Rotating machinery fault diagnosis method based on convolutional Transformer

Abstract

The rotating machinery fault diagnosis method fused with convolutional transformer performs fast Fourier transform and Hilbert transform on the vibration signal during the operation of the rotating machinery, and then combines it with the original sequence data of the vibration signal to form a multi-channel signal as the input of the diagnosis model, which can enrich the characteristic information of mechanical faults and enable the subsequent modules to better extract the fault features in the vibration signal. The diagnosis model has multiple layers of feature extraction modules, each of which is composed of a convolutional layer and a multi-head self-attention mechanism of transformer. It includes two convolutional layers and one multi-head self-attention layer, which combines the advantages of convolutional neural networks in processing local features and the ability of transformers in extracting long time series data features, and extracts fault features in frequency domain vibration signals more comprehensively and effectively, improving the accuracy and robustness of rotating machinery fault diagnosis.

Description

Technical Field

[0001] The present invention relates to a rotating machinery fault diagnosis method, in particular to a rotating machinery fault diagnosis method integrating a convolutional Transformer. Background Art

[0002] Rotating machinery is widely used in aerospace, shipbuilding and other fields, playing a vital and irreplaceable role in national defense and industry. However, rotating machinery often operates for a long time in harsh environments, and equipment components inevitably degrade, thus affecting the stable operation of the entire equipment. With the continuous advancement of rotating machinery in speed, scale and complexity, the demand for accurate and effective fault diagnosis has also increased accordingly.

[0003] Traditional fault diagnosis methods mainly consist of two key steps: feature extraction and fault identification. With the development of artificial intelligence, deep learning (DL) has shown great potential in adaptive feature extraction and model generalization performance. The deep learning method builds a deep neural network (DNN) to fully learn from a large number of vibration signals and reveal the hidden nonlinear relationship in the vibration signals, thereby realizing adaptive feature extraction and fault classification, so that feature extraction and fault identification are no longer isolated processes. In recent years, Transformer has shown significant superiority in research fields such as natural language processing and computer vision. The structure uses a multi-head self-attention (MSA) mechanism to complete efficient global feature extraction and long-distance feature modeling. Although Transformer has powerful feature extraction capabilities, it is difficult to achieve the same performance as convolutional neural network (CNN) when processing classification tasks with limited samples. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a rotating machinery fault diagnosis method integrating convolutional Transformer to improve the accuracy and robustness of rotating machinery fault diagnosis.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a rotating machinery fault diagnosis method integrating a convolutional Transformer, collecting vibration signals during the operation of the rotating machinery, and performing fast Fourier transform and Hilbert transform on the vibration signals respectively, combining the two sets of transformed sequence data with the original sequence data of the vibration signal to form a multi-channel signal including three sets of sequence data, and the constructed multi-channel signal is sent as input to a diagnosis model, and the diagnosis model outputs a diagnosis result.

[0006] The diagnostic model has multiple layers of feature extraction modules, each of which is composed of a fusion of a convolutional layer and a Transformer's multi-head self-attention mechanism, including two convolutional layers and one multi-head self-attention layer; the data input to the feature extraction module first passes through the first convolutional layer, and the data processed by the first convolutional layer is sent to the multi-head self-attention layer for feature extraction, and the other way forms a residual connection with the data output by the multi-head self-attention layer and then undergoes layer normalization processing; the data processed by layer normalization is processed by the second convolutional layer, and the other way forms a residual connection with the data processed by the second convolutional layer and is output after layer normalization processing.

[0007] The first convolution layer uses 1×1 convolution to perform channel adjustment and feature integration, and the second convolution layer uses 3×1 convolution to extract spatial features.

[0008] The step size of the 1×1 convolution is 2, and the step size of the 3×1 convolution is 1.

[0009] The feature extraction module is provided with 4 layers.

[0010] Furthermore, the multi-channel signal input to the diagnostic model is first processed by a convolution layer with a convolution kernel of 64×1 and then sent to the first-layer feature extraction module.

[0011] Furthermore, the data output by the feature extraction module of the last layer is processed by global average pooling and then outputs the diagnosis result through the softmax classification layer.

[0012] Furthermore, the collected vibration signal is converted into the frequency domain through fast Fourier transform to obtain the frequency domain characteristics of the data. At the same time, the frequency attributes of the signal that change with time are captured through Hilbert transform. The transformed data is combined with the original vibration signal to form a multi-channel signal containing the original time domain characteristics, frequency domain characteristics and frequency attributes that change with time.

[0013] The multi-head self-attention layer uses multiple sets of parallel self-attention to extract features, and its formula is:

[0014] , where: X is the input data, is the weight matrix, Indicates a connection operation. is the output of 1 to n groups of self-attention, n is the number of heads of multi-head self-attention;

[0015] The formula for each group of self-attention is:

[0016] ;

[0017] Among them, Q, K, and V are query matrix, key matrix, and value matrix respectively. is the transpose of the key matrix K, Represents the dimension of K.

[0018] The number of heads in the multi-head self-attention layer is 2.

[0019] The diagnostic model is trained using a data set consisting of vibration information collected in normal working conditions and various fault conditions.

[0020] The beneficial effect of the present invention is that the multi-channel signal formed by combining the vibration signal with the original vibration signal after fast Fourier transform and Hilbert transform respectively is used as the model input, which can enrich the mechanical fault feature information, so that the subsequent feature extraction module can better extract the fault features in the vibration signal. The feature extraction module uses convolution operation to extract the fault feature information in the local space, and extracts the global fault feature information through the multi-head self-attention mechanism. It combines the advantages of convolutional neural networks in processing local features and the ability of Transformer in extracting long time series data features, and can more comprehensively and effectively extract the fault features in the frequency domain vibration signal, thereby improving the accuracy and robustness of rotating machinery fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a structural principle diagram of the diagnostic model in the present invention.

[0022] Figure 2 It is a structural principle diagram of the feature extraction module of the present invention.

[0023] Figure 3 It is a flow chart of the diagnosis model training of the present invention.

[0024] Figure 4 It is a schematic diagram of vibration signal processing of the present invention.

[0025] Figure 5 This is a comparison chart of fault identification accuracy of different deep diagnosis models.

[0026] Figure 6 This is a comparison chart of the fault recognition rates of the present invention and the original Transformer model under four signal-to-noise ratios.

[0027] Figure 7 This is a comparison chart of the recognition accuracy of the present invention and the original Transformer model. DETAILED DESCRIPTION

[0028] The technical solution of the present invention is described clearly and completely below in conjunction with the accompanying drawings and specific implementation methods. The specific contents listed in the following embodiments are not limited to the technical features necessary for the technical problems to be solved by the technical solutions recorded in the claims. At the same time, the enumerated embodiments are only part of the present invention, not all embodiments.

[0029] The present invention integrates the convolutional transformer into the rotating machinery fault diagnosis method, collects the vibration signal of the rotating machinery during operation, constructs a fault diagnosis model to extract the characteristics of the vibration signal, and outputs the diagnosis result. Figure 1 As shown, it includes multiple layers of feature extraction modules for processing input signals to extract features. The data output by the last layer of feature extraction modules is subjected to global average pooling and softmax classification to output the diagnosis result.

[0030] Different from the usual method of using two-dimensional images as the input of the diagnostic model for fault diagnosis, the present invention uses the sequence data of the vibration signal as the input signal. In order to improve the detection accuracy, the vibration signal is first subjected to fast Fourier transform and Hilbert transform respectively to obtain two sets of transformed sequence data. Among them, the original time domain vibration signal is converted into the frequency domain by fast Fourier transform to obtain the frequency domain characteristics of the data. The frequency attributes of the vibration signal that change over time are captured by Hilbert transform. The two sets of transformed sequence data are combined with the original vibration signal sequence data into a multi-channel signal including 3 sets of sequence data. For example, the original vibration signal is 1×1024 in length. After fast Fourier transform, Hilbert transform and data splicing, the combined multi-channel signal data is 3×1024. The combined multi-channel signal simultaneously contains the time domain characteristics of the original vibration signal, the frequency domain characteristics after fast Fourier transform, and the time-varying frequency attributes captured by Hilbert transform. Using this multi-channel signal as the input of the diagnostic model improves the comprehensiveness and comprehensiveness of the fault feature information, laying the foundation for subsequent feature extraction and analysis.

[0031] like Figure 1 The multi-channel signal of the input diagnosis model shown in the figure first passes through a convolution layer with a large convolution kernel. In this embodiment, a convolution layer with a convolution kernel of 64×1 is selected to capture wider spatial information and features by utilizing the characteristic that a large convolution kernel can cover a larger receptive field, so as to facilitate subsequent fault diagnosis. After being processed by the convolution layer with a large convolution kernel, the input data is sent to the first-layer feature extraction module.

[0032] The feature extraction module is constructed by fusing the multi-head self-attention mechanism of the Transformer encoder with the convolutional layer, which is called FCTB (Fusion Convolutional Transformer Block). In this module, the feedforward neural network in the original Transformer encoder is replaced with a convolutional neural network, because the convolutional neural network can automatically extract and learn high-level features in the data, while the feedforward neural network may be limited by its simple structure and limited feature representation ability. At the same time, the convolutional neural network slides on the input data through the convolution kernel, performs weighted summation on each position of the input data, and extracts features from the data. Since the parameters of the convolution kernel are shared at different positions of the input data, this greatly reduces the number of parameters of the model, reduces the computational complexity, reduces the amount of calculation of the model, and improves the generalization ability of the model.

[0033] like Figure 2 As shown in the figure, the feature extraction module includes two convolutional layers, one multi-head self-attention layer and two residual connection layers. Convolution operation is used to extract fault feature information in local space, and the global fault feature information is extracted through the multi-head self-attention mechanism. By using residual connection, multiple modules can be stacked to increase the network depth, while alleviating the problem of gradient disappearance, thereby improving the accuracy and generalization ability of the model. The data input to the feature extraction module is first convolved through the first convolutional layer. The first convolutional layer uses 1×1 convolution to adjust the channel and integrate features, which can make the model more efficient and improve the feature expression ability without increasing the complexity of the model. The data processed by the first convolutional layer is sent to the multi-head self-attention layer, and the multi-head self-attention mechanism is used for feature extraction.

[0034] The multi-head self-attention layer uses multiple sets of parallel self-attention to extract features, and its formula is:

[0035] ;

[0036] Where: X is the input data, is the weight matrix, Indicates a connection operation. is the output of 1 to n groups of self-attention, and n is the number of heads of multi-head self-attention.

[0037] The formula for each group of self-attention is:

[0038] ;

[0039] Among them, Q, K, and V are query matrix, key matrix, and value matrix respectively. is the transpose of the key matrix K, Represents the dimension of K.

[0040] The output of the multi-head self-attention layer forms a residual connection with the output of the first convolution layer in the first residual connection layer, and is normalized by layer. The data processed by layer normalization is sent to the second convolution layer for convolution operation. The second convolution layer uses 3×1 convolution to extract spatial features, which can reduce the parameters and calculation amount of the model while maintaining spatial information. The data processed by the second convolution layer forms a residual connection with the data processed by layer normalization after the second residual connection layer and the first residual connection layer, and is used as the output of the feature extraction module of this layer after layer normalization. In this embodiment, the step size of the 1×1 convolution layer is 2, the step size of the 3×1 convolution layer is 1, and the number of multi-head self-attention heads is 2. The number of layers of the feature extraction module is selected as 4 layers. After the fault features are extracted by the 4-layer feature extraction module, the network structure is simplified by global average pooling, while reducing the risk of overfitting, so that the network is more focused on extracting global features related to the task. Finally, the fault diagnosis result is output through the softmax classification layer.

[0041] Taking the diagnosis model of the four-layer feature extraction module (FCTB) as an example, assuming that the vibration signal is 1×1024 in length, and the multi-channel signal data after data change and combined splicing is 3×1024, the diagnosis model network structure parameters are as follows:

[0042]

[0043] To train the diagnostic model, the vibration signals generated by the rotating machinery under specific working conditions are collected, such as the vibration signals of normal working state and various fault states, and data samples of different fault types are extracted from them. They are organized into a data set in a unified format, and the signals are processed and fused. Randomly extract the organized data set, 70% as the training set, 15% as the validation set, and the remaining 15% as the test set. Figure 1 As shown in the figure, the training set and validation set are used to train the constructed model. After the forward calculation of the diagnostic model network, the deviation value between the current output and the expected target is obtained. Next, the gradient of the parameters of each layer is continuously calculated and adjusted through the error back propagation algorithm, and the internal weights and biases of the network are gradually optimized. This process is iterated cyclically, so that the performance of the network on the training set is continuously improved until the maximum number of training rounds set is reached. Through the above iterative training process, the model parameters will gradually converge to the optimal solution, thereby obtaining a model with high fault diagnosis and classification accuracy. In order to test the performance of the diagnostic model, the reserved test set data is input into the trained diagnostic model. After layer-by-layer feature extraction and encoding, the diagnostic model outputs the corresponding fault type prediction label. Then the fault diagnosis recognition accuracy is calculated to evaluate the fault diagnosis performance of the diagnostic model.

[0044] To verify the effectiveness of the present invention, different models were tested using test data. The test data used the public CWRU bearing data set, and the test bearing was an SKF-6205-2RS deep groove ball bearing, with a sampling frequency of 12 kHz and a sampling time of 1s. A normal state type and three fault types were simulated using electrospark machining, including rolling element fault, inner ring fault and outer ring fault, with fault diameters of 0.007 inches (0.18 mm), 0.014 inches (0.36 mm), and 0.021 inches (0.54 mm). The corresponding loads of the motor are 0HP, 1HP, 2HP, and 3HP. The present invention uses data under a 1HP horsepower load as the experimental data set.

[0045] Considering the vibration of the rotating mechanical equipment and the interference of environmental noise when the rotating mechanical equipment is running, the present invention introduces Gaussian noise to the original vibration signal in the experimental data set. Four noise levels are considered in the test, namely, the signal-to-noise ratio values ​​are -6, -4, -2 and 0dB.

[0046] like Figure 4 As shown in the figure, the vibration signal is processed by sequential segmentation method, and the vibration signal is truncated into several segments using a non-overlapping sliding window of 1024 data points. After preprocessing, the CWRU bearing dataset consists of 1180 samples, including 10 different bearing states. The training set, validation set, and test set are randomly divided into 70%, 15%, and 15% ratios. The fault labels of the dataset are as follows:

[0047]

[0048] Verification of the influence of network depth on the diagnostic model: To verify the influence of the number of feature extraction module layers on the fault diagnosis results, the number of feature extraction modules (FCTB) in the diagnostic model of the present invention was set to 1, 2, 3, 4 and 5, respectively, and five diagnostic models were obtained, named I, II, III, IV and V. The data with a signal-to-noise ratio of -6dB in the experimental data set was used. In order to reduce the interference of random factors, the average results obtained from 10 repeated experiments were used to perform statistical analysis on the experimental results. The experimental results are shown in Figure 2. Figure 5 As shown in the attached figure, it can be seen that the recognition accuracy increases with the increase of network depth. However, beyond a certain depth, the accuracy tends to be stable, with no obvious improvement, and may even decrease.

[0049] The average recognition accuracy and standard deviation statistics of the five diagnostic models are shown in the following table:

[0050]

[0051] As can be seen from the above table, the diagnostic model IV has the best diagnostic effect and stability. The present invention selects a diagnostic model with 4 layers of feature extraction modules.

[0052] Input data type verification: In order to verify the influence of vibration signal data transformation on the diagnosis results, four types of data are selected as the input of the diagnosis model, including: (1) original vibration signal data; (2) data of the original vibration signal after Hilbert transform (HT); (3) data of the original vibration signal after fast Fourier transform (FFT); (4) multi-channel signal data of the present invention combining the two transformed data with the original vibration signal. The diagnostic accuracy of the diagnosis model is shown in the following table:

[0053]

[0054] When the original vibration signal data is used as the input of the diagnostic model, the bearing fault diagnosis performance is poor. This is because the fault characteristics reflected by the unprocessed vibration signal data are relatively simple and the information is not rich. After FFT and HT transformation processing, the average classification accuracy of mechanical fault diagnosis is significantly improved. However, after fusing the original vibration signal data with the data processed by the two transformations, the bearing fault diagnosis accuracy is the highest under the same diagnostic model. This shows that by using multi-channel input data containing the original vibration signal and its multi-domain transformation, the richness of the vibration signal characteristics can be effectively improved, and the fault diagnosis performance can be significantly improved.

[0055] The diagnostic model of the present invention is compared with the original Transformer model for verification: the diagnostic model of the present invention (FCTNN) adopts the improved feature extraction module of the present invention, and the Transformer model adopts the original Transformer encoder structure as the feature extraction module. The model input data adopts the multi-channel signal of the present invention. The results are as follows Figure 6 As shown in the figure, the original Transformer model is greatly affected by noise and the model robustness is poor. In any noise environment, the diagnostic model structure of the present invention has better model performance and stronger robustness than the original Transformer encoder structure.

[0056] In order to more intuitively display the experimental results of the above two models, the t-SNE technology is used to visualize the features extracted from the last layer of the diagnostic model when the vibration signal-to-noise ratio is -6dB. The results are as follows: Figure 7 As shown in the figure, the data after feature extraction by the unimproved original Transformer model overlap and mix, and the clustering effect is poor; while the clustering effect of the diagnosis model of the present invention is better, indicating that the method of the present invention has strong feature extraction and fault identification capabilities, and can improve the accuracy of fault diagnosis.

[0057] Comparison and verification of other models: The diagnostic model (FCTNN) of the present invention is compared with the WDCNN, CA-MCNN, 1DCNN-LSTM, ACNN-BiLSTM, and WCNN-BiGRU models in the prior art. The results are shown in the following table:

[0058]

[0059] The results show that as the noise intensity decreases, the fault diagnosis accuracy of each model gradually improves, but the FCTNN model proposed in the present invention achieves the highest test accuracy, precision, recall rate and f1 score under any signal-to-noise ratio condition, and has better noise resistance and robustness.

[0060] The above description of the specific implementation mode is only used to help understand the technical concept and core idea of ​​the present invention. Although the technical solution is described and illustrated using a specific preferred embodiment, it should not be understood as a limitation of the present invention itself. Those skilled in the art may make various changes in form and details without departing from the technical concept of the present invention. These easily conceived changes or substitutions should all be included in the protection scope of the present invention.

Claims

1. A rotating machinery fault diagnosis method integrating convolutional Transformer, characterized by: The vibration signals during the operation of the rotating machinery are collected, and the vibration signals are subjected to fast Fourier transform and Hilbert transform respectively, and the two sets of sequence data after the transform are combined with the original sequence data of the vibration signal to form a multi-channel signal including three sets of sequence data, and the formed multi-channel signal is sent as input to the diagnosis model, and the diagnosis model outputs the diagnosis result; The diagnostic model has a multi-layer feature extraction module, each of which is composed of a convolutional layer and a Transformer multi-head self-attention mechanism, including two convolutional layers and one multi-head self-attention layer; the data input to the feature extraction module first passes through the first convolutional layer, and the data processed by the first convolutional layer is sent to the multi-head self-attention layer for feature extraction, and the other path is connected to the data output by the multi-head self-attention layer to form a residual connection and then perform layer normalization processing; The data processed by layer normalization is processed by the second convolution layer in one way, and the other way forms a residual connection with the data processed by the second convolution layer and is output after layer normalization.

2. The rotating machinery fault diagnosis method fused with convolutional Transformer according to claim 1, characterized in that: The first convolution layer uses 1×1 convolution to perform channel adjustment and feature integration, and the second convolution layer uses 3×1 convolution to extract spatial features.

3. The rotating machinery fault diagnosis method fused with convolutional Transformer as claimed in claim 2, characterized in that: The step size of the 1×1 convolution is 2, and the step size of the 3×1 convolution is 1.

4. The rotating machinery fault diagnosis method fused with convolutional Transformer according to claim 1, characterized in that: There are 4 layers of feature extraction modules.

5. The rotating machinery fault diagnosis method fused with convolutional Transformer as claimed in claim 1, characterized in that: The multi-channel signal input to the diagnostic model is first processed by a convolution layer with a convolution kernel of 64×1, and then sent to the first-layer feature extraction module.

6. The rotating machinery fault diagnosis method fused with convolutional Transformer as claimed in claim 1, characterized in that: The data output by the feature extraction module in the last layer is processed by global average pooling and then outputs the diagnosis result through the softmax classification layer.

7. The rotating machinery fault diagnosis method fused with convolutional Transformer as claimed in claim 1, characterized in that: The multi-head self-attention layer uses multiple sets of parallel self-attention to extract features, and its formula is: , where: X is the input data, is the weight matrix, Indicates a connection operation. is the output of 1 to n groups of self-attention, n is the number of heads of multi-head self-attention; The formula for each group of self-attention is: ； Among them, Q, K, and V are query matrix, key matrix, and value matrix respectively. is the transpose of the key matrix K, Represents the dimension of K.

8. The rotating machinery fault diagnosis method fused with convolutional Transformer according to claim 1 or 7, characterized in that: The number of heads in the multi-head self-attention layer is 2.

9. The rotating machinery fault diagnosis method fused with convolutional Transformer as claimed in claim 1, characterized in that: The diagnostic model is trained using a data set consisting of vibration information collected in normal working conditions and various fault conditions.

Images