Aero-engine fault diagnosis method based on variational autoencoder generative adversarial network

Abstract

This invention discloses an aero-engine fault diagnosis method based on a variational autoencoder generative adversarial network. The method first acquires the vibration signals of key aero-engine components. Next, a deep variational autoencoder generative adversarial network is constructed, where the encoder structure enhances the ability to learn the features of the fault signals, thereby improving data generation quality. A gradient penalty mechanism is applied to the discriminator structure to prevent mode collapse during network training. The constructed generative adversarial network is used to generate samples similar to the acquired signals to expand the imbalanced dataset. Finally, the deep features extracted by the model are input into a convolutional neural network to obtain higher fault diagnosis results. This method is highly effective, simple to implement, and provides rapid diagnosis, effectively solving aero-engine fault diagnosis problems.

Description

Technical Field

[0001] The present invention belongs to the technical field of aero-engines, and in particular relates to a method for diagnosing faults of aero-engines. Background Art

[0002] Aircraft engine equipment is widely used in modern industry, with increasing levels of intelligence and integration. Rolling bearings, as key components of aircraft engines, are subject to long-term operation in the harsh environment of high altitude, high speed, and heavy loads. They are inevitably subject to various forms of degradation and even failure, which can easily lead to significant economic losses or safety accidents. Therefore, accurate and efficient fault diagnosis of rolling bearings is of great significance.

[0003] Aircraft engines often operate under harsh conditions, placing high demands on materials, machining precision, and equipment during operation. Despite this, critical engine components are susceptible to cracks, wear, fractures, corrosion, and other failures. Given the demands for aircraft engine safety, reliability, and economic efficiency, aircraft engine fault diagnosis has become a hot topic for governments, military enterprises, research institutions, and academics worldwide. Extensive research and application efforts have yielded fruitful results. By monitoring aircraft engine health and fault diagnosis, monitoring performance trends, predicting the likelihood of failure and remaining service life of key systems and components, and implementing condition-based maintenance, we can mitigate performance degradation and failure, extend service life, and avoid unexpected incidents and safety accidents caused by failures. This can also significantly improve operational efficiency and the economic benefits of maintenance.

[0004] Fault diagnosis technology has evolved into an independent, integrated information processing discipline, a product of interdisciplinary interpenetration, cross-fertilization, and mutual promotion. The integration of fault diagnosis technology with cutting-edge science is the future direction of aircraft engine fault diagnosis. Current trends in fault diagnosis technology are the precision and multidimensionality of sensors, the diversification of diagnostic models and theories, and the intelligence of diagnostic technology. Traditional health monitoring and fault diagnosis methods primarily include health monitoring methods based on signal processing and health monitoring methods based on traditional machine learning. Fault diagnosis methods based on signal processing rely primarily on advanced modern signal processing techniques to extract sensitive features of key components from complex, nonlinear sensor signals to achieve fault detection. However, in addition to relying on advanced signal processing techniques, health monitoring methods based on signal processing also require expert experience and prior knowledge. Intelligent diagnosis methods based on traditional machine learning primarily extract and select sensitive time-domain, frequency-domain, or time-frequency-domain features, then input them into shallow artificial neural networks or support vector machines to achieve fault prediction and identification. Commonly used methods include fuzzy reasoning, expert systems, artificial neural networks, and support vector machines. Intelligent diagnosis methods based on traditional machine learning suffer from two major drawbacks. First, they rely heavily on manual feature extraction and selection, which leads to low efficiency and poor model generalization. Second, artificial neural networks and support vector machines are shallow models with poor feature extraction performance and are prone to overfitting. Third, they are unable to handle big data.

[0005] In recent years, with the successful application of deep learning in fields such as image recognition and speech recognition, it has demonstrated significant advantages over traditional methods in terms of feature extraction and accuracy. Domestic and international scholars have successively developed popular deep learning models for intelligent fault diagnosis, including restricted Boltzmann machines, deep belief networks, convolutional neural networks, autoencoders, and recurrent neural networks. Research results show that deep learning overcomes the shortcomings of shallow models, enabling automatic acquisition of internal data information and adaptive extraction of effective features, which helps reduce data preprocessing and feature extraction steps in scientific research. It eliminates the need for manual feature extraction and can directly process raw sensor data, improving learning efficiency while reducing the impact of human error and subjective factors on the results. Layer-by-layer feature learning can improve the accuracy of fault prediction and fault identification.

[0006] In existing inventions, most data-driven fault diagnosis methods are based on the assumption that large-scale balanced training samples can be obtained, which means that a similar amount of labeled data can be obtained to train the model under different fault conditions. However, this assumption is usually unrealistic in engineering. In actual engineering, the training data under different health states is often unbalanced. For aircraft engines, considering economy and safety, they operate in a healthy state most of the time, so the vibration samples collected to represent the normal state are sufficient, but the operating frequency of the machine in the fault state is very low, and the corresponding fault samples are far less than normal samples. Therefore, there is a significant imbalance between normal samples and fault samples, which will seriously affect the accuracy of fault diagnosis. The problem of fault diagnosis under data imbalance has always been the focus of scholars.

[0007] In a 2014 paper titled "Generative Adversarial Nets," Ian Goodfellow of the University of Montreal proposed a deep learning unsupervised model—the Generative Adversarial Network (GAN). This was a completely novel idea, and GANs have since become a hot topic in academia. Recent research has achieved promising results in addressing unbalanced fault diagnosis, improving the training and data generation performance of GANs. However, GANs are still immature in fault diagnosis, and their effectiveness in diagnosing one-dimensional fault signals with unbalanced fault data is suboptimal. Due to the lack of auxiliary information in the deep features of the input data, once the discriminator is fooled, the GAN's game theory strategy can lead the generator to output data that is sometimes irrational. Furthermore, due to their limited ability to learn the deep features of the original vibration signal, existing GAN-based models struggle to maintain good data generation performance for all fault types, reducing the accuracy and robustness of fault detection.

[0008] Therefore, introducing generative adversarial networks into the field of fault diagnosis will undoubtedly bring new impetus to the research on fault feature learning and fault pattern recognition under data imbalance, and solve the shortcomings of traditional intelligent methods. Summary of the Invention

[0009] To overcome the shortcomings of the existing technology, the present invention provides an aircraft engine fault diagnosis method based on a variational autoencoder generative adversarial network. First, the vibration signals of key aircraft engine components are acquired. Second, a deep variational autoencoder generative adversarial network is constructed, in which the encoder structure enhances the feature learning ability of the fault signal, thereby improving the data generation quality. At the same time, a gradient penalty mechanism is applied to the discriminator structure to avoid mode collapse during the network training process. The constructed generative adversarial network is used to generate samples similar to the acquired signals to expand the imbalanced data set. Finally, the deep features extracted by the model are input into a convolutional neural network to obtain higher fault diagnosis results. The present invention has excellent results, is simple and easy to implement, and has rapid diagnosis, and can effectively solve aircraft engine fault diagnosis.

[0010] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0011] Step 1: Collect vibration acceleration signals of aircraft engine components to obtain fault signals. Then, normalize and perform sliding window sampling on the fault signals to construct model training and test sets.

[0012] Step 2: Construct a variational autoencoder generative adversarial network model to generate data for fault signals;

[0013] Step 2-1: Add the encoder to the deep generative adversarial network so that the encoder and generator form an autoencoder;

[0014] Step 2-2: Learn the distribution characteristics of input data through encoding and reconstruction;

[0015] During the training process, the encoder randomly samples from the real dataset x and then encodes it into the latent code space through the neural network. In order to make the posterior distribution q(z|x) obtained by the encoder as close as possible to the expected distribution, the KL divergence is selected to calculate the distance between the two distributions as part of the loss function. The objective function of the encoder is shown in formula (1):

[0016] L E =-E q(z|x) [log p(x|z)]+D KL (q(z|x)||p(z)) (1)

[0017] In the formula, the first part on the right is the reconstruction error of the variational autoencoder, and the second part is the prior distribution regularization; p(z) represents the prior distribution of the required encoding space z; D KL (·) represents the KL divergence between two variables;

[0018] Step 2-3: The principle of the constructed variational autoencoder generative adversarial network model is as shown in formula (2):

[0019]

[0020] Where G represents the generator, D represents the discriminator, and P da and P g Represent the prior distribution of the original sample distribution x and the noise z respectively, D(x) represents the output of the discriminator, and G(z) represents the generated data output by the generator;

[0021] Step 3: The generator samples from the latent encoding space z provided by the encoder; by making the decoder in the variational autoencoder and the generator in the generative adversarial network share parameters and train them together, they are combined into a whole. The loss function of the discriminator is as follows:

[0022]

[0023] Among them, the last term on the right side of formula (3) represents the gradient penalty term, λ represents the gradient penalty coefficient; D(y) represents the discriminant value of the discriminator on the generated data, P r Represents the distribution of real data;

[0024] Step 4: Use a feature matching algorithm in the generator to match the deep features of the generated data and measure the L2 distance between the mean values of the feature vectors. The improved generator loss function is as follows:

[0025]

[0026] Where h(x) represents the feature vector extracted from the middle layer of the discriminator, and the output of the second convolutional layer of the discriminator is selected as the feature h(x); w represents the weight coefficient of feature matching, and h(y) represents the output of the generated sample in the second convolutional layer of the discriminator;

[0027] Step 5: The data generated by the generator is provided to the discriminator together with the original data. Based on the gradient penalty loss function, the parameters of the discriminator are updated using the stochastic gradient descent method. The formula for applying the gradient penalty to the objective function of the discriminator is as follows:

[0028]

[0029] Where δ represents a weak perturbation in the normal distribution N;

[0030] Step 6: After the discriminator is trained, the generator updates its parameters based on the discriminator’s feedback and trains to generate more generated samples;

[0031] Step 7: Repeat steps 3 to 6. At the end of the training process, the generator samples from the normal distribution N(0,I) to generate synthetic samples.

[0032] Step 8: Using the data generated in step 7 as training data and the real data obtained in step 1 as test data, a convolutional neural network classifier is constructed for training and testing to diagnose the fault type of the bearing data.

[0033] Preferably, the p(z) is a standard normal distribution.

[0034] Preferably, the convolutional neural network classifier includes a convolution layer, a pooling layer and a fully connected layer;

[0035] The convolution layer extracts features from the input. It contains multiple feature maps, each of which consists of multiple neurons. To calculate a new feature map, the input feature map is convolved with a learnable convolution kernel, as shown below:

[0036]

[0037] represents the i-th feature map located in the n-th hidden layer, is the convolution kernel connecting the j-th input feature map and the i-th feature map, is the residual, · represents the convolution operation, and f(·) is the activation function;

[0038] The pooling layer is used to reduce the dimension of the feature map and achieve data compression. The output is represented as:

[0039]

[0040] σ(·) refers to the activation function, is the weight coefficient, p(·) is the pooling operation;

[0041] The function of the fully connected layer is to convert the three-dimensional feature map into a one-dimensional structure. The output is calculated by the following formula:

[0042]

[0043] θ(·) is the activation function, is the weight coefficient.

[0044] The beneficial effects of the present invention are as follows:

[0045] This paper proposes a variational autoencoder generative adversarial network (GAN) to better generate fault data, thereby improving the performance of imbalance fault diagnosis. This method combines the advantages of two deep generative models, the variational autoencoder and the GAN, leveraging the variational autoencoder's ability to learn the underlying features of the raw data while overcoming the unstable generation process and low data quality caused by random noise sampling. Simulation results demonstrate that the proposed method achieves higher accuracy and greater robustness than other methods in diagnosing imbalance faults in key aircraft engine components, and is therefore suitable for aircraft engine condition monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the rolling bearing fault diagnosis method of the present invention.

[0047] Figure 2 A time domain diagram of a rolling bearing vibration signal used and generated by an embodiment of the present invention.

[0048] Figure 3 A visualization of features extracted by the variational autoencoder generative adversarial network used in an embodiment of the present invention.

[0049] Figure 4 This is a diagram of the experimental diagnosis results of an embodiment of the present invention.

[0050] Figure 5 A comparison chart of the method of the present invention and other methods. DETAILED DESCRIPTION

[0051] The present invention will be further described below with reference to the accompanying drawings and examples.

[0052] This paper explores a new fault diagnosis method that can rapidly analyze nonlinear, nonstationary, and highly coupled aeroengine component data to promptly diagnose aeroengine faults. By augmenting unbalanced datasets with a generative adversarial network and then performing intelligent fault diagnosis based on balanced datasets, the method can rapidly and accurately complete aeroengine fault diagnosis tasks.

[0053] The present invention steps are as follows:

[0054] 1) By adding an encoder structure to a deep generative adversarial network (GAN), we construct a GAN consisting of an encoder, generator, and discriminator. The encoder extracts features from the original signal to enhance feature learning, thereby improving the quality of the generated data. By applying a gradient penalty to the discriminator network's objective function, we prevent mode collapse during network training. Ultimately, we establish an efficient and stable GAN fault diagnosis model.

[0055] 2) A feature matching algorithm is used in the generator to match the deep features of the generated data, speed up the model training process, and input the deep features extracted by the model into the convolutional neural network to obtain higher fault diagnosis results.

[0056] By adding an encoder structure to a deep generative adversarial network (GAN), we construct a GAN consisting of an encoder, decoder, and discriminator. The encoder extracts features from the original signal to enhance feature learning, thereby improving the quality of generated data. By applying a gradient penalty to the discriminator network's objective function, we prevent mode collapse during network training. Ultimately, we establish an efficient and stable GAN fault diagnosis model.

[0057] Step 1: Collect vibration acceleration signals of key aircraft engine components to obtain fault signals. Then, normalize and perform sliding window sampling on the fault signals to construct model training and test sets.

[0058] Step 2: The original vibration dataset lacks fault signals, resulting in data imbalance. Therefore, a generative adversarial network model is constructed to generate data for a small amount of fault signals.

[0059] Step 2-1: Add an encoder structure to the deep generative adversarial network to enhance the ability to learn features of real data, thereby improving the quality of data generation; the proposed method includes an encoder, a generator, and a discriminator, where the encoder and generator form an autoencoder structure;

[0060] Step 2-2: Learn the distribution characteristics of input data through encoding and reconstruction;

[0061] During the training process, the encoder randomly samples from the real dataset x and then encodes it into the latent coding space through the neural network. The latent vector contains the deep feature information of the input data. In order to make the posterior distribution q(z|x) obtained by the encoder as close as possible to the expected distribution, the KL divergence is selected to calculate the distance between the two distributions as part of the loss function. The objective function of the encoder is shown in formula (1):

[0062] L E =-E q(z|x) [log p(x|z)]+D KL (q(z|x)||p(z)) (1)

[0063] Where, the first part on the right is the reconstruction error of the variational autoencoder, and the second part is the prior distribution regularization. p(z) represents the prior distribution of the required encoding space z, and the standard normal distribution is usually chosen;

[0064] Step 2-3: The principle of the constructed variational autoencoder generative adversarial network model is as shown in formula (2):

[0065]

[0066] Among them, P da and P g Denote the prior distributions of the original sample distribution x and the noise z, respectively, and D(x) represents the output of the discriminator. The training goal of the generator is to minimize the second term of the loss function, while the training goal of the discriminator is to maximize the loss function, that is, to maximize the log-likelihood by optimizing between the generated samples and the real sample source. Through this adversarial training, the performance of the generator and discriminator continuously improves, ultimately aligning the distributions of fake samples and real data. Data augmentation addresses the imbalance of the original dataset.

[0067] Step 3: The generator samples from the latent encoding space z provided by the encoder; due to the pre-training of the variational autoencoder, the encoding space already contains deep feature information about the distribution of the input data, which is more conducive to generating high-quality fake data with the same distribution. By making the decoder in the variational autoencoder and the generator in the generative adversarial network share parameters and train them together, they are combined into a whole. The loss function of the discriminator is as follows:

[0068]

[0069] Among them, the last term on the right side of formula (3) represents the gradient penalty term, λ represents the gradient penalty coefficient, D(y) represents the discriminant value of the discriminator on the generated data, P r Represents the distribution of real data.

[0070] Step 4: A feature matching algorithm is used in the generator to match the deep features of the generated data to eliminate overfitting in network training. Deep feature matching reduces the statistical distance between real samples and generated data, and measures the L2 distance between the mean of feature vectors. The improved generator loss function is as follows:

[0071]

[0072] Where h(x) represents the feature vector extracted from the middle layer of the discriminator, the output of the second convolutional layer of the discriminator is selected as the feature h(x), w represents the weight coefficient of feature matching, and h(y) represents the output of the generated sample in the second convolutional layer of the discriminator.

[0073] Step 5: Provide the generated data together with the original data to the discriminator. Based on the gradient penalty loss function, the stochastic gradient descent method is used to update the parameters of the discriminator. By applying the gradient penalty formula to the objective function of the discriminator as shown in formula (5), this method can avoid the mode collapse phenomenon in network training.

[0074]

[0075] Step 6: After the discriminator is trained, the generator updates its parameters based on the discriminator’s feedback and trains to produce more generated samples that are similar to real samples.

[0076] Step 7: Repeat steps 3-6; at the end of the training process, the generator samples from the normal distribution N(0,I) to generate high-quality synthetic samples;

[0077] Step 8: Using the data generated in step 7 as training data and the real data obtained in step 1 as test data, a convolutional neural network classifier is constructed for training and testing to diagnose the fault type of the bearing data.

[0078] The convolutional neural network classifier consists of convolutional layers, pooling layers, and fully connected layers.

[0079] The convolution layer extracts features from the input. It contains multiple feature maps, each of which consists of multiple neurons. To calculate a new feature map, the input feature map is convolved with a learnable convolution kernel, as shown below:

[0080]

[0081] represents the i-th feature map located in the n-th hidden layer, is the convolution kernel connecting the j-th input feature map and the i-th feature map, is the residual, · represents the convolution operation, and f(·) is the activation function.

[0082] The pooling layer is used to reduce the dimension of the feature map and achieve data compression. The output is represented as:

[0083]

[0084] σ(·) refers to the activation function, is the weight coefficient, and p(·) is the pooling operation.

[0085] The function of the fully connected layer is to convert the three-dimensional feature map into a one-dimensional structure. The output is calculated by the following formula:

[0086]

[0087] θ(·) is the activation function, is the weight coefficient. Specific embodiment:

[0089] The present invention will be described in further detail below with reference to the accompanying drawings: Figure 1As shown, the present invention can be divided into two main parts. The first part is to construct a variational autoencoder generative adversarial network and then train the network model using unbalanced vibration data. The second part uses the trained network to generate a small amount of fault data and obtain diagnostic results based on the balanced data set through a convolutional neural network.

[0090] Reference Figure 2 As shown in the time domain comparison diagram of vibration signals of six types of bearing faults randomly selected, the horizontal axis represents time; the vertical axis represents vibration amplitude, the unit is m / s 2 .

[0091] Reference Figure 3 Figure 2 shows the features extracted by a deep convolutional neural network visualized using Tsne. Ten colors represent ten different bearing fault types. Triangles of the same color represent real samples, while squares represent generated samples. The figure shows that the characteristics of each fault can be accurately identified, and the generated data for similar faults closely resembles the original data, demonstrating the powerful data generation capabilities of the variational autoencoder generative adversarial network.

[0092] Reference Figure 4 As shown in the figure, the accuracy of various types of rolling bearing faults can be directly obtained from the figure.

[0093] Reference Figure 5 As shown in the figure, the bar chart compares the ten test results of the method proposed in the present invention with the ten test results of three comparison methods: support vector machine, stacked autoencoder, and convolutional neural network. It can be seen from the figure that the ten diagnostic results of the proposed method are significantly higher than the diagnostic results of other comparison methods, which further verifies the advanced nature of the method of the present invention.

[0094] Example 1: This example uses a dataset composed of vibration data from electric locomotive bearing failures. This dataset more closely resembles the actual operating conditions of bearings in engineering applications and features more complex failure modes. The dataset includes five fault conditions: inner race fault, outer race fault, rolling element fault, and mixed inner and outer race faults. These include three single faults and one combined fault. The sampling frequency is 12.8 kHz.

[0095] The invented method is used to diagnose vibration data failure of electric locomotive bearing fault. In order to avoid the randomness of the results, ten consecutive diagnostic results are selected to calculate their average value and standard deviation. That is, the higher the diagnostic average value, the better the diagnostic ability, and the lower the standard deviation, the stronger the diagnostic stability. Table 1 shows the comparison of the diagnostic results of the method of the present invention and three comparative methods: support vector machine, stacked autoencoder, and convolutional neural network. It can be seen from Table 1 that the fault diagnosis accuracy of the method of the present invention is higher and the robustness is better. Therefore, the method of the present invention can effectively complete the fault diagnosis task of key components of aircraft engines.

[0096] Table 1 Comparison of electric locomotive bearing diagnosis results

[0097]

[0098] Example 2: This example uses bearing failure data from the Case Western Reserve University laboratory for experimental analysis. Vibration data is collected using a high-precision accelerometer with a sampling frequency of 12,000 Hz and 48,000 samples per second for drive-end bearing failures. Fault conditions for the drive-end bearing under zero load include roller, inner race, and outer race failures. Damage diameters are 0.007 inches, 0.014 inches, and 0.021 inches, respectively. Data from nine faulty and normal states are used as training and test data.

[0099] The invented method was used to diagnose rolling bearing failures at Case Western Reserve University. To avoid the possibility of randomness in the results, ten consecutive diagnostic results were selected and their average and standard deviation were calculated. That is, the higher the diagnostic average, the better the diagnostic ability, and the lower the standard deviation, the stronger the diagnostic stability. The diagnostic results are as follows: Figure 4 Table 2 compares the diagnostic results of the method of the present invention with those of three comparative methods: support vector machine, stacked autoencoder, and convolutional neural network. As can be seen from Table 2, the fault diagnosis accuracy of the method of the present invention is higher and the robustness is better. Therefore, the method of the present invention can effectively complete the task of rolling bearing status detection and fault diagnosis.

[0100] Table 2 Comparison of rolling bearing diagnosis results

[0101]

Claims

1. An aero-engine fault diagnosis method based on variational autoencoder generative adversarial network, characterized in that: The steps include: Step 1: Collect vibration acceleration signals of aircraft engine components to obtain fault signals. Then, normalize and perform sliding window sampling on the fault signals to construct model training and test sets. Step 2: Construct a variational autoencoder generative adversarial network model to generate data for fault signals; Step 2-1: Add the encoder to the deep generative adversarial network so that the encoder and generator form an autoencoder; Step 2-2: Learn the distribution characteristics of input data through encoding and reconstruction; During the training process, the encoder randomly samples from the real dataset x and then encodes it into the latent code space through the neural network. In order to make the posterior distribution q(z|x) obtained by the encoder as close as possible to the expected distribution, the KL divergence is selected to calculate the distance between the two distributions as part of the loss function. The objective function of the encoder is shown in formula (1): L E =-E q(z|x) [log p(x|z)]+D KL (q(z|x)||p(z)) (1) In the formula, the first part on the right is the reconstruction error of the variational autoencoder, and the second part is the prior distribution regularization; p(z) represents the prior distribution of the required encoding space z; D KL (·) represents the KL divergence between two variables; Step 2-3: The principle of the constructed variational autoencoder generative adversarial network model is as shown in formula (2): Where G represents the generator, D represents the discriminator, and P da and P g Represent the prior distribution of the original sample distribution x and the noise z respectively, D(x) represents the output of the discriminator, and G(z) represents the generated data output by the generator; Step 3: The generator samples from the latent code space z provided by the encoder; the decoder in the variational autoencoder and the generator in the generative adversarial network are combined into a whole by making them share parameters and train together; the loss function of the discriminator is as follows: Among them, the last term on the right side of formula (3) represents the gradient penalty term, λ represents the gradient penalty coefficient; D(y) represents the discriminant value of the discriminator on the generated data, P r Represents the distribution of real data; Step 4: Use a feature matching algorithm in the generator to match the deep features of the generated data and measure the L2 distance between the mean values of the feature vectors. The improved generator loss function is as follows: Where h(x) represents the feature vector extracted from the middle layer of the discriminator, and the output of the second convolutional layer of the discriminator is selected as the feature h(x); w represents the weight coefficient of feature matching, and h(y) represents the output of the generated sample in the second convolutional layer of the discriminator; Step 5: The data generated by the generator is provided to the discriminator together with the original data. Based on the gradient penalty loss function, the parameters of the discriminator are updated using the stochastic gradient descent method. The formula for applying the gradient penalty to the objective function of the discriminator is as follows: Where δ represents a weak perturbation in the normal distribution N; Step 6: After the discriminator is trained, the generator updates its parameters based on the discriminator’s feedback and trains to generate more generated samples. Step 7: Repeat steps 3 to 6. At the end of the training process, the generator samples from the normal distribution N(0,I) to generate synthetic samples. Step 8: Using the data generated in step 7 as training data and the real data obtained in step 1 as test data, a convolutional neural network classifier is constructed for training and testing to diagnose the fault type of the bearing data.

2. The method for aircraft engine fault diagnosis based on variational autoencoder generative adversarial network according to claim 1, characterized in that: The p(z) is a standard normal distribution.

3. The method for aircraft engine fault diagnosis based on variational autoencoder generative adversarial network according to claim 1, characterized in that: The convolutional neural network classifier includes a convolution layer, a pooling layer and a fully connected layer; The convolution layer extracts features from the input. It contains multiple feature maps, each of which consists of multiple neurons. To calculate a new feature map, the input feature map is convolved with a learnable convolution kernel, as shown below: represents the i-th feature map located in the n-th hidden layer, is the convolution kernel connecting the j-th input feature map and the i-th feature map, is the residual, · represents the convolution operation, and f(·) is the activation function; The pooling layer is used to reduce the dimension of the feature map and achieve data compression. The output is represented as: σ(·) refers to the activation function, is the weight coefficient, p(·) is the pooling operation; The function of the fully connected layer is to convert the three-dimensional feature map into a one-dimensional structure. The output is calculated by the following formula: θ(·) is the activation function, is the weight coefficient.

Images