A cycle-gan-based aero-engine inter-shaft bearing fault diagnosis method under class imbalance condition

Abstract

The application discloses a kind of under the condition of aviation engine inter-shaft bearing fault diagnosis method of class imbalance based on Cycle-GAN, and its specific include: data acquisition and pre-processing, improvement and establish Cycle-GAN model, train improved Cycle-GAN model and complete signal conversion, complete the fault diagnosis of actual vibration signal by Cycle-GAN model.The application proposes a new fault diagnosis transfer learning method based on Cycle-GAN, by improved Cycle-GAN model, signal sample under known condition is converted into new signal sample under unknown condition, provide the conversion signal of data distribution closer to real signal, while also ensure that conversion signal retains the fault category information in original signal, solve the problem of fault data scarcity in actual scene, and use conversion signal to train classifier, so that the classifier can distinguish fault data under unknown condition.In practical engineering application, especially in the problem of aviation engine inter-shaft bearing fault diagnosis under the condition of class imbalance, the application has wide application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bearing fault diagnosis, and particularly relates to an aero-engine inter-shaft bearing fault diagnosis method under a class imbalance condition based on Cycle-GAN. BACKGROUND

[0002] Aero-engine is known as the "industrial flower", which is the core machine to provide power. Inter-shaft bearing, as the key component in the dual-rotor system of aero-engine, plays a core role in connecting and supporting. However, the inter-shaft bearing of aero-engine is installed between the high-pressure rotor and the low-pressure rotor. The unique installation location makes it long time in the harsh operating environment of high speed, high intensity impact and heavy load, which often leads to local failure of the bearing. Such failure will make the bearing show unique and complex vibration mode, and then directly affect the safety and reliability of the engine, leading to major safety accidents and endangering people's life and property safety. Therefore, efficient and accurate fault diagnosis method of inter-shaft bearing is of great significance to ensure the stable operation of the engine. In recent years, deep learning has made great achievements in computer vision and natural language processing, and has been widely used in fault diagnosis of industrial equipment and key components due to its powerful feature extraction capability. The effectiveness of deep learning-based fault diagnosis method is based on the assumption that the data should conform to the same distribution. However, in the actual industrial environment, the changes in working conditions such as load and speed will cause a large difference in data distribution, which will lead to poor generalization of the intelligent diagnosis model. As a sub-field of transfer learning, domain adaptation method is often used to solve this problem, which can transfer the knowledge of one domain to similar domains and reduce the data distribution difference between different domains, providing many effective methods for cross-domain fault diagnosis problems, including maximum mean discrepancy (MMD), domain adversarial neural network (DANN) and joint distribution adaptation (JDA). However, as a high-reliability industrial product, aero-engine usually runs in normal state, and mechanical failure is occasional and uncertain, which means that the bearing samples in normal state are abundant, while the bearing samples in fault state are scarce. In another case, the number of samples in a certain fault state is much smaller than that in other fault states, and it is difficult to collect enough fault samples of aero-engine inter-shaft bearing and construct an ideal data set. The sample data between different health states will have a class imbalance problem. When training the model with class imbalance data set, accurate diagnosis results cannot be obtained, which is an important factor limiting the practical application of intelligent diagnosis model in industrial scene. In order to realize fault diagnosis under class imbalance condition, many studies on class imbalance fault diagnosis have been carried out, among which generative adversarial network (GAN) has shown great application potential in solving such problems. GAN is a data generation model that can generate new sample data similar to the distribution of real data through the adversarial training of generator and discriminator. As a variant model of GAN, cycle generative adversarial network (Cycle-GAN) can convert labeled sample data from known scenarios to unknown scenarios without paired data, which is beneficial to solve the problem of no data in the actual scene of fault diagnosis.The application is improved based on Cycle-GAN, and the improved Cycle-GAN model is used to convert signals under known conditions into signals under unknown conditions, so as to reduce the data distribution difference between the converted signals and the real signals, improve the quality and quantity of the converted signals, and realize fault diagnosis of the aero-engine inter-shaft bearing under the class imbalance condition while the data characteristics and fault category information are reserved. The application proposes an unsupervised transfer learning network model based on Cycle-GAN, and the improved Cycle-GAN can be applied to sample data under different conditions, new samples under unknown conditions are generated through sample data under known conditions, and the model classifier is trained by using the new sample data to realize fault diagnosis of the sample under unknown conditions. SUMMARY

[0003] The application aims to solve the problem of fault diagnosis of the aero-engine inter-shaft bearing under the condition of insufficient fault data, and proposes a class imbalance condition-based aero-engine inter-shaft bearing fault diagnosis method based on Cycle-GAN.

[0004] The technical scheme of the application is: a class imbalance condition-based aero-engine inter-shaft bearing fault diagnosis method based on Cycle-GAN, comprising the following steps:

[0005] A. Obtain vibration signal samples of the engine inter-shaft bearing under normal and fault conditions under different working conditions, and divide them to obtain a training data set and a test data set;

[0006] B. Improve and establish a Cycle-GAN model;

[0007] C. Use the improved Cycle-GAN model to complete the conversion of the vibration signal under the known condition to the vibration signal under the unknown condition, and train the classifier of the Cycle-GAN model through the new converted signal;

[0008] D. Diagnose the actual bearing vibration signal by using the Cycle-GAN model trained in step C to obtain the result.

[0009] Further, the step A obtains vibration signal samples of the engine inter-shaft bearing under normal and fault conditions under different working conditions, and divides the training data set and the test data set, which specifically comprises the following steps:

[0010] A1. Select the bearing signal under the normal state under different working conditions as the source domain data (wherein represents signal data, represent the corresponding class label), for the simulation of fault diagnosis scenarios under the condition of class imbalance, randomly select the fault state bearing signal (signal that is not easy to obtain in reality) under one of the corresponding working conditions as the target domain data The above two types of data together constitute the training data set as the signal sample under the known condition;

[0011] A2. Select the same fault state bearing signal under the remaining working conditions in the corresponding working conditions as the test data set, as the signal sample under the unknown condition.

[0012] Further, the step B improves and establishes the Cycle-GAN model, and the improved Cycle-GAN model includes two generators and , two discriminators and and two classifiers and , specifically

[0013] B1. The generator of the improved Cycle-GAN model is composed of an encoder, a converter and a decoder, wherein the encoder is stacked by three convolutional layers, the converter is composed of three residual blocks connected in turn, and in each residual block, the first layer convolutional layer is replaced by an expansion convolutional layer, and the decoder is composed of three deconvolutional layers.

[0014] B2. The discriminator of the improved Cycle-GAN model is composed of four convolutional blocks and a convolutional layer, and each convolutional block is composed of a convolutional layer, an instance normalization layer and a LeakyReLU activation function.

[0015] B3. The classifier of the improved Cycle-GAN model is composed of a convolutional layer and a fully connected layer connected in turn.

[0016] Further, the step C uses the improved Cycle-GAN model of step B to complete the conversion of the vibration signal from the known condition to the unknown condition, and in the training process, the mapping directions of the two generators and are opposite, and the signal needs to be converted from the source domain to the target domain through forward conversion and reverse conversion, and the new converted signal is used to train the classifier, specifically including the following steps:

[0017] C1. First, the generator converts the source domain signal into the target domain converted signal at this time, the generator learns the mapping from the source domain to the target domain , and then uses the discriminator discriminate the converted signal from the target domain real signal , through and adversarial training makes the generated converted signal highly consistent with the data distribution of the target domain signal , and finally the generator converts the converted signal into a signal in the source domain , at this time the generator learns the mapping from the target domain to the source domain , and the forward conversion of the source domain signal should make the forward cycle hold;

[0018] C2. First, the target domain signal is converted into a source domain converted signal by the generator , at this time the generator learns the mapping from the target domain to the source domain , and then the discriminator discriminates the converted signal from the source domain real signal , through and adversarial training makes the data distribution of the generated converted signal highly consistent with the source domain signal , and finally the generator converts the converted signal into a signal in the target domain , at this time the generator learns the mapping from the source domain to the target domain , and the reverse conversion of the target domain signal should make the reverse cycle hold;

[0019] C3. The conversion of the signal from the known condition to the unknown condition is realized by steps C1 and C2, which improves the similarity between the converted signal and the real signal while ensuring that the samples generated by the two generators do not contradict each other, and in the training process, the target domain classifier is used to classify the target domain converted signal and the target domain real signal , to further ensure that the target domain classifier can correctly classify the real signal under unknown conditions.

[0020] ​Further, the step D utilizes the Cycle-GAN model trained in step C to diagnose the actual bearing vibration signal to obtain a result, specifically: taking the test data set divided in step A as the input of the target domain classifier of the Cycle-GAN model after training, realizing the fault diagnosis of the engine inter-shaft bearing under the condition of class imbalance.

[0021] The beneficial effects of the present application are: the present application proposes a new fault diagnosis transfer learning method based on Cycle-GAN, which converts the signal samples under known conditions into new signal samples under unknown conditions through the improved Cycle-GAN model, provides the converted signal with data distribution closer to the real signal, and also ensures that the converted signal retains the fault category information in the original signal, solves the problem of lack of fault data in the actual scene, and trains the classifier using the converted signal, so that the classifier can distinguish the fault data under unknown conditions. In actual engineering applications, especially in the problem of fault diagnosis of aero-engine inter-shaft bearings under the condition of class imbalance, the present application has a broad application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a flowchart of a fault diagnosis method for aero-engine inter-shaft bearings under the condition of class imbalance based on Cycle-GAN.

[0023] Figure 2 is a structure diagram of the improved Cycle-GAN model in the present application. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0025] As shown in Figure 1 is a flowchart of a fault diagnosis method for aero-engine inter-shaft bearings under the condition of class imbalance based on Cycle-GAN. The fault diagnosis method for aero-engine inter-shaft bearings under the condition of class imbalance based on Cycle-GAN comprises the following steps:

[0026] A. Obtain vibration signal samples of engine inter-shaft bearings under normal and fault conditions under different working conditions, and divide them to obtain a training data set and a test data set;

[0027] ​A1. Obtain bearing vibration signals under four different working conditions, specifically through sensors installed near the engine inter-shaft bearing to collect signals. The engine is usually in normal operation, and the bearing signal sample under normal state is abundant, that is, the bearing vibration signals under four working conditions in normal state are selected as source domain data. The fault occurs occasionally, resulting in a lack of fault signal samples, that is, the bearing signals under a fault state of one working condition are randomly selected as target domain data. The above two types of data together constitute a training data set as a known condition signal sample;

[0028] A2. Select the bearing signals under the same fault state of the remaining three working conditions as the test data set, as the unknown condition signal sample. According to steps A1 and A2, the data set is divided to simulate the fault diagnosis scene under the condition of class imbalance.

[0029] B. Improve and establish a Cycle-GAN model, as shown in Figure 2 The improved Cycle-GAN model includes two generators and , two discriminators and , and two classifiers and .

[0030] B1. The generator of the improved Cycle-GAN model is composed of an encoder, a converter and a decoder. The signal input into the generator first enters the encoder composed of three convolutional layers stacked together to extract signal features. Then it enters the converter composed of three residual blocks connected together. The first convolutional layer of each residual block is replaced by an expansion convolution. Without increasing the network parameters, the receptive field of the convolution kernel is further expanded. The expansion rate is set to 2, so that the generator can extract deeper features during signal conversion. The extracted signal features are converted from the source domain to the target domain through the converter. The converted features finally enter the decoder composed of three deconvolutional layers. The learned features are restored through the decoder and inversely synthesized into the final converted signal.

[0031] B2. The discriminator of the improved Cycle-GAN model is composed of four convolutional blocks and one convolutional layer. Each convolutional block is composed of a convolutional layer, an instance normalization layer and a LeakyReLU activation function. This structure is conducive to the discriminant analysis of each small segment of the signal. During training, the model can pay more attention to the details of the signal. At this time, the role of the discriminator is to distinguish between real signals and converted signals. Through the adversarial training of the discriminator and the generator, the generator generates converted signals highly similar to the real signals.

[0032] B3. The classifier of the improved Cycle-GAN model is sequentially connected by three convolutional layers and one fully connected layer, and the final fault diagnosis result is obtained by inputting the to-be-tested signal into the classifier.

[0033] B4. The size of the first layer convolutional kernel of the generator, discriminator and classifier is 64x1, and the large convolutional kernel can help the convolutional layer to better extract the global domain features of the bearing signal. An instance normalization layer is connected after each convolutional layer, which can normalize the feature channels, ensure sample independence, align the feature distribution, and improve the stability of model training. The LeakyReLU activation function is used to replace the initial ReLU activation function, which enhances the feature extraction capability and improves the robustness and computational efficiency of the model.

[0034] C. According to the data set division mode of step A1, the Cycle-GAN model is trained with source domain training data and target domain training data respectively. In the training process, the generator learns the mapping and respectively, learns the data distribution of the corresponding target domain signal by taking the respective source domain signal as input, realizes the conversion of the signal from the source domain to the target domain, i.e. through the forward conversion and reverse conversion of the signal to complete the conversion of the vibration signal under the known condition to the vibration signal under the unknown condition, the new converted data (bearing signal under four working condition fault states) should correspond to the bearing signal under normal state, and finally the new converted signal is used to train the classifier of the Cycle-GAN model;

[0035] C1. Through the generator the source domain signal is converted into the target domain converted signal , and the discriminator is used to distinguish the converted signal and the real signal of the target domain , through the adversarial training of and , the data distribution of the converted signal generated by the generator is highly consistent with the target domain signal , and the generator converts into the signal in the source domain , and the forward conversion of the source domain signal should make the forward cycle hold. The loss back-propagated from the discriminator will help the generator and the discriminator to update the parameters, and The adversarial training loss between and is expressed in the form of cross entropy as follows:

[0036] {L}_{GAN\left ( {{G}_{st},{D}_{t},S,T} \right )}={E}_{{x}_{t}\sim {P}_{T}\left ( {{x}_{t}} \right )}\left [ {log{D}_{t}\left ( {{x}_{t}} \right )} \right ]+{E}_{{x}_{s}\sim {P}_{s}\left ( {{x}_{s}} \right )}\left [ {log\left ( {1-{D}_{t}\left ( {{x}^{\^}_{t}} \right )} \right )} \right ] (1)

[0037] where, and denote the expectation of and respectively.

[0038] C2. Through the generator convert the target domain signal into the source domain converted signal , and then use the discriminator to distinguish the converted signal from the source domain real signal , through and adversarial training so that the data distribution of the generated converted signal is highly consistent with the source domain signal , and the generator converts into the signal in the target domain , and the reverse conversion of the target domain signal should make the reverse cycle hold; and the adversarial training loss between

[0039] ​​{L}_{GAN\left ( {{G}_{ts},{D}_{s},S,T} \right )}={E}_{{x}_{s}\sim {P}_{S}\left ( {{x}_{s}} \right )}\left [ {log{D}_{s}\left ( {{x}_{s}} \right )} \right ]+{E}_{{x}_{t}\sim {P}_{T}\left ( {{x}_{t}} \right )}\left [ {log\left ( {1-{D}_{s}\left ( {{x}^{\^}_{s}} \right )} \right )} \right ] (2)

[0040] When both the forward cycle and the reverse cycle are true, it is considered that the model has completed the learning of the cycle-consistent mapping function. The training process is constrained by the cycle-consistency loss function, which can prevent the loss of useful information during training to improve the similarity between the converted signal and the real signal, and ensure that the generated samples by the generator and do not contradict each other. The cycle-consistency loss function is represented as follows:

[0041] {L}_{cycle\left ( {{G}_{st,}{G}_{ts}} \right )}={E}_{{x}_{s}\sim {P}_{S}\left ( {{x}_{s}} \right )}\left [ {{||{x}^{\~}_{s}-{x}_{s}||}_{1}} \right ]+{E}_{{x}_{t}\sim {P}_{T}\left ( {{x}_{t}} \right )}\left [ {{||{x}^{\~}_{t}-{x}_{t}||}_{1}} \right ] (3)

[0042] In the formula, represents the norm.

[0043] C3. The conversion of the signal from the known condition to the unknown condition is realized by steps C1 and C2. In addition, in order to make the converted signal consistent with the real signal of the same category, the source domain classifier is trained to ensure that the signal retains the category information in the original signal during the conversion process. The standard cross-entropy is used to train :

[0044] (4)

[0045] wherein, and denote the number of samples of source domain and target domain signals, respectively, and denote the class label space of source domain and target domain, respectively.

[0046] In the training process, the target domain classifier is used to classify the target domain converted signals and the target domain real signals to further ensure that the target domain classifier can correctly classify the real signals under unknown conditions, and the is trained by standard cross-entropy:

[0047] (5)

[0048] wherein, denotes the out-of-distribution source domain (belonging to the source domain but not belonging to the target domain) class label space, denotes the number of out-of-distribution source domain signal samples.

[0049] D. The test data (real fault signals) obtained according to the data set division method of step A2 is input into the trained Cycle-GAN model target domain classifier to realize the fault diagnosis of the engine inter-shaft bearing under the condition of class imbalance.

[0050] Those skilled in the art will realize that the embodiments described herein are for the purpose of helping the reader understand the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspirations disclosed in the present application without departing from the spirit of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A Cycle-GAN-based method for diagnosing faults of an aero-engine inter-shaft bearing under class imbalance conditions, characterized in that, The method comprises the following steps: A. obtaining vibration signal samples of engine inter-shaft bearings in normal state and different fault states under different working conditions, and dividing to obtain a training data set and a test data set; B. Improve and establish Cycle-GAN model; the model includes two generators G st and G ts , two discriminators D s and D t , and two classifiers C s and C t , specifically: B1. the improved generator of the Cycle-GAN model is composed of an encoder, a converter and a decoder, wherein the encoder is stacked by three convolutional layers, the converter is composed of three residual blocks connected in sequence, in each residual block, an expansion convolution is used to replace the first convolutional layer, and the decoder is composed of three deconvolutional layers; B2. the improved discriminator of the Cycle-GAN model is composed of four convolutional blocks and a convolutional layer, each convolutional block is composed of a convolutional layer, an instance normalization layer and a rectified linear unit activation function with leakage; B3. the improved classifier of the Cycle-GAN model is composed of a convolutional layer and a fully connected layer connected in sequence; C.Using the improved Cycle-GAN model to complete the conversion of vibration signals from known conditions to unknown conditions, in the training process, the mapping directions of the two generators G st and G ts are opposite, and the conversion of signals from the source domain to the target domain needs to be realized through forward conversion and reverse conversion, and the new converted signals are used to train the classifier, which specifically includes the following steps: C1. First, the generator G st converts source domain signal x s into target domain converted signal At this time, the generator G st learns the mapping S→T from the source domain (S) to the target domain (T), and then utilizes the discriminator D t to distinguish the converted signal from the target domain real signal x t , through the adversarial training of G st and D t , so that the data distribution of the converted signal generated by G st is highly consistent with the target domain signal x t , and finally the generator G ts converts into the signal in the source domain At this time, the generator G ts learns the mapping T→S from the target domain to the source domain, and the forward conversion of the source domain signal x s should make the forward cycle hold. C2. First, the generator G ts converts the target domain signal x t into a source domain converted signal At this time, the generator G ts learns the mapping from the target domain to the source domain T→S, and then utilizes the discriminator D s to distinguish the converted signal from the source domain real signal x s , through the adversarial training of G ts and D s , so that the data distribution of the converted signal generated by G ts is highly consistent with the source domain signal x s , and finally the generator G st converts the converted signal into a signal in the target domain At this time, the generator G st learns the mapping from the source domain to the target domain S→T, and the reverse conversion of the target domain signal x t should make the reverse cycle hold. C3. The conversion of signals from known conditions to unknown conditions is achieved through steps C1 and C2, which improve the similarity between the converted signals and the real signals while ensuring that the samples generated by the two generators do not contradict each other. In the training process, the target domain classifier C t is used to classify the target domain converted signals and the target domain real signals x t to further ensure that the target domain classifier C t can correctly classify the real signals under unknown conditions; After the conversion, the classifier of the Cycle-GAN model is trained by using the new converted signal; D. using the trained Cycle-GAN model in step C to diagnose the actual bearing vibration signal to obtain the result.

2. The Cycle-GAN-based aero-engine inter-shaft bearing fault diagnosis method under class imbalance condition according to claim 1, wherein, The step A obtains vibration signal samples of engine inter-shaft bearings in normal state and different fault states under different working conditions, and divides to obtain a training data set and a test data set, and specifically comprises the following steps: A1. Select normal state bearing signals under different working conditions as source domain data {x s ,y s}, randomly select fault state bearing signals under one working condition in the corresponding working condition as target domain data {x t ,y t}, and the above two types of data jointly constitute a training data set as a signal sample under known conditions; A2. selecting the same fault state bearing signal under the remaining working conditions in the corresponding working condition as the test data set, as the signal sample under unknown conditions.

3. The Cycle-GAN based aero-engine inter-turbine bearing fault diagnosis method under class imbalance condition according to claim 1, wherein, The step D utilizes the Cycle-GAN model trained in the step C to diagnose the actual bearing vibration signal to obtain a result, specifically: taking the test data set divided in the step A as the input of the target domain classifier C of the Cycle-GAN model after training, and realizing the fault diagnosis of the engine inter-shaft bearing under the condition of class imbalance. t The step D utilizes the Cycle-GAN model trained in the step C to diagnose the actual bearing vibration signal to obtain a result, specifically: taking the test data set divided in the step A as the input of the target domain classifier C of the Cycle-GAN model after training, and realizing the fault diagnosis of the engine inter-shaft bearing under the condition of class imbalance.

Images