Permanent magnet generator eccentric fault diagnosis method based on multi-channel feature fusion

Abstract

The invention belongs to the field of permanent magnet generator fault diagnosis, and provides a permanent magnet generator eccentricity fault diagnosis method based on multi-channel feature fusion, and the method mainly comprises the steps: processing a one-dimensional three-phase current signal through matrix transformation, obtaining a recursive fusion graph, and carrying out the fault diagnosis through employing an improved ConvNeXt network; according to the method, the two-dimensional recursive fusion image, the CBAM attention mechanism and the FPN feature pyramid network are introduced, so that the fault feature extraction and diagnosis capability is remarkably improved; the two-dimensional recursion plot enhances the visualization effect of fault features through the time sequence dependence of space structure coding signals, and is beneficial to clearly identifying current fluctuation caused by eccentric faults; the CBAM mechanism optimizes focusing of the model on key feature areas and channels through a space and channel attention mechanism, and improves the accuracy of fault diagnosis; the FPN enables the network to recognize fault features on different scales through multi-scale feature fusion, and further improves the accuracy and robustness of fault diagnosis.

Description

Technical Field

[0001] The invention belongs to the field of permanent magnet generator fault diagnosis, and in particular relates to a permanent magnet generator eccentricity fault diagnosis method based on multi-channel feature fusion. Background Art

[0002] Permanent magnet synchronous generators are ideal for low-speed operation due to their high power density and high efficiency. However, due to the complexity of the working environment, permanent magnet synchronous generators are prone to a variety of faults, especially mechanical faults, among which rotor eccentricity faults account for the largest proportion. The unbalanced magnetic pull, vibration, and pulsating torque caused by rotor eccentricity faults may lead to a series of problems such as equipment vibration, mechanical wear, insulation damage, and even cause casualties in severe cases. Therefore, timely diagnosis of rotor eccentricity faults is of great significance to ensure the safe operation of the generator and reduce downtime.

[0003] With the rapid development of industrial Internet, computer technology and artificial intelligence, traditional manual fault diagnosis is gradually being replaced by intelligent fault diagnosis, and data-driven fault diagnosis methods are gradually becoming one of the solutions in the field of fault diagnosis of rotating machinery such as generators. Through the analysis and modeling of a large amount of historical fault data, data-driven fault diagnosis methods can realize automatic classification and diagnosis of faults. Traditional data-driven methods usually rely on feature extraction of raw data, and then use algorithms such as support vector machines, multi-layer perceptrons or decision trees to classify faults. Convolutional neural networks have become an indispensable research tool for data-driven fault diagnosis methods due to their excellent feature extraction capabilities.

[0004] In recent years, the convolutional neural network (CNN) has developed the ConvNeXt model, which has shown efficient performance while maintaining simplicity. However, in its own design, ConvNeXt may ignore the correlation between information at different spatial scales; at the same time, the feature extraction process of one-dimensional data is time-consuming and complex, and the impact of different features on the results varies greatly. In addition, feature dimensionality reduction may lead to the loss of important information, further affecting the accuracy of diagnosis.

[0005] In general, the application of convolutional neural networks in the research of permanent magnet generator fault diagnosis is expected to provide a more intelligent and efficient fault diagnosis solution for the generator industry. To this end, an improved ConvNeXt permanent magnet synchronous generator eccentricity fault diagnosis method based on multi-channel feature fusion is proposed. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a permanent magnet generator eccentricity fault diagnosis method based on multi-channel feature fusion to solve the problems in the prior art. The technical solution adopted by the present invention is: A method for diagnosing eccentricity fault of a permanent magnet generator based on multi-channel feature fusion comprises the following steps: Step S1: data acquisition, collecting three-phase current signals of a permanent magnet generator eccentricity fault simulation unit; Step S2: construct a recursive two-dimensional fusion graph, convert the three-phase current time series data into a two-dimensional recursive matrix according to the phase, and map the three two-dimensional recursive matrices to the RGB three channels to form a color image; Step S3: Add an attention mechanism and fuse the CBAM attention module before the LayerScale layer in the ConvNeXt Block module of the ConvNeXt network to obtain an improved ConvNeXt Block module; Step S4: Construct a multi-scale feature fusion pyramid, extract feature maps from each stage module, obtain feature maps of multiple scales through upsampling operations and lateral connections, and concatenate the feature maps into a feature vector after global average pooling; Step S5: Network construction, building a new network based on the original ConvNeXt network structure; Step S6: Network training, using the training set to train the new network, and using the test set to test it, to obtain an eccentricity fault classification and recognition model; Step S7: Identify the eccentricity fault of the permanent magnet generator through the eccentricity fault classification and identification model.

[0007] Further, step S2 includes: S21: The mutual information coefficient method is used to calculate the mutual information between the current signal and the corresponding delayed signal, and the optimal delay time is adaptively selected according to the first local minimum value of the mutual information: ; In the formula, is the mutual information; n is the current signal length; The delay time is set between [1, 100]; For the t Current signal of each point; is the probability density of the current signal; is the probability density of the delayed signal; is the joint probability density of the current signal and the delayed signal; Use optimal delay Reconstruct the phase space of the time series and obtain the reconstruction vector X ( t ): ; In the formula, m is the embedding dimension, is the optimal delay time; for each time point t , by calculating the reconstruction vector X ( t ) and the reconstructed vectors at other time points to determine their proximity in the phase space: ; In the formula, X i ( t ), X i ( t + k ) are the time points t and t + k The corresponding reconstruction vector X ( t ), X ( t + k ) i components; if the distance between two adjacent points in low-dimensional space is small, but its distance increases in high-dimensional space, these points are considered to be false neighbors: ; In the formula, factor is constant; as the dimension m As the dimension increases, the proportion of false neighbors gradually decreases and tends to be stable after a certain dimension. is the optimal embedding dimension; After the phase space is reconstructed, the recursiveness of the phase space points is determined by calculating the Euclidean distance between the reconstructed vectors, thereby constructing a two-dimensional recursive matrix: ; in ε is the minimum distance threshold, the element of the recursive matrix R R i,j Represents the time series i Point and j Recursive relations between points; H is the Heaviside function. If the distance between two points is less than or equal to ε, they are considered recursive points and marked as 1; otherwise, they are marked as 0. S22: Normalize the obtained two-dimensional recursive matrices one by one, map them to the three channels of R, G, and B, and fuse them into a color image.

[0008] Further, step S3 includes: S31: The ConvNeXt block module includes a deep convolution layer and two normal convolution layers; it also includes a LayerScale layer, which is located after the last convolution layer; S32: before the Layer Scale layer, a CBAM attention module is set, and the CBAM attention module includes a feature input terminal, a channel attention module, a spatial attention module, and a feature output terminal; The feature input end converts the input feature map into an initial feature map and passes it to the channel attention module. The channel attention module enhances the channel information of the initial feature map and passes the weight to the first feature mapping node. The first feature mapping node multiplies the channel attention weight with the initial feature map channel by channel to generate a feature map enhanced by channel attention and outputs it to the spatial attention module. The spatial attention module processes the feature map in the spatial domain and outputs the spatial attention weight to the second feature mapping node. The second feature mapping node multiplies the spatial attention weight with the feature map enhanced by channel attention pixel by pixel to generate the final feature map. The specific processing process of the CBAM attention module is as follows: For the input feature map ,in H and W are the height and width respectively, C is the number of channels; first, global average pooling and global maximum pooling are performed on the channel dimension of the feature map to capture the average information and maximum activation value of each channel. The two one-dimensional channel description vectors generated by global average pooling and global maximum pooling are No. n The channel expressions are expressed as: ; The two one-dimensional channel description vectors generated after the two poolings are sent to the shared MLP for processing and added, and then normalized by the Sigmoid activation function to generate the channel attention weight: ; Multiply the generated channel attention weights by the input features channel by channel to obtain the feature map of the first feature mapping node: ; The feature map of the first feature map node Average pooling and maximum pooling operations are performed on the spatial dimension to generate two two-dimensional feature maps. , these two-dimensional feature maps are located at ( i , j ) position expression is: ; The two feature maps are then concatenated in the channel dimension and a 7×7 convolution operation is performed to generate a spatial attention map. M s : ; Where [;] represents the channel connection operation, which multiplies the spatial attention weight by the input feature position by position to obtain the feature map of the second feature map node: .

[0009] Further, step S4 includes: S41: ConvNeXt network consists of 1 Stem, 4 Stages and Classifier Head; S42: Construct a multi-scale feature fusion pyramid and extract features from the four stage outputs of the ConvNeX network, which are represented as F 1, F 2, F 3. F 4, among which F 1 is the finest-grained feature map, F 4 is the coarsest-grained feature map; First, for each layer of feature map F l , first perform a 1×1 convolution operation to compress the feature dimension: ; Then the interpolation method is used to perform upsampling operation, so that The size and same: ; The upsampled high-level feature map is added and fused with the low-resolution feature map of the current layer to obtain a fused new feature map: ; Perform a 3×3 convolution operation on the fused feature map: ; Then, a global average pooling operation is used to transform the feature maps of each scale into a global feature vector: ; By concatenating the global feature vectors of each scale, we get the final feature vector containing all scale information: .

[0010] Further, step S6 includes: S61: Input image size, downsample through convolutional layer, and reduce feature map; S62: The feature map is sequentially downsampled by a factor of 2 through four new stage modules. S63: Take the output feature map of each stageF 1. F 2. F 3. F 4. Then, the output feature maps of each stage are aligned through a common convolution layer; the four feature maps are globally averaged pooled to obtain the output feature map , , , ; Finally, , , , Concatenate to get the final feature map ; S63: Feature Map Finally, the linear layer is used for output and type judgment.

[0011] The present invention has the following beneficial effects: 1. The two-dimensional recursive fusion graph can effectively enhance the fault characteristics, especially by encoding the temporal dependencies of the signal through the spatial structure. This allows the fault characteristics to be more clearly presented in the image data. The image form not only retains the temporal characteristics in the original data, but also enhances the data visualization and information expression capabilities through color coding, thereby improving the model's capabilities in feature extraction and anomaly detection.

[0012] 2. The introduction of the CBAM attention mechanism significantly improves the model's attention to key features. The spatial attention mechanism helps the network automatically focus on key areas in the two-dimensional recursive graph, such as local anomalies in the current waveform (such as spikes, current fluctuations, etc.). This mechanism can identify local fault features in the image and improve the accuracy of fault diagnosis. The channel attention mechanism helps the network identify which channels are most critical for eccentric fault detection, enhances the feature expression of these channels, and enables the network to pay more attention to signals that are valuable for diagnosis. Through this mechanism, the network can focus on the most critical feature areas, improving the accuracy and robustness of fault identification.

[0013] 3. Multi-scale feature fusion improves diagnostic capabilities. The introduction of the FPN feature pyramid network helps the network to better handle multi-scale features, especially in eccentric fault diagnosis, where fault features of different scales may be manifested in different frequency bands or time periods of the signal. FPN provides a multi-scale feature representation by fusing feature maps at different levels, enabling the network to identify and judge fault features at different scales, thereby improving the accuracy of overall fault diagnosis. At the same time, the spatial and channel attention mechanisms can guide the network to focus on the most important feature areas and channels, further enhancing the network's perception of key fault features. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The figure is a flow chart of the method of the present invention.

[0015] Figure 2 The present invention provides a process for converting three-phase current into a two-dimensional recursive fusion graph.

[0016] Figure 3 The ConvNeXt structure diagram and the CBAM attention mechanism in the ConvNeXtBlock position diagram provided for the embodiment of the present invention.

[0017] Figure 4 This is the structural diagram of the CBAM attention mechanism of the present invention, including the spatial attention mechanism and the channel attention mechanism structure.

[0018] Figure 5 This is a structural diagram of the ConvNeXt of the present invention after the multi-scale feature fusion method is applied. DETAILED DESCRIPTION

[0019] The following will be combined with the embodiments of the present invention Figure 1-Figure 5 , the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0020] The embodiment of the present invention provides an improved ConvNeXt permanent magnet generator eccentricity fault diagnosis method based on multi-channel feature fusion, please refer to Figure 1 , which is a flow chart of an improved ConvNeXt permanent magnet generator eccentricity fault diagnosis method based on multi-channel feature fusion according to an embodiment of the present invention. The method comprises the following steps: S1. Fault diagnosis data collection. The collected fault data is the three-phase current signal under different operating states of the generator, and the different operating states of the generator include the normal operating state of the generator, the radial static eccentricity fault operating state, the radial dynamic eccentricity fault operating state, the radial dynamic and static mixed eccentricity fault operating state, the axial static eccentricity fault operating state and the axial-radial three-dimensional composite static eccentricity fault operating state.

[0021] Exemplarily, this embodiment simultaneously collects the current signals of the three phases A, B, and C of the permanent magnet synchronous generator in six operating states, namely, normal operating state, radial static eccentricity fault operating state, radial dynamic eccentricity fault operating state, radial dynamic and static mixed eccentricity fault operating state, axial static eccentricity fault operating state, and axial-radial three-dimensional composite static eccentricity fault operating state, through three current probes at a sampling frequency of 5 kHz for 30 seconds each, and removes the beginning and end of 5 seconds each.

[0022] S2: Constructing a recursive 2D fusion graph The three-phase current time series data of each phase is converted into a two-dimensional matrix through a recursive graph, and the two-dimensional matrix corresponding to the three-phase currents of A, B, and C is mapped to the RGB three channels to form a color image.

[0023] In this example, step S2 is specifically divided into the following parts: S21: Convert the three-phase current time series data into a two-dimensional recursive matrix according to the phase. First, the mutual information coefficient method is used to calculate the mutual information between the current signal and the corresponding delayed signal, and the optimal delay time is adaptively selected according to the first local minimum value of the mutual information: ; In the formula, is the mutual information; n is the current signal length; The delay time is set between [1, 100]; For the t Current signal of each point; is the probability density of the current signal; is the probability density of the delayed signal; is the joint probability density of the two signals; Calculate different delays starting from the minimum value τ The mutual information under the value , we can get images of time delay. Usually, the mutual information increases with The delay corresponding to this local minimum value increases and gradually decreases, and a local minimum value appears at a certain point. It is the optimal delay, which represents the minimum redundancy of the signal while retaining the dynamic information of the signal.

[0024] Use optimal delay Reconstruct the phase space of the time series and obtain the reconstruction vector X ( t ): ; In the formula, m is the embedding dimension, is the optimal delay time. For each time point t , by calculating the reconstruction vector X ( t ) and the reconstructed vectors at other time points to determine their proximity in the phase space: ; In the formula, X i ( t ), X i( t + k ) are the time points t and t + k The corresponding reconstruction vector X ( t ), X ( t + k ) i components. If in a low-dimensional space (embedding dimension m ) is small, but in high-dimensional space (embedding dimension m +1), their distance increases, these points are considered to be false neighbors: ; Where factor is a constant (usually set to 1.2) that measures the growth rate of the distance. m As the dimension increases, the proportion of false neighbors should gradually decrease and stabilize after a certain dimension. is the optimal embedding dimension.

[0025] After the phase space is reconstructed, the recursiveness of the phase space points is determined by calculating the Euclidean distance between the reconstructed vectors, thereby constructing a two-dimensional recursive matrix: ; in ε is the minimum distance threshold, the element of the recursive matrix R R i,j Represents the time series i Point and j The recursive relationship between the points. H is the Heaviside function, which is used to determine whether two phase space points are within a specified distance threshold ε, thereby marking the recurrence relationship. Specifically, if the distance between two points is less than or equal to ε, they are considered to be recurrence points and marked as 1; otherwise, they are marked as 0.

[0026] S22: Map the three two-dimensional recursive matrices to the RGB three channels to form a color image. The obtained three-phase current recursive matrices are normalized one by one, mapped to the R, G, and B three channels and fused into a color image with a size of 224×224×3. After converting all fault category data into recursive fusion graphs, 80% of the data of each fault is used for the training set and 20% for the test set. And ensure that the data used for the test set does not exist in the training set and does not participate in any model training process. Table 1 shows the number of each type in the entire dataset.

[0027] Table 1 Number of various operating status types of permanent magnet synchronous generators Logo Generator operating status quantity 0 normal 2000 1 Radial static eccentricity 2000 2 Radial eccentricity 2000 3 Radial dynamic and static compound eccentric 2000 4 Axial static eccentricity 2000 5 Radial-axial three-dimensional mixed static eccentricity 2000 S3: Adding attention mechanism The CBAM (Convolutional Block Attention Module) attention module is fused before the Layer Scale layer in the ConvNeXt Block module of the ConvNeXt network to obtain the improved ConvNeXtBlock module. Through this fusion, the network can more effectively focus on the key parts in the feature map and enhance the expressiveness of the model.

[0028] In this example, step S3 is specifically divided into the following parts: S31: ConvNeXt block is the core component of the ConvNeXt network and its basic building block, with high reusability. ConvNeXt block includes multiple layers, including a deep convolution layer (DW convolution layer, convolution kernel is 7*7, stride is 1) and two ordinary convolution layers (convolution kernel is 1*1, stride is 1). In addition, it also includes LayerScale layer (layer scaling mechanism) and Drop mechanism. The LayerScale layer follows the last convolution layer, and obtains an adjustable scaling factor through learning to scale the output of the convolution operation to enhance the training stability and representation ability of the network; while the Drop mechanism is at the end, and sets the output of the neuron to zero with a certain probability, thereby avoiding the model's excessive dependence on specific neurons.

[0029] S32: Convolutional attention module based on CBAM attention mechanism Figure 4 As shown, this module is inserted before the Layer Scale layer in each ConvNeXt Block. The module includes a feature input terminal, a channel attention module, a spatial attention module, and a feature output terminal. The feature input terminal converts the input feature map into an initial feature map and passes it to the channel attention module. The channel attention module enhances the channel information of the initial feature map and passes the weight to the first feature mapping node. The first feature mapping node multiplies the channel attention weight with the initial feature map channel by channel to generate a feature map enhanced by the channel attention and outputs it to the spatial attention module. The spatial attention module processes the feature map in the spatial domain and outputs the spatial attention weight to the second feature mapping node, which multiplies the spatial attention weight with the feature map enhanced by the channel attention pixel by pixel to generate the final feature map. The specific processing process of the CBAM attention module is as follows: For the input feature map ,in H and W are the height and width respectively, Cis the number of channels. First, global average pooling and global maximum pooling are performed on the channel dimension of the feature map to capture the average information and maximum activation value of each channel. The two one-dimensional channel description vectors generated by global average pooling and global maximum pooling are No. n The channel expressions are expressed as: ; The two one-dimensional channel description vectors generated after the above two poolings are sent to the shared MLP for processing and added. The shared MLP is not only identical in structure, but also in weights (i.e., the same weights are used to process the two pooling results). First, the two one-dimensional vectors (length C ) is sent to a fully connected layer, which expands its dimension from C to a higher dimension D; then the output of the fully connected layer is usually transformed nonlinearly using the ReLU activation function; finally, the vector after ReLU activation is sent to the second fully connected layer, reduced back to the original dimension C, and the two output vectors are added. Then it is normalized by the Sigmoid activation function to generate the channel attention weight: ; Finally, the generated channel attention weights are multiplied by the generated channel attention weights and input features channel by channel to obtain the feature map of the first feature mapping node: ; The feature map of the first feature map node Average pooling and maximum pooling operations are performed on the spatial dimension to generate two two-dimensional feature maps. , these two-dimensional feature maps are located at ( i , j ) position expression is: ; The two feature maps are then concatenated in the channel dimension and a 7×7 convolution operation is performed to generate a spatial attention map. M s : ; In the formula, [;] represents the channel connection operation. Multiply the spatial attention weight with the feature map of the first feature map node position by position to obtain the feature map of the second feature map node: .

[0030] S4: Constructing a multi-scale feature fusion pyramid Feature maps are extracted from each ConvNeXt Block module, and feature maps of multiple scales are obtained through upsampling operations and lateral connections. The feature maps are then concatenated into a feature vector after global average pooling.

[0031] In this example, step S4 is specifically divided into the following parts: S41: ConvNeXt is a deep convolutional neural network, which consists of 1 Stem, 4 Stages and Classifier Head.

[0032] S42: Construct a multi-scale feature fusion pyramid, such as Figure 5 The features extracted from the output of the four stages of the ConvNeX network (each stage corresponds to a different feature scale) are expressed as F 1, F 2, F 3. F 4. These feature maps represent information of different scales, among which F 1 is the finest-grained feature map, F 4 is the coarsest-grained feature map.

[0033] First, for each layer of feature map F l (in l ∈{1,2,3,4), first perform a 1×1 convolution operation to compress the feature dimension, thereby generating a more compact and informative feature representation. This operation can help reduce the number of channels, suppress redundant information, and retain useful feature information: ; Next, an upsampling operation is performed to increase the high-level feature map to a higher resolution so that it is consistent with the feature size of the current layer. The upsampling operation usually uses interpolation to make The size and same: ; The upsampled high-level feature map is added and fused with the low-resolution feature map of the current layer to obtain a fused new feature map: ; It not only maintains low-level detail information (such as edges and textures), but also enhances high-level semantic information (such as object categories and contextual information). Perform a 3×3 convolution operation on the fused feature map: ; Then, a global average pooling operation is used to transform the feature maps of each scale into a global feature vector: ; This operation averages the pixels at all positions in the feature map, and finally reduces the feature map to a long vector, which represents the global information of that scale. Finally, by concatenating the global feature vectors of each scale, the final feature vector containing information of all scales is obtained: .

[0034] S5: Network Construction A new network is built based on the original ConvNeXt network structure, and its structure is as follows Figure 5 As shown; S6: Network Training The new network is trained using the training set and tested using the test set to obtain an eccentricity fault classification and recognition model. In this example, step S6 specifically includes the following contents: S61: The input image size is 224×224×3. It is first downsampled through a convolution layer with a kernel of 4×4 and a stride of 4, and the feature map is reduced to 56×56×96. This step preliminarily extracts low-level features in the image and reduces the spatial dimension; S62: The feature map passes through four new stage modules (named stage1, stage2, stage3, and stage4 respectively). The stage1, stage2, stage3, and stage4 modules are stacked 3 times, 3 times, 9 times, and 3 times respectively. The ConvNeXt Block module with the CBAM attention mechanism is added. The stage1 module does not have a downsampling layer. The stage2, stage3, and stage4 modules are preceded by a downsampling layer (consisting of a normalization layer and a convolution layer with a convolution kernel of 2×2 and a stride of 2) for 2x downsampling, downsampling the feature map to 28×28×192, 14×14×284, and 7×7×768 respectively. S63: Get the output feature map of each Stage F 1. F 2. F 3. F 4. Then, the output feature maps of each stage are aligned through a common convolution layer (the convolution kernel is 1×1 and the stride is 1), that is, , , , , whose sizes are 56×56×256, 28×28×256, 14×14×256, and 7×7×256 respectively; followed by the feature maps of stage2, stage3, and stage4 ( , , ) are upsampled to the same size as the previous layer (stage1, stage2, stage3) by interpolation method, that is, , , , whose sizes are 56×56×256, 28×28×256, and 14×14×256 respectively; then the feature maps of stage1, stage2, and stage3 ( , , ) and the feature maps after upsampling of stage2, stage3, and stage4 ( , , ) are added element by element to generate the fused feature map , , ; After upsampling and additive fusion, the four multi-scale feature maps ( , , , ) further extracts and enhances feature information through convolution operation (convolution kernel is 3×3, step size is 1), and smoothes the spatial information. The output feature map is , , , , the sizes are 56×56×256, 28×28×256, 14×14×256, and 7×7×256 respectively; these four feature maps are globally averaged pooled to obtain the output feature map , , , , the sizes are 1×1×256, 1×1×256, 1×1×256, 1×1×256 respectively; finally , , , Concatenate to get the final feature map , the size is 1×1×1024.

[0035] S63: Feature Map Finally, the linear layer is used for output and type judgment.

[0036] Step S7: Identify the eccentricity fault of the permanent magnet generator by using the eccentricity fault classification and identification model In this example, Table 2 shows the impact of adjusting different modules on the experimental results. It can be seen that after adding the RP module, the accuracy increased from 85.33% to 99.08%, and after adding the CBAM module, the accuracy increased from 93.29 to 96.71. The accuracy of the network model after the final improvement of the present invention is 99.08%.

[0037] Table 2 Accuracy, F1, and AUC data of some models Model Accuracy(%) F1 AUC RP+ConvNeXt 93.29 0.9327 0.9597 RP+CBAM-ConvNeXt 96.71 0.9670 0.9802 1D+CBAM-ConvNeXt-PFN 85.33 0.8535 0.9120 RP+CBAM-ConvNeXt-PFN 99.08 0.9907 0.9945 Table 3 shows the experimental results of different attention mechanisms. RP+CBAM-ConvNeXt exceeds RP+CAM-ConvNeXt and RP+SAM-ConvNeXt in terms of accuracy, F1 value and AUC, indicating that the CBAM mechanism performs better than CAM and SAM in this task. This may be because CBAM can better combine channel attention and spatial attention mechanisms, enhance the ability of feature expression, and thus improve the overall performance of the model.

[0038] Table 3 Comparison of experimental results of different attention mechanisms Model Accuracy(%) F1 AUC RP+CAM-ConvNeXt 94.79 0.9477 0.9688 RP+SAM-ConvNeXt 94.54 0.9453 0.9673 RP+CBAM-ConvNeXt 96.71 0.9670 0.9802 Table 4 shows the comparison of experimental results without network. In the comparison of different network architectures, RP+ConvNeXt outperforms RP+GoogLeNet and RP+ResNet, and achieves the best results in accuracy, F1 value and AUC. This shows that ConvNeXt is more effective in the task of diagnosing eccentric faults of permanent magnet synchronous generators, which may be due to its design that is more in line with the needs of current image processing tasks and has stronger feature extraction capabilities. The performance of GoogLeNet and ResNet in this task is relatively close, slightly lower than ConvNeXt.

[0039] Table 4 Experimental results of different networks Model Accuracy(%) F1 AUC RP+GoogLeNet 91.67 0.9164 0.9500 RP+ResNet 91.75 0.9172 0.9505 RP+ConvNeXt 93.29 0.9327 0.9597 In summary, the above-mentioned embodiment is based on the permanent magnet synchronous generator eccentricity fault diagnosis method based on two-dimensional recursive fusion graph and CBAM-ConvNeXt-FPN. This method significantly improves the fault feature extraction and diagnosis capabilities by introducing two-dimensional recursive fusion graph, CBAM attention mechanism and FPN feature pyramid network. The two-dimensional recursive graph encodes the signal timing dependency through spatial structure, enhances the visualization effect of fault features, and helps to more clearly identify current fluctuations caused by eccentricity faults; the CBAM mechanism optimizes the model's focus on key feature areas and channels through spatial and channel attention mechanisms, improves the feature expression ability of local anomalies and important channels, and thus improves the accuracy of fault diagnosis; FPN enables the network to identify fault features at different scales through multi-scale feature fusion, further improving the accuracy and robustness of fault diagnosis; finally, the effectiveness of the improvement is proved by various comparative experiments, and the final accuracy rate is 99.08%. The improved network model is also compared with other methods to prove that it has the best accuracy.

[0040] The embodiments described above are only descriptions of the preferred modes of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations, modifications, and substitutions made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A permanent magnet generator eccentricity fault diagnosis method based on multi-channel feature fusion, characterized in that: The following steps are involved: Step S1: data acquisition, collecting three-phase current signals of a permanent magnet generator eccentricity fault simulation unit; Step S2: construct a recursive two-dimensional fusion graph, convert the three-phase current time series data into a two-dimensional recursive matrix according to the phase, and map the three two-dimensional recursive matrices to the RGB three channels to form a color image; Step S3: Add an attention mechanism and fuse the CBAM attention module before the Layer Scale layer in the ConvNeXt Block module of the ConvNeXt network to obtain an improved ConvNeXt Block module; Step S4: Construct a multi-scale feature fusion pyramid, extract feature maps from each stage module, obtain feature maps of multiple scales through upsampling operations and lateral connections, and concatenate the feature maps into a feature vector after global average pooling; Step S5: Network construction, building a new network based on the original ConvNeXt network structure; Step S6: Network training, using the training set to train the new network, and using the test set to test it, to obtain an eccentricity fault classification and recognition model; Step S7: Identify the eccentricity fault of the permanent magnet generator through the eccentricity fault classification and identification model.

2. The method for diagnosing eccentricity fault of a permanent magnet generator based on multi-channel feature fusion according to claim 1 is characterized in that: Step S2 includes: S21: The mutual information coefficient method is used to calculate the mutual information between the current signal and the corresponding delayed signal, and the optimal delay time is adaptively selected according to the first local minimum value of the mutual information: ； In the formula, is the mutual information; n is the current signal length; The delay time is set between [1, 100]; For the t Current signal of each point; is the probability density of the current signal; is the probability density of the delayed signal; is the joint probability density of the current signal and the delayed signal; Use optimal delay Reconstruct the phase space of the time series and obtain the reconstruction vector X ( t ): ； In the formula, m is the embedding dimension, is the optimal delay time; for each time point t , by calculating the reconstruction vector X ( t ) and the reconstructed vectors at other time points to determine their proximity in the phase space: ； In the formula, X i ( t ), X i ( t + k ) are the time points t and t + k The corresponding reconstruction vector X ( t ), X ( t + k ) i components; if the distance between two adjacent points in low-dimensional space is small, but its distance increases in high-dimensional space, these points are considered to be false neighbors: ； In the formula, factor is constant; as the dimension m As the dimension increases, the proportion of false neighbors gradually decreases and tends to be stable after a certain dimension. is the optimal embedding dimension; After the phase space is reconstructed, the recursiveness of the phase space points is determined by calculating the Euclidean distance between the reconstructed vectors, thereby constructing a two-dimensional recursive matrix: ； in ε is the minimum distance threshold, the element of the recursive matrix R R i,j Represents the time series i Point and j Recursive relations between points; H is the Heaviside function. If the distance between two points is less than or equal to ε, they are considered recursive points and marked as 1; otherwise, they are marked as 0. S22: Normalize the obtained two-dimensional recursive matrices one by one, map them to the three channels of R, G, and B, and fuse them into a color image.

3. The method for diagnosing eccentricity fault of a permanent magnet generator based on multi-channel feature fusion according to claim 1 is characterized in that: Step S3 includes: S31: The ConvNeXt block module includes a deep convolution layer and two normal convolution layers; it also includes a LayerScale layer, which is located after the last convolution layer; S32: before the Layer Scale layer, a CBAM attention module is set, and the CBAM attention module includes a feature input terminal, a channel attention module, a spatial attention module, and a feature output terminal; The feature input end converts the input feature map into an initial feature map and passes it to the channel attention module. The channel attention module enhances the channel information of the initial feature map and passes the weight to the first feature mapping node. The first feature mapping node multiplies the channel attention weight with the initial feature map channel by channel to generate a feature map enhanced by channel attention and outputs it to the spatial attention module. The spatial attention module processes the feature map in the spatial domain and outputs the spatial attention weight to the second feature mapping node. The second feature mapping node multiplies the spatial attention weight with the feature map enhanced by channel attention pixel by pixel to generate the final feature map. The specific processing process of the CBAM attention module is as follows: For the input feature map ,in H and W are the height and width respectively, C is the number of channels; first, global average pooling and global maximum pooling are performed on the channel dimension of the feature map to capture the average information and maximum activation value of each channel. The two one-dimensional channel description vectors generated by global average pooling and global maximum pooling are No. n The channel expressions are expressed as: ； The two one-dimensional channel description vectors generated after the two poolings are sent to the shared MLP for processing and added, and then normalized by the Sigmoid activation function to generate the channel attention weight: ； Multiply the generated channel attention weights by the input features channel by channel to obtain the feature map of the first feature mapping node: ； The feature map of the first feature map node Average pooling and maximum pooling operations are performed on the spatial dimension to generate two two-dimensional feature maps. , these two-dimensional feature maps are located at ( i , j ) position expression is: ； The two feature maps are then concatenated in the channel dimension and a 7×7 convolution operation is performed to generate a spatial attention map. M s : ； Where [;] represents the channel connection operation, which multiplies the spatial attention weight by the input feature position by position to obtain the feature map of the second feature map node: 。 4. The method for diagnosing eccentricity fault of a permanent magnet generator based on multi-channel feature fusion according to claim 1 is characterized in that: Step S4 includes: S41: ConvNeXt network consists of 1 Stem, 4 Stages and Classifier Head; S42: Construct a multi-scale feature fusion pyramid and extract features from the four stage outputs of the ConvNeX network, which are represented as F 1, F 2, F 3. F 4, among which F 1 is the finest-grained feature map, F 4 is the coarsest-grained feature map; First, for each layer of feature map F l , first perform a 1×1 convolution operation to compress the feature dimension: ； Then the interpolation method is used to perform upsampling operation, so that The size and same: ； The upsampled high-level feature map is added and fused with the low-resolution feature map of the current layer to obtain a fused new feature map: ； Perform a 3×3 convolution operation on the fused feature map: ； Then, a global average pooling operation is used to transform the feature maps of each scale into a global feature vector: ； By concatenating the global feature vectors of each scale, we get the final feature vector containing all scale information: 。 5. The method for diagnosing eccentricity fault of a permanent magnet generator based on multi-channel feature fusion according to claim 4 is characterized in that: Step S6 includes: S61: Input image size, downsample through convolutional layer, and reduce feature map; S62: The feature map is sequentially downsampled by a factor of 2 through four new stage modules. S63: Take the output feature map of each stage F 1. F 2. F 3. F 4. Then, the output feature maps of each stage are aligned through a common convolution layer; the four feature maps are globally averaged pooled to obtain the output feature map , , , ; Finally, , , , Concatenate to get the final feature map ; S63: Feature Map Finally, the linear layer is used for output and fault type judgment.

Images