A Fault Diagnosis Method for Aeroengine Rotor System Based on Deep Learning
Through the combination of sliding window algorithm and deep residual shrinking network, the problem of fault feature extraction under high noise of aero engine bearing vibration signals is solved, and high-accurate fault diagnosis is achieved.
Patent Information
- Application Number
- CN202111237432.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-22
- Publication Date
- 2025-07-04
- Estimated Expiration
- 2041-10-22
AI Technical Summary
Due to high noise and complex transmission paths, the fault characteristics are difficult to extract, and the prior art is difficult to accurately diagnose faults in the context of high noise.
The sliding window algorithm is used to re-sample and construct the fault sample set, a deep residual shrinking network is established, a wide convolution kernel and a progressive semi-soft threshold function are used for data noise reduction, and fault identification is performed through a fully connected layer and a Softmax classifier.
It realizes accurate identification of fault characteristics in high-noise environments, improves the accuracy and noise immunity of fault diagnosis, and has efficient fault identification capabilities.
Smart Images
Figure CN113962264B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a rotor system of an aero-engine, and designs a fault diagnosis algorithm based on deep learning, belonging to the technical field of fault diagnosis. Background Art
[0002] As a core component of an aircraft, the aero-engine plays an important role in providing power for the aircraft, and has an important impact on the performance and reliability of the aircraft. Moreover, the working conditions of the aero-engine are relatively harsh and it is easy to fail. The rolling bearing is an important part of the rotor system of the aero-engine. If the rolling bearing is damaged, it may cause property losses and even casualties. Therefore, the research on the fault diagnosis of aero-engine bearings is of great significance.
[0003] The methods of fault diagnosis are mainly divided into those based on physical models, signal processing, and intelligent diagnosis. The fault diagnosis method based on a physical model has good robustness, but the premise of this method is to establish an accurate model for the target system. Therefore, this method is not easy to implement in practical applications; the diagnosis method based on signal processing does not require the establishment of a quantitative or qualitative mathematical model of the system. It only needs to collect the original data and extract the fault features from the original data through signal processing to diagnose the system fault; with the development of computational intelligence, deep learning has become a hot topic in the research field of fault detection. After preprocessing the original data, the diagnosis method based on a neural network inputs it into the neural network model, and finally directly obtains the fault diagnosis result.
[0004] Due to problems such as the complex vibration transmission path of the vibration signal of the aero-engine bearing, the obtained original vibration signal has the characteristics of high noise and multi-interference information, which causes the problem of difficult extraction of fault features. Therefore, how to extract effective fault features from the original data under the background of high noise interference is the focus of attention in the field of fault diagnosis. Summary of the Invention
[0005] Object of the Invention: In view of the above research background and based on the existing technical methods, aiming at the problems of high noise and difficult feature extraction of the vibration signal of the aero-engine, a fault diagnosis algorithm for the rotor system of the aero-engine based on deep learning is proposed.
[0006] Technical solution: To achieve the above-mentioned invention purpose, the present invention proposes a novel fault diagnosis algorithm for an aero-engine rotor system based on deep learning, which is characterized in that: the sliding window algorithm is used to repeatedly sample the original data to achieve data augmentation, and after normalization, a fault sample set is constructed. A deep residual shrinkage network is established for automatic feature extraction, and the first convolutional layer is set as a wide convolutional kernel. The progressive semi-soft threshold function algorithm is used as the shrinkage layer of the network to achieve data denoising. Through the fully connected layer and the Softmax classifier, the fault recognition and classification results are obtained, including the following specific steps:
[0007] Step 1) Through the vibration signal data acquisition system, collect the acceleration vibration signals of the rolling bearings of the aero-engine under different faults, and construct the original vibration signal samples;
[0008] Step 2) Sample and overlap the original data samples using a sliding window:
[0009]
[0010] where n is the number of samples after overlapping sampling, L raw is the length of the original data, L sample is the length of a single sample, that is, the window width, and P is the moving step of the sliding window, that is, the sampling interval;
[0011] The data after overlapping sampling is Data = [(x1, y1), …, (x i , y i ), …, (x n , y n )] T , Data is the data set after segmentation processing, x i is the data of a single vibration signal sample, and each sample contains 1024 sampling points. y i is the fault category label of the sample data;
[0012] Step 3) Normalize the data to achieve the unification of data dimensions:
[0013]
[0014] where x max and x min are the minimum value and the maximum value in respectively, and x′ is the sample after normalization;
[0015] Step 4) Randomly divide the samples obtained after normalization, and use 70% of the data as the training set and 30% of the data as the test set;
[0016] Step 5) Set up a new deep residual shrinkage network fault diagnosis model under Pytorch, and use the training set to train the fault diagnosis model. This network uses the Adam optimization algorithm and consists of a first-layer wide convolutional layer, several residual shrinkage modules, a fully connected layer, and a Softmax classifier. Finally, the fault classification result is output through the Softmax classifier. The structural parameters of the new deep residual shrinkage network fault diagnosis model are as follows:
[0017]
[0018]
[0019] Step 6) Input the training set into the trained deep residual shrinkage network model to identify the fault types of the test samples;
[0020] Beneficial effects: A fault diagnosis algorithm for an aero-engine rotor system based on deep learning proposed by the present invention uses an improved deep residual shrinkage network to propose an end-to-end fault diagnosis algorithm for an aero-engine rotor system based on deep learning. This algorithm has the following specific advantages:
[0021] (1) The present invention introduces a sliding window algorithm to achieve repeated interception sampling of the original data, expand the data samples, realize data augmentation, and improve the problem of fewer original data samples;
[0022] (2) The present invention sets the first-layer convolutional layer in the neural network model as a wide convolutional kernel, and the wide convolutional kernel can effectively extract short-time features in the samples, improving the feature extraction ability and anti-noise performance of the model;
[0023] (3) The present invention proposes a new deep residual shrinkage network model, which uses a progressive semi-soft threshold shrinkage function to fully retain the effective features of the signal on the basis of realizing vibration signal denoising, avoiding the signal distortion problem caused by the soft threshold function;
[0024] The method proposed by the present invention, as a fault diagnosis algorithm for an aero-engine rotor system based on deep learning, has certain practical application value, is easy to implement, has high accuracy, and can be widely applied to the bearing fault diagnosis of an aero-engine rotor system. Brief description of the drawings
[0025] Figure 1 is the flow chart of the diagnosis algorithm of the present invention;
[0026] Figure 2 is the schematic diagram of the basic module of the new deep residual shrinkage network of the present invention;
[0027] Figure 3 is the schematic diagram of the overall structure of the new deep residual shrinkage network of the present invention;
[0028] Figure 4 These are the training loss and accuracy curves of the novel deep residual shrinkage network of the present invention;
[0029] Figure 5 This is the confusion matrix of the novel deep residual shrinkage network of the present invention; Detailed implementation manners
[0030] The following further explains the present invention with reference to the accompanying drawings.
[0031] Figure 1 This is the fault diagnosis flowchart of an aero-engine rotor system based on deep learning, including the following specific steps:
[0032] Step 1) Through a vibration signal data acquisition system, collect the acceleration vibration signals of the rolling bearings of an aero-engine under different faults, and construct the original vibration signal samples;
[0033] Step 2) Use a sliding window to perform overlapping sampling on the original data samples:
[0034]
[0035] where n is the number of samples after overlapping sampling, L raw is the length of the original data, L sample is the length of a single sample, i.e., the window width, and P is the moving step of the sliding window, i.e., the sampling interval;
[0036] The data after overlapping sampling is Data = [(x1, y1), …, (x i , y i ), …, (x n , y n )] T , Data is the data set after segmentation processing, x i is the data of a single vibration signal sample, and each sample contains 1024 sampling points, y i is the fault category label of the sample data;
[0037] Step 3) Normalize the data to unify the data dimensions:
[0038]
[0039] where x max and x min are respectively the minimum and maximum values in, and x′ is the sample after normalization;
[0040] Step 4) Randomly divide the samples obtained after normalization, with 70% of the data as the training set and 30% of the data as the test set;
[0041] Step 5) Set up a new deep residual shrinkage network fault diagnosis model under Pytorch, and use the training set to train the fault diagnosis model;
[0042] The schematic diagram of the basic module of the deep residual shrinkage network is as Figure 2 shown;
[0043] The specific steps are as follows:
[0044] Step 5.1) After the input sample passes through the first convolutional layer, it enters the second convolutional layer through the ReLU activation function. The second convolutional layer constructs a substructure for obtaining the noise threshold;
[0045] Step 5.2) In this substructure, first take the absolute value of the input and perform global average pooling to obtain the mean parameter. Then, mine the features of the channels through two fully connected layers. Finally, obtain the attention weight parameter through the Sigmoid activation function, and each attention weight parameter acts on the feature vector of the corresponding feature channel;
[0046] Step 5.3) Multiply the attention weight parameter by the mean parameter to obtain the noise threshold. Therefore, each feature channel has an independent noise threshold;
[0047] Step 5.4) Finally, use the obtained threshold to perform progressive semi-soft threshold processing on the sample data. After adding the result of the threshold processing to the residual term of the cross-layer identity mapping, the final module output is obtained;
[0048] The formula for progressive semi-soft threshold is as follows:
[0049]
[0050] where τ is the threshold, and a' and a s are the sample data before and after progressive semi-soft threshold processing, respectively;
[0051] Step 5.5) The schematic diagram of the overall structure of the deep residual shrinkage network is as Figure 3 shown. The overall deep residual shrinkage network uses the Adam optimization algorithm and consists of a wide convolutional layer in the first layer, two residual shrinkage modules, a fully connected layer, and a Softmax classifier. Finally, the fault classification result is output through the Softmax classifier to achieve fault feature extraction in high-noise vibration signals. The structural parameters of the new deep residual shrinkage network fault diagnosis model are as follows:
[0052]
[0053] Step 6) Input the training set into the trained deep residual shrinkage network model to identify the fault diagnosis results of the test samples.
[0054] The fault diagnosis algorithm of the aero-engine of the present invention has a high fault recognition accuracy rate. The CWRU bearing data of the Bearing Data Center of Case Western Reserve University in the United States is used to verify this fault diagnosis algorithm, and finally an accuracy rate of 99.21% is obtained. This algorithm does not need to perform other noise reduction processing on the original data, and realizes end-to-end fault diagnosis through feature self-extraction. Figure 4 It is the training curve of the fault diagnosis model. It can be seen that the final accuracy rates of this fault diagnosis algorithm in the test set and the training set both reach more than 99%.
[0055] The confusion matrix is used to measure the accuracy of the fault diagnosis algorithm of the present invention. The confusion matrix calculates the number of correct classifications and incorrect classifications of the model to realize the performance evaluation of the diagnosis model. Figure 5 This is the confusion matrix of the fault diagnosis algorithm of the present invention. As Figure 5 shown, the abscissa of the confusion matrix is the fault diagnosis result of the diagnosis model, and the ordinate is the actual fault category label. It can be seen that the fault diagnosis algorithm of the present invention can effectively identify various faults of the rolling bearing, and the diagnosis effect is relatively accurate.
[0056] It can be seen that the deep residual shrinkage network integrates the progressive semi-soft threshold mechanism into the residual network, and automatically sets different thresholds for each threshold module with the help of the attention mechanism, realizes the processing of the original data, avoids the interference brought by noise information, and can realize the extraction of fault features from noise information.
Claims
1. A fault diagnosis method for an aero-engine rotor system based on deep learning, comprising the following specific steps: Step 1) Collect the acceleration vibration signals of the rolling bearings of the aero-engine under different faults through a vibration signal data acquisition system, and construct the original vibration signal samples; Step 2) Overlap-sample the original data samples by sampling sliding windows: Among them, n is the number of samples after overlapping sampling, L raw is the length of the original data, L sample is the length of a single sample, i.e., the window width, and P is the moving step of the sliding window, i.e., the sampling interval; The data after overlapping sampling is Data = [(x1, y1), …, (x i , y i ), …, (x n , y n )] T , Data is the data set after segmentation processing, x i is the single vibration signal sample data, each sample contains 1024 sampling points, y i is the fault class label of the sample data; Step 3) Normalize the data to unify the data dimensions: where x max and x min are the minimum and maximum values in, respectively, and x′ is the sample after normalization; Step 4) Randomly divide the samples obtained after normalization, with 70% of the data as the training set and 30% of the data as the test set; Step 5) Set up a new deep residual shrinkage network fault diagnosis model under Pytorch, and use the training set to train the fault diagnosis model. This network uses the Adam optimization algorithm and consists of a first-layer wide convolutional layer, several residual shrinkage modules, a fully connected layer, and a Softmax classifier. Finally, the fault classification result is output through the Softmax classifier. The specific steps are as follows: Step 5.1) After the input sample passes through the first convolutional layer, it enters the second convolutional layer through the ReLU activation function. The second convolutional layer constructs a sub-structure for obtaining the noise threshold; Step 5.2) In this sub-structure, first perform absolute value and global average pooling processing on the input to obtain the mean parameter, then mine the features of the channels through two fully connected layers, and finally obtain the attention weight parameter through the Sigmoid activation function. Each attention weight parameter acts on the feature vector of the corresponding feature channel; Step 5.3) Multiply the attention weight parameter by the mean parameter to obtain the noise threshold. Therefore, each feature channel has an independent noise threshold; Step 5.4) Finally, use the obtained threshold to perform progressive semi-soft threshold processing on the sample data. After adding the result of the threshold processing to the residual term of the cross-layer identity mapping, the final module output is obtained; The formula for progressive semi-soft threshold is as follows: where τ is the threshold, and a' and a s are the sample data before and after the progressive semi-soft threshold processing, respectively; Step 6) Input the training set into the trained deep residual shrinkage network model to identify the fault types of the test samples.