Rolling bearing fault diagnosis method and device

Abstract

The application relates to a rolling bearing fault diagnosis method, which comprises the following steps: obtaining a vibration signal of a bearing to be detected and inputting the vibration signal into a pre-trained fault diagnosis model, wherein the fault diagnosis model comprises, in sequence along a positive propagation direction, a first Inception-BN network, a second Inception-BN network, a third Inception-BN network, a pooling layer, a full connection layer and a Softmax layer; after the vibration signal of the bearing to be detected is input, first feature maps, second feature maps, third feature maps, local feature maps and complete feature maps are generated in sequence, and then the complete feature maps are sent into the Softmax layer to output a predicted fault type of the bearing to be detected. The BN layer is arranged in the Inception-BN network, the stability of the model under variable load is improved, a 1*1 convolution kernel is used as a bottleneck layer, the parameter quantity is reduced without influencing the network feature representation capability, and the fault diagnosis efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a method and equipment for diagnosing rolling bearing faults. Background Technology

[0002] Rolling bearings play a crucial role in various mechanical systems. However, due to harsh operating environments and wear caused by prolonged operation, rolling bearings are often threatened with failure. Timely fault diagnosis is essential for the maintenance and upkeep of mechanical systems, reducing unnecessary downtime and improving equipment reliability and lifespan.

[0003] Traditional rolling bearing fault diagnosis methods typically rely heavily on expert knowledge and are costly. However, these methods often suffer from insufficient sensitivity, low accuracy, and sensitivity to noise. In recent years, deep learning technology has made significant progress in signal processing and pattern recognition. Li Bin et al. proposed a deep belief network using a novel activation function, reducing the time cost of fault detection while maintaining high fault identification capabilities. Xiong et al. proposed a generative adversarial network based on a deep autoencoder. By using fault data synthesized by the autoencoder and adding a gradient penalty term to the generative adversarial network, they improved the training instability of the generative network, achieving intelligent fault diagnosis of rolling bearings. Tang Guiji et al. proposed a rolling bearing fault identification method based on a combination of AlexNet and Adaboost. This method converts vibration signals into time-frequency maps using wavelet transform, which are then input into an AlexNet-Adaboost classifier, achieving effective identification of rolling bearing faults under multiple loads.

[0004] However, these methods require a great deal of expertise in vibration signal processing and have certain limitations. Furthermore, existing fault diagnosis models suffer from low feature extraction accuracy and poor fault diagnosis accuracy due to their small receptive field. Simply increasing the receptive field of the model to improve diagnostic accuracy would lead to an increase in the computational load and number of parameters, resulting in low efficiency in fault diagnosis. Therefore, existing methods cannot simultaneously guarantee both diagnostic accuracy and diagnostic efficiency. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problem that while the existing technology increases the accuracy of model prediction, it also leads to a decrease in the efficiency of model diagnosis.

[0006] To solve the above-mentioned technical problems, the present invention provides a method for diagnosing rolling bearing faults, comprising:

[0007] The vibration signal of the bearing to be tested is acquired and input into a pre-trained fault diagnosis model. The pre-trained fault diagnosis model includes, in sequence along the forward propagation direction: a first Inception-BN network, a second Inception-BN network, a third Inception-BN network, a pooling layer, a fully connected layer, and a Softmax layer.

[0008] The vibration signal of the bearing to be tested is input into the first Inception-BN network. After obtaining multiple output feature maps of different sizes of the vibration signal, they are stitched together along the depth direction and then batch normalized to output the first feature map.

[0009] The first feature map is input into the second Inception-BN network to reduce its dimensionality, and multiple output feature maps of different scales of the first feature map after dimensionality reduction are obtained. The first feature map is then max-pooled and output by one-dimensional convolution. This output is then concatenated with the multiple output feature maps of different scales of the first feature map after dimensionality reduction along the depth direction, and then batch normalized to output the second feature map.

[0010] The second feature map is input into the third Inception-BN network to reduce its dimensionality and obtain multiple output feature maps of different scales of the second feature map after dimensionality reduction; the second feature map is max pooled and then output by one-dimensional convolution, and then concatenated with the multiple output feature maps of the second feature map at different scales along the depth direction, and then output as the third feature map after batch normalization.

[0011] The third feature map is input into the pooling layer for compression to obtain a local feature map;

[0012] The local feature map is input into a fully connected layer and weighted using a weight matrix to obtain a complete feature map.

[0013] The complete feature map is input into the Softmax layer, which outputs the predicted fault type of the bearing to be detected.

[0014] In one embodiment of the present invention, the first Inception-BN network includes:

[0015] The input layer takes the vibration signal of the bearing to be tested as its input.

[0016] Multiple parallel one-dimensional convolutional blocks with different kernel sizes are used, with the vibration signal of the bearing to be detected as the input and the output as output features at different scales.

[0017] The Maxpooling layer, which runs parallel to the one-dimensional convolutional block, takes the vibration signal of the bearing to be detected as input and outputs the pooling features.

[0018] The Filter Concatenation layer takes as input output features and pooling features of different scales, and outputs as the first stitched image that concatenates all inputs along the depth direction.

[0019] The BN layer takes the first stitched image as input and outputs the first feature map.

[0020] In one embodiment of the present invention, the plurality of parallel one-dimensional convolutional blocks with different kernel sizes include: a one-dimensional convolutional block with a kernel size of 3×1, a one-dimensional convolutional block with a kernel size of 5×1, and a one-dimensional convolutional block with a kernel size of 7×1.

[0021] In one embodiment of the present invention, the window size of the Max pooling layer is 3×1.

[0022] In one embodiment of the present invention, the second Inception-BN network includes:

[0023] The input layer takes the first feature map as its input.

[0024] The first branch includes:

[0025] The Bottleneck layer takes the first feature map as input and outputs a dimensionality-reduced feature.

[0026] Multiple convolutional blocks with different kernel sizes take the reduced-dimensionality feature map as input and output features at different scales as outputs.

[0027] The second branch includes:

[0028] The Max Pooling layer takes the first feature map as input and outputs the pooled features.

[0029] A one-dimensional convolutional block, whose input is the pooling features and whose output is the convolutional features;

[0030] The Filter Concatenation layer takes inputs of output features and convolutional features at different scales, and outputs a second stitched image that concatenates all inputs along the depth direction.

[0031] The BN layer takes the second stitched image as input and outputs the second feature map.

[0032] In one embodiment of the present invention, the third Inception-BN network includes:

[0033] The input layer takes the second feature map as its input.

[0034] The first branch includes:

[0035] The Bottleneck layer takes the second feature map as input and outputs a dimensionality-reduced feature.

[0036] Multiple convolutional blocks with different kernel sizes take the reduced-dimensionality feature map as input and output features at different scales as outputs.

[0037] The second branch includes:

[0038] The MaxPooling layer takes the second feature map as input and outputs pooled features.

[0039] A one-dimensional convolutional block, whose input is the pooling features and whose output is the convolutional features;

[0040] The Filter Concatenation layer takes inputs of output features and convolutional features at different scales, and outputs a third stitched image that connects all inputs along the depth direction.

[0041] The BN layer takes the third stitched image as input and outputs the third feature map.

[0042] In one embodiment of the invention, the Bottleneck layer includes convolutional blocks with a kernel size of 1x1.

[0043] In one embodiment of the present invention, the training process of the pre-trained fault diagnosis model includes:

[0044] Vibration signals of rolling bearings under different working conditions and various fault types are obtained, preprocessed, and a bearing fault dataset with real label categories is obtained. The dataset is divided into training dataset, validation dataset and test dataset according to a preset ratio.

[0045] The vibration signals from the training dataset are input into the fault diagnosis model to obtain the corresponding predicted fault types. The cross-entropy loss function between the predicted fault types and the true label categories is calculated. The Adam optimizer is used to optimize the cross-entropy loss function, and the fault diagnosis model is trained until the cross-entropy loss function converges, thus obtaining the pre-trained fault diagnosis model.

[0046] In one embodiment of the present invention, after obtaining the pre-trained fault diagnosis model, the method further includes:

[0047] The test set data is input into the pre-trained fault diagnosis model;

[0048] The AdaBN algorithm is used to perform a forward propagation calculation of the test set data from the input layer to the output layer to obtain the current mean and variance.

[0049] By using the current mean and variance, the mean and variance in all BN layers of the pre-trained fault diagnosis model are replaced, thus completing the adaptive adjustment of the pre-trained fault diagnosis model.

[0050] This invention also provides a rolling bearing fault diagnosis device, comprising:

[0051] A data acquisition device is connected to the bearing to be tested to acquire the vibration signal of the bearing to be tested;

[0052] The host computer is connected to the acquisition device and is used to execute the computer program to implement the steps of the rolling bearing fault diagnosis method described above, and to obtain the predicted fault type of the bearing to be tested.

[0053] The display device is connected to the host computer and is used to acquire and display the predicted fault type of the bearing to be tested.

[0054] The technical solution of the present invention has the following advantages compared with the prior art:

[0055] The rolling bearing fault diagnosis method of this invention incorporates BN layers in the first, second, and third Inception-BN networks to improve model stability under varying loads, standardize the output of each layer, and prevent gradient vanishing or exploding. In the Inception-BN network, different sized one-dimensional convolutional kernels are used in parallel within the same layer to replace the original two-dimensional convolutions for adaptive fault feature extraction, capturing output feature maps of different scales and levels. Furthermore, a parallel MaxPooling operation is introduced in the second and third Inception-BN networks, using 1×1 convolutional kernels as bottleneck layers. This reduces the number of parameters and computational complexity without affecting the network's feature representation capabilities, thereby improving fault diagnosis efficiency. Finally, Filter Concatenation is used to concatenate the output feature maps of different scales generated by convolution along the depth direction, expanding the receptive field of the convolutional kernels, increasing semantic information, improving network performance, and enhancing the accuracy of fault type prediction. Attached Figure Description

[0056] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein...

[0057] Figure 1 This is a model structure diagram of the fault diagnosis model for rolling bearings provided by the present invention;

[0058] Figure 2This is a structural diagram of the first Inception-BN network provided by the present invention;

[0059] Figure 3 This is a structural diagram of the second Inception-BN network provided by the present invention;

[0060] Figure 4 This is a comparison chart of the test results of the recognition rate of the rolling bearing fault diagnosis model on test set A under different numbers of training samples provided by the present invention.

[0061] Figure 5 'a' is a schematic diagram comparing the accuracy curves of the training set and the validation set. Figure 5 b is a schematic diagram comparing the loss curves of the training set and the validation set;

[0062] Figure 6 This is a schematic diagram of sensor signals with an inner ring damage size of 14 mils in the CWRU dataset under three different loads provided by this invention;

[0063] Figure 7 This is a schematic diagram comparing the recognition rates of the rolling bearing fault diagnosis model provided by this invention with its control model;

[0064] Figure 8 'a' represents a schematic diagram of the original input vibration signal. Figure 8 b is a schematic diagram of the output of the first Inception-BN network; Figure 8 c is a schematic diagram of the output of the second Inception-BN network; Figure 8 d represents the output diagram of the third Inception-BN network; Figure 8 The 'e' is a schematic diagram of the output of the Softmax layer. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0066] Reference Figure 1 As shown, the rolling bearing fault diagnosis model of the present invention, and the rolling bearing fault diagnosis method based on the fault diagnosis model, specifically include:

[0067] The vibration signal of the bearing to be tested is acquired and input into a pre-trained fault diagnosis model. The pre-trained fault diagnosis model includes, in sequence along the forward propagation direction: a first Inception-BN network, a second Inception-BN network, a third Inception-BN network, a pooling layer, a fully connected layer, and a Softmax layer.

[0068] The vibration signal of the bearing to be tested is input into the first Inception-BN network. After obtaining multiple output feature maps of different sizes of the vibration signal, they are stitched together along the depth direction and then batch normalized to output the first feature map.

[0069] The first feature map is input into the second Inception-BN network to reduce its dimensionality, and multiple output feature maps of different scales of the first feature map after dimensionality reduction are obtained. The first feature map is then max-pooled and output by one-dimensional convolution. This output is then concatenated with the multiple output feature maps of different scales of the first feature map after dimensionality reduction along the depth direction, and then batch normalized to output the second feature map.

[0070] The second feature map is input into the third Inception-BN network to reduce its dimensionality and obtain multiple output feature maps of different scales of the second feature map after dimensionality reduction; the second feature map is max pooled and then output by one-dimensional convolution, and then concatenated with the multiple output feature maps of the second feature map at different scales along the depth direction, and then output as the third feature map after batch normalization.

[0071] The third feature map is input into the pooling layer for compression to obtain a local feature map;

[0072] The local feature map is input into a fully connected layer and weighted using a weight matrix to obtain a complete feature map.

[0073] The complete feature map is input into the Softmax layer, which outputs the predicted fault type of the bearing to be detected.

[0074] The first, second, and third Inception-BN networks are all improvements on the original Inception network. The Inception network's design is inspired by the multi-scale feature extraction capabilities of the human visual system. Its core idea is to use two-dimensional convolutional kernels of different sizes in parallel within the same layer to capture feature information at different scales and levels. Simultaneously, it uses a 1x1 convolutional kernel as a bottleneck layer, reducing the model's computational complexity and improving diagnostic efficiency without affecting the network's feature representation capabilities. Finally, the multi-scale feature maps generated by the convolutions are concatenated along the dimensional direction for output. To further optimize the Inception network, this embodiment improves the original structure. After the bottleneck layer, three one-dimensional convolutions with kernel sizes of 3×1, 5×1, and 7×1 are used instead of the original two-dimensional convolutions for adaptive fault feature extraction. At the same time, another parallel MaxPooling operation is introduced, followed by a bottleneck layer to reduce dimensionality. This design ensures that the model captures features at different scales while achieving model effectiveness with a relatively small number of parameters.

[0075] Specifically, in this embodiment of the invention, the backbone network of the model is first constructed by stacking three one-dimensional Inception and BN combinations. The backbone network feeds the fault features extracted from the vibration signal into the pooling layer for compression to filter out more sensitive features. Finally, the fully connected layer generates a complete feature map from the filtered local features using a weight matrix, and feeds it into the softmax layer to obtain the fault diagnosis output. Referring to Table 1, the network structure parameters of the rolling bearing fault diagnosis model of this invention are shown.

[0076] Table 1 Model network structure parameters

[0077]

[0078] Specifically, refer to Figure 2 As shown, the first Inception-BN network includes:

[0079] The input layer takes the vibration signal of the bearing to be tested as its input.

[0080] Multiple parallel one-dimensional convolutional blocks with different kernel sizes are used, with the vibration signal of the bearing to be detected as the input and the output as output features at different scales.

[0081] The Maxpooling layer, which runs parallel to the one-dimensional convolutional block, takes the vibration signal of the bearing to be detected as input and outputs the pooling features.

[0082] The Filter Concatenation layer takes as input output features and pooling features of different scales, and outputs as the first stitched image that concatenates all inputs along the depth direction.

[0083] The BN layer takes the first stitched image as input and outputs the first feature map.

[0084] The multiple parallel one-dimensional convolutional blocks with different kernel sizes include: a one-dimensional convolutional block with a kernel size of 3×1, a one-dimensional convolutional block with a kernel size of 5×1, and a one-dimensional convolutional block with a kernel size of 7×1; the window size of the Max pooling layer is 3×1.

[0085] Specifically, in the embodiments of the present invention, reference is made to Figure 3 As shown, the second Inception-BN network and the third Inception-BN network have the same network structure; the second Inception-BN network takes the first feature map as input, and the third Inception-BN network takes the second feature map as input.

[0086] Specifically, the second Inception-BN network includes:

[0087] The input layer takes the first feature map as its input.

[0088] The first branch includes:

[0089] The Bottleneck layer takes the first feature map as input and outputs a dimensionality-reduced feature.

[0090] Multiple convolutional blocks with different kernel sizes take the reduced-dimensionality feature map as input and output features at different scales as outputs.

[0091] The second branch includes:

[0092] The Max Pooling layer takes the first feature map as input and outputs the pooled features.

[0093] A one-dimensional convolutional block, whose input is the pooling features and whose output is the convolutional features;

[0094] The Filter Concatenation layer takes inputs of output features and convolutional features at different scales, and outputs a second stitched image that concatenates all inputs along the depth direction.

[0095] The BN layer takes the second stitched image as input and outputs the second feature map.

[0096] Specifically, the third Inception-BN network includes:

[0097] The input layer takes the second feature map as its input.

[0098] The first branch includes:

[0099] The Bottleneck layer takes the second feature map as input and outputs a dimensionality-reduced feature.

[0100] Multiple convolutional blocks with different kernel sizes take the reduced-dimensionality feature map as input and output features at different scales as outputs.

[0101] The second branch includes:

[0102] The Max Pooling layer takes the second feature map as input and outputs the pooled features.

[0103] A one-dimensional convolutional block, whose input is the pooling features and whose output is the convolutional features;

[0104] The Filter Concatenation layer takes inputs of output features and convolutional features at different scales, and outputs a third stitched image that connects all inputs along the depth direction.

[0105] The BN layer takes the third stitched image as input and outputs the third feature map.

[0106] The Bottleneck layer consists of convolutional blocks with a kernel size of 1x1.

[0107] In this embodiment of the invention, Batch Normalization (BN) layers are added after the first, second, and third Inception-BN networks. Batch Normalization (BN) is an important neural network regularization technique. The core idea of ​​BN is to normalize the input data of each layer during the training iterations of deep networks. As training progresses, BN ensures that the output of each layer follows the same mean and variance, meaning that even if the distribution of the input data changes, the network can maintain efficient training and better generalization performance. Therefore, in this embodiment of the invention, BN layers are added to the network to improve the stability of the model under varying loads. This operation helps to normalize the input of each layer, prevent gradient vanishing or exploding, and accelerate the convergence process of the network. By maintaining a uniform distribution of the output of each layer, BN helps to make the model more reliable, especially under different load conditions.

[0108] Based on the above embodiments, in this embodiment, the training process of the pre-trained fault diagnosis model includes:

[0109] Vibration signals of rolling bearings under different working conditions and various fault types are obtained, preprocessed, and a bearing fault dataset with real label categories is obtained. The dataset is divided into training dataset, validation dataset and test dataset according to a preset ratio.

[0110] The vibration signals from the training dataset are input into the fault diagnosis model to obtain the corresponding predicted fault types. The cross-entropy loss function between the predicted fault types and the true label categories is calculated. The Adam optimizer is used to optimize the cross-entropy loss function, and the fault diagnosis model is trained until the cross-entropy loss function converges, thus obtaining the pre-trained fault diagnosis model.

[0111] Based on the above embodiments, in this embodiment, after obtaining the pre-trained fault diagnosis model, the method further includes:

[0112] The test set data is input into the pre-trained fault diagnosis model;

[0113] The AdaBN algorithm is used to perform a forward propagation calculation of the test set data from the input layer to the output layer to obtain the current mean and variance.

[0114] By using the current mean and variance, the mean and variance in all BN layers of the pre-trained fault diagnosis model are replaced, thus completing the adaptive adjustment of the pre-trained fault diagnosis model.

[0115] Existing rolling bearing fault diagnosis methods utilize neural network models to learn network weight parameters from training samples. However, the diagnostic performance of these models deteriorates when the distribution of test samples differs significantly from that of training samples. Furthermore, the vibration signal characteristics of rolling bearings change markedly under varying load conditions, leading to inaccurate fault diagnosis. Therefore, in this embodiment, the AdaBN algorithm is used to adjust model parameters during the testing phase of the fault diagnosis model. The AdaBN algorithm, developed based on the Batch Normalization (BN) concept, aims to address the inconsistency in distribution between the source and target domains. The algorithm works by adjusting the parameters of the BN layer in Inception-BN when a test signal is detected to be outside the same domain as the training signal after network training. This resolves the inconsistency between the source and target domains, enabling accurate fault diagnosis under varying load conditions in rolling bearings. The method first trains a fault diagnosis model using training samples. If the distribution of the training signal and the test signal is inconsistent, the test set is input into the fault diagnosis model, and only forward propagation of the data is performed. In this way, the mean and variance calculated from the target domain samples replace all the mean and variance in the original BN layer, thereby adjusting the target domain and the source domain to a space with approximately the same distribution, achieving the purpose of domain adaptation. This approach helps improve the model's performance in the target domain, enabling it to better adapt to the differences in feature distribution between different domains.

[0116] Based on the above embodiments, to verify the technical effects that the present invention can produce, in this embodiment, based on a 64-bit Windows 10 operating system, all algorithms use the deep learning framework PyTorch to build the network model. The basic information of the deep learning platform is as follows: the graphics card version is Nvidia RTX 1050 GPU, the CPU version is i7-7700HQ@2.80GHz, the Python version is 3.8, and the torch version is 1.7.1.

[0117] During the experiment, the training process of the rolling bearing fault diagnosis method based on the one-dimensional Inception-BN structure was as follows:

[0118] (1) Vibration datasets from the Case Western Reserve University (CWRU) Bearing Center were used and preprocessed. The test object in this experiment was the drive-end bearing, and the bearing model being diagnosed was a deep groove ball bearing SKF6205. The vibration signal was acquired at a frequency of 12kHz. The faults were fabricated by electrical discharge machining. The diagnosed bearing was divided into four states: rolling element damage (RF), outer ring damage (OF), inner ring damage (IF), and normal state (NO). The damage diameters for the four faults were 0.07 inch, 0.14 inch, and 0.21 inch, respectively, for a total of nine damage states and one normal state.

[0119] During the experiment, three datasets were prepared, as shown in Table 2. Datasets A, B, and C are CWRU datasets with loads of 1hp, 2hp, and 3hp, respectively. The sample length of each dataset is 512, and the samples are randomly divided into training, validation, and test sets in a ratio of 7:2:1.

[0120] Table 2 Description of Experimental Datasets

[0121]

[0122] (2) To allow the model to fully learn the features, the number of training epochs was set to 50. To better train the model, the Adam optimizer was used to optimize the cross-entropy loss function, and the initial learning rate was set to 0.001. Regarding the batch size, a smaller batch size yields better generalization, so the batch size was set to 64. The model training parameters are shown in Table 3.

[0123] Table 3 Model Training Parameter Settings

[0124]

[0125] (3) Model comprehensive performance analysis

[0126] In this embodiment, the rolling bearing fault diagnosis model was trained multiple times using training samples from dataset A with total sample sizes of 140, 700, 1400, and 3500, respectively. By observing the model's fault diagnosis capability under different sample conditions, the following conclusions were drawn: Since the initial weights of the neural network were randomly generated, this embodiment conducted 10 repeated experiments to verify the stability of the results. The experimental results of the rolling bearing fault diagnosis model's recognition rate on test set A under different numbers of training samples are as follows: Figure 4As shown, the recognition rate of the rolling bearing fault diagnosis model gradually increases with the increase in the number of training samples, while the standard deviation after 10 trials gradually decreases, indicating that the model's stability gradually improves. When the number of training samples reaches 3500, the recognition rate reaches 100% with a standard deviation of 0%, while when the number of training samples is 140, the recognition rate is only 95.2%. This clearly shows that the number of training samples has a significant impact on the model's diagnostic performance. Notably, when the number of training samples is 1400, the recognition rate can already reach 99.4%. This means that the rolling bearing fault diagnosis model can still achieve a high recognition rate when using less training data, demonstrating the model's strong ability to suppress overfitting. Such results provide useful insights for practical applications when faced with limited sample data.

[0127] Reference Figure 5 As shown, the training performance curves of the rolling bearing fault diagnosis model are presented on 3500 samples of training set A. The accuracy curve reflects the consistency between the predicted and true classes, while the loss value reveals the gap between the model's predictions and the actual values. Specifically, Figure 5 Figure 'a' shows a comparison of the accuracy curves for the training and validation sets, while... Figure 5 Figure b shows a comparison of the loss curves for the training and validation sets. Observation Figure 5 As can be seen, the training performance of the rolling bearing fault diagnosis model is very stable. Around the 10th iteration, the training accuracy approaches 100% and remains there until the end of training. Simultaneously, the loss value on the validation set decreases smoothly without drastic oscillations, eventually approaching zero, indicating that the model's predictions are very close to the target values. This result clearly demonstrates that the proposed rolling bearing fault diagnosis model possesses strong robustness, maintaining high accuracy during training and avoiding excessive oscillations, thus effectively achieving accurate predictions of the data.

[0128] (4) Model Cross-Load Capability Analysis

[0129] In practical electrical engineering, the drive-end bearing often operates under variable load conditions, which causes corresponding changes in the bearing's vibration characteristics. (Refer to...) Figure 6 The image shows the normalized sensor signals with an inner ring damage (IF) value of 14 mils in the CWRU dataset under three different loads. Figure 6 Observations reveal that the fault characteristics in the vibration signal change significantly under different loads, and there are also obvious differences in the amplitude and period of the fluctuations.

[0130] This situation may cause the classifier to fail to classify correctly when processing extracted features, thus requiring the model to have high domain adaptability. In this embodiment, the model is trained using parts A, B, and C of the CWRU dataset, respectively. The training and validation sets will be extracted from one of the loads, while signals from the other two loads will be used as the test set for validation. This design aims to ensure that the model can accurately classify faults under different load conditions, demonstrating good adaptability to changes in the working environment.

[0131] Reference Figure 7 The diagram shown is a comparison of the recognition rates of the rolling bearing fault diagnosis model provided by this invention and existing models; according to Figure 7 The experimental results show that the SVM algorithm has a low average diagnosis rate of less than 70%; the DNN achieves a recognition rate of approximately 80%; while the rolling bearing fault diagnosis model provided by this invention achieves an average recognition rate of 90.2%. Notably, by employing the AdaBN algorithm, the recognition rate of the rolling bearing fault diagnosis model is improved to 96%. Specifically, when using training set B (load of 2 hp) to diagnose test set A (load of 1 hp), the recognition rate of the rolling bearing fault diagnosis model is more than 13% higher than that of the SVM and DNN models. Furthermore, the AdaBN algorithm further improves the diagnosis rate of the rolling bearing fault diagnosis model, especially when trained on training set C and used for diagnosis on test sets A or B, the recognition rate is improved by more than 10%.

[0132] In summary, compared with traditional intelligent diagnostic algorithms and currently popular deep learning algorithms, the rolling bearing fault diagnosis model of this invention, with its one-dimensional Inception-BN model built based on the AdaBN algorithm, demonstrates outstanding adaptability across load domains. This not only achieves a high recognition rate under different load conditions but also highlights the model's powerful performance in coping with complex changes in working environments.

[0133] (5) Model visualization

[0134] Neural network models are often considered black boxes, with their internal workings difficult to explain. To better understand the bearing fault classification process, this embodiment employs the T-SNE (T-distributed Stochastic Neighbor Embedding) visualization method after dimensionality reduction of dataset C, outputting the results of each layer in the training process of the rolling bearing fault diagnosis model. This approach provides a more intuitive representation of the model's learning process in the dimensionality-reduced representation space.

[0135] This visualization method provides a clearer understanding of how the model internally learns and classifies bearing fault characteristics. It offers an intuitive way to gain insight into the complex feature learning process of neural network models when handling bearing fault classification tasks, thereby enhancing our understanding of model behavior. Figure 8 The T-SNE visualization method is used to demonstrate the network learning process, providing in-depth insights into how the model learns bearing fault characteristics.

[0136] in, Figure 8 'a' represents a schematic diagram of the original input vibration signal. Figure 8 b is a schematic diagram of the output of the first Inception-BN network; Figure 8 c is a schematic diagram of the output of the second Inception-BN network; Figure 8 d represents the output diagram of the third Inception-BN network; Figure 8 The output diagram of the Softmax layer is shown in Figure 'e'; from Figure 8 It can be observed that the features of the original vibration signal are initially mixed together and difficult to distinguish. However, after two layers of processing by the 1D-Inception module, some features begin to show a cohesive trend. With three layers of processing by the 1D-Inception module, it can be clearly seen that the distance between different types of features gradually increases, and each feature of the data shows obvious cohesion. Finally, after the output of the fully connected layer, it can be observed that the same features are effectively clustered together, while there is obvious separation between different features. This proves that the proposed one-dimensional Inception-BN model can effectively extract information related to category mapping. As the network deepens, the model's ability to learn features gradually increases, and the classification accuracy also gradually improves. Such visualization results highlight the model's gradual understanding and abstraction of data features during the learning process, indicating that the rolling bearing fault diagnosis model exhibits strong feature extraction and learning capabilities in vibration signal classification tasks.

[0137] Based on the above embodiments, this invention also provides a rolling bearing fault diagnosis device, comprising:

[0138] A data acquisition device is connected to the bearing to be tested to acquire the vibration signal of the bearing to be tested;

[0139] The host computer is connected to the acquisition device and is used to execute the computer program to implement the steps of the rolling bearing fault diagnosis method described above, and to obtain the predicted fault type of the bearing to be tested.

[0140] The display device is connected to the host computer and is used to acquire and display the predicted fault type of the bearing to be tested.

[0141] The rolling bearing fault diagnosis method of this invention incorporates BN layers in the first, second, and third Inception-BN networks to improve model stability under varying loads, standardize the output of each layer, and prevent gradient vanishing or exploding. In the Inception-BN network, different sized one-dimensional convolutional kernels are used in parallel within the same layer to replace the original two-dimensional convolutions for adaptive fault feature extraction, capturing output feature maps of different scales and levels. Furthermore, a parallel MaxPooling operation is introduced in the second and third Inception-BN networks, using 1×1 convolutional kernels as bottleneck layers. This reduces the number of parameters and computational complexity without affecting the network's feature representation capabilities, thereby improving fault diagnosis efficiency. Finally, Filter Concatenation is used to concatenate the output feature maps of different scales generated by convolution along the depth direction, expanding the receptive field of the convolutional kernels, increasing semantic information, improving network performance, and enhancing prediction accuracy.

[0142] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0143] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0144] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0145] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0146] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for diagnosing rolling bearing faults, characterized in that, include: The vibration signal of the bearing to be tested is acquired and input into the pre-trained fault diagnosis model; The pre-trained fault diagnosis model, along the forward propagation direction, includes, in sequence: a first Inception-BN network, a second Inception-BN network, a third Inception-BN network, a pooling layer, a fully connected layer, and a Softmax layer; The vibration signal of the bearing to be tested is input into the first Inception-BN network. After obtaining multiple output feature maps of different sizes of the vibration signal, they are stitched together along the depth direction and then batch normalized to output the first feature map. The first Inception-BN network includes: The input layer takes the vibration signal of the bearing to be tested as its input. Multiple parallel one-dimensional convolutional blocks with different kernel sizes are used, with the vibration signal of the bearing to be detected as the input and the output as output features at different scales. The Max pooling layer, which runs parallel to the one-dimensional convolutional block, takes the vibration signal of the bearing to be detected as input and outputs the pooling features. The Filter Concatenation layer takes as input output features and pooling features of different scales, and outputs as the first stitched image that concatenates all inputs along the depth direction. The BN layer takes the first stitched image as input and outputs the first feature map. The first feature map is input into the second Inception-BN network for dimensionality reduction, and multiple output feature maps of different scales of the dimensionality-reduced first feature map are obtained. The first feature map is then max-pooled and output via one-dimensional convolution. This output is concatenated along the depth direction with the multiple output feature maps of the dimensionality-reduced first feature map, and then batch-normalized to output the second feature map. The second Inception-BN network includes: The input layer takes the first feature map as its input. The first branch includes: The Bottleneck layer takes the first feature map as input and outputs a reduced-dimensional feature map. Multiple convolutional blocks with different kernel sizes take the reduced-dimensionality feature map as input and output features at different scales as outputs. The second branch includes: The Max Pooling layer takes the first feature map as input and outputs the pooled features. A one-dimensional convolutional block, whose input is the pooling features and whose output is the convolutional features; The Filter Concatenation layer takes inputs of output features and convolutional features at different scales, and outputs a second stitched image that concatenates all inputs along the depth direction. The BN layer takes the second stitched image as input and outputs the second feature map. The second feature map is input into the third Inception-BN network to reduce its dimensionality and obtain multiple output feature maps of different scales of the second feature map after dimensionality reduction; the second feature map is max pooled and then output by one-dimensional convolution, and then concatenated with the multiple output feature maps of the second feature map at different scales along the depth direction, and then output as the third feature map after batch normalization. The third feature map is input into the pooling layer for compression to obtain a local feature map; The local feature map is input into a fully connected layer and weighted using a weight matrix to obtain a complete feature map. The complete feature map is input into the Softmax layer, which outputs the predicted fault type of the bearing to be detected.

2. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The multiple parallel one-dimensional convolutional blocks with different kernel sizes include: a one-dimensional convolutional block with a kernel size of 3×1, a one-dimensional convolutional block with a kernel size of 5×1, and a one-dimensional convolutional block with a kernel size of 7×1.

3. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The window size of the Max pooling layer is 3×1.

4. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The third Inception-BN network includes: The input layer takes the second feature map as its input. The first branch includes: The Bottleneck layer takes the second feature map as input and outputs a reduced-dimensional feature map. Multiple convolutional blocks with different kernel sizes take the reduced-dimensionality feature map as input and output features at different scales as outputs. The second branch includes: The Max Pooling layer takes the second feature map as input and outputs the pooled features. A one-dimensional convolutional block, whose input is the pooling features and whose output is the convolutional features; The Filter Concatenation layer takes inputs of output features and convolutional features at different scales, and outputs a third stitched image that connects all inputs along the depth direction. The BN layer takes the third stitched image as input and outputs the third feature map.

5. The rolling bearing fault diagnosis method according to claim 4, characterized in that, The Bottleneck layer comprises convolutional blocks with a kernel size of 1x1.

6. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The training process of the pre-trained fault diagnosis model includes: Vibration signals of rolling bearings under different working conditions and various fault types are obtained, preprocessed, and a bearing fault dataset with real label categories is obtained. The dataset is divided into training dataset, validation dataset and test dataset according to a preset ratio. The vibration signals from the training dataset are input into the fault diagnosis model to obtain the corresponding predicted fault types. The cross-entropy loss function between the predicted fault types and the true label categories is calculated. The Adam optimizer is used to optimize the cross-entropy loss function, and the fault diagnosis model is trained until the cross-entropy loss function converges, thus obtaining the pre-trained fault diagnosis model.

7. The rolling bearing fault diagnosis method according to claim 6, characterized in that, After obtaining the pre-trained fault diagnosis model, the following is also included: Input the test set data into the pre-trained fault diagnosis model; The AdaBN algorithm is used to perform a forward propagation calculation of the test set data from the input layer to the output layer to obtain the current mean and variance. By using the current mean and variance, the mean and variance in all BN layers of the pre-trained fault diagnosis model are replaced, thus completing the adaptive adjustment of the pre-trained fault diagnosis model.

8. A rolling bearing fault diagnosis device, characterized in that, include: A data acquisition device is connected to the bearing to be tested to acquire the vibration signal of the bearing to be tested; A host computer is communicatively connected to the acquisition device and is used to execute a computer program to implement the steps of the rolling bearing fault diagnosis method as described in any one of claims 1 to 7, and to obtain the predicted fault type of the bearing to be tested. The display device is connected to the host computer and is used to acquire and display the predicted fault type of the bearing to be tested.

Images