Rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT

CN118898047APending Publication Date: 2024-11-05CHONGQING UNIV OF POSTS & TELECOMM

Patent Information

Application Number
CN202410928568.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-05

Smart Images

  • Figure CN118898047A_ABST
    Figure CN118898047A_ABST
Patent Text Reader

Abstract

The invention requests to protect a rolling bearing fault diagnosis method based on multi-sensor information fusion and an efficient diagnosis visual sense transformer (EDViT) network, and belongs to the technical field of bearing fault diagnosis. Aiming at the problems of incomplete feature extraction and low diagnosis efficiency in rolling bearing fault diagnosis, the invention provides a rolling bearing fault diagnosis method based on multi-sensor information fusion and efficient diagnosis of a visual transformer. According to the method, firstly, sensor information is merged by adopting a kurtosis-based weighted fusion method, then, a fusion signal is converted into a time-frequency image by using short-time Fourier transform (STFT) to serve as input of an EDViT network model, then, local and global features are sequentially extracted through a double-attention convolution module and a double-branch patch visual transformation module of the EDViT, and finally, the local and global features of the EDViT network model are extracted. And finally, fault classification is carried out by a classifier. The rolling bearing fault diagnosis method improves the precision of rolling bearing fault diagnosis, and the provided model has strong generalization ability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of intelligent diagnosis of rolling bearing faults, and in particular relates to a rolling bearing fault diagnosis method based on multi-sensor information fusion and a high-efficiency diagnostic visual transformer. Background Art

[0002] Under the background of intelligent manufacturing, mechanical equipment is developing along the trend of integration, automation and intelligence. In the development of mechanical equipment, the structure of rotating parts is becoming more and more sophisticated, and the functions are becoming more and more complex. It is an indispensable and important part of modern equipment. Among them, rolling bearings, as basic components, play the role of motion transmission, friction reduction and force bearing, and are one of the most important components of rotating mechanical transmission devices. The good operating state of rolling bearings plays a vital role in the smooth operation of the entire equipment. Therefore, it is very valuable to study the intelligent diagnosis method of rolling bearing faults.

[0003] Although traditional deep learning algorithms have achieved great success in fault diagnosis, they also have inherent limitations. Convolutional neural networks (CNNs) usually have difficulty capturing global information, and recurrent neural networks (RNNs) often face problems such as gradient vanishing or exploding, difficulty maintaining long-term memory, inability to parallelize, and slow training. With the further development of deep learning, the Transformer model came into being. Transformer relies on its unique multi-head self-attention (MSA) mechanism to achieve high parallelism and computational efficiency. It can capture long-distance dependencies and realize global information interaction. However, due to limited data samples or poor data quality, training the Transformer model is very difficult and prone to overfitting.

[0004] On the other hand, due to the increasing complexity of the internal structure of rotating machinery, multiple sensors need to be installed at different locations to capture information with different sensitivities and complementarities to ensure stable operation. The normal and fault signals of the bearing will overlap and modulate each other. However, most bearing fault diagnosis methods rely on single sensor data, which often cannot meet the needs of complex systems.

[0005] After searching, the application publication number CN118150167A, a method for analyzing the life evolution of aviation bearings based on multi-sensor information fusion, the existing methods for the life evolution of aviation bearings are relatively simple, and it is difficult to evolve through multi-sensor information fusion, and the effect is not good enough. At the same time, the existing methods are difficult to analyze the vibration signals collected by the sensors, and cannot correctly determine the cause of the fault. S1, through multiple groups of sensors to collect the parameters and vibration frequency of the aviation bearing during operation, at the same time establish a monitoring signal database under the operation state of the aviation bearing, and divide the monitoring signal database into a training set and a test set. This method integrates diversified information collection, can perform life evolution through multi-sensor information fusion, the evolution efficiency has been significantly improved, and can analyze the vibration signals collected by the sensors, greatly improving the accuracy of fault diagnosis, and facilitating the subsequent timely maintenance and replacement of corresponding accessories.

[0006] CN118150167A, a method for analyzing the life evolution of aviation bearings based on multi-sensor information fusion, does adopt diversified information collection as one, and integrates multi-sensor information. The calculation process based on wavelet transform and mixed Gaussian distribution model is relatively complicated, and the data volume requirement is high, and sufficient data is needed to train and verify the model. This model-based method feature extraction is not comprehensive, and often requires a lot of expert experience and domain knowledge to build and tune the model. This dependence increases the complexity and cost of the diagnosis process. At the same time, this method is only designed for aviation bearing fault types, and its generalization ability is limited, and its adaptability to dynamic working conditions is poor. When applied to different types of bearings or new fault modes, the dynamically changing working environment (such as load, speed) will affect the accuracy and reliability of fault feature extraction, and the effect of the model may be greatly reduced.

[0007] In order to solve the problem of incomplete feature extraction and poor generalization of traditional rolling bearing fault detection methods, this patent first uses the kurtosis weighted fusion method to fuse information from different sensors and make up for the completeness of the information. Secondly, the short-time Fourier transform is used to process the fused signal. This processing method can extract non-stationary and nonlinear features. Finally, combining the local feature extraction capability of the convolutional neural network and the global feature extraction capability of the visual transformer, an efficient diagnostic visual transformer network is proposed, which improves the accuracy of rolling bearing fault diagnosis. The proposed method has strong generalization ability. Summary of the invention

[0008] The present invention aims to solve the above problems of the prior art. A rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT is proposed. The technical solution of the present invention is as follows:

[0009] A rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT, comprising the following steps:

[0010] Step S1: using different sensors to collect original time domain signals of the bearing health and fault status to obtain a number of signal samples, and using a data fusion method based on kurtosis weighting to process the data samples from different sensors to obtain multi-sensor information fusion samples;

[0011] Step S2: Use STFT to convert the fused data samples into time-frequency images, then resize the time-frequency images into 224×224 RGB images and divide them into training set, validation set and test set according to the proportion;

[0012] Step S3: Use the training set to fully train the established network model, use the validation set to adjust and save the hyperparameters of the model, and then test it on the test set to obtain fault diagnosis results.

[0013] Furthermore, the step S1 uses different sensors to collect original time domain signals of the bearing health and fault status to obtain a number of signal samples, and uses a data fusion method based on kurtosis weighting to process the data samples from different sensors to obtain multi-sensor information fusion samples; specifically includes the following steps:

[0014] Step S1.1: Calculate the kurtosis values ​​of the original vibration signals of length N collected by m sensors respectively. The kurtosis formula is as follows:

[0015]

[0016] Among them, K i represents the kurtosis value of the i-th sensor, n represents the number of signal sample points, x i (n) represents the original one-dimensional vibration signal data, μ represents x i The mean of (n);

[0017] Step S1.2: m original vibration signals are weighted and fused according to the kurtosis values ​​of different sensors. The fused signal is expressed as follows:

[0018]

[0019] Among them, x(n) represents the fused signal data, K i Represents the kurtosis value of the i-th sensor.

[0020] Furthermore, the step S2 uses short-time Fourier transform STFT to convert the fused data samples into a time-frequency image, then adjusts the time-frequency image to a 224×224 RGB image, and divides it into a training set, a validation set, and a test set in proportion, specifically comprising the following steps:

[0021] Step S2.1: Perform sliding window overlapping sampling on the fusion signal of length N according to m data points as a group, with an overlap rate of r, and each sample contains m sampling points;

[0022] Therefore, the number of samples obtained is

[0023] Step S2.2: Use the Hanning window as the window function to perform STFT on each sample and extract the corresponding time-frequency diagram. The window width is set to 256.

[0024] Further, the steps of short-time Fourier transform STFT are as follows:

[0025] STFT divides each sample signal into many short time periods by using a window function, and performs Fourier transform on each short time period to effectively extract the time domain and frequency domain information of the non-stationary vibration signal.

[0026] The window function is expressed as ω(t). In order to transform the window function, the time variable τ is introduced to represent the center position of the current window. The transformed window function is expressed as ω(t-τ).

[0027] The Fourier transform formula is as follows:

[0028]

[0029] Where x(t) is a one-dimensional time domain signal, w is a frequency variable, and exp(-jwt) represents a complex conjugate operation;

[0030] Based on the above process, the STFT result of the fused signal sample can be given by the following integral:

[0031]

[0032] Where X(t,f) represents the amplitude value at time t and frequency f;

[0033] Step S2.3: Convert the time-frequency diagram obtained by STFT into a 224×224 RGB image and divide it into training set, validation set, and test set in a ratio of 6:2:2.

[0034] Furthermore, the network structure of the efficient diagnosis visual transformer in step S3 is specifically as follows:

[0035] Stage 1: Build a downsampling layer with a size of 4×4, 48 channels, and a stride of 4;

[0036] Stage 2: Build the dual attention convolution basic unit DAConv1, including:

[0037] (1) Depthwise separable convolution with size 7×7, number of channels 48, and stride 1;

[0038] (2) 1×1 convolution with size 1×1, number of channels 192, and stride 1;

[0039] (3) 1×1 convolution with size 1×1, number of channels 48, and stride 1;

[0040] (4) Dual Attention Mechanism

[0041] Stage 3: Build a downsampling layer with a size of 2×2, 96 channels, and a stride of 2;

[0042] Stage 4: Build the dual attention convolution basic unit DAConv2, including:

[0043] (1) Depthwise separable convolution with size 7×7, 96 channels, and stride 1;

[0044] (2) 1×1 convolution with size 1×1, number of channels 384, and stride 1;

[0045] (3) 1×1 convolution with size 1×1, number of channels 96, and stride 1;

[0046] (4) Dual Attention Mechanism

[0047] Stage 5: Build a downsampling layer with a size of 2×2, 192 channels, and a stride of 2;

[0048] Stage 6: Build the dual attention convolution basic unit DAConv3, including:

[0049] (1) Depthwise separable convolution with size 7×7, number of channels 192, and stride 1;

[0050] (2) 1×1 convolution with size 1×1, number of channels 768, and stride 1;

[0051] (3) 1×1 convolution with size 1×1, number of channels 192, and stride 1;

[0052] (4) Dual Attention Mechanism

[0053] Stage 7: Build a dual-branch visual transformer basic unit DPViT, where both branches contain a linear embedding layer, a position encoding layer, a cross transform encoder layer, a category label extraction layer, and a multi-layer perceptron layer.

[0054] Furthermore, the linear embedding layer uses patches of size 1 and 2 to convert the output feature map of Stage 6 into vectors of size 49×768 and 196×192 as patch labels;

[0055] On this basis, the position encoding layer introduces classification tags, adds position information to each vector, retains the spatial structure of the input feature map, and the merged vector sizes are 50×768 and 197×192 respectively;

[0056] The cross-transform encoder layer includes a multi-head cross-attention mechanism MCA, a layer norm LN, and a multi-layer perceptron MLP; MCA is used for the information of the entire two branches, LN normalizes the data of each channel in the batch, and MLP is used to feature map and transform the information at each position in the output sequence.

[0057] The category label extraction layer extracts the two category label vectors output by the cross transform encoder layer, and then converts the two category label vectors into 1×num_classes size through the multi-layer perceptron layer, where num_classes is the number of classification categories.

[0058] Finally, the classifier is used to weight the output results of the two branches to obtain the classification result.

[0059] Furthermore, the multi-head cross attention mechanism MCA mechanism operates as follows:

[0060] The classification mark of the small branch interacts with the patch mark of the large branch as a query mark, thereby integrating the semantic information of the large branch; conversely, the classification mark of the large branch interacts with the patch mark of the small branch to integrate the semantic information of the small branch; in order to match the dimension of the input sequence, the sequence sum of the two branches needs to be aligned through projection to obtain the sequence sum; the classification mark is processed by the cross attention module to obtain the integrated information; MCA superimposes the h-head cross attention CA to capture the multi-scale information of the feature map, and CA is defined as follows:

[0061]

[0062] Among them, Q, K, V are Query, Key and Value matrices respectively, D and h are the dimensions of the input vector and the number of heads of the attention mechanism respectively, and x cls and x are the aligned projections of the classification mark and patch mark, respectively, W q , W k , W v They are the learning weights of the Query, Key and Value matrices respectively.

[0063] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT as described in any one of the above is implemented.

[0064] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT as described in any one of the above.

[0065] The advantages and beneficial effects of the present invention are as follows:

[0066] 1) Claim 2 addresses the problem that most bearing fault diagnosis methods rely on single sensor data and cannot meet the needs of complex systems. It proposes a kurtosis weighted method to fuse the original vibration signals of different sensors to obtain information of different sensitivities and complementary information, thereby ensuring the integrity of the information and improving the accuracy and reliability of fault diagnosis of complex mechanical systems.

[0067] 2) Claims 3 and 4 use STFT to convert the fused signal into a time-frequency diagram, which can integrate the time domain and frequency domain information, effectively process non-stationary vibration signals, and facilitate robust feature extraction.

[0068] 3) Claims 5, 6 and 7 address the problems of low accuracy and efficiency in rolling bearing fault diagnosis by combining the advantages of CNN and visual transformers to propose an efficient diagnostic visual transformer network. The dual-attention convolution module integrates the effective channel attention mechanism and the position attention mechanism to capture information of different scales from different dimensions. The dual-branch patch visual transformation module focuses on different features from different angles through a multi-head cross-attention mechanism, comprehensively extracts feature information from the input sequence, and improves diagnostic efficiency and classification recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a schematic diagram of a rolling bearing fault diagnosis process provided by a preferred embodiment of the present invention;

[0070] Figure 2 It is a diagram of the framework of the efficient diagnostic visual transformer network;

[0071] Figure 3 Schematic diagram of the dual attention mechanism in the efficient diagnostic visual transformer network;

[0072] Figure 4 is a schematic diagram of the cross-transform encoder in the efficient diagnostic vision transformer network;

[0073] Figure 5is the t-SNE visualization result in the embodiment of the present invention;

[0074] Figure 6 It is a comparison chart of the accuracy of cross-domain experiments in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0076] The technical solution of the present invention to solve the above technical problems is:

[0077] Aiming at the problems of incomplete feature extraction, low diagnostic efficiency and unreliable diagnostic results in the rolling bearing fault diagnosis process, a rolling bearing fault diagnosis method based on multi-sensor information fusion and efficient diagnostic visual transformer is proposed. The technical solution of the present invention is as follows:

[0078] A rolling bearing fault diagnosis method based on multi-sensor information fusion and efficient diagnostic visual transformer. The process framework of this method is as follows Figure 1 As shown, the specific steps are as follows:

[0079] Step S1, under different working conditions, using different sensors to collect the original time domain vibration signals of the bearing health and fault status, to obtain a number of signal samples, using a data fusion strategy based on kurtosis weighting to process the data samples from different sensors, to obtain samples of multi-sensor information fusion.

[0080] The experimental data used in this embodiment is the bearing data set of Case Western Reserve University (CWRU). The experimental device consists of a 1.5 kW motor, a torque sensor and a power meter. The torque sensor adjusts the motor speed by adjusting the load. Accelerometers are installed at the drive end and the fan end to collect operating data at different positions of the bearing. The bearing model used is SKF6205, and the sampling frequency is 12KHz. Data is collected under three different load conditions: 1hp, 2hp and 3hp, and the corresponding motor speeds are 1772, 1750 and 1730RPM respectively. The bearing produces three types of faults through electrical discharge machining: pitting of the inner ring, rolling element and outer ring. Each fault type is divided into three severity levels (fault diameter: 0.007 inches, 0.014 inches, 0.021 inches).

[0081] The collected original data from the drive end and the fan end are fused through the kurtosis weighted fusion strategy to obtain fused data samples of three different working conditions.

[0082] In step S2, the fused data samples are converted into time-frequency images using STFT, and then the time-frequency images are resized to 224×224 RGB images and divided into training sets, validation sets, and test sets in proportion.

[0083] The fusion samples of three different working conditions are segmented using a sliding window of size m = 1024, with an overlap rate of r = 0.5. Each segmented sample contains 1024 data points. Different fault types contain 236 to 238 data samples. For the sake of uniformity, the first 230 samples are selected for the experiment.

[0084] The Hanning window is used as the window function, and the STFT is performed on the 230 samples after segmentation to extract the corresponding time-frequency diagram. The window width is set to 256. The steps of short-time Fourier transform are as follows:

[0085] STFT divides each sample signal into many short time periods by using a window function, and performs Fourier transform on each short time period, thereby effectively extracting the time domain and frequency domain information of the non-stationary vibration signal.

[0086] The window function is expressed as ω(t). In order to transform the window function, a time variable τ is introduced to represent the center position of the current window. The transformed window function is expressed as ω(t-τ).

[0087] The Fourier transform formula is as follows:

[0088]

[0089] Where x(t) is a one-dimensional time domain signal, w is a frequency variable, and exp(-jwt) represents a complex conjugate operation.

[0090] Based on the above process, the STFT result of the fused signal sample can be given by the following integral:

[0091]

[0092] Where X(t,f) represents the amplitude value at time t and frequency f.

[0093] The time-frequency graph obtained by STFT is converted into a 224×224 RGB image to obtain three different working condition data sets A, B, and C, corresponding to three different load conditions of 1hp, 2hp, and 3hp. Then the training set, validation set, and test set are divided into a ratio of 6:2:2. The composition of the data set is shown in Table 1.

[0094] Table 1 Composition of experimental dataset

[0095]

[0096] Step S3, building an efficient diagnostic visual transformer network, using the training set to fully train the established network model, using the validation set to adjust and save the hyperparameters of the model, and then testing it on the test set to obtain the fault diagnosis results.

[0097] Efficient diagnostic vision transformer network structure such as Figure 2 As shown, specifically:

[0098] Stage 1: Build a downsampling layer with a size of 4×4, 48 channels, and a stride of 4;

[0099] Stage 2: Build the dual attention convolution basic unit DAConv1, including:

[0100] (1) Depthwise separable convolution with size 7×7, number of channels 48, and stride 1

[0101] (2) 1×1 convolution with size 1×1, number of channels 192, and stride 1

[0102] (3) 1×1 convolution with size 1×1, number of channels 48, and stride 1

[0103] (4) Dual attention mechanism, its structure is as follows Figure 3 Shown

[0104] Stage 3: Build a downsampling layer with a size of 2×2, 96 channels, and a stride of 2;

[0105] Stage 4: Build the dual attention convolution basic unit DAConv2, including:

[0106] (1) Depthwise separable convolution with size 7×7, number of channels 96, and stride 1

[0107] (2) 1×1 convolution with size 1×1, number of channels 384, and stride 1

[0108] (3) 1×1 convolution with size 1×1, number of channels 96, and stride 1

[0109] (4) Dual Attention Mechanism

[0110] Stage 5: Build a downsampling layer with a size of 2×2, 192 channels, and a stride of 2;

[0111] Stage 6: Build the dual attention convolution basic unit DAConv3, including:

[0112] (1) Depthwise separable convolution with size 7×7, number of channels 192, and stride 1

[0113] (2) 1×1 convolution with size 1×1, number of channels 768, and stride 1

[0114] (3) 1×1 convolution with size 1×1, number of channels 192, and stride 1

[0115] (4) Dual Attention Mechanism

[0116] Stage 7: Build a dual-branch visual transformer basic unit DPViT, where both branches contain a linear embedding layer, a position encoding layer, a cross transform encoder layer, a category label extraction layer, and a multi-layer perceptron layer.

[0117] The linear embedding layer uses patch sizes of 1 and 2 to convert the output feature maps of Stage 6 into vectors of sizes 49×768 and 196×192 as patch labels, respectively.

[0118] On this basis, the position encoding layer introduces classification tags, adds position information to each vector, retains the spatial structure of the input feature map, and the merged vector sizes are 50×768 and 197×192 respectively.

[0119] The cross transform encoder layer structure is as follows Figure 4 As shown, it includes a multi-head cross attention mechanism (MCA), layer norm (LN) and a multi-layer perceptron (MLP). MCA is used for the information of the entire two branches, LN normalizes the data of each channel in the batch, and MLP is used to feature map and transform the information at each position in the output sequence. The proposed MCA mechanism operates as follows:

[0120] The classification mark of the small branch interacts with the patch mark of the large branch as a query mark, thereby integrating the semantic information of the large branch. Conversely, the classification mark of the large branch interacts with the patch mark of the small branch to integrate the semantic information of the small branch. In order to match the dimension of the input sequence, the sequence sum of the two branches needs to be aligned through projection to obtain the sequence sum. The classification mark is processed by the cross attention module to obtain the integrated information. MCA superimposes the h-head cross attention CA to capture the multi-scale information of the feature map. CA is defined as follows:

[0121]

[0122] Among them, Q, K, V are Query, Key and Value matrices respectively, D and h are the dimensions of the input vector and the number of heads of the attention mechanism respectively, and x cls and x are the aligned projections of the classification mark and patch mark, respectively, W q , W k , W v They are the learning weights of the Query, Key and Value matrices respectively.

[0123] The category label extraction layer extracts the two category label vectors output by the cross transform encoder layer, and then converts the two category label vectors into 1×10 size through the multi-layer perceptron layer.

[0124] Finally, the classifier is used to weight the output results of the two branches to obtain the classification result.

[0125] The specific parameters of the network structure are shown in Table 2.

[0126] Table 2 EDViT network structure parameters

[0127]

[0128]

[0129] Note: Patch Embedding means linear embedding, Position Encode means position encoding, CrossTransformer Encoder means cross transform encoder, Class Token Extraction means classification token extractor, MLP Head means multi-layer perceptron head

[0130] To verify the effectiveness of the proposed method, the present invention is verified on the CWRU dataset. The data used in the experiment come from vibration data under three different working conditions.

[0131] First, the effectiveness of multi-sensor information fusion is verified using dataset A. The time-frequency graph is used as the input of the EDViT network for full training, the hyperparameters of the model are adjusted and saved using the validation set, and then tested on the test set to obtain the fault diagnosis results. Table 3 shows the test accuracy of different sensors obtained in five repeated tests. The results show that under 1hp conditions, there are significant differences in the training accuracy (95.80%, 97.03%) and test accuracy (98.70%, 98.87%) of the drive-end and fan-end data. The diagnosis results obtained by relying on single sensor data are not reliable. Compared with single sensor data, the fused data of the two sensors achieved higher training accuracy (97.32%) and test accuracy (99.83%) on the EDViT model. This shows that the fusion information synergy ensures the complementarity of the data and provides more comprehensive and richer information, thereby enhancing the integrity and stability of the data and making the diagnosis results more accurate and reliable.

[0132] Table 3 Test accuracy of different sensor data

[0133]

[0134] Table 4 Test accuracy of different models

[0135]

[0136] In order to evaluate the superiority of the EDViT model, this example compares it with Vision Transformer (ViT), Data-efficient Image Transformer (DeiT), MobileNet, and ResNet. The test accuracy on dataset A is shown in Table 4, and the t-distributed Stochastic Neighbor Embedding (t-SNE) visualization classification results are shown in Table 4. Figure 5 The results show that the EDViT model has excellent feature extraction capabilities, fast convergence speed and high diagnostic accuracy. The classification accuracy of the proposed model is higher than that of other models, and the diagnostic results are more stable.

[0137] To further illustrate the generalization performance of the EDViT model, we conducted a cross-domain experiment. We selected a dataset for training and retained its training parameters, and then tested and validated it on the other two datasets. Six groups of tests (AB, AC, BA, BC, CA, CB) were conducted for different models, and each group of tests was repeated five times and the average was taken. The test results are shown in Figure 2. Figure 6 As shown. Figure 6 It can be seen that under different conditions, the domain generalization performance of the EDViT model exceeds that of other comparison models, with an average accuracy of 95.17%, which is 8.51%, 6.89%, 4.68% and 3.69% higher than other models respectively. Comparison with other methods shows that the EDViT model has strong generalization ability.

[0138] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0139] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. 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 technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic 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 temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0140] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0141] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT, characterized in that: The following steps are involved: Step S1: using different sensors to collect original time domain signals of the bearing health and fault status to obtain a number of signal samples, and using a data fusion method based on kurtosis weighting to process the data samples from different sensors to obtain multi-sensor information fusion samples; Step S2: Use STFT to convert the fused data samples into time-frequency images, then resize the time-frequency images into 224×224 RGB images and divide them into training set, validation set and test set according to the proportion; Step S3: Use the training set to fully train the established network model, use the validation set to adjust and save the hyperparameters of the model, and then test it on the test set to obtain fault diagnosis results.

2. The rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT according to claim 1 is characterized in that: The step S1 uses different sensors to collect the original time domain signals of the bearing health and fault status to obtain a number of signal samples, and uses a data fusion method based on kurtosis weighting to process the data samples from different sensors to obtain samples of multi-sensor information fusion; specifically, it includes the following steps: Step S1.1: Calculate the kurtosis values ​​of the original vibration signals of length N collected by m sensors respectively. The kurtosis formula is as follows: Among them, K i represents the kurtosis value of the i-th sensor, n represents the number of signal sample points, x i (n) represents the original one-dimensional vibration signal data, μ represents x i The mean of (n); Step S1.2: m original vibration signals are weighted and fused according to the kurtosis values ​​of different sensors. The fused signal is expressed as follows: Among them, x(n) represents the fused signal data, K i Represents the kurtosis value of the i-th sensor.

3. The rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT according to claim 2 is characterized in that: The step S2 uses short-time Fourier transform STFT to convert the fused data samples into a time-frequency image, then adjusts the time-frequency image to a 224×224 RGB image, and divides it into a training set, a validation set, and a test set in proportion, specifically comprising the following steps: Step S2.1: Perform sliding window overlapping sampling on the fusion signal of length N, with m data points as a group, and the overlapping rate is r. Each sample contains m sampling points. Therefore, the number of samples obtained is Step S2.2: Use the Hanning window as the window function to perform STFT on each sample and extract the corresponding time-frequency diagram. The window width is set to 256.

4. The rolling bearing fault diagnosis method based on claim 3 is characterized in that: The steps of short-time Fourier transform STFT are as follows: STFT divides each sample signal into many short time periods by using a window function, and performs Fourier transform on each short time period to effectively extract the time domain and frequency domain information of the non-stationary vibration signal. The window function is expressed as ω(t). In order to transform the window function, the time variable τ is introduced to represent the center position of the current window. The transformed window function is expressed as ω(t-τ). The Fourier transform formula is as follows: Where x(t) is a one-dimensional time domain signal, w is a frequency variable, and exp(-jwt) represents a complex conjugate operation; Based on the above process, the STFT result of the fused signal sample can be given by the following integral: Where X(t,f) represents the amplitude value at time t and frequency f; Step S2.3: Convert the time-frequency diagram obtained by STFT into a 224×224 RGB image and divide it into training set, validation set, and test set in a ratio of 6:2:

2.

5. The rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT according to claim 4 is characterized in that: The network structure of the efficient diagnosis visual transformer in step S3 is specifically as follows: Stage 1: Build a downsampling layer with a size of 4×4, 48 channels, and a stride of 4; Stage 2: Build the dual attention convolution basic unit DAConv1, including: (1) Depthwise separable convolution with size 7×7, number of channels 48, and stride 1; (2) 1×1 convolution with size 1×1, number of channels 192, and stride 1; (3) 1×1 convolution with size 1×1, number of channels 48, and stride 1; (4) Dual Attention Mechanism Stage 3: Build a downsampling layer with a size of 2×2, 96 channels, and a stride of 2; Stage 4: Build the dual attention convolution basic unit DAConv2, including: (1) Depthwise separable convolution with size 7×7, 96 channels, and stride 1; (2) 1×1 convolution with size 1×1, number of channels 384, and stride 1; (3) 1×1 convolution with size 1×1, number of channels 96, and stride 1; (4) Dual Attention Mechanism Stage 5: Build a downsampling layer with a size of 2×2, 192 channels, and a stride of 2; Stage 6: Build the dual attention convolution basic unit DAConv3, including: (1) Depthwise separable convolution with size 7×7, number of channels 192, and stride 1; (2) 1×1 convolution with size 1×1, number of channels 768, and stride 1; (3) 1×1 convolution with size 1×1, number of channels 192, and stride 1; (4) Dual Attention Mechanism Stage 7: Build a dual-branch visual transformer basic unit DPViT, where both branches contain a linear embedding layer, a position encoding layer, a cross transform encoder layer, a category label extraction layer, and a multi-layer perceptron layer.

6. The rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT according to claim 5, characterized in that: The linear embedding layer uses patches of size 1 and 2 to convert the output feature map of Stage 6 into vectors of size 49×768 and 196×192 as patch labels; On this basis, the position encoding layer introduces classification tags, adds position information to each vector, retains the spatial structure of the input feature map, and the merged vector sizes are 50×768 and 197×192 respectively; The cross-transform encoder layer includes a multi-head cross-attention mechanism MCA, a layer norm LN, and a multi-layer perceptron MLP; MCA is used for the information of the entire two branches, LN normalizes the data of each channel in the batch, and MLP is used to feature map and transform the information at each position in the output sequence. The category label extraction layer extracts the two category label vectors output by the cross transform encoder layer, and then converts the two category label vectors into 1×num_classes size through the multi-layer perceptron layer, where num_classes is the number of classification categories. Finally, the classifier is used to weight the output results of the two branches to obtain the classification result.

7. The rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT according to claim 6 is characterized in that: The multi-head cross attention mechanism MCA mechanism operates as follows: The classification mark of the small branch interacts with the patch mark of the large branch as a query mark, thereby integrating the semantic information of the large branch; conversely, the classification mark of the large branch interacts with the patch mark of the small branch to integrate the semantic information of the small branch; in order to match the dimension of the input sequence, the sequence sum of the two branches needs to be aligned through projection to obtain the sequence sum; the classification mark is processed by the cross attention module to obtain the integrated information; MCA superimposes the h-head cross attention CA to capture the multi-scale information of the feature map, and CA is defined as follows: Among them, Q, K, V are Query, Key and Value matrices respectively, D and h are the dimensions of the input vector and the number of heads of the attention mechanism respectively, and x cls and x are the aligned projections of the classification mark and patch mark, respectively, W q , W k , W v They are the learning weights of the Query, Key and Value matrices respectively.

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the rolling bearing fault diagnosis method based on multi-sensor information fusion and EDViT as described in any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Aviation bearing life evolution analysis method based on multi-sensor information fusion

    CN118150167A

Cited By

  • ViT cross-domain fault diagnosis method combining adaptive space attention and patch feature weighting

    CN119475102A

  • Wind turbine generator variable pitch bearing fault diagnosis method based on signal coding

    CN119646607A

  • Rolling bearing fault diagnosis method based on time-frequency fusion and double-branch deep network

    CN120992200A

  • Bearing fault diagnosis method based on fusion of cavity convolution and attention mechanism

    CN121431068A

  • RV reducer fault diagnosis method based on WGAN-RAG and SCViT multi-modal information fusion

    CN121524858A