008] As a preferred solution of the present invention, in S2, the process of constructing an iron loss analysis model considering PWM harmonics is as follows: Define a correction factor , considering the cumulative effect of each component caused by harmonics, the hysteresis loss and eddy current loss caused by PWM harmonics are expressed as: (4); In the formula, , They are hysteresis loss and eddy current loss caused by nth PWM harmonics respectively; is the hysteresis loss density; represents the eddy current loss caused by the nth harmonic; where, It is expressed as: (5); In the formula, is a constant coefficient between 0.6 and 0.7; is the magnetic flux density amplitude; N is the number of samples; is the fluctuation of the motor magnetic flux density of the ath sample; It is expressed as: (6); In the formula, f, They are the fundamental wave and nth harmonic frequency of the PWM waveform respectively; is the number of winding turns; S is the cross-sectional area of ​​the core; is the amplitude of the nth harmonic voltage; is the nth harmonic magnetic flux density amplitude, expressed as: (7).

[0009] As a preferred solution of the present invention, in S2, the improved iron loss analysis model that integrates both space harmonics and PWM harmonics is expressed as: (8); In the formula, represents the total iron loss in the stator core considering both space harmonics and PWM harmonics; , are the correction coefficients for the tooth and yoke regions, respectively; , are the hysteresis loss densities in the tooth and yoke regions, respectively; is the eddy current loss density in the tooth area; , are the longitudinal flux and normal component of the eddy current loss density in the yoke area respectively; i represents the order of the harmonic.

[0010] As a preferred embodiment of the present invention, in S3, the iron loss analysis model calibrated by the Pearson correlation coefficient is expressed as: (9); In the formula, is obtained by fitting the iron loss curve and the mechanical stress The associated hysteresis loss coefficient; is obtained by fitting the iron loss curve with the operating temperature The associated eddy current loss coefficient.

[0011] As a preferred embodiment of the present invention, in S3, the multi-physical mechanism coupled iron loss analysis model is expressed as: (10); In the formula, Represents the total iron loss in the stator core after the multi-physics mechanism is coupled; For consideration and The hysteresis loss coefficient; For consideration and The eddy current loss coefficient.

[0012] As a preferred embodiment of the present invention, in the S4, the selected characteristics are six key characteristics: load current, operating temperature, mechanical stress, mechanical speed, tooth magnetic flux density and yoke magnetic flux density.

[0013] As a preferred solution of the present invention, in S5, the loss function adopts the root mean square error RMES.

[0014] As a preferred embodiment of the present invention, in the S6, the CNN network includes an input layer, a convolution layer, an activation function, a pooling layer, a fully connected layer and an output layer, wherein there are two convolution layers, namely convolution layer 1 and convolution layer 2, the number of convolution kernels, kernel size and convolution stride of convolution layer 1 are 16, 2 and 2 respectively, and the number of convolution kernels, kernel size and convolution stride of convolution layer 2 are 32, 2 and 1 respectively.

[0015] As a preferred solution of the present invention, in S6, the multi-layer feature map output is expressed as: (11); In the formula, is the output of the jth convolution kernel of the hth layer; It is the qth feature map output by the h-1th layer; is the number of feature maps; Represents the convolution operation; is the jth convolution kernel of the hth layer; Represents the bias of the jth convolution kernel of the hth layer; is the ReLU activation function, expressed as: (12); In the formula, x represents the linear combination of the convolution kernel output; The feature map of the convolutional layer is expressed as: (13); In the formula, is the output of the hth layer; It is the feature map output by the h-1th layer; is the weight matrix of the hth layer; Represents the bias of the hth layer.

[0016] The algorithm involved in the present invention can be executed by an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the above algorithm calculation is implemented by executing the software by the processor.

[0017] The beneficial effects of the present invention are: The present invention predicts iron loss characteristics adapted to real scenarios by adopting a mechanism-data fusion driven iron loss modeling method that considers multi-physical field effects. First, a multi-physical mechanism coupled iron loss analysis model that simultaneously considers mechanical stress, temperature, harmonics, load current and frequency changes is used to predict a large amount of loss data under different working conditions; secondly, the predicted loss data is used to pre-train a convolutional neural network (CNN) so that it can grasp the basic dynamics of iron loss; finally, the pre-trained convolutional neural network is adjusted by the loss function, and a limited set of experimental data is selected for deep learning, and the complex characteristics of the iron loss process are extracted from a large amount of loss data; accurate prediction of iron loss characteristics in real scenarios is achieved, with high robustness and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 Schematic diagram of the Pearson correlation index of the iron loss coefficient in an embodiment of the present invention; Figure 2 (a) is a schematic diagram of the correlation between the iron loss coefficient and the operating temperature; Figure 2 (b) is a schematic diagram of the correlation between the iron loss coefficient and the mechanical stress; Figure 3 It is a diagram of offline training results of the CNN network during the verification process of the present invention; Figure 4 is a comparison chart of actual iron loss and model prediction value under different speed and load conditions during the verification process of the present invention; Figure 5 It is an error analysis diagram of the prediction results during the verification process of the present invention; Figure 5 (a) in the figure is the normalized processing diagram of the prediction residual; Figure 5 (b) in the figure is the residual kernel density estimation diagram; Figure 5 (c) is an annotated residual box plot; Figure 6 The comparison results of predicted iron loss under different current inputs during the verification process of the present invention are shown in Fig. Figure 1 ; Figure 6 (a) is the comparison result diagram under 32A input current; Figure 6 (b) is the comparison result diagram under 90A input current; Figure 6 (c) is the comparison result diagram under 180A input current; Figure 7 The comparison results of predicted iron loss under different current inputs during the verification process of the present invention are shown in Fig. Figure 2 ; Figure 7 (a) is the comparison result diagram under 285A input current; Figure 7 (b) is the comparison result diagram under 400A input current; Figure 7(c) is the comparison result diagram under 490A input current; Figure 8 is the comparison error of the predicted iron loss under different current inputs during the verification process of the present invention Figure 1 ; Figure 8 (a) is the comparative error diagram under 32A input current; Figure 8 (b) is the comparative error diagram under 90A input current; Figure 8 (c) is the comparative error diagram under 180A input current; Fig. 9 is the comparison error of the predicted iron loss under different current inputs during the verification process of the present invention Figure 2 ; Fig. 9 (a) is the comparative error diagram at 285A input current; Fig. 9 (b) is the comparative error diagram under 400A input current; Fig. 9 (c) in Figure 3 is the comparative error diagram at 490A input current. DETAILED DESCRIPTION

[0019] The embodiments of the present invention are further described below in conjunction with the accompanying drawings: like Figure 1 As shown, a permanent magnet synchronous motor mechanism-data hybrid driven iron loss modeling method includes the following steps: S1. Using Simplorer and Ansys software, generate three curves: pulse width modulation carrier PWM induced magnetic flux density curve, temperature curve and mechanical stress distribution curve under different load conditions and frequencies; S2. The finite element method is used to convert the magnetic flux density into a trapezoidal waveform to obtain an iron loss analysis model that takes into account spatial harmonics. The hysteresis loss is adjusted by defining a correction coefficient, and the Bertotti method is used to replace the eddy current loss with the cumulative effect of each component caused by harmonics to obtain an iron loss analysis model that takes into account PWM harmonics. The iron loss analysis model that takes into account spatial harmonics and the iron loss analysis model that takes into account PWM harmonics are integrated to obtain an improved iron loss analysis model that takes into account both spatial harmonics and PWM harmonics. S3. Based on the improved iron loss analysis model, the Pearson correlation coefficient is used to analyze the interaction between the iron loss coefficient and multiple physical factors, so as to calibrate the iron loss coefficient and obtain the calibrated iron loss analysis model; the iron loss coefficient is further defined as a function of the operating temperature and mechanical stress, and a multi-physical mechanism coupled iron loss analysis model that simultaneously considers mechanical stress, temperature, spatial harmonics, load current and frequency changes is obtained, and the iron loss data under different working conditions are calculated using the model; S4, based on the iron loss data under different working conditions, select features as the input of the convolutional neural network CNN; S5. Train the CNN and set the loss function to quantify the error between the predicted value and the true value of iron consumption during forward reasoning. Use the error to guide the CNN back-propagation algorithm and gradient adjustment. Then use the Adam optimizer (learning rate set to 0.001) to update the network parameters including the connection weights and biases of the convolutional layer and the fully connected layer, and refine the mapping relationship between the CNN input and output. S6. Based on the trained CNN, for the input features, multiple one-dimensional convolution kernels of two convolutional layers are used to achieve multi-feature fusion, and the multi-layer feature map output is obtained through the ReLU activation function, and various feature maps extracted by the convolutional layer are integrated in the fully connected layer; S7. Convert the integrated results of the fully connected layer into probability distribution through Softmax logistic regression for classification and final output.

[0020] Simplorer and Ansys software are two simulation tools widely used in engineering and scientific research. In this embodiment, they are used to generate PWM induced flux density curves, temperature curves, and mechanical stress distribution curves of permanent magnet synchronous motors under different load conditions and frequencies. Pulse width modulation carrier PWM, also known as pulse width modulation (Pulse Width Modulation), is a modulation technology widely used in power electronics and communications. It controls the average value of the signal by changing the duty cycle of the pulse signal (that is, the ratio of the duration of the pulse to the total cycle time). The Bertotti method is an analytical method for calculating eddy current losses in motors. It is mainly used to evaluate eddy current losses in conductors caused by alternating magnetic fields, especially in the core loss analysis of electromagnetic equipment such as transformers and motors.

[0021] In the study of neural networks, especially in multi-classification problems, Softmax logistic regression usually refers to the use of the Softmax function in the last layer of the network to convert the output into a probability distribution. It is a combination of the Softmax function and the output layer of the neural network for classification tasks. Specifically, in this embodiment, the Softmax function is used in the output layer of the neural network to perform probability prediction of multiple classifications, and the difference between the prediction and the true label can be measured by the cross entropy loss.

[0022] In S2, the process of constructing the iron loss analysis model considering spatial harmonics is: The magnetic flux density is converted into a trapezoidal waveform by the finite element method, and the iron loss density in the tooth and yoke regions is expressed as: (1); In the formula, , are the iron loss densities in the tooth and yoke regions, respectively; , are the empirical coefficients of the magnetic flux density amplitude in the tooth and yoke regions, respectively; , are the magnetic flux density amplitudes in the tooth and yoke regions, respectively; , T are the motor angular frequency and excitation period respectively; , The number of phases and slots per pole and phase respectively; are the derived fitting coefficients; , are the hysteresis loss coefficient and eddy current loss coefficient respectively; β is the magnet width coverage, expressed as: (2); In the formula, p is the number of poles; is the effective width of the magnet; r is the outer radius of the rotor; Determine the total iron losses in the stator core considering space harmonics , expressed as: (3); In the formula, , are the volumes of the tooth part and the yoke part respectively.

[0023] The excitation cycle of the motor refers to a complete cycle in which the excitation winding (or rotor winding for synchronous motors) receives the excitation current when the motor is operating normally. The teeth are the protruding parts of the motor stator core, which interact with the magnetic poles of the rotor to generate the torque of the motor. The yoke is the area in the motor stator core that connects the teeth, forming a closed loop of the motor magnetic circuit.

[0024] The process of constructing an iron loss analysis model considering PWM harmonics is as follows: Define a correction factor , considering the cumulative effect of each component caused by harmonics, the hysteresis loss and eddy current loss caused by PWM harmonics are expressed as: (4); In the formula, , They are hysteresis loss and eddy current loss caused by nth PWM harmonics respectively; is the hysteresis loss density; represents the eddy current loss caused by the nth harmonic; where, It is expressed as: (5); In the formula, is a constant coefficient between 0.6 and 0.7; is the magnetic flux density amplitude; N is the number of samples; is the fluctuation of the motor magnetic flux density of the ath sample; It is expressed as: (6); In the formula, f, They are the fundamental wave and nth harmonic frequency of the PWM waveform respectively; is the number of winding turns; S is the cross-sectional area of ​​the core; is the amplitude of the nth harmonic voltage; is the nth harmonic magnetic flux density amplitude, expressed as: (7).

[0025] The improved iron loss analysis model that takes into account both space harmonics and PWM harmonics is expressed as: (8); In the formula, represents the total iron loss in the stator core considering both space harmonics and PWM harmonics; , are the correction coefficients for the tooth and yoke regions, respectively; , are the hysteresis loss densities in the tooth and yoke regions, respectively; is the eddy current loss density in the tooth area; , are the longitudinal flux and normal component of the eddy current loss density in the yoke area respectively; i represents the order of the harmonic.

[0026] In S3, the iron loss analysis model calibrated by the Pearson correlation coefficient is expressed as: (9); In the formula, is obtained by fitting the iron loss curve and the mechanical stress The associated hysteresis loss coefficient; is obtained by fitting the iron loss curve with the operating temperature The associated eddy current loss coefficient.

[0027] The multi-physics mechanism coupled iron loss analysis model is expressed as: (10); In the formula, Represents the total iron loss in the stator core after the multi-physics mechanism is coupled; For consideration and The hysteresis loss coefficient; For consideration and The eddy current loss coefficient.

[0028] The Pearson correlation coefficient quantifies the strength and direction of the linear relationship between the coefficient and the influencing factor, thus facilitating a more efficient and insightful fitting process. The magnitude of its value approaching 1 indicates that there is a robust correlation between the input and output variables. Using finite element method data, based on multifaceted physical data over a wide range of operating speeds from 500 rpm to 5500 rpm, the Pearson correlation index linking the iron loss coefficients (including eddy current loss coefficient and hysteresis loss coefficient) is derived, as shown in Figure 2 As shown, it can be seen that the hysteresis loss coefficient has an obvious correlation with the stress level, while the eddy current loss coefficient is significantly affected by temperature changes.

[0029] In S4, the selected features are load current, operating temperature, mechanical stress, mechanical speed, tooth magnetic flux density and yoke magnetic flux density. In S5, the loss function uses the root mean square error RMES.

[0030] In S6, the CNN network includes an input layer, a convolutional layer, an activation function, a pooling layer, a fully connected layer, and an output layer. There are two convolutional layers, namely convolutional layer 1 and convolutional layer 2. The number of convolutional kernels, kernel size, and convolutional stride of convolutional layer 1 are 16, 2, and 2, respectively, and the number of convolutional kernels, kernel size, and convolutional stride of convolutional layer 2 are 32, 2, and 1, respectively.

[0031] The multi-layer feature map output is expressed as: (11); In the formula, is the output of the jth convolution kernel of the hth layer; It is the qth feature map output by the h-1th layer; is the number of feature maps; Represents the convolution operation; is the jth convolution kernel of the hth layer; Represents the bias of the jth convolution kernel of the hth layer; is the ReLU activation function, expressed as: (12); In the formula, x represents the linear combination of the convolution kernel output; The feature map of the convolutional layer is expressed as: (13); In the formula, is the output of the hth layer; It is the feature map output by the h-1th layer; is the weight matrix of the hth layer; Represents the bias of the hth layer.

[0032] The verification process is: MATLAB was used as a simulation tool for the experiment, and Ubuntu 16.04 with Python 3.6 and TensorFlow 1.5 were used as the deep learning framework. The computer used in the experiment was configured with a 3GHz Core i7 1185G7 processor and 16GB of memory. The joint simulation of Simplorer and Ansys was run on a Linux-based high-performance computing cluster with 56 CPU cores and up to 1.5TB of memory per node. This setting is well suited for processing large bioinformatics datasets, complex physical simulations, and multidimensional datasets.

[0033] The training data was obtained offline. Considering the importance of multidisciplinary coupled modeling, 6 input variables were carefully selected from the thresholds of their respective fields. These variables are essential for the subsequent iron loss simulation calculation of the training data. Through simulation and experiments, a total of 5000 sets of training data were obtained, each containing 6 inputs and 1 output. Since the training set drives the learning process of the CNN model, and the test data is crucial to evaluate the prediction accuracy of the model in actual operation, in order to evaluate the prediction performance of the iron loss model, an additional 700 customized data points are required to measure the 14% proportion of the test set. Incorporating the actual measured data into the test set significantly improves the authenticity, generalization ability and robustness of the model. Considering the slight influence of stator temperature and motor stress on the speed change rate, the main changes in the measured data are mainly determined by the speed and load current, where each variable contributes about 24 instances to the test data set.

[0034] The experimental data was measured using an IPMSM experimental prototype (IPMSM is a special type of permanent magnet synchronous motor PMSM, whose permanent magnets are embedded inside the rotor) in conjunction with traditional sinusoidal PWM and field-oriented controllers. The key technical indicators of the IPMSM experimental prototype are shown in Table 1. The experimental platform is constructed by a converter, a load motor, an electric control cabinet, an oscilloscope, a PWM control circuit, a power analyzer, a DC power supply, a dynamometer, a dynamic torque / speed sensor, and a computer. The DC power supply is used to provide the required DC power to the test motor, the computer is used to write the control program, the electric control cabinet is used to simulate different load conditions, and the oscilloscope is used to record data. Other experimental prototypes and other known experimental platforms with similar functions can also be used for verification.

[0035] Table 1 Key technical indicators of the experimental prototype project value project value Number of poles 8 Core Materials 35WW360 Number of slots 48 Core length 108mm Number of stages 3 Stator outer diameter 196mm Magnet per pole 2 Stator inner diameter 135mm Rated Power 20kW Rotor outer diameter 134mm Rated speed 3600rpm Air gap length 0.5mm Maximum Power 40kW Tooth width 6.35mm Maximum speed 5500rpm Magnet Width 19.5mm Rated torque 53Nm V-shaped angle 145° Maximum Torque 180Nm Phase resistance 0.0052Ω The offline training results of the CNN network (except the traditional CNN, all other CNNs refer to the CNN in this embodiment) using training and test data are shown in Figure 3As shown in Figure 2, it can be seen that the loss function, i.e., the root mean square error of the output variable, stops decreasing after about 1700 iterations. During this period, each parameter of the CNN network reaches its optimal value through the back propagation of the training parameters and the use of gradient descent.

[0036] The actual iron loss under different speed and load conditions is obtained through the experimental platform, and compared with the predicted value of the proposed CNN model. Figure 4 As shown in the figure, the blue "*" symbol represents the actual iron loss of the motor, while the magenta "." symbol represents the iron loss value predicted by the proposed CNN model. By comparing the predicted value with the actual iron loss value, the scatter plot shows that the two types of symbols have good overlap.

[0037] The error analysis results are summarized as follows: Figure 5 As shown. Figure 5 As shown in (a) of Figure 2, the normalized residuals between the predicted and actual iron losses are small under various operating conditions. Figure 5 As can be seen in (b), the kurtosis of the histogram and kernel density is very high, about 0.22, and the residual is approximately 0, indicating that the residual is relatively concentrated. Although the kernel density shows slight fluctuations, the residual distribution is closely consistent with the normal distribution, verifying the stable prediction ability of the proposed iron loss model. Figure 5 The box plot of (c) shows that the maximum and minimum values ​​of the residual are 5.59% and -5.29% respectively, while the median is around 0.13%, the 25th percentile is -1.22%, and the 75th percentile is 1.55%. The results show that the proposed hybrid mechanism-data driven model (hybrid model in the figure) can effectively predict the iron loss of the experimental prototype with satisfactory accuracy and stability.

[0038] To further verify the superiority of the proposed method, the measured iron loss of the IPMSM prototype was compared with the predicted iron loss of the proposed hybrid data-driven model, the traditional CNN model and the advanced analytical iron loss model in the literature (referred to as the analytical model in the figures and tables) under six different current inputs.

[0039] The comparison results are as follows Figure 6 and Figure 7 As shown, the comparison error is Figure 8 and Fig. 9 shown. Figures 6 to 9 The data analysis is shown in Table 2, where EM m and Eam (m=32, 90, 180, 285, 400 and 490) represent the maximum and absolute average errors of the iron loss relative to the measured data at phase currents of 32A, 90A, 180A, 285A, 400A and 588A, respectively.

[0040] Table 2 Figures 6 to 9 Data analysis project The present invention proposes Traditional CNN Analytical Model project The present invention proposes Traditional CNN Analytical Model <![CDATA[EM 32 ]]> -3.92% -7.12% -3.35% <![CDATA[EM 285 ]]> 3.78% 6.45% -4.36% <![CDATA[EA 32 ]]> 1.17% 3.84% 2.67% <![CDATA[EA 285 ]]> 2.11% 3.39% 2.85% <![CDATA[EM 90 ]]> 3.21% 4.95% -4.38% <![CDATA[EM 400 ]]> -4.98% 6.18% -4.86% <![CDATA[EA 90 ]]> 1.54% 2.50% 2.08% <![CDATA[EA 400 ]]> 1.55% 3.86% 2.75% <![CDATA[EM 180 ]]> 3.76% 2.76% -5.46% <![CDATA[EM 490 ]]> 3.95% 6.95% -5.43% <![CDATA[EA 180 ]]> 1.52% 2.91% 2.90% <![CDATA[EA 490 ]]> 1.83% 3.21% 3.84% like Figure 6 and Figure 7 As shown in the figure, the proposed hybrid mechanism-data driven model shows excellent performance in predicting the iron loss of the IPMSM prototype. Under different operating speeds and different current inputs, compared with the traditional CNN model and the improved analytical iron loss model, the predicted iron loss curve is closest to the measured value, and has better accuracy and stability.

[0041] like Figure 8 and Fig. 9 As shown in the figure, under all 6 different input phase currents, the maximum errors of the iron loss model developed using the traditional CNN and the improved analytical method are -7.12% and -5.46%, respectively, while the maximum error using the proposed hybrid data-driven method is only -4.98% at an input current of 400A. The maximum errors of the remaining 5 input currents are all less than 4%. In addition, the prediction accuracy and stability of the proposed method are also better than other models. Under different operating conditions, the absolute average error of the iron loss predicted by the proposed model is within 2.11%, which is only about half of the traditional CNN and 70% of the improved analytical method in the literature. All the results well verify the satisfactory performance of the proposed hybrid mechanism-data-driven iron loss model under various operating conditions.

Claims

1. A permanent magnet synchronous motor mechanism-data hybrid driven iron loss modeling method, characterized in that The following steps are involved: S1. Using Simplorer and Ansys software, generate three curves: pulse width modulation carrier PWM induced magnetic flux density curve, temperature curve and mechanical stress distribution curve under different load conditions and frequencies; S2. The finite element method is used to convert the magnetic flux density into a trapezoidal waveform to obtain an iron loss analysis model that takes into account spatial harmonics. The hysteresis loss is adjusted by defining a correction coefficient, and the Bertotti method is used to replace the eddy current loss with the cumulative effect of each component caused by harmonics to obtain an iron loss analysis model that takes into account PWM harmonics. The iron loss analysis model that takes into account spatial harmonics and the iron loss analysis model that takes into account PWM harmonics are integrated to obtain an improved iron loss analysis model that takes into account both spatial harmonics and PWM harmonics. S3. Based on the improved iron loss analysis model, the Pearson correlation coefficient is used to analyze the interaction between the iron loss coefficient and multiple physical factors, so as to calibrate the iron loss coefficient and obtain the calibrated iron loss analysis model; the iron loss coefficient is further defined as a function of the operating temperature and mechanical stress, and a multi-physical mechanism coupled iron loss analysis model that simultaneously considers mechanical stress, temperature, spatial harmonics, load current and frequency changes is obtained, and the iron loss data under different working conditions are calculated using the model; S4, based on the iron loss data under different working conditions, select features as the input of the convolutional neural network CNN; S5. Train CNN and set the loss function to quantify the error between the predicted value and the true value of iron consumption during forward reasoning. Use the error to guide the CNN back-propagation algorithm and gradient adjustment. Then use the Adam optimizer to update the network parameters including the connection weights and biases of the convolutional layer and the fully connected layer, and refine the mapping relationship between the CNN input and output. S6. Based on the trained CNN, for the input features, multiple one-dimensional convolution kernels of two convolutional layers are used to achieve multi-feature fusion, and the multi-layer feature map output is obtained through the ReLU activation function, and various feature maps extracted by the convolutional layer are integrated in the fully connected layer; S7. Convert the integrated results of the fully connected layer into probability distribution through Softmax logistic regression for classification and final output.

2. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 1 is characterized in that: In S2, the process of constructing an iron loss analysis model taking into account spatial harmonics is as follows: The magnetic flux density is converted into a trapezoidal waveform by the finite element method, and the iron loss density in the tooth and yoke regions is expressed as: （1）； In the formula, , are the iron loss densities in the tooth and yoke regions, respectively; , are the empirical coefficients of the magnetic flux density amplitude in the tooth and yoke regions, respectively; , are the magnetic flux density amplitudes in the tooth and yoke regions, respectively; , T are the motor angular frequency and excitation period respectively; , The number of phases and slots per pole and phase respectively; are the derived fitting coefficients; , are the hysteresis loss coefficient and eddy current loss coefficient respectively; β is the magnet width coverage, expressed as: （2）； In the formula, p is the number of poles; is the effective width of the magnet; r is the outer radius of the rotor; Determine the total iron losses in the stator core considering space harmonics , expressed as: （3）； In the formula, , are the volumes of the tooth part and the yoke part respectively.

3. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 2 is characterized in that: In S2, the process of constructing the iron loss analysis model considering PWM harmonics is: Define a correction factor , considering the cumulative effect of each component caused by harmonics, the hysteresis loss and eddy current loss caused by PWM harmonics are expressed as: （4）； In the formula, , They are hysteresis loss and eddy current loss caused by nth PWM harmonics respectively; is the hysteresis loss density; represents the eddy current loss caused by the nth harmonic; where, It is expressed as: （5）； In the formula, is a constant coefficient between 0.6 and 0.7; is the magnetic flux density amplitude; N is the number of samples; is the fluctuation of the motor magnetic flux density of the ath sample; It is expressed as: （6）； In the formula, f, They are the fundamental wave and nth harmonic frequency of the PWM waveform respectively; is the number of winding turns; S is the cross-sectional area of ​​the core; is the amplitude of the nth harmonic voltage; is the nth harmonic magnetic flux density amplitude, expressed as: （7）。 4. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 3 is characterized in that: In S2, the improved iron loss analysis model that takes into account both space harmonics and PWM harmonics is expressed as: （8）； In the formula, represents the total iron loss in the stator core considering both space harmonics and PWM harmonics; , are the correction coefficients for the tooth and yoke regions, respectively; , are the hysteresis loss densities in the tooth and yoke regions, respectively; is the eddy current loss density in the tooth area; , are the longitudinal flux and normal component of the eddy current loss density in the yoke area respectively; i represents the order of the harmonic.

5. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 4 is characterized in that: In S3, the iron loss analysis model calibrated by the Pearson correlation coefficient is expressed as: （9）； In the formula, is obtained by fitting the iron loss curve and the mechanical stress The associated hysteresis loss coefficient; is obtained by fitting the iron loss curve with the operating temperature The associated eddy current loss coefficient.

6. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 5 is characterized in that: In S3, the multi-physical mechanism coupled iron loss analysis model is expressed as: （10）； In the formula, Represents the total iron loss in the stator core after the multi-physics mechanism is coupled; For consideration and The hysteresis loss coefficient; For consideration and The eddy current loss coefficient.

7. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 1 is characterized in that: In the above-mentioned S4, the selected features are six key features: load current, operating temperature, mechanical stress, mechanical speed, tooth magnetic flux density and yoke magnetic flux density.

8. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 1 is characterized in that: In S5, the loss function uses the root mean square error RMES.

9. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 1 is characterized in that: In the S6, the CNN network includes an input layer, a convolution layer, an activation function, a pooling layer, a fully connected layer and an output layer, wherein there are two convolution layers, namely convolution layer 1 and convolution layer 2, the number of convolution kernels, kernel size and convolution stride of convolution layer 1 are 16, 2, and 2 respectively, and the number of convolution kernels, kernel size and convolution stride of convolution layer 2 are 32, 2, and 1 respectively.

10. The iron loss modeling method of a permanent magnet synchronous motor mechanism-data hybrid drive according to claim 7, characterized in that: In the above S6, the multi-layer feature map output is expressed as: （11）； In the formula, is the output of the jth convolution kernel in the hth layer; It is the qth feature map output by the h-1th layer; is the number of feature maps; Represents the convolution operation; is the jth convolution kernel of the hth layer; Represents the bias of the jth convolution kernel of the hth layer; is the ReLU activation function, expressed as: （12）； In the formula, x represents the linear combination of the convolution kernel output; The feature map of the convolutional layer is expressed as: （13）； In the formula, is the output of the hth layer; It is the feature map output by the h-1th layer; is the weight matrix of the hth layer; Represents the bias of the hth layer.

Images