Bearing fault diagnosis method based on WSDAE-WDCNN-SENet

Abstract

The application discloses a bearing fault diagnosis method based on WSDAE-WDCNN-SENet, and relates to the technical field of bearing fault diagnosis.The diagnosis method comprises the following steps: S1, sampling the key vibration signal of a bearing by using a sliding window, and dividing the key vibration signal into a training set, a verification set and a test set; S2, learning the features of the vibration signal by using a WDCNN model and a WSDAE model respectively, and performing feature information fusion by using an SENet; and S3, inputting the feature information obtained after the fusion into Softmax for classification diagnosis, and generating a fault diagnosis result report.The bearing fault diagnosis method based on WSDAE-WDCNN-SENet is designed to propose a weight transition module (WSC), to add an original signal path, and to integrate a sparse network structure, a BN normalization and a Dropout layer into the module, so that the path can capture feature weight information, the important feature weight of the original signal can be further improved and amplified, and other deep learning networks can be facilitated to learn features.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method based on WSDAE-WDCNN-SENet. Background Technology

[0002] Mechanical equipment contains many small but crucial components, and bearings, as one of the key components of rotating machinery, directly affect the overall operation of the equipment. Therefore, how to quickly and accurately extract fault characteristics in rotating machinery and effectively diagnose faults in mechanical equipment, especially bearing faults, is an issue that cannot be ignored. How to accurately and efficiently diagnose faults under noise interference remains a major research focus.

[0003] Artificial intelligence-based fault diagnosis methods have been widely applied to bearing fault diagnosis and have achieved certain results. Currently, most bearing fault diagnosis is based on the monitoring and analysis of various state parameters to determine its operating status, fault location, and wear degree. Generally, bearing fault diagnosis can be accomplished through vibration signal acquisition, feature extraction, and classification. Wavelet transform can be used to obtain more ideal signal decomposition results, thereby achieving higher accuracy in fault diagnosis. The ensemble empirical mode decomposition method has certain advantages over the original method in suppressing mode confusion and fault diagnosis.

[0004] However, since the vibration signals collected by rotating machinery are easily interfered with by other background noise or other noise, many technical methods can only process bearing vibration signals under noise-free conditions, making it difficult to effectively distinguish valuable bearing vibration signals under different health conditions. Furthermore, many current dual-channel and multi-channel fault diagnosis network methods directly connect the outputs of each channel, ignoring the fact that the output characteristics of each channel are of different importance under different environments. Therefore, we propose a bearing fault diagnosis method based on WSDAE-WDCNN-SENet. Summary of the Invention

[0005] The purpose of this invention is to provide a bearing fault diagnosis method based on WSDAE-WDCNN-SENet, which overcomes the shortcomings of the prior art and solves the above-mentioned problems.

[0006] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0007] A bearing fault diagnosis method based on WSDAE-WDCNN-SENet, the diagnosis method includes:

[0008] S1: Slide window sampling is performed on the key vibration signals of the bearing, and the data is divided into training set, validation set and test set;

[0009] S2: Use the WDCNN model and the WSDAE model to learn the characteristics of the vibration signal respectively, and use SENet to fuse the feature information;

[0010] S3: Input the fused feature information into Softmax for classification and diagnosis, and generate a fault diagnosis result report;

[0011] The WSDAE model integrates WSC into the SDAE network structure layer and consists of an encoder, a WSC module, a hidden layer, and a decoder.

[0012] The WSC consists of two paths: a weight learning path and a raw feature path. Input data enters the raw feature path, other neural network paths, and the weight learning path, respectively. In the raw feature path, the data x is directly connected from the input to the output. In the other neural network paths, the data x is processed by the neural network to generate a feature vector F. n (x), in the weight learning path, the data x flows sequentially through a sparse fully connected (SFC) layer, a batch normalization (BN) layer, a ReLU activation function, a fully connected (FC) layer, a BN layer, and a Sigmoid activation function. The ultimate goal of the weight learning path is to generate a weight matrix where each element is between 0 and 1.

[0013] W(x) = S(b2 + R(b1 + w) s x)),

[0014] Where S(·) and R(·) represent the activation functions Sigmoid and ReLU, respectively, b1 and b2 represent the network bias terms, and w s represents a sparse matrix that controls the sparsity of the network, and x represents the input value.

[0015] Preferably, the window length in the sliding window sampling is 2048, and the step size is 28.

[0016] Preferably, the method for learning the characteristics of the vibration signal in S2 is as follows:

[0017] S21: Establish the WDCNN and WSDAE models;

[0018] S22: Connect the outputs of the WSDAE and WDCNN network models to their respective fully connected layers, reshape them into multiple feature channels, and then input them into the SENet module;

[0019] S23: After the output of the pooling layer in the SENet module, the importance of each feature channel is obtained through a series of layer processing;

[0020] S24: Multiply the importance by the feature channel to obtain the final feature information.

[0021] Preferably, the pooling layer in the SENet module uses average pooling to obtain the global feature description of each channel and applies max pooling to extract local channel features. At the same time, a Dropout layer is added to the SENet module.

[0022] Preferably, the WDCNN model in S2 has 5 convolutional and pooling layers. The kernel size of the first convolutional layer is 64×1, and the kernel size of the remaining convolutional layers is 3×1.

[0023] Preferably, the encoder has the following structure in sequence: a sparse fully connected (SFC) layer, a batch normalized (BN) layer, a ReLU activation function, a fully connected (FC) layer, a BN layer, and a ReLU activation function; the hidden layer has the following structure: an FC layer, a BN layer, and a ReLU activation function; and the decoder has the following structure: an FC layer, a BN layer, a ReLU activation function, an FC layer, a BN layer, and a Tanh activation function.

[0024] Preferably, the weight learning path of the WSC is connected to the input end of the encoder and the hidden layer to generate the feature matrix of the hidden layer, and a WSC module is connected to the output end of the hidden layer to further learn the signal features.

[0025] Preferably, the loss function of the WSDAE model combines reconstruction loss and sparsity penalty term;

[0026] The reconstruction loss employs mean squared error (MSE), which is defined as:

[0027]

[0028] Where n is the number of samples, x i and These are the true value and the model prediction value of the i-th sample, respectively;

[0029] The sparsity penalty term is expressed as:

[0030]

[0031] Where λ is the regularization parameter;

[0032] The total loss function is defined as:

[0033] L WSDAE =L mse +L sparse ,

[0034] When L mse and L sparse When it is at its minimum, that is, λ and When equilibrium is reached, L WSDAEIt can be minimized to achieve the optimal model.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] This invention proposes a weighted transition module (WSC) that incorporates the original signal path and integrates it with a sparse network structure, combining BN normalization and Dropout layers. This allows the path to capture feature weight information, further enhancing and amplifying the important feature weights of the original signal, thus facilitating feature learning by other deep learning networks.

[0037] Based on the WSC attention module, this invention proposes a weighted denoising autoencoder network (WSDAE). By inserting feature weight paths from the WSC module into the traditional sparse denoising autoencoder network structure, the efficiency of feature extraction in the autoencoder model is improved, noise information contained in the input signal is eliminated, and the generalization performance of the network model is enhanced.

[0038] This invention proposes a dual-channel attention mechanism feature fusion network (WSDAE-WDCNN-SENet), which utilizes two network channels, WDCNN and WSDAE, to simultaneously extract feature information from the input signal and store it in multiple channels. Average pooling and max pooling layers are integrated into the SENet module, and the importance of feature information from multiple channels is captured, filtered, and fused. Related noise experiments demonstrate the robustness and superiority of the WSDAE-WDCNN-SENet fault diagnosis model. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall process of the bearing fault diagnosis method based on WSDAE-WDCNN-SENet;

[0040] Figure 2 This is a structural diagram of the dual-channel attention mechanism feature fusion network model (WSDAE-WDCNN-SENet).

[0041] Figure 3 This is a structural diagram of the Weighted Switching Module (WSC).

[0042] Figure 4 The structure diagram of the WSDAE heavy sparse noise reduction autoencoder;

[0043] Figure 5 The flowchart shows the feature fusion network model of the dual-channel attention mechanism (WSDAE-WDCNN-SENet). Detailed Implementation

[0044] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0045] Please refer to Figure 1 As shown in Figure 5, this embodiment presents a bearing fault diagnosis method based on WSDAE-WDCNN-SENet. The diagnosis method includes:

[0046] S1: Slide window sampling is performed on the key vibration signals of the bearing, and the data is divided into training set, validation set and test set;

[0047] In the sliding window sampling, the window length is 2048 and the step size is 28.

[0048] S2: Use the WDCNN model and the WSDAE model to learn the characteristics of the vibration signal respectively, and use SENet to fuse the feature information;

[0049] The method for learning the characteristics of vibration signals in S2 is as follows:

[0050] S21: Establish the WDCNN and WSDAE models;

[0051] S22: Connect the outputs of the WSDAE and WDCNN network models to their respective fully connected layers, reshape them into multiple feature channels, and then input them into the SENet module;

[0052] S23: After the output of the pooling layer in the SENet module, the importance of each feature channel is obtained through a series of layer processing;

[0053] S24: Multiply the importance by the feature channel to obtain the final feature information;

[0054] The WDCNN model has 5 convolutional and pooling layers. The kernel size of the first convolutional layer is 64×1, and the kernel size of the remaining convolutional layers is 3×1.

[0055] The pooling layers in the SENet module use average pooling to obtain the global feature description of each channel and max pooling to extract local channel features. Dropout layers are also added to the SENet module.

[0056] S3: Input the fused feature information into Softmax for classification and diagnosis, and generate a fault diagnosis result report;

[0057] The WSDAE model integrates WSC into the SDAE network structure layer and consists of an encoder, WSC module, hidden layer and decoder;

[0058] The encoder's structure consists of: a sparse fully connected (SFC) layer, a batch normalization (BN) layer, a ReLU activation function, a fully connected (FC) layer, a BN layer, and a ReLU activation function; the hidden layer's structure is: an FC layer, a BN layer, and a ReLU activation function; the decoder's structure is: an FC layer, a BN layer, a ReLU activation function, an FC layer, a BN layer, and a Tanh activation function.

[0059] WSC consists of two paths: a weight learning path and a raw feature path. Input data enters the raw feature path, other neural network paths, and the weight learning path, respectively. In the raw feature path, data x is directly connected from the input to the output. In other neural network paths, data x is processed by the neural network to generate a feature vector F. n (x), the data x in the weight learning path flows sequentially through a sparse fully connected (SFC) layer, a batch normalization (BN) layer, a ReLU activation function, a fully connected (FC) layer, a BN layer, and a Sigmoid activation function. The ultimate goal of the weight learning path is to generate a weight matrix where each element is between 0 and 1.

[0060] W(x) = S(b2 + R(b1 + w) s x)),

[0061] Where S(·) and R(·) represent the activation functions Sigmoid and ReLU, respectively, b1 and b2 represent the network bias terms, and w s represents a sparse matrix that controls the sparsity of the network, and x represents the input value;

[0062] Through the above process, a weight matrix W(x) is generated. Each element in this weight matrix represents the importance of the corresponding feature, thereby strengthening the network's attention to important feature information and improving the network model's performance.

[0063] The weight learning path of WSC is connected to the input of the encoder and the hidden layer to generate the feature matrix of the hidden layer. A WSC module is then connected to the output of the hidden layer to further learn the signal features.

[0064] The loss function of the WSDAE model combines reconstruction loss and sparsity penalty term;

[0065] The reconstruction loss uses mean squared error (MSE), which is defined as:

[0066]

[0067] Where n is the number of samples, x i and These are the true value and the model prediction value of the i-th sample, respectively;

[0068] The sparsity penalty term is represented as:

[0069]

[0070] Where λ is the regularization parameter;

[0071] The total loss function is defined as:

[0072] L WSDAE =L mse +L sparse ,

[0073] When L mse and L sparse When it is at its minimum, that is, λ and When equilibrium is reached, L WSDAE It can be minimized to achieve the optimal model;

[0074] The WSDAE loss function combines reconstruction loss and sparsity penalty term. This design enables the model to capture key features of the data and perform accurate reconstruction, while also encouraging the model to generate sparsity, thereby enhancing the model's generalization ability.

[0075] This embodiment proposes a Weighted Transition Module (WSC), which incorporates the original signal path and integrates it with a sparse network structure, Batch Normalization (BN), and Dropout layers. This allows the path to capture feature weight information, further amplifying the important feature weights of the original signal, facilitating feature learning by other deep learning networks. Based on the WSC attention module, a Weighted Denoising Autoencoder Network (WSDAE) is proposed. By inserting the feature weight path from the WSC module into the traditional sparse denoising autoencoder network structure, the efficiency of feature extraction in the autoencoder model is improved, noise information in the input signal is eliminated, and the generalization performance of the network model is enhanced.

[0076] This embodiment proposes a dual-channel attention mechanism feature fusion network (WSDAE-WDCNN-SENet), which uses two network channels, WDCNN and WSDAE, to simultaneously extract feature information from the input signal and store it in multiple channels. Average pooling layers and max pooling layers are integrated into the SENet module, and the importance of feature information from multiple channels is captured, filtered, and fused. Related noise experiments are conducted to demonstrate the robustness and superiority of the WSDAE-WDCNN-SENet fault diagnosis model.

[0077] This embodiment selects some advanced and representative fault diagnosis models, including ResNet, WDCNN, and RNN-WDCNN. These models will be compared with the proposed WSDAE-WDCNN-SENet to verify the superiority of the proposed method. To ensure the fairness of the comparison, all algorithms adopt early stopping to control the number of iterations, and all use Adam as the optimizer. Each algorithm runs 10 times and the average accuracy is taken. Two datasets are used to compare the methods.

[0078] Dataset A includes one normal state (NC) and three fault states: rolling fault (BF), inner race fault (IR), and outer race fault (OR). Each fault state is further subdivided into three different categories based on the fault diameter. Including the normal state, there are a total of 10 different state categories. For each bearing state, 400 samples were collected. The ratio of training set, validation set, and test set is [0.7:0.2:0.1].

[0079] The comparison results are shown in the table below:

[0080]

[0081] As shown in the table above, compared with ResNet and WDCNN, the RNN-WDCNN model, which connects the LSTM in RNN to WDCNN, has improved accuracy. This indicates that dual-channel technology plays a role in improving fault diagnosis accuracy. However, this method only considers the simple superposition of features extracted by the two models for diagnosis, ignoring the fact that the effectiveness of features obtained by different models varies in different environments. This embodiment introduces the SENet module on the basis of simple dual-channel parallel connection, and proposes the WSDAE-WDCNN-SENet model. Compared with RNN-WDCNN, it further improves the accuracy of fault diagnosis. Moreover, in noisy environments, the proposed model performs better than other models and has a more efficient structure, indicating the importance of feature weight learning for fault diagnosis. It also proves that the WSDAE-WDCNN-SENet model proposed in this embodiment has better robustness and versatility.

[0082] Dataset B contains one normal state (NC) and four fault states: rolling fault (BF), inner race fault (IR), outer race fault (OR), and combined fault (CF). In the experiment, the motor speed was kept constant at 20Hz, the sensor was positioned horizontally at the input end of the planetary gearbox, and the sampling frequency was 12800Hz. A total of 400 samples were collected for each bearing state. The ratio of training set, validation set, and test set was [0.7:0.2:0.1].

[0083] The comparison results are shown in the table below:

[0084]

[0085] As shown in the table above, the WSDAE-WDCNN-SENet model outperforms other models under different conditions, demonstrating the model's generality and effectiveness. Similar to the experimental results of dataset A, the experimental results of both dual-channel models are superior to the traditional single-channel model, which further proves the feasibility of dual-channel models in fault diagnosis problems.

[0086] However, unlike the results of dataset A, the WDCNN model did not perform stably on this dataset. Although ResNet performed relatively stably on this dataset, its network structure was complex and its running time was long. Furthermore, the average accuracy of both ResNet and WDCNN models was less than 95%. This indicates that traditional single-structure network architectures are difficult to adapt to fault diagnosis problems of different environments and varying degrees of complexity.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A bearing fault diagnosis method based on WSDAE-WDCNN-SENet, characterized in that, The diagnostic method includes: S1: Slide window sampling is performed on the key vibration signals of the bearing, and the data is divided into training set, validation set and test set; S2: Use the WDCNN model and the WSDAE model to learn the characteristics of the vibration signal respectively, and use SENet to fuse the feature information; S3: Input the fused feature information into Softmax for classification and diagnosis, and generate a fault diagnosis result report; The WSDAE model integrates WSC into the SDAE network structure layer and consists of an encoder, a WSC module, a hidden layer, and a decoder. The WSC consists of two paths: a weight learning path and a raw feature path. Input data enters the raw feature path, other neural network paths, and the weight learning path, respectively. In the raw feature path, the data x is directly connected from the input to the output. In the other neural network paths, the data x is processed by the neural network to generate a feature vector F. n (x), in the weight learning path, the data x flows sequentially through a sparse fully connected (SFC) layer, a batch normalization (BN) layer, a ReLU activation function, a fully connected (FC) layer, a BN layer, and a Sigmoid activation function. The ultimate goal of the weight learning path is to generate a weight matrix where each element is between 0 and 1. W(x)=S(b2+R(b1+w s x)), Where S(·) and R(·) represent the activation functions Sigmoid and ReLU, respectively, b1 and b2 represent the network bias terms, and w s represents a sparse matrix that controls the sparsity of the network, and x represents the input value; The weight learning path of the WSC is connected to the input of the encoder and the hidden layer to generate the feature matrix of the hidden layer, and a WSC module is connected to the output of the hidden layer to further learn the signal features.

2. The bearing fault diagnosis method based on WSDAE-WDCNN-SENet according to claim 1, characterized in that: The window length in the sliding window sampling is 2048, and the step size is 28.

3. The bearing fault diagnosis method based on WSDAE-WDCNN-SENet according to claim 1, characterized in that, The method for learning the characteristics of the vibration signal in S2 is as follows: S21: Establish the WDCNN and WSDAE models; S22: Connect the outputs of the WSDAE and WDCNN network models to their respective fully connected layers, reshape them into multiple feature channels, and then input them into the SENet module; S23: After the output of the pooling layer in the SENet module, the importance of each feature channel is obtained through a series of layer processing; S24: Multiply the importance by the feature channel to obtain the final feature information.

4. The bearing fault diagnosis method based on WSDAE-WDCNN-SENet according to claim 3, characterized in that: The pooling layer in the SENet module uses average pooling to obtain the global feature description of each channel and applies max pooling to extract local channel features. At the same time, a Dropout layer is added to the SENet module.

5. The bearing fault diagnosis method based on WSDAE-WDCNN-SENet according to claim 1, characterized in that: The WDCNN model in S2 has 5 convolutional layers and 5 pooling layers. The kernel size of the first convolutional layer is 64×1, and the kernel size of the remaining convolutional layers is 3×1.

6. The bearing fault diagnosis method based on WSDAE-WDCNN-SENet according to claim 1, characterized in that: The encoder has the following structure in sequence: Sparse Fully Connected (SFC) layer, Batch Normalized (BN) layer, ReLU activation function, Fully Connected (FC) layer, BN layer, ReLU activation function; the hidden layer has the following structure: FC layer, BN layer, ReLU activation function; the decoder has the following structure: FC layer, BN layer, ReLU activation function, FC layer, BN layer, Tanh activation function.

7. The bearing fault diagnosis method based on WSDAE-WDCNN-SENet according to claim 1, characterized in that: The loss function of the WSDAE model combines reconstruction loss and sparsity penalty term; The reconstruction loss employs mean squared error (MSE), which is defined as: Where n is the number of samples, x i and These are the true value and the model prediction value of the i-th sample, respectively; The sparsity penalty term is expressed as: Where λ is the regularization parameter; The total loss function is defined as: L WSDAE =L mse +L sparse , When L mse and L sparse When it is at its minimum, that is, λ and When equilibrium is reached, L WSDAE It can be minimized to achieve the optimal model.

Images