Motor initial turn-to-turn short circuit fault diagnosis method based on dynamic sparse convolution residual network

Abstract

The invention discloses a motor initial turn-to-turn short circuit fault diagnosis method based on a dynamic sparse convolution residual network, and belongs to the technical field of motor turn-to-turn short circuit fault diagnosis. The method comprises the following steps: firstly, providing a new dynamic sparse convolution residual module for extracting a local feature and long-range periodic dependency relationship from a time sequence signal; the architecture adopts a novel dynamic sparse convolution filter which has a dynamic receptive field and relatively low computational complexity. Secondly, providing an improved extrusion-excitation mechanism, and adaptively enhancing the characterization capability of the model on fault-related discrimination features; the provided intelligent diagnosis architecture is suitable for various motors, and has universality and robustness. The method can be used for early turn-to-turn short circuit fault diagnosis of the motor under multiple operation working conditions and data scarcity conditions, original current signals are directly adopted, and end-to-end intelligent fault diagnosis and severity estimation of the motor are achieved.

Description

Technical Field

[0001] This invention relates to the field of motor inter-turn short-circuit fault diagnosis technology, specifically to a method and system for diagnosing initial inter-turn short-circuit faults in motors using a dynamic sparse convolutional residual network. Background Technology

[0002] As a core component of electromechanical energy conversion, electric motors are widely used in industry, intelligent robotics, and transportation. Electrical equipment can fail due to the combined effects of electrical, mechanical, thermal, and environmental stresses. Failure to detect motor faults in a timely manner can lead to severe damage to the motor and its drive system. Inter-turn short circuit (ITSC) faults in stator windings are a typical electrical fault, accounting for approximately 30% of all faults. Winding short circuits generate enormous heat, causing insulation failure and potentially spreading to other windings. In severe cases, they can lead to phase-to-phase or phase-to-ground short circuits and irreversible demagnetization. Therefore, timely detection of initial winding short circuit faults is crucial to preventing catastrophic motor damage and improving equipment reliability, requiring efficient and reliable fault diagnosis technologies.

[0003] Currently, ITSC fault diagnosis for electric motors is mainly divided into model-based methods and data-driven methods. Model-based methods estimate motor parameters by establishing mathematical models or observers and analyze the residuals between model predictions and actual values ​​as fault diagnosis indicators. However, electric motors are highly coupled nonlinear systems, making it difficult to establish accurate mathematical models under various operating conditions. The universality of algorithms is limited by the differences in models and parameters between different motors. Data-driven methods identify motor fault modes by extracting and analyzing fault features from historical and real-time operating data. Current data from the motor stator windings is often used as the signal source for ITSC fault diagnosis. Current current feature signal analysis methods, such as Fast Fourier Transform, wavelet analysis, Short-Time Fourier Transform, Hilbert-Huang Transform, and Wigner-Weil distribution, can extract fault symptoms without precise mathematical modeling. However, these methods rely on expert experience and prior knowledge, resulting in limited diagnostic efficiency. Subsequently, shallow machine learning methods, such as Adaline, K-Nearest Neighbors, and Support Vector Machines, have been introduced and combined with signal analysis to achieve automatic association between fault features and fault modes. These methods still rely on previous feature engineering. With the increasing demands on the operation of complex motor systems, fault diagnosis methods require stronger noise immunity and rapid response capabilities to enable early warning of faults. In recent years, fault diagnosis methods based on deep representation learning have shown great potential. Deep learning models can automatically extract optimal feature representations suitable for diagnostic tasks from complex raw data. These end-to-end fault diagnosis methods no longer rely on expert knowledge and precise mathematical modeling, offering greater flexibility. Summary of the Invention

[0004] This invention proposes a method for diagnosing initial inter-turn short circuit faults (ITSCs) in motors using dynamic sparse convolutional residual networks. This method is used to diagnose the initial ITSC faults of various motors under different operating conditions and assess their severity. It achieves high diagnostic accuracy, strong noise resistance, and robustness while maintaining low computational cost.

[0005] To achieve the technical objective, the present invention adopts the following technical solution, including the following steps:

[0006] Step S1: Build experimental platforms for induction motors and permanent magnet synchronous motors.

[0007] Step S2: Based on the motor test platform, collect the motor stator current under various operating conditions through the data acquisition system, including normal motor operation data and data on different fault severity levels.

[0008] Step S3: Integrate the data collected in Step 2 and perform data preprocessing, dividing the dataset into training set and test set according to the proportion.

[0009] Step S4: Based on PyTorch, a dynamic sparse convolutional residual network model for extracting inter-turn short-circuit fault diagnosis is designed. This model includes a dynamic sparse convolutional residual module, an improved squeeze excitation module, and a classification layer. First, to enable the convolutional filter to more flexibly focus on the local and long-range dependencies of multi-scale features in time-series signals, a novel sparse convolutional paradigm with a dynamic offset Δ is constructed. By combining multiple non-overlapping convolutional modules, convolutional kernels with differentiated receptive fields are formed, thus accommodating both local and long-range dependent features. Each sub-convolutional module learns local representations, while multiple sub-convolutional modules collaboratively learn long-term information in the time-series signal. Furthermore, by introducing a dynamic offset Δ between different sub-convolutional kernels, information redundancy can be avoided, and multi-scale features can be extracted more efficiently. Second, based on sparse convolution, a dynamic sparse convolutional residual module (DSCR) is constructed. The convolutional layer adopts a dual convolutional mode combining point convolution (PWConv) and DSConv to decouple channel and temporal feature representations. Specifically, PWConv is used to expand channel dimensions to enrich feature representations and integrate cross-channel information. DSConv focuses on extracting discriminative temporal features. In DSConv, each input channel is independently convolved through a dedicated filter, and the input and output feature maps maintain the same spatial resolution to avoid information loss; then, an improved squeeze-excitation module is used to adaptively enhance fault-related discriminative features; finally, the results are output through a classification layer.

[0010] Step S5: Use the training set to train the model. During the training process, the cross-entropy loss function is introduced to calculate the error between the predicted label and the true label. At the same time, the error is used for backpropagation to adjust the parameters and structure of the neural network. The experiment is repeated to train the model until the set number of iterations is reached and then training is stopped.

[0011] Step S6: Test the trained model using the test set. Input the test samples into the network model to obtain the accuracy of the diagnostic results.

[0012] Furthermore, in step 2, when collecting the motor current, to simulate faults of varying severity, an ITSC fault is injected through the stator winding inter-turn short circuit (ITSC) control module, with phase A designated as the faulty phase. Wires are led out from the internal windings to the motor housing and connected with short-circuit resistors of different values ​​to simulate ITSC faults of varying severity.

[0013] Furthermore, in step 3, when partitioning the dataset, data augmentation and normalization are employed to fully utilize the data and extract more effective features; Each sample is the original signal. A subsequence, starting index is Defined as:

[0014] ;

[0015] in, For the first Each sample corresponds to the original signal. subsequence of For the first Each sample corresponds to

[0016] The number of signal points; when When windows overlap, the number of samples increases; when At the same time, the windows do not overlap. A series of samples are generated by sliding the window. ,in It represents the total number of samples generated.

[0017] The normalization operation is defined as:

[0018] ;

[0019] Where X represents the original signal, X pLet X represent the preprocessed signal, max(X) be the maximum value of the original signal, and min(X) be the minimum value of the original signal. The original data is sampled by overlapping sampling using the data augmentation method of equal interval sliding window to obtain the sample dataset. 1000 samples are set for each working condition, and let α be the proportion of training samples. The sample length is 2048 data points.

[0020] Step S41: To enable convolutional filters to more flexibly focus on the local and long-range dependencies of multi-scale features in time-series signals, thereby improving feature extraction quality, we construct a sparse convolution paradigm with a dynamic receptive field. A dynamic offset Δ is introduced between sub-convolutional kernels to avoid information redundancy and extract multi-scale features more efficiently. The unpadded dynamic sparse convolution can be represented as:

[0021] ;

[0022] in, Let i be the output signal or feature map value at position i of the sparse convolution, where i is the index of the output signal and x is the input signal. Here, represents the weights of the sparse convolution kernel, and s represents the stride of the convolution operation. For the output The offset applied to the input signal by the b-th sparse connection, where m is the depth of the sparse convolution kernel and n is the width of the sparse convolution kernel; a new set of dynamic offsets is generated each time a sliding convolution with stride s is performed; sub-convolutional blocks do not overlap and are arranged sequentially to ensure the continuity and smoothness of time series features; to avoid receptive field overlap, the offset difference between adjacent sub-convolutional blocks must be at least greater than the size n of the sub-convolution kernel; a specific offset ∆ i,j The network dynamically generates the input x; based on DSConv, a dynamic sparse convolutional residual DSCR module is constructed, defined as:

[0023] ;

[0024] Among them W PWConv and W DSConv represents the weights of PWConv and DSConv respectively; BN represents the batch normalization operation; σ represents the LReLU nonlinear activation function; x is the initial input feature map; w is the set of all learnable parameters in the model; x1 is the first intermediate feature map; F is the residual function; y is the final output feature map; Conv_1 is the first convolution operation; and Conv_2 is the second convolution operation.

[0025] Step S42: The data for inter-turn short-circuit fault diagnosis mainly consists of single-channel time-series signals, and fault features are extracted through multi-channel filters within the network. Therefore, the feature information carried by different channels plays different roles in fault identification. To learn more fault-related discrimination information from the input time-series signals, we propose an improved squeeze excitation (ISE) module to adaptively recalibrate the weights of each feature channel. During feature compression, hybrid multi-head pooling is introduced to generate multi-scale channel statistics, replacing simple global average pooling; specifically, feature dimensions are compressed simultaneously through dual-branch multi-head mean pooling and multi-head max pooling, and the information flow is balanced using a learnable weight factor λ. The generated statistics... It can be represented as:

[0026] ;

[0027] Wherein represents Multi-head pooling window size, Indicates the number of heads. For learnable weight factors, The feature statistics generated for the c-th channel, where i is the iteration index. The length of the i-th element, where j is the internal iteration index. Input data points, To sum all x_i(j) values ​​in the i-th element, To find the maximum value of all x_i(j) values ​​in the i-th element, the global statistic is described by the piecewise maximum, while the highest value among the piecewise means represents the local statistic. To fully utilize the information aggregated during the squeezing process, a simple gating mechanism consisting of two fully connected layers is used to capture inter-channel dependencies. The generated channel recalibration vector z′ is defined as:

[0028] ;

[0029] in W1 represents the weights of the first fully connected layer, and W2 represents the weights of the second fully connected layer. This represents the sigmoid function, which maps vector values ​​to the interval [0,1]. z is the input vector, and F... ` For the first level of linear transformation, F `` For the second level of linear transformation, z ` The channel weight vectors are recalibrated by performing element-wise multiplication with the input to achieve corresponding channel-level feature recalibration. Furthermore, residual connections are introduced to alleviate the vanishing gradient problem and enhance model stability.

[0030] Step S43: Feature extraction consists of three DSCR and ISE modules used to extract fault features. The DSCR and ISE modules are stacked alternately to achieve feature extraction and calibration, thereby completing the discriminative fault feature learning more efficiently. Each DSConv layer deploys four 1×3 sub-convolutional modules, and the offset Δ is dynamically updated according to formula (2); the stride of the first sub-convolutional module is set to 1; finally, the fault identification and classification of the feature map are completed by the classifier.

[0031] The present invention has the following beneficial effects:

[0032] 1. This invention proposes a highly efficient and robust deep neural network, called a dynamic sparse residual network, for early inter-turn short circuit fault diagnosis in motors. This method exhibits excellent diagnostic performance under multiple operating conditions, scarce samples, and high noise conditions.

[0033] 2. This invention proposes a novel DSConv convolutional layer with dynamic receptive field and sparse multi-scale characteristics, which effectively captures local detail features and long-range periodic features in current signals; the improved squeezing excitation mechanism enhances fault-related discrimination features and suppresses irrelevant features.

[0034] 3. This invention proposes a general intelligent diagnostic framework applicable to the initial stage of ITSC faults. This framework has been verified on various motor types and does not require modification of the model structure or hyperparameters, demonstrating effective cross-machine adaptability. The model is applicable under different operating speeds, high-noise environments, and limited sample sizes. Attached Figure Description

[0035] Figure 1 : Block diagram of motor inter-turn short circuit fault diagnosis algorithm.

[0036] Figure 2 Dynamic sparse convolution residual module and improved squeezing incentive mechanism.

[0037] Figure 3 : Dynamic sparse convolution.

[0038] Figure 4 Figure 1 shows the fault diagnosis results of the induction motor under different noise levels. (a) shows the fault diagnosis effect of the induction motor, (b) shows the fault diagnosis effect of the five-phase permanent magnet synchronous motor, and (c) shows the fault diagnosis effect of the three-phase permanent magnet synchronous motor.

[0039] Figure 5 Inter-turn fault diagnosis confusion matrix, (a) is the inter-turn fault diagnosis confusion matrix under different noise levels, and (b) is the inter-turn fault diagnosis confusion matrix for different methods.

[0040] Figure 6 T-SNE: Inter-turn fault diagnosis under different noise levels.

[0041] Figure 7 Visualization of Grad-CAM++ with different convolutions.

[0042] Figure 8 Thermal visualization.

[0043] Figure 9 The following are Grad-CAM++ visualizations under different motor conditions: (a) is a Grad-CAM++ visualization of an induction motor, (b) is a Grad-CAM++ visualization of a five-phase permanent magnet synchronous motor, and (c) is a Grad-CAM++ visualization of a three-phase permanent magnet synchronous motor. Detailed Implementation

[0044] The following will be combined with the appendix Figures 1-9 The technical solution of the present invention will be further described below.

[0045] Specific embodiment 1: A method for diagnosing inter-turn short-circuit faults in motors using dynamic sparse residual networks, the algorithm flowchart of which is shown below. Figure 1 As shown in the accompanying drawings, the specific implementation of this embodiment will now be described in detail.

[0046] Step S1: Set up the motor experimental platform. Taps with different short-circuit turns were extracted from the motor stator winding and connected to a short-circuit resistor to simulate inter-turn short-circuit faults of different severity. A total of five different inter-turn short-circuit fault levels were set.

[0047] Step S2: Based on the motor experimental platform, collect motor stator current signals under various operating conditions through the data acquisition system, including data on normal motor operation and different fault severity.

[0048] Step S21, Experimental Platform 1: Experimental fault prototype ABB QABP-90L induction motor with a rated power of 2.2 kW. Wires are led out from the internal windings to the motor casing, and short-circuit resistors of different values ​​are connected to simulate ITSC faults of different severity; under three different operating conditions of 720 rpm, 840 rpm, and 960 rpm, the load is 1 N·m; the number of short-circuit turns is set to 0, 5, and 10 turns, and the short-circuit resistance values ​​are 0.3, 0.6, 0.9, and 1.2 Ω, respectively; a detailed description of the IM dataset is shown in Table I.

[0049] Step S22, Experimental Platform Two: Experimental Fault Prototype Five-Phase Permanent Magnet Synchronous Motor, rated power 2.0kW, number of pole pairs 11. Load set to 1.5N·m; speed set to 300 rpm, 500 rpm and 700 rpm; short-circuit resistance value 0.2Ω, short-circuit turns set to 0, 1, 2, 3 and 5 turns respectively; detailed description of the dataset is shown in Table I.

[0050] Step S23, Experimental Platform Three: The three-phase permanent magnet synchronous motor dataset provided by the Korea Advanced Institute of Science and Technology (KAIST) is a publicly available dataset for inter-turn short-circuit fault diagnosis; the PMSM has a rated power of 1.0kW, a rated speed of 3000r / min, and 4 pole pairs; the load is set to 1.5N·m; a short-circuit resistance is artificially introduced between windings 1 and 2. Five sets of samples with short-circuit resistances of 0, 2, 3, 4, and 6Ω were selected for experimental verification; a detailed description of the dataset is shown in Table 1.

[0051] Table 1 shows the experimental dataset:

[0052] ;

[0053] Step S3: Organize the sample dataset from Step 2, dividing it into a training set and a test set. To fully utilize the data and extract more effective features, data augmentation and normalization methods are employed. Equal-interval sliding window data augmentation technology is used to perform overlapping sampling on the original data, resulting in the sample dataset. 500 samples are set for each working condition, with each sample containing 2048 data points. The data collected in Step 2 is then integrated, and the data from different working conditions are divided into a training set and a test set in a 4:1 ratio.

[0054] Step S4: Design an intelligent diagnostic model for extracting inter-turn short-circuit fault diagnosis based on PyTorch. The aforementioned dynamic sparse convolutional residual network model includes a dynamic sparse convolutional residual module, an improved squeeze-excitation mechanism, and a classification layer. First, to enable the convolutional filter to more flexibly focus on the local and long-range dependencies of multi-scale features in time-series signals, we construct a novel sparse convolutional paradigm with a dynamic offset Δ. By combining multiple non-overlapping convolutional modules, convolutional kernels with differentiated receptive fields are formed, thus accommodating both local and long-range dependent features. Each sub-convolutional module learns local representations, while multiple sub-convolutional modules collaboratively learn long-term information in time-series signals. Furthermore, by introducing dynamic offsets between different sub-convolutional kernels... The offset ∆ avoids information redundancy and extracts multi-scale features more efficiently. Secondly, a dynamic sparse convolution residual module is constructed based on sparse convolution. The convolutional layer adopts a dual convolution mode combining point convolution and DSConv to decouple channel and temporal feature representations. Specifically, PWConv is used to expand the channel dimension to enrich feature representation and integrate cross-channel information. DSConv focuses on extracting discriminative temporal features. In DSConv, each input channel is independently convolved through a dedicated filter, and the input and output feature maps maintain the same spatial resolution to avoid information loss. Then, an improved squeeze-excitation module is used to adaptively enhance fault-related discriminative features. Finally, the results are output through a classification layer.

[0055] To enable convolutional filters to more flexibly focus on the local and long-range dependencies of multi-scale features in time-series signals, thereby improving feature extraction quality, we construct a sparse convolution paradigm with a dynamic receptive field. A dynamic offset Δ is introduced between sub-convolutional kernels to avoid information redundancy and extract multi-scale features more efficiently. The unpadded dynamic sparse convolution can be represented as:

[0056] ;

[0057] in, Let i be the output signal or feature map value at position i of the sparse convolution, where i is the index of the output signal and x is the input signal. Here, represents the weights of the sparse convolution kernel, and s represents the stride of the convolution operation. For the output The offset applied to the input signal by the b-th sparse connection, where m is the depth of the sparse convolution kernel and n is the width of the sparse convolution kernel; a new set of dynamic offsets is generated each time a sliding convolution with stride s is performed; sub-convolutional blocks do not overlap and are arranged sequentially to ensure the continuity and smoothness of time series features; to avoid receptive field overlap, the offset difference between adjacent sub-convolutional blocks must be at least greater than the size n of the sub-convolution kernel; a specific offset ∆ i,j The network dynamically generates the input x; based on DSConv, a Dynamic Sparse Convolutional Residual (DSCR) module is constructed, defined as:

[0058] ;

[0059] Among them W PWConv and W DSConv represents the weights of PWConv and DSConv respectively; BN represents the batch normalization operation; σ represents the LReLU nonlinear activation function; x is the initial input feature map; w is the set of all learnable parameters in the model; x1 is the first intermediate feature map; F is the residual function; y is the final output feature map; Conv_1 is the first convolution operation; and Conv_2 is the second convolution operation.

[0060] Step S42: ITSC fault diagnosis data mainly consists of single-channel time-series signals, and fault features are extracted through multi-channel filters within the network. Therefore, the feature information carried by different channels plays different roles in fault identification. To learn more fault-related discrimination information from the input time-series signals, we propose an ISE module to adaptively recalibrate the weights of each feature channel. During feature compression, hybrid multi-head pooling is introduced to generate multi-scale channel statistics, replacing simple global average pooling; specifically, feature dimensions are compressed simultaneously through dual-branch multi-head mean pooling and multi-head max pooling, and the information flow is balanced using a learnable weight factor λ. The generated statistics... It can be represented as:

[0061] ;

[0062] Wherein represents Multi-head pooling window size, Indicates the number of heads. For learnable weight factors, The feature statistics generated for the c-th channel, where i is the iteration index. The length of the i-th element, where j is the internal iteration index. Input data points, To sum all x_i(j) values ​​in the i-th element, To find the maximum value of all x_i(j) values ​​in the i-th element, the global statistic is described by the piecewise maximum, while the highest value among the piecewise means represents the local statistic. To fully utilize the information aggregated during the squeezing process, a simple gating mechanism consisting of two fully connected layers is used to capture inter-channel dependencies. The generated channel recalibration vector z′ is defined as:

[0063] ;

[0064] in This represents the weights of the first fully connected layer. Indicates the weights of the second fully connected layer. This represents the sigmoid function, which maps vector values ​​to the interval [0,1]. z is the input vector, and F... ` For the first level of linear transformation, F `` For the second-level linear transformation, z′ can also be considered as a channel weight vector. Element-wise multiplication with the input achieves recalibration of the corresponding channel-level features. Furthermore, residual connections are introduced to alleviate the vanishing gradient problem and enhance model stability. The output of the ISE module can be expressed as: ;

[0065] Where X is the input feature vector. Y represents the channel feature correction value, and Y represents the output result. This refers to matrix multiplication.

[0066] Step S43: Feature extraction consists of three DSCR and ISE modules used to extract fault features. The DSCR and ISE modules are stacked alternately to achieve feature extraction and calibration, thereby completing the discriminative fault feature learning more efficiently. Each DSConv layer deploys four 1×3 sub-convolutional modules, and the offset Δ is dynamically updated according to formula (2); the stride of the first sub-convolutional module is set to 1; finally, the fault identification and classification of the feature map are completed by the classifier.

[0067] Step S44: The output of the bidirectional gated recurrent unit is fed into the fully connected layer and then into the Softmax output layer to obtain the final diagnostic result. The specific parameters of the dynamic sparse convolutional residual network model are shown in Table 2.

[0068] Table 2 shows the network parameter settings:

[0069] ;

[0070] Step S5 involves training the network model using the training set. During training, the cross-entropy loss function is used to calculate the error between the predicted and true labels. This error is then used for backpropagation to adjust the network parameters and structure. The training process is repeated until the set number of iterations is reached, at which point training stops. The model training uses the Adam optimization algorithm, with 100 iterations, an initial learning rate of 1e-4, and a batch size of 64.

[0071] Step S6: Perform performance testing for fault diagnosis using a test set. Input the proportionally segmented test samples into the network to obtain the final diagnostic results. Add white Gaussian noise with different signal-to-noise ratios (SNR) to the original current signal to simulate signal interference in a real environment and verify the robustness of the model. The signal-to-noise ratio is defined as follows:

[0072] ;

[0073] in and These represent the original signal power and the noise power, respectively.

[0074] Step S7, the effect of the proposed method on experimental platform one is as follows: Figure 4 As shown in (a), the proposed method achieves high prediction accuracy in various noise scenarios, even under high noise conditions with a signal-to-noise ratio of -10 dB, the diagnostic accuracy still reaches 96.24%. Furthermore, this model outperforms other models under all noise ratio conditions. Figure 5 (a) presents the confusion matrix results for the six methods at SNR=-10 dB. Compared to other models, the proposed method outperforms all noise ratio conditions. It still achieves 96.24% even under high noise conditions of SNR=-10 dB. Compared to other models, this model has a smaller number of misdiagnosed samples in each health state, with a misdiagnosis rate of less than 10%. Figure 6The paper demonstrates the visualization of the hidden layer output using t-SNE under different noise levels. The feature distributions of different fault states exhibit discrete and significantly overlapping characteristics. The proposed method displays relatively clear boundaries under all fault states, indicating that the feature representations learned by the proposed model have stronger intra-class aggregation and inter-class separation capabilities, providing a more reliable basis for the accurate identification and classification of fault states.

[0075] Step S8, the effect of the proposed method on experimental platform two is as follows: Figure 4 As shown in (b), the proposed method achieves an accuracy of 99.47% on the FPPMSSM dataset with no noise additives. Figure 5 As shown in (b), the algorithm performance decreases due to noise, and the false alarm rate is positively correlated with the noise intensity. This is particularly evident in labels 1Z and 2Z, where the features of the less faulty turns 1 and 2 are similar to those of healthy states, and the noise further blurs the classification boundaries. The core feature clusters of each category remain relatively intact under low to high noise conditions. Even at a signal-to-noise ratio of -10 dB, the feature distributions of each category are still recognizable, with an accuracy of 94.92%. Overall, this method demonstrates excellent diagnostic performance and robustness in noisy environments.

[0076] Step S9, the effect of the proposed method on experimental platform three is as follows: Figure 4 As shown in (c), the accuracy of the proposed model remained above 95% under all test conditions.

[0077] Step S9, Visualization Analysis. Heatmaps are used to visualize local and global feature information to analyze the mechanism by which the model learns potential fault characteristics. Taking healthy samples and 1-turn short-circuit fault samples in the FPPMSM dataset as an example, ... Figure 8 As shown, the heatmaps of healthy and faulty samples exhibit differentiated periodic texture features, with faulty samples generally displaying higher feature values ​​than healthy samples. Healthy samples exhibit a uniform and stable feature distribution. The anomalous information caused by faults is captured and amplified during network feature extraction, resulting in more significant textures and feature intensities observed in the heatmaps. Grad-CAM++ is used to construct a mapping relationship between fault types and signal patterns. Figure 9 The visualization results of Grad-CAM++ for three datasets are shown. The color levels in the graph correspond to the weight values ​​output by Grad-CAM++, describing the contribution of each time interval in the current waveform to the fault classification decision. Figure 9 It is evident that the model exhibits differentiated attention to different fault types. The overall activation intensity of fault signals is higher than that of healthy signals. Due to differences in operating conditions and sampling factors, the signals in the three datasets show varying levels of glitches and jitter. Notably, compared to... Figure 7 Standard convolution and dilated convolution produce different results. Figure 7In dynamic sparse convolution, the highlighted regions are more concentrated in the peak areas where the signal waveform changes drastically. This indicates that the model can learn more vivid feature representations, thus dynamically adapting to diagnostic tasks in different scenarios. Furthermore, the highlighted regions exhibit a periodic trend, suggesting that the model can continuously enhance its feature learning ability in key segments.

[0078] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0079] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for diagnosing initial inter-turn short-circuit faults in motors using dynamic sparse convolutional residual networks, characterized in that, Including the following steps: Step S1: Based on the constructed motor experimental platform, collect the motor stator current under various operating conditions through the data acquisition system, including normal motor operation data and data on different fault severity. Step S2: Integrate the data collected in Step 2 and perform data preprocessing, dividing the dataset into training set and test set according to the proportion; Step S3: Based on PyTorch, design a dynamic sparse convolutional residual network model for extracting inter-turn short-circuit fault diagnosis. The constructed dynamic sparse convolutional residual network model includes a dynamic sparse convolutional residual module, an improved squeezing excitation module, and a classification layer. Step S4: Use the training set to train the model. During the training process, the cross-entropy loss function is introduced to calculate the error between the predicted label and the true label. At the same time, the error is used for backpropagation to adjust the parameters and structure of the neural network. The experiment is repeated to train the model until the set number of iterations is reached and then training is stopped. Step S5: Test the trained model using the test set. Input the test samples into the network model to obtain the accuracy of the diagnostic results.

2. The method for diagnosing initial inter-turn short-circuit faults in a motor using a dynamic sparse convolutional residual network according to claim 1, characterized in that, In step 1, when collecting the motor current, in order to simulate faults of different severity, the stator winding ITSC fault is injected through the stator winding inter-turn short circuit control module (ITSC). Phase A is set as the fault phase. Wires are led out from the internal winding to the motor casing and short-circuit resistors of different values ​​are connected to simulate ITSC faults of different severity.

3. The method for diagnosing initial inter-turn short-circuit faults in a motor using a dynamic sparse convolutional residual network according to claim 1, characterized in that, In step 2, when partitioning the dataset, data augmentation and normalization are employed to fully utilize the data and extract more effective features; Each sample is the original signal. A subsequence, starting index is Defined as: ； in, For the first Each sample corresponds to the original signal. subsequence of For the first Each sample corresponds to The number of signal points; when When windows overlap, the number of samples increases; when At that time, the windows do not overlap; by using a sliding window, a series of samples are generated. ,in This represents the total number of generated samples. The normalization operation is defined as follows: ； Where X represents the original signal, X p Let X represent the preprocessed signal, max(X) be the maximum value of the original signal, and min(X) be the minimum value of the original signal. The original data is sampled by overlapping sampling using the data augmentation method of equal interval sliding window to obtain the sample dataset. 1000 samples are set for each working condition, and let α be the proportion of training samples. The sample length is 2048 data points.

4. The method for diagnosing initial inter-turn short-circuit faults in a motor using a dynamic sparse convolutional residual network according to claim 1, characterized in that, The specific process of step S3 is as follows: Step S31: To enable the convolutional filter to more flexibly focus on the local and long-range dependencies of multi-scale features in the time series signal, thereby improving the feature extraction quality, a sparse convolutional paradigm with a dynamic receptive field is constructed. A dynamic offset Δ is introduced between the sub-convolutional kernels to avoid information redundancy and extract multi-scale features more efficiently. The unpadded dynamic sparse convolution can be represented as: ； in, Let i be the output signal or feature map value at position i of the sparse convolution, where i is the index of the output signal and x is the input signal. Here, represents the weights of the sparse convolution kernel, and s represents the stride of the convolution operation. For the output The offset applied to the input signal by the b-th sparse connection, where m is the depth of the sparse convolution kernel and n is the width of the sparse convolution kernel; a new set of dynamic offsets is generated each time a sliding convolution with stride s is performed; sub-convolutional blocks do not overlap and are arranged sequentially to ensure the continuity and smoothness of time series features; to avoid receptive field overlap, the offset difference between adjacent sub-convolutional blocks must be at least greater than the size n of the sub-convolution kernel; a specific offset ∆ i,j The network dynamically generates the data based on the input x; a dynamic sparse convolution residual module DSCR is constructed based on dynamic sparse convolution DSConv, defined as: ； Among them W PWConv and W DSConv represents the weights of point convolution PWConv and dynamic sparse convolution DSConv, respectively; BN represents the batch normalization operation; σ represents the LReLU non-linear activation function; x is the initial input feature map; w is the set of all learnable parameters in the model; x1 is the first intermediate feature map; F is the residual function; y is the final output feature map; Conv1 is the first convolution operation; and Conv2 is the second convolution operation. Step S32: The ITSC fault diagnosis data mainly consists of single-channel time-series signals. Fault features are extracted through multi-channel filters within the network. Therefore, the feature information carried by different channels plays different roles in fault identification. To learn more fault-related discrimination information from the input time-series signals, an improved squeeze excitation module (ISE) adaptive method is proposed to recalibrate the weights of each feature channel. During feature compression, hybrid multi-head pooling is introduced to generate multi-scale channel statistics, replacing simple global average pooling. Specifically, the feature dimension is compressed simultaneously through dual-branch multi-head mean pooling and multi-head max pooling, and the information flow is balanced by a learnable weight factor λ, generating statistics. It can be represented as: ； Among them, it means Multi-head pooling window size, Indicates the number of heads. For learnable weight factors, The feature statistics generated for the c-th channel, where i is the iteration index. The length of the i-th element, where j is the internal iteration index. Input data points, To sum all x_i(j) values ​​in the i-th element, To find the maximum value of all x_i(j) values ​​in the i-th element; the global statistic is described by the piecewise maximum value, while the highest value among the piecewise means represents the local statistic; to fully utilize the aggregated information during the squeezing process, a simple gating mechanism consisting of two fully connected layers is used to capture the inter-channel dependencies; the generated channel recalibration vector z′ is defined as: ； in W1 represents the weights of the first fully connected layer, and W2 represents the weights of the second fully connected layer. This represents the sigmoid function, which maps vector values ​​to the interval [0,1]. z is the input vector, and F... ` For the first level of linear transformation, F `` For the second level of linear transformation, z ` The channel weight vector is used to recalibrate the corresponding channel-level features by performing element-wise multiplication with the input; in addition, residual connections are introduced to alleviate the gradient vanishing problem and enhance model stability. Step S33: Feature extraction consists of three DSCR and ISE modules, used to extract fault features. The DSCR and ISE modules are stacked alternately to achieve feature extraction and calibration, thereby completing the discriminative fault feature learning more efficiently. Each DSConv layer deploys four 1×3 sub-convolutional modules, and the offset Δ is dynamically updated according to the formula; the stride of the first sub-convolutional module is set to 1; finally, the fault identification and classification of the feature map are completed through the classifier.

Images