A small sample bearing fault diagnosis method based on dual graph network

Abstract

The present invention relates to a small-sample bearing fault diagnosis method based on a dual-graph network. The method involves selecting a test sample and a training sample of bearing faults, preprocessing them, collecting data points, and constructing a metadata set. A convolutional neural network is then used to extract features from the metadata set. The extracted features and the annotation information of the training samples are then input into the dual-graph network to determine the fault type of the test sample. Compared with existing technologies, this method offers advantages such as high accuracy and wide applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a small sample bearing fault diagnosis method based on a dual-graph network. Background Art

[0002] In recent years, industrial facility health diagnostics has received extensive attention in the field of smart manufacturing. Specifically, the diagnosis of rotating machinery faults is crucial because they are key components of facilities and are very sensitive to failure.

[0003] Small-sample fault diagnosis involves fault diagnosis and classification with limited labeled data. Many deep learning methods have been proposed, such as meta-learning, embedding learning, and generative modeling, but few have been successful in small-sample bearing fault diagnosis. However, directly applying traditional fault diagnosis methods to small-sample fault diagnosis can lead to problems such as overfitting and repeated training, which severely impacts the accuracy of fault diagnosis. Summary of the Invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a small sample bearing fault diagnosis method based on a dual-graph network.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A small sample bearing fault diagnosis method based on a dual graph network includes the following steps:

[0007] S1. Selecting a test sample and a training sample of a bearing fault, wherein both the test sample and the training sample have labeling information;

[0008] S2, preprocess the test samples and training samples to construct a metadata dataset;

[0009] S3, extract features of the metadata dataset using convolutional neural networks;

[0010] S4. Input the extracted features and the annotation information of the training samples in step S3 into a dual-graph network, wherein the dual-graph network includes a point graph and a distribution graph, calculate the final edge features of the point graph, and perform class prediction for each test sample when the loss value of the dual-graph network is minimized.

[0011] Furthermore, the pre-processing step includes:

[0012] C1. Randomly select W*W continuous signal data points consisting of N continuous signal segments and N continuous signal points from all samples;

[0013] C2. Normalize the values ​​of all continuous signal data points, and fill the normalized values ​​with the pixel values ​​at the corresponding positions in all sample images to obtain a filled image;

[0014] C3. Convert the filled image into a grayscale image and stack it to form an image of size W*W*M as the metadata set.

[0015] Furthermore, the class prediction is calculated by feeding the edge into the function, and the calculation expression is:

[0016]

[0017] Where, Represents the final edge feature of the point graph, x i represents the characteristics of the i-th test sample, Represents x i The probability distribution of the fault type of the sample represented by y j Represents the annotation information of the j-th training sample, onehot represents one-hot encoding; NK represents the dimension of distribution-level node features.

[0018] Furthermore, the final edge feature calculation steps of the point graph are as follows:

[0019] B1. Obtain the initial node features of the point graph based on the features of the metadata dataset;

[0020] B2. Based on the initial node features of the point graph, the initial edge features of the point graph are calculated in combination with the encoding network;

[0021] B3. Based on the labeled information of the training samples, the initial node features of the distribution graph are obtained, and based on the initial edge features of the point graph, the final node features of the distribution graph are iteratively calculated;

[0022] B4. Based on the final node features of the distribution graph, the edge features of the distribution graph are calculated in combination with the encoding network;

[0023] B5. Based on the edge features of the distribution graph, the final node features of the point graph are calculated in combination with the aggregation network, and the final edge features of the point graph are calculated in combination with the encoding network;

[0024] B6. Based on the final edge features of the point graph, predict the class of each test sample.

[0025] Furthermore, the dual-graph network loss value is a weighted sum of point loss and distribution loss.

[0026] Furthermore, the point loss expression is as follows:

[0027]

[0028] In the formula, pg represents the point graph, l represents the number of iterations of the network, and x i represents the i-th sample, y i Indicates the labeling information of the i-th sample, L crossentropy represents the cross entropy loss function.

[0029] Furthermore, the distribution loss expression is as follows:

[0030]

[0031] In the formula, dg represents the distribution graph, l represents the number of iterations of the network, onehot represents the single hot encoding, and y i Represents the annotation information of the i-th sample, NK represents the dimension of the distribution-level node feature, Indicates the final edge feature in the distribution graph at the lth iteration, L crossentropy represents the cross entropy loss function.

[0032] Furthermore, the test samples and training samples are selected from the Case Western Reserve bearing dataset.

[0033] Furthermore, the initial node features of the distribution graph The calculation expression is as follows:

[0034]

[0035] Where, represents a real number, || represents the concatenation operator, δ(·) represents the Kronecker function, y i and y j represents the annotation information, and NK represents the dimension of the distribution-level node feature.

[0036] Furthermore, the ratio of the number of test samples to training samples is 7:3.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] 1. The present invention selects the Case Western Reserve bearing dataset as test and training samples, and uses a convolutional neural network for feature extraction, which facilitates the extraction of various types of features, and uses a dual-graph network for fault diagnosis and prediction, which has good accuracy. The dual-graph network combines distribution propagation, effectively combining labeled data with features, has strong applicability, and can be transferred to other labeled data.

[0039] 2. The present invention preprocesses both the test samples and the training samples and extracts grayscale images, thereby further improving the accuracy of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1It is a schematic diagram of the process of the present invention.

[0041] Figure 2 Schematic diagram of the pretreatment process in step S2 of the present invention.

[0042] Figure 3 This is a schematic diagram of converting a point graph into a distribution graph in a dual-graph network of the present invention.

[0043] Figure 4 Schematic diagram of converting a distribution graph into a point graph in the dual-graph network of the present invention. DETAILED DESCRIPTION

[0044] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0045] This embodiment provides a small sample bearing fault diagnosis method based on a dual-graph network, such as Figure 1 As shown, the following steps are included:

[0046] Step S1: 50 samples are extracted from each of the 16 fault types in the public Case Western Reserve bearing dataset. To ensure the accuracy of the test results, training samples and test samples are selected in a ratio of 7:3.

[0047] Step S2: preprocess the training and test samples. Figure 2 As shown, W*W continuous signal data points consisting of N continuous signal segments and N continuous signal points are randomly extracted from the original fault signal.

[0048] Normalize the value of each original fault signal data point from 0 to 255. Therefore, each data point fills each pixel in the converted sample image accordingly, and the W signal segments fill the W rows of the image in sequence. The pixel value of the sample image can be obtained by the following formula:

[0049]

[0050] Where value(i,j) is the pixel value of the sample image, τ is the value of signal normalization, and Round is the closest integer function.

[0051] Finally, the sample images are converted into grayscale images and stacked to form W*W*3 images to construct the metadata dataset.

[0052] Step S3: Use a convolutional neural network to extract features from the metadata dataset, and concatenate the extracted features with the annotation information of all samples to establish a support set and a query set, respectively. The support set is the training sample, and the query set is the test sample. The support set can be expressed as:

[0053] S={(x1,y1),(x2,y2),...,(x N×K ,y N×K )},

[0054] The query set can be represented as:

[0055] Q={(x N×K+1 ,y N×K+1 ),...,(x N×K+T ,y N×K+T )}

[0056] Where x i Represents the feature, y i Represents the labeling information, N represents the number of fault categories, K represents that each fault category has K samples, and T represents the number of samples in the query set, that is, the number of test samples.

[0057] It is worth noting that y in the query set Q i It only affects the comparison of results after the prediction is completed and does not participate in the training calculation process of the network.

[0058] Step S4: Input the support set and query set into the dual-graph network. The dual-graph network consists of a point graph and a distribution graph. The point graph is denoted as The distribution map is recorded as Where G represents the graph network, V represents the node feature set of all iterations, E represents the edge feature set of all iterations, pg represents the point graph, dg represents the distribution graph, l represents the number of iterations of the network, v represents the node feature, e represents the edge feature, i, j = 1,..., T, T represents the number of samples in the query set.

[0059] Initialize node features in the point graph using features from the support set extracted by the convolutional neural network This can be called the initial node feature, which can be obtained by the following formula:

[0060]

[0061] Where, m represents the dimension of feature embedding, represents a real number, x i Represents the features of the support set.

[0062] Instance-level similarity is calculated by node features in the dot graph, that is, edge features in the dot graph. The edge features in the dot graph in the first round of iteration It can be obtained by the following formula:

[0063]

[0064] Where, represents an encoding network that converts instance-level similarity into a specific scale.

[0065] Edge features in the graph at the time of the first iteration That is, the initial edge feature can be obtained by the following formula:

[0066]

[0067] Where, represents an encoding network that converts instance-level similarity into a specific scale, Represents a real number. In order to combine the edge feature information with the point graph as a whole, it is also necessary to Perform normalization operation.

[0068] After the above conversion from point map to distribution map, such as Figure 3 As shown, the node features in the distribution graph during the first round of iteration That is, the initial node characteristics of the distribution graph can be obtained by the following formula:

[0069]

[0070] Where, || represents the concatenation operator, δ(·) represents the Kronecker function, and y i and y j represents the annotation information, and NK represents the dimension of the distribution-level node feature.

[0071] Node features in the distribution graph at the lth iteration That is, the final node feature of the distribution graph can be obtained by the following formula:

[0072]

[0073] Where, Represents an aggregate network of distribution graphs, applies a join operation between two features, and then performs conversion.

[0074] Then calculate the distribution level similarity between the node features in the distribution graph, that is, the edge features in the distribution graph. In the first round of iteration, the edge features in the distribution graph It can be obtained by the following formula:

[0075]

[0076] Where, Represents an encoding network that transforms distribution-level similarity.

[0077] Edge features in the distribution graph at the lth iteration It can be obtained by the following formula:

[0078]

[0079] Where, Represents an encoding network that transforms distribution-level similarity, and then Perform normalization operation.

[0080] Finally, the distribution map is converted to a point map, such as Figure 4 As shown, a final point graph is generated again to build a more discriminative node representation.

[0081] Node features in the graph at the time point of the lth iteration It can also be called the final node feature and can be obtained by the following formula:

[0082]

[0083] Where, Represents an aggregate network of point graphs.

[0084] Finally, the class prediction for the query set can be calculated by feeding the edges of the final l iterations into the softmax function:

[0085]

[0086] Where, Represents the final edge feature of the point graph. The calculation method is the same as that of the initial edge feature. i represents the sample features of the query set, Represents x i The probability distribution of the fault type of the sample represented by y j represents the labeling information of the j-th sample in the support set, and onehot represents unique hot encoding. By comparing the fault type corresponding to the largest value in the probability distribution with the labeling information in the query set, its accuracy can be verified.

[0087] In order to ensure the reliability of the prediction, it is necessary to introduce the concept of dual-graph network loss value. First, the point loss of the lth iteration can be obtained by the following formula:

[0088]

[0089] Where, L crossentropy is the cross entropy loss function;

[0090] In order to facilitate convergence, the distribution loss is included. The distribution loss of the lth iteration can be obtained as follows:

[0091]

[0092] Where, Represents the edge features in the l-th iteration distribution graph;

[0093] The loss function of the overall dual-graph network is the weighted sum of point loss and distribution loss, which can be obtained as follows:

[0094]

[0095] Where, Represents the number of iterations of the dual-graph network, weight λ p and λ d are set to 1.0 and 0.1 respectively. The class prediction is made for the query set when the loss function value is minimized.

[0096] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A small sample bearing fault diagnosis method based on dual graph network, characterized in that: The following steps are involved: S1. Selecting a test sample and a training sample of a bearing fault, wherein both the test sample and the training sample have labeling information; S2, preprocess the test samples and training samples to construct a metadata dataset; S3, extract features of the metadata dataset using convolutional neural networks; S4. Input the features extracted in step S3 and the annotation information of the training samples into a dual-graph network, wherein the dual-graph network includes a point graph and a distribution graph, calculate the final edge features of the point graph, and perform class prediction for each test sample when the loss value of the dual-graph network is minimized; The pre-processing step comprises: C1. Randomly select W*W continuous signal data points consisting of N continuous signal segments and N continuous signal points from all samples; C2. Normalize the values ​​of all continuous signal data points, and fill the normalized values ​​with the pixel values ​​at the corresponding positions in all sample images to obtain a filled image; C3. Convert the filled image into a grayscale image and stack it to form an image of size W*W*M as the metadata set; The edge feature of the point graph is obtained based on the square of the difference of the node features, the edge feature of the distribution graph is obtained based on the square of the difference of the node features, and the loss value of the dual-graph network is the weighted sum of the point loss and the distribution loss; The point loss expression is as follows: In the formula, pg represents the point graph, l represents the number of iterations of the network, and x i represents the i-th sample, y i Indicates the labeling information of the i-th sample, L crossentropy represents the cross entropy loss function; The distribution loss expression as follows: In the formula, dg represents the distribution graph, l represents the number of iterations of the network, onehot represents the single hot encoding, and y i Represents the annotation information of the i-th sample, NK represents the dimension of the distribution-level node feature, Indicates the final edge feature in the distribution graph at the lth iteration, L crossentropy represents the cross entropy loss function.

2. A small sample bearing fault diagnosis method based on dual graph network according to claim 1, characterized in that: The class prediction is calculated by feeding the edge into the function, and the calculation expression is: Where, Represents the final edge feature of the point graph, x i represents the characteristics of the i-th test sample, Represents x i The probability distribution of the fault type of the sample represented by y j Represents the annotation information of the j-th training sample, onehot represents one-hot encoding; NK represents the dimension of distribution-level node features.

3. The small sample bearing fault diagnosis method based on dual graph network according to claim 1 is characterized in that: The final edge feature calculation steps of the point graph are as follows: B1. Obtain the initial node features of the point graph based on the features of the metadata dataset; B2. Based on the initial node features of the point graph, the initial edge features of the point graph are calculated in combination with the encoding network; B3. Based on the labeled information of the training samples, the initial node features of the distribution graph are obtained, and based on the initial edge features of the point graph, the final node features of the distribution graph are iteratively calculated; B4. Based on the final node features of the distribution graph, the edge features of the distribution graph are calculated in combination with the encoding network; B5. Based on the edge features of the distribution graph, the final node features of the point graph are calculated in combination with the aggregation network, and the final node features of the point graph are calculated in combination with the encoding network; B6. Based on the final node features of the point graph, predict the class of each test sample.

4. The small sample bearing fault diagnosis method based on dual graph network according to claim 1 is characterized in that: The test samples and training samples are selected from the Case Western Reserve bearing dataset.

5. The small sample bearing fault diagnosis method based on dual graph network according to claim 4 is characterized in that: Initial node features of the distribution graph The calculation expression is as follows: Where, represents a real number, || represents the concatenation operator, δ(·) represents the Kronecker function, y i and y j represents the annotation information, and NK represents the dimension of the distribution-level node feature.

6. The small sample bearing fault diagnosis method based on dual graph network according to claim 1 is characterized in that: The ratio of the number of test samples to training samples is 7:3.

Images