Measurement-based cross-domain few-sample bearing fault diagnosis meta-learning relation network method

Abstract

The invention discloses a measurement-based cross-domain few-sample bearing fault diagnosis meta-learning relation network method, which comprises the following steps of: converting vibration signals from different working conditions into two-dimensional time-frequency images, dividing data samples into a meta-training set and a meta-test set according to a meta-learning training strategy, and further subdividing each set into a support set and a query set; a residual shrinkage non-local feature extraction module is designed and used for extracting and fusing features from a support set and a query set, and a similarity score between the support set and the query set is calculated by adopting a neural network of nonlinear measurement. According to the bearing fault diagnosis method, rapid and accurate bearing fault diagnosis can be realized under the condition of few samples, even under the unknown working condition and the condition of limited data samples.

Description

Technical Field

[0001] The present invention relates to the field of bearing fault diagnosis and a metric-based few-sample meta-learning technology, and in particular to a metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method. Background Art

[0002] Deep learning techniques have recently shown great potential in the field of bearing fault diagnosis. However, these methods must meet two key conditions: sufficient training data sets are available and the distribution characteristics between training and test data are similar. However, in practical applications, bearing fault data is very limited and the working conditions are variable, which greatly affects the accuracy of fault diagnosis. Therefore, solving the cross-domain few-sample problem is crucial.

[0003] Ren et al. proposed a few-shot generative adversarial network (GAN) that can solve the overfitting problem in the training process under severe data imbalance by utilizing classes with rich samples to enhance the diversity and accuracy of generated fault samples. Yang et al. proposed a conditional GAN ​​(CGAN) that generates labeled data by augmenting the original small-shot dataset and uses a two-dimensional convolutional neural network (2-D-CNN) for fault type classification, demonstrating considerable accuracy under limited samples. Liang et al. proposed a semi-supervised GAN-based method for diagnosing faults in rotating machinery using small labeled samples. Although GAN-based methods can alleviate the problems associated with data imbalance and scarcity of labeled samples, there is still a lack of reliable methods to evaluate the quality of generated sample data.

[0004] Zhang et al. first applied the twin network to few-sample bearing fault diagnosis. This method combined sample pairs to increase the training samples, alleviating the overfitting problem caused by insufficient samples. Feng et al. introduced a prototype network that exploits similarity and combines adversarial domain adaptation for cross-domain fault diagnosis. However, the similarity scores in the model are calculated using a fixed linear metric, which may degrade the diagnostic performance in complex scenarios. In addition, traditional deep transfer learning methods usually require extensive parameter fine-tuning for specific target tasks. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide a metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method that can solve the problem of cross-domain few-sample bearing fault diagnosis.

[0006] Technical solution: A metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method of the present invention comprises the following steps:

[0007] Step 1: Obtain bearing vibration signals under different working conditions, transform the one-dimensional time domain signal in the source domain bearing data set into a two-dimensional time-frequency diagram through continuous wavelet transform (CWT), and generate multiple meta-tasks according to the meta-learning training strategy. Each meta-task contains a support set and a query set. The support set contains known category faults, and the query set contains unknown category faults.

[0008] Step 2: construct and train the RSNL relationship network, which includes an RSNL feature extraction module, a feature fusion module and an NL relationship module;

[0009] First, the support set and query set in the meta-task obtained in step 1 are input into the RSNL relational network, and the RSNL feature extraction module based on residual shrinkage non-locality extracts features from the support set and query set respectively to obtain respective fault features. The RSNL feature extraction module can not only reduce the noise interference in the vibration signal, but also capture the internal correlation in the data and solve the remote dependency problem, thereby extracting effective signal features and improving the fault diagnosis accuracy; the RSNL feature extraction module is provided with an RS component and an RSNL model, the RS component is combined with a soft threshold operation and an attention mechanism to realize automatic determination of the threshold, the RSNL model is combined with an RS component (existing in the prior art) and an NL component, and the RSNL model extracts global information and key details of the fault features;

[0010] Then, the fault features of the support set and the query set are input into the feature fusion module for cross fusion to obtain the fused fault features. At the same time, a corresponding fault label is generated for each fused fault feature. The fused fault feature label of two features of the same type is recorded as 1, and the fused fault feature label of two features of different types is recorded as 0.

[0011] Next, the fused fault features are input into the NL relation module based on the NL component (NL component introduces the existing relation module to obtain the NL relation module), the corresponding relation score is obtained, and the cross entropy loss is calculated by combining the relation score with the fault label;

[0012] Finally, the model parameters of the RSNL feature extraction module and the NL relationship module are updated respectively, and training is performed until the model loss converges;

[0013] Step 3: Generate a meta-task in the target domain bearing dataset to be predicted. The generated meta-task includes a support set and a query set. The support set and the query set are respectively input into the RSNL relational network trained in step 2. After feature extraction, feature fusion, and relation score prediction, the category to which the target domain data belongs is finally output.

[0014] Furthermore, step 1 uses the CWT method to convert the one-dimensional time-domain vibration signal of the bearing into a two-dimensional time-frequency diagram, which not only enriches the fault information, but also enables the relational network model to better process the bearing data; when generating meta-tasks using the meta-learning training strategy, the N-way K-shot training method is adopted, N is the number of categories of each generated task, K is the number of support set samples of each category, and K is used as the number of samples of the known category; meta-learning can acquire cross-domain prior knowledge by learning from multiple different tasks through this training method, thereby showing good generalization and adaptability in the target domain.

[0015] Furthermore, when the RSNL feature extraction module processes the support set and the query set, it first performs convolution, batch normalization, ReLu activation function and maximum average pooling operations in sequence, and then inputs the RS component, RSNL model, RS component and RSNL model in sequence.

[0016] Furthermore, the NL component is based on the image filtering algorithm non-local means NLM algorithm, and the NL operation in the NL component is defined as:

[0017]

[0018] for where x i Represents the multi-channel feature vector of the i-th point in the feature map to be calculated, x j Represents x i The multi-channel feature vector of the jth point in the feature map of the neighborhood, f i Represents x i The multi-channel feature vector of the i-th point in the feature map is calculated; v(x j ) represents the transformation function used to obtain the mapping representation of the feature vector at the jth point in the feature map, g(x i ,x j ) represents the feature vector x i and x j Calculate the similarity between them;

[0019] Here we use the dot product method to calculate the similarity g(x i ,x j ), the formula is as follows:

[0020]

[0021] in(·) T represents the transpose of a vector or matrix, Represents a normalization operation;

[0022] According to the above two formulas, when all feature vectors x in the feature map are calculated at the same time iThe output v(x j ), the matrix form formula can be used as follows:

[0023]

[0024] Where F represents the eigenvector f after all calculations i The matrix composed of i Represents all feature vectors x from the feature map to be calculated i The matrix composed of j Represents all feature vectors x from adjacent feature maps j The matrix of the group layer; here X i and X j are all two-dimensional matrices; since the feature map input to the NL component is a three-dimensional feature map tensor Therefore, it is necessary to Perform a flattening operation, namely:

[0025] When calculating the similarity between feature vectors in a feature map, the dot product method only considers the product of two feature vectors, which limits its ability to capture complex relationships in the original feature map. We further introduce an extended dot product method, called embedded dot product. The feature map composed of feature vectors Through nonlinear mapping θ(·) and φ(·) are embedded into a new space for computing similarity, which is defined as:

[0026]

[0027]

[0028] The embedding operation can not only capture the complex relationships in the feature graph and extract more meaningful features, but also reduce the dimension and improve the computational efficiency.

[0029] The NL component formula is finally derived from the above formula:

[0030]

[0031] Here, the embedded operations θ(·) and φ(·) are implemented using 1×1 2D convolutions, which can accurately map each feature vector in the feature map to a new space. In order to better represent each feature vector in the feature map and to facilitate matching with the calculated similarity weight matrix, v(·) is also implemented using 1×1 2D convolutions.

[0032] In order to obtain a relatively stable similarity weight, the softmax function is used.

[0033] Furthermore, the specific method for the feature fusion module in step 2 to cross-fuse the fault features of the support set and the query set is: channel-fusing the multi-channel fault features of each sample in the support set with the multi-channel fault features of each sample in the query set to ensure that each sample in the support set and each sample in the query set are fused with each other.

[0034] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0035] 1. The present invention is achieved by developing a metric-based meta-learning relational network that can quickly adapt to new tasks by leveraging meta-knowledge learned from known working condition data.

[0036] 2. The RSNL feature extraction module of the present invention includes a new residual shrinkage non-local module, which is composed of residual shrinkage and non-local components. It can use soft thresholds to reduce noise interference and is good at capturing global information, thereby effectively solving long-distance dependency problems.

[0037] 3. In addition, the NL component of the RSNL feature extraction module is integrated in the NL relationship module, which further enhances the nonlinear measurement ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the overall steps of the present invention;

[0039] Figure 2 It is a schematic diagram of the network structure of the present invention;

[0040] Figure 3 It is a structural diagram of the RSNL feature extraction module of the present invention;

[0041] Figure 4 is a structural diagram of the RS component of the present invention;

[0042] Figure 5 is a structural diagram of the NL component of the present invention;

[0043] Figure 6 A schematic diagram of a one-dimensional time domain signal of an embodiment;

[0044] Figure 7 A two-dimensional time-frequency diagram visualization diagram of an embodiment;

[0045] Figure 8 This is a feature visualization diagram of each model in the embodiment based on T-SNE dimensionality reduction. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0047] like Figure 1 As shown, the metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method of the present invention includes the following steps:

[0048] Step 1: Obtain bearing vibration signals under different working conditions, transform the one-dimensional time domain signal in the source domain bearing data set into a two-dimensional time-frequency graph through continuous wavelet transform (CWT), and generate multiple meta-tasks from the two-dimensional time-frequency graph according to the meta-learning training strategy. Each meta-task contains a support set and a query set. The support set contains known category faults, and the query set contains unknown category faults.

[0049] Step 2: construct and train the RSNL relationship network, which includes an RSNL feature extraction module, a feature fusion module and an NL relationship module;

[0050] First, the support set and query set in the meta-task obtained in step 1 are input into the RSNL relational network, and the RSNL feature extraction module based on residual shrinkage non-locality extracts features from the support set and query set respectively to obtain respective fault features; the RSNL feature extraction module is provided with an RS component and an RSNL model, the RS component is combined with a soft threshold operation and an attention mechanism to realize automatic determination of the threshold, the RSNL model is combined with an RS component and an NL component, and the RSNL model extracts global information and key details of the fault features;

[0051] Then, the fault features of the support set and the query set are input into the feature fusion module for cross fusion to obtain the fused fault features. At the same time, a corresponding fault label is generated for each fused fault feature. The fused fault feature label of two features of the same type is recorded as 1, and the fused fault feature label of two features of different types is recorded as 0.

[0052] Next, the fused fault features are input into the NL relation module based on the NL component to obtain the corresponding relation score, and the cross entropy loss is calculated by combining the relation score with the fault label;

[0053] Finally, the model parameters of the RSNL feature extraction module and the NL relationship module are updated respectively, and training is performed until the model loss converges;

[0054] Step 3: Generate a meta-task in the target domain bearing dataset to be predicted. The generated meta-task includes a support set and a query set. The support set and the query set are respectively input into the RSNL relational network trained in step 2. After feature extraction, feature fusion, and relation score prediction, the category to which the target domain data belongs is finally output.

[0055] like Figures 2 to 5As shown, when the RSNL feature extraction module of this embodiment processes the support set and the query set, it first performs convolution (convolution kernel is 3×3), 2D batch normalization with 64 channels, ReLu activation function and maximum average pooling (3 channels) operations in sequence, and then inputs the RS component, RSNL model, RS component and RSNL model in sequence. The input and output channel sizes of the RSNL feature extraction module are 3 and 64 respectively.

[0056] The NL component of this embodiment is based on the image filtering algorithm non-local mean NLM algorithm, and the NL operation in the NL component is defined as:

[0057]

[0058] for x i Represents the multi-channel feature vector of the i-th point in the feature map to be calculated, x j Represents x i The multi-channel feature vector of the jth point in the feature map of the neighborhood, f i Represents x i The multi-channel feature vector of the i-th point in the feature map is calculated; v(x j ) represents the transformation function used to obtain the mapping representation of the feature vector at the jth point in the feature map, g(x i ,x j ) represents the feature vector x i and x j Calculate the similarity between them;

[0059] Here we use the embedded dot product method to calculate the similarity g(x i ,x j ), the formula is as follows:

[0060]

[0061] in(·) T represents the transpose of a vector or matrix, Represents a normalization operation;

[0062]

[0063] Finally, the NL component formula is:

[0064]

[0065] v(·), the embedded operations θ(·) and φ(·) are implemented using 1×1 2D convolutions.

[0066] The present invention constructs a metric-based meta-learning relational network (RSNL relational network), which can quickly adapt to new tasks by utilizing meta-knowledge learned from known working condition data; the RSNL feature extraction module includes a new residual shrinkage non-local module, which consists of residual shrinkage and non-local components. The RSNL feature extraction module can use soft thresholds to reduce noise interference and is good at capturing global information, thereby effectively solving long-distance dependency problems. In addition, the proposed NL component is integrated into the relational module, further enhancing the nonlinear metric capability of the model.

[0067] This embodiment performs a certain bearing vibration information (such as Figure 6 and Figure 7 ) to perform a diagnostic analysis, including the following steps:

[0068] Step 1: Convert the one-dimensional time domain signal in the source domain bearing dataset into a two-dimensional time-frequency diagram through continuous wavelet transform (CWT) with a size of 3*224*224. Generate multiple meta-tasks according to the meta-learning training strategy. Each meta-task contains a support set and a query set. The support set is a known category, and the query set is an unknown category.

[0069] Step 2: Input the support set and the query set into the RSNL relational network of the training number respectively, and the RSNL feature extraction module extracts the corresponding fault features respectively.

[0070] The multi-channel fault features of each sample in the support set are channel-fused with the multi-channel fault features of each sample in the query set, and then the fault features of the support set and the query set are cross-fused to ensure that each sample in the support set is fused with each sample in the query set to obtain a fused fault feature. At the same time, a corresponding label is generated for each fused feature. Two features of the same type are fused into "1", and two features of different types are fused into "0".

[0071] The fused fault features are input into the NL relation module to obtain the corresponding relation score, which is then jointly calculated with the fault label to calculate the cross entropy loss. Then, the model parameters of the RSNL feature extraction module and the NL relation module are updated respectively, and training is performed until the model loss converges.

[0072] like Figure 3 to Figure 4 As shown, the RSNL feature extraction module of this embodiment is provided with two groups of RS components and RSNL models. The RSNL model combines RS components and NL components. The RS components combine soft threshold operation and attention mechanism to automatically determine the threshold, thereby reducing noise interference in the vibration signal. The NL component can not only consider the global information and key details of the fault characteristics, but also capture the internal correlation in the data and solve the remote dependency problem.

[0073] The input of the NL component is the feature map Where C represents the number of channels, while H and W are the height and width of the feature size, respectively.

[0074] First, use f Conv2d (·) The feature map is embedded into the new space, and then flattened to obtain the query matrix Q and key matrix K. In addition, through f Conv2d (·) Process the feature map to obtain a representation of the feature map and apply a flattening operation to obtain the value matrix V. Compute the similarity score between each element in the feature map by multiplying the transpose of the key matrix K with the query matrix Q.

[0075] Then, the softmax operation is applied for normalization to obtain the global weight matrix. Then, the transpose of the value matrix V is multiplied by the global weight matrix obtained by the query matrix Q and the key matrix K. After transposition and unflattening operations, the NL feature map is obtained.

[0076] Finally, to preserve the information in the input feature map and prevent gradient vanishing or gradient exploding during training, a residual structure is added; the number of channels is increased to C by using a 2D convolution operation, ensuring that the resulting NL feature map matches the spatial dimensions of the input feature map. The NL feature map is element-wise added to the original feature, producing the output

[0077] In order to verify the performance of the technical solution of the present invention, this embodiment applies the technical solution of the present invention to specific data sets for verification. The data sets used include: Case Western Reserve University (CWRU) bearing data set, Paderborn University (PU) bearing data set, and self-built platform (PT) bearing data set.

[0078] CWRU dataset: This dataset is widely used to evaluate the performance of various methods in bearing fault diagnosis. The vibration signal data from the bearing is sampled at a frequency of 12kHz. The single-point damage is caused by electrospark machining. Four different bearing health states. In addition, four different load working conditions are evaluated. For each load condition, there are 10 different health states. The model learning process uses 200 subsamples for each class of each task. Fault detection under various working conditions can be regarded as different tasks, and each task represents a multi-class classification problem. The description of each task is shown in Table 1.

[0079] Table 1

[0080] Task load Number of categories Sample length Total number of samples <![CDATA[T0]]> 0 10 2048 200*10 <![CDATA[T1]]> 1 10 2048 200*10 <![CDATA[T2]]> 2 10 2048 200*10 <![CDATA[T3]]> 3 10 2048 200*10

[0081] PT Dataset: The bearing fault test bench of the self-built platform consists of a motor, a main shaft, a gearbox, and two supporting bearing seats. The bearings are divided into five categories according to the specific fault location: normal bearings, inner ring fault bearings, outer ring fault bearings, rolling element fault bearings, and cage fault bearings. In order to verify the cross-domain fault diagnosis performance of the technical solution, the dataset is configured for four different rotation speeds (750rpm, 1000rpm, 1250rpm, and 1500rpm), resulting in four different working conditions. For each condition, a subsampling window is used to generate 100 subsamples, each containing 10240 data points, and the sampling frequency is 65536Hz. The description of each task is shown in Table 2.

[0082] Table 2

[0083]

[0084] PU dataset: This experiment is designed to verify the detection of bearing health status under various operating parameter settings as different multi-class classification tasks using bearing vibration data with artificial EDM damage under four different working conditions. Each task includes three different types of health status: undamaged (UD), inner race (IR) fault, and outer race (OR) fault. For each task, 200 samples per class are generated by subsampling. The sampling frequency is 64KHz. The description of each task is shown in Table 3

[0085] Table 3

[0086]

[0087]

[0088] The cross-domain few-sample bearing fault diagnosis meta-learning relational network of the present invention runs under the PyTorch deep learning framework. The hardware uses a single Nvidia GeForce RTX 4070Ti GPU, 32G running memory, and the operating system is Windows 10. Based on different scenarios, the number of samples in the support set is specified as 1, 5, and 10 (K = 1, 5, and 10). The model training adopts an optimization algorithm with a learning rate of 0.001 and a decay of 50% every 1000 iterations until convergence and stopping training.

[0089] The comparison results of the ablation test diagnostic performance of the technical solution of the present invention on the CWRU dataset are shown in Table 4.

[0090] Table 4

[0091] Model CWT RS NL <![CDATA[T0→T1]]> <![CDATA[T0→T2]]> <![CDATA[T0→T3]]> <![CDATA[T2→T3]]> RN 91.63％ 90.61％ 88.93％ 89.62％ Model I √ √ 98.83％ 96.64％ 95.78％ 98.40％ Model II √ √ 98.76％ 96.39％ 96.05％ 94.86％ Model III √ √ 93.36％ 93.05％ 90.41％ 91.34％ RSNLRN √ √ √ 98.93％ 98.43％ 97.68％ 98.96％

[0092] As can be seen from Table 4, all five models achieved an accuracy of more than 88% in a single scenario, demonstrating the good generalization ability of the classic RN. First, the RSNLRN network of the present invention is compared with Model I and Model II. The results show that under the T0→T3 working condition, the classification accuracy of the RN with the designed RS and NL components is improved by 1.9% and 1.63% respectively compared with the latter two models. The RSNL module not only alleviates noise interference, but also captures global features and enhances long-distance dependencies. Compared with Model III, the introduction of the CWT module significantly enhances the diagnostic accuracy of the RSNLRN.

[0093] To further verify the fault diagnosis performance of the network model of the present invention, in the T0→T3 experiment of a single task, this embodiment uses the t-distributed stochastic neighbor embedding (TSNE) method to reduce the dimension of the feature space of the query set samples. The corresponding visualization results after dimensionality reduction in this embodiment are as follows: Figure 8 As shown, Figure 8 (a), (b) and (c) are the effect diagrams of Model I, Model II and Model III respectively. Figure 8 (d) is a model effect diagram of the present invention.

[0094] The comparison results of the diagnostic performance of the inventive technical solution under variable working conditions and noise conditions on the CWRU dataset are shown in Table 5.

[0095] Table 5

[0096]

[0097]

[0098] In Table 5, from left to right, it is obvious that the performance of all models decreases as the noise increases. On the contrary, the higher the SNR, the higher the accuracy of all models. From top to bottom, the RSNLRN method of the present invention consistently outperforms most other compared methods under different working conditions. In the T0→T1 experiment, the accuracy of the ML-based method remains relatively stable in the range of 0dB to 8dB SNR. Even in the case of single learning, the accuracy change remains within 20%, and the present invention achieves the highest accuracy. However, as the difference between the working conditions increases, the accuracy of all models decreases. This decline can be attributed to the limited adaptive learning ability of the model when faced with a small number of tasks under increasingly complex new conditions.

[0099] The comparative experimental results of the few-sample fault classification of the present invention under different working conditions of the PT dataset are shown in Table 6.

[0100] Table 6

[0101]

[0102]

[0103] As can be seen from Table 6, the best accuracies of the present invention in 1-shot, 5-shot and 10-shot learning scenarios are 95.62%, 96.01% and 98.83%, respectively, all of which exceed the performance of the other four ML-based methods. Since both PN and SN methods rely on fixed linear metrics to calculate sample similarity, they suffer some accuracy loss in fault diagnosis under complex conditions. In contrast, the present invention utilizes a learnable nonlinear classifier, which provides greater flexibility.

[0104] The performance comparison of different methods of the present invention in different task combinations on the PU dataset is shown in Table 7.

[0105] Table 7

[0106]

[0107] As can be seen from Table 7, by training on data from four different working conditions, the network model of the present invention achieves a classification accuracy of more than 99% in all cases, except for the combination In addition, it surpasses the other four methods. These results show that the proposed method shows superior performance in this experiment.

[0108] The above experimental results show that the present invention has higher accuracy and efficiency in bearing fault detection under different operating conditions, and effectively alleviates the overfitting problem that often occurs due to limited sample size.

Claims

1. A metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method, characterized in that: The following steps are involved: Step 1: Obtain bearing vibration signals under different working conditions, transform the one-dimensional time domain signal in the source domain bearing data set into a two-dimensional time-frequency graph through continuous wavelet transform (CWT), and generate multiple meta-tasks from the two-dimensional time-frequency graph according to the meta-learning training strategy. Each meta-task contains a support set and a query set. The support set contains known category faults, and the query set contains unknown category faults. Step 2: construct and train the RSNL relationship network, which includes an RSNL feature extraction module, a feature fusion module and an NL relationship module; First, the support set and query set in the meta-task obtained in step 1 are input into the RSNL relational network, and the RSNL feature extraction module based on residual shrinkage non-locality extracts features from the support set and query set respectively to obtain respective fault features; the RSNL feature extraction module is provided with an RS component and an RSNL model, the RS component is combined with a soft threshold operation and an attention mechanism to realize automatic determination of the threshold, the RSNL model is combined with an RS component and an NL component, and the RSNL model extracts global information and key details of the fault features; Then, the fault features of the support set and the query set are input into the feature fusion module for cross fusion to obtain the fused fault features. At the same time, a corresponding fault label is generated for each fused fault feature. The fused fault feature label of two features of the same type is recorded as 1, and the fused fault feature label of two features of different types is recorded as 0. Next, the fused fault features are input into the NL relation module based on the NL component to obtain the corresponding relation score, and the cross entropy loss is calculated by combining the relation score with the fault label; Finally, the model parameters of the RSNL feature extraction module and the NL relationship module are updated respectively, and training is performed until the model loss converges; Step 3: Generate a meta-task in the target domain bearing dataset to be predicted. The generated meta-task includes a support set and a query set. The support set and the query set are respectively input into the RSNL relational network trained in step 2. After feature extraction, feature fusion, and relation score prediction, the category to which the target domain data belongs is finally output.

2. The metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method according to claim 1 is characterized in that: Step 1: In step 1, the one-dimensional time-domain vibration signal of the bearing is converted into a two-dimensional time-frequency diagram using the CWT method; When generating meta-tasks using the meta-learning training strategy, the N-way K-shot training method is adopted, where N is the number of categories of tasks generated each time, K is the number of support set samples for each category, and K is used as the number of samples of known categories.

3. The metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method according to claim 1 is characterized in that: When the RSNL feature extraction module processes the support set and query set, it first performs convolution, batch normalization, ReLu activation function and maximum average pooling operations in sequence, and then inputs the RS component, RSNL model, RS component and RSNL model in sequence.

4. According to the metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method of claim 1 or 3, the NL component is based on the image filtering algorithm non-local mean NLM algorithm, and the NL operation in the NL component is defined as: for x i Represents the multi-channel feature vector of the i-th point in the feature map to be calculated, x j Represents x i The multi-channel feature vector of the jth point in the feature map of the neighborhood, f i Represents x i The multi-channel feature vector of the i-th point in the feature map is calculated; v(x j ) represents the transformation function used to obtain the mapping representation of the feature vector at the jth point in the feature map, g(x i ,x j ) represents the feature vector x i and x j Calculate the similarity between them; Here we use the dot product method to calculate the similarity g(x i ,x j ), the formula is as follows: in(·) T represents the transpose of a vector or matrix, Represents a normalization operation; When all feature vectors x in the feature map are calculated simultaneously i The output v(x j ), the matrix form formula is as follows: in, F represents the eigenvector f after all calculations i The matrix composed of i Represents all feature vectors x from the feature map to be calculated i The matrix composed of j Represents all feature vectors x from adjacent feature maps j The matrix of the group layer; Since the feature map input to the NL component is a three-dimensional feature map tensor Therefore, it is necessary to Perform a flattening operation, namely:

5. The metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method according to claim 5 is characterized in that: When using the dot product method to calculate the similarity between feature vectors in the feature map, the embedded dot product is introduced. The specific method is: Feature map composed of feature vectors Through nonlinear mapping θ(·) and φ(·) are embedded into a new space for computing similarity, which is defined as: Finally, the theoretical formula of NL component is obtained as: v(·), the embedded operations θ(·) and φ(·) are implemented using 1×1 2D convolutions.

6. The metric-based cross-domain few-sample bearing fault diagnosis meta-learning relational network method according to claim 1 is characterized in that: The specific method of the feature fusion module in step 2 for cross-fusing the fault features of the support set and the query set is: channel-fusing the multi-channel fault features of each sample in the support set with the multi-channel fault features of each sample in the query set to ensure that each sample in the support set and each sample in the query set are fused with each other.

Images