Fault diagnosis method and system based on multi-scale residual attention network
Patent Information
- Application Number
- CN202310296694.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-23
- Publication Date
- 2026-08-21
- Estimated Expiration
- 2043-03-23
AI Technical Summary
[0006]针对现有技术的不足，本发明提供了一种基于多尺度残差注意力网络的故障诊断方法及系统，解决了现有故障诊断技术存在精度低的问题
[0026] This invention provides a fault diagnosis method and system based on a multi-scale residual attention network. Compared with existing technologies, it has the following advantages:
Smart Images
Figure CN116595313B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault diagnosis technology, specifically to a fault diagnosis method and system based on a multi-scale residual attention network. Background Technology
[0002] Against the backdrop of new-generation information technology, the application of high-end equipment is becoming increasingly diversified, and its precision and performance requirements are also becoming higher and higher. With the vigorous development of the information technology industry, the demand for equipment performance and maintenance has become extremely urgent and important, especially for components such as ship propulsion equipment, turbines, and gearboxes. Once a failure occurs, it can lead to serious consequences and cause incalculable losses.
[0003] Currently, equipment fault diagnosis generally includes three main steps: raw data acquisition and processing, feature extraction of the data (e.g., manual feature annotation based on expert experience), and fault classification using relevant models. Among these, feature extraction and fault classification are two crucial steps affecting the accuracy of fault diagnosis.
[0004] However, traditional fault diagnosis techniques typically rely on expert experience and manual feature annotation during feature extraction, which introduces a degree of subjectivity, leading to poor classification performance and low accuracy. Furthermore, much current research focuses on extracting single-scale features from a single perspective. However, when machinery operates normally, the timing, cause, and location of faults are random, and signals from damage at different locations may exhibit coupling during acquisition. Therefore, single-scale feature extraction may miss important local features, also affecting classification accuracy. In addition, noise is constantly present during equipment manufacturing and data acquisition, and existing models have poor noise resistance and weak generalization capabilities. Therefore, using existing models for fault classification also results in low classification accuracy. Summary of the Invention
[0005] (a) Technical problems to be solved
[0006] To address the shortcomings of existing technologies, this invention provides a fault diagnosis method and system based on multi-scale residual attention networks, which solves the problem of low accuracy in existing fault diagnosis technologies.
[0007] (II) Technical Solution
[0008] To achieve the above objectives, the present invention provides the following technical solution:
[0009] Firstly, this invention proposes a fault diagnosis method based on a multi-scale residual attention network, the method comprising:
[0010] Acquire signal data from the device to be diagnosed, and preprocess the signal data;
[0011] A fault diagnosis model based on a multi-scale residual attention network is constructed. The multi-scale residual attention network includes a wide convolutional module, a multi-scale convolutional module, and a residual attention module connected in sequence. The wide convolutional module uses dilated convolutions to replace the small convolutional kernels of the original neural network to extract long-term feature information of the signal data. The multi-scale convolutional module uses a soft thresholding function as a residual branch on the original Res2block module and uses two depthwise separable convolutions to replace the regular small convolutions in each branch of the original Res2block module. The residual attention module adds a hybrid pooling strategy to the channel attention mechanism of the CBAM attention mechanism, and uses the dot product result of the channel attention mechanism and the spatial attention mechanism with the added hybrid pooling strategy as the second residual connection branch, while the input of the residual attention module is directly used as the first residual connection branch.
[0012] Fault diagnosis is achieved based on the trained fault diagnosis model.
[0013] Preferably, the preprocessing includes data augmentation of the signal data using a sliding window technique.
[0014] Preferably, the hybrid pooling strategy is a weighted operation on global max pooling and global min pooling.
[0015] Preferably, the depthwise separable convolution includes channel convolution and point convolution; the channel convolution determines the number of convolution kernels based on the dimension of the input channels and extracts a number of feature maps equal to the number of input channels; the point convolution uses 1x1 convolution kernels to extract features, and the number of convolution kernels is set to the number of feature maps to be obtained.
[0016] Preferably, the fault diagnosis model further includes a classification module, which is used for classifying and outputting fault diagnosis results, and the classification module includes an adaptive average pooling module and a convolution module.
[0017] Secondly, this invention also proposes a fault diagnosis system based on a multi-scale residual attention network, the system comprising:
[0018] The data acquisition and processing module is used to acquire signal data of the device to be diagnosed and to preprocess the signal data.
[0019] A fault diagnosis model acquisition module is used to construct a fault diagnosis model based on a multi-scale residual attention network. The multi-scale residual attention network includes a wide convolutional module, a multi-scale convolutional module, and a residual attention module connected in sequence. The wide convolutional module uses dilated convolutions to replace the small convolutional kernels of the original neural network to extract long-term feature information from the signal data. The multi-scale convolutional module uses a soft thresholding function as a residual branch on the original Res2block module, and replaces the regular small convolutions in each branch of the original Res2block module with two depthwise separable convolutions. The residual attention module adds a hybrid pooling strategy to the channel attention mechanism of the CBAM attention mechanism, and uses the dot product result of the channel attention mechanism and the spatial attention mechanism with the added hybrid pooling strategy as the second residual connection branch, while directly using the input of the residual attention module as the first residual connection branch.
[0020] The fault diagnosis module is used to perform fault diagnosis based on the trained fault diagnosis model.
[0021] Preferably, the data acquisition and processing module performs preprocessing operations including data enhancement of the signal data using sliding window technology.
[0022] Preferably, the hybrid pooling strategy is a weighted operation on global max pooling and global min pooling.
[0023] Preferably, the depthwise separable convolution includes channel convolution and point convolution; the channel convolution determines the number of convolution kernels based on the dimension of the input channels and extracts a number of feature maps equal to the number of input channels; the point convolution uses 1x1 convolution kernels to extract features, and the number of convolution kernels is set to the number of feature maps to be obtained.
[0024] Preferably, the fault diagnosis model further includes a classification module, which is used for classifying and outputting fault diagnosis results, and the classification module includes an adaptive average pooling module and a convolution module.
[0025] (III) Beneficial Effects
[0026] This invention provides a fault diagnosis method and system based on a multi-scale residual attention network. Compared with existing technologies, it has the following advantages:
[0027] This invention constructs a fault diagnosis model based on a multi-scale residual attention network and uses the trained fault diagnosis model for equipment fault diagnosis. The multi-scale residual attention network includes sequentially connected wide convolutional modules, multi-scale convolutional modules, and residual attention modules. The fault diagnosis method of this invention can extract features implicit in equipment signal data from multiple scales, while also suppressing noise interference in the equipment signal data to a certain extent. Compared with traditional fault diagnosis techniques, it improves the accuracy and robustness of fault diagnosis. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] Figure 1 This is a flowchart of a fault diagnosis method based on a multi-scale residual attention network according to the present invention;
[0030] Figure 2 This is an embodiment of a fault diagnosis method based on a multi-scale residual attention network according to the present invention.
[0031] Figure 3 This is a structural diagram of the original Res2Net module in an embodiment of the present invention;
[0032] Figure 4 This is a structural diagram of the multi-scale convolution module in an embodiment of the present invention;
[0033] Figure 5 This is a structural diagram of the CBAM attention mechanism before its improvement in this embodiment of the invention;
[0034] Figure 6 This is a structural diagram of the improved channel attention mechanism in an embodiment of the present invention;
[0035] Figure 7 This is a structural diagram of the improved spatial attention mechanism in an embodiment of the present invention;
[0036] Figure 8 This is a structural diagram of the residual attention module in an embodiment of the present invention;
[0037] Figure 9 This is a simplified diagram of the overall structure of the multi-scale residual attention network in an embodiment of the present invention;
[0038] Figure 10 This is a structural diagram of the multi-scale residual attention network in an embodiment of the present invention. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0040] This application provides a fault diagnosis method and system based on a multi-scale residual attention network, which solves the problem of low accuracy in existing fault diagnosis technologies, achieves high-precision diagnosis of equipment faults, and thus improves the efficiency of equipment maintenance.
[0041] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:
[0042] To address the low accuracy issues of existing fault diagnosis technologies due to various reasons, this application proposes a multi-scale residual attention network based on deep learning. This network mainly includes a wide convolutional module, a multi-scale module, and a residual attention module. The wide convolutional module extracts long-term features of the acquired equipment signals; the multi-scale module samples the acquired equipment signals at different granularities, providing both global and local detailed information; and the residual attention module assigns relatively large weights to important features in the acquired equipment signals and relatively small weights to noise, focusing on the main features of the signal and ignoring secondary features (noise interference). This technical solution effectively solves the problem of equipment fault diagnosis, improves the accuracy of fault diagnosis, and exhibits high robustness.
[0043] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.
[0044] The technical solutions of this application include, but are not limited to, fault diagnosis of components and equipment such as ship propulsion systems, turbines, and gearboxes, and are also applicable to fault diagnosis of other high-end equipment. For ease of explanation, the following embodiments of this application will use fault diagnosis of ship propulsion systems as an example.
[0045] Example 1:
[0046] Firstly, this invention proposes a fault diagnosis method based on a multi-scale residual attention network, see [link to relevant documentation]. Figure 1-2 The method includes:
[0047] S1. Acquire the signal data of the device to be diagnosed, and preprocess the signal data;
[0048] S2. Construct a fault diagnosis model based on a multi-scale residual attention network; the multi-scale residual attention network includes a wide convolutional module, a multi-scale convolutional module, and a residual attention module connected in sequence; wherein, the wide convolutional module includes replacing the small convolutional kernels of the original neural network with dilated convolutions to extract long-term feature information of the signal data; the multi-scale convolutional module includes using a soft thresholding function as a residual branch on the original Res2block module, and using two depthwise separable convolutions to replace the regular small convolutions of each branch in the original Res2block module; the residual attention module includes adding a hybrid pooling strategy to the channel attention mechanism of the CBAM attention mechanism, and using the dot product result of the channel attention mechanism and the spatial attention mechanism with the added hybrid pooling strategy as the second residual connection branch, and using the input of the residual attention module directly as the first residual connection branch;
[0049] S3. Fault diagnosis is achieved based on the trained fault diagnosis model.
[0050] As can be seen, this embodiment constructs a fault diagnosis model based on a multi-scale residual attention network and uses the trained fault diagnosis model to diagnose equipment faults. The multi-scale residual attention network includes sequentially connected wide convolutional modules, multi-scale convolutional modules, and residual attention modules. The fault diagnosis method of this embodiment can extract the features implicit in the equipment signal data from a multi-scale perspective, and can also suppress the interference of noise signals in the equipment signal data to a certain extent, thus improving the accuracy and robustness of fault diagnosis.
[0051] The following is in conjunction with the appendix Figure 1-10 The following details the implementation process of an embodiment of the present invention, including explanations of the specific steps S1-S3.
[0052] S1. Acquire the signal data of the device to be diagnosed and preprocess the signal data.
[0053] Signal data collection. During equipment operation, the ship's main propulsion equipment generates signals due to vibration. Accelerators are installed at various locations on the ship's main propulsion equipment; these locations are generally the fault locations, such as the outer and inner rings and rolling elements of bearings, which are prone to damage. The corresponding vibration signals are collected and then processed and classified.
[0054] Data Augmentation. Considering the insufficient amount of vibration signal data, we need to augment the original vibration signal data. This embodiment employs a sliding window technique for data augmentation, expanding the dataset without affecting data classification. The essence of the sliding window technique is to sample by moving the window. After setting the window size, sampling is performed sliding from the original signal, preserving the original data signal while increasing the dataset sample size, achieving significant results on small datasets.
[0055] Divide the dataset. Randomly divide the data-augmented signal data into a training set and a test set. The training set is used to train the neural network, and the test set is used to test the model's characteristics. In this embodiment, the ratio of the training set to the test set is generally 8:2 or 7:3, which can be selected according to actual needs.
[0056] S2. Construct a fault diagnosis model based on a multi-scale residual attention network; the multi-scale residual attention network includes a wide convolutional module, a multi-scale convolutional module, and a residual attention module connected in sequence; wherein, the wide convolutional module includes replacing the small convolutional kernels of the original neural network with dilated convolutions to extract long-term feature information of the signal data; the multi-scale convolutional module includes using a soft thresholding function as a residual branch on the original Res2block module, and using two depthwise separable convolutions to replace the regular small convolutions of each branch in the original Res2block module; the residual attention module includes adding a hybrid pooling strategy to the channel attention mechanism of the CBAM attention mechanism, and using the dot product result of the channel attention mechanism and the spatial attention mechanism with the added hybrid pooling strategy as the second residual connection branch, and using the input of the residual attention module directly as the first residual connection branch.
[0057] To address the low classification accuracy of existing fault diagnosis technologies, this embodiment constructs a fault diagnosis model based on a multi-scale residual attention network. The fault diagnosis model includes a sequentially connected wide convolutional module, a multi-scale convolutional module, and a residual attention module. Specifically, signal data is first input into the wide convolutional module, then to the multi-scale convolutional module, and finally to the residual attention mechanism module to complete network feature learning of the signal data. The fault diagnosis result is then output through the classification module. Specifically:
[0058] (1) Construct a wide convolution module. The wide convolution module includes replacing the small convolutions of the original neural network with dilated convolutions to extract long-term feature information of the signal data.
[0059] Considering that classic neural networks mostly use small convolutional kernels (similar to 3x3, 1x1) to extract features from signal data, this embodiment proposes a convolutional structure with expanded wide convolutional kernels. This structure extracts long-term correlated multi-scale short-term features from noisy signals through sparse sampling and suppresses noise interference through filtering. The core of wide convolution is to improve the model's robustness to noise by mapping more long-term correlated information and a larger receptive field. In this embodiment, a 5×5 dilated convolution is used instead of a typical 3×3 convolutional kernel, with a dilation rate set to 2. Therefore, the 5×5 dilated convolution kernel is equivalent to a 9×9 convolution kernel, expanding the receptive field without increasing parameters. Dilated convolution adds dilation-1 elements (holes) between the kernel elements to expand the receptive field and extract signal features.
[0060] (2) Construct a multi-scale convolution module. The multi-scale convolution module includes using a soft threshold function as a residual branch on the original Res2block module, and using two depthwise separable convolutions to replace the regular small convolutions in each branch of the original Res2block module.
[0061] In industrial settings, equipment fault signals often possess multiple characteristics. Extracting features from ship propulsion equipment using a single-scale fault diagnosis model frequently results in low accuracy and poor robustness. To address this issue of low diagnostic accuracy caused by single-scale feature extraction, this embodiment proposes a multi-scale feature extraction method. Therefore, an improved Res2block module is designed, and a multi-scale convolution module is constructed based on this improved module. The multi-scale convolution module performs a convolution operation on a signal segment to obtain a feature map, which is also a feature vector. This feature information is divided into n equal parts, each of which undergoes different processing. For example... Figure 3 As shown, the original Res2block module divides the feature vector into four sub-feature vectors. The first sub-feature vector is left unprocessed. The second sub-feature vector is processed by a 3x3 convolution kernel to obtain a new feature vector. The third sub-feature vector is added to the feature vector after the second convolution, and then processed by another 3x3 convolution kernel to obtain a new feature vector. The final sub-feature vector is convolved with the result of the third sub-feature vector and added to it, again processed by a 3x3 convolution kernel. Finally, the feature information is integrated and aggregated before being sent to the next module. Figure 4As shown, the improved Res2block module (i.e., the multi-scale convolution module) mainly includes: First, a soft threshold function is added to the multi-scale convolution module to remove redundant features obtained from the previous layer, facilitating the backpropagation of gradients. Soft thresholding is a function that shrinks the input data towards zero and is often used in signal denoising algorithms. Specifically, the soft threshold can set the features in a certain interval [-τ, τ] to 0, and this branch can suppress noise interference to a certain extent. Second, considering the result of the first segment feature vector after convolution, all 3x3 convolution kernels are replaced with two depthwise separable convolution kernels.
[0062] Depthwise separable convolution consists of channel convolution and point convolution. Channel convolution determines the number of kernels based on the dimension of the input channels, extracting a number of feature maps equal to the number of input channels. Point convolution, on the other hand, uses 1x1 kernels to extract features, with the number of kernels set to the desired number of feature maps. The parameters of depthwise separable convolution are significantly smaller than those of ordinary convolution. This configuration reduces the overall parameters of the multi-scale convolution module, improves model training speed, and avoids overfitting.
[0063] (3) Constructing a residual attention module. The residual attention module includes adding a hybrid pooling strategy to the channel attention mechanism of the CBAM attention mechanism, and using the dot product result of the channel attention mechanism and the spatial attention mechanism with the added hybrid pooling strategy as the second residual connection branch, while using the input of the residual attention module directly as the first residual connection branch.
[0064] After solving the multi-scale feature extraction problem in fault diagnosis through the above-mentioned multi-scale convolution module, another key point of this embodiment is to solve the problem of vibration signals containing noise and other interference factors.
[0065] In real-world industrial environments, significant noise levels are often present, necessitating the consideration of noise impact when analyzing vibration signals from ship propulsion systems. Therefore, this embodiment constructs a residual attention module. This module is an improvement upon the CBAM attention mechanism, incorporating three pooling operations. This allows the module to focus on important features while simultaneously processing noise signals, assigning them less weight during network training, thereby improving the model's robustness and accuracy.
[0066] The original CBAM attention mechanism mainly consists of two parts: channel attention and spatial attention. Figure 5 As shown. The essence of channel attention mechanism is to assign different weights to different channels, making the model focus more on effective channel information. Different channel weights are assigned for different tasks, learning the importance of different channels. The essence of spatial attention mechanism is to locate the target and perform some transformations or obtain weights. A mixed pooling strategy is added to the residual attention module. This mixed pooling strategy uses global max pooling and global min pooling for weighted operations. Adding the mixed pooling strategy to the channel attention mechanism can extract richer feature information, such as... Figure 6 As shown.
[0067] In addition, such as Figure 7-8 As shown, after designing the Channel Attention and Spatial Attention mechanisms, two residual connection branches are set up: the first residual branch is derived from the input features fed into the residual attention module, and the second residual connection is derived from the dot product of the Channel Attention and Spatial Attention mechanisms. Typically, the attention mechanism in a neural network plays a bridging role. In this embodiment, the result of the attention mechanism is used as the residual; that is, the product of the attention weights of the two mechanisms is used as the residual branch input to the next layer of the network, which is equivalent to learning the feature information that the attention focuses on. This residual strategy has achieved good results, enhancing the robustness of the fault diagnosis model.
[0068] A fault diagnosis model is constructed based on a multi-scale residual attention network. The three main modules—the wide convolutional module, the multi-scale convolutional module, and the residual attention module—are fused together to complete the construction of a fault diagnosis model for the ship's main propulsion equipment.
[0069] In this embodiment, the fault diagnosis model includes three main modules: a wide convolutional module, a multi-scale convolutional module, and a residual attention module. It also includes a classification module, which is used for classifying and outputting fault diagnosis results. This classification module is mainly composed of an adaptive average pooling module and a convolutional module. The probability of each category is obtained through the Softmax function to achieve fault diagnosis. Specifically, as shown... Figure 9-10As shown, the collected signal is input into a wide convolutional module, where the signal data passes through to obtain the first layer of features. This first layer feature is then passed to a multi-scale convolutional module, where it is divided into four sub-features. These sub-features are then passed through two depthwise separable convolutional modules, with the feature information from the previous segment being incorporated into the feature information of the next segment, making the contextual features more closely linked and demonstrating the characteristics of information extraction at multiple scales. The features extracted from these four branches are then fused. The fused features are then input into a residual attention mechanism module. The residual attention network first passes through a channel attention module to extract channel attention features before being fed into a spatial attention mechanism. The channel attention weights are multiplied by the spatial attention weights, and this multiplication is used as a residual branch in the training of subsequent modules. After the signal data undergoes network feature learning through the multi-scale residual attention network, the fault diagnosis result is finally output through a classification module.
[0070] S3. Fault diagnosis is achieved based on the trained fault diagnosis model.
[0071] After the fault diagnosis model is built, in order to further ensure the accuracy of the fault diagnosis results, we first train and then optimize the built fault diagnosis model. Specifically:
[0072] (1) Fault diagnosis model parameter settings. The experiment used the Tesla V100 deep learning platform and GPU acceleration for training. The model training time was set to 100 epochs, the optimizer was Adam, the learning rate was 0.001, the batch size was set to 128, and the loss function was cross-entropy loss. Each experiment was run 10 times to avoid randomness issues.
[0073] (2) Model Training, Saving, and Testing. First, the training set data divided in S1 is input into the fault diagnosis model built on a multi-scale residual attention network. The model processes the input data in batches according to the batch size. After training, the classification effect is judged based on the magnitude of the cross-entropy loss function. The model parameters are updated using the gradient backpropagation principle. When the training iterations reach 100, the accuracy of each model classification is recorded, compared, and the optimal model is saved. The test set data is used for testing, and the classification performance of the fault diagnosis model is evaluated based on the test set accuracy index and the magnitude of the loss.
[0074] (3) Fault diagnosis and classification. The test set data is input into the trained fault diagnosis model based on a multi-scale residual attention network. The classification effect of the model is measured by evaluation indicators such as accuracy, loss size, confusion matrix and F1 scores. When the expected training effect is achieved, the optimized fault diagnosis model can be used to diagnose the faults of the ship's main propulsion equipment.
[0075] This completes the entire process of the fault diagnosis method based on a multi-scale residual attention network in this embodiment.
[0076] To verify the accuracy of the fault diagnosis method based on a multi-scale residual attention network proposed in this embodiment in fault classification, the inventors tested it on the CWRU dataset from Case Western Reserve University. This dataset is a 10-class fault detection task. The fault diagnosis model based on a multi-scale residual attention network proposed in this embodiment was used for prediction and classification. The experiment showed that its classification accuracy was 100%, which demonstrates the effectiveness of the model and its applicability to data from actual high-end equipment. Furthermore, by simulating noise in a real-world background using Gaussian noise, the result also achieved an accuracy of 97.50% (when SNR = -4), indicating that the model can achieve a high fault diagnosis effect even when the signal contains noise.
[0077] Furthermore, the fault diagnosis model based on multiscale residual attention network (MRANet) proposed in this embodiment is compared with the parameters of classic networks (such as MolileNet, VGG, CNN, ResNet, transformer, Unet), and the comparison results are shown in Table 1 below.
[0078] Table 1 Comparison of Model Parameters
[0079] Parameters 10289 3215625 139357544 17302420
[0080] As shown in Table 1, the fault diagnosis model constructed based on the multi-scale residual attention network in this embodiment effectively improves the model classification accuracy while having a very small number of parameters, greatly reducing training costs and time costs, and saving resources and time.
[0081] Example 2:
[0082] Secondly, the present invention also provides a fault diagnosis system based on a multi-scale residual attention network, the system comprising:
[0083] The data acquisition and processing module is used to acquire signal data of the device to be diagnosed and to preprocess the signal data.
[0084] A fault diagnosis model acquisition module is used to construct a fault diagnosis model based on a multi-scale residual attention network. The multi-scale residual attention network includes a wide convolutional module, a multi-scale convolutional module, and a residual attention module connected in sequence. The wide convolutional module uses dilated convolutions to replace the small convolutional kernels of the original neural network to extract long-term feature information from the signal data. The multi-scale convolutional module uses a soft thresholding function as a residual branch on the original Res2block module, and replaces the regular small convolutions in each branch of the original Res2block module with two depthwise separable convolutions. The residual attention module adds a hybrid pooling strategy to the channel attention mechanism of the CBAM attention mechanism, and uses the dot product result of the channel attention mechanism and the spatial attention mechanism with the added hybrid pooling strategy as the second residual connection branch, while directly using the input of the residual attention module as the first residual connection branch.
[0085] The fault diagnosis module is used to perform fault diagnosis based on the trained fault diagnosis model.
[0086] Optionally, the data acquisition and processing module may perform preprocessing operations including data enhancement of the signal data using sliding window technology.
[0087] Optionally, the hybrid pooling strategy is to perform a weighted operation on global max pooling and global min pooling.
[0088] Optionally, the depthwise separable convolution includes channel convolution and point convolution; the channel convolution determines the number of convolution kernels based on the dimension of the input channels and extracts a number of feature maps equal to the number of input channels; the point convolution uses 1x1 convolution kernels to extract features, and the number of convolution kernels is set to the number of feature maps to be obtained.
[0089] Optionally, the fault diagnosis model further includes a classification module, which is used for classifying and outputting fault diagnosis results, and the classification module includes an adaptive average pooling module and a convolution module.
[0090] It is understood that the fault diagnosis system based on multi-scale residual attention network provided in this embodiment of the invention corresponds to the fault diagnosis method based on multi-scale residual attention network described above. The explanation, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the fault diagnosis method based on multi-scale residual attention network, and will not be repeated here.
[0091] In summary, compared with existing technologies, it has the following beneficial effects:
[0092] 1. This invention constructs a fault diagnosis model based on a multi-scale residual attention network and uses the trained fault diagnosis model for equipment fault diagnosis. The multi-scale residual attention network includes sequentially connected wide convolutional modules, multi-scale convolutional modules, and residual attention modules. The fault diagnosis method of this invention can extract features implicit in equipment signal data from multiple scales, while also suppressing noise interference in the equipment signal data to a certain extent. Compared with traditional fault diagnosis techniques, this improves the accuracy and robustness of fault diagnosis.
[0093] 2. This invention utilizes a multi-scale convolution module to extract hidden features from collected equipment-related data at multiple scales. This addresses the problem of insufficient feature extraction in traditional fault diagnosis models that extract fault signal features from only a single scale, resulting in low accuracy and poor robustness in fault diagnosis. The multi-scale module of this invention can significantly improve the classification accuracy of the fault diagnosis model, thereby enhancing the overall accuracy of fault diagnosis.
[0094] 3. The residual attention module in this invention can effectively focus on noise signals in the vibration signals of equipment such as ship propulsion equipment, thereby giving greater weight to important signals and less weight to noise signals, effectively improving the model's noise resistance and generalization ability.
[0095] 4. The fault diagnosis model based on multi-scale residual attention network proposed in this invention has far fewer parameters than some classic networks. While effectively improving the classification accuracy of the model, it greatly reduces training cost and time cost, saving resources and time.
[0096] 5. In the case of limited vibration signal data, this invention expands the number of training set samples based on the sliding window technique, thereby expanding the dataset and improving the classification accuracy of fault diagnosis through data augmentation.
[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A fault diagnosis method based on a multi-scale residual attention network, characterized in that, The method includes: Acquire signal data from the device to be diagnosed, and preprocess the signal data; A fault diagnosis model based on a multi-scale residual attention network is constructed. The multi-scale residual attention network includes a wide convolutional module, a multi-scale convolutional module, and a residual attention module connected in sequence. The wide convolutional module replaces the small convolutional kernels of the original neural network with dilated convolutions to extract long-term feature information of the signal data. The multi-scale convolutional module uses a soft thresholding function as a branch after convolving the original input on the original Res2block module, and replaces the regular small convolutions of each branch in the original Res2block module with two depthwise separable convolutions. The residual attention module adds a hybrid pooling strategy to the channel attention mechanism of the CBAM attention mechanism, and uses the dot product result of the channel attention mechanism and the spatial attention mechanism with the added hybrid pooling strategy as the second residual connection branch, while the input of the residual attention module is directly used as the first residual connection branch. Fault diagnosis is achieved based on the trained fault diagnosis model.
2. The method as described in claim 1, characterized in that, The preprocessing includes data augmentation of the signal data using a sliding window technique.
3. The method as described in claim 1, characterized in that, The hybrid pooling strategy involves weighting the global max pooling and global min pooling operations.
4. The method as described in claim 1, characterized in that, The depthwise separable convolution includes channel convolution and point convolution; the channel convolution determines the number of convolution kernels based on the dimension of the input channels and extracts a number of feature maps equal to the number of input channels; the point convolution uses 1x1 convolution kernels to extract features, and the number of convolution kernels is set to the number of feature maps to be obtained.
5. The method as described in claim 1, characterized in that, The fault diagnosis model further includes a classification module, which is used for classifying and outputting fault diagnosis results, and the classification module includes an adaptive average pooling module and a convolution module.
6. A fault diagnosis system based on a multi-scale residual attention network, characterized in that, The system includes: The data acquisition and processing module is used to acquire signal data of the device to be diagnosed and to preprocess the signal data. A fault diagnosis model acquisition module is used to construct a fault diagnosis model based on a multi-scale residual attention network. The multi-scale residual attention network includes a wide convolutional module, a multi-scale convolutional module, and a residual attention module connected in sequence. The wide convolutional module uses dilated convolutions to replace the small convolutional kernels of the original neural network to extract long-term feature information from the signal data. The multi-scale convolutional module uses a soft thresholding function as a branch after convolving the original input on the original Res2block module, and replaces the regular small convolutions of each branch in the original Res2block module with two depthwise separable convolutions. The residual attention module adds a hybrid pooling strategy to the channel attention mechanism of the CBAM attention mechanism, and uses the dot product result of the channel attention mechanism and the spatial attention mechanism with the added hybrid pooling strategy as the second residual connection branch, while directly using the input of the residual attention module as the first residual connection branch. The fault diagnosis module is used to perform fault diagnosis based on the trained fault diagnosis model.
7. The system as described in claim 6, characterized in that, The data acquisition and processing module performs preprocessing operations, including using sliding window technology to enhance the signal data.
8. The system as described in claim 6, characterized in that, The hybrid pooling strategy involves weighting the global max pooling and global min pooling operations.
9. The system as described in claim 6, characterized in that, The depthwise separable convolution includes channel convolution and point convolution; the channel convolution determines the number of convolution kernels based on the dimension of the input channels and extracts a number of feature maps equal to the number of input channels; the point convolution uses 1x1 convolution kernels to extract features, and the number of convolution kernels is set to the number of feature maps to be obtained.
10. The system as described in claim 6, characterized in that, The fault diagnosis model further includes a classification module, which is used for classifying and outputting fault diagnosis results, and the classification module includes an adaptive average pooling module and a convolution module.