Fault identification system and method for multiple detection devices under small sample condition

Abstract

The invention provides a fault identification system for multiple detection devices under a small sample condition, and the system comprises a graph structure construction module which extracts time domain features from time domain signals of multiple detection devices of a gearbox, and generates frequency domain features through Fourier transform; respectively constructing a time domain graph and a frequency domain graph based on K nearest neighbor, splicing the time domain graph and the frequency domain graph into a positive graph, and screening unique neighborhoods of the frequency domain graph to construct a negative graph; the feature fusion module is used for inputting the positive frequency domain graph, the negative frequency domain graph and the multi-layer GraphSAGE and outputting a fusion node feature matrix; the fusion feature graph construction module is used for calculating K neighbor edge weights based on the fusion features and constructing a weighted adjacency matrix; the feature learning module is used for fusing the feature and the weighted adjacency matrix, inputting the fused feature and weighted adjacency matrix into the alternately stacked GraphSAGE-Transform layer, and outputting a high-order feature; and the fault identification module compresses the features into node-level vectors through global pooling, extracts nonlinear features through MLP, inputs the nonlinear features into a Softmax classifier, and outputs a fault probability. Through multi-graph fusion, comparative learning and dynamic weight, the accuracy and robustness of fault recognition of multiple detection devices under the small sample condition are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-detection device fault diagnosis, and in particular to a multi-detection device fault identification system and method under small sample conditions. Background Art

[0002] With the development of intelligent industrial equipment, mechanical system fault diagnosis technology has become a core link in ensuring production safety. As a key component of power transmission, the coordinated diagnosis of signals from multiple detection devices on gearboxes is of great significance for early fault warning.

[0003] While traditional data-driven approaches have made progress under sufficient sample conditions, they still face two major technical bottlenecks when faced with the limited sample size found in real industrial scenarios: 1. Deep learning methods rely on large amounts of labeled data for feature learning, but obtaining early-stage gearbox failure samples is expensive and involves a single operating condition. When training with fewer than 50 groups / classes of samples, traditional CNN and RNN models are prone to falling into local optima, resulting in a sudden drop in feature extraction dimensionality and confusion about failure modes. 2. Existing feature-level fusion methods employ simple concatenation or static weighting strategies, failing to effectively exploit topological correlations between time- and frequency-domain features.

[0004] In summary, feature-level fusion, as one of the core methods of multi-channel signal processing, currently lacks a reasonable feature screening and weight distribution mechanism, making it difficult to reflect complex relationships and feature differences. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of the prior art and provide a multi-detection device fault identification system under small sample conditions, comprising:

[0006] A graph structure construction module extracts time domain features from the time domain signals of the multiple detection devices of the gearbox, performs Fourier transform on the time domain signals of the multiple detection devices of the gearbox to extract frequency domain features, constructs a time domain graph and a frequency domain graph based on the time domain features and the frequency domain features using the K-nearest neighbor rule, concatenates the time domain graph and the frequency domain graph to obtain a positive graph, compares the neighborhood node sets of the time domain graph and the frequency domain graph, and filters out the neighborhood nodes unique to the frequency domain graph to construct a negative graph;

[0007] The feature fusion module inputs the positive image, negative image, and frequency domain image into the graph variational autoencoder respectively. In the graph variational autoencoder, it undergoes multi-layer GraphSAGE encoding to generate potential distribution, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion of channel weights to obtain the fused node feature matrix;

[0008] The fusion feature graph construction module calculates the edge weights of the fused features based on the K-nearest neighbor rule according to the fusion node feature matrix, and constructs an adjacency matrix with edge weights based on the edge weights of the fused features;

[0009] The feature learning module inputs the fused node feature matrix and the adjacency matrix with edge weights into alternating stacks of GraphSAGE and Transformer layers. The GraphSAGE layer updates the node representation by weighted aggregation of neighborhood features. The Transformer layer allows edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with a gating mechanism, outputting the node features after feature learning.

[0010] The fault identification module inputs the node features after feature learning into the global average pooling layer to aggregate them into a pooled global feature vector. After extracting high-level nonlinear features from the pooled global feature vector through a multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier to obtain the fault category probability.

[0011] Furthermore, in the graph structure construction module, time domain features are extracted from the time domain signals of the multiple detection devices of the gearbox, the time domain signals of the multiple detection devices of the gearbox are subjected to Fourier transform to extract frequency domain features, and the specific method for constructing the time domain graph and the frequency domain graph based on the K-nearest neighbor rule according to the time domain features and the frequency domain features is as follows:

[0012] The time domain signal of the gearbox multi-detection device is recorded as X = {x1, x2, ..., x i};

[0013] Extract time domain features from the time domain signals of the gearbox multi-detection device, each sample x i Generate time domain feature vector d t Represents the time domain feature dimension, R is a real number, where the time domain features include mean, variance, peak, peak-to-peak value, root mean square, shape factor, kurtosis, and impulse factor;

[0014] Perform Fourier transform FFT on the time domain signal, that is, for each sample x i Perform Fourier transform to obtain the spectrum F i =FFT(x i )∈R X , X is the spectrum resolution, and the following frequency domain features are constructed: Spectrum amplitude: Spectrum amplitude = |F i |; Band energy: divide the frequency band and calculate the energy of each band; frequency domain statistics: spectrum centroid, spectrum bandwidth, spectrum flatness. Then, the frequency domain feature vector is obtained based on the frequency domain characteristics. d f Represents the frequency domain feature dimension. Frequency domain features include spectrum amplitude, frequency band energy, and frequency domain statistics.

[0015] Use Euclidean distance to calculate the feature space distance between node i and node j:

[0016]

[0017] Among them, D t (i,j) is the time domain feature distance matrix, D f (i, j) is the frequency domain feature distance matrix;

[0018] Constructing time domain graph G based on K nearest neighbors t =(V,E t ) and the frequency domain graph G f =(V,E f ), V is a node set, each node V in the node set V i Corresponding to a gearbox detection signal sample x i , E t is the edge set of the time domain graph, E f is the edge set of the frequency domain graph;

[0019] Based on the time domain feature distance matrix D t Construct the adjacency matrix A of the time domain graph t , based on the frequency domain feature distance matrix D f Construct the adjacency matrix A of the frequency domain graph f , as shown below:

[0020]

[0021] Among them, Topk means for each node i, from D t (i,:) or D f In (i,:), K nodes with the smallest distance are selected as neighbors, and the : symbol represents all elements in the corresponding row or column.

[0022] Furthermore, in the graph structure construction module, the time domain graph and the frequency domain graph are spliced together to obtain a positive graph, and the neighborhood node sets of the time domain graph and the frequency domain graph are compared and the neighborhood nodes unique to the frequency domain graph are screened out to construct a negative graph. The specific method is:

[0023] Splice the time domain image and frequency domain image into the positive image G pos =(V,E pos ), E pos is the edge set of the positive graph. The specific method is: splice the time domain graph and the frequency domain graph into the positive graph G pos =(V,E pos ), E pos is the edge set of the positive graph. The specific method is: the positive graph feature vector The edge set E of the main graph pos =E t ∩E f , the adjacency matrix A of the positive graph pos =A f ∩A t ;

[0024] By comparing the differences in neighborhood nodes between the time domain graph and the frequency domain graph, the unique neighborhood nodes in the frequency domain are screened to construct the negative graph G. neg =(V,E neg ), E neg The edge set of the negative graph is as follows: for each node i, the neighboring node set N of node i in the time domain graph is t (i) = {j | A t (i,j)=1}, the neighboring node set N of node i in the frequency domain graph f (i) = {j | A f (i,j)=1}, the neighborhood node set N of the negative graph neg (i) is the neighborhood node set N in the frequency domain f (i) Neighborhood node set N relative to the time domain t (i) A set of unique neighborhood nodes, node features of the negative graph Negative graph adjacency matrix A neg as follows:

[0025]

[0026] Furthermore, in the feature fusion module, the positive image, negative image, and frequency domain image are respectively input into the graph variational autoencoder, and the graph variational autoencoder undergoes multi-layer GraphSAGE encoding to generate potential distribution, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion processing of channel weights to obtain the fusion node feature matrix. The specific method is as follows:

[0027] Perform multi-layer graph convolution on the positive image, negative image, and frequency domain image, and gradually aggregate neighborhood information. The specific steps are as follows:

[0028]

[0029] in, They are respectively the positive graph node features, negative graph node features, and frequency domain graph node features output after neighborhood aggregation and nonlinear transformation. is the input positive graph node feature, negative graph node feature, frequency domain graph node feature, W (l) is the learnable multi-layer graph convolution weight, σ is the activation function, and i is the node v i The serial number, N pos (i) N neg (i) N f (i) are nodes v i In the positive graph, negative graph, and frequency domain graph, the domain node set j is the node v iThe serial number of the domain node; AGGREGATE represents neighbor feature aggregation, which aggregates all neighbor features of node i to generate an aggregate vector for capturing local neighborhood information; CONCAT represents feature concatenation, which concatenates the node's own features with the aggregated neighbor features to form a richer joint representation;

[0030] Then, the mean and covariance matrix of the latent variables are calculated independently for each graph through the multi-layer GraphSAGE encoder. The formula is as follows:

[0031] Main image:

[0032] Negative image:

[0033] Frequency domain plot:

[0034] Covariance matrix:

[0035] in, is the feature matrix of the positive graph nodes after neighborhood aggregation and nonlinear transformation, d pos is the characteristic dimension of the positive image, is the negative graph node feature matrix after neighborhood aggregation and nonlinear transformation, d neg is the feature dimension of the negative graph, is the frequency domain graph node feature matrix after neighborhood aggregation and nonlinear transformation, d f is the characteristic dimension of the frequency domain graph, E pos 、E neg 、E f are the edge sets of the positive graph, negative graph, and frequency domain graph, respectively. is the mean of the latent variables of the positive graph, N is the number of nodes, d z is the latent space dimension, μ neg is the mean of the negative graph latent variable, μ f is the mean of the latent variable in the frequency domain graph, is the logarithmic variance of the latent variable of the positive graph, is the log variance of the latent variable of the negative graph, is the logarithmic variance of the latent variable in the time domain graph, is the dth diagonal in the covariance matrix of the latent variable z The variance of the dimensions, Ender is a multi-layer GraphSAGE encoder;

[0036] Then, the latent variables are sampled from the posterior distribution of the positive image, negative image, and frequency domain image. The formula is as follows:

[0037]

[0038] Where ⊙ represents element-by-element multiplication, ∈~N(0,I) represents the noise sampled from the standard normal distribution, is the identity matrix, z pos is the latent variable of the positive graph, z neg is the latent variable of the negative graph, z f is the latent variable of the frequency domain graph, and exp converts the logarithmic variance into variance;

[0039] Then the latent variable z pos 、z neg 、z f Use multi-layer GraphSAGE decoders to reconstruct the original features and output the reconstructed frequency domain graph feature matrix l is the channel number:

[0040]

[0041] Calculate the channel weight using the following formula:

[0042]

[0043] Among them, α l is the weight of the frequency domain feature moment in channel l, ε is a very small positive number, is the mean square error of the lth channel, s l Indicates the standard deviation of the node feature of the lth channel, s l =std(F l );

[0044] Based on the weights of c channels, the multi-graph reconstruction features are weightedly fused to obtain the fused node feature matrix F fused :

[0045]

[0046] Calculate the mean square error L between the reconstructed features and the original features of the positive image, negative image, and frequency domain image recon , measures the difference between the reconstructed features and the original features:

[0047]

[0048] Regularize the latent space, L KL is a regularization term that constrains the posterior distribution to be close to the standard normal distribution:

[0049]

[0050] Based on the contrast loss L contrast , enhance the similarity between the positive image and the frequency domain image, and suppress the correlation between the negative image and the positive image and the frequency domain image:

[0051]

[0052] Among them, sim represents cosine similarity;

[0053] Total loss L total for:

[0054] L total =L recon +βL KL +γL contrast

[0055] Among them, β is the first hyperparameter; γ is the second hyperparameter, and the first hyperparameter β and the second hyperparameter γ are used to control the strength of regularization and contrastive learning respectively;

[0056] The training is performed according to the following steps: Forward propagation is: input the node features and adjacency matrix of the positive graph, negative graph, and frequency domain graph, the encoder generates the mean and variance, reparameterizes the sampling to obtain the latent variables, the decoder reconstructs the features, calculates the fusion weights, and outputs the fusion feature matrix F fused ; Back propagation is: calculate the total loss and update the parameters of the encoder, decoder and channel weights.

[0057] Furthermore, in the fusion feature graph construction module, the edge weights of the fused features are calculated based on the K-nearest neighbor rule according to the fusion node feature matrix, and the adjacency matrix with edge weights is constructed according to the edge weights of the fused features. The specific method is:

[0058] First calculate the Euclidean distance matrix of the fused features:

[0059]

[0060] Among them, D fused (i,j) is the fusion node feature matrix F fused The Euclidean distance between nodes i and j, is the fused feature vector of node i, is the fusion node feature matrix F fused The nth eigenvalue of the i-th node in , is the fusion node feature matrix F fused The nth eigenvalue of the jth node in ;

[0061] Select neighbor sets based on K nearest neighbors:

[0062] N fused (i) = {j | D fused (i,j)≤TopK(D fused (i,:),K)}

[0063] Among them, Nfused (i) is the fusion node feature matrix F fused The set of the nearest K neighbors of node i, TopK(D fused (i,:),K) is the fusion node feature matrix F fused The threshold of the smallest K distances among all sample distances of node i;

[0064] Calculate edge weight W based on Euclidean distance fused (i,j), the formula is as follows:

[0065]

[0066] Among them, η is the bandwidth parameter, which is used to control the weight decay speed;

[0067] Construct a weighted adjacency matrix A fused (i,j)∈R N×N , only the weights of the first K neighbors of each node are retained, and the remaining edge weights are 0:

[0068]

[0069] Furthermore, in the feature learning module, the fused node feature matrix and the adjacency matrix with edge weights are input into the alternately stacked GraphSAGE and Transformer layers. The specific method of the GraphSAGE layer updating the node representation by weighted aggregation of neighborhood features is as follows:

[0070] The fusion node feature matrix F fused ∈R N×d and the adjacency matrix A with edge weights fused (i,j)∈R N×N Input alternately stacked GraphSAGE and Transformer layers;

[0071] First, add position information to the node feature:

[0072]

[0073] Among them, PE(i,2k) is the k-th dimension position code of node i;

[0074] For the GraphSAGE layer, each layer updates the node representation by weighted aggregation of neighbor features:

[0075]

[0076] Among them, H (l) ∈R N×d is the node feature of the lth layer, W self and W neighare the first parameter matrix and the second parameter matrix that can be learned, respectively, and σ is the activation function.

[0077] Furthermore, in the feature learning module, the Transformer layer allows edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with the gating mechanism. The specific method for outputting the node features after feature learning is:

[0078] For the Transformer layer, the adjacency matrix A with weights fused (i,j) is added to the attention calculation as a bias term:

[0079]

[0080] A=H (l) W A , B=H (l) W B ,C=H (l) W C

[0081] Among them, A, B, and C are query matrix, key matrix, and value matrix respectively, and W A 、W B and W C are the learnable first projection matrix, the learnable second projection matrix and the learnable third projection matrix respectively; ε is a very small positive number; log(A fused (i,j)+ε) is used to convert edge weights into attention biases;

[0082] In the Transformer layer, the gating coefficients are used to dynamically adjust the influence of edge weights:

[0083] G=σ(W g ·H (l) ),H (l+1) =G⊙Attention(A,B,C)+(1-G)⊙H (l)

[0084] Among them, W g is the learnable gating parameter matrix; G is the gating coefficient matrix;

[0085] After alternately stacking GraphSAGE and Transformer layers, the node features H after feature learning are output (L) ∈R N×d , N is the number of nodes, d is the feature dimension, and L is the total number of layers for feature learning.

[0086] Furthermore, in the fault identification module, the node features after feature learning are input into the global average pooling layer to aggregate into a pooled global feature vector. After the high-level nonlinear features are extracted from the pooled global feature vector through a multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier. The specific method for obtaining the fault category probability is as follows:

[0087] The node feature H after feature learning (L) ∈R N×d Enter the average pooling layer to convert the features of each node into a global feature vector at the graph level:

[0088]

[0089] in, is the eigenvector of the i-th node, h global is the global feature vector after pooling;

[0090] The pooled global feature vector h global Input MLP for feature extraction, specifically in MLP, high-level features are further extracted through fully connected layers and nonlinear activation functions:

[0091] h mlp =ReLU(W mlp ·h global +b mlp )

[0092] in, is the weight matrix of MLP, b mlp is the bias vector of MLP, d hidden is the hidden layer dimension, It is the high-level nonlinear feature vector output by MLP;

[0093] The high-level nonlinear feature vector h output by MLP mlp Input Softmax classifier, the high-level nonlinear feature vector h mlp Mapping to class probability distribution:

[0094]

[0095] in, is the weight matrix of the Softmax classifier, b cls ∈R M is the bias vector of the Softmax classifier, p m ∈R M , p m is the probability of the mth type of failure, M is the number of failure categories;

[0096] The difference between the predicted probability and the true label is then calculated through the loss function:

[0097]

[0098] Among them, L cls is the cross entropy loss; y∈{0,1} M is the one-hot encoding of the true label.

[0099] A method for identifying faults of multiple detection devices under small sample conditions, comprising:

[0100] Extracting time domain features from the time domain signals of the multiple detection devices of the gearbox, performing Fourier transform on the time domain signals of the multiple detection devices of the gearbox to extract frequency domain features, constructing a time domain graph and a frequency domain graph based on the K-nearest neighbor rule according to the time domain features and the frequency domain features, splicing the time domain graph and the frequency domain graph to obtain a positive graph, comparing the neighborhood node sets of the time domain graph and the frequency domain graph, and filtering out the neighborhood nodes unique to the frequency domain graph to construct a negative graph;

[0101] The positive image, negative image, and frequency domain image are input into the graph variational autoencoder respectively. In the graph variational autoencoder, the latent distribution is generated through multi-layer GraphSAGE encoding, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion of channel weights to obtain the fused node feature matrix.

[0102] According to the fusion node feature matrix, the edge weights of the fused features are calculated based on the K-nearest neighbor rule, and the adjacency matrix with edge weights is constructed according to the edge weights of the fused features;

[0103] The fused node feature matrix and the adjacency matrix with edge weights are input into the alternating stack of GraphSAGE and Transformer layers. The GraphSAGE layer updates the node representation by weighted aggregation of neighborhood features. The Transformer layer allows edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with a gating mechanism, outputting the node features after feature learning.

[0104] The node features after feature learning are input into the global average pooling layer to aggregate into a pooled global feature vector. After the high-level nonlinear features are extracted from the pooled global feature vector through a multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier to obtain the fault category probability.

[0105] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the above-mentioned method for identifying faults of multiple detection devices under small sample conditions.

[0106] The beneficial effects of the present invention are:

[0107] 1. Through multi-image fusion, contrastive learning, and dynamic weight design, this invention significantly improves the accuracy and robustness of multi-detection device fault identification under small sample conditions, while reducing data dependence and having high industrial application value.

[0108] 2. By leveraging the topological complementarity of time-domain and frequency-domain graphs, we construct positive and negative comparison graphs, innovatively achieving bidirectional mapping of the time-frequency feature space. We utilize positive graphs to preserve cross-modal shared features, while negative graphs enhance unique frequency-domain features. Neighborhood difference comparison learning enables the model to be sensitive to anomalies even with small sample sizes, thereby improving fault classification accuracy.

[0109] 3. The K-nearest neighbor relationship based on the graph structure reduces the dependence on a large amount of labeled data. Combining graph variational autoencoders and contrastive learning improves the generalization of the model with a small number of samples.

[0110] 4. The channel weight mechanism based on variational graph autoencoders breaks through the limitations of traditional static fusion: the weight distribution guided by reconstruction error improves the gain of important feature channels, and the latent space contrast loss function effectively decouples normal and fault modes.

[0111] 5. In the GraphSAGE-Transformer hybrid architecture, the gated attention mechanism enhances the contribution of key neighborhood nodes. Position encoding is combined with dynamic edge weights to address the periodic offset problem of rotating machinery signals. A weighted adjacency matrix is introduced into the Transformer to enhance attention to key neighborhood nodes, improve fault identification accuracy, and reduce the risk of overfitting through an alternating stacking structure.

[0112] 6. Jointly optimize reconstruction loss, KL divergence, and contrast loss to ensure collaborative optimization of the model in feature learning and classification tasks.

[0113] 7. The K-nearest neighbor graph relies only on local similarity and can be constructed without a large number of samples, reducing data requirements. By generating latent variables, the model's generalization ability to the data distribution is enhanced, alleviating the problem of overfitting in small samples. By comparing positive and negative images, discriminative features are learned from a limited number of samples, reducing reliance on labeled data. The feature weights of multiple images are adaptively adjusted to avoid manual design and improve model flexibility under small sample conditions. The global pooling + MLP structure has a small number of parameters, making it suitable for small sample training and avoiding overfitting caused by complex models. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 This is a block diagram of the multi-detection device fault identification system under small sample conditions of the present invention;

[0115] Figure 2 A schematic diagram of the structure of the graph in the present invention;

[0116] Figure 3It is a device for identifying failures of multiple detection devices under small sample conditions of the present invention;

[0117] Figure 4 is the F1 score of each channel in Example 1 of the present invention. DETAILED DESCRIPTION

[0118] In order to make the technical problems, technical solutions and beneficial effects to be solved by this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0119] Example 1

[0120] refer to Figure 1 and Figure 3 , a multi-detection device fault identification system under small sample conditions, comprising:

[0121] A graph structure construction module extracts time domain features from the time domain signals of the multiple detection devices of the gearbox, performs Fourier transform on the time domain signals of the multiple detection devices of the gearbox to extract frequency domain features, constructs a time domain graph and a frequency domain graph based on the time domain features and the frequency domain features using the K-nearest neighbor rule, concatenates the time domain graph and the frequency domain graph to obtain a positive graph, compares the neighborhood node sets of the time domain graph and the frequency domain graph, and filters out the neighborhood nodes unique to the frequency domain graph to construct a negative graph;

[0122] The feature fusion module inputs the positive image, negative image, and frequency domain image into the graph variational autoencoder respectively. In the graph variational autoencoder, it undergoes multi-layer GraphSAGE encoding to generate potential distribution, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion of channel weights to obtain the fused node feature matrix;

[0123] The fusion feature graph construction module calculates the edge weights of the fused features based on the K-nearest neighbor rule according to the fusion node feature matrix, and constructs an adjacency matrix with edge weights based on the edge weights of the fused features;

[0124] The feature learning module inputs the fused node feature matrix and the adjacency matrix with edge weights into alternating stacks of GraphSAGE and Transformer layers. The GraphSAGE layer updates the node representation by weighted aggregation of neighborhood features. The Transformer layer allows edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with a gating mechanism, outputting the node features after feature learning.

[0125] The fault identification module inputs the node features after feature learning into the global average pooling layer to aggregate them into a pooled global feature vector. After extracting high-level nonlinear features from the pooled global feature vector through a multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier to obtain the fault category probability.

[0126] (1) Collect the time domain signals of multiple detection devices of the gearbox, including data from the speed control, motor, and acceleration detection devices. Normalize the collected data.

[0127] (2) In the graph structure construction module, time domain features are extracted from the time domain signals of the multiple detection devices of the gearbox, and the time domain signals of the multiple detection devices of the gearbox are subjected to Fourier transform to extract frequency domain features. According to the time domain features and frequency domain features, a time domain graph and a frequency domain graph are constructed based on the K nearest neighbor rule. The time domain graph and the frequency domain graph are spliced to obtain a positive graph. The neighborhood node sets of the time domain graph and the frequency domain graph are compared and the neighborhood nodes unique to the frequency domain graph are screened out. The specific method for constructing a negative graph is as follows:

[0128] The time domain signal of the gearbox multi-detection device is recorded as X = {x1, x2, ..., x i}.

[0129] Extract time domain features from the time domain signals of the gearbox multi-detection device, each sample x i Generate time domain feature vector d t represents the time domain feature dimension, and R is a real number. Time domain features include mean, variance, peak, peak-to-peak value, root mean square (RMS), crest factor, kurtosis, and impulse factor.

[0130] Perform Fourier transform FFT on the time domain signal, that is, for each sample x i Perform Fourier transform to obtain the spectrum F i =FFT(x i )∈R X (X is the spectrum resolution), construct the following frequency domain features: Spectrum amplitude: Spectrum amplitude = |F i |; Band energy: divide the frequency band (such as the gear meshing frequency octave band) and calculate the energy of each band; frequency domain statistics: spectrum centroid, spectrum bandwidth, spectrum flatness. Then, the frequency domain feature vector is obtained based on the frequency domain characteristics. d f Represents the frequency domain feature dimension.

[0131] Use Euclidean distance to calculate the feature space distance between node i and node j:

[0132]

[0133] Among them, Dt (i,j) is the time domain feature distance matrix, D f (i, j) is the frequency domain feature distance matrix;

[0134] Constructing time domain graph G based on K nearest neighbors t =(V,E t ) and the frequency domain graph G f =(V,E f ), V is a node set, each node V in the node set V i Corresponding to a gearbox detection signal sample x i , E t is the edge set of the time domain graph, E f is the edge set of the frequency domain graph. For the edge set E of the time domain graph t ={(v i ,v j )|v j ∈TopK(D t (i,:))}, if K = 3, the time domain nearest neighbors of node v1 are v3, v5, v7; for the edge set E of the frequency domain graph f ={(v i ,v j )|v j ∈TopK(D f (i,:))}, if K = 3, the time domain nearest neighbors of node v1 are v2, v4, v6. TopK means that for each node i, from D t (i,:) or D f In (i,:), K nodes with the smallest distance are selected as neighbors, and the : symbol represents all elements in the corresponding row or column.

[0135] Based on the time domain feature distance matrix D t Construct the adjacency matrix A of the time domain graph t , based on the frequency domain feature distance matrix D f Construct the adjacency matrix A of the frequency domain graph f , as shown below:

[0136]

[0137] Splice the time domain image and frequency domain image into the positive image G pos =(V,E pos ), E pos is the edge set of the positive graph. The specific method is: the positive graph eigenvector The edge set E of the main graph pos =E t ∩E f , the adjacency matrix A of the positive graph pos =A f ∩A t ;

[0138] By comparing the differences in neighborhood nodes between the time domain graph and the frequency domain graph, the unique neighborhood nodes in the frequency domain are screened to construct the negative graph G. neg =(V,E eng ), E neg The edge set of the negative graph is as follows: for each node i, the neighboring node set N of node i in the time domain graph is t (i) = {j | A t (i,j)=1}, the neighboring node set N of node i in the frequency domain graph f (i) = {j | A f (i,j)=1}, so the neighborhood node set N of the negative graph neg (i) is the neighborhood node set N in the frequency domain f (i) Neighborhood node set N relative to the time domain t (i) A set of unique neighborhood nodes. Node features of negative graphs Negative graph adjacency matrix A neg as follows:

[0139]

[0140] The schematic diagram of graph structure generation is as follows Figure 2 shown.

[0141] Through the above steps, we clarify the extraction methods of time-domain and frequency-domain features, build a complementary graph structure, provide a foundation for subsequent fusion, and enhance feature diversity. Using negative graphs to highlight unique neighborhoods in the frequency domain enhances contrastive learning, helps the model distinguish between normal and abnormal patterns, and improves its ability to discriminate under small sample sizes.

[0142] (3) In the feature fusion module, the positive image, negative image, and frequency domain image are input into the graph variational autoencoder respectively. In the graph variational autoencoder, the latent distribution is generated through multi-layer GraphSAGE encoding, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion processing of channel weights. The specific method to obtain the fusion node feature matrix is as follows:

[0143] Perform multi-layer graph convolution on the positive image, negative image, and frequency domain image, and gradually aggregate neighborhood information. The specific steps are as follows:

[0144]

[0145] in, They are respectively the positive graph node features, negative graph node features, and frequency domain graph node features output after neighborhood aggregation and nonlinear transformation. is the input positive graph node feature, negative graph node feature, frequency domain graph node feature, W (l)is the learnable multi-layer graph convolution weight, σ is the activation function, and i is the node v i The serial number, N pos (i) N neg (i) N f (i) are nodes v i In the domain node set in the positive graph, negative graph, and frequency domain graph, j is the serial number of the domain node of node vi; AGGREGATE represents neighbor feature aggregation, which aggregates all neighbor features of node i to generate an aggregate vector to capture local neighborhood information; CONCAT represents feature concatenation, which concatenates the node's own features with the aggregated neighbor features to form a richer joint representation;

[0146] Then, the mean and covariance matrix of the latent variables are calculated independently for each graph through the multi-layer GraphSAGE encoder. The formula is as follows:

[0147] Main image:

[0148] Negative image:

[0149] Frequency domain plot:

[0150] Covariance matrix:

[0151] in, is the feature matrix of the positive graph nodes after neighborhood aggregation and nonlinear transformation, d pos is the characteristic dimension of the positive image, is the negative graph node feature matrix after neighborhood aggregation and nonlinear transformation, d neg is the feature dimension of the negative graph, is the frequency domain graph node feature matrix after neighborhood aggregation and nonlinear transformation, d f is the characteristic dimension of the frequency domain graph, E pos 、E neg 、E f are the edge sets of the positive graph, negative graph, and frequency domain graph, respectively. is the mean of the latent variables of the positive graph, N is the number of nodes, d z is the latent space dimension, μ neg is the mean of the negative graph latent variable, μ f is the mean of the latent variable in the frequency domain graph, is the logarithmic variance of the latent variable of the positive graph, is the log variance of the latent variable of the negative graph, is the logarithmic variance of the latent variable in the time domain graph, is the dth diagonal in the covariance matrix of the latent variable zThe variance of the dimensions, Encoder is a multi-layer GraphSAGE encoder;

[0152] Then, the latent variables are sampled from the posterior distribution of the positive image, negative image, and frequency domain image. The formula is as follows:

[0153]

[0154]

[0155] Where ⊙ represents element-by-element multiplication, ∈~N(0,I) represents the noise sampled from the standard normal distribution, is the identity matrix, z pos is the latent variable of the positive graph, z neg is the latent variable of the negative graph, z f is the latent variable of the frequency domain graph, and exp converts the logarithmic variance into variance;

[0156] Then the latent variable z pos 、z neg 、z f Use multi-layer GraphSAGE decoders to reconstruct the original features and output the reconstructed frequency domain graph feature matrix l is the channel number:

[0157]

[0158] Calculate the channel weight using the following formula:

[0159]

[0160] Among them, α l is the weight of the frequency domain feature moment in channel l, ε is a very small positive number, is the mean square error of the lth channel, s l Indicates the standard deviation of the node feature of the lth channel, s l =std(F l );

[0161] Based on the weights of c channels, the multi-graph reconstruction features are weightedly fused to obtain the fused node feature matrix F fused :

[0162]

[0163] Calculate the mean square error L between the reconstructed features and the original features of the positive image, negative image, and frequency domain image recon , measures the difference between the reconstructed features and the original features:

[0164]

[0165] Regularize the latent space, L KL is a regularization term that constrains the posterior distribution to be close to the standard normal distribution:

[0166]

[0167] Based on the contrast loss L contrast , enhance the similarity between the positive image and the frequency domain image, and suppress the correlation between the negative image and the positive image and the frequency domain image:

[0168]

[0169] Among them, sim represents cosine similarity;

[0170] Total loss L total for:

[0171] L total =L recon +βL KL +γL contrast

[0172] Among them, β is the first hyperparameter; γ is the second hyperparameter, and the first hyperparameter β and the second hyperparameter γ are used to control the strength of regularization and contrastive learning respectively;

[0173] The training is performed according to the following steps: Forward propagation is: input the node features and adjacency matrix of the positive graph, negative graph, and frequency domain graph, the encoder generates the mean and variance, reparameterizes the sampling to obtain the latent variables, the decoder reconstructs the features, calculates the fusion weights, and outputs the fusion feature matrix F fused ; Back propagation is: calculate the total loss and update the parameters of the encoder, decoder and channel weights.

[0174] This method captures local and global graph structure through multi-layer neighborhood aggregation, dynamically learning the contributions of different modalities while avoiding manual weighting. It also utilizes negative graphs to enhance the model's sensitivity to abnormal patterns. It effectively integrates time, frequency, and contrast information, improving the robustness and accuracy of tasks such as gearbox fault detection. The latent variable generation and dynamic weighted fusion based on a graph variational autoencoder reduce noise interference and improve the robustness of feature representation, making it suitable for data-scarce scenarios.

[0175] (4) In the fusion feature graph construction module, the edge weights of the fused features are calculated based on the K-nearest neighbor rule according to the fusion node feature matrix, and the specific method of constructing the adjacency matrix with edge weights according to the edge weights of the fused features is as follows:

[0176] Fusion node feature matrix F fusesd ∈R N×d , N is the number of nodes, and d is the dimension of the fused features.

[0177] First calculate the Euclidean distance matrix of the fused features:

[0178]

[0179] Among them, D fused (i,j) is the fusion node feature matrix F fused The Euclidean distance between nodes i and j, is the fused feature vector of node i, is the fusion node feature matrix F fused The nth eigenvalue of the i-th node in , is the fusion node feature matrix F fused The nth eigenvalue of the jth node in .

[0180] Select neighbor sets based on K nearest neighbors:

[0181] N fused (i) = {j | D fused (i,j)≤TopK(D fused (i,:),K)}

[0182] Among them, N fused (i) is the fusion node feature matrix F fused The set of the nearest K neighbors of node i, TopK(D fused (i,:),K) is the fusion node feature matrix F fused The threshold for the smallest top K distances among all sample distances of node i.

[0183] Calculate edge weight W based on Euclidean distance fused (i, j), the smaller the distance, the greater the weight, indicating the similarity between nodes. The calculation method is as follows:

[0184]

[0185] Among them, η is the bandwidth parameter, which is used to control the weight decay speed (can be set to the mean of all distances).

[0186] Construct a weighted adjacency matrix A fused (i,j)∈R N×N , only the weights of the first K neighbors of each node are retained, and the remaining edge weights are 0:

[0187]

[0188] (5) In the feature learning module, the fused node feature matrix and the adjacency matrix with edge weights are input into the alternately stacked GraphSAGE and Transformer layers. The GraphSAGE layer updates the node representation by weighted aggregation of neighborhood features. The Transformer layer allows the edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with the gating mechanism. The specific method for outputting the node features after feature learning is as follows:

[0189] The fusion node feature matrix F fused ∈R N×d and the adjacency matrix A with edge weights fused (i,j)∈R N×N Input is alternating stacks of GraphSAGE and Transformer layers.

[0190] First, add position information to the node features to enhance the model's perception of node order:

[0191]

[0192] Among them, PE(i,2k) is the k-th dimension position code of node i.

[0193] For the GraphSAGE layer, each layer updates the node representation by weighted aggregation of neighbor features:

[0194]

[0195] Among them, H (l) ∈R N×d is the node feature of the lth layer, W self and W neigh are the first parameter matrix and the second parameter matrix that can be learned, respectively, and σ is the activation function.

[0196] For the Transformer layer, the adjacency matrix A with weights fused (i,j) is added to the attention calculation as a bias term:

[0197]

[0198] A=H (l) W A , B=H (l) W B ,C=H (l) W C

[0199] Among them, A, B, and C are query matrix, key matrix, and value matrix respectively, and W A 、W B and W CThey are the learnable first projection matrix, the learnable second projection matrix and the learnable third projection matrix respectively; ε is a very small positive number (such as 1e -5 ), to prevent taking logarithms about zero; log(A fused (i,j)+ε) is used to convert edge weights into attention biases, where higher weights lead to higher attention scores.

[0200] In the Transformer layer, the gating coefficient (Sigmoid function) is used to dynamically adjust the influence of edge weights:

[0201] G=σ(W g ·H (l) ),H (l+1) =G⊙Attention(A,B,C)+(1-G)⊙H (l)

[0202] Among them, W g is a learnable gating parameter matrix; G is a gating coefficient matrix (values between 0 and 1). Higher-weighted edges correspond to larger feature updates. The gating mechanism determines how much information from edge weights to retain based on the current feature state.

[0203] After feature learning (alternating stacking of GraphSAGE and Transformer layers), the node features H after feature learning are output (L) ∈R N×d , N is the number of nodes, d is the feature dimension, and L is the total number of layers of feature learning (such as the number of layers of GraphSAGE+Transformer alternately stacked).

[0204] GraphSAGE and Transformer are stacked alternately, combined with local neighborhood aggregation and global attention mechanism to fully explore node relationships and improve feature learning efficiency.

[0205] (6) In the fault identification module, the node features after feature learning are input into the global average pooling layer to aggregate into a pooled global feature vector. After the high-level nonlinear features are extracted from the pooled global feature vector by the multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier. The specific method to obtain the fault category probability is as follows:

[0206] The node feature H after feature learning (L) ∈R N×d Enter the average pooling layer to convert the features of each node into a global feature vector at the graph level:

[0207]

[0208] in, is the eigenvector of the i-th node, h global is the global feature vector after pooling.

[0209] The pooled global feature vector h global Input MLP (Multi-layer Perceptron) for feature extraction. Specifically, in MLP, high-level features are further extracted through fully connected layers and nonlinear activation functions:

[0210] h mlp =ReLU(W mlp ·h global +b mlp )

[0211] in, is the weight matrix of MLP, b mlp is the bias vector of MLP, d hidden is the hidden layer dimension, It is the high-level nonlinear feature vector output by MLP.

[0212] The high-level nonlinear feature vector h output by MLP mlp Input Softmax classifier, the high-level nonlinear feature vector h mlp Mapping to class probability distribution:

[0213]

[0214] in, is the weight matrix of the Softmax classifier, b cls ∈R M is the bias vector of the Softmax classifier, p m ∈R M , p m is the probability of the mth type of failure, M is the number of failure categories;

[0215] The difference between the predicted probability and the true label is then calculated through the loss function:

[0216]

[0217] Among them, L cls is the cross entropy loss; y∈{0,1} M One-hot encoding is a method for converting categorical variables (CategoricalData) into binary vectors. Its core idea is to assign a unique vector to each category, with only one position in the vector being 1 (indicating "activation") and the rest being 0.

[0218] Global pooling is combined with MLP to compress node-level features into high-level representations, reduce the number of parameters, and adapt to small sample classification requirements.

[0219] like Figure 4 As shown, channel 1, channel 2, channel 3, all channels (based on the dynamic weighted fusion of channel 1, channel 2, channel 3, that is, the F obtained in this embodiment fused ) is the final output after the dynamic fusion of the three. In this embodiment, the Spectra-Quest mechanical fault simulation test bench is used to collect gearbox data. The data source of this embodiment is the gearbox data of Huazhong University of Science and Technology. The multi-sensor data of the gearbox is obtained by speed control sensor, motor sensor, and acceleration sensor. 11 types of fault state data are collected, including missing teeth, broken teeth under five loads (including 20, 25, 30, 35, and 40Hz) and normal gears under 20Hz load. In order to verify the effectiveness of the method proposed in the present invention, each fault state of the fault samples in the training set is set to 1, 3, 5, 10, and 20 respectively, while the normal samples are 30; the samples of each fault type in the test set are 30. The five training sets of 1, 3, 5, 10, and 20 correspond to Figure 4 In the field of deep learning, the F1 score is the harmonic mean of precision and recall, which is used to comprehensively reflect the model's ability to identify positive examples (such as fault classes). Data imbalance often occurs in small sample classification. The F1 score is relatively insensitive to data imbalance and can more objectively reflect the comprehensive performance of the model on minority and majority classes. Figure 4 It can be seen that the method proposed in the present invention maintains the highest F1 score compared with the method using data from any channel for model training.

[0220] Example 2

[0221] A method for identifying faults of multiple detection devices under small sample conditions, comprising:

[0222] Extracting time domain features from the time domain signals of the multiple detection devices of the gearbox, performing Fourier transform on the time domain signals of the multiple detection devices of the gearbox to extract frequency domain features, constructing a time domain graph and a frequency domain graph based on the K-nearest neighbor rule according to the time domain features and the frequency domain features, splicing the time domain graph and the frequency domain graph to obtain a positive graph, comparing the neighborhood node sets of the time domain graph and the frequency domain graph, and filtering out the neighborhood nodes unique to the frequency domain graph to construct a negative graph;

[0223] The positive image, negative image, and frequency domain image are input into the graph variational autoencoder respectively. In the graph variational autoencoder, the latent distribution is generated through multi-layer GraphSAGE encoding, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion of channel weights to obtain the fused node feature matrix.

[0224] According to the fusion node feature matrix, the edge weights of the fused features are calculated based on the K-nearest neighbor rule, and the adjacency matrix with edge weights is constructed according to the edge weights of the fused features;

[0225] The fused node feature matrix and the adjacency matrix with edge weights are input into the alternating stack of GraphSAGE and Transformer layers. The GraphSAGE layer updates the node representation by weighted aggregation of neighborhood features. The Transformer layer allows edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with a gating mechanism, outputting the node features after feature learning.

[0226] The node features after feature learning are input into the global average pooling layer to aggregate into a pooled global feature vector. After the high-level nonlinear features are extracted from the pooled global feature vector through a multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier to obtain the fault category probability.

[0227] Example 3

[0228] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the method for identifying faults of multiple detection devices under small sample conditions in Example 2.

[0229] The contents not described in detail in this specification belong to the prior art known to those skilled in the art. It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0230] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0231] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0232] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0233] 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 its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art may still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A multi-detection device fault identification system under small sample conditions, characterized in that: include: A graph structure construction module extracts time domain features from the time domain signals of the multiple detection devices of the gearbox, performs Fourier transform on the time domain signals of the multiple detection devices of the gearbox to extract frequency domain features, constructs a time domain graph and a frequency domain graph based on the time domain features and the frequency domain features using the K-nearest neighbor rule, concatenates the time domain graph and the frequency domain graph to obtain a positive graph, compares the neighborhood node sets of the time domain graph and the frequency domain graph, and filters out the neighborhood nodes unique to the frequency domain graph to construct a negative graph; The feature fusion module inputs the positive image, negative image, and frequency domain image into the graph variational autoencoder respectively. In the graph variational autoencoder, it undergoes multi-layer GraphSAGE encoding to generate potential distribution, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion of channel weights to obtain the fused node feature matrix; The fusion feature graph construction module calculates the edge weights of the fused features based on the K-nearest neighbor rule according to the fusion node feature matrix, and constructs an adjacency matrix with edge weights based on the edge weights of the fused features; The feature learning module inputs the fused node feature matrix and the adjacency matrix with edge weights into alternating stacks of GraphSAGE and Transformer layers. The GraphSAGE layer updates the node representation by weighted aggregation of neighborhood features. The Transformer layer allows edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with a gating mechanism, outputting the node features after feature learning. The fault identification module inputs the node features after feature learning into the global average pooling layer to aggregate them into a pooled global feature vector. After extracting high-level nonlinear features from the pooled global feature vector through a multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier to obtain the fault category probability.

2. The multi-detection device fault identification system under small sample conditions according to claim 1, characterized in that: In the graph structure construction module, time domain features are extracted from the time domain signals of the multiple detection devices of the gearbox, the time domain signals of the multiple detection devices of the gearbox are subjected to Fourier transform to extract frequency domain features, and the specific method for constructing the time domain graph and the frequency domain graph based on the K-nearest neighbor rule according to the time domain features and the frequency domain features is as follows: The time domain signal of the gearbox multi-detection device is recorded as X = {x1, x2, ..., x i }; Extract time domain features from the time domain signals of the gearbox multi-detection device, each sample x i Generate time domain feature vector d t Represents the time domain feature dimension, R is a real number, where the time domain features include mean, variance, peak, peak-to-peak value, root mean square, shape factor, kurtosis, and impulse factor; Perform Fourier transform FFT on the time domain signal, that is, for each sample x i Perform Fourier transform to obtain the spectrum F i =FFT(x i )∈R X , X is the spectrum resolution, and the following frequency domain features are constructed: Spectrum amplitude: Spectrum amplitude = |F i |; Band energy: divide the frequency band and calculate the energy of each band; frequency domain statistics: spectrum centroid, spectrum bandwidth, spectrum flatness. Then, the frequency domain feature vector is obtained based on the frequency domain characteristics. d f Represents the frequency domain feature dimension. Frequency domain features include spectrum amplitude, frequency band energy, and frequency domain statistics. Use Euclidean distance to calculate the feature space distance between node i and node j: Among them, D t (i,j) is the time domain feature distance matrix, D f (i, j) is the frequency domain feature distance matrix; Constructing time domain graph G based on K nearest neighbors t =(V,E t ) and the frequency domain graph G f =(V,E f ), V is a node set, each node V in the node set V i Corresponding to a gearbox detection signal sample x i , E t is the edge set of the time domain graph, E f is the edge set of the frequency domain graph; Based on the time domain feature distance matrix D t Construct the adjacency matrix A of the time domain graph t , based on the frequency domain feature distance matrix D f Construct the adjacency matrix A of the frequency domain graph f , as shown below: Among them, TopK means for each node i, from D t (i,:) or D f In (i,:), K nodes with the smallest distance are selected as neighbors, and the : symbol represents all elements in the corresponding row or column.

3. The multi-detection device fault identification system under small sample conditions according to claim 2, characterized in that: In the graph structure construction module, the time domain graph and the frequency domain graph are spliced together to obtain a positive graph, and the specific method for comparing the neighborhood node sets of the time domain graph and the frequency domain graph and filtering out the neighborhood nodes unique to the frequency domain graph to construct a negative graph is as follows: Splice the time domain image and frequency domain image into the positive image G pos =(V,E pos ), E pos is the edge set of the positive graph. The specific method is: the positive graph eigenvector The edge set E of the main graph pos =E t ∩E f , the adjacency matrix A of the positive graph pos =A f ∩A t ; By comparing the differences in neighborhood nodes between the time domain graph and the frequency domain graph, the unique neighborhood nodes in the frequency domain are screened to construct the negative graph G. neg =(V,E neg ), E neg The edge set of the negative graph is as follows: for each node i, the neighboring node set N of node i in the time domain graph is t (i) = {j | A t (i,j)=1}, the neighboring node set N of node i in the frequency domain graph f (i) = {j | A f (i,j)=1}, the neighborhood node set N of the negative graph neg (i) is the neighborhood node set N in the frequency domain f (i) Neighborhood node set N relative to the time domain t (i) A set of unique neighborhood nodes, node features of the negative graph Negative graph adjacency matrix A neg as follows:

4. The multi-detection device fault identification system under small sample conditions according to claim 3, characterized in that: In the feature fusion module, the positive image, negative image, and frequency domain image are input into the graph variational autoencoder respectively. In the graph variational autoencoder, the latent distribution is generated through multi-layer GraphSAGE encoding, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion processing of channel weights. The specific method to obtain the fusion node feature matrix is as follows: Perform multi-layer graph convolution on the positive image, negative image, and frequency domain image respectively, and gradually aggregate neighborhood information. The specific steps are as follows: in, They are respectively the positive graph node features, negative graph node features, and frequency domain graph node features output after neighborhood aggregation and nonlinear transformation. is the input positive graph node feature, negative graph node feature, frequency domain graph node feature, W (l) is the learnable multi-layer graph convolution weight, σ is the activation function, and i is the node v i The serial number, N pos (i) N neg (i) N f (i) are nodes v i In the positive graph, negative graph, and frequency domain graph, the domain node set j is the node v i The serial number of the domain node; AGGREGATE represents neighbor feature aggregation, which aggregates all neighbor features of node i to generate an aggregate vector for capturing local neighborhood information; CONCAT represents feature concatenation, which concatenates the node's own features with the aggregated neighbor features to form a richer joint representation; Then, the mean and covariance matrix of the latent variables are calculated independently for each graph through the multi-layer GraphSAGE encoder. The formula is as follows: Main image: Negative image: Frequency domain plot: Covariance matrix: in, is the feature matrix of the positive graph nodes after neighborhood aggregation and nonlinear transformation, d pos is the characteristic dimension of the positive image, is the negative graph node feature matrix after neighborhood aggregation and nonlinear transformation, d neg is the feature dimension of the negative graph, is the frequency domain graph node feature matrix after neighborhood aggregation and nonlinear transformation, d f is the characteristic dimension of the frequency domain graph, E pos 、E neg 、E f are the edge sets of the positive graph, negative graph, and frequency domain graph, respectively. is the mean of the latent variables of the positive graph, N is the number of nodes, d z is the latent space dimension, μ neg is the mean of the negative graph latent variable, μ f is the mean of the latent variable in the frequency domain graph, is the log variance of the latent variable of the positive graph, is the log variance of the negative graph latent variable, is the logarithmic variance of the latent variable in the time domain graph, is the dth diagonal in the covariance matrix of the latent variable z The variance of the dimensions, Encoder is a multi-layer GraphSAGE encoder; Then, the latent variables are sampled from the posterior distribution of the positive image, negative image, and frequency domain image. The formula is as follows: Where ⊙ represents element-by-element multiplication, ∈~N(0,I) represents the noise sampled from the standard normal distribution, is the identity matrix, z pos is the latent variable of the positive graph, z neg is the latent variable of the negative graph, z f is the latent variable of the frequency domain graph, and exp converts the logarithmic variance into variance; Then the latent variable z pos 、z neg 、z f Use multi-layer GraphSAGE decoders to reconstruct the original features and output the reconstructed frequency domain graph feature matrix l is the channel number: Calculate the channel weight using the following formula: Among them, α l is the weight of the frequency domain feature moment in channel l, ε is a very small positive number, is the mean square error of the lth channel, s l Indicates the standard deviation of the node feature of the lth channel, s l =std(F l ); Based on the weights of c channels, the multi-graph reconstruction features are weightedly fused to obtain the fused node feature matrix F fused :

5. The multi-detection device fault identification system under small sample conditions according to claim 4, characterized in that: Calculate the mean square error L between the reconstructed features and the original features of the positive image, negative image, and frequency domain image recon , measures the difference between the reconstructed features and the original features: Regularize the latent space, L KL is a regularization term that constrains the posterior distribution to be close to the standard normal distribution: Based on the contrast loss L contrast , enhance the similarity between the positive image and the frequency domain image, and suppress the correlation between the negative image and the positive image and the frequency domain image: Among them, sim represents cosine similarity; Total loss L total for: L total =L recon +βL KL +γL contrast Among them, β is the first hyperparameter, γ is the second hyperparameter, and the first hyperparameter β and the second hyperparameter γ are used to control the strength of regularization and contrastive learning respectively; The training is performed according to the following steps: Forward propagation is: input the node features and adjacency matrix of the positive graph, negative graph, and frequency domain graph, the encoder generates the mean and variance, reparameterizes the sampling to obtain the latent variables, the decoder reconstructs the features, calculates the fusion weights, and outputs the fusion feature matrix F fused ; Back propagation is: calculate the total loss and update the parameters of the encoder, decoder and channel weights.

6. The multi-detection device fault identification system under small sample conditions according to claim 5, characterized in that: In the fusion feature graph construction module, the specific method of calculating the edge weights of the fused features based on the K-nearest neighbor rule according to the fusion node feature matrix and constructing the adjacency matrix with edge weights according to the edge weights of the fused features is as follows: First calculate the Euclidean distance matrix of the fused features: Among them, D fused (i,j) is the fusion node feature matrix F fused The Euclidean distance between nodes i and j, is the fused feature vector of node i, is the fusion node feature matrix F fused The nth eigenvalue of the i-th node in , is the fusion node feature matrix F fused The nth eigenvalue of the jth node in ; Select neighbor sets based on K nearest neighbors: N fused (i)={j∣D fused (i,j)≤TopK(D fused (i,:),K)} Among them, N fused (i) is the fusion node feature matrix F fused The set of the nearest K neighbors of node i, TopK(D fused (i,:),K) is the fusion node feature matrix F fused The threshold of the smallest K distances among all sample distances of node i; Calculate edge weight W based on Euclidean distance fused (i,j), the formula is as follows: Among them, η is the bandwidth parameter, which is used to control the weight decay speed; Construct a weighted adjacency matrix A fused (i,j)∈R N×N , only the weights of the first K neighbors of each node are retained, and the remaining edge weights are 0:

7. The multi-detection device fault identification system under small sample conditions according to claim 6, characterized in that: In the feature learning module, the fused node feature matrix and the adjacency matrix with edge weights are input into the alternately stacked GraphSAGE and Transformer layers. The GraphSAGE layer updates the node representation by weighted aggregation of neighborhood features. The Transformer layer allows the edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with the gating mechanism. The specific method for outputting the node features after feature learning is as follows: The fusion node feature matrix F fused ∈R N×d and the adjacency matrix A with edge weights fused (i,j)∈R N×N Input alternately stacked GraphSAGE and Transformer layers; First, add position information to the node feature: Among them, PE(i,2k) is the k-th dimension position code of node i; For the GraphSAGE layer, each layer updates the node representation by weighted aggregation of neighbor features: Among them, H (l) ∈R N×d is the node feature of the lth layer, W self and W neigh are the first parameter matrix and the second parameter matrix that can be learned, respectively, and σ is the activation function. For the Transformer layer, the adjacency matrix A with weights fused (i,j) is added to the attention calculation as a bias term: A=H (l) W A ,B=H (l) W B ,C=H (l) W C Among them, A, B, and C are query matrix, key matrix, and value matrix respectively, and W A 、W B and W C are the learnable first projection matrix, the learnable second projection matrix and the learnable third projection matrix respectively; ε is a very small positive number; log(A fused (i,j)+ε) is used to convert edge weights into attention biases; In the Transformer layer, the gating coefficients are used to dynamically adjust the influence of edge weights: G=σ(W g ·H (l) ),H (l+1) =G⊙Attention(A,B,C)+(1-G)⊙H (l) Among them, W g is the learnable gating parameter matrix; G is the gating coefficient matrix; After alternately stacking GraphSAGE and Transformer layers, the node features H after feature learning are output (L) ∈R N×d , N is the number of nodes, d is the feature dimension, and L is the total number of layers for feature learning.

8. The multi-detection device fault identification system under small sample conditions according to claim 7, characterized in that: In the fault identification module, the node features after feature learning are input into the global average pooling layer to aggregate into a pooled global feature vector. After the high-level nonlinear features are extracted from the pooled global feature vector through a multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier. The specific method for obtaining the fault category probability is as follows: The node feature H after feature learning (L) ∈R N×d Enter the average pooling layer to convert the features of each node into a global feature vector at the graph level: in, is the eigenvector of the i-th node, h global is the global feature vector after pooling; The pooled global feature vector h global Input MLP for feature extraction, specifically in MLP, high-level features are further extracted through fully connected layers and nonlinear activation functions: h mlp =ReLU(W mlp h global +b mlp ) in, is the weight matrix of MLP, b mlp is the bias vector of MLP, d hidden is the hidden layer dimension, It is the high-level nonlinear feature vector output by MLP; The high-level nonlinear feature vector h output by MLP mlp Input Softmax classifier, the high-level nonlinear feature vector h mlp Mapping to class probability distribution: in, is the weight matrix of the Softmax classifier, b cls ∈R M is the bias vector of the Softmax classifier, p m ∈R M , p m is the probability of the mth type of failure, M is the number of failure categories; The difference between the predicted probability and the true label is then calculated through the loss function: Among them, L cls is the cross entropy loss; y∈{0,1} M is the one-hot encoding of the true label.

9. A method for identifying faults of multiple detection devices under small sample conditions, characterized in that: include: Extracting time domain features from the time domain signals of the multiple detection devices of the gearbox, performing Fourier transform on the time domain signals of the multiple detection devices of the gearbox to extract frequency domain features, constructing a time domain graph and a frequency domain graph based on the K-nearest neighbor rule according to the time domain features and the frequency domain features, splicing the time domain graph and the frequency domain graph to obtain a positive graph, comparing the neighborhood node sets of the time domain graph and the frequency domain graph, and filtering out the neighborhood nodes unique to the frequency domain graph to construct a negative graph; The positive image, negative image, and frequency domain image are input into the graph variational autoencoder respectively. In the graph variational autoencoder, the latent distribution is generated through multi-layer GraphSAGE encoding, reparameterized sampling, multi-layer GraphSAGE decoding reconstruction, and dynamic weighted fusion of channel weights to obtain the fused node feature matrix. According to the fusion node feature matrix, the edge weights of the fused features are calculated based on the K-nearest neighbor rule, and the adjacency matrix with edge weights is constructed according to the edge weights of the fused features; The fused node feature matrix and the adjacency matrix with edge weights are input into the alternating stack of GraphSAGE and Transformer layers. The GraphSAGE layer updates the node representation by weighted aggregation of neighborhood features. The Transformer layer allows edge weights to participate in the scaled dot product attention calculation and dynamically adjusts the weight influence in combination with a gating mechanism, outputting the node features after feature learning. The node features after feature learning are input into the global average pooling layer to aggregate into a pooled global feature vector. After the high-level nonlinear features are extracted from the pooled global feature vector through a multi-layer perceptron, the high-level nonlinear features are input into the Softmax classifier to obtain the fault category probability.

10. A computer program product, characterized in that It includes a computer program / instruction, characterized in that when the computer program / instruction is executed by a processor, it implements the method for identifying faults of multiple detection devices under small sample conditions as described in claim 9.

Images