A bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion

Abstract

This invention discloses a bearing fault diagnosis method based on an improved residual network using multi-sensor signal fusion, belonging to the field of bearing fault diagnosis technology. The steps of this invention are as follows: 1. Using multiple sensors, acquire synchronous vibration signal data of bearing components at different locations and normalize the data; 2. Extract one-dimensional vibration signals from the dataset, convert them into two-dimensional signals, and then fuse the two-dimensional signals obtained from different locations into a multi-channel input; 3. Using a diagnostic model containing three dual-connected attention residual modules, learn the fault features in the multi-channel input data, perform differential classification on the extracted channel features, and calculate the loss value during model training; 4. Based on the obtained loss value, update the weight parameters of the entire network using an error backpropagation algorithm until the maximum number of updates is reached. This invention exhibits good fault classification accuracy under varying operating conditions and noisy environments.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and specifically to a bearing fault diagnosis method based on an improved residual network using multi-sensor signal fusion. Background Technology

[0002] As a critical component of rotating machinery, bearing failure directly impacts equipment performance and can even lead to safety accidents. Therefore, accurately diagnosing the health status of bearings is of great importance for the smooth operation of equipment.

[0003] With the widespread application of information technology, vibration data collected in various industrial scenarios is increasingly being used in fault monitoring. This has led to the rapid development of data-driven fault diagnosis technology, which is now widely used in actual production. Traditional fault diagnosis methods mainly consist of three parts: vibration signal preprocessing, feature extraction, and fault state classification. While traditional methods have some value in improving fault diagnosis results, they suffer from drawbacks such as requiring extensive prior knowledge for data preprocessing, insufficient feature extraction, and poor generalization performance across different scenarios.

[0004] Researchers have developed various deep network models using deep learning theory for adaptive learning of fault data, enabling in-depth mining of nonlinear relationships between data and thus eliminating reliance on expert knowledge. Convolutional neural networks (CNNs), as a typical multi-level feedforward neural network, possess sparse connection structures and weight-sharing characteristics conducive to model training, and have therefore been introduced into the field of fault diagnosis for in-depth research. These methods enhance feature extraction capabilities by constructing deep diagnostic models, improving classification accuracy to some extent. However, as network depth increases, model training becomes difficult, hindering further improvements in classification performance. On one hand, deep networks experience gradient vanishing during training, leading to overfitting; on the other hand, these network models only learn from signal data collected from a single sensor, which contains limited fault feature information, making it difficult for the model to adapt to operating conditions beyond constant load or speed variations. These factors limit the further development of CNNs in fault diagnosis. Therefore, it is necessary to explore bearing fault diagnosis methods that can effectively extract fault feature information and exhibit good fault diagnosis performance under varying operating conditions. Summary of the Invention

[0005] 1. The technical problem that the invention aims to solve

[0006] Considering the influencing factors such as variable load, variable speed, and environmental noise interference under complex operating conditions, this invention proposes a bearing fault diagnosis method based on an improved residual network using multi-sensor signal fusion. This method simplifies the data processing and increases the amount of fault feature information contained in the input data. Simultaneously, the deep network composed of stacked DARMs can effectively mine deep feature information contained in the fused signal. Utilizing the attention mechanism (Sequeeze-and-Excitation Networks, SE) of the DARM allows the model to learn more discriminative fault features, thereby improving the model's recognition performance in complex environments.

[0007] 2. Technical Solution

[0008] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0009] The present invention provides a bearing fault diagnosis method based on an improved residual network using multi-sensor signal fusion, comprising the following steps:

[0010] Step 1: Use multiple sensors to acquire synchronous vibration signal data of bearing components at different positions under different operating conditions, and normalize the data;

[0011] Step 2: Extract one-dimensional vibration signals from the dataset at a fixed length, convert the one-dimensional vibration signals into two-dimensional signals, and then fuse the two-dimensional signals obtained from different positions into a multi-channel input.

[0012] Step 3: Using a diagnostic model containing three bi-connected attention residual modules, learn the fault features in the multi-channel input data, perform differential segmentation on the extracted channel features, and calculate the loss value during the model training process using the cross-entropy loss function.

[0013] Step 4: Based on the obtained loss value, update the weight parameters of the entire network using the backpropagation algorithm until the maximum number of updates set by the network is reached.

[0014] Step 5: Apply the trained model to bearing fault diagnosis and output the fault classification results.

[0015] 3. Beneficial effects

[0016] Compared with existing known technologies, the technical solution provided by this invention has the following significant advantages:

[0017] (1) This invention addresses the problems of insufficient feature information acquired by a single sensor and the lack of importance differentiation in fault features extracted by deep convolutional neural networks. It proposes a bearing fault diagnosis method based on an improved residual network using multi-sensor signal fusion. First, vibration signals collected from different locations are converted into multi-channel inputs using a sensor information fusion strategy to obtain more comprehensive fault feature information. Then, a dual-connected residual network is designed for the fused multi-channel input to enhance the model's feature extraction. Simultaneously, a channel attention mechanism module is introduced to assign different weights to the output features, making the extracted features more discriminative and improving the recognition accuracy. Finally, global average pooling is used to reduce network parameters and mitigate the gradient vanishing problem in deep networks. The proposed method is applied to a bearing dataset under complex working conditions. Experiments show that this method has good fault classification accuracy under varying working conditions and noisy environments.

[0018] (2) The sensor information fusion strategy adopted in this invention simplifies the data processing process and improves the fault feature information contained in the input data.

[0019] (3) In this invention, the deep network composed of stacked DARMs can effectively mine the deep feature information contained in the fused signal, and the channel attention mechanism of DARM can enable the model to learn more discriminative fault features, thereby improving the model's recognition performance in complex environments. Attached Figure Description

[0020] Figure 1 This is a flowchart of the fault diagnosis process of the present invention;

[0021] Figure 2 This is a schematic diagram of data reconstruction according to the present invention;

[0022] Figure 3 This is a schematic diagram of the dual-connection residual network of the present invention;

[0023] Figure 4 This is a model diagram of the bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion according to the present invention;

[0024] Figure 5 This is a comparison chart of the diagnostic accuracy of the variable load method of this invention;

[0025] Figure 6 This is a comparison chart of the noise change diagnosis accuracy results of this invention;

[0026] Figure 7 This is a comparison chart of the diagnostic accuracy of the present invention in a variable speed experiment with noise variations;

[0027] Figure 8 This is a dimension-reduced visualization of the diagnostic process of this invention. Detailed Implementation

[0028] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments.

[0029] Example 1

[0030] This embodiment provides a bearing fault diagnosis method based on an improved residual network using multi-sensor signal fusion. The method uses sensors at different locations as signal sources and employs a data fusion strategy to fuse vibration signals collected by a single sensor at a corresponding time into a multi-channel input. Then, a dual-connected attention residual module is designed to enable the network to differentiate feature information, making the extracted fault features more discriminative. By stacking this module, the network is deepened, allowing the model to deeply mine hidden fault features in the fused signal. Finally, global average pooling is used to mitigate the gradient vanishing problem caused by deep networks.

[0031] like Figure 1 As shown, the method specifically includes the following steps:

[0032] Step 1: Using multiple sensors, acquire synchronous vibration signal data of bearing components at different positions under different operating conditions, normalize the data, and divide it into training set and test set.

[0033] This embodiment uses the Case Western Reserve University Bearing Experiment Dataset (CWRU) to detail the bearing fault diagnosis method based on improved residual networks under multi-sensor signal fusion provided by this invention. The CWRU bearing dataset introduces single-point faults into the inner ring, rolling elements, and outer ring of the test bearing using electrical sparks, with fault sizes of 0.007, 0.014, and 0.021 inches, respectively. Therefore, the bearing fault states can be divided into 9 types, resulting in 10 different bearing operating states when combined with the normal state. In the experiment, the data acquisition frequency was 12kHz, and vibration signals were collected at three different locations—the fan end, the base end, and the drive end—under loads of 0, 1, and 2kW, corresponding to speeds of 1797, 1772, and 1750 r / min. The collected data were sampled non-overlappingly with 1024 sampling points per sample length, with 80% used as the training set and 20% as the test set. The experimental data is described in Table 1.

[0034] Table 1. Description of CWRU Data

[0035]

[0036] Normalizing the collected vibration signal data to a dimensionless range can prevent gradient vanishing during training and accelerate network convergence. The specific normalization operation is as follows:

[0037]

[0038] In the formula, This represents the i-th sample point of the k-th signal segment; max(x k ) and min(x k ) represent the maximum and minimum values ​​of the k-th signal segment, respectively.

[0039] To verify the effectiveness of the present invention under complex working conditions, three datasets, A, B, and C, were created, as shown in Table 2.

[0040] Table 2 Description of Experimental Datasets

[0041]

[0042] Step 2: Extract one-dimensional vibration signals from the dataset at a fixed length, convert them into two-dimensional signals, and then fuse the two-dimensional signals obtained from different locations into a multi-channel input.

[0043] Step 2.1, Data format conversion. For example... Figure 2 As shown, the original signal with 1024 sampling points is truncated into 32 segments of equal length, with each segment consisting of 32 sampling points. The truncated signals are then stacked sequentially to obtain a 32×32 two-dimensional reconstructed signal.

[0044] Step 2.2, Multi-sensor data fusion. The signals collected by each sensor are converted according to the method described in Step 2.1 to obtain a two-dimensional reconstructed signal with a data format of [32, 32, 1], where 1 represents a single channel. To effectively utilize information from different sensors, this embodiment converts the information collected at the same time into two-dimensional data, and then fuses the signals collected from different locations in parallel to form a multi-channel input, thereby increasing the fault feature information contained in the model input. The fused input data format is [32, 32, 3], where 3 represents three channels.

[0045] Step 3: Using a diagnostic model containing three dual-connected attention residual modules (DARMs), learn the fault features in the multi-channel input data and perform differential segmentation on the extracted channel features. The loss value during model training is then calculated using the cross-entropy loss function.

[0046] Specifically, the structure of the DARM designed in this invention is as follows: Figure 3As shown, this structure constructs a residual module with dual connections based on the idea of ​​skip connections, which is used to deepen the diagnostic network. Simultaneously, to enable the dual-connected residual module to distinguish the importance of features from different channels and improve the efficiency of feature learning, a channel attention mechanism module is connected after each dual-connected residual module, thus forming a DARM. The model structure parameters of the diagnostic method proposed in this invention are shown in Table 3.

[0047] Table 3 shows the proposed model structure parameters.

[0048]

[0049] Step 3.1: The preprocessed multi-channel input first passes through two layers of wide-kernel convolutional neural networks with a kernel size of 5*5 and a number of 32 to extract shallow features from the fault signal. The final output size is [24, 24, 32].

[0050] Step 3.2: The extracted features are input into a deep feature extraction structure composed of three stacked DARMs. A dual-connected residual network optimized with an attention mechanism is used to extract high-level feature information from the multi-channel input layer by layer. By repeatedly reusing previous information, gradient vanishing caused by unidirectional information flow is overcome, thereby achieving deep mining of the input information. The multi-channel data input of DARM1 is x. Fault features are extracted through convolution 1 and convolution 3 operations. The output of convolution 1 is then connected to the input x in a skip connection, resulting in the high-level feature y.

[0051] y = f1(w1*x + b1)∪x

[0052] In the formula, the size of the output y is [24, 24, 64], f1 is the activation function of the convolution operation; w1 and b1 are the corresponding weight coefficients and bias terms; ∪ is the parallel operation.

[0053] The y is processed through convolution 2, and its output is then fused with the up-dimensional output obtained from convolution 3 to obtain the high-level feature Y:

[0054]

[0055] In the formula, the size of the output Y is [24, 24, 64], and f2 and f3 are the activation functions of the convolution operation; This is an information addition and fusion operation; w2, w3 and b2, b3 are the corresponding weight coefficients and bias terms.

[0056] Next, for the channel features extracted by the dual-connected residual network, an attention mechanism module is used to perform differential partitioning, thereby distinguishing the importance of different channel feature information and obtaining the weight coefficients of different channel features. First, a global average pooling (GAP) operation, i.e., a compression operation F, is performed on each channel. sq (·), which transforms the H×W×C input features into a 1×1×C feature map along the channel direction. The calculation process is as follows:

[0057]

[0058] In the formula, c is the compressed channel number; z c Y is the compressed value of the c-th channel; c (i,j) is an i×j dimensional two-dimensional matrix, i=1,2…H, j=1,2…W.

[0059] Using activation operation F ex (·,·) Learn the feature weights for each channel. This step is crucial for the SE module to establish the relationship between channel features and training performance. The goal is to ensure that the learned weights incentivize important feature mappings and suppress unimportant ones. Specifically, after compression, a two-layer fully connected gating mechanism is used to learn the weights for different channels. The gating unit calculation method is as follows:

[0060] s c =σ(W2δ(W1z) c ))

[0061] In the formula, δ represents the ReLU activation function; σ represents the sigmoid activation function; W1 and W2 are the network parameters of the two fully connected layers, respectively; s c It is the gating unit obtained after the activation operation.

[0062] Through the product operation F scale (·,·) Output gating unit s c With input feature map submap Y c vector product:

[0063]

[0064] In the formula, obtained through DARM1 The weighted feature map enhances the discriminative power of the dual-connected residual network in extracting fault features, and its output size is [24, 24, 64].

[0065] Finally, the high-level feature map obtained from DARM1 is reduced in dimensionality through pooling, and the output size is [12, 12, 64].

[0066] Step 3.3: Pass the dimensionality-reduced output of DARM1 through DARM2 and DARM3 sequentially, and perform pooling dimensionality reduction on each. The final output size is [3, 3, 256].

[0067] Step 3.4: Input the high-level feature maps extracted layer by layer through the three DARMs into the GAP network layer to reduce the network training parameters and prevent overfitting.

[0068] Step 3.5: Use the softmax layer to output the fault category and calculate the loss value during model training using the cross-entropy loss function.

[0069]

[0070] In the formula, n represents the number of categories in the classification task; x represents the output of the last layer of the fully connected layer. The classification task is implemented using the softmax function, and the softmax processing ensures that the final output satisfies the probability distribution based on the number of categories in the classification task, with a sum of 1.

[0071] The expression for the cross-entropy loss function is:

[0072]

[0073] In the formula, Loss represents the value of the cross-entropy loss function; n represents the number of samples; y i This represents the label of a sample in the i-th dimension using one-hot encoding; that is, when the sample is i, y... i =1, and the rest of the positions are 0; This represents the predicted label output in one-hot encoded form.

[0074] Step 4: Based on the obtained loss value, update the weight parameters of the entire network using the backpropagation algorithm to reduce the loss function value and improve diagnostic accuracy until the maximum number of updates set by the network is reached.

[0075] This process mainly utilizes the chain rule, which involves solving the loss function with respect to the reciprocals of each weight parameter in the network, and then using the gradient descent algorithm to iteratively update the weight parameters in the network, thereby achieving model optimization during the training process.

[0076] The fault diagnosis method proposed in this invention, based on sensor information fusion and a double-connected attention residual network, increases the number of fault features in the input signal, improves the model's feature learning effect, and thus enhances fault diagnosis accuracy. To evaluate the diagnostic performance of the proposed model in complex working conditions from multiple perspectives, it is compared with Wide-Kernel Convolutional Neural Network (WDCNN), Ordinary Residual Network (ResNet), Double-Connected Residual Network (DRN), and the Improved Residual Network Under Single Sensor Input (IRN-SSF) method using the diagnostic model proposed in this embodiment. The parameters of the selected comparison models are shown in Table 4.

[0077] Table 4 Comparison of Model Structure Parameters

[0078]

[0079]

[0080] To verify the actual diagnostic effect of the proposed method under load variations, this embodiment uses a dataset with a constant load as the training set and datasets with different loads as the test set to simulate load variations. For example, AB represents training on dataset A and testing on dataset B. Each experiment is performed 10 times, and the average value is taken as the experimental result. Finally, under load variation conditions, the diagnostic results of different methods are obtained, such as... Figure 5 As shown, the proposed method demonstrates higher diagnostic accuracy than the comparative method under all load variations, with an average diagnostic accuracy exceeding 94%.

[0081] Meanwhile, to verify the noise resistance performance of the proposed method under complex working conditions, Gaussian white noise ranging from 0 to 8 dB was added to the test set of dataset B to test the noise resistance capabilities of different methods. Each experiment was conducted 5 times, and the mean of the experimental results was compared with... Figure 6 As shown in the figure, when the noise intensity is 0-8 dB, the diagnostic accuracy of the proposed method is higher than that of other diagnostic models, with an average recognition accuracy exceeding 98%. Furthermore, when the SNR changes from 0 dB to 8 dB, only the diagnostic model of this invention can consistently maintain a diagnostic accuracy of over 93%, indicating that the proposed method still has a strong ability to extract fault features in noisy environments.

[0082] Example 2

[0083] This embodiment of the bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion is basically the same as Embodiment 1, except that this embodiment uses a self-made bearing experimental dataset to provide a detailed description of the bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion provided by this invention. In this embodiment, electrical discharge machining (EDM) is used to process different fault sizes on the inner ring, outer ring, and rolling elements of the bearing. Accelerometers placed at different locations are used to collect bearing vibration signals at three operating conditions: 900 r / min, 1200 r / min, and 1500 r / min, with a sampling frequency of 10.24 kHz, resulting in datasets A, B, and C. The collected vibration signals are divided into non-overlapping segments of 1024 samples each, with 80% used as the training set and 20% as the test set. The composition of each dataset is shown in Table 5.

[0084] Table 5 Description of Safety Engineering Big Data

[0085]

[0086] To evaluate the diagnostic performance of the proposed model under complex working conditions from multiple perspectives, it is compared with Wide Kernel Convolutional Neural Network (WDCNN), Ordinary Residual Network (ResNet), Double-Connected Residual Network (DRN), and the Improved Residual Network Under Single Sensor Input (IRN-SSF) method using the diagnostic model proposed in this embodiment. The parameters of the selected comparison models are shown in Table 6.

[0087] Table 6 Comparison of Model Structure Parameters

[0088]

[0089] Considering the variations in bearing speed and the interference of noise environment during operation, this embodiment designs a variable speed experiment with varying noise levels. The specific experimental process is as follows: the diagnostic model is trained using data collected at one speed, and then the trained model is tested using data collected at another speed. For example, AB represents training in dataset A and testing the trained model in dataset B. Simultaneously, to test the noise resistance performance of the diagnostic model under varying operating conditions, Gaussian white noise with different SNRs is added to the test dataset.

[0090] To avoid randomness in the experiment, the experiment was repeated 5 times for each noise environment condition change, and the average value was taken. The experimental results are shown in Table 7. Figure 7 As shown.

[0091] In Table 7, 0, 2, 4, and 6 represent the signal-to-noise ratio (SNR) with Gaussian white noise added to the test set. As shown in Table 7, under the same rotational speed variation, the proposed method achieves a higher average accuracy in fault type identification than other methods in noise environments with an SNR of 0–6. Furthermore, under noise environments with the same SNR, the proposed method also achieves a higher average accuracy in fault identification when rotational speed varies compared to the comparative methods. Particularly in high-noise environments with an SNR of 0, the proposed method achieves an accuracy exceeding 90% compared to IRN-SSF and DRN, further demonstrating that the diagnostic method constructed in this invention possesses strong feature learning capabilities without relying on manual feature extraction. Moreover, when testing bearings with an SNR of 0–6 and varying rotational speed, the average accuracy of the model in this invention reaches 96.46%, indicating that the proposed method has good fault identification accuracy and anti-interference performance under complex operating conditions.

[0092] Table 7. Comparison of Diagnostic Accuracy of Model Experiments

[0093]

[0094]

[0095] To intuitively demonstrate the diagnostic process of the proposed method, this embodiment employs the t-SNE dimensionality reduction algorithm to reduce the high-dimensional features in the input vibration signal, the output of the two-layer wide-kernel convolutional layer, the outputs of the three DARMs in the model, and the output of the last layer of the fully connected network to two dimensions, thereby achieving visualization of the feature extraction of the diagnostic method. The training set used in the experiment is A, and then a test set B with added noise intensity and an SNR of 0 dB is used to obtain the visualization results of the trained diagnostic model at different stages of feature extraction, such as... Figure 8 As shown.

[0096] It can be seen that the separability of the diagnostic model for different fault types gradually increases with the depth of the network. In particular, after learning from three DARM modules, the model's ability to distinguish fault types is significantly enhanced, indicating that the DARM used in this invention has good feature extraction capabilities and can extract fault features that reflect the bearing's operating state from vibration signals. Furthermore, the visualization results of high-dimensional features output by DARM modules at different depths show that the deep network formed by stacking three DARM modules in the diagnostic model of this invention is more conducive to the model's differentiation of different types of faults. In addition, the visualization results of the entire diagnostic process show that the model trained on dataset A, when tested on dataset B in a noisy environment with an SNR of 0 dB, can still clearly identify the boundaries between different types of faults, further demonstrating that the proposed method has good recognition performance in noisy and variable operating conditions.

Claims

1. A bearing fault diagnosis method based on an improved residual network using multi-sensor signal fusion, characterized in that, Includes the following steps: Step 1: Use multiple sensors to acquire synchronous vibration signal data of bearing components at different positions under different operating conditions, and normalize the data; Step 2: Extract one-dimensional vibration signals from the dataset at a fixed length, convert the one-dimensional vibration signals into two-dimensional signals, and then fuse the two-dimensional signals obtained from different positions into a multi-channel input. Step 3: Using a diagnostic model containing three bi-connected attention residual modules, learn the fault features in the multi-channel input data, perform differential segmentation on the extracted channel features, and calculate the loss value during the model training process using the cross-entropy loss function. Step 4: Based on the obtained loss value, update the weight parameters of the entire network using the backpropagation algorithm until the maximum number of updates set by the network is reached. Step 5: Apply the trained model to bearing fault diagnosis and output the fault classification results.

2. The bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion according to claim 1, characterized in that: Step 1 involves non-overlapping sampling with 1024 sampling points as one sample length, and normalizing the collected original synchronous vibration signal data to the same dimensionless range.

3. The bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion according to claim 2, characterized in that: In step 2, the original signal with a length of 1024 sampling points is truncated into 32 segments of equal length, with each segment consisting of 32 sampling points. The truncated signals are then stacked sequentially to obtain a 32×32 two-dimensional reconstructed signal.

4. The bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion according to claim 3, characterized in that: In step 2, the signals collected by each sensor at the same time are converted into two-dimensional reconstructed signals, and then the signals collected at different locations are fused in parallel to form a multi-channel input.

5. A bearing fault diagnosis method based on an improved residual network under multi-sensor signal fusion according to any one of claims 1-4, characterized in that: The diagnostic model described in step 3 includes two convolutional layers connected in sequence and three dual-connected attention residual modules DARM1, DARM2, and DARM3. Each of DARM1, DARM2, and DARM3 is followed by a pooling layer. After the pooling layer connected to DARM3, a global average pooling layer and a Softmax classification layer are connected.

6. The bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion according to claim 5, characterized in that: The dual-connected attention residual module (DARM) is built on skip connections to construct residual modules with dual connections. After each dual-connected residual module, a channel attention mechanism module is connected.

7. The bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion according to claim 6, characterized in that: The dual-connection residual module includes convolution 1, convolution 2, and convolution 3. Multi-channel input data x is processed by convolution 1 and convolution 3 to extract fault features. The output of convolution 1 is then connected to the input x in a jump connection to obtain high-level features y. y is then processed by convolution 2, and its output is connected to the up-dimensional output obtained by convolution 3 in a jump connection to obtain high-level features Y.

8. The bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion according to claim 7, characterized in that: The channel attention mechanism module performs a compression operation F on each channel. sq (·) transforms the H×W×C input features into a 1×1×C feature map along the channel direction; after compression, a gate mechanism consisting of two fully connected layers is used to learn the feature weights of different channels; through the product operation F scale (·,·) is the vector product of the output gating unit and the input feature map subgraph. This vector product is the weighted feature map obtained by the dual-connected attention residual module (DARM).

9. The bearing fault diagnosis method based on an improved residual network under multi-sensor signal fusion according to claim 8, characterized in that, The specific steps of step 3 are as follows: Step 3.1: The preprocessed multi-channel input is passed through a two-layer wide-kernel convolutional neural network to extract shallow features from the fault signal; Step 3.2: The extracted features are input into a deep feature extraction structure composed of three stacked DARMs. The high-level feature information of the multi-channel input is extracted layer by layer using a dual-connected residual module optimized by the attention mechanism. For the channel features extracted by the dual-connected residual module, the channel attention mechanism module is used to perform differential partitioning. The high-level feature map obtained by DARM1 is reduced in dimensionality by pooling operation. Step 3.3: Pass the dimensionality-reduced output of DARM1 through DARM2 and DARM3 in sequence, and perform pooling dimensionality reduction on each of them; Step 3.4: Input the high-level feature maps extracted layer by layer through the three DARMs into the global average pooling layer; Step 3.5: Use the softmax layer to output the fault category and calculate the loss value during model training using the cross-entropy loss function.

10. The bearing fault diagnosis method based on improved residual network under multi-sensor signal fusion according to claim 8, characterized in that: Step 4 utilizes the chain rule to optimize the model by solving the loss function with respect to the reciprocals of each weight parameter in the network and using the gradient descent algorithm to iteratively update the weight parameters in the network.

Images