Parallel double-path bearing fault diagnosis method and system based on deep learning

Abstract

The invention discloses a parallel double-path bearing fault diagnosis method and system based on deep learning, and the method comprises the steps: collecting a fault vibration signal of a rolling bearing, and carrying out the feature engineering processing of the fault vibration signal; the feature engineering processing comprises variational mode decomposition and continuous wavelet transform, and intrinsic mode functions of different fault types and corresponding wavelet time-frequency diagrams are generated. Feature data obtained through processing serve as input of a CNN branch in the two-way model, and the other branch adopts an original signal to be input to an MLP branch. And finally, forming a fault diagnosis result by combining the outputs of the two branches through the full connection layer. The method has the characteristics of low sample demand, high diagnosis accuracy and strong robustness, is suitable for bearing fault detection under the condition of limited data samples, and remarkably improves the efficiency and accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a parallel dual-path bearing fault diagnosis method and system based on deep learning. Background Art

[0002] Rolling bearings are key components of mechanical equipment and play a vital role in industries such as metallurgy, chemical industry, aerospace, and high-end machinery such as aircraft and high-speed trains. Bearings usually operate under high loads and harsh conditions, and undetected faults may cause significant economic losses and safety risks. Therefore, the development of efficient and reliable fault diagnosis models is crucial to improving the reliability and safety of these components. Effective monitoring and diagnostic technologies can not only detect faults in a timely manner and prevent accidents, but also optimize maintenance plans, reduce unnecessary downtime, and significantly improve productivity and economic benefits. These technologies are indispensable to ensure production safety and maintain optimal equipment performance.

[0003] However, since bearings may fail after months or years, data collection usually takes a long time and the failure samples are often small, which requires the model to maintain high accuracy even with fewer samples.

[0004] In this regard, although fault injection experiments can provide labeled data, the cost of fault injection experiments is high, and the frequent changes in the operating conditions of mechanical systems further complicate data collection. Therefore, signal processing techniques are crucial because they can effectively extract important information from complex signals by analyzing raw data, thereby improving the accuracy and efficiency of fault detection. Summary of the invention

[0005] The present invention aims to solve one of the technical problems existing in the related art at least to a certain extent.

[0006] One object of the present invention is to provide a parallel dual-path bearing fault diagnosis method based on deep learning, which effectively extracts features by processing the original small sample data, and at the same time constructs a deep learning model that integrates multiple channels to improve the accuracy and robustness of bearing fault diagnosis.

[0007] Another object of the present invention is to provide a parallel dual-path bearing fault diagnosis system based on deep learning.

[0008] In order to achieve the above-mentioned purpose, the present invention provides a parallel dual-path bearing fault diagnosis method based on deep learning, comprising:

[0009] S1, collecting vibration signals of the bearing during operation;

[0010] S2, after performing variational mode decomposition and continuous wavelet transform on the vibration signal, fusion processing is performed to obtain a feature set;

[0011] S3. Construct a dual-link deep learning network framework, wherein the first link adopts a residual CNN network framework combined with an SE attention mechanism, and the second link adopts an MLP framework combined with a multi-head attention mechanism; the feature set is used as the input of the first link, and the vibration signal is used as the input of the second link. The diagnostic results output by the dual links are combined through a fully connected layer to train and obtain a fault diagnosis model;

[0012] S4. After the operating vibration signal of the bearing to be detected is processed by step S2, it is input into the trained fault diagnosis model together with the operating vibration signal to obtain the fault diagnosis result of the bearing to be detected.

[0013] A further preferred technical solution of the present invention is that, after performing variational modal decomposition and continuous wavelet transform on the vibration signal in step S2, fusion processing is performed to obtain a feature set; comprising:

[0014] S21, performing variational modal decomposition on the vibration signal to obtain a number of inherent modal components;

[0015] S22, performing continuous wavelet transform on each intrinsic modal component to obtain a time-frequency diagram corresponding to each intrinsic modal component;

[0016] S23, performing fusion processing on the obtained time-frequency graph.

[0017] Preferably, when performing variational modal decomposition in step S21, the number of decomposition modes K is first set, and a different decomposition K value is used for each fault type to obtain a different number of natural modal components.

[0018] Preferably, the number K of decomposition modes is determined as follows:

[0019] When performing variational modal decomposition, start with a smaller K value, calculate the center frequency of each mode obtained by each decomposition, and then gradually increase the K value. If further increasing the K value does not lead to a significant change in the center frequency, or the center frequencies of adjacent modes are too close, it indicates that modal mixing begins to occur at this K value, and the optimal K value is finally determined.

[0020] Preferably, when performing continuous wavelet transform on each intrinsic modal component in step S22, complex Morlet wavelet (cmor) is used for transform.

[0021] Preferably, when the obtained time-frequency graph is fused in step S23, a 1×1 convolution layer is used to unify the channel dimension, and the 1×1 convolution layer linearly combines features from different fault types in different channels.

[0022] Another aspect of the present invention provides a parallel dual-path bearing fault diagnosis system based on deep learning, comprising:

[0023] A data acquisition module is used to collect vibration signals of the bearing during operation;

[0024] The data processing module is used to perform variational mode decomposition and continuous wavelet transform on the vibration signal, and then perform fusion processing to obtain a feature set;

[0025] A model building module is used to build a dual-link deep learning network framework, wherein the first link adopts a residual CNN network framework combined with an SE attention mechanism, and the second link adopts an MLP framework combined with a multi-head attention mechanism; and is used to use the feature set as the input of the first link, and the original signal as the input of the second link, and the diagnosis results output by the dual links are combined through a fully connected layer to train and obtain a fault diagnosis model;

[0026] The fault diagnosis module is connected to the data processing module and the model building module respectively, and is used to send the operating vibration signal of the bearing to be detected to the data processing module for processing, and input the processed feature set together with the operating vibration signal into the fault diagnosis model trained by the model building module to obtain the fault diagnosis result of the bearing to be detected.

[0027] Another aspect of the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, which enable a computer to execute the above-mentioned parallel dual-path bearing fault diagnosis method based on deep learning.

[0028] Yet another aspect of the present invention provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus, and the processor calls the logic instructions in the memory to execute the above-mentioned parallel dual-path bearing fault diagnosis method based on deep learning.

[0029] On the other hand, the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the above-mentioned parallel dual-path bearing fault diagnosis method based on deep learning.

[0030] Beneficial effects: The present invention has the characteristics of low sample requirements, high diagnostic accuracy, and strong robustness, and is suitable for bearing fault detection when data samples are limited. At the same time, the present invention collects the vibration signal of the bearing during operation, performs variational mode decomposition and continuous wavelet transform on the vibration signal of the bearing during operation, and generates wavelet time-frequency diagrams respectively, including: time-frequency diagrams of the intrinsic mode function of different K values ​​for each fault type. The processed features are input into the CNN branch network combined with the attention mechanism, the original signal is input into the MLP combined with the attention mechanism, and finally the fault diagnosis structure is fully connected and output, which effectively improves the diagnostic efficiency and accuracy of bearing faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of a parallel dual-path bearing fault diagnosis method based on deep learning of the present invention;

[0032] Figure 2 is a frequency amplitude spectrum of a variational modal decomposition of a vibration sensor under a bearing inner race fault in an embodiment of the present invention;

[0033] Figure 3 is a schematic diagram of a structure in which channel normalization is performed through a 1×1 convolutional layer in an embodiment of the present invention;

[0034] Figure 4 This is the process of combining feature engineering in the CNN branch of the attention mechanism in the embodiment of the present invention;

[0035] Figure 5 It is a schematic diagram of the residual CNN structure combined with the attention mechanism;

[0036] Figure 6 This is a schematic diagram of the MLP structure combined with the attention mechanism;

[0037] Figure 7 It is the time-frequency diagram of multiple IMFs;

[0038] Figure 8 is the box plot of ten experiments;

[0039] Fig. 9 This is a T-sne visualization of the feature engineering and CNN models.

[0040] in, Fig. 9 (a) is the T-sne visualization of CWT+CNN. Fig. 9 (b) is the T-sne visualization of VMD+CWT+CNN;

[0041] Fig.10 This is a confusion matrix diagram comparing feature engineering and CNN models;

[0042] in, Fig.10(a) is the CWT+CNN confusion matrix result, Fig.10 (b) is the VMD+CWT+CNN confusion matrix result;

[0043] Fig.11 is a comparison diagram of the ablation experimental model;

[0044] Fig.12 It is a line chart of accuracy;

[0045] in, Fig.12 (a) is the line chart of VMD+CWT+CNN accuracy. Fig.12 (b) is a line graph of the accuracy of the two-way model;

[0046] Fig.13 It is the T-sne visualization of the two-path model;

[0047] in, Fig.13 (a) is the T-sne visualization of VMD+CWT+CNN. Fig.13 (b) is the T-sne visualization of the two-path model;

[0048] Fig.14 It is the confusion matrix diagram of the two-way model;

[0049] in, Fig.14 (a) is the confusion matrix diagram of VMD+CWT+CNN (b) is the confusion matrix diagram of the two-way model; DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0051] Combine the following Figure 1-Figure 14 The present invention describes a parallel dual-path bearing fault diagnosis method and system based on deep learning.

[0052] Embodiment 1: This embodiment provides a parallel dual-path bearing fault diagnosis method based on deep learning, comprising:

[0053] S1, collecting vibration signals of the bearing during operation;

[0054] S2, after performing variational mode decomposition and continuous wavelet transform on the vibration signal, fusion processing is performed to obtain a feature set;

[0055] S3. Construct a dual-link deep learning network framework, wherein the first link adopts a residual CNN network framework combined with an SE attention mechanism, and the second link adopts an MLP framework combined with a multi-head attention mechanism; the feature set is used as the input of the first link, and the vibration signal is used as the input of the second link. The diagnostic results output by the dual links are combined through a fully connected layer to train and obtain a fault diagnosis model;

[0056] S4. After the operating vibration signal of the bearing to be detected is processed by step S2, it is input into the trained fault diagnosis model together with the operating vibration signal to obtain the fault diagnosis result of the bearing to be detected.

[0057] Step S2 is feature engineering processing, vibration signal variational mode decomposition and continuous wavelet processing, and the generated two-dimensional time-frequency diagram is combined along the channel dimension to create a fusion signal for feature engineering. It includes:

[0058] S21, performing variational mode decomposition (VMD) on the vibration signal to obtain a number of intrinsic mode components (IMFs);

[0059] S22, performing continuous wavelet transform (CWT) on each intrinsic mode component (IMF) to obtain a time-frequency diagram corresponding to each intrinsic mode component (IMF);

[0060] S23, performing fusion processing on the obtained time-frequency graph.

[0061] For each fault type, a different decomposition K value is used to obtain a different number of IMFs. Then, each inherent modal component is subjected to continuous wavelet transformation to obtain a time-frequency diagram, which is then stacked to obtain the data.

[0062] When processing signals received from sensors, decomposing nonlinear signals, such as wavelet transform (WT), wavelet packet transform (WPT), short-time Fourier transform (STFT), empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD) and variational mode decomposition (VMD), are used. So continuous wavelet transform (CWT) plays a big role in the analysis of unsteady signals. WT and WPT have difficulty in accurately identifying signal discontinuities, and the computational cost can be large. EMD and EEMD often face problems of mode mixing and high computational cost. In contrast, VMD provides robust mode decomposition with better frequency separation and stability, making it ideal for analyzing complex signals.

[0063] Variational Mode Decomposition (VMD) is an adaptive signal decomposition method that decomposes a complex signal into multiple modal components, each with a specific center frequency. These components together constitute the original signal.

[0064] The core idea of ​​VMD is to decompose the signal f(t) into several modes u k (t), each mode has a narrowband spectrum characteristic. The optimization goal is to minimize the bandwidth of each mode:

[0065]

[0066] Where: u k (t) is the kth mode. ω k is the center frequency of the mode. δ(t) is the Dirac function. j is the imaginary unit.

[0067] VMD iteratively updates u via the alternating direction method of multipliers (ADMM) k and ω k :

[0068]

[0069] Its continuous wavelet transform is defined as WT Ψ f(α, τ), the formula is as follows

[0070]

[0071] Among them, α is the translation factor, which determines the position of the time-frequency window in the time domain; τ is the scale factor, which determines the size of the time-frequency window and its position in the frequency domain; f(t) is the source time variable signal; R∈(-∞, +∞); t is the time variable; express The complex conjugate of α，τ (t) is the wavelet basis function (also called mother wavelet) after translation and scale change, and its expression is:

[0072]

[0073] In general, it is advantageous to select a wavelet basis function that closely matches the fault characteristics of the signal. The complex Morlet wavelet (CMOR) is particularly suitable for this purpose due to its similarity to the impulse waveform characteristics of bearing faults. Therefore, we chose the CMOR wavelet because of its excellent adaptability.

[0074] Variational mode decomposition (VMD) and continuous wavelet transform (CWT) are an effective combination of signal processing methods, which aims to extract important features from signals and improve the accuracy and robustness of data analysis. First, the input signal is decomposed into multiple IMFs by VMD. Each IMF corresponds to a narrowband frequency component. VMD can effectively avoid the problem of mode mixing and improve the accuracy of modal separation. The VMD algorithm requires the number of decomposition modes K to be set in advance. Choosing a K value that is too large or too small will affect the decomposition result, resulting in frequency aliasing or insufficient decomposition. Here, the default value of the VMD parameter α is 2000. Different decomposition K values ​​are used for each fault type to obtain different numbers of IMFs. Then, each IMF is subjected to continuous wavelet transformation to obtain a time-frequency diagram, which is then stacked to obtain data. Figure 2 A plot showing the effect of decomposing a signal into different IMFs.

[0075] In the embodiment of the present invention, a fusion algorithm based on variational mode decomposition and continuous wavelet transform is provided. In view of the characteristics of modal aliasing of different components of variational mode decomposition and noise in components under strong noise, continuous wavelet transform is used. Experimental results show that the data processing method combining variational mode decomposition and continuous wavelet transform has the effect of enhancing feature extraction and has a good fusion effect.

[0076] In variational mode decomposition (VMD), choosing the appropriate number of modes K is the key to ensuring the signal decomposition effect. Too small K may lead to the loss of important signal features, while too large K will introduce redundant information. Therefore, the center frequency-based method provides an effective way to determine the K value. Start VMD decomposition from a smaller K value, and calculate the center frequency ω of each mode obtained in each decomposition. k , which can usually be obtained by analyzing the frequency spectrum of the mode. Increase the K value and repeat the decomposition until the center frequencies of adjacent modes are found to be significantly different. When K is further increased, if the center frequencies of the decomposed modes no longer change significantly, or the adjacent center frequencies are too close, the current K value can be considered to be the appropriate number of modes. The K value with the last significant change in center frequency is selected as the final number of modes. Subsequently, the continuous wavelet transform is applied to the generated IMF to obtain the time-frequency diagram, further enhancing the data analysis process.

[0077] After the K value is determined by variational mode decomposition (VMD), the K value of each fault type may be different, which leads to the difference in the channel dimension of the time-frequency images generated by the intrinsic mode functions (IMFs) of different types of faults after continuous wavelet transform (CWT), and cannot be directly fused. In order to solve this problem, Figure 3As shown, this embodiment uses 1×1 convolution to unify the channel dimension. 1×1 convolution achieves consistency in the number of channels by linearly combining the features of different channels, so that features from different fault types can be effectively fused and analyzed. In addition, 1×1 convolution can compress or enhance specific features while maintaining the integrity of input features, helping the network to effectively select and combine important features and improve the representation ability of the model. Compared with other convolution operations, 1×1 convolution has a lower computational complexity and can significantly enhance the integration and transmission capabilities of features while maintaining computational efficiency. In this way, the information in the CWT time-frequency diagrams of different IMFs can be more effectively fused, the feature expression capability can be improved, and the model can more accurately capture the subtle differences between different fault types, thereby improving the accuracy of fault diagnosis. Therefore, 1×1 convolution provides an effective solution for the fusion of features of different fault types, simplifies the feature integration process, enhances the expression and analysis capabilities of features, and provides strong support for complex signal processing tasks.

[0078] Step S3 is to build and train the model. The specific process is as follows:

[0079] (1) Building a CNN network

[0080] In order to verify the superiority of the small sample feature fusion method, a CNN model with residual and SE attention mechanism was selected. The processing flow chart of the first link is as follows Figure 4 The network structure is shown in Figure 5 As shown in Figure 1. The first link is developed by enhancing the traditional CNN architecture by adding residual blocks and combining the SE attention mechanism. The introduction of residual blocks enhances the feature extraction capability of the model and alleviates the problem of gradient vanishing. The SE attention mechanism enhances the expressiveness of learned features and accurately locates objects of interest, resulting in more accurate and robust performance.

[0081] The residual CNN model consists of 12 layers, including input layer, convolution layer, residual layer, normalization layer, ReLU layer, pooling layer, fully connected layer, Softmax layer and output layer. Two layers of convolution are added to the residual block, and then the residual is connected to the input features, activated by the ReLU activation function, and finally output. The SE attention mechanism is added after the output of the residual convolution network, and the spatial dimension of each channel is compressed into a single statistic through global average pooling. A bottleneck structure consisting of two fully connected layers is used to learn the relative importance of each channel. Specifically, the first fully connected layer reduces the number of channels to a scaling factor of the original (1 / 16), and then passes through the ReLU activation function, and the second fully connected layer restores the number of channels to the original number of channels. Finally, the activation of each channel is mapped to a weight coefficient between [0,1] through the Sigmoid function. The specific parameters of the CNN network are shown in Table 1.

[0082] Table 1 Residual CNN parameters

[0083]

[0084] (2) Building an MLP network

[0085] In the dual-path model, the second link extracts features from the original signal using a multi-layer perceptron (MLP) and a multi-head attention mechanism. The MLP processes these features in a nonlinear way, capturing complex patterns and relationships in the data. The multi-head attention mechanism improves accuracy by identifying feature correlations and focusing on multiple aspects simultaneously.

[0086] The MLP network structure is as follows Figure 6 As shown. A multi-head attention mechanism is integrated after the first MLP module of the MLP network. The input features are projected into multiple subspaces through a linear layer, and independent query, key, and value matrices are created for each attention head. For each head, the dot product between the query and the key is calculated and divided by a scaling factor (usually the square root of the dimension of the key) to obtain the attention score. A Softmax function is applied to the attention score to obtain the weight, which is then multiplied by the value to obtain the attention output. The attention outputs of all heads are concatenated together and transformed through a linear layer to restore the original dimension. Thus, an enhanced MLP model is formed. This integration solves the limitations of CNN models (which usually require a lot of data), enabling the model to focus on key features and capture complex dependencies more effectively, even with a small data set. The specific parameters of the MLP network are shown in Table 2.

[0087] Table 2 MLP parameters

[0088] LayerName InputDimension Outputdimension Other Parmameters Fc1 1024 512 ReLU Multi-HeadAttention 512 512 Head=8 FC2 512 256 ReLU FC3 156 128 ReLU Dropout 128 128 Rate = 0.3

[0089] (3) Model training: The mixed data set after data set processing is selected and batched and input into the dual-path model for training. 20 samples of each fault type are used for training and 80 samples are used for testing. During the training process, the network architecture and parameters are adjusted according to the training results.

[0090] Overall, the integration of residual CNN with attention mechanism and MLP model with attention mechanism creates a powerful architecture for signal analysis. Including: The data processed by feature engineering is input into the CNN branch, which enhances feature extraction with its residual block and SE attention mechanism. At the same time, the original signal is input into the MLP branch, which adopts multi-head attention to capture complex dependencies and patterns. After feature extraction, the outputs of the two models are combined in the fully connected layer to form the final integrated classification output. This approach takes advantage of the strengths of both models, thereby achieving efficient and effective data analysis and classification.

[0091] Embodiment 2: This embodiment provides a parallel dual-path bearing fault diagnosis system based on deep learning, comprising:

[0092] A data acquisition module is used to collect vibration signals of the bearing during operation;

[0093] The data processing module is used to perform variational mode decomposition and continuous wavelet transform on the vibration signal, and then perform fusion processing to obtain a feature set;

[0094] A model building module is used to build a dual-link deep learning network framework, wherein the first link adopts a residual CNN network framework combined with an SE attention mechanism, and the second link adopts an MLP framework combined with a multi-head attention mechanism; and is used to use the feature set as the input of the first link, and the original signal as the input of the second link, and the diagnosis results output by the dual links are combined through a fully connected layer to train and obtain a fault diagnosis model;

[0095] The fault diagnosis module is connected to the data processing module and the model building module respectively, and is used to send the operating vibration signal of the bearing to be detected to the data processing module for processing, and input the processed feature set together with the operating vibration signal into the fault diagnosis model trained by the model building module to obtain the fault diagnosis result of the bearing to be detected.

[0096] Embodiment 3: This embodiment provides a non-transitory computer-readable storage medium on which computer instructions are stored. The computer instructions enable a computer to execute a parallel dual-path bearing fault diagnosis method based on deep learning. The method includes the following steps:

[0097] S1, collecting vibration signals of the bearing during operation;

[0098] S2, after performing variational mode decomposition and continuous wavelet transform on the vibration signal, fusion processing is performed to obtain a feature set;

[0099] S3. Construct a dual-link deep learning network framework, wherein the first link adopts a residual CNN network framework combined with an SE attention mechanism, and the second link adopts an MLP framework combined with a multi-head attention mechanism; the feature set is used as the input of the first link, and the vibration signal is used as the input of the second link. The diagnostic results output by the dual links are combined through a fully connected layer to train and obtain a fault diagnosis model;

[0100] S4. After the operating vibration signal of the bearing to be detected is processed by step S2, it is input into the trained fault diagnosis model together with the operating vibration signal to obtain the fault diagnosis result of the bearing to be detected.

[0101] Embodiment 4: This embodiment provides an electronic device, which may include: a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor may call the logic instructions in the memory to execute a parallel dual-path bearing fault diagnosis method based on deep learning, and the method includes the following steps:

[0102] S1, collecting vibration signals of the bearing during operation;

[0103] S2, after performing variational mode decomposition and continuous wavelet transform on the vibration signal, fusion processing is performed to obtain a feature set;

[0104] S3. Construct a dual-link deep learning network framework, wherein the first link adopts a residual CNN network framework combined with an SE attention mechanism, and the second link adopts an MLP framework combined with a multi-head attention mechanism; the feature set is used as the input of the first link, and the vibration signal is used as the input of the second link. The diagnostic results output by the dual links are combined through a fully connected layer to train and obtain a fault diagnosis model;

[0105] S4. After the operating vibration signal of the bearing to be detected is processed by step S2, it is input into the trained fault diagnosis model together with the operating vibration signal to obtain the fault diagnosis result of the bearing to be detected.

[0106] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0107] Embodiment 5: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a parallel dual-path bearing fault diagnosis method based on deep learning, which includes the following steps:

[0108] S1, collecting vibration signals of the bearing during operation;

[0109] S2, after performing variational mode decomposition and continuous wavelet transform on the vibration signal, fusion processing is performed to obtain a feature set;

[0110] S3. Construct a dual-link deep learning network framework, wherein the first link adopts a residual CNN network framework combined with an SE attention mechanism, and the second link adopts an MLP framework combined with a multi-head attention mechanism; the feature set is used as the input of the first link, and the vibration signal is used as the input of the second link. The diagnostic results output by the dual links are combined through a fully connected layer to train and obtain a fault diagnosis model;

[0111] S4. After the operating vibration signal of the bearing to be detected is processed by step S2, it is input into the trained fault diagnosis model together with the operating vibration signal to obtain the fault diagnosis result of the bearing to be detected.

[0112] In order to verify the technical effects that can be produced by the above embodiments, experimental verification was carried out. The experiment was based on PyCharm using the PyTorch 2.2.2 framework to train and test the model. The computing tasks were performed on a 64-bit Windows 11 operating system equipped with an Intel(R) Core(TM) i5-9300H CPU @ 2.40GHz, an NVIDIA RTX 1660Ti GPU, and 16GB RAM.

[0113] (1) Dataset introduction:

[0114] The experiment uses the Case Western Reserve University (CWRU) rolling bearing fault dataset, which is widely used in bearing fault diagnosis based on vibration signals. The CWRU test bench consists of a 1.5kW motor, a torque sensor, and a dynamometer. The motor load power is 0HP, 1HP, 2HP, and 3HP, corresponding to speeds of 1797rpm, 1772rpm, 1750rpm, and 1730rpm (1 horsepower = 735.49875 watts). The rolling bearing model used is SKF6205-2RS JEM SKF, and the sampling frequency is 12kHz. Single-point faults with diameters of 0.007mm and 0.014. 0.021mm and 0.021mm are introduced on the inner ring, outer ring, and rolling element of the bearing by electrical discharge machining. The fault positions on the outer ring are at 3 o'clock, 6 o'clock, and 12 o'clock, and the experiment uses the data at the 6 o'clock position. Data of 9 fault states and 1 normal state of the rolling bearing are selected, and each state has 233 sets of samples. Each group consists of 1024 sampling points. The experimental sample data set is detailed in Table 3.

[0115] Table 3 CWRU experimental sample dataset

[0116]

[0117]

[0118] (2) Data processing

[0119] The VMD algorithm requires the number of decomposition modes K to be set in advance. Choosing a K value that is too large or too small will affect the decomposition result, resulting in frequency aliasing or insufficient decomposition. Here, the default value of the VMD parameter is 2000. The vibration signal of the inner loop fault In(4) is taken as an example. After VMD decomposition of the vibration signal, the center frequencies of each modal component under different K values ​​are shown in Table 4.

[0120] Table 4 Center frequencies corresponding to different K values

[0121]

[0122]

[0123] When the mode number K is 6, the center frequencies of 551.56 Hz and 580.17 Hz are close to each other, and mode aliasing may occur. It is more appropriate to select the mode number K as 5.

[0124] A group of vibration signals is selected for VMD analysis in each state, and the number of vibration modes in other categories is determined as follows: 6(0), 6(1), 6(2), 6(3), 5(5), 8(6), 6(7), 5(8), and 5(9). Figure 7 As shown, sub-signals of different scales exhibit unique vibration characteristics in different frequency bands.

[0125] (3) Model parameter setting

[0126] In order to allow the model to fully learn the features, the number of training rounds (Epoch) is set to 20. In order to better train the model, the Adam optimizer is used to optimize the cross entropy loss function and the initial learning rate is set to 0.003. In terms of the batch size setting, a smaller batch size can achieve better generalization, so the batch size is set to 32.

[0127] (4) Comparison of ablation trials using different treatment methods

[0128] In order to fully evaluate the performance of the proposed multi-scale signal processing, the present invention selected a CNN model with a basic continuous wavelet transform for performance test comparison. Their models are the same, only the processing methods are different. Table 5 summarizes the specific information of the comparison model. It can be seen that the average accuracy of the experimental results using the feature engineering data processing method of this paper has increased by 0.326 percentage points.

[0129] Table 5. Comparison of ablation experiments

[0130]

[0131]

[0132] The box plots of the ten experiments are shown in Figure 8 The results of one of the experiments are randomly selected for T-SNE visualization and confusion matrix visualization, as shown in Figure 1. Fig. 9 and 10 shown.

[0133] (5) Dual-path model comparison

[0134] In order to verify the superiority of the proposed dual-link modeling method, the CWRU dataset was used to perform intra-model and inter-model comparisons. First, a single model residual CNN was selected for fault diagnosis after being processed with the feature engineering data proposed in the present invention. Another model was to perform a separate VMD decomposition of the one-dimensional signal and the hybrid dual-path model proposed in the present invention on different test sets (A, B, and C), and each training set contained 20 samples in each category. Ten rounds of experiments were performed on each dataset to calculate the average performance. Fig.11As shown in Figure 3, the accuracy of the proposed model on different test sets exceeds that of the single-channel model. This shows that fusing the original fault signals can significantly improve the generalization ability of the model, especially in complex scenarios.

[0135] For further validation, 160 test sets are randomly selected for diagnosis of the one-way and two-way models.

[0136] The line graph of accuracy is as follows Fig.12 As shown in Figure 2. T-SNE visualization and confusion matrix analysis are shown in Figure 2. Fig.13 and Fig.14 The T-SNE visualization shows that the two-channel model effectively clusters the data, making the categories more distinguishable. The confusion matrix highlights the model's ability to accurately classify each category, demonstrating the robustness and reliability of the proposed two-channel approach. This analysis confirms that the approach not only improves accuracy but also ensures that the model generalizes well to new, unseen data.

[0137] Then, the models are compared longitudinally, and the superiority of the method is further verified through comparative experiments. Again, the CWRU dataset is used as the experimental dataset. Different models are used to prove the accuracy, and the results are shown in Table 6.

[0138] Table 6 Experimental comparison of different models

[0139]

[0140] The experimental results show that the feature engineering method exhibits excellent robustness. The accuracy comparison experiment on a small sample data set shows that the proposed method has an average accuracy of 99.34% in ten rounds of experiments, while the traditional CWT method achieves an accuracy of 98.97%. This shows that the proposed method provides better feature extraction on small data sets.

[0141] The hybrid dual-channel model proposed in this invention, in which the MLP branch processes the original signal and the CNN branch processes the signal processed by the feature engineering described in the previous section, achieves 100% accuracy on the Case Western Reserve University dataset using the same training and test sets.

[0142] In summary, the present invention has the characteristics of high accuracy, strong robustness, good domain adaptability, etc., and can be used for bearing fault detection in a small sample environment.

[0143] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0144] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A parallel dual-path bearing fault diagnosis method based on deep learning, characterized in that: include: S1, collecting vibration signals of bearings during operation; S2, after performing variational mode decomposition and continuous wavelet transform on the vibration signal, fusion processing is performed to obtain a feature set; S3. Build a dual-link deep learning network framework, where the first link adopts the residual CNN network framework combined with the SE attention mechanism, and the second link adopts the MLP framework combined with the multi-head attention mechanism; The feature set is used as the input of the first link, the vibration signal is used as the input of the second link, the diagnosis results output by the two links are combined through a fully connected layer, and a fault diagnosis model is obtained by training; S4. After the operating vibration signal of the bearing to be detected is processed by step S2, it is input into the trained fault diagnosis model together with the operating vibration signal to obtain the fault diagnosis result of the bearing to be detected.

2. The parallel dual-path bearing fault diagnosis method based on deep learning according to claim 1 is characterized in that: Step S2, after performing variational modal decomposition and continuous wavelet transform on the vibration signal, performs fusion processing to obtain a feature set; includes: S21, performing variational modal decomposition on the vibration signal to obtain a number of inherent modal components; S22, performing continuous wavelet transform on each intrinsic mode component to obtain a time-frequency diagram corresponding to each intrinsic mode component; S23, performing fusion processing on the obtained time-frequency graph.

3. The parallel dual-path bearing fault diagnosis method based on deep learning according to claim 2 is characterized in that: When performing variational modal decomposition in step S21, the number of decomposition modes K is first set, and different decomposition K values ​​are used for each fault type to obtain different numbers of natural modal components.

4. The parallel dual-path bearing fault diagnosis method based on deep learning according to claim 3 is characterized in that: The number of decomposition modes K is determined as follows: When performing variational modal decomposition, start with a smaller K value, calculate the center frequency of each mode obtained by each decomposition, and then gradually increase the K value. If further increasing the K value does not lead to a significant change in the center frequency, or the center frequencies of adjacent modes are too close, it indicates that modal mixing begins to occur at this K value, and the optimal K value is finally determined.

5. The parallel dual-path bearing fault diagnosis method based on deep learning according to claim 2 is characterized in that: When performing continuous wavelet transform on each intrinsic modal component in step S22, complex Morlet wavelet (cmor) is used for transform.

6. The parallel dual-path bearing fault diagnosis method based on deep learning according to claim 2 is characterized in that: When the obtained time-frequency graph is fused in step S23, a 1×1 convolution layer is used to unify the channel dimension, and the 1×1 convolution layer linearly combines the features from different fault types in different channels.

7. A parallel dual-path bearing fault diagnosis system based on deep learning, characterized in that: include: A data acquisition module is used to collect vibration signals of the bearing during operation; The data processing module is used to perform variational mode decomposition and continuous wavelet transform on the vibration signal, and then perform fusion processing to obtain a feature set; A model building module is used to build a dual-link deep learning network framework, wherein the first link adopts a residual CNN network framework combined with an SE attention mechanism, and the second link adopts an MLP framework combined with a multi-head attention mechanism; and is used to use the feature set as the input of the first link, and the original signal as the input of the second link, and the diagnosis results output by the dual links are combined through a fully connected layer to train and obtain a fault diagnosis model; The fault diagnosis module is connected to the data processing module and the model building module respectively, and is used to send the operating vibration signal of the bearing to be detected to the data processing module for processing, and input the processed feature set together with the operating vibration signal into the fault diagnosis model trained by the model building module to obtain the fault diagnosis result of the bearing to be detected.

8. A non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions enable a computer to execute the deep learning-based parallel dual-path bearing fault diagnosis method described in any one of claims 1 to 6.

9. An electronic device, comprising: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus, and the processor calls the logic instructions in the memory to execute the parallel dual-path bearing fault diagnosis method based on deep learning as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer executes the parallel dual-path bearing fault diagnosis method based on deep learning as described in any one of claims 1 to 6.

Images