A motor fault diagnosis method based on SK-CA-ResNet and stepped unfreezing fine-tuning

Abstract

The present application relates to a kind of motor fault diagnosis methods based on SK-CA-ResNet and step thaw fine-tuning, belong to motor fault diagnosis field.The method aims at solving the problem that the early weak fault feature extraction of permanent magnet synchronous motor stator winding is insufficient, and the characteristic frequency band is offset under variable working condition and the model generalization is poor, and the problem that the target domain sample is scarce, and the cross-domain migration is prone to negative migration.The technical scheme is: obtaining motor voltage signal and converting into time-frequency diagram by continuous wavelet transform;Residual network SK-CA-ResNet model is constructed by fusing selective convolution kernel and coordinate attention mechanism, to dynamically adapt multi-scale feature and lock time-frequency information;First, train the model using source domain data, then use step partial thaw strategy.The method can effectively improve the diagnosis accuracy of early weak fault, and adapt to the characteristic offset under variable working condition.

Description

Technical Field

[0001] This invention belongs to the field of motor fault diagnosis and relates to a motor fault diagnosis method based on SK-CA-ResNet and step-by-step defrosting fine-tuning. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs), with their high power density, high efficiency, and excellent dynamic response performance, are widely used in key areas such as new energy vehicles, industrial automation equipment, and renewable energy power generation systems. Among various motor faults, stator winding inter-turn short circuits are one of the most common and insidious. If not diagnosed in time, they can easily lead to serious secondary disasters and safety accidents. However, in actual industrial production environments, PMSMs often operate under dynamically changing conditions, which causes a significant shift in the distribution of fault characteristics. Furthermore, industrial sites face the challenge of scarce fault samples. These factors combined make traditional model-driven fault diagnosis methods, due to their poor robustness to changes in operating conditions and external disturbances, difficult to achieve accurate generalization in practical applications.

[0003] Therefore, more adaptable data-driven methods have gradually become the mainstream research direction. However, at the diagnostic network level, existing data-driven methods generally face two core challenges. First, when faced with minor faults of varying degrees in the early stages, the network struggles to effectively handle multi-scale differences in features, and time and frequency information is often processed in isolation, resulting in a lack of multi-scale dynamic adaptation capability and high-precision time-frequency feature locking capability, which easily leads to underfitting. Second, regarding the feature frequency band shift and target domain sample scarcity caused by cross-operational conditions, some existing cross-domain transfer diagnostic methods, while attempting to reduce the distribution difference between the source and target domains to some extent, are prone to negative transfer due to the difficulty in effectively preserving the general knowledge already learned by the source domain model during the transfer process, leading to a serious technical bottleneck in the model's generalization ability. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a motor fault diagnosis method based on SK-CA-ResNet and step-by-step defrosting fine-tuning.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A motor fault diagnosis method based on SK-CA-ResNet and step-by-step defrosting fine-tuning, the method comprising the following steps: S1: Acquire the motor's operating signals under source domain and target domain operating conditions; S2: Perform a continuous wavelet transform (CWT) on the running signal obtained from S1 to generate the corresponding two-dimensional time-frequency spectrum; S3: Construct a diagnostic model of a residual network (SK-CA-ResNet) that integrates selective kernel-coordinate attention mechanism; S4: Train the SK-CA-ResNet diagnostic model using time-frequency spectrum samples under source domain conditions to obtain a pre-trained source domain model; S5: A step-by-step partial unfreezing strategy is adopted to fine-tune the pre-trained source domain model using a small number of time-frequency spectrum samples under the target domain conditions, so as to obtain a fault diagnosis model adapted to the target domain conditions. S6: After processing the motor operation signal to be diagnosed by S2, input it into the fault diagnosis model adapted to the target domain operating conditions, and output the fault diagnosis result.

[0006] Furthermore, in S1, the operating signal is the stator voltage signal.

[0007] Furthermore, in S2, the transformation formula for the continuous wavelet transform is: ,in, As a scale factor, The translation factor is... For time variables, For the input running signal, It is the complex conjugate of the mother wavelet function.

[0008] Furthermore, in S3, the SK-CA-ResNet diagnostic model uses the deep residual network ResNet as its backbone and embeds selective convolution kernel SK modules and coordinate attention CA modules in its residual blocks.

[0009] Furthermore, the selective convolution kernel SK module generates the output feature matrix in the following manner. : Convolution operations with kernels of different sizes are performed on the input feature map to generate a multi-scale feature flow; The aggregated features are obtained by summing the elements of the multi-scale feature stream. ; Aggregation features Perform global average pooling along the spatial dimension to obtain channel statistics. ,in These represent the height and width of the feature map, respectively. For channel indexing; Based on the channel statistics Attention weights corresponding to features at different scales are calculated using the Softmax function. ; The multi-scale feature stream is weighted and summed with the corresponding attention weights to obtain the output feature matrix. .

[0010] Furthermore, the coordinate attention (CA) module processes the input feature map in the following way. Perform calibration: One-dimensional global pooling is performed on the input feature map along both the height and width directions to obtain the feature vector along the height direction. and the feature vector in the width direction ; Convolution and nonlinear transformation are performed on the concatenated feature vectors to generate highly directional attention weights. Attention weights in width direction ; The height-direction attention weights and width-direction attention weights are multiplied by the original input feature map to obtain the calibrated output feature map. .

[0011] Furthermore, in S5, the step-by-step partial thawing strategy specifically includes: Freeze the shallow network parameters in the pre-trained source domain model used to extract general low-level features; Unfreeze the deep network parameters and classifier layer parameters used to extract task-specific semantic features in the pre-trained source domain model; Force unfreeze the parameters of all coordinate attention (CA) modules in the pre-trained source domain model.

[0012] Furthermore, in step S5, the parameters for thawing are fine-tuned using a differential learning rate, wherein the deep network parameters and the coordinate attention (CA) module parameters use a first learning rate, and the classifier layer parameters use a second learning rate that is higher than the first learning rate.

[0013] Furthermore, in S4 and S5, the goal of the training and fine-tuning is to minimize the loss function between the fault category prediction results output by the SK-CA-ResNet diagnostic model and the true labels.

[0014] Furthermore, the source domain operating condition is that the motor operates at a first speed, and the target domain operating condition is that the motor operates at a second speed different from the first speed.

[0015] The beneficial effects of this invention are as follows: (1) By constructing a residual network model that integrates selective convolution kernels and coordinate attention mechanisms, this invention can dynamically adapt to multi-scale feature differences in fault signals and accurately lock key information in time-frequency spectra. This significantly enhances the feature extraction capability for early weak inter-turn short-circuit faults in permanent magnet synchronous motors, thereby effectively improving the accuracy and reliability of diagnosis and reducing the risk of missed detections and false alarms.

[0016] (2) To address the problem of fault harmonic frequency band shifting in the time-frequency diagram caused by actual variable operating conditions, the step-by-step partial unfreezing fine-tuning strategy proposed in this invention has unique advantages. This strategy enables the model to adaptively focus on the fault feature region that has shifted under the new operating conditions by forcibly unfreezing and recalibrating the coordinate attention module in the network, thus overcoming the disadvantage of traditional fine-tuning methods that are prone to failure when the frequency band shifts.

[0017] (3) By combining rigid locking of low-level general features, targeted unfreezing of high-level semantic features, and differential learning rate mechanism, this invention can still achieve efficient cross-domain knowledge transfer and feature distribution alignment even under conditions of extremely scarce target domain samples. This method retains valuable diagnostic knowledge from the source domain while rapidly adapting to the target domain, fundamentally alleviating overfitting, catastrophic forgetting, and negative transfer phenomena in scenarios with few samples, and providing a practical solution for achieving high reliability health status assessment of motors under conditions of scarce industrial field samples.

[0018] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 Flowchart for fine-tuning the source domain partial unfreezing; Figure 2 Time-frequency plots of CWT at 20Hz and 40Hz for different ITSC severity levels; Figure 3 This is the confusion matrix of the five diagnostic methods on the test set. Detailed Implementation

[0020] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0021] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0022] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0023] 1. Continuous wavelet transform Continuous Wavelet Transform (CWT) is a multi-resolution mathematical analysis tool. Compared to the fixed time-frequency window of the Short-Time Fourier Transform, CWT can effectively address the problem of large multi-scale feature differences in motor fault signals by adaptively scaling and shifting basis functions, and provides excellent display and feature analysis capabilities in both the time and frequency domains.

[0024] For any stator current signal belonging to a square-integrable space Its continuous wavelet transform coefficients are strictly defined as the inner product of the signal and the mother wavelet function: (1) In the formula, a This is the scale factor, corresponding to the reciprocal of the frequency; b The translation factor determines the absolute position of the wavelet window on the time axis; This is the energy normalization coefficient; This represents the complex conjugate of the mother wavelet.

[0025] This invention selects the complex Morlet wavelet, which is formed by coupling a Gaussian envelope with a complex exponential modulation signal. It exhibits optimal localization characteristics in both the time and frequency domains, and its mathematical expression is: (2) in, The bandwidth parameter determines the decay rate of the Gaussian window; The center frequency is used. Complex wavelet transform can extract not only the instantaneous amplitude of a signal but also preserve instantaneous phase information. The magnitude of its output wavelet coefficients is... This allows for a rigorous characterization of the signal at different scales. a With time b The energy density distribution on the constructed two-dimensional phase plane provides rigorous physical input features for subsequent deep two-dimensional convolutional networks. Among these features... For bandwidth parameters, The center frequency is [value]. The magnitudes of its output wavelet coefficients can rigorously characterize the signal at different scales. a With time b The energy density distribution on the constructed two-dimensional phase plane provides physical input features for subsequent deep networks.

[0026] 2. SK-CA-RESNET model To address the challenges of significant multi-scale feature differences and insufficient feature extraction due to the independence of temporal and frequency information in inter-turn short-circuit fault diagnosis, traditional convolutional neural networks (CNNs) suffer from fixed receptive fields and are prone to disrupting the time-frequency manifold topology, hindering accurate diagnosis. Therefore, this invention proposes the SK-CA-ResNet diagnostic model, using a deep residual network (ResNet) as its backbone and deeply fusing selective kernels (SK) and coordinate attention (CA).

[0027] 2.1 ResNet Network As networks become deeper, gradient vanishing and network degradation worsen dramatically. ResNet introduces an identity mapping, transforming implicit nonlinear fitting into explicit residual approximation. Assuming the input is x, the residual network causes the convolutional layers to fit the residual function. Residual block output via cross-layer quick connection. y Reconstructed as: (3) In the formula, This represents the weight matrix. After element-wise addition and activation, it is passed to the next layer. During backpropagation, the chain rule applies: (4) The presence of the constant 1 in the above formula enables the gradient to bypass the nonlinear layer without loss, overcoming the gradient vanishing dilemma and giving the network the potential to extract deep abstract physical features.

[0028] 2.2SK Module The SK module aims to endow the network with the ability to adaptively adjust the local receptive field when faced with features of different scales. First, it generates a multi-scale feature stream through convolutions with different kernel sizes or dilation rates. and Subsequently, the aggregated features are obtained by summing and merging elements one by one. It is then compressed into channel statistics by performing global average pooling (GAP) along the spatial dimension. s , its first c Each element is calculated as follows: (5) Statistic s Feeding a dimensionality reduction network to extract compact features z Based on this, attention gating weights at spatial scales are calculated using Softmax. and The final output feature matrix V is a dynamically soft-weighted sum of information at different scales: (6) The SK module can dynamically determine the size of the receptive field based on the time-frequency bandwidth of the fault signal, which greatly enhances the robustness of the model to distorted fault characteristics under multiple operating conditions.

[0029] 2.3CA Module Traditional channel attention uses global two-dimensional pooling to compress spatial information, directly disrupting the binary heterogeneous physical topology of the time-frequency graph. The CA module decouples the two-dimensional pooling into two independent one-dimensional directional feature codes, endowing the network with robust location awareness. For the... c One channel, height H Horizontal slices and width W The vertical slice output is decoupled as follows: (7) (8) This decoupling perfectly captures the energy intensity of a specific frequency band and the transient envelope of a fault at a specific moment. The features from both directions are concatenated and input into a 1×1 convolution to generate an intermediate map, which is then split and attention weights are obtained through independent convolutions and the sigmoid function. and Finally, multiplicative calibration is performed to complete the element-wise physical domain reconstruction of the feature map: (9) The CA module ensures that the key time-frequency spatial relationships are not lost during the dimensionality reduction process, enabling it to accurately locate weak fault ripples on the time-frequency plane.

[0030] 3. Stepped partial thawing and fine-tuning strategy In deep convolutional neural networks, different depths of network layers extract features at different levels of abstraction: shallow networks extract general low-level features such as edges and textures, exhibiting high cross-domain sharing in time-frequency maps at different rotation speeds; deep networks extract task-specific semantic features, responsible for constructing high-dimensional decision boundaries. When the target domain has few samples, global fine-tuning strategies are prone to overfitting and catastrophic forgetting due to the large parameter space, destroying the physical diagnostic knowledge accumulated by the source domain model.

[0031] Therefore, the present invention proposes a method such as Figure 1 The stepped partial unfreezing strategy shown achieves a perfect balance between knowledge preservation and target domain adaptation by locking asymmetric parameters of different network modules. Specifically, this strategy includes the following three core mechanisms: First, the underlying general knowledge is rigidly locked. During the fine-tuning stage, the parameters of the network backbone starting layer and shallow residual modules are frozen, gradient backpropagation is cut off, the basic texture perception capability of the source domain is inherited, and the number of optimization parameters is reduced to curb overfitting. Second, the high-level semantic features are unfrozen in a targeted manner. The deep residual modules and fully connected classifier layers are unfrozen, and the parameters are updated using a small number of target domain samples to quickly fit the target domain decision boundary. Finally, the CA module is forcibly unfrozen and recalibrated. In response to the fault harmonic frequency band shift caused by the change in rotation speed, the weights of the CA modules at each level are unfrozen, and the target domain time-frequency coordinate attention is relearned through training with a small number of samples to accurately focus on the fault features after the shift.

[0032] To facilitate smooth knowledge transfer, a differential learning rate mechanism is introduced: a very small base learning rate is assigned to the unfrozen deep convolutional and CA modules to protect the feature space structure of the source domain and achieve alignment of feature micro-perturbations; a higher learning rate is assigned to the classifier layer to enable it to respond quickly to the target domain label signal.

[0033] 4. Overall Diagnostic Process for Cross-Domain Transfer Learning like Figure 1 As shown, this invention constructs a complete cross-domain transfer learning diagnostic framework based on SK-CA-RESNET and a step-by-step partial thawing fine-tuning strategy. The specific transfer learning experimental steps are as follows: Step 1: Source Domain Model Training. Sufficient health and fault current signals of varying degrees are collected under a single baseline operating condition. These signals are converted into two-dimensional time-frequency feature maps using CWT. The source domain training set is then input into the initialized SK-CA-ResNet network. The SK module's multi-scale sensing and the CA module's coordinate localization are used to capture deep fault manifolds. After training to convergence, the optimal network weights are saved as a knowledge base for transfer learning.

[0034] Step 2: Target Domain Small Sample Construction. Collect similar current signals under the target operating conditions and perform preprocessing consistent with the source domain. Randomly select 2% of the target domain data as the support set to guide network fine-tuning; the remaining 98% of the data serves as the test set to verify the model's cross-domain generalization ability.

[0035] Step 3: Cross-domain fine-tuning and validation. Instantiate an SK-CA-ResNet network with the same structure as the source domain, load the pre-trained weights, and execute a step-by-step partial unfreezing strategy to freeze the low-level modules and activate the high-level semantic modules and CA modules. Use the support set to drive the model forward propagation, and combine it with differential learning rate for backpropagation fine-tuning. Finally, input the test set into the network to predict the fault category and validate the model performance.

[0036] To verify the results, a high-precision PMSM stator winding inter-turn short-circuit fault dataset was used as an experimental example. The basic prototype used in the experiment was a 36-slot, 8-pole double-layer lap-winding built-in permanent magnet synchronous motor.

[0037] This experimental example selects the voltage signal of the faulted phase as the core analysis object. Two typical operating conditions were chosen: 20Hz (source domain) and 40Hz (target domain). Each condition included the motor's healthy state and five different degrees of inter-turn short-circuit conditions. The fault severity was defined by the short-circuit turns ratio, which were 1.98%, 5.95%, 12.30%, 20.64%, and 45.63%. To focus on the mid-to-low frequency characteristics most significantly affected by the fault and improve computational efficiency, the original 1MHz voltage signal was preprocessed: first, a 5th-order Butterworth low-pass filter was used to filter out signals above 300Hz, and then the equivalent sampling rate was reduced to 2kHz. Based on this, a continuous wavelet transform (CWT) was introduced to convert the one-dimensional voltage signal into a two-dimensional time-frequency spectrum. The CWT time-frequency spectra of the voltage signals of the inter-turn short-circuit faulted phases with different degrees of severity under the 20Hz condition are shown below. Figure 2 As shown.

[0038] To verify the superiority of the proposed SK-CA-ResNet model in extracting time-frequency spectrum features and diagnosing inter-turn short-circuit faults of varying severity, a benchmark comparison experiment was conducted under a single source domain condition (20Hz). The preprocessed sample dataset collected under the 20Hz condition was divided into training, validation, and test sets in an 8:1:1 ratio. Four methods—1D-CNN, 2D-CNN, ResNet-18, and ResNet-SE—were compared with the experimental results. The comparison results and confusion matrix results are shown in Table 1.

[0039] Table 1 Performance Comparison of Each Model

[0040] Figure 3 The confusion matrix of the five diagnostic methods on the test set is shown. To intuitively compare the performance differences of different strategies in typical few-sample scenarios, a detailed comparative experiment was conducted on the diagnostic accuracy of the five transfer strategies on a 40Hz blind test set when the number of single-class fault samples in the target domain reached a set maximum of 15. The experimental results are shown in Table 2.

[0041] Table 2. Comparison of diagnostic accuracy of different transfer strategies when the number of samples in a single class is 15.

[0042] Comparing the experimental results in Table 2, it can be seen that the accuracy of the direct testing scheme is only 17.8%, indicating that the frequency band shift caused by the cross-condition has caused the source domain model to almost completely fail. Even with an increase in the number of target domain samples, the traditional fine-tuning strategy, because it absolutely freezes the fully convolutional network, still adheres to the old 20Hz frequency coordinate system, resulting in an accuracy of only 73.6%. Although full-network fine-tuning gives the network the ability to relearn features, improving the accuracy to 89.5%, the large number of unfreezing parameters under the constraint of only 15 samples still leads to a certain degree of overfitting. The direct training method of the target domain, because it completely abandons prior knowledge and trains from scratch using only 15 samples, is prone to getting trapped in local optima, with an accuracy of 86.2%. In contrast, the step-by-step unfreezing strategy proposed in this invention achieves an accuracy of up to 93.3%, effectively preserving basic visual knowledge while accurately locking the harmonic resonance band under the new condition through the directional reconstruction CA module.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A motor fault diagnosis method based on SK-CA-ResNet and step-by-step defrosting fine-tuning, characterized in that: The method includes the following steps: S1: Acquire the motor's operating signals under source domain and target domain operating conditions; S2: Perform continuous wavelet transform (CWT) on the running signal obtained from S1 to generate the corresponding two-dimensional time-frequency spectrum; S3: Construct a residual network SK-CA-ResNet diagnostic model that integrates selective convolutional kernels and coordinate attention mechanisms; S4: Train the SK-CA-ResNet diagnostic model using time-frequency spectrum samples under source domain conditions to obtain a pre-trained source domain model; S5: A step-by-step partial unfreezing strategy is adopted to fine-tune the pre-trained source domain model using a small number of time-frequency spectrum samples under the target domain conditions, so as to obtain a fault diagnosis model adapted to the target domain conditions. S6: After processing the motor operation signal to be diagnosed by S2, input it into the fault diagnosis model adapted to the target domain operating conditions, and output the fault diagnosis result.

2. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 1, characterized in that: In S1, the operating signal is the stator voltage signal.

3. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 1 or 2, characterized in that: In S2, the transformation formula for the continuous wavelet transform is: ,in, As a scale factor, The translation factor is... For time variables, For the input running signal, It is the complex conjugate of the mother wavelet function.

4. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 1, characterized in that: In S3, the SK-CA-ResNet diagnostic model uses the deep residual network ResNet as its backbone and embeds selective convolution kernel SK modules and coordinate attention CA modules in its residual blocks.

5. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 4, characterized in that: The selective convolution kernel (SK) module generates the output feature matrix in the following manner. : Convolution operations with kernels of different sizes are performed on the input feature map to generate a multi-scale feature flow; The aggregated features are obtained by summing the elements of the multi-scale feature stream. ; Aggregation features Perform global average pooling along the spatial dimension to obtain channel statistics. ,in These represent the height and width of the feature map, respectively. For channel indexing; Based on the channel statistics Attention weights corresponding to features at different scales are calculated using the Softmax function. ; The multi-scale feature stream is weighted and summed with the corresponding attention weights to obtain the output feature matrix. .

6. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 4, characterized in that: The coordinate attention (CA) module processes the input feature map in the following ways. Perform calibration: One-dimensional global pooling is performed on the input feature map along both the height and width directions to obtain the feature vector along the height direction. and the feature vector in the width direction ; Convolution and nonlinear transformation are performed on the concatenated feature vectors to generate highly directional attention weights. Attention weights in width direction ; The height-direction attention weights and width-direction attention weights are multiplied by the original input feature map to obtain the calibrated output feature map. .

7. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 1, characterized in that: In S5, the step-by-step partial thawing strategy specifically includes: Freeze the shallow network parameters in the pre-trained source domain model used to extract general low-level features; Unfreeze the deep network parameters and classifier layer parameters used to extract task-specific semantic features in the pre-trained source domain model; Force unfreeze the parameters of all coordinate attention (CA) modules in the pre-trained source domain model.

8. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 7, characterized in that: In step S5, the parameters for thawing are fine-tuned using a differential learning rate. The parameters of the deep network and the coordinate attention (CA) module use a first learning rate, while the parameters of the classifier layer use a second learning rate that is higher than the first learning rate.

9. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 1, characterized in that: In S4 and S5, the goal of the training and fine-tuning is to minimize the loss function between the fault category prediction results output by the SK-CA-ResNet diagnostic model and the true labels.

10. The motor fault diagnosis method based on SK-CA-ResNet and stepped thawing fine-tuning according to claim 1, characterized in that: The source domain operating condition is that the motor operates at a first speed, and the target domain operating condition is that the motor operates at a second speed, which is different from the first speed.

Images