Rolling bearing fault diagnosis method and system, electronic equipment and storage medium

Abstract

The invention relates to the technical field of mechanical fault diagnosis and intelligent operation and maintenance, in particular to a rolling bearing fault diagnosis method and system, electronic equipment and a storage medium, and the method comprises the steps: obtaining a one-dimensional vibration signal of a rolling bearing; inputting the signals to an image branch and a time sequence branch respectively; the image branch converts a signal into a time frequency image through MTF, and spatial features are extracted through FDConv frequency domain dynamic convolution, ResNet and CBAM attention mechanisms in sequence; the time sequence branch performs segmented stacking on the signals, and extracts time features through a GRU network; and fusing the double-branch features and inputting into a classifier to realize fault type identification. According to the method, the perception capability of weak fault features is enhanced through frequency domain adaptive convolution and an attention mechanism, and the accuracy and robustness of bearing fault diagnosis under complex working conditions are remarkably improved by combining time-frequency double-branch feature fusion. The average diagnosis accuracy on two public data sets reaches 99.52% and 98.61% respectively.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis and intelligent operation and maintenance technology, and in particular to a method, system, electronic device and storage medium for diagnosing rolling bearing faults. Background Technology

[0002] Rolling bearings are core components of rotating machinery, and their health directly affects the safe operation of the entire equipment. Traditional fault diagnosis methods (such as vibration spectrum analysis and wavelet transform) rely heavily on expert knowledge and manual feature extraction, making it difficult to effectively capture early, subtle fault characteristics under complex operating conditions and high-noise environments.

[0003] In recent years, deep learning techniques, especially convolutional neural networks (CNNs), have been widely used in bearing fault diagnosis. These methods typically convert one-dimensional vibration signals into two-dimensional time-frequency images (e.g., through short-time Fourier transform, wavelet transform, etc.) for processing, or directly use one-dimensional CNNs to process the signals. However, such methods have the following limitations: 1. Insufficient utilization of time and frequency information: A single CNN structure cannot fully utilize the dynamic evolution of signals in the time domain and the fine structural features in the frequency domain at the same time.

[0004] 2. Limited feature extraction capability: Traditional convolution kernels are fixed and lack the ability to adaptively perceive different frequency components in bearing fault signals, and are not sensitive to weak features.

[0005] 3. Weak anti-interference ability: In the context of strong noise, the model has difficulty focusing on the key feature regions most relevant to the fault.

[0006] Therefore, there is an urgent need for a robust diagnostic scheme that can integrate complementary information in the time and frequency domains and has strong feature extraction and selection capabilities. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method, system, electronic device, and storage medium for diagnosing rolling bearing faults.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: Design a method for fault diagnosis of rolling bearings, including: acquiring a one-dimensional vibration signal of the rolling bearing, and performing bi-branch feature extraction and fusion on the one-dimensional vibration signal. One branch is the image branch, which first encodes the one-dimensional vibration signal into a two-dimensional time-frequency image through Markov transfer field (MTF), and then uses a ResNet network that integrates frequency domain dynamic convolution (FDConv) and convolutional block attention module (CBAM) to extract spatial features from the image; One branch is the time-series branch, which first segments and stacks the one-dimensional vibration signal to form a two-dimensional sequence, and then uses a gated recurrent unit (GRU) network to extract time features from the sequence; Finally, the features extracted from the two branches (spatial features and temporal features) are fused together, and a fully connected classifier is used to complete the diagnosis of the fault type.

[0009] Preferably, the "segmented stacking" in the timing branch specifically means: dividing the one-dimensional vibration signal x of length L into K segments, each segment of length S, K = L / S, thereby reconstructing a two-dimensional matrix of K rows and S columns as the input of the gated recurrent unit (GRU) network.

[0010] Preferably, the operation process of the frequency domain dynamic convolution (FDConv) is as follows: the convolution kernel weights are divided into multiple frequency band groups in the frequency domain; inverse discrete Fourier transform is performed on the coefficients of each frequency band to obtain multiple spatial convolution kernels; weights are then dynamically generated according to the input features, and these convolution kernels are weighted and fused into an adaptive dynamic convolution kernel.

[0011] Preferably, the Convolutional Block Attention (CBAM) module performs channel attention and spatial attention sequentially: channel attention calculates channel weights through global pooling and fully connected layers; spatial attention calculates spatial weights through pooling and convolution in the channel dimension.

[0012] Preferably, the feature fusion method for the two branches is to directly concatenate the spatial feature vector and the temporal feature vector.

[0013] Based on a rolling bearing fault diagnosis method, this invention also proposes a rolling bearing fault diagnosis system, comprising: Sensors used to acquire one-dimensional vibration signals; An image branch processing module, whose input is connected to a sensor, includes an MTF encoder, an FDConv module, a ResNet network, and a CBAM module in sequence; A timing branch processing module, whose input is connected to a sensor, and which sequentially includes a signal stacking unit and a GRU network; A feature fusion module is used to connect and fuse the outputs of two branch processing modules (image branch processing module and temporal branch processing module); A fault classifier whose input is connected to a feature fusion module and whose output is the fault type.

[0014] Based on a rolling bearing fault diagnosis method, the present invention also proposes an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the rolling bearing fault diagnosis method described above.

[0015] Based on a rolling bearing fault diagnosis method, the present invention also proposes a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it is used to implement the steps of the rolling bearing fault diagnosis method described above.

[0016] The present invention provides a method, system, electronic device, and storage medium for diagnosing rolling bearing faults, the advantages of which are as follows: 1. Multidimensional information complementarity: Through the dual-branch design of MTF image and original time series signal, the time-frequency structure information and time dynamic information of fault are utilized at the same time, forming a more comprehensive feature representation.

[0017] 2. Excellent feature extraction capability: The introduced FDConv module endows the model with adaptive learning capability in the frequency domain, which significantly enhances the sensitivity to the perception of fault components of different frequencies, especially weak high-frequency features.

[0018] 3. Strong noise resistance and focusing ability: The introduction of the CBAM attention mechanism enables the model to automatically filter and strengthen the channels and spatial regions most relevant to the fault, effectively improving robustness in noisy environments.

[0019] 4. High accuracy and robustness: As shown in the experiment, the present invention achieved an average diagnostic accuracy of over 98.5% on two publicly available bearing datasets, which is significantly better than traditional methods and single-branch models, demonstrating excellent potential for engineering applications.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an overall flowchart of the fault diagnosis method of the present invention; Figure 2 This is an overall flowchart of the fault diagnosis system of the present invention; Figure 3 This is a schematic diagram of the internal structure of the GRU unit in an embodiment of the present invention; Figure 4 This is a schematic diagram of the FDCA-ResNet architecture in an embodiment of the present invention; Figure 5This is a schematic diagram of the FDConv frequency domain dynamic convolution module structure in an embodiment of the present invention; Figure 6 This is a schematic diagram of the CBAM attention mechanism structure in an embodiment of the present invention; Figure 7 This is a schematic diagram of the BJTU-RAO bearing fault simulation experimental platform in an embodiment of the present invention; Figure 8 This is a schematic diagram of the HUST bearing failure simulation experimental platform in an embodiment of the present invention; Figure 9 This is a diagram showing the confusion matrix results of two datasets in an embodiment of the present invention, where (a) is Dataset A and (b) is Dataset B; Figure 10 This is a T-SNE feature visualization diagram of an embodiment of the present invention, where (a) is Dataset A and (b) is Dataset B; Figure 11 This is a comparison chart of the ablation test performance of embodiments of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Reference Figure 1-11 A method for diagnosing rolling bearing faults includes the following steps: A one-dimensional vibration signal of a rolling bearing is acquired, and a two-branch feature extraction and fusion process is performed on the one-dimensional vibration signal: an image branch and a temporal branch. First, the input one-dimensional vibration signal is encoded using Markov transfer field coding through the image branch, converting it into a two-dimensional time-frequency image. The resulting time-frequency image is then subjected to frequency domain dynamic convolution, residual network feature extraction, and weighting by a convolutional block attention module to obtain a spatial feature vector. Next, the input one-dimensional vibration signal is segmented and stacked through the temporal branch to construct a two-dimensional sequence matrix, which is then input into a gated recurrent unit network to extract deep time-dependent features, resulting in a temporal feature vector. Finally, the spatial and temporal feature vectors are concatenated and fused, and then input into a fully connected classifier to achieve intelligent fault type diagnosis.

[0025] The "segmented stacking" in the timing branch is as follows: the one-dimensional vibration signal x of length L is divided into K non-overlapping segments, each segment of length S, where K = L / S, thereby reconstructing a two-dimensional matrix X of K rows and S columns, which serves as the input to the gated recurrent unit (GRU) network.

[0026] First, the convolution kernel weight parameters These coefficients are considered as frequency domain spectral coefficients. They are sorted according to the Euclidean norm of the Fourier index and uniformly divided into G non-overlapping frequency band groups. The inverse discrete Fourier transform is performed on the coefficients of the g-th group to obtain G groups of spatial convolution kernels. Attention weights are adaptively generated according to the input feature map, and the G groups of spatial convolution kernels are weighted and fused to form input-related dynamic convolution kernels.

[0027] The convolutional block attention module sequentially performs channel attention and spatial attention mechanisms on the input feature map. For the input feature map F, the channel attention mechanism generates channel description vectors through global average pooling and global max pooling, which are then processed by a shared fully connected layer and activation function to generate channel attention weights. The spatial attention mechanism generates spatial attention weights by performing average pooling and max pooling along the channel dimension, concatenating the results, and then applying a 7×7 convolution and sigmoid activation. The final output feature is .

[0028] The spatial feature vector extracted from the image branch and the temporal feature vector extracted from the temporal branch are fused through a concatenation operation.

[0029] Based on the same inventive concept, the present invention provides a rolling bearing fault diagnosis system, comprising: Sensors used to acquire one-dimensional vibration signals; An image branch processing module, whose input is connected to a sensor, includes an MTF encoder, an FDConv module, a ResNet network, and a CBAM module in sequence; A timing branch processing module, whose input is connected to a sensor, and which sequentially includes a signal stacking unit and a GRU network; A feature fusion module is used to connect and fuse the outputs of two branch processing modules (image branch processing module and temporal branch processing module); A fault classifier whose input is connected to a feature fusion module and whose output is the fault type.

[0030] Based on the same inventive concept, the present invention provides an electronic device comprising a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; The processor is used to execute the steps of the above-described rolling bearing fault diagnosis method when running a computer program.

[0031] Based on the same inventive concept, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by at least one processor, implements the steps of the above-described rolling bearing fault diagnosis method.

[0032] I. Experimental Setup All experiments were conducted on a device equipped with an NVIDIA GeForce RTX 4090 Laptop GPU, using Python 3.9.0 and PyTorch 1.13.0. Balanced accuracy, weighted average F1 score, and precision were used as performance metrics for the model.

[0033] II. Methods and Implementation Details 1. Signal Preprocessing: After obtaining the raw one-dimensional vibration signal from the dataset, it is standardized. Each sample is truncated to a fixed length L = 1024. Subsequently, the data is split into two channels for parallel processing.

[0034] 2. Image branch construction: MTF encoding: Converts a one-dimensional signal sample of length 1024 into a 64×64 two-dimensional MTF time-frequency image.

[0035] An improved ResNet backbone: Based on ResNet-18, FDConv and CBAM modules are incorporated into each residual block. The architecture of this improved ResNet backbone (FDCA-ResNet) is as follows: Figure 4 As shown. The FDConv module (its structure is as follows) Figure 5 (As shown) The standard 3×3 convolution kernel is divided into G=4 frequency bands in the frequency domain, and frequency domain dynamic convolution is achieved through iDFT reconstruction and attention weighting. The CBAM module (its structure is shown) Figure 6 As shown, the feature map output by FDConv is then subjected to channel and spatial attention weighting in sequence to focus on key features.

[0036] 3. Timing branch construction: Signal stacking: Divide a one-dimensional signal of length 1024 into K=16 segments, each segment of length S=64, and reconstruct it into a two-dimensional sequence matrix of 16×64.

[0037] GRU Network: A two-layer GRU network is used, with 128 hidden layer units. The internal structure of the GRU unit is as follows: Figure 3 As shown, the network ultimately extracts the hidden state at the last time step. As a temporal feature vector.

[0038] Feature fusion and classification: The image branch outputs a 256-dimensional spatial feature vector, and the temporal branch outputs a 128-dimensional temporal feature vector.

[0039] The two feature vectors are directly concatenated into a 384-dimensional joint feature vector.

[0040] The joint feature vector is input into a fully connected network with 128 hidden layers, and finally the probability distribution corresponding to different fault types is output through the Softmax function.

[0041] III. Experimental Results To verify the effectiveness of the proposed model, we conducted experiments on two publicly available datasets. The BJTU-RAO experimental platform used is as follows: Figure 7 As shown, the HUST experimental platform is as follows: Figure 8 As shown in the table below, the detailed diagnostic performance on both datasets is as follows: To verify the contributions of each core component, we conducted systematic ablation experiments. The results show that the complete model exhibits optimal performance. Removing the GRU temporal branch resulted in an accuracy decrease of approximately 1.54% and 1.74% on datasets A and B, respectively; replacing FDConv with standard convolutions resulted in an accuracy decrease of approximately 2.28% and 2.49%, respectively; and removing the CBAM attention module resulted in an accuracy decrease of approximately 3.66% and 4.38%, respectively. This clearly demonstrates that the dual-branch architecture, FDConv frequency-domain convolutions, and the CBAM attention mechanism all contribute indispensable improvements to model performance, and that there is a significant synergistic effect among the three. The accuracy comparison of the five model configurations in the ablation experiments is shown below. Figure 11 As shown.

[0042] The model's detailed diagnostic performance is presented using a confusion matrix, such as... Figure 9 As shown in the figure. To more intuitively evaluate the model's feature extraction capability, we used T-SNE technology to visualize the deep features of the test set, and the results are as follows. Figure 10 As shown, the results indicate that the deep features extracted by the model of this invention exhibit clear clustering effects with high cohesion and high separation in the feature space. The feature cluster boundaries of different fault categories are distinct, which intuitively verifies the model's powerful feature learning and discrimination capabilities.

[0043] This invention can effectively integrate the static structure of time-frequency images with the dynamic temporal evolution of the original signal. Through frequency domain adaptive learning and attention focusing mechanism, it significantly improves the diagnostic accuracy and robustness of rolling bearing faults under complex working conditions, especially early and weak faults, providing an effective solution for intelligent operation and maintenance in industrial sites.

[0044] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0045] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for diagnosing rolling bearing faults, characterized in that, include: Obtain the one-dimensional vibration signal of the rolling bearing, and perform bi-branch feature extraction and fusion on the one-dimensional vibration signal: One branch is the image branch, which first encodes the one-dimensional vibration signal into a two-dimensional time-frequency image through Markov transfer field (MTF), and then uses a ResNet network that integrates frequency domain dynamic convolution (FDConv) and convolutional block attention module (CBAM) to extract spatial features from the image; One branch is the time-series branch, which first segments and stacks the one-dimensional vibration signal to form a two-dimensional sequence, and then uses a gated recurrent unit (GRU) network to extract time features from the sequence; Finally, the features extracted from the two branches (spatial features and temporal features) are fused together, and a fully connected classifier is used to diagnose the fault type.

2. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The "segmented stacking" in the timing branch specifically means: dividing the one-dimensional vibration signal x of length L into K segments, each segment of length S, K = L / S, thereby reconstructing a two-dimensional matrix of K rows and S columns as the input of the gated recurrent unit (GRU) network.

3. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The operation process of the frequency domain dynamic convolution (FDConv) is as follows: the convolution kernel weights are divided into multiple frequency band groups in the frequency domain; inverse discrete Fourier transform is performed on the coefficients of each frequency band to obtain multiple spatial convolution kernels; then weights are dynamically generated according to the input features, and these convolution kernels are weighted and fused into an adaptive dynamic convolution kernel.

4. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The Convolutional Block Attention (CBAM) module executes channel attention and spatial attention sequentially: channel attention calculates channel weights through global pooling and fully connected layers; spatial attention calculates spatial weights through pooling and convolution along the channel dimension.

5. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The feature fusion method for the two branches is to directly concatenate the spatial feature vector and the temporal feature vector.

6. A rolling bearing fault diagnosis system according to any one of claims 1 to 5, characterized in that, include: Sensors used to acquire one-dimensional vibration signals; An image branch processing module, whose input is connected to a sensor, includes an MTF encoder, an FDConv module, a ResNet network, and a CBAM module in sequence; A timing branch processing module, whose input is connected to a sensor, and which sequentially includes a signal stacking unit and a GRU network; A feature fusion module is used to connect and fuse the outputs of two branch processing modules (image branch processing module and temporal branch processing module); A fault classifier whose input is connected to a feature fusion module and whose output is the fault type.

7. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements a rolling bearing fault diagnosis method as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a rolling bearing fault diagnosis method as described in any one of claims 1 to 5.

Images