CN114169396A - Construction method and application of training data generation model for aircraft fault diagnosis - Google Patents

Abstract

The invention discloses a training data generation model construction method for aircraft fault diagnosis and application, belonging to the technical field of aircraft fault diagnosis, and comprising S1, constructing a training data generation model comprising a generator and a discriminator; s2, inputting a pre-collected aircraft sample data set into a training data generation model for training, and enabling a generator and a discriminator to play games with each other until Nash balance is achieved; the generator is based on a VAE model, input data are reconstructed by carrying out maximum likelihood estimation on data distribution, data similar to original data are generated, and meanwhile, the authenticity of the generated data is judged through a discriminator, so that the reliability and the authenticity of the data are improved; by adopting the training data generation model provided by the invention, real and accurate sample generation can be realized, so that the technical problems of data shortage and class unbalance are solved under the condition of ensuring the accuracy of fault diagnosis.

Description

Construction method and application of training data generation model for aircraft fault diagnosis

technical field

The invention belongs to the technical field of aircraft fault diagnosis, and more particularly, relates to a training data generation model construction method and application for aircraft fault diagnosis.

Background technique

Once the aircraft fails during flight, it will cause serious economic losses and casualties. Fault diagnosis technology is very important in order to be able to provide a plan for the fault to be controlled so that the fault can be controlled.

The fault diagnosis technology is a state identification technology that uses the current state information and historical conditions of the equipment to evaluate the equipment state through a certain analysis method. Fault diagnosis is often divided into two categories: model-based methods and data-based methods; among them, the model-based method establishes an accurate mathematical model to diagnose faults according to the abnormal state quantities of the aircraft, but in the actual application process, There are many model uncertainties and external disturbances that cannot be modeled and are therefore difficult to apply. In the data-based method, the diagnosis of aircraft faults is mostly based on deep learning models. With the rapid development of deep learning and its wide migration in various fields, structures such as CNN, LSTM, and Transformer have long been used in the field of fault diagnosis. . But deep learning-based methods suffer from two fatal flaws, namely lack of data and class imbalance. On the one hand, the lack of data is a common problem for all deep learning tasks. On the other hand, in industrial systems represented by aircraft, failures are avoided. Therefore, the failure data is very small compared to normal data, and The data on the failure of a specific component is even less, which seriously affects the diagnosis effect of the deep learning diagnosis network. It is well known that the choice of dataset has an impact on the results far beyond the network structure itself.

In order to solve the above defects, the aircraft data is simulated in the existing research, but there is still a difference in the data distribution between the results of the computer simulation and the actual data obtained, and there are risks in the application of the aircraft. Therefore, for the problems of lack of data and class imbalance, the data is often not simulated directly, but the data is expanded based on the actual collection of a small number of sample data. It is called the SMOTE method. However, this method will make the data redundant and easily cause the problem of model overfitting, which is not conducive to the training of the subsequent fault diagnosis model and affects the accuracy of the subsequent fault diagnosis. Another idea is to group the data, that is, divide the unbalanced training data set into multiple balanced data sets according to certain rules, and integrate the multiple classifiers trained on the balanced data set according to a certain learning method Together, in order to eliminate the class imbalance problem, but such methods will lead to new imbalanced data grouping problems in the process of data grouping, fail to fundamentally solve the problem, and also affect the accuracy of subsequent fault diagnosis.

SUMMARY OF THE INVENTION

Aiming at the above defects or improvement needs of the prior art, the present invention provides a training data generation model construction method and application for aircraft fault diagnosis, so as to solve the problem that the prior art cannot guarantee the accuracy of fault diagnosis. Technical issues with lack of data and class imbalance.

In order to achieve the above purpose, the present invention provides a method for constructing a training data generation model for aircraft fault diagnosis, comprising the following steps:

S1. Build a training data generation model;

The training data generation model includes: a generator and a discriminator; wherein, the generator is a generator based on the VAE model, and is used to sequentially encode and decode the input aircraft sample data to reconstruct it and obtain the same as the aircraft sample data. Similar dummy data; the discriminator is used to determine whether the input data is real aircraft data;

S2. Input the pre-collected aircraft sample data set into the training data generation model for training, so that the generator and the discriminator play games with each other until a Nash equilibrium is reached; the aircraft sample data set includes aircraft fault sample data.

Further preferably, the above generator includes: a cascaded first MLP-Mixer, a VAE model and a second MLP-Mixer;

The first MLP-Mixer is used to expand the dimension of the input aircraft sample data, and integrate the features of different channels of the expanded aircraft sample data, thereby mapping the aircraft sample data into depth data features;

The VAE model is used to encode the depth data features through the encoder to obtain the corresponding mean and variance, then resample in the normal distribution corresponding to the obtained mean and variance to obtain hidden variables, and pass the decoder to the hidden variables. Perform reconstruction to obtain pseudo data features similar to depth data features;

The second MLP-Mixer is used to map the pseudo data features according to different channels, so as to inversely map the pseudo data features into pseudo data similar to the aircraft sample data.

Further preferably, step S2 includes:

S21. For each sample data in the pre-collected aircraft sample data set, after generating the pseudo data corresponding to the sample data through the generator, input the sample data and the corresponding pseudo data into the discriminator, respectively, to obtain the sample data and the corresponding pseudo data. The pseudo data of the aircraft are respectively judged as the predicted probability of the real data of the aircraft;

S22. Obtain the loss value of the generator by calculating the sum of the expression difference and distribution difference between the sample data and the corresponding pseudo data; by calculating the sample data and the corresponding pseudo data, the predicted probability and corresponding The sum of the differences between the true probabilities of , and the loss value of the discriminator is obtained;

S23. Determine whether the sum of the loss value of the generator and the loss value of the discriminator reaches the minimum or whether the number of iterations reaches the preset number of iterations. If so, the training of the training data generation model is completed, and the operation ends; otherwise, the generator and the discriminator are updated. parameters in , go to step S21.

Further preferably, the loss value of the generator is:

L _G =MSE+KLD

为第i个样本数据所对应的伪数据；

为第i个样本数据；KLD为样本数据分布和对应的伪数据分布的KL散度；μ_i和σ_i分别为编码器输出的第i个样本数据所对应的均值和方差。Among them, MSE is the expression difference between the sample data and the corresponding pseudo data; N is the number of sample data in the aircraft sample data set;

is the pseudo data corresponding to the i-th sample data;

is the i-th sample data; KLD is the KL divergence of the sample data distribution and the corresponding pseudo-data distribution; μ _i and σ _i are the mean and variance corresponding to the i-th sample data output by the encoder, respectively.

Further preferably, the loss value of the discriminator is:

为样本数据被判别为飞行器真实数据的预测概率；

为样本数据为飞行器真实数据的真实概率；

为样本数据所对应的伪数据被判别为飞行器真实数据的预测概率。in,

is the predicted probability that the sample data is judged to be the real data of the aircraft;

is the real probability that the sample data is the real data of the aircraft;

The pseudo data corresponding to the sample data is judged as the predicted probability of the real data of the aircraft.

In a second aspect, the present invention provides a training data generation method for aircraft fault diagnosis, including:

By controlling the expansion times of the sample data in the pre-collected aircraft sample data set, the aircraft sample data set is expanded, so that the number of sample data in the expanded aircraft sample data set reaches the preset number, and the aircraft fault sample data and aircraft normal samples The amount of data is balanced; the obtained expanded aircraft sample data set is the training data set used for aircraft fault diagnosis;

Wherein, the expanded sample data is aircraft fault sample data and/or aircraft normal sample data;

Augmentation methods for sample data include:

Input the sample data into the generator of the training data generation model constructed by using the training data generation model construction method provided by the first aspect of the present invention, and obtain pseudo data corresponding to the sample data, so as to expand the sample data; The fault information label of is the same as the fault information label of its corresponding sample data.

In a third aspect, the present invention provides a method for constructing an aircraft fault diagnosis model, including:

The training data set generated by the training data generation method provided by the second aspect of the present invention is input into the machine learning model for training, and an aircraft fault diagnosis model is obtained.

Further preferably, the above-mentioned machine learning model is a Transformer model.

In a fourth aspect, the present invention provides a method for diagnosing aircraft faults, comprising: inputting flight sampler data into an aircraft fault diagnosis model constructed by using the method for constructing an aircraft fault diagnosis model provided in the third aspect of the present disclosure, and obtaining Aircraft failure information.

In a fifth aspect, the present invention also provides a machine-readable storage medium, where the machine-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are called and executed by a processor, the machine can Executing the instructions causes the processor to implement the method for constructing a training data generation model provided in the first aspect of the present invention, the method for generating training data provided in the second aspect, the method for constructing an aircraft fault diagnosis model provided in the third aspect, and the third aspect of the present invention. One or more of the aircraft fault diagnosis methods provided in the four aspects.

In general, through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

1. The present invention provides a method for constructing a training data generation model for aircraft fault diagnosis, and the constructed training data generation model includes a generator and a discriminator; wherein, the generator is a generator based on the VAE model, and the The maximum likelihood estimation of the distribution is performed to realize the reconstruction of the input data, generate data similar to the original data, and at the same time judge the authenticity of the generated data through the discriminator; through the mutual game between the generator and the discriminator, thereby increasing the data The training data generation model provided by the present invention can realize real and accurate sample generation, so as to solve the technical problems of lack of data and unbalanced categories under the condition of ensuring the accuracy of fault diagnosis.

2. In the training data generation model construction method provided by the present invention, the MLP-Mixer is introduced into the generator, and when training within the network, the information in different channels in the spatial domain is fused to play the role of feature enhancement, the present invention By expanding the dimensions of the input aircraft sample data and integrating the features of different channels of the expanded aircraft sample data, the aircraft sample data is mapped to depth data features, which greatly improves the feature extraction ability and improves the accuracy of the model.

Description of drawings

1 is a flowchart of a method for constructing a training data generation model for aircraft fault diagnosis provided by Embodiment 1 of the present invention;

2 is a schematic structural diagram of a training data generation model provided in Embodiment 1 of the present invention;

3 is a schematic structural diagram of a generator provided in Embodiment 1 of the present invention;

4 is a schematic structural diagram of a VAE model provided in Embodiment 1 of the present invention;

Fig. 5 is provided by the comparative experiment 1 of the present invention respectively using the aircraft sample data set that is actually collected and the data set expanded by the training data generation model provided by the present invention to train the same aircraft fault diagnosis model, after the same aircraft fault diagnosis model is performed The experimental results obtained by comparing the four faulty actuators with the real fault diagnosis values after fault diagnosis;

Fig. 6 adopts the model that adds MLP-Mixer and the model that does not add MLP-Mixer to provide for comparative experiment 2 of the present invention respectively after the same aircraft fault diagnosis model is trained, after carrying out the fault diagnosis of four fault actuators of the aircraft, Experimental results compared with real fault diagnosis values.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

Embodiment 1,

A method for building a training data generation model for aircraft fault diagnosis, as shown in Figure 1, includes the following steps:

S1. Build a training data generation model;

Among them, as shown in Figure 2, the training data generation model includes: a generator and a discriminator; wherein, the generator is a generator based on the VAE model, and is used to encode and decode the input aircraft sample data in turn, so as to Carry out reconstruction, obtain the pseudo data similar to the aircraft sample data; The discriminator is used to distinguish whether the input data is real aircraft data; In the present embodiment, the training data generation model is recorded as the Adversarial Variational AutoEncoding Generator model (AdversarialVariational AutoEncoding Generator , AVAEG).

Preferably, in an optional implementation manner, as shown in FIG. 3 , the above generator includes: a cascaded first MLP-Mixer, a VAE model and a second MLP-Mixer;

The first MLP-Mixer is used to expand the dimension of the input aircraft sample data, and integrate the features of different channels of the expanded aircraft sample data, thereby mapping the aircraft sample data into depth data features; it should be noted that this embodiment The sample data of the aircraft used in this paper includes the attitude data and fault data of the aircraft. The dimensions of these data are small, but the amount of information contained is very large. When judging the failure of the aircraft by comprehensively considering the data of different state quantities, the above data needs to be processed to extract deeper feature information. The present invention introduces MLP-Mixer into the generator, which can enable the generator to extract different features to enhance feature extraction capabilities.

The VAE model is used to encode the depth data features through the encoder to obtain the corresponding mean and variance, then resample in the normal distribution corresponding to the obtained mean and variance to obtain hidden variables, and pass the decoder to the hidden variables. Perform reconstruction to obtain pseudo data features similar to depth data features;

可以被平分为两个部分，即μ_i,σ_i∈R^h，h表示隐变量维度；μ_i和σ_i分别为编码器输出的第i个样本数据所对应的均值和方差。Specifically, the VAE model adopted in this embodiment is shown in FIG. 4 , wherein the encoder includes a convolution layer; the output of the encoder corresponding to the i-th sample data is

It can be divided into two parts, namely μ _i , σ _i ∈ R ^h , h represents the hidden variable dimension; μ _i and σ _i are the mean and variance corresponding to the ith sample data output by the encoder, respectively.

The output of the encoder is resampled to obtain a hidden variable _{zi ,} which is a composite function of the output _zi of the encoder and a random variable ε that satisfies the Gaussian distribution, namely:

ε_i～N(0,1)，f_h(·)为隐变量函数。Among them, _zi ∈ R ^h ,

ε _i ～N(0,1), f _h (·) is a hidden variable function.

In this embodiment, the decoder includes a deconvolution layer; the latent variable _zi is input into the decoder to obtain pseudo data features, and the dimensions of the pseudo data features are the same as the depth data features of the input encoder;

The second MLP-Mixer is used to map the pseudo data features according to different channels, so as to inversely map the pseudo data features into pseudo data similar to the aircraft sample data.

S2. Input the pre-collected aircraft sample data set into the training data generation model for training, so that the generator and the discriminator play games with each other until a Nash equilibrium is reached; the aircraft sample data set includes aircraft fault sample data.

同时也采集到了每个样本数据的故障信息，具体包括飞行器的四个执行机构的故障情况所构成的故障信息向量c^r。记飞行器样本数据集为

其中，

N为数据集数据容量，feature_num为状态信息向量的元素个数，label_num为故障信息向量的元素个数。Specifically, in this embodiment, a small data set including aircraft fault sample data is pre-collected in a real aircraft scenario, which is recorded as the aircraft sample data set; each sample data includes aircraft state information, specifically including the aircraft state quaternion and the state information vector composed of the attitude angle

At the same time, the fault information of each sample data is also collected, specifically including the fault information vector ^r formed by the fault conditions of the four actuators of the aircraft. The aircraft sample data set is recorded as

in,

N is the data capacity of the dataset, feature_num is the number of elements of the state information vector, and label_num is the number of elements of the fault information vector.

经过生成器得到对应的伪数据

接着，将样本数据

与对应的伪数据

分别送入判别器，令判别器鉴别真伪。训练过程是生成器与判别器的博弈，当两者理论上达到纳什均衡时停止训练。During the training phase, the sample data in the aircraft sample dataset

Obtain the corresponding pseudo data through the generator

Next, the sample data

with the corresponding dummy data

They are respectively sent to the discriminator, so that the discriminator can identify the authenticity. The training process is a game between the generator and the discriminator, and the training stops when the two theoretically reach the Nash equilibrium.

Specifically, a schematic diagram of the training phase of AVAEG is shown in FIG. 3 . In an optional implementation manner, step S2 includes:

S21. For each sample data in the pre-collected aircraft sample data set, after generating the pseudo data corresponding to the sample data through the generator, input the sample data and the corresponding pseudo data into the discriminator, respectively, to obtain the sample data and the corresponding pseudo data. The pseudo data of the aircraft are respectively judged as the predicted probability of the real data of the aircraft;

S22. Obtain the loss value of the generator by calculating the sum of the expression difference and distribution difference between the sample data and the corresponding pseudo data; by calculating the sample data and the corresponding pseudo data, the predicted probability and corresponding The sum of the differences between the true probabilities of , and the loss value of the discriminator is obtained;

Specifically, the loss value of the generator is:

L _G =MSE+KLD

为第i个样本数据所对应的伪数据；

为第i个样本数据；KLD为样本数据分布和对应的伪数据分布的KL散度；μ_i和σ_i分别为编码器输出的第i个样本数据所对应的均值和方差。Among them, MSE is the expression difference between the sample data and the corresponding pseudo data; N is the number of sample data in the aircraft sample data set;

is the pseudo data corresponding to the i-th sample data;

is the i-th sample data; KLD is the KL divergence of the sample data distribution and the corresponding pseudo-data distribution; μ _i and σ _i are the mean and variance corresponding to the i-th sample data output by the encoder, respectively.

Specifically, the loss value of the discriminator is:

为样本数据被判别为飞行器真实数据的预测概率；

为样本数据为飞行器真实数据的真实概率；

为样本数据所对应的伪数据被判别为飞行器真实数据的预测概率。in,

is the predicted probability that the sample data is judged to be the real data of the aircraft;

is the real probability that the sample data is the real data of the aircraft;

The pseudo data corresponding to the sample data is judged as the predicted probability of the real data of the aircraft.

S23. Determine whether the sum of the loss value of the generator and the loss value of the discriminator reaches the minimum or whether the number of iterations reaches a preset number of iterations (the preset number of iterations Epoch generally takes a value of 100), if so, the training data generation model training is completed , the operation ends; otherwise, update the parameters in the generator and the discriminator, and go to step S21.

Specifically, the loss function of the training data generation model is: L=L _G +L _D ; the parameters in the generator and the loss value are reversely updated according to the loss function value of the data generation model to train the training data generation model.

Embodiment 2,

A training data generation method for aircraft fault diagnosis, comprising:

By controlling the expansion times of the sample data in the pre-collected aircraft sample data set, the aircraft sample data set is expanded, so that the number of sample data in the expanded aircraft sample data set reaches the preset number, and the aircraft fault sample data and aircraft normal samples The amount of data is balanced; the obtained expanded aircraft sample data set is the training data set used for aircraft fault diagnosis;

Wherein, the expanded sample data is aircraft fault sample data and/or aircraft normal sample data;

Augmentation methods for sample data include:

The sample data is input into the generator of the training data generation model constructed by the training data generation model construction method provided in Embodiment 1, and the pseudo data corresponding to the sample data is obtained, so as to expand the sample data; the fault of the pseudo data The information label is the same as the failure information label of the corresponding sample data.

Specifically, take a scenario in which the collected aircraft sample data set contains less sample data and the number of fault sample data and normal sample data is unbalanced as an example;

中的各样本数据分别到生成器中，将每一个样本数据喂入生成器M(本实施例中M取值为100)次，得到M个伪数据，每一个生成的伪数据直接根据原样本数据分配故障信息向量(故障信息标签)。飞行器样本数据集

通过这种方法得到的生成数据集

的数据量被扩充了100倍，合并飞行器样本数据集

和生成数据集

得到新的数据集

从而有效地解决了数据集太小的问题。First, the aircraft sample dataset

The sample data in the generator are respectively sent to the generator, and each sample data is fed into the generator M (the value of M in this embodiment is 100) times to obtain M pseudo data, and each generated pseudo data is directly based on the original sample. Data assigns fault information vectors (fault information labels). Aircraft sample dataset

Generated datasets obtained by this method

The amount of data has been expanded by a factor of 100, combining aircraft sample data sets

and generate dataset

get new dataset

This effectively solves the problem that the dataset is too small.

Further, in order to solve the problem of class imbalance, the number of times the aircraft fault sample data is fed into the generator can be controlled according to the quantity ratio Q of the aircraft normal sample data and the aircraft fault sample data in the new data set, so that the aircraft normal sample data and the aircraft fault can be matched. The number of sample data is balanced; in this embodiment, each fault sample is fed to the generator Q times to achieve class balance.

The related technical solutions are the same as those in Embodiment 1, and are not repeated here.

Embodiment 3,

A method for constructing an aircraft fault diagnosis model, comprising:

The training data set generated by using the training data generation method provided in Embodiment 2 is input into a machine learning model (deep learning model) for training, and an aircraft fault diagnosis model is obtained.

Specifically, the loss function of the aircraft fault diagnosis model is:

为飞行器故障诊断模型输出的第k个训练数据的飞行器故障信息的预测值；c_k为第k个训练数据的飞行器故障信息的真实值。Among them, K is the number of training data in the training data set;

is the predicted value of the aircraft fault information of the kth training data output by the aircraft fault diagnosis model; _ck is the actual value of the aircraft fault information of the kth training data.

Based on the loss function value of the aircraft fault diagnosis model, the parameters in the aircraft fault diagnosis model are reversely updated, and the above process is repeated until the network converges, thereby minimizing the loss function value of the aircraft fault diagnosis model.

The aircraft fault diagnosis model may be a commonly used machine learning model (deep learning model); preferably, in an optional implementation manner, the above-mentioned machine learning model is a Transformer model.

The related technical solutions are the same as those in Embodiment 2, and are not repeated here.

Embodiment 4,

A method for diagnosing aircraft faults, comprising: inputting flight sampler data into an aircraft fault diagnosis model constructed by using the method for constructing an aircraft fault diagnosis model provided in Embodiment 3 to obtain aircraft fault information.

The related technical solution is the same as that of Embodiment 3, and will not be repeated here.

Embodiment 5,

A machine-readable storage medium storing machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement the machine-executable instructions Among the training data generation model construction method provided in Example 1, the training data generation method provided in Example 2, the construction method of the aircraft fault diagnosis model provided in Example 3, and the aircraft fault diagnosis method provided in Example 4 one or more.

The related technical solutions are the same as those of Embodiment 1 to Embodiment 4, and will not be repeated here.

In order to further illustrate the training data generation model provided by the present invention, detailed description is given below in conjunction with Comparative Experiment 1 and Comparative Experiment 2:

Comparative experiment 1.

In this experiment, after training the same aircraft fault diagnosis model using the actual collected aircraft sample data set (small data set) and the data set expanded by the training data generation model provided by the present invention, the four After the fault diagnosis of the faulty actuators (respectively denoted as f _a 1, f _a 2, f _a 3 and f _a 4) is carried out, they are compared with the real fault diagnosis values, and the experimental results shown in Figure 5 are obtained; among them, the horizontal The coordinates are the sampling points obtained by sampling 100 times the data of the aircraft flying continuously for 1 s; the ordinate is the diagnostic value at the sampling point. In this experiment, based on the above model, each sampling point was diagnosed separately, and then arranged to obtain the results shown in Figure 5; it can be seen from Figure 5 that compared with the failure of the aircraft fault diagnosis model trained with a small data set As for the diagnosis result, the fault diagnosis result of the aircraft fault diagnosis model obtained by training the expanded data set of the training data generation model provided by the present invention is closer to the real fault diagnosis value. In addition, calculate the mean square error MSE of the fault diagnosis result of the aircraft fault diagnosis model obtained by training with the small data set and the fault diagnosis result of the aircraft fault diagnosis model obtained by training the data set after the training data generation model provided by the present invention is expanded, The diagnostic MSE corresponding to the small data set is 0.025256619 and the diagnostic MSE after data expansion is 0.0022036468. It can be seen from the above experiments that the training data generation model provided by the present invention has an order of magnitude optimization for the accuracy of the aircraft diagnosis result.

Comparative experiment 2,

In this experiment, after training the same aircraft fault diagnosis model using the training data sets generated by the first experimental model and the second experimental model, the four fault actuators of the aircraft (respectively denoted as f _a 1, f _a 2 , f _a 3 and f _a 4) after fault diagnosis, compare with the real fault diagnosis value to obtain the experimental results shown in Figure 6; wherein, the first experimental model and the second experimental model are respectively provided by the present invention. The training data generation model (the model with MLP-Mixer added) and the model after removing the MLP-Mixer in the training data generation model provided by the present invention (the model without MLP-Mixer). Similarly, the abscissa is the sampling point obtained by sampling 100 times the data of the aircraft flying continuously for 1 s; the ordinate is the diagnostic value at the sampling point. In this experiment, fault diagnosis is performed for each sampling point based on the above model, and then arranged to obtain the results shown in Figure 6; as can be seen from Figure 6, compared with the fault diagnosis based on the model without MLP-Mixer As a result, the fault diagnosis result obtained based on the training data generation model provided by the present invention (the model with MLP-Mixer) is closer to the real fault diagnosis value. In addition, calculate the mean square error MSE of the fault diagnosis result obtained based on the model without MLP-Mixer and the fault diagnosis result based on the training data generation model provided by the present invention (the model with MLP-Mixer), and obtain the result without the addition of MLP-Mixer. The diagnostic MSE corresponding to the MLP-Mixer model is 0.0013495281 and the diagnostic MSE corresponding to the MLP-Mixer model is 0.0005102753. It can be seen from the above experiments that by introducing the MLP-Mixer in the present invention, the accuracy of the diagnostic result of the aircraft can be greatly improved, and the performance is better.

Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (10)

1. A training data generation model construction method for aircraft fault diagnosis is characterized by comprising the following steps:

s1, building a training data generation model;

the training data generation model includes: a generator and a discriminator; the generator is based on a VAE model and is used for sequentially carrying out encoding and decoding operations on input aircraft sample data so as to reconstruct the aircraft sample data and obtain pseudo data similar to the aircraft sample data; the discriminator is used for discriminating whether the input data is real aircraft data;

s2, inputting a pre-collected aircraft sample data set into the training data generation model for training, and enabling the generator and the discriminator to play games with each other until Nash equilibrium is achieved; the aircraft sample data set includes aircraft fault sample data.

2. The training data generation model construction method according to claim 1, wherein the generator includes: a first MLP-Mixer, a VAE model, and a second MLP-Mixer in cascade;

the first MLP-Mixer is used for expanding dimensionality of input aircraft sample data and integrating features of different channels of the expanded aircraft sample data, so that the aircraft sample data is mapped into depth data features;

the VAE model is used for coding the depth data features through a coder to obtain a mean value and a variance corresponding to the depth data features, resampling is carried out in normal distribution corresponding to the obtained mean value and variance to obtain a hidden variable, and the hidden variable is reconstructed through a decoder to obtain pseudo data features similar to the depth data features;

the second MLP-Mixer is used for mapping the pseudo data characteristics according to different channels, so that the pseudo data characteristics are reversely mapped into pseudo data similar to the aircraft sample data.

3. The training data generative model construction method according to claim 1 or 2, wherein said step S2 comprises:

s21, respectively generating pseudo data corresponding to the sample data for each sample data in the aircraft sample data set through the generator, and respectively inputting the sample data and the corresponding pseudo data into the discriminator to obtain the prediction probability that the sample data and the corresponding pseudo data are respectively discriminated as the real data of the aircraft;

s22, calculating the sum of the expression difference and the distribution difference between the sample data and the corresponding pseudo data to obtain the loss value of the generator; obtaining a loss value of the discriminator by calculating the sum of differences between the predicted probability of the sample data and the corresponding true probability of the aircraft true data respectively judged by the sample data and the corresponding pseudo data;

s23, judging whether the sum of the loss value of the generator and the loss value of the discriminator reaches the minimum or whether the iteration number reaches the preset iteration number, if so, finishing training of the training data generation model, and ending the operation; otherwise, the parameters in the generator and the discriminator are updated, and the process goes to step S21.

4. The training data generation model construction method according to claim 3, wherein the loss value of the generator is:

L_G＝MSE+KLD

wherein MSE is the expression difference between the sample data and the corresponding dummy data; n is the number of sample data in the aircraft sample data set;

the data is the pseudo data corresponding to the ith sample data;

the ith sample data; KLD is KL divergence of sample data distribution and corresponding pseudo data distribution; mu.s_iAnd σ_iRespectively, the mean and the variance corresponding to the ith sample data output by the encoder.

5. The training data generative model construction method according to claim 3, wherein the loss value of the discriminator is:

wherein,

judging the sample data as the prediction probability of the real data of the aircraft;

the sample data is the true probability of the true data of the aircraft;

and judging the pseudo data corresponding to the sample data as the prediction probability of the real data of the aircraft.

6. A method of generating training data for aircraft fault diagnosis, comprising:

expanding the aircraft sample data set by controlling the expansion times of the pre-collected sample data in the aircraft sample data set, so that the number of the sample data in the expanded aircraft sample data set reaches a preset number, and the number of the aircraft fault sample data and the number of the aircraft normal sample data reach a balance; the extended aircraft sample data set is a training data set for aircraft fault diagnosis;

the extended sample data is aircraft fault sample data and/or aircraft normal sample data;

the method for expanding the sample data comprises the following steps:

inputting the sample data into a generator of a training data generation model constructed by adopting the training data generation model construction method according to any one of claims 1 to 5 to obtain pseudo data corresponding to the sample data so as to expand the sample data; and the fault information label of the pseudo data is the same as the fault information label of the sample data corresponding to the pseudo data.

7. A method for constructing an aircraft fault diagnosis model is characterized by comprising the following steps:

inputting a training data set generated by the training data generation method of claim 6 into a machine learning model for training to obtain an aircraft fault diagnosis model.

8. The method of constructing an aircraft fault diagnosis model of claim 7, wherein the machine learning model is a Transformer model.

9. An aircraft fault diagnosis method, comprising: inputting flight sampler data into the aircraft fault diagnosis model constructed by the aircraft fault diagnosis model construction method according to claim 7 or 8 to obtain aircraft fault information.

10. A machine-readable storage medium having stored thereon machine-executable instructions that, when invoked and executed by a processor, cause the processor to perform one or more of the training data generation model construction method of any one of claims 1 to 5, the training data generation method of claim 6, the aircraft fault diagnosis model construction method of any one of claims 7 to 8, and the aircraft fault diagnosis method of claim 9.

Images