Bearing fault diagnosis method based on dual-channel dense connection network fusion model
By using a dual-channel densely connected network fusion model (DC-DenseNet-GRU), the spatial and temporal features of bearing faults are extracted simultaneously, solving the accuracy and stability problems of bearing fault diagnosis under strong noise environment and achieving highly reliable fault identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-07
AI Technical Summary
Existing bearing fault diagnosis methods struggle to accurately identify complex faults in high-noise environments, and traditional models lack robustness and generalization ability under noise interference, resulting in diagnostic accuracy and stability that fail to meet practical requirements.
A dual-channel densely connected network fusion model (DC-DenseNet-GRU) is adopted. By setting DenseNet channels and GRU channels in parallel, deep spatial features and temporal dependency features of bearing faults are extracted respectively. Feature fusion is performed in the convergence layer, and finally the fault state classification result is obtained through a classifier.
It significantly improves diagnostic accuracy and stability in high-noise environments, can accurately identify complex fault states, forms an inherent redundancy and complementarity mechanism, and improves the robustness and reliability of the diagnostic system.
Smart Images
Figure CN121808338A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of mechanical fault diagnosis technology, and in particular to a bearing fault diagnosis method based on a dual-channel dense connection network fusion model. Background Technology
[0002] As a crucial component of rotating machinery, rolling bearings directly impact the safety and reliability of the entire equipment system. Statistics show that 45% to 55% of failures in rotating machinery originate from bearing failure. Therefore, accurate and timely diagnosis of bearing faults is essential for predictive maintenance and preventing catastrophic accidents and economic losses. In actual industrial environments, the acquisition of bearing vibration signals is highly susceptible to interference from background noise. Especially under harsh operating conditions, strong noise can drown out subtle early fault characteristics, leading to a sharp drop in the signal-to-noise ratio, posing a significant challenge to the effective extraction and identification of fault features. Furthermore, bearings may exhibit complex fault modes where multiple components, such as the inner ring, outer ring, and rolling elements, fail simultaneously. The vibration signal characteristics of these components are coupled, further increasing the difficulty of accurate diagnosis.
[0003] Currently, intelligent diagnostic methods based on deep learning technology have become a research focus in this field. Among these, Convolutional Neural Networks (CNNs) are widely used due to their excellent spatial feature extraction capabilities. Specifically, this can be achieved by converting one-dimensional signals into two-dimensional time-frequency maps and inputting them into a CNN, or by building multi-scale, multi-channel CNN models to fuse different features. However, bearing vibration signals are inherently nonlinear and non-stationary time-series data, and traditional CNN models have inherent limitations in capturing their long-term time dependencies. To overcome this problem, researchers have attempted to introduce Recurrent Neural Networks (RNNs), especially Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRUs), to model the temporal dependencies of signals. For example, existing CNN-GRU or CNN-LSTM hybrid models strive to simultaneously leverage the advantages of CNNs in spatial feature extraction and the capabilities of RNNs in time-series modeling. Although these methods have made some progress, they typically employ a serial structure, with features being passed sequentially in a single channel, which can easily lead to information loss or error accumulation. In addition, networks such as DenseNet, which achieve feature reuse through dense connections, have also been used for fault diagnosis to enhance feature propagation. However, existing single models or simple serial hybrid models are still insufficient under strong noise interference, and their robustness and generalization ability need to be further enhanced. Specifically, they are difficult to simultaneously take into account the deep and effective fusion of spatial local impact characteristics and temporal dynamic evolution trend information in extreme noise environments, and lack a complementary mechanism when a certain feature extraction channel fails due to noise. As a result, the diagnostic accuracy and stability in complex industrial scenarios cannot meet the needs of practical applications.
[0004] Therefore, there is an urgent need for a new intelligent diagnostic model that can deeply integrate spatiotemporal features and has strong noise resistance to solve the problem of accurate diagnosis of rolling bearings, especially complex faults, in high-noise environments. Summary of the Invention
[0005] To address the aforementioned shortcomings in existing technologies, this application provides a bearing fault diagnosis method based on a dual-channel densely connected network fusion model, which solves the problems of insufficient bearing fault feature extraction and low diagnostic accuracy under strong noise environments.
[0006] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows: This application provides a bearing fault diagnosis method based on a dual-channel densely connected network fusion model, including: S1: Acquire the original vibration signal of the bearing; S2: Construct a dual-channel DC-DenseNet-GRU model and train it using the original vibration signal. The training process includes: A1: The original vibration signal is processed using the DenseNet channel in the dual-channel DC-DenseNet-GRU model to extract the deep spatial features of bearing faults and output the first feature vector. A2: The original vibration signal is segmented using the GRU channel in the dual-channel DC-DenseNet-GRU model, and the segmented signal sequence is processed using the gated cyclic unit GRU network to extract the time-dependent features of bearing faults and output the second feature vector. A3: The first and second feature vectors are fused at the convergence layer to obtain a feature vector that fuses spatiotemporal features; A4: Input the feature vector that integrates spatiotemporal features into the classifier to obtain the bearing fault state classification result; S3: Based on the trained dual-channel DC-DenseNet-GRU model, the test data is diagnosed to obtain the bearing fault state classification results.
[0007] Furthermore, the dual-channel DC-DenseNet-GRU model includes: parallel DenseNet channels and GRU channels, a convergence layer, and a classifier; The DenseNet channel includes an initial convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, two densely connected blocks, a transition layer, and a global average pooling layer. Each densely connected block contains multiple densely connected layers. The GRU channel includes a two-layer gated recurrent unit (GRU) network and a fully connected layer, wherein the hidden state dimension of the first layer GRU is 32 and the hidden state dimension of the second layer GRU is 1. The classifier consists of a fully connected layer, a ReLU activation function, and a regularized Dropout.
[0008] Furthermore, the calculation formula for the gated recurrent unit (GRU) network is as follows:
[0009]
[0010]
[0011]
[0012] in, To update the door, To reset the door, In the candidate hidden state, For the gated recurrent unit (GRU) network The final hidden state of time. For activation function, , and These are the weight matrices corresponding to the update gate, reset gate, and candidate hidden states, respectively. , and These are the bias vectors for the update gate, reset gate, and candidate hidden state, respectively. Indicates the previous moment Hidden state With the current input vector To splice, This represents the concatenated vector of the previous hidden state and the current input under the control of the reset gate.
[0013] Furthermore, the densely connected block includes: The first densely connected block The input to the densely connected layer is from the first The first to the second The outputs of densely connected layers are concatenated along the channel dimension, where the first... The specific expression for the output of the densely connected layer is:
[0014] In the formula, Indicates the first The output of the densely connected layer, Represents a nonlinear transformation function. Indicates the first layer to the first The feature maps of densely connected layers are stitched together along the channel dimension.
[0015] Furthermore, the process of using the DenseNet channel in the dual-channel DC-DenseNet-GRU model to process the original vibration signal, extract the deep spatial features of the bearing fault, and output the first feature vector includes: B1: The original vibration signal is sequentially passed through an initial convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer for preliminary feature transformation; B2: Input the pre-transformed features into densely connected blocks and transition layers for deep feature extraction and reuse; B3: Perform global average pooling on the extracted deep spatial features to compress and obtain the first feature vector.
[0016] Furthermore, the original vibration signal is processed and segmented using the GRU channel in the dual-channel DC-DenseNet-GRU model, and the segmented signal sequence is processed using a gated recurrent unit (GRU) network to extract the time-dependent features of bearing faults, outputting a second feature vector, including: C1: Divide the original vibration signal into multiple signal segments of fixed length; C2: Each signal segment is sequentially input into a two-layer gated recurrent unit (GRU) network to extract the temporal features within the signal segment; C3: Combines the temporal features of all signal segments and outputs a second feature vector through a fully connected layer.
[0017] Furthermore, the step of inputting the feature vector fused with spatiotemporal features into the classifier to obtain the bearing fault state classification result includes: The fully connected layer performs dimensionality reduction on the feature vectors that fuse spatiotemporal features and filters out key fault features; it then introduces a nonlinear transformation on the key fault features using the ReLU function; the key fault features after the nonlinear transformation are then subjected to regularized Dropout processing; and the output layer maps the key fault features after regularized Dropout processing to fault state classification results.
[0018] The beneficial effects of this application are: This application presents a bearing fault diagnosis method based on a dual-channel densely connected network fusion model. Utilizing the dual-channel DC-DenseNet-GRU model, it can simultaneously and fully extract the spatial local features and temporally dynamic dependent features contained in vibration signals. Through feature fusion, it achieves deep complementarity of spatiotemporal information, effectively addressing noise interference and accurately identifying the complex fault states of rolling bearings. This significantly improves the stability and robustness of the diagnostic system under strong noise pollution. Therefore, this method maintains extremely high diagnostic accuracy under both strong noise and variable operating conditions, providing an effective solution for achieving highly reliable intelligent bearing fault diagnosis in industrial settings. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.
[0020] Figure 1 This is a flowchart illustrating a bearing fault diagnosis method based on a dual-channel dense connection network fusion model, as provided in an embodiment of this application.
[0021] Figure 2 This is a schematic diagram of the structure of a dual-channel DC-DenseNet-GRU model provided in an embodiment of this application.
[0022] Figure 3 This is a schematic diagram of the structure of a DenseNet channel provided in an embodiment of this application.
[0023] Figure 4 This is a schematic diagram of the structure of a GRU channel provided in an embodiment of this application.
[0024] Figure 5 A comparison chart of the test accuracy of different models under different loads provided in the embodiments of this application.
[0025] Figure 6 The confusion matrix of different models provided in the embodiments of this application under 0HP load.
[0026] Figure 7 The confusion matrix of different models provided in the embodiments of this application under 1HP load. Detailed Implementation
[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.
[0028] Example 1: Taking the Case Western Reserve University (CWRU) bearing dataset as an example, this paper provides a detailed description of a bearing fault diagnosis method based on a dual-channel densely connected network fusion model provided in this application. This method can be found in [reference needed]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating a bearing fault diagnosis method based on a dual-channel dense connection network fusion model provided in this application, including: S1: Acquire and preprocess the one-dimensional raw vibration signal of the bearing.
[0029] In one embodiment of this application, vibration signals collected by the drive-end accelerometer are obtained from the CWRU data center at a sampling frequency of 48kHz. The data includes 10 states (1 normal and 9 faults), and the fault types include inner ring, outer ring, and rolling element faults. Each fault has three damage diameters (0.007, 0.014, and 0.021 inches). Data were collected under four loads: 0, 1, 2, and 3 horsepower (HP).
[0030] Overlap sampling is performed on the original continuous signal to construct a sample set. Each sample has a length of 1024 data points, with an overlap offset of 28 points. The dataset is then randomly divided into a training set and a test set in a 7:3 ratio.
[0031] S2: Construct a dual-channel DC-DenseNet-GRU model and train the dual-channel DC-DenseNet-GRU model using a one-dimensional original vibration signal.
[0032] In one embodiment of this application, a construction is performed as follows: Figure 2 The dual-channel DC-DenseNet-GRU model shown includes: parallel DenseNet and GRU channels, a pooling layer, and a classifier. The structures of the DenseNet and GRU channels can be found in [reference needed]. Figure 3 and Figure 4The DenseNet channel consists of an initial convolutional layer (kernel size 7, stride 2, padding 3, output channels 64), a batch normalization layer, a ReLU activation function layer, a max pooling layer (kernel size 3, stride 2, padding 1), two dense blocks, a transition layer, and global average pooling. Each dense block contains multiple dense blocks, with a growth rate of 32. The specific expression for the output of the densely connected layer is:
[0033] In the formula, Indicates the first The output of the densely connected layer, Represents a nonlinear transformation function. Indicates the first layer to the first The feature maps of densely connected layers are stitched together along the channel dimension.
[0034] The GRU channel consists of two layers of gated recurrent unit (GRU) networks and a fully connected layer. The hidden state dimension of the first GRU layer is 32, and the hidden state dimension of the second GRU layer is 1. The calculation formula for this gated recurrent unit (GRU) network is as follows:
[0035]
[0036]
[0037]
[0038] in, To update the door, To reset the door, In the candidate hidden state, For the gated recurrent unit (GRU) network The final hidden state of time. For activation function, , and These are the weight matrices corresponding to the update gate, reset gate, and candidate hidden states, respectively. , and These are the bias vectors for the update gate, reset gate, and candidate hidden state, respectively. Indicates the previous moment Hidden state With the current input vector To splice, This represents the concatenated vector of the previous hidden state and the current input under the control of the reset gate.
[0039] Furthermore, a dual-channel DC-DenseNet-GRU model is trained using the one-dimensional original vibration signal, including: A1: The DenseNet channel in the dual-channel DC-DenseNet-GRU model is used to process the one-dimensional original vibration signal, extract the deep spatial features of bearing faults, and output the first feature vector.
[0040] In one embodiment of this application, the DenseNet channels first pass through an initial convolutional layer (kernel size of 7, stride of 2, padding of 3, and output channel number of 64), and then sequentially pass through a batch normalization layer, a ReLU activation function, and a max pooling layer for preliminary feature transformation. The features after preliminary transformation are input into at least one densely connected block for deep feature extraction and reuse. Each densely connected block contains multiple densely connected layers. Finally, global average pooling is used to compress the spatial dimension of the extracted deep spatial features to 1, outputting a 64-dimensional first feature vector.
[0041] The processor, through the DenseNet network and its dense connection structure, enables the reuse and fusion of multi-level features, effectively capturing spatial patterns such as local impacts and peaks in vibration signals, and extracting robust features even in noisy environments.
[0042] A2: The one-dimensional original vibration signal is segmented using the GRU channel in the dual-channel DC-DenseNet-GRU model, and the segmented signal sequence is processed using the gated cyclic unit GRU network to extract the time-dependent features of bearing faults and output the second feature vector.
[0043] In one embodiment of this application, the GRU channel first divides the input one-dimensional signal of length 1024 into 8 segments, each segment of length 128; each segment of the signal is sequentially passed through two layers of GRU network to extract temporal features, wherein the hidden state dimension of the first layer of GRU is 32 and the hidden state dimension of the second layer of GRU is 1; after merging the output results of each segment, the data is sent to a fully connected layer, and finally a 64-dimensional second feature vector is produced.
[0044] The processor processes the segmented sequence signal through the GRU network, and uses its update gate and reset gate mechanism to effectively model the long-range time dependence of the vibration signal, overcome the gradient problem of traditional recurrent neural networks, and can accurately capture the dynamic evolution trend of fault characteristics.
[0045] A3: The first 64-dimensional feature vector and the second 64-dimensional feature vector are concatenated in the convergence layer to form a 128-dimensional joint feature vector.
[0046] A4: The feature vector, which integrates spatiotemporal features, is input into the classifier to obtain the bearing fault state classification result. The classifier consists of a fully connected layer, a ReLU activation function, and regularized Dropout. During the classifier processing stage, a multi-layer neural network converts the integrated features into bearing fault classification results. First, the fully connected layer performs dimensionality reduction on the feature vector to filter out key fault features. Then, the ReLU activation function introduces a nonlinear transformation to enhance the model's ability to identify complex fault modes. The Dropout layer prevents overfitting and improves the model's generalization performance by randomly deactivating neurons. Finally, the output layer generates a 10-dimensional classification result, corresponding to different bearing fault states, including normal operation, inner race fault, outer race fault, and rolling element fault, among other operating conditions.
[0047] In one embodiment of this application, the model is built using the Python 3.8.0 programming language and the PyTorch deep learning framework. On the Windows 11 operating system, the model is trained using an Intel i7-13700H processor and an NVIDIA GeForce GTX4060 graphics card, with the Adam optimizer and the cross-entropy loss function as the training objective.
[0048] S3: Based on the trained dual-channel DC-DenseNet-GRU model, the test data is diagnosed to obtain the bearing fault state classification results.
[0049] In one embodiment of this application, test set data is fed into a trained model for diagnostic processing. The diagnostic accuracy under different load conditions is shown in Table 1. Figure 5 As shown, the accuracy of the dual-channel DC-DenseNet-GRU model proposed in this application is higher than 99%, which is significantly better than other comparative models.
[0050] Table 1. Diagnostic accuracy under different load conditions
[0051] To test the noise immunity performance, Gaussian white noise with different signal-to-noise ratios (-7dB, 0dB, 7dB) was added to the test data. The results are shown in Tables 2 and 3. Figure 6 and Figure 7 As shown, the model proposed in this application maintains the highest diagnostic accuracy and the strongest robustness under different noise environments.
[0052] Table 2. Accuracy of different models under different noise environments and at 0HP and 1HP loads.
[0053] Table 3. Accuracy of different models under different noise environments and at 2HP and 3HP loads.
[0054] This application presents a bearing fault diagnosis method based on a dual-channel densely connected network fusion model. Utilizing the dual-channel DC-DenseNet-GRU model, it can simultaneously and fully extract the spatial local features and temporally dynamic dependent features contained in vibration signals. Through feature fusion, it achieves deep complementarity of spatiotemporal information, effectively addressing noise interference and accurately identifying the complex fault states of rolling bearings. Furthermore, the dual-channel DC-DenseNet-GRU model forms an inherent redundancy and complementarity mechanism; when strong noise interference affects one channel, the other channel can still provide effective feature information, thus significantly improving the stability and robustness of the diagnostic system under strong noise pollution. Therefore, this method maintains extremely high diagnostic accuracy under both strong noise and variable operating conditions, providing an effective solution for highly reliable intelligent bearing fault diagnosis in industrial settings. In addition, by constructing a dual-channel deep learning architecture and inputting standardized vibration signal data, this method can be extended to other rotating machinery fault diagnosis tasks, making the solution have a wider range of industrial applications.
[0055] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.
Claims
1. A bearing fault diagnosis method based on a dual-channel densely connected network fusion model, characterized in that, include: S1: Acquire the original vibration signal of the bearing; S2: Construct a dual-channel DC-DenseNet-GRU model and train it using the original vibration signal. The training process includes: A1: The original vibration signal is processed using the DenseNet channel in the dual-channel DC-DenseNet-GRU model to extract the deep spatial features of bearing faults and output the first feature vector. A2: The original vibration signal is segmented using the GRU channel in the dual-channel DC-DenseNet-GRU model, and the segmented signal sequence is processed using the gated cyclic unit GRU network to extract the time-dependent features of bearing faults and output the second feature vector. A3: The first and second feature vectors are fused at the convergence layer to obtain a feature vector that fuses spatiotemporal features; A4: Input the feature vector that integrates spatiotemporal features into the classifier to obtain the bearing fault state classification result; S3: Based on the trained dual-channel DC-DenseNet-GRU model, the test data is diagnosed to obtain the bearing fault state classification results.
2. The bearing fault diagnosis method based on a dual-channel densely connected network fusion model according to claim 1, characterized in that, The dual-channel DC-DenseNet-GRU model includes: parallel DenseNet channels and GRU channels, a convergence layer, and a classifier; The DenseNet channel includes an initial convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, two densely connected blocks, a transition layer, and a global average pooling layer. Each densely connected block contains multiple densely connected layers. The GRU channel includes a two-layer gated recurrent unit (GRU) network and a fully connected layer, wherein the hidden state dimension of the first layer GRU is 32 and the hidden state dimension of the second layer GRU is 1. The classifier consists of a fully connected layer, a ReLU activation function, and a regularized Dropout.
3. The bearing fault diagnosis method based on a dual-channel densely connected network fusion model according to claim 2, characterized in that, The calculation formula for the gated recurrent unit (GRU) network is as follows: in, To update the door, To reset the door, In the candidate hidden state, For the gated recurrent unit (GRU) network The final hidden state of time. For activation function, , and These are the weight matrices corresponding to the update gate, reset gate, and candidate hidden states, respectively. , and These are the bias vectors for the update gate, reset gate, and candidate hidden state, respectively. Indicates the previous moment Hidden state With the current input vector To splice, This represents the concatenated vector of the previous hidden state and the current input under the control of the reset gate.
4. The bearing fault diagnosis method based on a dual-channel densely connected network fusion model according to claim 2, characterized in that, The densely connected block includes: The first densely connected block The input to the densely connected layer is from the first The first to the second The outputs of densely connected layers are concatenated along the channel dimension, where the first... The specific expression for the output of the densely connected layer is: In the formula, Indicates the first The output of the densely connected layer, Represents a nonlinear transformation function. Indicates the first layer to the first The feature maps of densely connected layers are stitched together along the channel dimension.
5. The bearing fault diagnosis method based on a dual-channel densely connected network fusion model according to claim 2, characterized in that, The process utilizes the DenseNet channel in the dual-channel DC-DenseNet-GRU model to process the original vibration signal, extract the deep spatial features of the bearing fault, and output the first feature vector, including: B1: The original vibration signal is sequentially passed through an initial convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer for preliminary feature transformation; B2: Input the pre-transformed features into densely connected blocks and transition layers for deep feature extraction and reuse; B3: Perform global average pooling on the extracted deep spatial features to compress and obtain the first feature vector.
6. The bearing fault diagnosis method based on a dual-channel densely connected network fusion model according to claim 2, characterized in that, The original vibration signal is segmented using the GRU channel in the dual-channel DC-DenseNet-GRU model, and the segmented signal sequence is processed using a gated recurrent unit (GRU) network to extract the time-dependent features of bearing faults, outputting a second feature vector, including: C1: Divide the original vibration signal into multiple signal segments of fixed length; C2: Each signal segment is sequentially input into a two-layer gated recurrent unit (GRU) network to extract the temporal features within the signal segment; C3: Combines the temporal features of all signal segments and outputs a second feature vector through a fully connected layer.
7. The bearing fault diagnosis method based on a dual-channel densely connected network fusion model according to claim 2, characterized in that, The step of inputting the feature vector fused with spatiotemporal features into the classifier to obtain the bearing fault state classification result includes: The fully connected layer performs dimensionality reduction on the feature vectors that fuse spatiotemporal features and filters out key fault features; it then introduces a nonlinear transformation on the key fault features using the ReLU function; the key fault features after the nonlinear transformation are then subjected to regularized Dropout processing; and the output layer maps the key fault features after regularized Dropout processing to fault state classification results.