An ECA_ResNet-based rolling bearing fault diagnosis method

Abstract

The application discloses a kind of based on ECA_ResNet Rolling bearing fault diagnosis method, specific diagnosis method is: first, one-dimensional vibration signal is converted into three-channel color time-frequency chart by using short-time Fourier and pseudo-color processing method combination, time-frequency chart is used as input sample set;Second, ECA_ResNet bearing fault diagnosis model is constructed, the model includes convolution layer, maximum pooling layer, improved residual module and fully connected layer, shallow edge feature extraction and compression are carried out to data based on convolution layer and pooling layer, network depth is increased by constructing improved residual module, and Dropout is introduced between two connection layers to inhibit overfitting;Then, ECA_ResNet bearing fault diagnosis model is trained, after the sample of inputting into the ECA_ResNet bearing fault diagnosis model that well-trained after joining Gaussian white noise, diagnosis result is obtained.The diagnosis method can not only solve the problem that gradient dispersion, explosion and network degradation will be caused by too high network depth, but also can improve the bearing fault diagnosis accuracy of model under strong noise and variable load.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, specifically to a rolling bearing based on ECA_ResNet.

[0002] Fault diagnosis methods. Background Technology

[0003] With the rapid advancement of industrial manufacturing, the bearing industry is also developing at a breakneck pace. As a core component in equipment such as aerospace, high-end machine tools, and advanced medical devices, bearings play a crucial role in the operation of the entire mechanical equipment. Problems encountered by bearings during operation, such as foreign objects, load impacts, and poor lubrication, exacerbate equipment wear, reduce bearing life, and have a significant impact on the safety and stability of the entire rotating machinery. Currently, about half of rotating machinery failures are related to rolling bearings. Therefore, developing a highly accurate and robust bearing fault diagnosis method is of profound significance for improving industrial production efficiency, reducing costs, and eliminating safety hazards.

[0004] Currently, deep learning has been widely applied in the field of bearing fault diagnosis. Its biggest feature is adaptive learning of data features, eliminating the shortcomings of manual feature extraction and improving the efficiency and accuracy of rolling bearing fault diagnosis. Eren et al. (EREN L, INCE T, KIRANY AZ SA generic intelligent bearing fault diagnosis system using compact adaptive 1D CNN classifier[J]. Journal of Signal Processing Systems, 2019, 91(2):179-189) proposed an adaptive 1D CNN bearing fault diagnosis model with a compact structure and good diagnostic performance. Zhang et al. (Zhang W, Peng G, Li C, et al. A New Deep Learning Model for Fault Diagnosis with Good Anti-Noise and Domain Adaptation Ability on Raw Vibration Signals[J]. Sensors, 2017, 17(2):425.) designed a deep convolutional neural network WDCNN with "wide convolutional kernels in the first layer and small convolutional kernels in the remaining layers" to extract short-time features of the signal and suppress high-frequency noise. The above methods are all based on one-dimensional signals for bearing fault diagnosis. However, one-dimensional signals are time-series signals, and feature value loss often occurs when extracting features from non-stationary signals, reducing diagnostic accuracy. The weight-sharing and local connectivity characteristics of Convolutional Neural Networks (CNNs) make them more suitable for feature extraction and classification of two-dimensional images.Wen et al. (WEN L, Li X, GAO L, et al. A new convolutional neural network-based data-driven fault diagnosis method [J]. IEEE Transactions on Industrial Electronics, 2017, 65(7):5990-5998) converts the signal into a grayscale image and inputs it into a convolutional neural network model for fault classification; Zhao Xiaoqiang et al. (Zhao Xiaoqiang, Zhang Qingqing. Improved AlexNet method for fault diagnosis of rolling bearing under varying working conditions[J]. Vibration. Testing and Diagnosis, 2020, 40(03):472-480+623) convert the signal into an image using transverse interpolation and improves upon AlexNet to improve the fault diagnosis accuracy under varying working conditions; Wu Chenfang et al. (Wu Chenfang, Yang Shixi, Huang Haizhou et al. Research on a fault diagnosis method for rolling bearing based on an improved LeNet-5 model[J]. Vibration and Shock, 2021, 40(12):55-61) convert the signal into a grayscale image using signal stacking and improves upon LeNet5 to enhance the model's generalization ability.

[0005] Although the above bearing fault diagnosis methods have achieved good results, there are still shortcomings in terms of model structure and performance: First, the feature extraction capability of convolutional neural networks is affected by the network depth. However, excessive depth can lead to problems such as gradient vanishing, explosion, and network degradation, affecting the accuracy of the model. Second, in actual industrial production, environmental noise and bearing load are variable, and the vibration signals acquired by sensors are inevitably interfered with, increasing the difficulty of extracting key features of the model. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a rolling bearing fault diagnosis method based on ECA_ResNet.

[0007] This fault diagnosis method not only solves the problems of gradient vanishing, explosion, and network degradation caused by excessive network depth, but also improves the accuracy of bearing fault diagnosis under strong noise and variable load.

[0008] To achieve the above objectives, the specific solution adopted by the present invention is as follows:

[0009] A rolling bearing fault diagnosis method based on ECA_ResNet mainly includes the following steps:

[0010] Step 1: Collect one-dimensional vibration signals of rolling bearings under different conditions, and convert the one-dimensional vibration signals into three-channel color time-frequency maps using short-time Fourier transform and pseudo-color processing methods. Summarize all three-channel color time-frequency maps into a sample set, and randomly divide the samples in the sample set to obtain a training set, a validation set, and a test set. The sample ratio of the training set, validation set, and test set is 7:2:1.

[0011] Step 2: Construct the ECA_ResNet bearing fault diagnosis model. The ECA_ResNet bearing fault diagnosis model includes convolutional layers, max pooling layers, improved residual modules, and fully connected layers. The convolutional layers are used to obtain shallow edge features of the sample image. The max pooling layer reduces the load and number of parameters by downsampling the feature map, while retaining the main features of the data. The improved residual module is used to deepen the number of network layers and adaptively obtain deep key information. The fully connected layer is used to arrange the features learned by the convolutional and pooling layers into a one-dimensional vector, and then the softmax function is used to classify the collected features to achieve fault diagnosis.

[0012] Step 3: Input the color time-frequency image of the training set into the ECA_ResNet bearing fault diagnosis model for training, and use the validation set to adjust the model hyperparameters until the network converges. Save the optimal model, which is the ECA_ResNet bearing fault diagnosis model with the best diagnostic effect.

[0013] Step 4: Add Gaussian white noise with different signal-to-noise ratios to the test set to simulate different levels of environmental noise in actual industrial production, test the robustness of the model, and input the test set samples with added Gaussian white noise into the ECA_ResNet bearing fault diagnosis model with the best diagnostic effect obtained in Step 3 to obtain the final fault diagnosis result.

[0014] Furthermore, in step 1, the one-dimensional vibration signal of the rolling bearing under different conditions includes the one-dimensional vibration signal of the bearing under normal, inner ring fault, outer ring fault, and rolling element fault conditions.

[0015] Furthermore, in step 1, the one-dimensional vibration signal is sequentially truncated into several short signals of fixed length using overlapping sampling. The signals are then processed using short-time Fourier transform to analyze their time-domain and frequency-domain characteristics, generating a single-channel grayscale image.

[0016] Furthermore, in step 1, the pseudo-color processing method uses pcolormesh from matplotlib, which can map single-channel grayscale values ​​to a three-channel color image.

[0017] Furthermore, in step 2, the improved residual module is to embed the effective channel attention network ECA-Net structure into the residual network module.

[0018] Furthermore, the computation process of the effective channel attention network ECA-Net is as follows: First, the features are aggregated by global average pooling to obtain global channel information. Then, 1D convolution with a kernel size of k is used to replace the fully connected layer to generate channel weights and obtain the interdependencies between channels. Finally, the original input features and channel weights are multiplied by a dot product to obtain features with channel attention.

[0019] Furthermore, the ECA_ResNet bearing fault diagnosis model constructed in step 2) contains two fully connected layers. By introducing the Dropout technique between the two fully connected layers, some neurons are made to pause working with a certain probability, thereby suppressing model overfitting.

[0020] Furthermore, in step 3), when training the ECA_ResNet bearing fault diagnosis model, the color time-frequency image is randomly vertically flipped, the batch size is 32, the number of iterations is 30, the cross-entropy loss is used as the error loss function, Adam is the optimizer, the learning rate is 0.001, and the learning rate decay mechanism StepLR is used to adjust the learning rate, which is halved every 10 iterations.

[0021] Furthermore, in step 4), Gaussian white noise of 2dB, 4dB, 6dB, 8dB, and 10dB is added to the test set.

[0022] Beneficial effects:

[0023] This invention utilizes a combination of short-time Fourier transform and pseudo-color processing to obtain a sample set of color time-frequency maps. While acquiring the time-frequency domain features of the original data, it transforms single-channel grayscale images into three-channel color time-frequency maps, enhancing the information features of the sample data. The ECA_ResNet bearing fault diagnosis model constructed in this invention embeds an effective channel attention mechanism into the residual network module, solving the gradient vanishing and network degradation phenomena caused by increasing network depth. Simultaneously, it strengthens key channels and weakens irrelevant channels, achieving cross-channel interaction and adaptively predicting deep key features, enhancing the model's noise resistance and generalization ability. Furthermore, it introduces the Dropout technique to suppress model overfitting, enhancing model generalization and improving the accuracy of bearing fault diagnosis. Attached Figure Description

[0024] Figure 1 This is a flowchart of the bearing fault diagnosis process in this invention.

[0025] Figure 2 This is a structural diagram of the ECA_ResNet bearing fault diagnosis model in this invention.

[0026] Figure 3This is a structural diagram of the improved residual module of the present invention.

[0027] Figure 4 This is a structural diagram of the residual module.

[0028] Figure 5 This is a diagram of an effective channel attention network structure.

[0029] Figure 6 Time-frequency diagrams for different fault types.

[0030] Figure 7 This is a schematic diagram comparing the bearing fault diagnosis accuracy of different models in Example 1 under test sets with different levels of noise.

[0031] Figure 8 The confusion matrix diagram for Example 1 using the model of the present invention with a test set noise of 8dB.

[0032] Figure 9 This is a visualization of the image features extracted from the test set data using the model in this invention, as shown in Example 1.

[0033] Figure 10 This is a schematic diagram comparing the bearing fault diagnosis accuracy of different models under variable load conditions in Example 2. Detailed Implementation

[0034] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. 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 protection scope of the present invention.

[0035] This invention proposes a rolling bearing fault diagnosis method based on ECA_ResNet. Please refer to [link / reference]. Figure 1First, a combination of short-time Fourier transform and pseudo-color processing methods is used to convert one-dimensional vibration signals into three-channel color time-frequency images, enhancing the data information features. The converted images are then used as the input sample set. Second, an ECA_ResNet bearing fault diagnosis model is constructed. This model includes convolutional layers, max-pooling layers, an improved residual module, and fully connected layers. Shallow edge feature extraction and compression are performed on the data based on the convolutional and pooling layers. An improved residual module is constructed to increase network depth, address network degradation, establish connections between feature channels, and improve the model's ability to adaptively extract deep key features under strong noise and variable loads. Dropout is introduced between two connected layers to suppress overfitting and increase model generalization. Then, the ECA_ResNet bearing fault diagnosis model is trained and validated to obtain the ECA_ResNet bearing fault diagnosis model with the best diagnostic performance. Finally, test set samples with added Gaussian white noise are input into the ECA_ResNet bearing fault diagnosis model with the best diagnostic performance to obtain the final fault diagnosis result.

[0036] The following is a detailed explanation of each step.

[0037] Step 1: Data Processing

[0038] One-dimensional vibration signals of rolling bearings under different conditions were collected. Overlap sampling was used to sequentially truncate the one-dimensional vibration signals into several short signals of fixed length. The short signals were processed by short-time Fourier transform technology to analyze the time-domain and frequency-domain characteristics of the signals. The pseudo-color processing method of pcolormesh in matplotlib was used to map the single-channel grayscale image generated by the short-time Fourier transform into a three-channel color time-frequency image. The three-channel color time-frequency image contains rich color information, which enhances the data information features. All three-channel color time-frequency images were summarized into a sample set. The samples in the sample set were randomly divided to obtain a training set, a validation set, and a test set. The sample ratio of the training set, validation set, and test set was 7:2:1.

[0039] The one-dimensional vibration signals of rolling bearings under different conditions include those under normal, inner ring fault, outer ring fault, and rolling element fault conditions. Specifically, based on the location and diameter of the rolling bearing fault, the one-dimensional vibration signals are categorized into ten types: normal, inner ring (IR007, IR014, IR021), outer ring (OR007, OR014, OR021), and rolling element (B007, B014, B021).

[0040] The Short-Time Fourier Transform (SFT) connects the time and frequency domains, fully capturing time-frequency characteristics. It is simple, efficient, and produces analysis results free from cross-contamination. The SFT formula is as follows:

[0041]

[0042] Where t is time; v is frequency; j is an imaginary unit; f(t) represents the input signal; and g(t-τ) represents the sliding window centered at time τ; STFT f (t,v) represents the Fourier transform of the signal f(t) multiplied by a window function g(t-τ) centered at τ.

[0043] Step 2: Construct the ECA_ResNet bearing fault diagnosis model

[0044] Construct the ECA_ResNet bearing fault diagnosis model. Please refer to the structure diagram of the ECA_ResNet bearing fault diagnosis model. Figure 2 The model mainly consists of one or more convolutional layers, max-pooling layers, improved residual modules, and fully connected layers. First, after two convolutional and max-pooling operations, the convolutional operation obtains shallow edge features from the time-frequency map, while max-pooling reduces the load and number of parameters by downsampling the feature map, thus retaining the main features of the data, removing redundant information, and improving the model's prediction speed. Second, three improved residual modules deepen the network layers. In the residual modules (such as… Figure 4 As shown) an internally embedded effective channel attention network (such as Figure 5 As shown, this not only avoids gradient vanishing and network degradation caused by increasing network depth, but also effectively achieves cross-channel interaction, enhancing feature channels useful for the current task and suppressing feature channels that are not very useful for the current task, adaptively acquiring deep key features. The improved residual module structure diagram is shown below. Figure 3 As shown in Table 1, Dropout is added between the two fully connected layers to pause some neurons with a certain probability, allowing them to learn more robust information, improve the model's generalization ability, suppress overfitting, and use a Softmax layer as a classifier to achieve fault diagnosis. The parameter settings of the ECA_ResNet bearing fault diagnosis model are shown in Table 1.

[0045] Table 1 Model Parameter Settings

[0046]

[0047]

[0048] Both the convolutional and pooling layers are standard convolutional and pooling structures found in conventional CNN networks. The convolutional kernels are 5×5 and 3×3 in size, respectively, and use convolution operations to obtain shallow edge features from the time-frequency map. The pooling layer uses max pooling, which retains the main features of the data and removes redundant information, thereby improving the model's prediction speed.

[0049] The improved calculation formula for the residual module is as follows:

[0050] H(x)′=F scale (F(x),ω)+x (2)

[0051] Where H(x)′ represents the unknown mapping of the improved residual module, F scale (F(x),ω) represents the dot product operation between F(x) and ω, where F(x) represents the residual mapping, x represents the identity mapping, and ω represents the channel weight.

[0052] The improved residual module embeds the effective channel attention network ECA-Net structure into the residual network module. By improving the residual module, the connection between feature channels within the residual is strengthened, the ability of the model to extract key features at a deep level is enhanced, redundant information interference is effectively suppressed, and the problems of gradient vanishing and network degradation are solved.

[0053] The calculation formula for the residual network module is as follows:

[0054] H(x)=F(x)+x (3)

[0055] Where H(x) represents the unknown mapping, F(x) represents the residual mapping, and x represents the identity mapping.

[0056] The computation process of an effective channel attention network, ECA-Net, is as follows:

[0057] (1) Features are aggregated by global average pooling to obtain global channel information. The formula for global average pooling is as follows:

[0058]

[0059] Where, x i represents the i-th feature map with an input size of H×W, and y represents the global feature obtained after global average pooling;

[0060] (2) Use 1D convolution with kernel size k to replace fully connected layers to generate channel weights, obtain the interdependencies between channels, avoid dimensionality reduction, and reduce computational parameters. The 1D convolution formula is as follows:

[0061] ω=σ(C1D k (y)) (5)

[0062] Where ω represents the channel weights, σ ​​represents the Sigmoid function, C1D is 1D convolution, y represents the global features obtained after global average pooling, and k is the kernel size. The kernel size adaptively changes through the channel dimension, allowing layers with a larger number of channels to achieve better interaction. The adaptive function formula is as follows:

[0063]

[0064] Where c represents the channel dimension, b = 1, γ = 2;

[0065] (3) Perform a dot product operation between the original input features and the channel weights to obtain features with channel attention.

[0066] Step 3: Model Training

[0067] The color time-frequency images of the training set are input into the ECA_ResNet bearing fault diagnosis model for training. The model hyperparameters are adjusted using the validation set until the network converges. The optimal model is then saved, which yields the ECA_ResNet bearing fault diagnosis model with the best diagnostic performance.

[0068] In training the ECA_ResNet bearing fault diagnosis model, the color time-frequency image was randomly vertically flipped, the batch size was 32, the number of iterations was 30, the cross-entropy loss was used as the error loss function, Adam was used as the optimizer, the learning rate was 0.001, and the learning rate decay mechanism StepLR was used to adjust the learning rate. The learning rate was halved every 10 iterations. By adopting the learning rate decay mechanism, the training effect was improved and the training time cost was reduced.

[0069] Step 4, Diagnosis

[0070] Gaussian white noise with different signal-to-noise ratios was added to the test set to simulate different levels of environmental noise in actual industrial production and to test the robustness of the model. The test set samples with added Gaussian white noise were then input into the ECA_ResNet bearing fault diagnosis model with the best diagnostic effect obtained in step 3 to obtain the final fault diagnosis result.

[0071] The Gaussian white noise added to the test set was 2dB, 4dB, 6dB, 8dB, and 10dB, respectively. The formula for calculating Gaussian white noise is:

[0072]

[0073] Among them, SNR dB Indicates signal-to-noise ratio, P signal P represents the signal amplitude. noise Indicates the noise amplitude.

[0074] The data used in the following examples comes from the publicly available bearing dataset from Case Western Reserve University Bearing Data Center [EB / OL]. 2018. https: / / csegroups.case.edu /

[0075] The bearing data center (pages / download-data-file) contains CWRU bearing data, which is a globally recognized standard dataset for bearing fault diagnosis. The test bench consists of a horsepower motor, torque sensor / encoder, dynamometer, and control electronics. The tested bearing model is SKF6205 motor bearing, and the corresponding motor loads are 0HP, 1HP, 2HP, and 3HP. The sampling frequencies are 12kHz and 48kHz.

[0076] Example 1

[0077] A rolling bearing fault diagnosis method based on ECA_ResNet mainly includes the following steps:

[0078] Step 1: Accelerometer data from a drive-end unit with a sampling frequency of 12kHz and a load of 2HP was used as experimental data. Bearing faults were determined using electrical discharge machining (EDM), including inner ring faults, outer ring faults, and rolling element faults, with fault diameters of 0.1778mm, 0.3556mm, and 0.5334mm, respectively. Based on the fault location and diameter, the bearing data were categorized into ten types: Normal, Inner Ring (IR007, IR014, IR021), Outer Ring (OR007, OR014, OR021), and Rolling Element (B007, B014, B021). Three-channel color time-frequency plots of rolling bearings under different conditions are shown below. Figure 6 As shown, 6(a) is an image of bearing B007, 6(b) is an image of bearing B014, 6(c) is an image of bearing B021, 6(d) is an image of bearing IR007, 6(e) is an image of bearing IR014, 6(f) is an image of bearing IR021, 6(g) is an image of bearing OR007, 6(h) is an image of bearing OR014, 6(i) is an image of bearing OR021, and 6(j) is an image of a normal bearing. To address the problem of insufficient bearing samples, overlapping sampling is used to extract the signal and enhance the data. The overlapping sampling calculation formula is as follows:

[0079]

[0080] Where N is the number of samples in a single class; L is the length of the one-dimensional signal; L1 is the length of the signal segment; and D is the step size.

[0081] This invention uses a single signal segment length of 1024 and 1000 samples per class of signal, for a total of 10,000 sample data. The bearing signal is converted into a three-channel color time-frequency image using a combination of short-time Fourier transform and pseudo-color processing, serving as the model input sample set. The image resolution is 64×64, with the training set, validation set, and test set ratio being 7:2:1. The experimental dataset parameters are shown in Table 2.

[0082] Table 2 Experimental Dataset Parameter Table

[0083]

[0084]

[0085] Step 2: Construct the ECA_ResNet bearing fault diagnosis model

[0086] Step 3: Input the color time-frequency images from the training set into the ECA_ResNet bearing fault diagnosis model for training. Adjust the model's hyperparameters using the validation set until the network converges. Save the optimal model, which yields the ECA_ResNet bearing fault diagnosis model with the best diagnostic performance. During training, the data is randomly vertically flipped; the batch size is 32; the number of iterations is 30; cross-entropy loss is used as the error loss function; Adam is the optimizer; the learning rate is 0.001; and a learning rate decay mechanism (StepLR) is used to adjust the learning rate, halving it every 10 iterations. This learning rate decay mechanism improves training effectiveness and reduces training time costs.

[0087] Step 4: To simulate different levels of environmental noise in actual industrial production, Gaussian white noise with a signal-to-noise ratio (SNR) of 2dB, 4dB, 6dB, 8dB, and 10dB is added to the test set. The test set with added noise is then put into the model to obtain the classification results, verify the model performance, and realize the rolling bearing fault diagnosis task.

[0088] To verify the noise immunity of the present invention, the experimental results of Example 1 were compared with those of Lenet5, AlexNet, MobileNetV2, improved Lenet5 (Wu Chenfang, Yang Shixi, Huang Haizhou, et al. Research on a rolling bearing fault diagnosis method based on an improved LeNet-5 model [J]. Vibration and Shock, 2021, 40(12): 55-61.), and SE-CNN (WANG H, XU J, YANR, et al. A new intelligent bearing fault diagnosis method using SDPrepresentation and SE-CNN [J]. IEEE Transactions on Instrumentation and Measurement, 2019, 69(5): 2377-2389.). The experimental results are as follows: Figure 7 As shown in the figure, the average accuracy of this invention under noise conditions reaches 97.5%, which is higher than several other deep learning methods. This indicates that the rolling bearing fault diagnosis method proposed in this invention has good noise resistance and robustness.

[0089] To further verify the fault diagnosis method in this invention's ability to learn features and classify faults under noise, a confusion matrix of the model's classification results was plotted when the signal-to-noise ratio of the test set was 8dB. The visualization results are shown below. Figure 8 As shown. The fault diagnosis method in this invention introduces a channel attention mechanism to increase the model's sensitivity to important features, eliminate the interference of redundant information, and effectively achieve bearing fault classification.

[0090] To verify the key feature extraction capability of the fault diagnosis method in this invention, the feature image extracted by the fault diagnosis method in this invention is visualized in two dimensions using the t-SNE dimensionality reduction algorithm. The visualization results are as follows: Figure 9 As shown. By Figure 9 It can be seen that the fault diagnosis method in this invention has sufficient feature extraction and clear fault classification.

[0091] Example 2

[0092] In real industrial environments, bearing load conditions are highly variable, requiring bearing fault diagnosis methods to have high generalization ability. To further verify the generalization performance of this invention, bearing fault diagnosis experiments under variable loads were conducted. First, bearing data with a sampling frequency of 12kHz and loads of 0HP, 1HP, and 2HP were used as model training samples. Then, bearing data under other loads were used as model test samples. The signal processing method was the same as the data processing process in Example 1. For example, "0-->1" represents using data under 0HP load as the model training sample set and data under 1HP load as the model test sample set.

[0093] The ECA-ResNet bearing fault diagnosis model proposed in this invention is compared with WDCNN (Zhang W, Peng G, Li C, et al. A New Deep Learning Model for Fault Diagnosis with Good Anti-Noise and Domain Adaptation Ability on Raw Vibration Signals[J]. Sensors, 2017, 17(2):425.), SE_CNN, and MobileNetv2. The average of five repeated experiments is used as the experimental result. The bearing fault diagnosis results of each method under variable load are as follows: Figure 10 As shown in the figure, SE_CNN, WDCNN, and MobileNetv2 exhibit low accuracy in bearing fault diagnosis under variable load conditions, indicating poor model generalization performance. In contrast, the fault diagnosis method in this invention achieves an average accuracy of 93.69%, representing improvements of 6.97%, 3.39%, and 1.75% respectively compared to the other three methods. This demonstrates that the fault diagnosis method in this invention possesses strong generalization capabilities and is applicable to bearing fault diagnosis under variable load conditions.

[0094] As demonstrated by the above embodiments, the ECA-ResNet-based rolling bearing fault diagnosis method proposed in this invention exhibits better noise resistance and generalization compared to other depth methods. Therefore, the ECA-ResNet-based rolling bearing fault diagnosis method proposed in this invention has high practical value and can be used for actual rolling bearing fault diagnosis tasks.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention in any way. All equivalent transformations or modifications made in accordance with the essence of the present invention should be covered within the protection scope of the present invention.

Claims

1. An ECA_ResNet-based rolling bearing fault diagnosis method, characterized in that, The method mainly comprises the following steps: Step 1: Collect one-dimensional vibration signals of rolling bearings in different conditions, convert the one-dimensional vibration signals into three-channel color time-frequency diagrams by using short-time Fourier transform and pseudo-color processing method, aggregate all the three-channel color time-frequency diagrams into a sample set, and randomly divide the samples in the sample set to obtain a training set, a verification set and a test set, wherein the sample proportion in the training set, the verification set and the test set is 7:2:1; Step 2: Construct an ECA_ResNet bearing fault diagnosis model, which comprises a convolution layer, a maximum pooling layer, an improved residual module and a full connection layer; the convolution layer is used to obtain shallow edge features of a sample image, the maximum pooling layer is used to reduce the load and the number of parameters by down-sampling the feature map while retaining the main features of the data; the improved residual module is used to deepen the network layers and adaptively obtain deep key information; and the full connection layer is used to arrange the features learned by the convolution layer and the pooling layer into a one-dimensional vector, and then use a softmax function to classify the collected features to realize fault diagnosis; Step 3: input the color time-frequency diagram of the training set into the ECA_ResNet bearing fault diagnosis model for training, and adjust the model hyperparameters by using the verification set until the network converges, and save the optimal model, that is, obtain the ECA_ResNet bearing fault diagnosis model with the optimal diagnosis effect; Step 4: add Gaussian white noise with different signal-to-noise ratios in the test set to simulate different degrees of environmental noise in actual industrial production, detect the robustness of the model, input the test set sample with Gaussian white noise into the ECA_ResNet bearing fault diagnosis model with the optimal diagnosis effect obtained in step 3, and obtain the final fault diagnosis result.

2. The ECA_ResNet-based rolling bearing fault diagnosis method according to claim 1, characterized in that, In step 1, the one-dimensional vibration signals of rolling bearings in different conditions include one-dimensional vibration signals of bearings in normal, inner ring fault, outer ring fault and rolling element fault states.

3. The ECA_ResNet-based rolling bearing fault diagnosis method of claim 1, wherein, In step 1, the one-dimensional vibration signals are sequentially cut into a plurality of fixed-length short signals by using overlap sampling, the signals are processed by using short-time Fourier transform to analyze the time domain characteristics and frequency domain characteristics of the signals, and a single-channel gray image is generated.

4. The ECA_ResNet-based rolling bearing fault diagnosis method of claim 3, wherein, In step 1, the pseudo-color processing method uses pcolormesh in matplotlib, and the pseudo-color processing method can map the single-channel gray value to a three-channel color diagram.

5. The ECA_ResNet-based rolling bearing fault diagnosis method of claim 1, wherein, In step 2, the improved residual module is an effective channel attention network ECA-Net structure embedded in a residual network module.

6. The ECA_ResNet-based rolling bearing fault diagnosis method according to claim 5, characterized in that, The operation process of the effective channel attention network ECA-Net is as follows: first, the feature is aggregated by global average pooling to obtain channel global information, then a 1D convolution with a convolution kernel size of k is used to replace the full connection to generate channel weights and obtain the mutual dependence relationship between channels, and finally the point product operation of the original input feature and the channel weight is performed to obtain the feature with channel attention.

7. The ECA_ResNet-based rolling bearing fault diagnosis method of claim 1, wherein, The ECA_ResNet bearing fault diagnosis model constructed in step 2) contains two fully connected layers. By introducing the Dropout technique between the two fully connected layers, part of the neurons are suspended from working with a certain probability, thereby inhibiting model overfitting.

8. The ECA_ResNet-based rolling bearing fault diagnosis method of claim 1, wherein, In step 3), when training the ECA_ResNet bearing fault diagnosis model, the color time-frequency diagram is randomly vertically flipped, the batch size is 32, the number of iterations is 30 times, the cross-entropy loss is used as the error loss function, Adam is used as the optimizer, the learning rate is 0.001, and the learning rate decay mechanism StepLR is used to adjust the learning rate, and the learning rate is halved every 10 iterations.

9. The ECA_ResNet-based rolling bearing fault diagnosis method of claim 1, wherein, In step 4), the Gaussian white noise added to the test set is 2dB, 4dB, 6dB, 8dB, and 10dB, respectively.

Images