MODWPT-based bearing fault diagnosis method and apparatus, and electronic device

Abstract

The invention provides a bearing fault diagnosis method and device based on MODWPT and electronic equipment. Comprising the following steps: acquiring a bearing signal, and decomposing the bearing signal into bearing time-frequency information; a feature extraction network model is constructed, the feature extraction network model comprises a ConvNext model, a CBAM attention module, a GGL graph generation layer, a GCN graph convolutional network and a feature extraction network, and the ConvNext model is used for extracting bearing fault feature information; the CBAM attention module is used for integrating and screening the bearing fault feature information; the GGL graph generation layer is used for learning a data structure to construct a feature instance graph; the GCN graph convolutional network is used for eliminating the influence of information one-sidedness and focus flattening; and performing feature extraction on the bearing time-frequency information by using the feature extraction network model to obtain bearing feature information, and determining a bearing fault according to the bearing feature information. The problem that bearing fault features are difficult to effectively extract from complex noise in the prior art is solved.

Description

Technical Field

[0001] This application relates to the technical field of bearing fault diagnosis. Specifically, it relates to a bearing fault diagnosis method, device, computer-readable storage medium, and electronic device based on MODWPT. Background Art

[0002] Machine fault diagnosis is of great significance in modern industry. Faults in key machine components can lead to equipment failures, causing serious casualties and economic losses. Bearings are precision standard mechanical components widely used in various machines. However, they are vulnerable to vibration, shock, erosion, and wear under harsh conditions, as well as other external influences, such as long-term exposure to heavy loads and alternating loads, high temperatures, and high speeds, making them one of the most vulnerable components in equipment. Therefore, it is a hot topic in the current research field to timely and effectively identify the bearing characteristics of rolling equipment before accidents occur to ensure the normal operation of the equipment.

[0003] In the past few years, fault diagnosis based on deep learning has achieved remarkable success, but there are still two main issues to consider: data distribution and data availability. The first is that deep learning-based models must follow the same distribution for training data and test data. However, in engineering practical application scenarios, according to different industrial production requirements, the load and speed of rolling bearings during operation cannot be kept consistent, and even show significant differences. This will lead to a significant decline in the accuracy of fault diagnosis based on deep learning models. The second is that the performance of deep learning models is mainly adjusted by using labeled data to adjust the model weights of the network. However, labeling the collected data one by one will waste a lot of time and effort. At the same time, in actual fault diagnosis tasks, the noise components in the equipment operation environment are quite complex and may have different effects on the model's ability to extract relevant features. This makes it difficult for existing deep learning-based methods to effectively extract bearing fault features from complex noise. Summary of the Invention

[0004] The main purpose of this application is to provide a bearing fault diagnosis method, device, computer-readable storage medium, and electronic device based on MODWPT, so as to at least solve the problem that existing deep learning-based methods are difficult to effectively extract bearing fault features from complex noise.

[0005] To achieve the above object, according to one aspect of the present application, a bearing fault diagnosis method based on MODWPT is provided, including: obtaining a bearing signal, and using the maximum overlap discrete wavelet packet transform (MODWPT) to decompose the bearing signal into bearing time-frequency information at different resolutions; constructing a feature extraction network model, wherein the feature extraction network model includes a ConvNext model, a CBAM attention module, a GGL graph generation layer, a GCN graph convolutional network, and a feature extraction network. The ConvNext model is used to extract bearing fault feature information from the bearing time-frequency information; the CBAM attention module is used to integrate and screen the bearing fault feature information to obtain target feature information; the GGL graph generation layer is used to learn the data structure from the target feature information to construct a feature instance graph; the GCN graph convolutional network is used to model the constructed feature instance graph to eliminate the influence that key feature information is vulnerable to information one-sidedness and focus flattening; the feature extraction network is used to perform feature extraction processing on the result of the GCN graph convolutional network modeling; using the feature extraction network model to perform feature extraction processing on the bearing time-frequency information to obtain bearing feature information, and determining bearing faults according to the bearing feature information.

[0006] Optionally, the CBAM attention module includes a channel attention module and a spatial attention module. The channel attention module squeezes the obtained bearing fault feature information spatially, and at the same time sums the information features spatially element by element through max pooling and average pooling, so as to obtain the weight feature of channel attention. The spatial attention module uses max pooling and average pooling on the obtained bearing fault feature information in the channel dimension to obtain two different information feature descriptions, and then stacks the two information feature descriptions together to obtain the weight feature of spatial attention.

[0007] Optionally, using the maximum overlap discrete wavelet packet transform (MODWPT) to decompose the bearing signal into bearing time-frequency information at different resolutions includes: using the first formula: Determine the MODWPT decomposition coefficient level of the bearing signal, where h(t) is a low-pass filter, g(t) is a high-pass filter, is the z-th decomposition coefficient of the j-th level, z = 0, 1, 2…, 2 j-1 , t is the continuous time change parameter of the previous moment, and k is the continuous time change parameter of the next moment; decompose the bearing signal into the bearing time-frequency information at different resolutions according to the MODWPT decomposition coefficient level.

[0008] Optionally, determining a bearing fault according to the bearing feature information includes: constructing a bearing fault diagnosis model, wherein the domain adaptation adversarial learning framework of the bearing fault diagnosis model combines an entropy-conditional domain adversarial network ECDAN with an improved joint discriminant probability maximum mean discrepancy JMMD strategy, and adjusts network parameters of the entropy-conditional domain adversarial network through a gradient reversal layer during the backpropagation process. Using the bearing fault diagnosis model to determine the bearing fault according to the bearing feature information.

[0009] Optionally, according to the second formula: Determine the JMMD strategy L JMMD (P, Q), where P is the source domain distribution, Q is the target domain distribution, |L| is the number of layers in the corresponding set, and H l Is the l-th layer of the reproducing kernel Hilbert space, Is the feature map of the tensor product in the Hilbert space; z sl The activation generated by the source domain at the l-th layer; z tl Is the activation generated by the target domain at the l-th layer.

[0010] Optionally, before using the bearing fault diagnosis model to determine the bearing fault according to the bearing feature information, the method further includes: training the bearing fault diagnosis model with sample data to obtain a trained bearing fault diagnosis model, where the sample data is divided into source domain bearing fault samples with fault labels and target domain bearing fault samples without fault labels, and determining whether the bearing fault diagnosis model is trained completed according to minimizing the classification loss and maximizing the discriminant loss.

[0011] Optionally, after determining the bearing fault according to the bearing feature information, the method further includes: determining a solution to the bearing fault according to the bearing fault, and uploading the bearing fault and the solution to the user terminal.

[0012] According to another aspect of the present application, a bearing fault diagnosis device based on MODWPT is provided, including: a decomposition unit, configured to obtain a bearing signal and decompose the bearing signal into bearing time-frequency information at different resolutions by using maximum overlap discrete wavelet packet transform (MODWPT); a construction unit, configured to construct a feature extraction network model, wherein the feature extraction network model includes a ConvNext model, a CBAM attention module, a GGL graph generation layer, a GCN graph convolutional network, and a feature extraction network. The ConvNext model is used to extract bearing fault feature information from the bearing time-frequency information; the CBAM attention module is used to integrate and screen the bearing fault feature information to obtain target feature information; the GGL graph generation layer is used to learn the data structure from the target feature information to construct a feature instance graph; the GCN graph convolutional network is used to model the constructed feature instance graph to eliminate the influence of key feature information being vulnerable to information one-sidedness and focus flattening; the feature extraction network is used to perform feature extraction processing on the result of the modeling processing of the GCN graph convolutional network; a determination unit, configured to perform feature extraction processing on the bearing time-frequency information by using the feature extraction network model to obtain bearing feature information, and determine a bearing fault according to the bearing feature information.

[0013] According to still another aspect of the present application, a computer-readable storage medium is provided. The computer-readable storage medium includes a stored program, wherein when the program runs, it controls the device where the computer-readable storage medium is located to execute any one of the above-mentioned bearing fault diagnosis methods based on MODWPT.

[0014] According to yet another aspect of the present application, an electronic device is provided, including: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors. The one or more programs include those for executing any one of the above-mentioned bearing fault diagnosis methods based on MODWPT.

[0015] Applying the technical solution of the present application, bearing signals are acquired, and the maximum overlap discrete wavelet packet transform (MODWPT) is used to decompose the bearing signals into bearing time-frequency information at different resolutions; a feature extraction network model is constructed. The feature extraction network model includes a ConvNext model, a CBAM attention module, a GGL graph generation layer, a GCN graph convolutional network, and a feature extraction network. The ConvNext model is used to extract bearing fault feature information from the bearing time-frequency information; the CBAM attention module is used to integrate and screen the bearing fault feature information to obtain target feature information; the GGL graph generation layer is used to learn the data structure from the target feature information to construct a feature instance graph; the GCN graph convolutional network is used to model the constructed feature instance graph to eliminate the influence of key feature information being vulnerable to information one-sidedness and focus flattening; the feature extraction network is used to perform feature extraction processing on the result of the GCN graph convolutional network modeling; the feature extraction network model is used to perform feature extraction processing on the bearing time-frequency information to obtain bearing feature information, and the bearing fault is determined according to the bearing feature information. By using MODWPT to decompose the bearing signals into time-frequency information at different resolutions; secondly, using ConvNext ensures smoother network gradients and accelerated convergence, guaranteeing that the network can quickly extract the key fault features of the bearing. And by adding an improved attention mechanism to the feature extraction network, the channel attention mechanism and the spatial attention mechanism are used to integrate and screen the rich fault feature information at different scales, and weights are assigned to the integrated and screened bearing information features. The more critical the information, the greater the weight assigned; then a graph generation layer is proposed to learn the data structure from the extracted ConvNext features, and an instance graph is constructed by mining the relationships between the sample structure features. The graph convolutional network (GCN) can model the structural information propagated along the weighted edges of the graph, eliminating the influence of key information features being vulnerable to information one-sidedness and focus flattening, ensuring that the feature extraction network can comprehensively extract key fault information and solving the problem that it is difficult for the prior art to effectively extract bearing fault features from complex noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings forming a part of this application are used to provide a further understanding of the application. The schematic embodiments and descriptions thereof of the application are used to explain the application and do not constitute an improper limitation of the application. In the drawings:

[0017] Figure 1 The hardware structure block diagram of a mobile terminal for executing a bearing fault diagnosis method based on MODWPT provided in an embodiment of the present application is shown;

[0018] Figure 2 The flowchart of a bearing fault diagnosis method based on MODWPT provided in an embodiment of the present application is shown;

[0019] Figure 3 Schematic diagram showing the outer ring bearing fault signal provided according to an embodiment of the present application;

[0020] Figure 4 Schematic diagram showing the bearing fault signals of each node of MODWPT provided according to an embodiment of the present application;

[0021] Figure 5 Schematic diagram showing the ConvNext network model provided according to an embodiment of the present application;

[0022] Figure 6 Schematic diagram showing the CBAM attention module provided according to an embodiment of the present application;

[0023] Figure 7 Schematic diagram showing the MRF-GCN network structure parameters provided according to an embodiment of the present application;

[0024] Figure 8 Schematic diagram showing the CDAN structure provided according to an embodiment of the present application;

[0025] Figure 9 Schematic diagram showing the MWCGCNECDJD fault condition migration structure provided according to an embodiment of the present application;

[0026] Figure 10a Schematic diagram showing the fault classification result of the MWGCNECDJD model provided according to an embodiment of the present application;

[0027] Figure 10b Schematic diagram showing the fault classification result of the MWCGCNJD model provided according to an embodiment of the present application;

[0028] Figure 10c Schematic diagram showing the fault classification result of the MWCGCNECDJD model provided according to an embodiment of the present application;

[0029] Figure 10d Schematic diagram showing the fault classification result of the MWCGCNECDJD model provided according to an embodiment of the present application;

[0030] Figure 11 Schematic diagram showing the MWCGCNECDJD and other model condition migration diagnosis structure provided according to an embodiment of the present application;

[0031] Figure 12 Block diagram showing the structure of a bearing fault diagnosis device based on MODWPT provided according to an embodiment of the present application.

[0032] Among them, the above-mentioned drawings include the following reference numerals:

[0033] 102. Processor; 104. Memory; 106. Transmission device; 108. Input / output device. Detailed implementation manners

[0034] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0035] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0036] It should be noted that the terms "first", "second", etc. in the description and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so as to describe the embodiments of the present application here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those clearly listed steps or units, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0037] For the convenience of description, some nouns or terms related to the embodiments of the present application are described below:

[0038] MODWPT: Maximum Overlap Discrete Wavelet Packet Transform;

[0039] JMMD: Joint Discriminant Probability Maximum Mean Discrepancy;

[0040] MWCGCN: Feature extraction network combining MODWPT, graph convolution, ConvNext and CBAM attention module;

[0041] CDAN: Domain Adversarial Network. The idea of CDAN comes from the Generative Adversarial Network, which ensures that the neural network can extract domain-insensitive feature information through adversarial training;

[0042] ECDAN: Improved adversarial neural network with entropy condition added.

[0043] As introduced in the background art, it is difficult for the prior art to effectively extract bearing fault features from complex noises. To solve the problem that it is difficult for the prior art to effectively extract bearing fault features from complex noises, embodiments of the present application provide a bearing fault diagnosis method, device, computer-readable storage medium and electronic device based on MODWPT.

[0044] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0045] The method embodiments provided in the embodiments of the present application can be executed on a mobile terminal, a computer terminal or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 is a hardware structure block diagram of a mobile terminal of a bearing fault diagnosis method based on MODWPT according to an embodiment of the present invention. As Figure 1 shown, the mobile terminal may include one or more ( Figure 1 only one is shown in the figure) processors 102 (the processor 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. Among them, the above-mentioned mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those of ordinary skill in the art can understand that Figure 1 the structure shown is only schematic and does not limit the structure of the above-mentioned mobile terminal. For example, the mobile terminal may further include more or fewer components than Figure 1 shown in the figure, or have a different configuration from Figure 1 shown in the figure.

[0046] The memory 104 can be used to store computer programs, such as software programs and modules of application software, such as the computer program corresponding to the bearing fault diagnosis method based on MODWPT in the embodiments of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, implements the above-mentioned method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely disposed relative to the processor 102, and these remote memories can be connected to the mobile terminal through a network. Examples of the above-mentioned network include but are not limited to the Internet, intranet, local area network, mobile communication network, and combinations thereof. The transmission device 106 is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by a communication provider of the mobile terminal. In one instance, the transmission device 106 includes a network adapter (Network Interface Controller, abbreviated as NIC), which can be connected to other network devices through a base station and thus can communicate with the Internet. In one instance, the transmission device 106 may be a radio frequency (Radio Frequency, abbreviated as RF) module, which is used to communicate with the Internet wirelessly.

[0047] In this embodiment, a bearing fault diagnosis method based on MODWPT running on a mobile terminal, a computer terminal, or a similar computing device is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0048] Figure 2 It is a flowchart of the bearing fault diagnosis method based on MODWPT according to the embodiments of the present application. As Figure 2 shown, the method includes the following steps:

[0049] Step S201, acquire a bearing signal, and decompose the bearing signal into bearing time-frequency information at different resolutions by using the maximum overlap discrete wavelet packet transform MODWPT;

[0050] Specifically, MODWPT is a wavelet transform that combines the maximum overlap discrete intensity and a wavelet packet grouping method. Different from traditional wavelet decomposition methods, MODWPT can effectively eliminate the phenomenon of frequency band crossing, thus obtaining a more accurate wavelet decomposition. In addition, MODWPT allows for flexible adjustment of the number of groups and wavelets in each group according to signal characteristics, making it a general tool for different applications. Another advantage of MODWPT over the discrete wavelet packet transform (DWPT) is that it does not involve downsampling by a factor of 2, thus generating a uniform frequency output band. Generally speaking, MODWPT is a valuable tool in signal processing and fault diagnosis applications, providing better performance compared to traditional wavelet decomposition methods.

[0051] Step S202: Construct a feature extraction network model. Specifically, the feature extraction network model includes a ConvNext model, a CBAM attention module, a GGL graph generation layer, a GCN graph convolutional network, and a feature extraction network. The ConvNext model is used to extract bearing fault feature information from the bearing time-frequency information; the CBAM attention module is used to integrate and screen the bearing fault feature information to obtain target feature information; the GGL graph generation layer is used to learn the data structure from the target feature information to construct a feature instance graph; the GCN graph convolutional network is used to model the constructed feature instance graph to eliminate the influence of key feature information being vulnerable to information one-sidedness and focus flattening; the feature extraction network is used to perform feature extraction processing on the result of the GCN graph convolutional network's modeling processing.

[0052] Specifically, this model uses ConvNext to ensure smoother network gradients and network convergence speed, guaranteeing that the network can quickly extract key bearing fault features. Secondly, by adding an improved attention mechanism to the feature extraction network, the channel attention mechanism and the spatial attention mechanism are used to integrate and screen rich fault feature information at different scales, and weights are assigned to the integrated and screened bearing information features. The more critical the information, the greater the weight assigned. Then, a graph generation layer (GGL) is proposed to learn the data structure from the extracted ConvNext features. Finally, an instance graph is constructed by mining the relationships between sample structure features. The graph convolutional network (GCN) can model the structural information propagated along the graph weighted edges, eliminating the influence of key information features being vulnerable to information one-sidedness and focus flattening, ensuring that the feature extraction network can comprehensively extract key fault information and enhancing the overall fault feature extraction ability of the network.

[0053] Step S203: Use the above feature extraction network model to perform feature extraction processing on the bearing time-frequency information to obtain bearing feature information, and determine the bearing fault according to the bearing feature information.

[0054] Through this embodiment, the bearing signal is decomposed into time-frequency information at different resolutions by using MODWPT; secondly, ConvNext is used to ensure smoother network gradients and accelerated convergence, guaranteeing that the network can quickly extract the key fault features of the bearing. By adding an improved attention mechanism to the feature extraction network, the channel attention mechanism and the spatial attention mechanism are used to integrate and screen the rich fault feature information at different scales, and weights are assigned to the integrated and screened bearing information features. The more critical the information, the greater the weight assigned; then a graph generation layer is proposed to learn the data structure from the extracted ConvNext features, and an instance graph is constructed by mining the relationships between the sample structure features. The graph convolutional network (GCN) can model the structural information propagated along the weighted edges of the graph, eliminating the influence that the key information features are vulnerable to information one-sidedness and focus flattening, ensuring that the feature extraction network can comprehensively extract the key fault information and solving the problem that it is difficult to effectively extract the bearing fault features from complex noise in the prior art.

[0055] In the specific implementation process, the above-mentioned CBAM attention module includes a channel attention module and a spatial attention module. The above-mentioned channel attention module squeezes the obtained bearing fault feature information spatially, and at the same time adds the information features spatially element by element through max pooling and average pooling, so as to obtain the weight feature of the channel attention. The above-mentioned spatial attention module uses max pooling and average pooling in the channel dimension to obtain two different information feature descriptions of the obtained bearing fault feature information, and then stacks the two above-mentioned information feature descriptions together to obtain the weight feature of the spatial attention.

[0056] The method CBAM (Convolutional Block Attention Module) represents a convolutional attention mechanism module, which is a module that combines spatial and channel attention mechanisms. The CBAM attention mechanism module consists of a channel attention module and a spatial attention module. The channel attention module first squeezes the obtained feature information spatially, and at the same time adds the information features spatially element by element through max pooling and average pooling, so as to obtain the weight feature of the channel attention. The spatial attention module is a supplement to the channel attention module, and the spatial attention module focuses on the position of the information features. First, max pooling and average pooling are used in the channel dimension to obtain two different information feature descriptions, and then they are stacked together. After passing through a convolutional layer, weights are assigned to the obtained features. The more critical the information, the greater the weight. Finally, the key information features are obtained by backtracking the receptive field to the original information.

[0057] Specifically, the maximum overlap discrete wavelet packet transform (MODWPT) is used to decompose the bearing signal into bearing time-frequency information at different resolutions, including: using the first formula: Determine the MODWPT decomposition coefficient level of the above bearing signal, where h(t) is a low-pass filter, g(t) is a high-pass filter, is the z-th decomposition coefficient of the j-th level, z = 0, 1, 2…, 2 j-1 , t is the continuous time change parameter at the previous moment, k is the continuous time change parameter at the next moment; decompose the above bearing signal into the above bearing time-frequency information at different resolutions according to the above MODWPT decomposition coefficient level.

[0058] More specifically, determining the bearing fault according to the above bearing characteristic information includes: constructing a bearing fault diagnosis model, where the domain adaptation adversarial learning framework of the above bearing fault diagnosis model combines the entropy-conditional domain adversarial network (ECDAN) with an improved joint discriminant probability maximum mean discrepancy (JMMD) strategy, and adjusts the network parameters of the above entropy-conditional domain adversarial network through a gradient reversal layer during the backpropagation process. Use the above bearing fault diagnosis model to determine the above bearing fault according to the above bearing characteristic information.

[0059] This method uses a deep entropy-conditional domain adversarial multi-scale neural network (ECDAN) combined with the JMMD method to constrain the distributions of source domain and target domain data in the high-dimensional kernel space. The distribution difference between the joint distributions of source domain data and target domain data is minimized using the JMMD distance, and the conditional distributions are aligned using the entropy-conditional domain adversarial loss. In addition, an adaptive factor can dynamically measure the relative importance of the two distributions to adapt to fault diagnosis under different cross-domain tasks.

[0060] Furthermore, according to the second formula: Determine the above JMMD strategy L JMMD (P, Q), where P is the source domain distribution, Q is the target domain distribution, |L| is the number of layers in the corresponding set, H l is the l-th layer of the reproducing kernel Hilbert space, is the feature map of the tensor product in the Hilbert space; z sl is the activation generated by the source domain at the l-th layer; z tl is the activation generated by the target domain at the l-th layer.

[0061] The previous work of this method in domain adaptation mainly focused on realizing the marginal domain fusion of mapped features, and the Maximum Mean Discrepancy (MMD) was very effective in this regard. However, the differences in joint distributions exist in multi-layer feature extraction and predicted labels. As the model is continuously optimized, the pseudo-labels of the target domain dataset can also be predicted. The key to achieving this goal lies in the assumption that most target domain samples are similar to source domain samples and they have the same labels. Based on reliable pseudo-labels and network optimization techniques, this algorithm can assist other poor target domain samples. Therefore, on the basis of MMD, the Joint Maximum and Mean Discrepancy based on Predicted Labels (JMMD) was explored.

[0062] Furthermore, before determining the bearing fault according to the above bearing feature information by using the above bearing fault diagnosis model, the method further includes: training the above bearing fault diagnosis model with sample data to obtain a trained bearing fault diagnosis model, where the above sample data is divided into source domain bearing fault samples with fault labels and target domain bearing fault samples without fault labels, and determining whether the above bearing fault diagnosis model is trained based on minimizing the classification loss and maximizing the discriminative loss.

[0063] The sample data of this method is divided into source domain bearing fault samples (with labels) and target domain bearing fault samples (without given labels) carrying fault samples. In addition, the following assumptions are made for the source domain and target domain of the bearing fault sample set: (1) The feature spaces are the same; (2) The label spaces are the same; (3) The probability distributions are different. The cross-condition bearing fault diagnosis method based on CDAN aims to help target domain samples obtain corresponding labels. It mainly determines the labels of target domain bearing faults by using source domain bearing sample data and labels. CDAN mainly includes a feature extractor F, a fault classifier G, and a domain discriminator D. F extracts features of different fault sizes in the source domain and target domain containing the fault sample set. G classifies the features of different network layer sizes in the source domain extracted by F and obtains the classification loss by calculating the cross-entropy loss between the predicted labels and the true labels. D determines whether the features extracted by F come from the source domain or the target domain and calculates their binary cross-entropy loss to obtain the discriminative loss.

[0064] Among them, the process of using the bearing fault diagnosis model to identify bearing faults is as follows: First, when the two-domain data enters the ECDAN feature extraction network, the JMMD algorithm calculates the feature data between different categories in the source domain and the target domain in the ECDAN feature extraction network, and obtains the distance weight factor between the two-domain feature data in the last fully connected layer, realizing the mapping of the ECDAN feature data in the feature space; Second, minimize the loss function of the JMMD for calculating the two-domain feature data, and at the same time use this part of the loss function as a part of the backpropagation to optimize the feature extraction network in the ECDAN, improving the maximum inter-domain transferability and inter-class discriminability of the two-domain data in the ECDAN network, reducing the probability distribution difference of the two-domain deep features, so that the deep features have better cross-domain invariance and fault state discrimination ability; Then, the ECDAN network makes the two-domain feature data generate fierce confrontation through the domain discriminator, and further aligns the distribution distance between the two domains in this way; Finally, when the ECDAN network performs backpropagation, it jointly optimizes the ECDAN network parameters by combining the loss function of the JMMD, reduces the difference between the two-domain data distributions, realizes the transfer of the source domain data to train the target domain data, and completes the final fault classification.

[0065] Specifically, after determining the bearing fault according to the above bearing feature information, the above method further includes: determining a solution to the above bearing fault according to the above bearing fault, and uploading the above bearing fault and the above solution to the user terminal.

[0066] In order to enable those skilled in the art to more clearly understand the technical solution of the present application, the implementation process of the bearing fault diagnosis method based on MODWPT of the present application will be described in detail below with specific embodiments.

[0067] This embodiment relates to a specific bearing fault diagnosis method based on MODWPT. First, the bearing signal is decomposed into time-frequency information at different resolutions using MODWPT. Secondly, ConvNext is used to ensure smoother network gradients and accelerated convergence, guaranteeing that the network can quickly extract key bearing fault features. By adding an improved attention mechanism to the feature extraction network, the channel attention mechanism and spatial attention mechanism are used to integrate and screen rich fault feature information at different scales, and weights are assigned to the integrated and screened bearing information features. The more critical the information, the greater the weight assigned. Then, a graph generation layer (GGL) is proposed to learn the data structure from the extracted ConvNext features, and an instance graph is constructed by mining the relationships between sample structure features. The graph convolutional network (GCN) can model the structural information propagated along the weighted edges of the graph, eliminating the influence of key information features being vulnerable to information one-sidedness and focus flattening, ensuring that the feature extraction network can comprehensively extract key fault information. Finally, JMMD is combined with the ECDAN network to maximize inter-domain transferability and inter-class discriminability, optimize the deep feature extraction network, reduce the probability distribution difference of the deep features in the two domains, and make the deep features have better cross-domain invariance and fault state discrimination ability.

[0068] The technical content of the bearing fault diagnosis method based on the attention graph convolutional entropy conditional domain adversarial joint network specifically based on MODWPT is as follows:

[0069] 1. Maximal overlap discrete wavelet packet transform (MODWPT):

[0070] MODWPT is a wavelet transform that combines maximal overlap discrete intensity and wavelet packet grouping methods. Different from traditional wavelet decomposition methods, MODWPT can effectively eliminate the phenomenon of frequency band crossing, thus obtaining a more accurate wavelet decomposition. In addition, MODWPT allows for flexible adjustment of the number of groups and the number of wavelets in each group according to signal characteristics, making it a general tool for different applications. Another advantage of MODWPT over the discrete wavelet packet transform (DWPT) is that it does not involve downsampling by a factor of 2, thus producing a uniform frequency output band. Generally speaking, MODWPT is a valuable tool in signal processing and fault diagnosis applications, providing better performance compared to traditional wavelet decomposition methods. The MODWPT decomposition coefficient level is given by the following formula:

[0071]

[0072] where h(t) and g(t) are the low-pass and high-pass filters respectively, is the z coefficient of the jth level, z = 0, 1, 2…, 2 j-1 .

[0073] All coefficients j in the horizontal direction can be expressed as the following matrix:

[0074]

[0075] Generally, combining traditional deep learning with signal decomposition methods can improve the performance of the model. Based on this, we combine signal decomposition methods with deep learning to improve the performance of the network model.

[0076] Taking the fault of the outer ring of the bearing as an example, the signal samples of the outer ring of the original vibration signal are as Figure 3 shown, and the bearing fault signals of each node after MODWPT decomposition are as Figure 4 shown.

[0077] 2. ConvNext Graph Convolutional Spatiotemporal Attention Deep Feature Extraction Network:

[0078] Many fault feature extraction network models often pay little attention to the extracted fault features themselves, which makes the features vulnerable to defects such as information one-sidedness and focus flattening, resulting in a bottleneck in the improvement of model performance. In response to this, the present invention designs a MWCGCN feature extraction network model. This model uses ConvNext to ensure smoother network gradients and network convergence speed, ensuring that the network can quickly extract key bearing fault features. Secondly, by adding an improved attention mechanism to the feature extraction network, the channel attention mechanism and the spatial attention mechanism are used to integrate and screen rich fault feature information at different scales, and weights are assigned to the integrated and screened bearing information features. The more critical the information, the greater the weight assigned; then a graph generation layer (GGL) is proposed to learn the data structure from the extracted ConvNext features, and finally an instance graph is constructed by mining the relationships between sample structure features. The graph convolutional network (GCN) can model the structural information propagated along the graph weighted edges, eliminating the influence of key information features being vulnerable to information one-sidedness and focus flattening, ensuring that the feature extraction network can comprehensively extract key fault information and improving the overall fault feature extraction ability of the network.

[0079] The ConvNext network has achieved enhancements in multiple aspects compared to traditional deep learning networks, including the overall structure, depth convolution, inverted bottleneck, large convolutional kernels, GELU activation function, and LN layer. Regarding the overall structure, the Stem layer of the ConvNext network uses convolutional kernels of the same size and a four-step convolutional operation similar to that of the Swin-transform. Regarding convolution, the ConvNext network adopts the principle of depth convolution design, separating the number of input and output channels to reduce the parameter size of the designed depth convolution, which is significantly smaller than that of traditional convolution. In addition, ConvNext adopts a bottleneck design similar to ResNet. The researchers obtained inspiration from the transform network model and designed the block module in ConvNext as an inverted bottleneck structure; in addition, ConvNext replaced the conventional BN layer with an LN layer and reduced the number of normalization layers, thus eliminating redundancy. The schematic diagram of the ConvNext network model is as shown in Figure 5 . In each convolutional block, the LN layer is located after the initial convolutional layer, and these improvements have enhanced the overall performance and efficiency of the ConvNext network.

[0080] CBAM (Convolutional Block Attention Module) represents a convolutional attention mechanism module, which is a module that combines spatial and channel attention mechanisms. Its schematic diagram is as shown in Figure 6 . As can be seen from the figure, the CBAM attention mechanism module consists of a channel attention module and a spatial attention module. The channel attention module first squeezes the obtained feature information spatially, and at the same time adds the spatial information features element-wise through max-pooling and average-pooling to obtain the weight features of channel attention. The channel attention feature representation is shown in Equation 3, where σ represents the sigmoid function, W 0 and W 1 are the weights of the MLP, and are the spatial information features of average-pooling and max-pooling.

[0081]

[0082] The spatial attention module is a complement to the channel attention module. The spatial attention module focuses on the position of information features. First, two different information feature descriptions are obtained by using max pooling and average pooling in the channel dimension, and then they are stacked together. After passing through a convolutional layer, weights are assigned to the obtained features, and the more critical the information, the greater the weight. Finally, the key information features are obtained by backtracking to the original information through the receptive field. The spatial attention feature representation is shown in Equation 4, where σ represents the sigmoid function and f represents the convolutional operation.

[0083]

[0084] Figure 7 It is a parameter diagram of the GCN network structure. GCN is a kind of non-Euclidean space data, which contains more information. In the context of large-scale unavailability of migration bearing data, if more fault information can be extracted from the graph data, a more reliable feature extraction network can be established. Assume the graph G(A, X), where A is its adjacency matrix and X is its node features. The Laplacian matrix of its symmetric normalized graph L = I N - D -1 / 2 AD -1 / 2 , where D is the diagonal matrix of A; I N is the identity matrix. The graph filter g θ = diag(θ) can smooth the input signal x ∈ R□ to obtain a new definition of graph convolution, and its expression is as follows:

[0085]

[0086] In the formula, θ represents the learnable parameter; U T x represents the Fourier transform of node x in the graph; U gθ represents mapping the convolution kernel to the frequency domain.

[0087] However, the graph convolution operation in Equation 5 depends on the eigenvalue decomposition of L, so the operation is very complex. To solve this problem, the Chebyshev polynomial of the eigenvalue diagonal matrix is used to approximate the convolution kernel, and a new convolution kernel expression 6 is obtained:

[0088]

[0089] In the formula, k is the order of the Chebyshev polynomial, which also determines the range of the node neighborhood; Λ is the eigenvalue of the Laplacian matrix. Through the redefined GCN operation, k can aggregate the jump distance to achieve graph smoothing. The standard GCN can only aggregate information within a fixed receptive field. To solve this problem, a multi-receptive field GCN (MRF-GCN) is used to obtain a powerful feature representation and embed the data structure information into the feature representation to enhance the feature representation of the network.

[0090] 3. Design of Diagnostic Model Based on Entropy-Conditional Domain-Adversarial Transfer

[0091] Although the proposed network model has achieved good results in the bearing dataset under the same working conditions, in engineering practice, the fault types are unknown, and only known data can be used for training to judge the faults of unknown equipment. However, this leads to a sharp decline in the accuracy of variable working condition faults. Therefore, the present invention provides a Multi-scale Wavelet Packet Transform-based Attention Graph Convolution Entropy-Conditional Domain-Adversarial Joint Network (MWCGCNECDJD).

[0092] The present invention uses a Deep Entropy-Conditional Domain-Adversarial Multi-scale Neural Network (ECDAN) combined with the JMMD method to constrain the distributions of source domain and target domain data in the high-dimensional kernel space. The JMMD distance is used to minimize the distribution difference between the joint distributions of source domain data and target domain data, and the entropy-conditional domain-adversarial loss is used to align the conditional distributions. In addition, an adaptive factor can dynamically measure the relative importance of the two distributions to adapt to fault diagnosis under different cross-domain tasks.

[0093] The structure diagram of the CDAN network is as shown in Figure 8 Figure [Figure number not provided in the original]. The CDAN network draws on the idea of the conditional generative adversarial network, connecting the features of the source domain and target domain as well as the labels of the source domain and target domain. By merging the features extracted by the feature extractor and the predictions of the classifier into a form of cross-covariance, the CDAN can improve the model's ability to handle multi-modal structures in the source domain and target domain.

[0094] The sample data is divided into source domain bearing fault samples (with labels) and target domain bearing fault samples carrying fault samples (without labels). And represent the samples and their corresponding labels in the source domain bearing fault samples, represent the samples in the target domain of the fault sample set. n s and n t respectively represent the total number of samples in the source domain and target domain. In addition, the following assumptions are made for the source domain and target domain of the bearing fault sample set: (1) The feature spaces are the same, (2) The label spaces are the same, and (3) The probability distributions are different. The cross-working-condition bearing fault diagnosis method based on CDAN aims to help the target domain samples obtain the corresponding labels mainly by using the source domain bearing sample data and labels to judge the labels of the target domain bearing faults. The CDAN mainly includes a feature extractor F, a fault classifier G, and a domain discriminator D. Their corresponding parameters are denoted as θ f , θ g and θ d。F extracts features of different fault sizes in the source domain and target domain containing the fault sample set. G classifies the features of different network layer sizes in the source domain extracted by F, and obtains the classification loss by calculating the cross-entropy loss between the predicted label and the true label. D determines whether the features extracted by F are from the source domain or the target domain, and calculates their binary cross-entropy loss to obtain the discriminative loss.

[0095] For the bearing fault diagnosis method based on CDAN, it is by minimizing the classification loss θ f ,θ g and then maximizing the discriminative loss θ d 。CDAN realizes the above process by inserting a Gradient Reversal Layer (GRL) in the middle of F and D. During the training process, GRL is equivalent to a constant mapping for forward propagation. For backpropagation, the gradient is reversed by multiplying by a negative parameter, which makes domain adversarial possible. The min-max game between the objective functions of the generator and discriminator of CDAN is represented as follows:

[0096]

[0097]

[0098] λ is the weight coefficient that balances G and D, L(.,.) is the cross-entropy loss, and h represents the joint variable of the features extracted by F and the prediction result of g.

[0099] In the process of using CDAN for cross-condition diagnosis of bearing faults, when migrating the bearing fault samples in the source domain to the target domain, there will inevitably be some fault samples that are difficult to migrate. To solve the uncertainty of bearing fault samples in prediction, the entropy condition can be used to quantify it. The main method is to assign a smaller weight to the bearing fault samples with higher entropy in prediction, and assign a larger weight to the bearing fault samples with lower entropy in prediction. CDAN with entropy condition is called ECDAN. The formula for calculating the entropy condition is as follows:

[0100]

[0101] Here C represents the total number of categories of bearing fault samples, and g c represents the probability that the fault sample belongs to category c.

[0102] Now, the objective function of ECDAN is as follows:

[0103]

[0104]

[0105] Previous work in domain adaptation mainly focused on achieving marginal domain fusion of mapped features, and the Maximum Mean Discrepancy (MMD) was very effective in this regard. However, the differences in joint distributions exist in multi-layer feature extraction and predicted labels. As the model is continuously optimized, pseudo-labels for the target domain dataset can also be predicted. The key to achieving this goal lies in the assumption that most target domain samples are similar to source domain samples and have the same labels. Based on reliable pseudo-labels and network optimization techniques, this algorithm can assist other poor target domain samples. Therefore, the present invention explores the Joint Maximum and Mean Discrepancy (JMMD) based on predicted labels on the basis of MMD. The JMMD expression is as follows:

[0106]

[0107] In the formula: P—the source domain distribution; Q—the target domain distribution; |L|—the number of layers in the corresponding set; H l —the l-th layer of the Reproducing Kernel Hilbert Space; —the feature mapping of the tensor product in the Hilbert space; z sl —the activation generated by the source domain at the l-th layer; z tl —the activation generated by the target domain at the l-th layer.

[0108] To ensure that the feature extraction network can fully extract the fault information of the two domains, the present invention improves the ECDAN network and designs an attention graph convolutional entropy conditional domain adversarial joint network based on MODWPT (MWCGCNECDJD), as Figure 9 shown.

[0109] Improvement 1: Use entropy conditioning in the CDAN network to solve the uncertainty of predicting bearing fault samples, assign smaller weights to bearing fault samples with higher entropy, and assign larger weights to bearing fault samples with lower entropy.

[0110] Improvement 2: Traditional metric distance methods ignore the distinguishability between different categories, resulting in insufficient classification performance of the model. To reduce the probability distribution difference of the deep features in the two domains and make the deep features have better cross-domain invariance and fault state distinguishability. The present invention adds the JMMD metric distance method to ECDAN to enhance the maximum inter-domain transferability and inter-class discriminability of the two-domain data in the ECDAN network and reduce the probability distribution difference of the two-domain deep features.

[0111] Improvement 3: Replace the fully connected network structure of the channel attention module with a one-dimensional convolutional structure, which can better aggregate the information between channels. At the same time, replace the convolutional kernel in the spatial attention module with a dilated convolution to increase the spatial receptive field to obtain a wider information feature.

[0112] Improvement 4: Replace the ECDAN feature extraction network with the MWCGCN network designed in the present invention to fully extract the time-frequency feature information of the two domains at different time scales.

[0113] Label classification loss of MWCGCNECDJD:

[0114]

[0115] In the formula: represents the source domain data; represents the source domain label.

[0116] L of MWCGCNECDJD d (θ f , θ d ) is expressed as follows:

[0117]

[0118] In the formula: d i represents the binary variable of the domain class; represents the target domain data.

[0119] Therefore, the total loss function of MWCGCNECDJD is expressed as shown in formula (15):

[0120] L = L c (θ f , θ c ) + λ 1 L JMMD + λ 2 L d (θ f , θ d ) (15);

[0121] In the formula, λ 1 represents the trade-off parameter of L JMMD ; λ 2 represents the trade-off parameter of L d (θ f , θ d ).

[0122] The specific implementation process of combining the ECDAN and JMMD distance measurement methods is as follows: First, when the data of the two domains enter the ECDAN feature extraction network, the JMMD algorithm calculates the feature data between different categories of the source domain and the target domain in the ECDAN feature extraction network, and obtains the distance weight factor between the feature data of the two domains in the last fully connected layer, realizing the mapping of the ECDAN feature data in the feature space; Second, minimize the loss function of the JMMD for calculating the feature data of the two domains, and at the same time use this part of the loss function as a part of the backpropagation to optimize the feature extraction network in the ECDAN, improving the maximum inter-domain transferability and inter-class discriminability of the two-domain data in the ECDAN network, reducing the probability distribution difference of the two-domain deep features, and making the deep features have better cross-domain invariance and fault state discrimination ability; Then, the ECDAN network makes the feature data of the two domains have a fierce confrontation through the domain discriminator, and further aligns the distribution distance between the two domains in this way; Finally, when the ECDAN network performs backpropagation, it jointly optimizes the ECDAN network parameters in combination with the loss function of the JMMD, reduces the difference between the distribution of the two-domain data, realizes the migration of the source-domain data to the target-domain data for training, and completes the final fault classification.

[0123] The specific embodiment for verifying the migration diagnosis effect of the bearing fault diagnosis method based on the MODWPT attention graph convolutional entropy conditional domain adversarial joint network between variable working conditions is as follows:

[0124] The experimental bearing model is SER205-16 of NSK. Laser was used to engrave three different degrees of damage of 0.1mm, 0.2mm, and 0.3mm on the inner ring, outer ring, and rolling elements of the bearing respectively. Three different acquisition speeds were set at a signal sampling frequency of 16kHz, which were 900r / min, 1200r / min, and 1800r / min respectively. According to the different speeds, three working condition data sets were set from low to high, denoted as data set A, data set B, and data set C. The collected data was divided into source domain data (with labels corresponding to bearing data) and target domain data (without labels corresponding to bearing data). The bearing data collected was divided into a total of 10 fault types according to the degree of damage to the inner ring, outer ring, and rolling element bearings by electric discharge machining, and the labels corresponding to the fault types were 0-9.

[0125] Comparative experiments were carried out on the experimental bench through ablation experiments:

[0126] MWGCNECDJD model: Compared with the MWCGCNECDJD model, the CBAM attention module was not added.

[0127] MWCGCNECD model: Compared with the MWCGCNECDJD model, the JMMD distance measurement was not added.

[0128] MWCGCNJD model: Compared with the MWCGCNECDJD model, only the JMMD metric distance is used, and the ECDAN adversarial network is not used.

[0129] In the first set of experiments, the experimental results of the MWCGCNECDJD model, the MWGCNECDJD, MWCGCNECD, and MWCGCNJD models are shown in Table 1. It can be seen from the table that:

[0130] (1) Compared with the MWCGECNCD model, the average classification accuracy of the proposed MWCGCNECD model has increased by 4.85%. This shows that adding the JMMD metric distance to the MWGCNCDJD network model can jointly align the data distributions of the source domain and the target domain by the ECDAN, reduce the difference in the data feature distributions between the two domains, and improve the overall fault diagnosis rate of the network model.

[0131] (2) Compared with the MWCGCNJD model, the average classification accuracy of the proposed MWCGCNECDJD model has increased by 5.61%. This shows that adding the domain discriminator network to the MWCGCN network model can reduce the difference in the feature distributions between the source domain and the target domain, make the feature category representations of the two domains more discriminative, and improve the domain adaptation ability of the deep learning network.

[0132] (3) Compared with the MWGCNECDJD model, although the average classification accuracy of the proposed MWCGCNECDJD model has only increased by 1.84%, this indicates that after the feature extraction network lacks the CBAM attention mechanism to weight the key information, it is vulnerable to the influence of information one-sidedness and focus flattening, resulting in the loss of key fault feature information of the bearing and insufficient feature extraction ability of the MWCGCN network.

[0133] Table 1 Comparison of the migration accuracy of models under various working conditions

[0134]

[0135]

[0136] To compare the deep feature extraction capabilities of the four models, the t-distributed stochastic neighbor embedding (t-SNE) algorithm and the confusion matrix, which are commonly used, are adopted to visualize the fault classification results of the MWGCNECDJD, MWCGCNECD, MWCGCNJD, and the proposed MWCGCNECDJD models under the B-C transfer task. The results are as Figure 10a 、 Figure 10b 、 Figure 10c 、 Figure 10dAs shown, it can be seen from the t-SNE diagram that different colors and different shapes represent different types of faults. It can be seen that these ten fault types are identified and clustered, but there may still be some misclassifications, which may be due to the complexity and similarity between different fault types. The proposed MWCGCNECDJD model can basically cluster the faults completely together. This indicates that the proposed method has strong feature extraction ability and can effectively capture and represent the unique patterns and features of each fault type. The experimental results further show that the MWCGCNECDJD model can better align the distributions between the two domains, reduce the distribution differences between the two domains, enhance the ability to integrate data information from the two domains, and improve the fault classification accuracy of the data from the two domains.

[0137] In the second group of experiments, in order to verify the diagnostic effectiveness of the proposed MMWCGCNECDJD model in a variable working condition environment, several excellent rolling bearing fault diagnosis methods were selected for comparison, including: TCA, JDA, and DAN. The fault classification accuracies of the four models in 6 working condition migration experiments were compared, and the experimental results are as Figure 11 shown. From Figure 11 it can be intuitively obtained that:

[0138] (1) Compared with the three network models of TCA, JDA, and DAN, the MWCGCNECDJD model has better stability.

[0139] (2) The MWCGCNECDJD model maintains a high fault classification accuracy in 6 migration tasks, and is far higher than the three network models of TCA, JDA, and DAN.

[0140] (3) In the variable working condition mode, compared with other migration network models, the MWCGCNECDJD model can not only minimize the distribution difference between the source domain data and the joint distribution of the target domain data by combining JMMD and ECDAN, but also improve the network's ability to learn domain-invariant features.

[0141] The embodiments of this application have the following advantages:

[0142] 1. A MWCGCNECDJD variable working condition deep feature extraction network model combining ECDAN and JMMD is designed, which has strong feature extraction ability, can effectively capture and represent the unique patterns and features of each fault type. At the same time, this model can better align the distributions between the two domains, reduce the distribution differences between the two domains, enhance the ability to integrate data information from the two domains, and improve the fault classification accuracy of the data from the two domains.

[0143] 2. Introduce the CBAM attention mechanism into the MWCGCNECDJD network, assign higher weights to the key information of the bearing damage mechanism, and at the same time suppress inefficient or invalid feature information, so that the network optimizes the weights of key information during the backpropagation process and improves the overall feature expression ability of the network model.

[0144] 3. Add a graph convolutional network to the network. The GCN can model the structural information propagated along the weighted edges of the graph, eliminate the influence of the key information features being vulnerable to information one-sidedness and focus flattening, and ensure that the feature extraction network can comprehensively extract key fault information;

[0145] 4. To solve the uncertainty of bearing fault samples in prediction, add an entropy condition to the CDAN network to quantify the sample weights. Assign smaller weights to bearing fault samples with higher entropy, and assign larger weights to bearing fault samples with lower entropy, thereby improving the separability of migrated fault samples.

[0146] The embodiment of the present application also provides a bearing fault diagnosis device based on MODWPT. It should be noted that the bearing fault diagnosis device based on MODWPT in the embodiment of the present application can be used to execute the bearing fault diagnosis method based on MODWPT provided by the embodiment of the present application. This device is used to implement the above embodiments and preferred implementation manners, and those that have been described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that can achieve a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0147] The following introduces the bearing fault diagnosis device based on MODWPT provided by the embodiment of the present application.

[0148] Figure 12 is a schematic diagram of the bearing fault diagnosis device based on MODWPT according to the embodiment of the present application. As Figure 12 shown, the device includes:

[0149] A decomposition unit 1201, configured to obtain a bearing signal, and decompose the bearing signal into bearing time-frequency information at different resolutions by using the maximum overlap discrete wavelet packet transform MODWPT;

[0150] The building unit 1202 is used to build a feature extraction network model. The feature extraction network model includes a ConvNext model, a CBAM attention module, a GGL graph generation layer, a GCN graph convolutional network, and a feature extraction network. The ConvNext model is used to extract bearing fault feature information from the bearing time-frequency information; the CBAM attention module is used to integrate and screen the bearing fault feature information to obtain target feature information; the GGL graph generation layer is used to learn the data structure from the target feature information to build a feature instance graph; the GCN graph convolutional network is used to model the built feature instance graph to eliminate the influence of key feature information being vulnerable to information one-sidedness and focus flattening; the feature extraction network is used to perform feature extraction processing on the result of the GCN graph convolutional network modeling.

[0151] The determination unit 1203 is used to perform feature extraction processing on the bearing time-frequency information by using the feature extraction network model to obtain bearing feature information, and determine the bearing fault according to the bearing feature information.

[0152] In this embodiment, the bearing signal is decomposed into time-frequency information at different resolutions by using MODWPT; secondly, ConvNext is used to ensure smoother network gradients and accelerated convergence, ensuring that the network quickly extracts key bearing fault features. And by adding an improved attention mechanism to the feature extraction network, the channel attention mechanism and the spatial attention mechanism are used to integrate and screen rich fault feature information at different scales, and weights are assigned to the integrated and screened bearing information features. The more critical the information, the greater the weight assigned. Then a graph generation layer is proposed to learn the data structure from the extracted ConvNext features, and an instance graph is constructed by mining the relationships between sample structure features. The graph convolutional network (GCN) can model the structural information propagated along the weighted edges of the graph, eliminating the influence of key information features being vulnerable to information one-sidedness and focus flattening, ensuring that the feature extraction network can comprehensively extract key fault information, and solving the problem that it is difficult to effectively extract bearing fault features from complex noises in the prior art.

[0153] As an optional solution, the CBAM attention module includes a channel attention module and a spatial attention module. The channel attention module squeezes the obtained bearing fault feature information spatially, and at the same time sums up the information features spatially by using maximum pooling and average pooling element by element, so as to obtain the weight feature of channel attention. The spatial attention module uses maximum pooling and average pooling on the obtained bearing fault feature information in the channel dimension to obtain two different information feature descriptions, and then stacks the two information feature descriptions together to obtain the weight feature of spatial attention.

[0154] An alternative solution is that the decomposition unit includes a first determination module and a decomposition module; the first determination module is configured to use the first formula: to determine the MODWPT decomposition coefficient level of the bearing signal, where h(t) is a low-pass filter, g(t) is a high-pass filter, is the z-th decomposition coefficient of the j-th level, z = 0, 1, 2…, 2 j-1 , t is the continuous time change parameter at the previous moment, and k is the continuous time change parameter at the next moment; the decomposition module is configured to decompose the bearing signal into the bearing time-frequency information at different resolutions according to the MODWPT decomposition coefficient level.

[0155] An alternative solution is that the determination unit includes a construction module and a second determination module; the construction module is configured to construct a bearing fault diagnosis model, where the domain adaptation adversarial learning framework of the bearing fault diagnosis model combines the entropy-conditional domain adversarial network ECDAN with an improved joint discriminant probability maximum mean discrepancy JMMD strategy, and adjusts the network parameters of the entropy-conditional domain adversarial network through a gradient reversal layer during the backpropagation process; the second determination module is configured to use the bearing fault diagnosis model to determine the bearing fault according to the bearing feature information.

[0156] An alternative solution is that the determination unit includes a third determination module, which is configured to use the second formula: to determine the JMMD strategy L JMMD (P, Q), where P is the source domain distribution, Q is the target domain distribution, |L| is the number of layers in the corresponding set, and H l is the l-th layer of the reproducing kernel Hilbert space, is the characteristic mapping of the tensor product in the Hilbert space; z sl is the activation generated by the source domain at the l-th layer; z tl is the activation generated by the target domain at the l-th layer.

[0157] An alternative solution is that the determination unit further includes a training module, which is configured to train the bearing fault diagnosis model with sample data to obtain a trained bearing fault diagnosis model before using the bearing fault diagnosis model to determine the bearing fault according to the bearing feature information, where the sample data is divided into source domain bearing fault samples with fault labels and target domain bearing fault samples without fault labels, and determines whether the bearing fault diagnosis model is trained completed according to minimizing the classification loss and maximizing the discriminant loss.

[0158] An alternative solution is that the device further includes an upload unit, which is configured to determine a solution to the bearing fault according to the bearing fault after determining the bearing fault according to the bearing feature information, and upload the bearing fault and the solution to the user terminal.

[0159] The above-mentioned bearing fault diagnosis device based on MODWPT includes a processor and a memory. The above-mentioned decomposition unit, construction unit, determination unit, etc. are all stored in the memory as program units, and the corresponding functions are realized by the processor executing the above-mentioned program units stored in the memory. The above-mentioned modules are all located in the same processor; or, the above-mentioned each module is located in different processors in any combined form.

[0160] The processor contains a kernel, and the kernel retrieves the corresponding program units from the memory. One or more kernels can be set, and by adjusting the kernel parameters, the problem that it is difficult to effectively extract the bearing fault characteristics from complex noises in the prior art can be solved.

[0161] The memory may include non-permanent memory in a computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM), and the memory includes at least one storage chip.

[0162] An embodiment of the present invention provides a computer-readable storage medium. The above-mentioned computer-readable storage medium includes a stored program, wherein when the above-mentioned program runs, it controls the device where the above-mentioned computer-readable storage medium is located to execute the above-mentioned bearing fault diagnosis method based on MODWPT.

[0163] An embodiment of the present invention provides a processor. The above-mentioned processor is used to run a program, wherein when the above-mentioned program runs, it executes the above-mentioned bearing fault diagnosis method based on MODWPT.

[0164] An embodiment of the present invention provides an electronic device. The device includes a processor, a memory, and a program stored on the memory and executable on the processor. When the processor executes the program, it realizes at least the steps of the bearing fault diagnosis method based on MODWPT.

[0165] The device in this article can be a server, a PC, a PAD, a mobile phone, etc.

[0166] This application also provides a computer program product, which is suitable for executing a program initialized with at least the steps of the bearing fault diagnosis method based on MODWPT when executed on a data processing device.

[0167] Obviously, those skilled in the art should understand that the various modules or steps of the present invention described above can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed over a network composed of multiple computing devices. They can be implemented by program codes executable by the computing device. Thus, they can be stored in a storage device and executed by the computing device. And in some cases, the steps shown or described can be executed in a different order than here, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. In this way, the present invention is not limited to any specific combination of hardware and software.

[0168] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0169] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified functions in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0170] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the specified functions in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0171] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide for implementing the specified functions in the flowFigure 1 one or more processes and / or blocks Figure 1 steps of functions specified in one or more blocks

[0172] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0173] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.

[0174] Computer-readable media includes permanent and non-permanent, removable and non-removable media and can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0175] It should also be noted that the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising the element.

[0176] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A bearing fault diagnosis method based on MODWPT, characterized in that: include: Acquire the bearing signal and use the maximum overlap discrete wavelet packet transform (MODWPT) to decompose the bearing signal into bearing time-frequency information at different resolutions. Constructing a feature extraction network model, wherein the feature extraction network model includes a ConvNext model, a CBAM attention module, a GGL graph generation layer, a GCN graph convolutional network and a feature extraction network, wherein the ConvNext model is used to extract bearing fault feature information from the bearing time-frequency information; the CBAM attention module is used to integrate and filter the bearing fault feature information to obtain target feature information; the GGL graph generation layer is used to learn the data structure from the target feature information to construct a feature instance graph; the GCN graph convolutional network is used to model the constructed feature instance graph to eliminate the vulnerability of key feature information to information one-sidedness and focus flattening; the feature extraction network is used to perform feature extraction processing on the result of the GCN graph convolutional network modeling processing; The feature extraction network model is used to perform feature extraction processing on the bearing time-frequency information to obtain bearing feature information, and the bearing fault is determined based on the bearing feature information.

2. The method according to claim 1, characterized in that: The CBAM attention module includes a channel attention module and a spatial attention module. The channel attention module squeezes the acquired bearing fault feature information in space, and at the same time adds the spatial information features element by element through maximum pooling and average pooling to obtain the weighted features of the channel attention. The spatial attention module uses maximum pooling and average pooling on the channel dimension to obtain two different information feature descriptions, and then stacks the two information feature descriptions together to obtain the weighted features of the spatial attention.

3. The method according to claim 1, characterized in that The maximum overlap discrete wavelet packet transform (MODWPT) is used to decompose the bearing signal into bearing time-frequency information at different resolutions, including: Using the first formula: Determine the MODWPT decomposition coefficient level of the bearing signal, where h(t) is a low-pass filter, g(t) is a high-pass filter, and is the z-th decomposition coefficient of the j-th level, z=0,1,2…,2 j-1 , t is the continuous time variation parameter at the previous moment, k is the continuous time varying parameter at the next moment; The bearing signal is decomposed into the bearing time-frequency information at different resolutions according to the MODWPT decomposition coefficient level.

4. The method according to claim 1, characterized in that: Determining a bearing fault according to the bearing characteristic information includes: Constructing a bearing fault diagnosis model, wherein the domain adaptive adversarial learning framework of the bearing fault diagnosis model combines an entropy conditional domain adversarial network (ECDAN) with an improved joint discriminant probability maximum mean difference (JMMD) strategy, and adjusts the network parameters of the entropy conditional domain adversarial network through a gradient reversal layer during back propagation; The bearing fault diagnosis model is used to determine the bearing fault according to the bearing characteristic information.

5. The method according to claim 4, characterized in that include: According to the second formula: Determine the JMMD strategy L JMMD (P,Q), where P is the source domain distribution, Q is the target domain distribution, |L| is the number of layers in the corresponding set, and H l 为 The lth level of the reproducing kernel Hilbert space, 为 Eigenmaps of tensor products in Hilbert space; z sl The activation generated by the source domain at layer l; z tl 为 The activations produced by the target domain at layer l.

6. The method according to claim 4, characterized in that Before using the bearing fault diagnosis model to determine the bearing fault according to the bearing characteristic information, the method further includes: The bearing fault diagnosis model is trained using sample data to obtain a trained bearing fault diagnosis model, wherein the sample data is divided into source domain bearing fault samples with fault labels and target domain bearing fault samples without fault labels, and whether the bearing fault diagnosis model is trained is determined based on minimizing classification loss and maximizing discrimination loss.

7. The method according to claim 1, characterized in that After determining the bearing fault according to the bearing characteristic information, the method further includes: A solution to the bearing fault is determined according to the bearing fault, and the bearing fault and the solution are uploaded to a user terminal.

8. A bearing fault diagnosis device based on MODWPT, characterized in that: include: A decomposition unit is used to obtain bearing signals and decompose the bearing signals into bearing time-frequency information at different resolutions using maximum overlap discrete wavelet packet transform (MODWPT); A construction unit is used to construct a feature extraction network model, wherein the feature extraction network model includes a ConvNext model, a CBAM attention module, a GGL graph generation layer, a GCN graph convolution network and a feature extraction network, wherein the ConvNext model is used to extract bearing fault feature information from the bearing time-frequency information; the CBAM attention module is used to integrate and filter the bearing fault feature information to obtain target feature information; the GGL graph generation layer is used to learn the data structure from the target feature information to construct a feature instance graph; the GCN graph convolution network is used to model the constructed feature instance graph to eliminate the vulnerability of key feature information to information one-sidedness and focus flattening; the feature extraction network is used to perform feature extraction on the result of the GCN graph convolution network modeling process; A determination unit is used to perform feature extraction processing on the bearing time-frequency information using the feature extraction network model to obtain bearing feature information, and determine the bearing fault according to the bearing feature information.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the bearing fault diagnosis method based on MODWPT according to any one of claims 1 to 7.

10. An electronic device, characterized in that: include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs include a method for executing the MODWPT-based bearing fault diagnosis method described in any one of claims 1 to 7.

Images