A multi-channel feature fusion-based eccentric fault diagnosis method for permanent magnet generator

Abstract

The application belongs to the field of permanent magnet generator fault diagnosis, and provides a permanent magnet generator eccentric fault diagnosis method based on multi-channel feature fusion, the application mainly processes one-dimensional three-phase current signals through matrix transformation to obtain a recursive fusion graph, and adopts an improved ConvNeXt network to perform fault diagnosis; the application introduces a two-dimensional recursive fusion graph, a CBAM attention mechanism and an FPN feature pyramid network, significantly improves fault feature extraction and diagnosis capability; the two-dimensional recursive graph encodes the time sequence dependence of signals through a space structure, enhances the visualization effect of fault features, and helps to more clearly identify current fluctuations caused by eccentric faults; the CBAM mechanism optimizes the focusing of the model on key feature areas and channels through a space and channel attention mechanism, improves the accuracy of fault diagnosis; the FPN fuses multi-scale features, so that the network can identify fault features at different scales, further improves the accuracy and robustness of fault diagnosis.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of permanent magnet generator fault diagnosis, and particularly relates to a permanent magnet generator eccentric fault diagnosis method based on multi-channel feature fusion. BACKGROUND

[0002] Permanent magnet synchronous generators are ideal choices for low-speed working conditions 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 various faults, especially mechanical faults, among which the proportion of rotor eccentricity fault is the largest. The unbalanced magnetic pull, vibration and pulsating torque caused by rotor eccentricity fault may cause a series of problems such as equipment vibration, mechanical wear and tear, insulation damage, and even personnel injury accidents in severe cases. Therefore, timely diagnosis of rotor eccentricity fault is of great significance to ensure the safe operation of the generator and reduce downtime.

[0003] Under the background of 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 have gradually become one of the solutions in the field of rotating machinery fault diagnosis such as generators. Through 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 original data, and then use algorithms such as support vector machine, multilayer perception or decision tree to classify faults. Convolutional neural network has become an indispensable research tool for data-driven fault diagnosis methods due to its excellent feature extraction capability.

[0004] In recent years, convolutional neural network (CNN) has developed ConvNeXt model, which exhibits efficient performance while maintaining simplicity. However, ConvNeXt may ignore the relevance between different spatial scale information in its own design; at the same time, one-dimensional data is time-consuming and complex in feature extraction process, and different features have different influences on the results. In addition, feature dimension reduction may cause loss of important information, further affecting the diagnosis accuracy.

[0005] Overall, the application of convolutional neural network in permanent magnet generator fault diagnosis research is expected to provide more intelligent and efficient fault diagnosis solutions for the generator industry. Therefore, an improved ConvNeXt permanent magnet synchronous generator eccentric fault diagnosis method based on multi-channel feature fusion is proposed. SUMMARY

[0006] To solve the above technical problems, the application provides a permanent magnet generator eccentric fault diagnosis method based on multi-channel feature fusion to solve the problems in the prior art, and the technical scheme adopted by the application is as follows:

[0007] A permanent magnet generator eccentric fault diagnosis method based on multi-channel feature fusion, comprising the following steps:

[0008] Step S1: data acquisition, acquiring three-phase current signals of the permanent magnet generator eccentric fault simulation unit;

[0009] 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 RGB three channels to form a color image;

[0010] Step S3: add attention mechanism, fuse CBAM attention module in front of LayerScale layer in ConvNeXt Block module of ConvNeXt network, and get improved ConvNeXt Block module;

[0011] Step S4: build a multi-scale feature fusion pyramid, extract feature maps from each stage module, get feature maps of multiple scales through upsampling operation and horizontal connection, and splice the feature maps into a feature vector after global average pooling;

[0012] Step S5: network building, build a new network based on the original ConvNeXt network structure;

[0013] Step S6: network training, train the new network using the training set, and test it using the test set to obtain an eccentric fault classification and recognition model;

[0014] Step S7: identify the eccentric fault of the permanent magnet generator through the eccentric fault classification and recognition model.

[0015] Further, step S2 comprises:

[0016] S21: adopt the mutual information coefficient method to calculate the mutual information of the current signal and the corresponding delayed signal, and adaptively select the optimal delay time according to the first local minimum value of the mutual information:

[0017] ;

[0018] In the formula, is the mutual information; n is the length of the current signal; is the delay time set between 1 and 100; is the current signal of the first t 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;

[0019] Optimal delay time Phase space reconstruction is performed on the time series to obtain reconstructed vectors X t ):

[0020]

[0021] wherein, m is the embedding dimension, is the optimal delay time; for each time point t , the neighborhood relationship in phase space is determined by calculating the Euclidean distance between the reconstructed vector X t ) and the reconstructed vectors of other time points:

[0022]

[0023] wherein, X i t ), X i t k are the first components of the reconstructed vectors t t k corresponding to the time points X t X t k ; if the distance between two adjacent points in low-dimensional space is small, but the distance between them increases in high-dimensional space, these points are considered to be false neighbors: i

[0024]

[0025] wherein, factor is a constant; with the increase of the dimension m , the proportion of false neighbors gradually decreases and tends to be stable after a certain dimension, which is the optimal embedding dimension

[0026] After phase space reconstruction, the recurrence of phase space points is determined by calculating the Euclidean distance between the reconstructed vectors, so as to construct a two-dimensional recurrence matrix:

[0027]

[0028] wherein ε is the minimum distance threshold, and the element R i,j of the recurrence matrix R represents the distance between the first i point and the first​​​​​​​​​​​​​​​​​j The recursive relationship between points; H For 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.

[0029] S22: Normalize the obtained two-dimensional recursive matrix one by one, map it to the R, G, and B channels, and fuse it into a color image.

[0030] Furthermore, step S3 includes:

[0031] S31: The ConvNeXt block module includes one deep convolutional layer and two regular convolutional layers; it also includes a LayerScale layer, which is placed after the last convolutional layer.

[0032] S32: Set up the CBAM attention module before the Layer Scale layer. The CBAM attention module includes a feature input end, a channel attention module, a spatial attention module, and a feature output end.

[0033] 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 initial feature map with channel information and passes the weights to the first feature mapping node. The first feature mapping node multiplies the channel attention weights with the initial feature map channel by channel to generate a channel-attention-enhanced feature map, which is then output to the spatial attention module. The spatial attention module processes this feature map in the spatial domain and outputs the spatial attention weights to the second feature mapping node. The second feature mapping node multiplies the spatial attention weights with the channel-attention-enhanced feature map pixel by pixel to generate the final feature map. The specific processing procedure of the CBAM attention module is as follows:

[0034] For the input feature map ,in H and W These are the height and width, respectively. C Given the number of channels; firstly, global average pooling and global max pooling are performed on the channel dimensions of the feature map to capture the average information and maximum activation value of each channel. Global average pooling and global max pooling generate two one-dimensional channel description vectors. The n Each channel expression is represented as follows:

[0035] ;

[0036] The two one-dimensional channel description vectors generated after the two pooling operations are fed into a shared MLP and added together. Then, they are normalized using the Sigmoid activation function to generate channel attention weights.

[0037] ;

[0038] The generated channel attention weights are multiplied channel by channel with the input features to obtain the feature map of the first feature mapping node:

[0039] ;

[0040] The feature map of the first feature mapping node Two two-dimensional feature maps are generated by performing average pooling and max pooling operations in the spatial dimension, respectively. These two two-dimensional feature maps are located at ( i , j The expression for the position is:

[0041] ;

[0042] Then, these two feature maps are concatenated along the channel dimension and a spatial attention map is generated through a 7×7 convolution operation. M s :

[0043] ;

[0044] In the formula, [;] represents the channel connection operation, which multiplies the spatial attention weights with the input features position by position to obtain the feature map of the second feature mapping node:

[0045] .

[0046] Furthermore, step S4 includes:

[0047] S41: The ConvNeXt network consists of 1 Stem, 4 Stages, and a Classifier Head;

[0048] S42: Construct a multi-scale feature fusion pyramid, extracting features from the outputs of the four stages of the ConvNeX network, represented as follows: F 1, F 2, F 3, F 4, of which F 1 is the finest-grained feature map. F 4 is the coarsest-grained feature map;

[0049] First, for each layer of feature maps F l First, a 1×1 convolution operation is performed to compress the feature dimension:

[0050] ;

[0051] Next, an upsampling operation is performed using interpolation, so that... Size and same:

[0052] ;

[0053] The upsampled high-level feature map is then added to and fused with the low-resolution feature map of the current layer to obtain a new fused feature map:

[0054] ;

[0055] Perform a 3×3 convolution operation on the fused feature map:

[0056] ;

[0057] Then, a global average pooling operation is used to transform the feature map at each scale into a global feature vector:

[0058] ;

[0059] By concatenating the global feature vectors from each scale, we obtain the final feature vector containing information from all scales:

[0060] .

[0061] Furthermore, step S6 includes:

[0062] S61: Input image size, downsampled through convolutional layers to reduce the feature map size;

[0063] S62: The feature map is downsampled by a factor of 2 through four new stage modules in sequence;

[0064] S63: Extract the output feature map of each stage F 1. F 2. F 3. F 4. Then, each stage's output feature map is channel-aligned using a regular convolutional layer; these four feature maps are then subjected to global average pooling to obtain the final output feature map. , , , Finally, , , , The final feature map is obtained by splicing. ;

[0065] S63: Feature Map Finally, the output is generated after the linear layer, and the type is determined.

[0066] The present invention has the following beneficial effects:

[0067] 1. Two-dimensional recursive fusion graphs can effectively enhance fault features, especially by encoding the temporal dependencies of signals through spatial structure. This allows fault features to be more clearly represented in image data. The image format not only preserves the temporal features of the original data but also enhances the visualization and information expression capabilities of the data through color encoding, thereby improving the model's capabilities in feature extraction and anomaly detection.

[0068] 2. The introduction of the CBAM attention mechanism significantly enhances the model's focus on key features. The spatial attention mechanism helps the network automatically focus on key regions in the two-dimensional recursive graph, such as local anomalies in the current waveform (e.g., spikes, current fluctuations). This mechanism can identify local fault features in the image, improving the accuracy of fault diagnosis. The channel attention mechanism helps the network identify which channels are most critical for eccentricity fault detection, enhancing the feature representation of these channels and enabling the network to focus more on signals valuable for diagnosis. Through this mechanism, the network can focus on the most critical feature regions, improving the accuracy and robustness of fault identification.

[0069] 3. Multi-scale Feature Fusion Enhances Diagnostic Capabilities. Introducing the Feature Pyramid Network (FPN) helps the network better handle multi-scale features, especially in eccentricity fault diagnosis, where fault features at different scales may manifest in different frequency bands or time periods of the signal. FPN provides a multi-scale feature representation by fusing feature maps from different levels, enabling the network to identify and judge fault features at different scales, thereby improving the overall accuracy of fault diagnosis. Simultaneously, spatial and channel focus mechanisms guide the network to focus on the most important feature regions and channels, further enhancing the network's ability to perceive key fault features. Attached Figure Description

[0070] Figure 1 This is a flowchart of the method of the present invention.

[0071] Figure 2 The present invention provides a process for converting three-phase current into a two-dimensional recursive fusion diagram.

[0072] Figure 3 The diagram shows the ConvNeXt structure and the location of the CBAM attention mechanism in the ConvNeXtBlock, as provided in the embodiments of the present invention.

[0073] Figure 4 This is a structural diagram of the CBAM attention mechanism of the present invention, including the spatial attention mechanism and the channel attention mechanism structure.

[0074] Figure 5 This is a structural diagram of ConvNeXt after its application based on a multi-scale feature fusion method. Detailed Implementation

[0075] The following will be based on embodiments of the present invention. Figures 1-5 The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0076] This invention provides an improved method for diagnosing eccentricity faults in ConvNeXt permanent magnet generators based on multi-channel feature fusion. Please refer to [link to relevant documentation]. Figure 1 This is a flowchart 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 includes the following steps:

[0077] S1. Fault Diagnosis Data Acquisition. The acquired fault data consists of three-phase current signals under different generator operating conditions. These different generator operating conditions include normal generator operation, radial static eccentricity fault operation, radial dynamic eccentricity fault operation, radial dynamic-static mixed eccentricity fault operation, axial static eccentricity fault operation, and axial-radial three-dimensional composite static eccentricity fault operation.

[0078] For example, in this embodiment, the current signals of the three phases A, B, and C of the permanent magnet synchronous generator are simultaneously collected for 30 seconds each through three current probes at a sampling frequency of 5 kHz. These signals represent six different operating states of the generator: normal operation, radial static eccentricity fault, radial dynamic eccentricity fault, radial dynamic-static mixed eccentricity fault, axial static eccentricity fault, and axial-radial three-dimensional composite static eccentricity fault. The first and last 5 seconds are then removed.

[0079] S2: Construct a recursive two-dimensional fusion graph

[0080] The three-phase current timing data of each phase is converted into a two-dimensional matrix through a recursive graph, and the two-dimensional matrices corresponding to the three-phase currents of A, B, and C are mapped to the RGB three channels to form a color image.

[0081] In this example, step S2 is specifically divided into the following parts:

[0082] S21: Convert the three-phase current timing data into a two-dimensional recursive matrix according to phase. First, use the mutual information coefficient method to calculate the mutual information between the current signal and the corresponding delayed signal. Adaptively select the optimal delay time based on the first local minimum of the mutual information.

[0083] ;

[0084] In the formula, For mutual information content; n The length of the current signal; The delay time set between [1, 100]; For the first t Current signal at each point; Let be the probability density of the current signal; This represents the probability density of the delayed signal. Let be the joint probability density of the two signals;

[0085] Calculate different delays starting from the minimum value. τ Mutual information under value This allows us to obtain individual images related to the delay. Typically, the mutual information increases with... It increases and then gradually decreases, eventually reaching a local minimum at some point. The delay corresponding to this local minimum is... It is the optimal delay, which represents the minimum redundancy of the signal while preserving the dynamic information of the signal.

[0086] Use optimal delay The time series is reconstructed in phase space to obtain the reconstructed vector. X ( t ):

[0087] ;

[0088] In the formula, m For the embedding dimension, This is the optimal delay time. For each time point... t By calculating the reconstructed vector X ( t The Euclidean distance between the reconstructed vectors at different time points and other time points is used to determine their proximity in phase space.

[0089] ;

[0090] In the formula, X i ( t ), X i ( t + k () are time points. t and t + k corresponding reconstructed vector X ( t ), X ( t + k )i Each component. If in a low-dimensional space (embedding dimension)... m In a dimensional space, the distance between two adjacent points is small, but in a high-dimensional space (embedded dimension), the distance between two adjacent points is small. m If the distance between them increases in +1), then these points are considered to have a pseudo-nearest neighbor relationship:

[0091] ;

[0092] In the formula, `factor` is a constant (usually set to 1.2) used to measure the rate of increase in distance. As the dimension increases... m As the dimension increases, the proportion of false nearest neighbors should gradually decrease and stabilize after a certain dimension, at which point the dimension... That is the optimal embedding dimension.

[0093] After reconstructing the phase space, 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:

[0094] ;

[0095] in ε The elements of the recursion matrix R are the minimum distance threshold. R i,j Represents the time series number 1 i The point and the first j The recursive relationship between points. H The Heaviside function is used to determine whether two points in phase space are within a specified distance threshold ε, thus marking the recursion relationship. Specifically, 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.

[0096] S22: Map the three two-dimensional recursive matrices to the RGB three channels to form a color image. Normalize each of the obtained three-phase current recursive matrices, map them to the R, G, and B channels, and fuse them into a single color image of size 224×224×3. After converting all fault category data into a recursive fusion graph, allocate 80% of the data for each fault category to the training set and 20% to the test set. 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 quantity of each type in the entire dataset.

[0097] Table 1. Number of various operating state types for permanent magnet synchronous generators

[0098] Identification Generator operating state Number 0 Normal 2000 1 Radial static eccentricity 2000 2 Radial dynamic eccentricity 2000 3 Radial dynamic-static combined eccentricity 2000 4 Axial static eccentricity 2000 5 Radial-axial three-dimensional mixed static eccentricity 2000

[0099] S3: Enhanced attention mechanisms

[0100] The ConvNeXt Block module is improved by fusing the CBAM (Convolutional Block Attention Module) attention module before the Layer Scale layer in the ConvNeXt Block module. This fusion allows the network to focus more effectively on key parts of the feature map, enhancing the model's expressive power.

[0101] In this example, step S3 is specifically divided into the following parts:

[0102] S31: The ConvNeXt block is the core component and basic building block of the ConvNeXt network, possessing high reusability. The ConvNeXt block comprises multiple layers, specifically one deep convolutional layer (DW convolutional layer, 7x7 kernel, stride 1) and two regular convolutional layers (1x1 kernel, stride 1). Additionally, it includes a LayerScale layer (layer scaling mechanism) and a Drop mechanism. The LayerScale layer follows the last convolutional layer, learning an adjustable scaling factor to scale the output of the convolutional operation, enhancing the network's training stability and representational power. The Drop mechanism, located last, sets the neuron's output to zero with a certain probability, thus preventing the model from over-relying on specific neurons.

[0103] S32: Constructing a convolutional attention module based on the CBAM attention mechanism, as follows Figure 4 As shown, this module is inserted before the Layer Scale layer in each ConvNeXt Block. This module includes a feature input, a channel attention module, a spatial attention module, and a feature output. The feature input transforms the input feature map into an initial feature map and passes it to the channel attention module. The channel attention module enhances the initial feature map with channel information and passes the weights to the first feature mapping node. The first feature mapping node multiplies the channel attention weights with the initial feature map channel-by-channel to generate a channel-attention-enhanced feature map, which is then output to the spatial attention module. The spatial attention module processes this feature map in the spatial domain and outputs the spatial attention weights to the second feature mapping node, which multiplies the spatial attention weights with the channel-attention-enhanced feature map pixel-by-pixel to generate the final feature map. The specific processing procedure of the CBAM attention module is as follows:

[0104] For the input feature map ,in H and W These are the height and width, respectively. CLet be the number of channels. First, global average pooling and global max pooling are performed on the channel dimensions of the feature map to capture the average information and maximum activation value of each channel. Global average pooling and global max pooling generate two one-dimensional channel description vectors. The n Each channel expression is represented as follows:

[0105] ;

[0106] The two one-dimensional channel description vectors generated after the two pooling operations are fed into a shared MLP for processing and then summed. The shared MLP is not only structurally identical, but its weights are also shared (i.e., the same weights are used to process the two pooling results). First, the two one-dimensional vectors (of length 1) are processed... C The vector is fed into a fully connected layer, which expands its dimension from C to a higher dimension D. Then, the ReLU activation function is typically used to perform a non-linear transformation on the output of the fully connected layer. Finally, the ReLU-activated vector is fed into a second fully connected layer, shrinking it back to the original dimension C, and the two output vectors are added together. Then, it is normalized using the Sigmoid activation function to generate the channel attention weights.

[0107] ;

[0108] Finally, the generated channel attention weights are multiplied channel by channel with the input features to obtain the feature map of the first feature mapping node:

[0109] ;

[0110] The feature map of the first feature mapping node Two two-dimensional feature maps are generated by performing average pooling and max pooling operations in the spatial dimension, respectively. These two two-dimensional feature maps are located at ( i , j The expression for the position is:

[0111] ;

[0112] Then, these two feature maps are concatenated along the channel dimension and a spatial attention map is generated through a 7×7 convolution operation. M s :

[0113] ;

[0114] In the formula, [;] denotes the channel connection operation. The spatial attention weights are multiplied position-by-position with the feature map of the first feature mapping node to obtain the feature map of the second feature mapping node:

[0115] .

[0116] S4: Constructing a multi-scale feature fusion pyramid

[0117] Feature maps are extracted from each ConvNeXt Block module, and then multi-scale feature maps are obtained through upsampling and lateral concatenation. Finally, the feature maps are concatenated into a single feature vector after global average pooling.

[0118] In this example, step S4 is specifically divided into the following parts:

[0119] S41: ConvNeXt is a deep convolutional neural network that consists of one Stem, four Stages, and a Classifier Head.

[0120] S42: Construct a multi-scale feature fusion pyramid, such as Figure 5 As shown. Features are extracted from the outputs of the four stages of the ConvNeX network (each stage corresponds to a different feature scale), as represented by: F 1, F 2, F 3, F 4. These feature maps represent information at different scales, among which... F 1 is the finest-grained feature map. F 4 is the coarsest-grained feature map.

[0121] First, for each layer of feature maps F l (in l For features ∈{1,2,3,4), a 1×1 convolution operation is first performed to compress the feature dimension, resulting in a more compact and informative feature representation. This operation helps reduce the number of channels, suppress redundant information, and retain useful feature information.

[0122] ;

[0123] Next, an upsampling operation is performed to upsample the high-level feature map to a higher resolution, making it consistent with the feature size of the current layer. Upsampling typically uses interpolation to achieve this. Size and same:

[0124] ;

[0125] The upsampled high-level feature map is then added to and fused with the low-resolution feature map of the current layer to obtain a new fused feature map.

[0126] ;

[0127] It not only preserves low-level details (such as edges and textures) but also enhances high-level semantic information (such as object category and contextual information). A 3×3 convolution operation is then performed on the fused feature map:

[0128] ;

[0129] Then, a global average pooling operation is used to transform the feature map at each scale into a global feature vector:

[0130] ;

[0131] This operation averages the pixels across all locations in the feature map, ultimately reducing the feature map to a long vector representing global information at that scale. Finally, by concatenating the global feature vectors from each scale, a final feature vector containing information from all scales is obtained.

[0132] .

[0133] S5: Network Setup

[0134] A new network is built based on the original ConvNeXt network structure, and its structure is as follows: Figure 5 As shown;

[0135] S6: Network Training

[0136] The new network was trained using the training set and tested using the test set to obtain the eccentricity fault classification and recognition model.

[0137] In this example, step S6 specifically includes the following:

[0138] S61: The input image size is 224×224×3. First, it is downsampled by a convolutional layer with a kernel of 4×4 and a stride of 4, reducing the feature map to 56×56×96. This step initially extracts low-level features from the image and reduces the spatial dimension.

[0139] S62: The feature map passes through four new stage modules (named stage1, stage2, stage3, and stage4 respectively). Stage1, stage2, stage3, and stage4 are stacked 3, 3, 9, and 3 times respectively, with ConvNeXt Block modules with CBAM attention mechanism added. Stage1 module does not have a downsampling layer. Stage2, stage3, and stage4 modules are all preceded by a downsampling layer (consisting of a normalization layer and a convolutional layer with a kernel of 2×2 and a stride of 2) to perform a 2x downsampling, downsampling the feature map to 28×28×192, 14×14×284, and 7×7×768 respectively.

[0140] S63: Extract the output feature map for each Stage F 1. F 2. F 3. F 4. Then, each stage's output feature map is channel-aligned by passing it through a regular convolutional layer (1×1 kernel, stride 1). , , , Their dimensions 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 ( , , Each stage (stage1, stage2, stage3) is upsampled to the same size as the previous stage using interpolation methods. , , Their dimensions 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 from stage 2, stage 3, and stage 4 ( , , ) Perform element-wise addition to generate the fused feature map. , , After completing upsampling and additive fusion, four multi-scale feature maps ( , , , The feature map is further extracted and enhanced through convolution operations (3×3 kernel, stride 1), and the spatial information is smoothed. , , , The four feature maps have dimensions of 56×56×256, 28×28×256, 14×14×256, and 7×7×256, respectively. Global average pooling is then applied to each of these four feature maps to obtain the output feature map. , , , The dimensions are 1×1×256, 1×1×256, 1×1×256, and 1×1×256 respectively; finally, , , , The final feature map is obtained by splicing. The dimensions are 1×1×1024.

[0141] S63: Feature Map Finally, the output is generated after the linear layer, and the type is determined.

[0142] Step S7: Identify eccentric faults in permanent magnet generators using an eccentric fault classification and identification model.

[0143] In this example, Table 2 shows the impact of adjusting different modules on the experimental results. It can be seen that adding the RP module increased the accuracy from 85.33% to 99.08%, and adding the CBAM module increased the accuracy from 93.29% to 96.71%. The final improved network model of this invention has an accuracy of 99.08%.

[0144] Table 2 shows the accuracy, F1 score, and AUC data for some models.

[0145] 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

[0146] Table 3 compares the experimental results of different attention mechanisms. RP+CBAM-ConvNeXt outperforms RP+CAM-ConvNeXt and RP+SAM-ConvNeXt in accuracy, F1 score, 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, enhancing the ability to represent features and thus improving the overall performance of the model.

[0147] Table 3 Comparison of experimental results for different attention mechanisms

[0148] 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

[0149] Table 4 shows a comparison of the experimental results for different network architectures. Among the comparisons of different network architectures, RP+ConvNeXt outperforms RP+GoogLeNet and RP+ResNet, achieving the best results in accuracy, F1 score, and AUC. This indicates that ConvNeXt is more effective in the task of diagnosing eccentricity faults in permanent magnet synchronous generators, possibly due to its design being more suited to the needs of current image processing tasks and possessing stronger feature extraction capabilities. GoogLeNet and ResNet, on the other hand, perform relatively similarly in this task, slightly lower than ConvNeXt.

[0150] Table 4 Experimental results for different networks

[0151] 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

[0152] In summary, the permanent magnet synchronous generator eccentricity fault diagnosis method based on two-dimensional recursive fusion graph and CBAM-ConvNeXt-FPN described in the above embodiments significantly improves fault feature extraction and diagnosis capabilities by introducing a two-dimensional recursive fusion graph, CBAM attention mechanism, and FPN feature pyramid network. The two-dimensional recursive graph enhances the visualization of fault features by encoding signal temporal dependencies through spatial structure, facilitating clearer identification of current fluctuations caused by eccentricity faults. The CBAM mechanism optimizes the model's focus on key feature regions and channels through spatial and channel attention mechanisms, improving the feature representation ability of local anomalies and important channels, thereby increasing the accuracy of fault diagnosis. FPN, through multi-scale feature fusion, enables the network to identify fault features at different scales, further improving the accuracy and robustness of fault diagnosis. Finally, various comparative experiments demonstrate the effectiveness of the improvement, achieving a final accuracy of 99.08%. The improved network model is also compared with other methods, proving that it has the best accuracy.

[0153] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, substitutions, or variations made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for diagnosing eccentricity faults in permanent magnet generators based on multi-channel feature fusion, characterized in that, Includes the following steps: Step S1: Data acquisition, acquiring the three-phase current signal of the permanent magnet generator eccentric fault simulation unit; Step S2: Construct a recursive two-dimensional fusion graph, convert the three-phase current timing data into a two-dimensional recursive matrix according to 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 by fusing the CBAM attention module before the Layer Scale layer in the ConvNeXt Block module of the ConvNeXt network to obtain the 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 and lateral connection, and concatenate the feature maps into a feature vector after global average pooling; Step S5: Network setup, build a new network based on the original ConvNeXt network structure; Step S6: Network training. Train the new network using the training set and test it using the test set to obtain the eccentricity fault classification and recognition model. Step S7: Identify eccentric faults in the permanent magnet generator using the eccentric fault classification and identification model; Step S3 includes: S31: The ConvNeXt block module includes one deep convolutional layer and two regular convolutional layers; it also includes a LayerScale layer, which is placed after the last convolutional layer. S32: Set up the CBAM attention module before the Layer Scale layer. The CBAM attention module includes a feature input end, a channel attention module, a spatial attention module, and a feature output end. 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 initial feature map with channel information and passes the weights to the first feature mapping node. The first feature mapping node multiplies the channel attention weights with the initial feature map channel by channel to generate a channel-attention-enhanced feature map, which is then output to the spatial attention module. The spatial attention module processes this feature map in the spatial domain and outputs the spatial attention weights to the second feature mapping node. The second feature mapping node multiplies the spatial attention weights with the channel-attention-enhanced feature map pixel by pixel to generate the final feature map. The specific processing procedure of the CBAM attention module is as follows: For the input feature map F∈R C×H×W Where H and W are the height and width, respectively, and C is the number of channels; firstly, global average pooling and global max pooling are performed on the channel dimensions of the feature map to capture the average information and maximum activation value of each channel. The two one-dimensional channel description vectors R generated by global average pooling and global max pooling are... C×1×1 The nth channel expression is represented as follows: The two one-dimensional channel description vectors generated after the two pooling operations are fed into a shared MLP and added together. Then, they are normalized using the Sigmoid activation function to generate channel attention weights. M c (F)=σ(W1(W0(F avg ))+W1(W0(F max ))) The generated channel attention weights are multiplied channel by channel with the input features to obtain the feature map of the first feature mapping node: The feature map F′∈R of the first feature mapping node C×H×W Two two-dimensional feature maps R are generated by performing average pooling and max pooling operations in the spatial dimension, respectively. 1×H×W The expression for these two two-dimensional feature maps located at position (i,j) is: Then, these two feature maps are concatenated along the channel dimension and a spatial attention map M is generated through a 7×7 convolution operation. s : M s (F′)=σ(f 7×7 ([F′ avg ;F′ max ])) In the formula, [;] denotes the channel connection operation, which multiplies the spatial attention weights with the input features position by position to obtain the feature map of the second feature mapping node: Step S4 includes: S41: The ConvNeXt network consists of 1 Stem, 4 Stages, and a Classifier Head; S42: Construct a multi-scale feature fusion pyramid to extract features from the outputs of the four stages of the ConvNeX network, denoted as F1, F2, F3, and F4, where F1 is the finest-grained feature map and F4 is the coarsest-grained feature map. First, for each layer of feature map F l First, a 1×1 convolution operation is performed to compress the feature dimension: F l ′=Conv1×1(F l ) Next, an upsampling operation is performed using interpolation, so that F′ l+1 Size and F l 'same: F l up =UpSample(F′ l+1 ) The upsampled high-level feature map is then added to and fused with the low-resolution feature map of the current layer to obtain a new fused feature map: F l p =F l up +F l ′ Perform a 3×3 convolution operation on the fused feature map: F l Conv =Conv3×3(F l p ) Then, a global average pooling operation is used to transform the feature map at each scale into a global feature vector: F l global =Global Avg Pooling(F l Conv ) By concatenating the global feature vectors from each scale, we obtain the final feature vector containing information from all scales: Step S6 includes: S61: Input image size, downsampled through convolutional layers to reduce the feature map size; S62: The feature map is downsampled by a factor of 2 through four new stage modules in sequence; S63: Take the output feature maps F1, F2, F3, and F4 of each stage, and then perform channel alignment on the output feature maps of each stage by passing them through a regular convolutional layer; then perform global average pooling on these four feature maps to obtain the output feature map F1. global , Finally, F1 global , The final feature map is obtained by splicing. S63: Feature Map Finally, the output is generated after the linear layer, and the fault type is determined. Step S2 includes: S21: Using the mutual information coefficient method, the mutual information between the current signal and the corresponding delayed signal is calculated, and the optimal delay time is adaptively selected based on the first local minimum of the mutual information. In the formula, I(τ) is the mutual information; n is the length of the current signal; τ is the delay time set between [1, 100]; x(t) is the current signal at point t; p(x(t)) is the probability density of the current signal; p(x(t+τ)) is the probability density of the delayed signal; p(x(t), x(t+τ)) is the joint probability density of the current signal and the delayed signal. Using the optimal delay τ * The time series is reconstructed in phase space to obtain the reconstructed vector X(t): X(t)=[x(t),x(t+τ * ),...,x(t+(m-2)τ * ),x(t+(m-1)τ * )] In the formula, m is the embedding dimension, and τ * This is the optimal delay time; for each time point t, the proximity relationship in phase space is determined by calculating the Euclidean distance between the reconstructed vector X(t) and the reconstructed vectors at other time points: In the formula, X i (t), X i (t+k) are the i-th components of the reconstructed vectors X(t) and X(t+k) corresponding to time points t and t+k, respectively; if the distance between two adjacent points is small in the low-dimensional space, but the distance increases in the high-dimensional space, these points are considered to be pseudo-nearest neighbors: d(X(t),X(t+k)) (m+1) >d(X(t),X(t+k)) (m) ×factor In the formula, factor is a constant; as the dimension m increases, the proportion of false nearest neighbors gradually decreases and tends to stabilize after a certain dimension, where m is the dimension. * The optimal embedding dimension; After reconstructing the phase space, 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: Where ε is the minimum distance threshold, and the elements R of the recursion matrix R are... i,j This represents the recursive relationship between the i-th and j-th points in the time series; H is the Heaviside function, which considers two points as recursive points and marks them as 1 if the distance between them is less than or equal to ε; otherwise, it marks them as 0. S22: Normalize the obtained two-dimensional recursive matrix one by one, map it to the R, G, and B channels, and fuse it into a color image.

Images