Aero-engine main shaft bearing transfer learning fault diagnosis method based on BI-sandwich model and MMSD algorithm
Through the sandwich network model of multi-scale convolutional neural network and BSA module, combined with the maximum mean square difference and MMSD algorithm, the domain offset problem in the fault diagnosis of aero engine spindle bearings is solved, and high accuracy is achieved under complex operating conditions Fault identification demonstrates better robustness and generalization capabilities.
Patent Information
- Application Number
- CN202410988548.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2024-11-08
AI Technical Summary
The domain offset problem of aircraft engine spindle bearings under complex operating conditions, which affects the generalization performance of traditional machine learning fault diagnosis methods, and it is unrealistic to frequently retrain the model.
Transfer learning is performed using a sandwich network model of multi-scale convolutional neural network and BSA module. The domain deviation between the source domain and the target domain is reduced through the maximum mean square difference and MMSD algorithm, and the E-Loss loss function is introduced to enhance the Lu of the model. Awesomeness and generalization ability.
It improves the accuracy and robustness of fault diagnosis of aircraft engine spindle bearings, can effectively identify fault types under different operating conditions, and performs better than other transfer learning methods.
Smart Images

Figure SMS_1 
Figure SMS_3 
Figure SMS_5
Abstract
Description
Technical Field
[0001] The present invention relates to a fault diagnosis technology for an aero-engine main shaft bearing, and in particular to a method for performing fault diagnosis on an aero-engine main shaft bearing by utilizing a transfer learning method. Background Art
[0002] Aircraft engines are one of the key components of aircraft and are closely related to the performance, efficiency and safety of aircraft. The main shaft bearing is the core component that supports the high-speed rotation of the aircraft engine shaft system, and its importance in aircraft engines cannot be ignored. The main shaft bearing is directly exposed to high-temperature airflow and gas flow channels, and at the same time bears the huge centrifugal force and inertia force brought by high-speed rotating fans, turbines and other rotating components. In addition, the main shaft bearing needs to resist high-frequency vibrations such as unbalanced vibration and gas pulsation. During the operation of the aircraft, the main shaft bearing will experience different working conditions, including starting, acceleration, cruising, landing, etc. Under complex working conditions, the main shaft bearing will inevitably produce performance degradation and even cause various failures. For example, there will be failure types such as contact fatigue, dents, wear, corrosion, misalignment, and contamination. Whether the engine main shaft fails or not directly affects it, so timely and accurate diagnosis and resolution of the main shaft bearing failure is crucial to ensure aviation safety.
[0003] With the vigorous development of the aerospace industry and the continuous improvement of flight safety requirements, the requirements for the stability of aircraft engines are getting higher and higher, which requires the main shaft fault inspection and repair technology to be faster, more timely and more effective. In recent years, the development of intelligent fault diagnosis technology has made aircraft engine maintenance more intelligent and efficient, providing a more reliable, safe and economical solution for the aviation industry. Specifically, the fault diagnosis technology of aircraft engine main shaft bearings can be roughly divided into two categories: model-based methods and data-driven methods. Model-based methods more specifically refer to the use of physical models of the system for fault diagnosis and state estimation. For example, Kalman filtering, particle filtering, etc. are model-based filtering algorithms that rely on accurate modeling of system dynamics to diagnose aircraft engine faults. Although model-based methods can obtain relatively accurate and interpretable models. However, model-based methods may face problems such as inaccurate models, challenges to complex systems, and deterministic requirements for fault information in some scenarios. In contrast, data-based methods are more flexible and practical, and can adapt to a variety of systems and working conditions, and can learn system patterns from a large amount of measured data. With the advent of the big data era, researchers have turned their attention to data-driven methods for fault diagnosis of aircraft engine main shaft bearings.
[0004] Specifically, the method of aircraft engine main shaft bearing fault diagnosis based on machine learning may not be ideal in actual application scenarios. Because traditional machine learning algorithms assume that training data and test data come from the same distribution, in actual applications, this assumption may not always hold true, especially when facing distribution shift. Faced with machines with complex working conditions, their speed, load, pressure, lubrication conditions, etc. may change during operation. Obviously, it is unrealistic to obtain enough labeled data under various working conditions. Therefore, these fault diagnosis algorithms that achieve good diagnostic results may only be used in laboratories, and are still a long way from actual engineering applications. Data-driven aircraft engine main shaft bearing faults have the following problems that need to be solved:
[0005] 1) The same model of aircraft engine main shaft bearings have large differences in operating speed, workload, service life, maintenance, operations of different pilots, environmental noise, etc., which directly leads to large domain offsets in the collected bearing data. 2) Data-based fault diagnosis methods require sufficient labeled data under normal and various fault conditions to train the model. However, there is often a large distribution difference between the newly detected fault data and the data used for model training. Significant data differences greatly affect the generalization performance of traditional machine learning fault diagnosis methods.
[0006] 3) Obviously, it is unrealistic to frequently retrain the model with new data from actual measurements to enhance the generalization ability of the model, not only because it is difficult to obtain actual data, but also because data labeling requires a lot of manpower and material resources.
[0007] The bearing fault diagnosis method based on transfer learning provides a feasible solution to the above problems and is expected to be applied in engineering practice. Summary of the invention
[0008] The present invention provides a novel aircraft engine main shaft bearing fault diagnosis method by transfer learning, comprising the following steps:
[0009] Step 1: Obtain the original vibration signals under multiple working conditions, and divide the collected data samples into a source domain training set, a source domain verification set, and a target domain verification set.
[0010] Step 2: Input the source domain samples into the sandwich network model based on the embedded multi-scale convolutional neural network and BSA module for network training and verification. Use the multi-scale separable convolutional neural network to obtain sample features of different fine-grained levels, and use the BSA module for global feature information.
[0011] The embedded multi-scale convolutional neural network includes:
[0012] First, the cross-channel information is integrated. The specific operation is to use a convolution layer with a kernel size of 1 to process the input signal. The integration and output z is:
[0013]
[0014] in Represents the convolution weight value of the convolution kernel.
[0015] The integrated data enters the multi-scale parallel separable convolution. The one-to-one correspondence between input and output is a unique operation of separable convolution, which reduces the amount of model calculation of the network. Its output is as follows:
[0016]
[0017] in represents the weights of the separable convolution with kernel size kl.
[0018] The BSA module:
[0019] This module designs a BSA module based on broadcast operation to obtain global feature information, which is characterized by only needing to perform one-dimensional traversal of data samples. The method is as follows:
[0020] Compute the time fraction τ, the key matrix κ, and the value matrix v:
[0021]
[0022] Among them, ReLU(·) is the activation function, w τ ,w κ ,w v is the weight parameter.
[0023] The broadcast operation matches the time score τ with the key matrix κ. Then, the summation operation is performed over the time dimension to obtain the feature weight γ:
[0024]
[0025] in, Represents a broadcast operation.
[0026] After the broadcast operation, the value matrix of the weights is passed in, and important features are obtained from the global scope of the signal. The output features are mapped and can be expressed as:
[0027]
[0028] in, Represents the convolution weight.
[0029] The sandwich model specifically comprises:
[0030] Input layer, middle layer, output layer. The middle layer is arranged as a small sandwich structure, consisting of 3 identical modules connected in series. (The small and large sandwich structures can be changed to inner-sandwich, outer-sandwich structures).
[0031] The input layer performs "slimming" processing on a large number of input signals, reduces the time dimension of the data, and extracts data features. The input layer consists of two average pooling layers, a convolutional layer, a BN and a GELU layer.
[0032] The middle layer is the content layer of the big sandwich model and the core of the model. The signal processed by the input layer enters the SMC, and the local features of different fine-grained levels are obtained with the help of multi-channel parallel convolutional neural networks, and the key information of the signal is extracted using the BSA module. In addition, cross-layer connections are added to make full use of the characteristics of the signal and optimize the model performance. The small sandwich module consists of 3 feature extraction modules connected in series, and the output is processed by layer normalization and enters the next layer.
[0033] The output layer is composed of an average pooling layer and a linear layer. The output layer performs average pooling on the data, which can integrate the information within the batch to improve the robustness of the model and make it more robust to slight changes in the input data.
[0034] Step 3: Input the target domain samples into the network that has been trained with the source domain training set and verified with the source domain verification set. Use the trained network to diagnose the target domain data, introduce the calculation of the second-order statistic of maximum mean square difference to reduce the domain deviation between the source domain and the target domain, and further optimize the fault diagnosis effect of transfer learning. In addition, E-Loss is proposed, which introduces the loss based on the MMSD algorithm to better focus on the offset between the source domain and the target domain.
[0035] The MMSD algorithm is defined as follows:
[0036] The working mechanism of MMD shows that the kernel function k(x, ) is a sample mapping in the infinite-dimensional RKHS. The difference in distribution between the mean square statistics of two domains in RKHS can be easily measured by and
[0037]
[0038] Among them, H is the Hilbert space, and according to the properties of the space, is a new Hilbert space.
[0039] After the above derivation and the processing to reduce the computational complexity, the square of MMSD is defined as:
[0040]
[0041] In order to apply the MMSD metric in the above formula to actual situations, the partial empirical MMSD statistic is given as follows:
[0042]
[0043] Furthermore, the specific method of E-Loss in S3 is as follows:
[0044] The traditional cross entropy loss function calculates the loss by measuring the difference between the probability distribution predicted by the model and the true label:
[0045]
[0046] Where N is the number of samples, C is the number of categories, and y ij is the true label of the jth category of the ith sample, which is 1 if the ith sample belongs to the jth category, otherwise it is 0; is the model's predicted probability of the jth category for the i-th sample;
[0047] Consider L calculated by the MMSD formula DA , the enhanced loss that considers both classification loss and mean square loss is as follows:
[0048] E-Loss=Cross-Entropy Loss+αL DA (9)
[0049] Among them, α is the weight factor. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 Flow chart of the present invention
[0051] Figure 2 It is a multi-scale separable convolution
[0052] Figure 3 For BSA module
[0053] Figure 4 Bi-Sandwich Model
[0054] Figure 5 University of Turin Testbed
[0055] Figure 6 . Schematic diagram of bearing defects on the test bench of the University of Turin
[0056] Figure 7 Confusion matrix for case 1
[0057] Figure 8 . For case 1 t-SNE clustering diagram
[0058] Fig. 9 .ROC curve for case 1
[0059] Fig.10 . This is the accuracy ring diagram of the ablation experiment in case 1
[0060] Fig.11 . Scoring chart for case 1F1 DETAILED DESCRIPTION
[0061] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0062] The bearing data set was collected from the aviation high-speed bearing test bench built by the DIRG laboratory of the Department of Mechanical and Aeronautical Engineering of the Polytechnic University of Turin. Figure 5 The test bench consists of a spindle, bearings, cooling circuit, sleds, loading springs, etc.
[0063] The two bearings used in the experiment are both special aviation high-speed bearings, and their specific parameters are shown in Table 1. The bearing fault at B1 is machined by Rockwell hardness tool to produce a bearing defect (such as Figure 6 The bearing health types and corresponding labels are shown in Table 2.
[0064] Table 1 Bearing parameters
[0065]
[0066] Table 2 Bearing defect details
[0067]
[0068] Design a transfer learning bearing fault diagnosis study under different loads. Select 18 data files of 7 different health conditions of bearings with loads of 1000N, 1400N, and 1800N at a speed of 200HZ, and the sampling length is set to 1024. A total of 6 different transfer learning tasks are designed, as shown in Table 3. The number of samples in the source domain training set, the number of samples in the source domain validation set, and the number of samples in the target domain validation set for each migration task is 2730, 684, and 684. The running equipment is as follows: the software is Pytorch1.10.2; the CPU is i5-12400F; the GPU is RTX3050.
[0069] Set the learning rate to 1×10 -4, a smaller batch size of 16 was selected, and the Adam optimizer was used for parameter optimization. A total of 300 epochs were trained, and each migration task was repeated 5 times.
[0070] Table 3 Transfer tasks and outcomes
[0071]
[0072] The accuracy of our method in the above migration tasks is shown in Table 3. Figure 7 The confusion matrix of Case 1 more intuitively represents the fault identification performance of this method. It can be seen that this method has good fault diagnosis performance in 6 tasks. Figure 8 To further illustrate the superiority of the present invention, it can be seen that the present invention can separate the fault types well.
[0073] To prove the superiority of the present invention, the present invention is compared with popular transfer learning bearing fault diagnosis methods such as FTNN, CNN_1D, MK-CNN, Ada-Resnet, and DA. In each case, 10 experiments are performed and the average value is taken. The experimental results are shown in Table 4 and Fig. 9 In addition, Fig.11 The classification of each type of fault by the evaluation model shows that the present invention has a good performance in each type of fault in each migration task, with the highest F1 score.
[0074] In order to further illustrate the effectiveness of the present invention, an ablation experiment was conducted on the method in this paper. In order to study the effect of the MMSD algorithm, it was compared with the current distance measurement method. This paper sets up the sandwich model without a distance measurement function, the sandwich model + MMD, the sandwich model + MK-MMD, and the method of the present invention. Verification was performed under each group of migration tasks, and the results are shown in Table 4 and Fig.10 shown.
[0075] Table 4. Ablation experiments on the Politecnico di Torino dataset
[0076]
[0077]
Claims
1. A method for fault diagnosis of aero-engine main shaft bearing by transfer learning, characterized in that A new fault diagnosis method for aero-engine main shaft bearings based on BI-sandwich model and MMSD algorithm.
2. As described in claim 1, the BI-sandwich model is based on the BSA module, a multi-scale convolutional neural network, including a large sandwich module and a small Meiji structure.
3. As described in claim 2, the big sandwich module consists of an input layer, a middle layer and an output layer.
4. The big sandwich input layer as described in claim 3 consists of two average pooling layers, a convolutional layer, a BN and a GELU layer.
5. In the large sandwich middle layer as described in claim 3, the signal processed by the input layer enters the SMC, and the local features of different fine-grained levels are obtained by means of a multi-channel parallel convolutional neural network, and the key information of the signal is extracted using the BSA module. In addition, cross-layer connections are added to make full use of the characteristics of the signal and optimize the model performance. The small sandwich module consists of three feature extraction modules connected in series, and the output is processed by layer normalization and enters the next layer.
6. The large sandwich output layer as claimed in claim 3 is composed of an average pooling layer and a linear layer. The output layer performs average pooling on the data, which can integrate the information within the batch to improve the robustness of the model and make it more robust to slight changes in the input data.
7. The MMSD algorithm as claimed in claim 1 is based on a mean square value, which is defined as:
8. The BSA module as claimed in claim 2 is designed based on broadcast operation, characterized in that Only one-dimensional traversal of the data samples is required to reduce the computational complexity of the network while obtaining global information. Formula 1 represents the time fraction τ, the key matrix κ and the value matrix v, Formula 2 represents the feature weight, and Formula 3 represents the feature map.
9. As described in claim 2, a multi-scale separable convolutional neural network is designed, characterized in that MSSC reorganizes the structure of cross-channel convolution and separable convolution to further reduce the amount of computation.
10. As described in claim 9, the multi-scale separable convolutional neural network construction comprises: First, the cross-channel information is integrated. The specific operation is to use a convolution layer with a kernel size of 1 to process the input signal. The integration and output z is: The integrated data enters the multi-scale parallel separable convolution. The one-to-one correspondence between input and output is a unique operation of separable convolution, which reduces the amount of model calculation of the network. Its output is as follows:
11. The second-order statistic of mean square value is introduced, and the maximum mean square difference metric is constructed to measure the difference between different domains. Its value is embedded in the loss to minimize the loss while reducing the domain offset. E-Loss=Cross-Entropy Loss+αL DA (10) Formula 9 is the traditional cross entropy loss function, which calculates the loss by measuring the difference between the probability distribution predicted by the model and the true label. Formula 10 is the L calculated by the MMSD formula (13). DA , we get the enhanced loss that considers both classification loss and mean square loss.
