A small sample bearing fault detection classification method and system based on deep learning

Abstract

This invention belongs to the field of industrial fault detection, specifically a small-sample bearing fault detection classification method and system based on deep learning, including the following steps: 1) Acquiring bearing vibration signals and sending them to a host computer for processing; 2) The host computer uses a multidimensional convolutional network combined with an attention mechanism as a feature extraction network model to extract features from the input bearing vibration signals; 3) The signal is processed by the feature extraction network model to extract feature vectors, which are then mapped to a metric space using a nonlinear method; 4) The feature extraction network model measures the similarity of the feature vectors in the metric space, and then classifies the input bearing vibration signals to complete the bearing fault detection classification. This invention not only solves the problem of bearing fault detection with small samples, but also effectively addresses the issue of similarity measurement between bearing fault signal samples. Compared with other methods, this invention has higher accuracy.

Description

Technical Field

[0001] This invention belongs to the field of industrial fault detection, specifically a small-sample bearing fault detection and classification method and system based on deep learning. Background Technology

[0002] In recent years, bearings have become critical components of industrial equipment, significantly impacting the safety of industrial physical systems. Bearing failures can lead to equipment downtime and accidents, posing substantial risks to production safety. Traditional bearing fault detection relies on methods such as vibration signal analysis, which is inefficient and dependent on expert judgment.

[0003] To achieve automatic detection and classification of bearing faults, proposing a bearing fault detection method based on deep learning is of great significance. In recent years, deep learning has achieved excellent results in the field of fault diagnosis, as it can learn the implicit features in signals.

[0004] However, obtaining large amounts of bearing fault data is practically difficult, making the few-sample problem a major challenge in bearing fault detection. Furthermore, some methods may overlook important features in bearing vibration signals, resulting in insufficient detection capabilities.

[0005] Existing methods require a large amount of data for training, lack effective methods for small samples, and generally extract features through a single dimension. Some important spatial and channel information is not given priority, and there is a lack of methods that start from multiple scales and dimensions. Summary of the Invention

[0006] The purpose of this invention is to provide a bearing fault detection method based on deep learning suitable for small sample sizes, which can improve the accuracy of bearing fault detection under small sample conditions.

[0007] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0008] A few-sample bearing fault detection and classification method based on deep learning includes the following steps:

[0009] 1) Collect bearing vibration signals and send them to the host computer for processing;

[0010] 2) The host computer uses a multidimensional convolutional network combined with an attention mechanism as the feature extraction network model to extract features from the input bearing vibration signal;

[0011] 3) The signal is processed by a feature extraction network model to extract feature vectors, and the extracted feature vectors are mapped to the metric space using a nonlinear method;

[0012] 4) The feature extraction network model measures the similarity of feature vectors in the metric space, and then classifies the input bearing vibration signal to complete the bearing fault detection classification.

[0013] The multidimensional convolution uses 1DCNN and 2DCNN networks; both 1DCNN and 2DCNN networks are based on the WDCNN network with the addition of CBAM, reducing the number of fully connected layers.

[0014] The structure of the 1DCNN network includes:

[0015] The architecture of WDCNN consists of multiple 1D convolutional layers with a stride of 1 and max pooling layers;

[0016] The convolutional layers use 1D convolutional kernels with different dilation rates, all with a kernel size of 3 and a total of 64 kernels.

[0017] The max pooling layer has a pooling window size of 2, and the feature map length is halved after each pooling operation. After multiple convolution-pooling operations, the feature map length is gradually halved from its original length until it becomes a fixed length L, but the number of channels remains unchanged at 64 channels. At this point, the feature map is regarded as a matrix of size L×64. Finally, a softmax function is added for classification.

[0018] The structure of the 2DCNN network includes:

[0019] The first convolutional kernel is 7×7 with 64 channels. The second layer is a CBAM module with a 7×7 kernel and the same number of channels. The third layer has a 2×2 pooling kernel. The fourth convolutional kernel is 5×5 with 32 features. The second layer is a CBAM module with a 7×7 kernel. The sixth layer has a 2×2 pooling kernel. Finally, a fully connected layer is added.

[0020] In step 2), the feature extraction of the input bearing vibration signal specifically involves:

[0021] A 1DCNN and 2DCNN combined with an attention mechanism model are used as the feature extraction network to extract features from the input bearing vibration signal;

[0022] Among them, the 1DCNN convolutional neural network is used to extract the time-domain features of bearing vibration signals;

[0023] A 2D CNN convolutional neural network was used to extract the time-frequency features of bearing vibration signals.

[0024] An attention mechanism model is used to focus on important features, including vibration signal characteristics at different locations of the bearing and important features of different frequency channels, in order to improve the network's feature extraction capabilities and thus improve the bearing fault detection capabilities.

[0025] Step 4) includes the following steps:

[0026] (1) To measure the similarity of samples, a non-linear mapping function is selected, namely the sigmoid activation function. This function will be used to map feature vectors to a new space; the absolute distance between the two is obtained, i.e.:

[0027]

[0028] D=∑|x i -x i+1 |

[0029] Where xi is the vibration signal of the i-th bearing, D is the absolute distance between two samples, and f(x) i () is the sigmoid activation function;

[0030] (2) The similarity between samples is obtained based on the absolute distance between them;

[0031] Select a set number of samples from each category in the training set as the support set;

[0032] Suppose we need to calculate x t To conduct the test, a set number of samples from each category are selected as the support set T:

[0033] T = {(X1, Y1)......(X i Y i )}

[0034] Then, find x t The test sample is classified using the sample most similar to T, i.e.:

[0035] M(x t (X1, X2, ..., X i ))=argmax(P(x t x m )), x m ∈T

[0036] y t =y m

[0037] Where P() represents calculating the similarity between two samples, y m It is the category corresponding to the sample with the highest similarity;

[0038] (3) Compare the sample to be tested with the sample in the support set. The category with the highest similarity is determined as the category of the sample to be tested, i.e., bearing fault detection classification.

[0039] A classification system based on a deep learning-based few-sample bearing fault detection classification method includes: a signal acquisition module, a feature extraction module, an attention module, and a classification and recognition module;

[0040] The signal acquisition module is used to acquire the vibration signal of the bearing and send the bearing vibration signal to the feature extraction module for feature extraction.

[0041] The attention module is used to generate an attention mechanism model and send it to the feature extraction module. This helps the feature extraction module focus on important features, including vibration signal features at different positions of the bearing and important features of different frequency channels, in order to improve the network's feature extraction capabilities and thus improve the bearing fault detection capabilities.

[0042] The feature extraction module is used to receive the attention mechanism model transmitted by the attention module, and extract the bearing vibration signal by combining the 1DCNN network and the 2DCNN network; and send the extracted feature vector to the classification and recognition module.

[0043] The classification and recognition module is used to map feature vectors to a metric space through a nonlinear activation function, and then obtain similarity to measure the similarity between samples in order to classify bearing fault detection.

[0044] The present invention has the following beneficial effects and advantages:

[0045] 1. This invention proposes a method for small-sample bearing fault detection based on deep learning, which can not only solve the problem of bearing fault detection under small sample conditions, but also effectively solve the problem of similarity between bearing fault signal measurement samples. Compared with other methods, this invention has higher accuracy.

[0046] 2. The present invention uses cross-entropy as the loss function for the entire network and updates the convolution kernel and parameters through backpropagation algorithm, with the aim of improving the classification effect.

[0047] 3. The attention mechanism model of the present invention is used to focus on important features including vibration signal characteristics at different positions of the bearing and important features of different frequency channels, so as to improve the network feature extraction capability and thus improve the bearing fault detection capability. Attached Figure Description

[0048] Figure 1 This is a flowchart of the bearing fault classification and detection method of the present invention;

[0049] Figure 2 This is a structural diagram of the CBAM (attention mechanism) in this invention;

[0050] Figure 3 The present invention provides a 1DCNN+CBAM (attention mechanism) network architecture for bearing fault classification and detection feature extraction.

[0051] Figure 4 The present invention provides a 2DCNN+CBAM (attention mechanism) network architecture for bearing fault classification and detection feature extraction. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0053] This invention is based on a deep learning-based few-sample bearing fault detection and classification system, comprising: a signal acquisition module, a feature extraction module, an attention module, and a classification and recognition module;

[0054] The signal acquisition module is used to acquire the vibration signal of the bearing and send the bearing vibration signal to the feature extraction module for feature extraction.

[0055] The attention module generates an attention mechanism model and sends it to the feature extraction module. This helps the feature extraction module focus on important features, including vibration signal features at different locations on the bearing and features from different frequency channels, thereby improving the network's feature extraction capabilities and ultimately enhancing bearing fault detection capabilities.

[0056] The feature extraction module receives the attention mechanism model transmitted by the attention module, and combines 1DCNN and 2DCNN networks to extract the bearing vibration signal; and sends the extracted feature vector to the classification and recognition module.

[0057] The classification and recognition module is used to map feature vectors to a metric space through a non-linear activation function, and then obtain similarity to measure the similarity between samples for bearing fault detection classification.

[0058] like Figure 1 The diagram shows a flowchart of the bearing fault classification and detection method of the present invention. The present invention provides a small-sample bearing fault detection and classification method based on deep learning, which includes the following steps:

[0059] 1) Collect bearing vibration signals and send them to the host computer for processing;

[0060] 2) The host computer uses a multidimensional convolutional network combined with an attention mechanism as the feature extraction network model to extract features from the input bearing vibration signal;

[0061] 3) The signal is processed by a feature extraction network model to extract feature vectors, and the extracted feature vectors are mapped to the metric space using a nonlinear method;

[0062] 4) The feature extraction network model measures the similarity of feature vectors in the metric space, and then classifies the input bearing vibration signal to complete the bearing fault detection classification.

[0063] like Figure 2 The diagram shows the structure of CBAM (attention mechanism) in this invention. This invention uses 1DCNN and 2DCNN networks for multidimensional convolution. Both 1DCNN and 2DCNN networks are based on the WDCNN network with the addition of CBAM, reducing the number of fully connected layers.

[0064] like Figure 3 The diagram shows the 1DCNN+CBAM network architecture used in this invention, which is an improvement upon the WDCNN architecture. The WDCNN architecture consists of multiple 1D convolutional layers with a stride of 1 and max-pooling layers. The convolutional layers use 1D convolutional kernels with different dilation rates, all with a kernel size of 3, for a total of 64 kernels.

[0065] The max-pooling layer has a pooling window size of 2, and the feature map length is halved after each pooling operation. After multiple convolution-pooling operations, the feature map length is gradually halved from its original length until it becomes a fixed length L, while the number of channels remains unchanged at 64. At this point, the feature map can be viewed as an L×64 matrix. Finally, a softmax function is added for classification. The entire network uses cross-entropy as the loss function and updates the convolution kernels and parameters through backpropagation to improve classification performance.

[0066] This invention employs a CBAM (Hybrid Attention Mechanism) approach, which involves convolutional layers 1 and 2 of WDCNN. The final structure is shown in Table 3. The CBAM structure is as follows: Figure 2 As shown;

[0067] like Figure 4 The diagram shows the 2DCNN+CBAM network architecture used in this invention. The first layer has a 7×7 kernel size and 64 channels. The second layer is a CBAM module with a 7×7 kernel and the same number of channels. The third layer has a 2×2 pooling kernel. The fourth layer has a 5×5 kernel size and 32 features. The second layer is a CBAM module with a 7×7 kernel. The sixth layer has a 2×2 pooling kernel. Finally, a fully connected layer is added.

[0068] like Figure 1 As shown, in step 2), feature extraction is performed on the input bearing vibration signal, specifically as follows:

[0069] A 1DCNN and 2DCNN combined with an attention mechanism model are used as the feature extraction network to extract features from the input bearing vibration signal;

[0070] Among them, the 1DCNN convolutional neural network is used to extract the time-domain features of the bearing vibration signal;

[0071] A 2D CNN convolutional neural network was used to extract the time-frequency features of bearing vibration signals.

[0072] An attention mechanism model is used to focus on important features, including vibration signal characteristics at different locations of the bearing and important features of different frequency channels, in order to improve the network's feature extraction capabilities and thus improve the bearing fault detection capabilities.

[0073] In step 4), the feature extraction network model measures the similarity of feature vectors in the metric space, and then classifies the input bearing vibration signal to complete the bearing fault detection classification, including the following steps:

[0074] (1) To measure the similarity of samples, a non-linear mapping function is selected, namely the sigmoid activation function. This function will be used to map feature vectors to a new space; the absolute distance between the two is obtained, i.e.:

[0075]

[0076] D=∑|x i -x i+1 |

[0077] Where xi is the vibration signal of the i-th bearing, D is the absolute distance between two samples, and f(x) i () is the sigmoid activation function;

[0078] (2) The similarity between samples is obtained based on the absolute distance between them;

[0079] Select a set number of samples from each category in the training set as the support set;

[0080] Suppose we need to calculate x t To conduct the test, a set number of samples from each category are selected as the support set T:

[0081] T = {(X1, Y1)......(X i Y i )}

[0082] Then, find x t The test sample is classified using the sample most similar to T, i.e.:

[0083] M(x t (X1, X2, ..., X i ))=argmax(P(x t x m )), x m ∈T

[0084] y t =y m

[0085] Where P() represents calculating the similarity between two samples, y m It is the category corresponding to the sample with the highest similarity;

[0086] (3) Compare the sample to be tested with the sample in the support set. The category with the highest similarity is determined as the category of the sample to be tested, i.e., bearing fault detection classification.

[0087] In summary, this invention not only solves the problem of bearing fault detection with small sample sizes, but also effectively addresses the issue of similarity between bearing fault signal measurement samples. Compared with other methods, this invention has higher accuracy.

[0088] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, extensions, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A few-sample bearing fault detection and classification method based on deep learning, characterized in that, Includes the following steps: 1) Collect bearing vibration signals and send them to the host computer for processing; 2) The host computer uses a multidimensional convolutional network combined with an attention mechanism as the feature extraction network model to extract features from the input bearing vibration signal; In step 2), the feature extraction of the input bearing vibration signal specifically involves: A 1DCNN and 2DCNN combined with an attention mechanism model are used as the feature extraction network to extract features from the input bearing vibration signal; Among them, the 1DCNN convolutional neural network is used to extract the time-domain features of bearing vibration signals; A 2D CNN convolutional neural network was used to extract the time-frequency features of bearing vibration signals. An attention mechanism model is used to focus on important features, including vibration signal characteristics at different locations of the bearing and important features of different frequency channels, in order to improve the network's feature extraction capabilities and thus improve the bearing fault detection capabilities. 3) The signal is processed by a feature extraction network model to extract feature vectors, and the extracted feature vectors are mapped to the metric space using a nonlinear method; 4) The feature extraction network model measures the similarity of feature vectors in the metric space, and then classifies the input bearing vibration signal to complete the bearing fault detection classification.

2. The method for small-sample bearing fault detection and classification based on deep learning according to claim 1, characterized in that, The multidimensional convolution uses 1DCNN and 2DCNN networks; both 1DCNN and 2DCNN networks are based on the WDCNN network with the addition of CBAM, reducing the number of fully connected layers.

3. The method for small-sample bearing fault detection and classification based on deep learning according to claim 2, characterized in that, The structure of the 1DCNN network includes: The architecture of WDCNN consists of multiple 1D convolutional layers with a stride of 1 and max pooling layers; The convolutional layers use 1D convolutional kernels with different dilation rates, all with a kernel size of 3 and a total of 64 kernels. The max pooling layer has a pooling window size of 2, and the feature map length is halved after each pooling operation. After multiple convolution-pooling operations, the feature map length is gradually halved from its original length until it becomes a fixed length L, but the number of channels remains unchanged at 64 channels. At this point, the feature map is regarded as a matrix of size L×64. Finally, a softmax function is added for classification.

4. The method for small-sample bearing fault detection and classification based on deep learning according to claim 2, characterized in that, The structure of the 2DCNN network includes: The first convolutional kernel is 7×7 with 64 channels. The second layer is a CBAM module with a 7×7 kernel and the same number of channels. The third layer has a 2×2 pooling kernel. The fourth convolutional kernel is 5×5 with 32 features. The second layer is a CBAM module with a 7×7 kernel. The sixth layer has a 2×2 pooling kernel. Finally, a fully connected layer is added.

5. The method for small-sample bearing fault detection and classification based on deep learning according to claim 1, characterized in that, Step 4 includes the following steps: (1) To measure the similarity of samples, a non-linear mapping function is selected, namely the sigmoid activation function. This function will be used to map the feature vectors to a new space; the absolute distance between the two is obtained, i.e.: ； ； in, Let i be the vibration signal of the i-th bearing. The absolute distance between two samples. It is the sigmoid activation function; (2) The similarity between samples is obtained based on the absolute distance between them; Select a set number of samples from each category in the training set as the support set; Assuming it is necessary to To conduct the test, a set number of samples from each category are selected as the support set T: ； Then, find The test sample is classified using the sample most similar to T, i.e.: ； ； Where P() represents calculating the similarity between two samples. It is the category corresponding to the sample with the highest similarity; (3) Compare the sample to be tested with the sample in the support set. The category with the highest similarity is determined as the category of the sample to be tested, i.e., bearing fault detection classification.

6. The classification system of the deep learning-based few-sample bearing fault detection classification method according to claim 1, characterized in that, include: Signal acquisition module, feature extraction module, attention module, classification and recognition module; The signal acquisition module is used to acquire the vibration signal of the bearing and send the bearing vibration signal to the feature extraction module for feature extraction. The attention module is used to generate an attention mechanism model and send it to the feature extraction module. This helps the feature extraction module focus on important features, including vibration signal features at different positions of the bearing and important features of different frequency channels, in order to improve the network's feature extraction capabilities and thus improve the bearing fault detection capabilities. The feature extraction module is used to receive the attention mechanism model transmitted by the attention module, and extract the bearing vibration signal by combining the 1DCNN network and the 2DCNN network; and send the extracted feature vector to the classification and recognition module. The classification and recognition module is used to map feature vectors to a metric space through a nonlinear activation function, and then obtain similarity to measure the similarity between samples in order to classify bearing fault detection.

Images