Permanent magnet motor small sample demagnetization fault diagnosis method based on binary information fusion and EMC-Net

Abstract

The invention discloses a permanent magnet motor small sample demagnetization fault diagnosis method based on binary information fusion and EMC-Net. According to the scheme, stator current and torque time sequence data of a motor are compressed and reconstructed through compressed sensing, so that expansion generation of limited data samples is achieved, and meanwhile the problem that fault data are deficient is solved. Furthermore, a multi-source visual heterogeneous information fusion strategy is provided, the strategy utilizes a two-dimensional mapping algorithm to visualize time series data so as to avoid a complex signal processing process, multi-source visual heterogeneous data are fused according to visual saliency and a weighted least square theory, and then enhancement and complementation of demagnetization features are achieved. And finally, an EMC-Net fault diagnosis and identification model is designed. According to the model, ConvNeXt is taken as a basic framework, a network can adaptively lock potential key fault information in input data by embedding an efficient multi-scale attention mechanism, and meanwhile, demagnetization characteristics of different scales are effectively captured by utilizing a parallel sub-network structure, so that the fault diagnosis performance is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of permanent magnet motor fault diagnosis, and in particular relates to a small sample demagnetization fault diagnosis method for permanent magnet motors based on binary information fusion and EMC-Net. Background Art

[0002] The health of the permanent magnets in a permanent magnet synchronous motor's rotor is crucial to its safe operation. However, the rare earth magnets used in these permanent magnets, such as neodymium and samarium-cobalt, can experience irreversible demagnetization under high temperatures, overload, and vibration, leading to reduced motor output and potentially gradual failure. Therefore, accurately monitoring the health of the permanent magnets is crucial to maintaining stable motor operation and preventing premature and persistent deterioration.

[0003] Common permanent magnet demagnetization faults can be roughly divided into uniform demagnetization faults and localized demagnetization faults. Uniform demagnetization faults are more common in extreme overload and weak magnetic failure conditions. Currently, more accurate diagnosis and compensation can be achieved by constructing appropriate observers. Localized demagnetization faults, which do not occur in all rotor permanent magnets, are generally caused by local high temperatures, frequent overloads, and other factors. In addition to causing a decrease in the average output of the motor, the back-electromotive force distortion caused by localized demagnetization faults can exacerbate the unbalanced magnetic pull of the motor, leading to serious noise, vibration, and harshness issues. Therefore, early diagnosis and fault location of localized demagnetization faults are crucial to preventing demagnetization faults from developing into major safety accidents.

[0004] In real-world industrial applications, motors operate in a healthy state most of the time, making fault data collection difficult and limiting the number of valid fault samples. This can lead to overfitting or underfitting of network models. Furthermore, after a permanent magnet synchronous motor fails, the distribution of its internal electromagnetic and temperature fields becomes even more complex. Due to the lack of ideal datasets, many key research questions remain regarding intelligent fault diagnosis for permanent magnet synchronous motors. Furthermore, most traditional fault diagnosis methods rely on a single type of sensor data. For motor fault diagnosis, multivariate information can be used to enrich fault feature descriptions, enhance the robustness of network models, and improve diagnostic accuracy.

[0005] Aiming at these problems in traditional methods, the present invention proposes a small sample demagnetization fault diagnosis method for permanent magnet motors based on binary information fusion and EMC-Net. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, the present invention provides a small-sample demagnetization fault diagnosis method for permanent magnet motors based on binary information fusion and EMC-Net. Based on compressed sensing theory, this method compresses and reconstructs the motor's stator current and torque time-series data to achieve expanded generation of limited data samples while overcoming the problem of insufficient fault data. Secondly, a multi-source visual heterogeneous information fusion strategy is designed. This strategy utilizes a Markov transition field algorithm to visualize the time-series data, avoiding complex signal processing. The multi-source visual heterogeneous data is then fused based on visual saliency and weighted least squares theory to enhance and complement the demagnetization features. Finally, the fused image is input into the EMC-Net neural network for local demagnetization fault diagnosis.

[0007] The technical solution adopted by this invention is a small sample demagnetization fault diagnosis method for permanent magnet motors based on binary information fusion and EMC-Net. The specific content is as follows:

[0008] Step 1: Assume that the motor timing data f=[f1 f2…f N ] T It has sparse characteristics in the Fourier transform domain, so it can be expressed as

[0009] f=ψs

[0010] Where s is a sparse or nearly sparse N×1-dimensional signal. If the number of non-zero values ​​K in s is much smaller than the signal length N, the signal s is called K-sparse. Ψ is an N×N Fourier matrix, which can be expressed as

[0011]

[0012] Where ω = e 2π c / N, c is a positive integer. When there is an uncorrelated M×N dimensional measurement matrix Φ, the original high-dimensional time series data f can be linearly projected into an M×1 dimensional low-dimensional observation vector y, that is, y=Φf=Φψs=As

[0013] Where A is the sensing matrix. A random Gaussian matrix is ​​selected as the measurement matrix, and each element in the matrix independently obeys the normal distribution. The probability density function p(θ) of this matrix can be expressed as

[0014]

[0015] Where θ is the sampling point of the time domain data. The advantage of the random Gaussian matrix is ​​that it can be compressed at a lower ratio. A higher reconstruction accuracy is achieved, that is, when the sampling length M ≥ rKlog(N / K) (r is a very small constant), the original time series data can be reconstructed with a higher probability.

[0016] Signal reconstruction is an important part of compressive sensing. According to the above discussion, the signal reconstruction problem can be solved under the framework of the l1 - norm minimization, that is

[0017]

[0018] However, since M << N, solving the above equation is an under - determined problem and is not easy to solve. To solve this problem, the orthogonal matching pursuit method is used to approximately solve the above equation. Orthogonal matching pursuit is a non - linear adaptive algorithm. The core idea of this algorithm is to select the column vectors in the sensing matrix A in a greedy iterative manner, so that the selected column vectors a j (j = 1, 2,..., n) are most relevant to the current residual r, then subtract the relevant part from the compressed observation vector y and repeat the iteration until the iteration number l is equal to the sparsity K and then stop. The orthogonal matching pursuit method estimates the sparse coefficients s of the signal in the Fourier transform domain, and obtains the estimated sparse coefficients After that, the original high - dimensional signal can be recovered and reconstructed through the following formula.

[0019]

[0020] Since the constructed random Gaussian matrix has strong randomness and diversity, when the compressive sensing algorithm is executed each time, a reconstructed signal that is similar but slightly different in statistical characteristics from the original signal can be obtained Furthermore, the purpose of small - sample sensor data augmentation can be achieved. In addition, in the frequency domain, compressive sensing approximates the original signal by retaining the key coefficients with larger amplitudes and discarding the coefficients with smaller amplitudes. This mechanism can suppress the electromagnetic harmonics and background noise in the original time - series data, thereby improving the signal quality.

[0021] Step (2): The Markov transfer field is a method to visually represent the hidden information in time - domain data through Markov transition probabilities. This method can retain the correlation information between the amplitude and phase in the time series, while enhancing the richness of the fault information and realizing the enhanced display of the demagnetization characteristics [[ID=A22]]

[0022] According to the numerical value of the time - series data f of the permanent - magnet synchronous motor, it is divided into H j quantile bins, and each data point is mapped to the corresponding quantile region. Subsequently, based on the principle of the first - order Markov chain, the transition probabilities between each quantile region are calculated to obtain the Markov transition matrix Q, which can be expressed as

[0023]

[0024] In the formula, q [[ID=A33]] ij is from bin H i to bin H jFinally, a two-dimensional image is generated by linearly mapping the value range of each element in the matrix Q from [0, 1] to [0, 255].

[0025] Step 3): By utilizing visual saliency and weighted least squares method to fuse the Markov transfer field images of the permanent magnet motor current signal and torque signal, complementary enhancement of the demagnetization features is achieved.

[0026] The specific steps of visual saliency mapping and least weighted squares method are:

[0027] 1. Use rolling guided filter and Gaussian filter to perform multi-scale decomposition. The rolling guided filter mainly consists of two parts. First, the Gaussian filter is used to remove small structures in the input image T. After the input image T passes through the Gaussian filter, the output image U can be expressed as

[0028] U=GaussianFilter(T,σ f )

[0029] Where, GaussianFilter(T,σ f ) is the standard deviation σ f is a Gaussian filter with scaling parameters.

[0030] At the same time, the guided filter is used to iteratively restore the edge information of the image, which can be expressed as

[0031]

[0032] Where, is the guided filter, P v is the guide image, σ k is the regularization parameter. When the number of iterations v = 1, the initial image P 1 is the output image U of the Gaussian filter. The rolling guided filter is implemented by the above two equations, which can be further expressed as

[0033] I=RGF(T,σ f ,σ k , V)

[0034] Where V is the total number of iterations and I is the output value of the rolling guidance filter.

[0035] Therefore, the multi-scale decomposition mathematical model combining rolling guide filter and Gaussian filter can be expressed as

[0036]

[0037] d e =I e-1 -I e , e=1,2,...,L-1

[0038]

[0039] d e =I e-1 -I e , e=L

[0040] Where, I e and d e are the filtered output image and image detail layer of the e-th layer respectively, and L is the number of multi-scale decomposition layers. The initial image I 0 = T. In addition, the image base layer B = I L is obtained by Gaussian filter processing. At the same time, set the scale parameter to obtain increasingly coarse images.

[0041] 2. Fusion of base layers. Based on the multi-scale decomposition method, the Markov transfer field image of the permanent magnet synchronous motor current and torque can be decomposed into base layers B1 and B2. Furthermore, the visual saliency technology is used to fuse the base layers of the image, giving priority to retaining the significant areas of demagnetization features in the image to avoid losing the contrast information of the image. The basic idea of ​​this method is to define the pixel-level saliency by the contrast between a single pixel value and all other pixel values ​​in the image. If T g is the intensity value of pixel g in image T, then the saliency value V(g) of pixel g can be expressed as

[0042] V(g)=|T g -T1|+|T g -T2|+…+|T g -T o |

[0043] Where O is the total number of pixel values ​​in the image T. If the intensity values ​​of two pixel values ​​are the same, then their significance values ​​are equal. In this case, the above formula can be expressed as

[0044]

[0045] Where κ is the pixel intensity value of the image, Z κ is the number of pixels with intensity equal to κ, is the number of gray levels.

[0046] After processing by visual saliency technology, the saliency maps V1 and V2 of the Markov transfer field images of different sensor data can be obtained, and then the fusion process of B1 and B2 can be expressed as follows:

[0047]

[0048] Where, and are the fused basic layer image and weight coefficient respectively.

[0049] 3. Fusion of detail layers. To fuse the detail layers of the current and torque Markov transfer field images and Define the weight coefficient W e for

[0050]

[0051] In order to reduce the noise components in the image, a Gaussian filter is used to filter W e Processing is as shown below

[0052]

[0053] Where, W e The output after smoothing by Gaussian filter, standard deviation σ f It is usually set to 2. According to the maximum absolute value principle, the fusion detail layer M of the e-th layer can be obtained. e ,Right now

[0054]

[0055] Based on the weighted least squares method, the fusion details of the e-th layer are optimized to obtain the fusion detail layer D of the e-th layer. e , which can be expressed as

[0056]

[0057] Where, and q are the spatial adaptive weights and the spatial positions of the image pixels, respectively. is a square window centered at pixel g, and μ is usually set to a small constant to prevent it from being divided by zero. The purpose is to minimize D e and M e The Euclidean distance between them, and the second term is set The purpose is to make D e More inclined to α is used to globally control the emphasis of the image fusion model on the two terms in the above formula.

[0058] The matrix representation of the above formula is

[0059]

[0060] Where A e is determined by the weight β e Minimize the D of the above formula eIt is given by the unique solution of the following linear system, namely

[0061]

[0062] Where E is the identity matrix. Since Ae is a diagonal matrix, the above formula can be further expressed as

[0063]

[0064] In the above formula, since (E+αA e ) is a diagonal matrix, so the fusion detail layer D e It can be expressed as

[0065]

[0066] Finally, the base layers are fused according to the following formula With detail layer D 1 , D 2 ,...,D L , the final fused Markov transfer field image can be obtained.

[0067]

[0068] Where FMTF is the fused Markov transfer field image of different sensor data.

[0069] Step 4): By combining the ConvNeXt neural network with an efficient multi-scale attention mechanism, an EMC-Net is designed to capture the demagnetization feature information hidden in the fused Markov transfer field image to achieve accurate diagnosis and identification of demagnetization faults.

[0070] First, the Markov transfer field fusion image will go through four cascade modules consisting of downsampling layers and EMC modules. The downsampling layer is used to gradually compress the feature map size and extract higher-level semantic features, while the EMC module is used to achieve efficient multi-level visual feature extraction through convolution operations and attention mechanisms. It can be simply expressed as

[0071] Q(X)=F(X)+X

[0072] In the formula, X and Q(X) are the input and output feature maps of the EMC module, respectively, and F(X) is the residual, representing the error between X and Q(X). Through the identity mapping connection in the EMC module, the input demagnetization feature information can be propagated forward more quickly. The EMC module utilizes depthwise separable convolution to improve the model's computational speed and adopts larger convolution kernels to capture a wider range of contextual feature information. Secondly, layer normalization is used instead of batch normalization to address the problem of excessive deviation in normalized statistics when the data sample size is too small. Furthermore, layer scaling stabilizes network training and dynamically calibrates feature importance.

[0073] In order to enhance the network's ability to perceive demagnetization feature information, the parallel sub-network structure in the efficient multi-scale attention mechanism is used to capture spatial semantic information at different scales. The efficient multi-scale attention mechanism converts the input feature map into (C, H and W are the number of channels, height and width of the input image respectively) It is divided into G sub-features along the channel dimension, that is, X = [X0, X Υ ,…,X G-1 ], To learn demagnetization features of different dimensions separately. Furthermore, for each sub-feature, an efficient multi-scale attention mechanism extracts attention weight information for the grouped feature maps through three parallel branches. In the first two branches, the efficient multi-scale attention mechanism module average pools the input feature maps along both the height and width directions. The pooling results in both directions are then aggregated, and a 1×1 convolution operation is performed simultaneously. Nonlinear fitting is performed using a sigmoid activation function to enable cross-channel information interaction between the two branches. In the third branch, a 3×3 convolution is used to mine fault feature information, aiming to capture multi-scale feature representations. In the cross-space learning phase, two-dimensional global average pooling is used to process the outputs of the 1×1 and 3×3 convolution branches, and the softmax function is used to normalize the feature weights of branches at different scales. Matrix multiplication is then used to aggregate the feature information of branches at different scales across space to obtain the final feature representation.

[0074] At the end of the EMC module, Drop Path regularization randomly removes subpaths from the multi-branch structure of the network to alleviate network degradation and prevent overfitting. Finally, the feature extraction results of the cascade module are input into the classification head for global average pooling. A fully connected layer maps the distributed features to the sample label space to obtain the final classification and recognition results.

[0075] Compared with the existing technical solutions, the technical solutions of the present invention have the following beneficial effects:

[0076] This invention is based on a data-driven diagnostic method. First, by introducing compressed sensing theory, it effectively suppresses electromagnetic harmonics and background noise in complex industrial environments in time-domain signals, while simultaneously enhancing and expanding current and torque time-domain data. This effectively solves the problem of low diagnostic accuracy of deep learning models due to a lack of fault data. Secondly, using the Markov transfer field algorithm as a medium, one-dimensional data is converted into two-dimensional images to visualize fault features, avoiding complex signal processing. Then, the innovative introduction of visual saliency and weighted least squares methods fuses and transforms the Markov transfer field images of binary heterogeneous sensor data, achieving complementary enhanced display of demagnetization fault features. Finally, an EMC-Net permanent magnet synchronous motor local demagnetization fault diagnosis model based on an attention mechanism is proposed. This model accurately captures the key information hidden in the two-dimensional image through the attention mechanism, and can achieve simultaneous diagnosis and separation of multiple demagnetization modes. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a schematic diagram of the principle of the present invention based on compressed sensing signal sampling;

[0078] Figure 2 It is a flow chart of two-dimensional image conversion based on Markov transition field of the present invention;

[0079] Figure 3 It is a binary information fusion graph based on visual saliency mapping and least weighted squares method of the present invention;

[0080] Figure 4 This is the overall structure diagram of the EMC-Net network of the present invention;

[0081] Figure 5 It is a structural diagram of the EMC module of the present invention;

[0082] Figure 6 This is the structural diagram of the efficient multi-scale attention mechanism of the present invention; DETAILED DESCRIPTION

[0083] The following describes in detail a method for diagnosing small-sample demagnetization faults of permanent magnet motors based on binary information fusion and EMC-Net in conjunction with the embodiments and drawings.

[0084] The method of the present invention comprises the steps of:

[0085] Step 1: Assume that the motor timing data f=[f1 f2…f N ] T It has sparse characteristics in the Fourier transform domain, so it can be expressed as

[0086] f = ψs (1) Where s is a sparse or approximately sparse N×1 dimensional signal. If the number K of non-zero values in s is much smaller than the signal length N, the signal s is called K-sparse. Ψ is an N×N Fourier matrix, which can be expressed as

[0087]

[0088] where ω = e 2πc / N , and c is a positive integer. Figure 1 Fig. shows the signal sampling principle based on compressive sensing. When there is an M×N dimensional measurement matrix Φ that is uncorrelated, the original high-dimensional time-series data f can be linearly projected into an M×1 dimensional low-dimensional observation vector y, that is

[0089] y = Φf = Φψs = As (3)

[0090] where A is the sensing matrix. A random Gaussian matrix is selected as the measurement matrix, and each element in the matrix independently follows a normal distribution. The probability density function p(θ) of this matrix can be expressed as

[0091]

[0092] where θ is the time-domain data sampling point. The advantage of the random Gaussian matrix is that it can achieve a high reconstruction accuracy at a low compression ratio That is, when the sampling length M ≥ rK log(N / K) (r is a very small constant), it can be guaranteed to reconstruct the original time-series data with a high probability.

[0093] Signal reconstruction is an important part of compressive sensing. According to the above discussion, the signal reconstruction problem corresponding to equation (3) can be solved under the l1-norm minimum framework, that is

[0094]

[0095] However, since M << N, solving (5) is an underdetermined problem and is not easy to solve. To solve this problem, the orthogonal matching pursuit method is used to approximately solve (5). Orthogonal matching pursuit is a non-linear adaptive algorithm. The core idea of this algorithm is to greedily and iteratively select the column vectors in the sensing matrix A such that the selected column vectors a j (j = 1, 2,..., n) are most correlated with the current residual r, and then subtract the relevant part from the compressed observation vector y and repeat the iteration until the iteration number l is equal to the sparsity K and then stop. The orthogonal matching pursuit method estimates the sparse coefficient s of the signal in the Fourier transform domain, and obtains the estimated sparse coefficient After that, the original high-dimensional signal can be restored and reconstructed through equation (6).

[0096]

[0097] Since the constructed random Gaussian matrix has strong randomness and diversity, each time the compressed sensing algorithm is executed, a reconstructed signal with similar statistical characteristics to the original signal but with slight differences can be obtained. This can achieve the goal of expanding small sample sensor data. Furthermore, in the frequency domain, compressed sensing achieves a close approximation of the original signal by retaining key coefficients with large amplitudes and discarding coefficients with smaller amplitudes. This mechanism can suppress electromagnetic harmonics and background noise in the original time series data, thereby improving signal quality.

[0098] Step 2): Markov transition field is a method to visualize the hidden information in time domain data by using Markov transition probability. This method can preserve the correlation between amplitude and phase in the time series, while improving the richness of fault information and achieving enhanced display of demagnetization characteristics. Taking the stator current signal as an example, the Markov transition field two-dimensional image conversion process is as follows: Figure 2 shown.

[0099] According to the numerical value of the time series data f of the permanent magnet synchronous motor, it is divided into H j quantile bins and map each data point to the corresponding quantile region. Then, based on the first-order Markov chain principle, the transition probability between each quantile region is calculated to obtain the Markov transition matrix Q, which can be expressed as

[0100]

[0101] Where q ij For box H i To box H j Finally, a two-dimensional image is generated by linearly mapping the value range of each element in the matrix Q from [0, 1] to [0, 255].

[0102] Step 3): By utilizing visual saliency and weighted least squares method to fuse the Markov transfer field images of the permanent magnet motor current signal and torque signal, complementary enhancement of the demagnetization features is achieved.

[0103] The specific steps of visual saliency mapping and least weighted squares method are:

[0104] 1. Use rolling guided filter and Gaussian filter to perform multi-scale decomposition. The rolling guided filter mainly consists of two parts. First, the Gaussian filter is used to remove small structures in the input image T. After the input image T passes through the Gaussian filter, the output image U can be expressed as

[0105] U=GaussianFilter(T,σ f ) (8)

[0106] Where, GaussianFilter(T,σ f ) is the standard deviation σ f is a Gaussian filter with scaling parameters.

[0107] At the same time, the guided filter is used to iteratively restore the edge information of the image, which can be expressed as

[0108]

[0109] Where, is the guided filter, P v is the guide image, σ k is the regularization parameter. When the number of iterations v = 1, the initial image P 1 is the output image U of the Gaussian filter. The rolling guided filter is implemented by equations (8) and (9), which can be further expressed as

[0110] I=RGF(T,σ f ,σ k , V) (10) where V is the total number of iterations and I is the output value of the rolling guidance filter.

[0111] Therefore, the multi-scale decomposition mathematical model combining rolling guide filter and Gaussian filter can be expressed as

[0112]

[0113] d e =I e-1 -I e , e=1,2,...,L-1 (12)

[0114]

[0115] d e =I e-1 -I e , e=L (14)

[0116] Where, I e and d e are the filtered output image and image detail layer of the e-th layer respectively, and L is the number of multi-scale decomposition layers. The initial image I 0 = T. In addition, the image base layer B = I L is obtained by Gaussian filter processing. At the same time, set the scale parameter to obtain increasingly coarse images.

[0117] 2. Fusion of base layers. Based on the multi-scale decomposition method, the Markov transfer field image of the permanent magnet synchronous motor current and torque can be decomposed into base layers B1 and B2. Furthermore, the visual saliency technology is used to fuse the base layers of the image, giving priority to retaining the significant areas of demagnetization features in the image to avoid losing the contrast information of the image. The basic idea of ​​this method is to define the pixel-level saliency by the contrast between a single pixel value and all other pixel values ​​in the image. If T g is the intensity value of pixel g in image T, then the saliency value V(g) of pixel g can be expressed as

[0118] V(g)=|T g -T1|+|T g -T2|+…+|T g -T o | (15)

[0119] Where O is the total number of pixel values ​​in image T. If the intensity values ​​of two pixel values ​​are the same, their significance values ​​are equal. In this case, equation (15) can be expressed as

[0120]

[0121] Where κ is the pixel intensity value of the image, Z k is the number of pixels with intensity equal to κ, is the number of gray levels.

[0122] After processing by visual saliency technology, the saliency maps V1 and V2 of the Markov transfer field images of different sensor data can be obtained, and then the fusion process of B1 and B2 can be expressed as follows:

[0123]

[0124]

[0125] Where, and are the fused basic layer image and weight coefficient respectively.

[0126] 3. Fusion of detail layers. To fuse the detail layers of the current and torque Markov transfer field images and Define the weight coefficient W e for

[0127]

[0128] In order to reduce the noise components in the image, a Gaussian filter is used to filter W e Processing is as shown below

[0129]

[0130] Where, W e The output after smoothing by Gaussian filter, standard deviation σ f It is usually set to 2. According to the maximum absolute value principle, the fusion detail layer M of the e-th layer can be obtained. e ,Right now

[0131]

[0132] Based on the weighted least squares method, the fusion details of the e-th layer are optimized to obtain the fusion detail layer D of the e-th layer. e , which can be expressed as

[0133]

[0134] Where, and q are the spatial adaptive weights and the spatial positions of the image pixels, respectively. is a square window centered at pixel g, and μ is usually set to a small constant to prevent it from being divided by zero. The purpose is to minimize D e and M e The Euclidean distance between them, and the second term is set The purpose is to make D e More inclined to α is used to globally control the emphasis of the image fusion model on the two terms in the above formula.

[0135] The matrix representation of the above formula is

[0136]

[0137] Where A e is determined by the weight β e Minimize D in formula (24) e It is given by the unique solution of the following linear system, namely

[0138]

[0139] Where E is the unit matrix. e is a diagonal matrix, so Equation (25) can be further expressed as

[0140]

[0141] In formula (26), since (E+αA e ) is a diagonal matrix, so the fusion detail layer D e It can be expressed as

[0142]

[0143] Finally, the base layers are fused according to the following formula With detail layer D 1 , D 2 ,...,D L , the final fused Markov transfer field image can be obtained.

[0144]

[0145] Where FMTF is the fused Markov transfer field image of different sensor data.

[0146] In summary, the visual saliency map and the least weighted squares algorithm can achieve effective fusion and feature complementation of multi-sensor heterogeneous information. The fused Markov transfer field image is as follows: Figure 3 As shown. Figure 3 It can be seen that before fusion, the Markov transfer field image features between the test prototypes showed little difference, showing a certain degree of similarity. However, after fusing the heterogeneous multi-sensor data using visual saliency mapping and the least weighted squares algorithm, the fused image showed a more distinct separability in terms of texture structure and feature distribution, thereby improving the deep learning network model's ability to identify permanent magnet motor demagnetization faults.

[0147] Step 4): By combining the ConvNeXt neural network with an efficient multi-scale attention mechanism, an EMC-Net is designed to capture the demagnetization feature information hidden in the fused Markov transfer field image to achieve accurate diagnosis and identification of demagnetization faults, such as Figure 4 shown.

[0148] First, the Markov transfer field fusion image will go through four cascade modules consisting of downsampling layers and EMC modules. The downsampling layer is used to gradually compress the feature map size and extract higher-level semantic features, while Figure 5 The EMC module in the model achieves efficient multi-level visual feature extraction through convolution operations and attention mechanisms, which can be simply expressed as

[0149] Q(X)=F(X)+X (29)

[0150] In the formula, X and Q(X) are the input and output feature maps of the EMC module, respectively, and F(X) is the residual, representing the error between X and Q(X). Through the identity mapping connection in the EMC module, the input demagnetization feature information can be propagated forward more quickly. The EMC module utilizes depthwise separable convolution to improve the model's computational speed and adopts larger convolution kernels to capture a wider range of contextual feature information. Secondly, layer normalization is used instead of batch normalization to address the problem of excessive deviation in normalized statistics when the data sample size is too small. Furthermore, layer scaling stabilizes network training and dynamically calibrates feature importance.

[0151] In order to enhance the network's ability to perceive demagnetization feature information, the parallel sub-network structure in the efficient multi-scale attention mechanism is used to capture spatial semantic information at different scales, such as Figure 6 As shown. The efficient multi-scale attention mechanism takes the input feature map (C, H and W are the number of channels, height and width of the input image respectively) It is divided into G sub-features along the channel dimension, that is, X = [X0, X Υ ,…,X G-1 ], To learn demagnetization features of different dimensions separately. Furthermore, for each sub-feature, an efficient multi-scale attention mechanism extracts attention weight information for the grouped feature maps through three parallel branches. In the first two branches, the efficient multi-scale attention mechanism module average pools the input feature maps along both the height and width directions. The pooling results in both directions are then aggregated, and a 1×1 convolution operation is performed simultaneously. Nonlinear fitting is performed using a sigmoid activation function to enable cross-channel information interaction between the two branches. In the third branch, a 3×3 convolution is used to mine fault feature information, aiming to capture multi-scale feature representations. In the cross-space learning phase, two-dimensional global average pooling is used to process the outputs of the 1×1 and 3×3 convolution branches, and the softmax function is used to normalize the feature weights of branches at different scales. Matrix multiplication is then used to aggregate the feature information of branches at different scales across space to obtain the final feature representation.

[0152] At the end of the EMC module, Drop Path regularization randomly removes subpaths from the multi-branch structure of the network to alleviate network degradation and prevent overfitting. Finally, the feature extraction results of the cascade module are input into the classification head for global average pooling. A fully connected layer maps the distributed features to the sample label space to obtain the final classification and recognition results.

[0153] From the above description, it can be seen that the present invention proposes a small-sample demagnetization fault diagnosis method for permanent magnet motors based on binary information fusion and EMC-Net. While enhancing and expanding limited data, it also realizes the enhancement and complementarity of binary information features, thereby limitedly improving the diagnostic accuracy and generalization ability of the intelligent model.

[0154] The above description of the functions and working processes of the present invention in combination with the drawings in the specification is only one of the preferred implementation cases, but the present invention is not limited to the above specific functions and working processes. The above specific implementation methods are only illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which are all within the scope of protection of the present invention.

Claims

1. A small sample demagnetization fault diagnosis method for permanent magnet motors based on binary information fusion and EMC-Net, characterized in that: The following steps are involved: 1) A data augmentation strategy based on compressed sensing is proposed. Leveraging the randomness and diversity of random Gaussian matrices, a compressed sampling operation is performed on the motor current and torque time series data. This is then iteratively reconstructed using an orthogonal matching pursuit algorithm to augment the limited data sample. The enhanced data retains the key features of the original data to the greatest extent possible. 2) A multi-source visual heterogeneous information fusion strategy is proposed. Using a Markov transfer field algorithm, the motor's one-dimensional time-domain signals are converted into two-dimensional visualizations without complex signal processing. This allows for rapid and effective visualization of varying degrees of demagnetization faults. Furthermore, based on visual saliency and weighted least squares theory, the two-dimensional visualizations are fused to enhance and complement the demagnetization signature information. 3) We designed an efficient multi-scale convolutional neural network. Based on the ConvNeXt classification model, we leveraged the powerful feature extraction capabilities of convolutional neural networks to classify and identify two-dimensional images. By incorporating an efficient multi-scale attention mechanism into the ConvNeXt classification model, we constructed EMC-Net for demagnetization fault diagnosis. This model selectively ignores irrelevant information and focuses on features that are relevant to the target classification task, effectively improving the model's generalization capabilities.

Images