CN111597760B - A method for obtaining the deviation value of gas path parameters under the condition of small sample - Google Patents

Abstract

The application discloses a method for obtaining a gas circuit parameter deviation value under a small sample condition, and belongs to the technical field of health management and monitoring of an aircraft engine. The method comprises the steps of constructing an engine sample data set, carrying out normalization pretreatment on data, constructing a depth field self-adaptive gas circuit parameter deviation value regression model, training the depth field self-adaptive gas circuit parameter deviation value regression model by using a target field engine and a source field engine training set, testing an engine sample extracted from the target field engine test set by using the trained depth field self-adaptive gas circuit parameter deviation value regression model, analyzing a regression effect, obtaining small sample new model aeroengine ACARS data, and obtaining a gas circuit parameter deviation value by using the stored gas circuit parameter deviation value regression model. The method realizes the cross-working condition and cross-machine type establishment of the gas circuit parameter deviation value model, and further obtains the monitoring autonomy of the engine.

Description

A method for obtaining the deviation value of gas path parameters under the condition of small sample

technical field

The present application belongs to the technical field of aero-engine monitoring and health management, and relates to a method for obtaining the deviation value of gas path parameters, in particular to a method for obtaining the deviation value of gas path parameters under the condition of small samples.

Background technique

Air circuit parameter monitoring is an important technical means for the health management of aero-engines, especially civil aviation engines. As a kind of heat engine, the core components of aero-engines are air system components, such as compressors, combustion chambers, turbines, etc. The thermal parameters of the air path components reflect the performance state of the engine. The commonly used air path parameters are: EGT, FF, N1, N2, etc. These parameters are collected by airborne equipment and transmitted to the aircraft monitoring base in the form of ACARS (Aircraft CommunicationAddressing and Reporting System) messages. The aero-engine gas path analysis method is that the state parameters are collected by the airborne monitoring system in the take-off phase (Take off) and the cruise phase (Cruise) of the aircraft, respectively, and sent to the monitoring base in the form of ACARS messages. The original state parameters are obtained by analysis, and then the deviation value of the gas path parameters is calculated by a specific model, and then the deviation value is smoothed to facilitate the observation of the deviation change trend, so as to understand the performance of the engine. Finally, in order to "pre-diagnose" the engine, it is also necessary to predict the deviation value of the engine gas path parameters. It can be seen that establishing an accurate model of engine gas path parameter deviation value is the premise of engine state monitoring and fault diagnosis.

However, due to the complex working environment and diverse models of civil aviation engines, the model of gas path parameter deviation value lacks universality and the new model has the problem of lack of usable information. That is, the prior art lacks a regression model for the deviation value of civil aviation engine gas path parameters under cross-working conditions and cross-models, and a corresponding method for mining the deviation value of engine gas path parameters for realizing knowledge transfer and reuse.

SUMMARY OF THE INVENTION

The technical problem solved by this application is a domain adaptation problem in the field of aero-engine health management. The purpose of the present invention is to provide a method for obtaining the deviation value of gas path parameters under the condition of small sample. By transferring the associated knowledge of flight parameters learned under different aircraft types to each other, a model of gas path parameter deviation value can be established under different operating conditions and aircraft types, and then the autonomy of engine monitoring can be obtained.

In order to achieve the above purpose of the invention, the present invention adopts the following technical solutions.

A method for obtaining the deviation value of gas path parameters under the condition of small sample, comprising:

Step 1: Collect source domain and target domain aero-engine ACARS data, construct an engine sample data set, and divide the engine sample data set into a training set and a test set;

Step 2 normalizes the data of the training set and the test set of the source domain and target domain engine;

Step 3: constructing a regression model of the deviation value of the depth domain adaptive gas path parameter, and the depth domain adaptive gas path parameter deviation value regression model is composed of three parts: a feature extraction module, a domain adaptive module and a regression module;

Specifically include:

Step 3.1 By stacking multi-layer Res-BP residual learning modules, build a gas path high-order feature extraction module, model architecture and hyperparameters refer to the gas path parameter deviation value regression model established under the condition of a large sample;

Step 3.2 Build a deep domain adaptation module, stacking multi-layer multi-core mean difference adaptation layers to measure the distribution difference between the source domain and target domain engines;

Step 3.3 build a regression module of the gas path parameter deviation value, and connect the regression module after the depth domain adaptive module to realize the gas path parameter deviation value regression;

Step 4 uses the target domain engine and the source domain engine training set to train the above-mentioned deep domain adaptive gas path parameter deviation value regression model;

Step 5: Test the engine samples extracted from the target domain engine test set by using the trained deep domain adaptive air path parameter deviation value regression model;

Step 6: Analyze the regression effect of the regression model of the self-adaptive air path parameter deviation value in the depth field, and evaluate it. If the accuracy meets the requirements, save the current depth field self-adaptive air path parameter deviation value regression model, and return if it is unqualified. step 1;

Step 7: Acquire the ACARS data of a small sample of the aero-engine of the new model, and obtain the deviation value of the gas path parameter by using the saved regression model of the deviation value of the gas path parameter, and obtain the regression result.

Optionally, Z-score normalization preprocessing is used in step 2.

Optionally, in step 3.2, in order to fully exploit the domain invariant features and achieve the optimal depth domain adaptation effect, a domain confrontation mechanism is introduced, and the domain discriminator and the gradient inversion layer are connected after the multi-layer multi-core mean difference adaptation layer. to maximize domain confusion.

Optionally, the feature extraction module realizes the feature extraction of the air path state by stacking three Res-BP residual learning blocks, and the feature extraction module has ten layers in total, including one input layer; each Res-BP residual learning block is mainly composed of It consists of three fully-connected layers and a residual identity shortcut; the output of each fully-connected layer first adopts batch regularization, and then activates each neuron through the SELU activation function.

Optionally, the depth domain adaptation module is composed of a domain confusion layer and three MK-MMD domain adaptation layers; the three MK-MMD domain adaptation layers are connected with the feature extraction module, and the MK-MMD metric source is used. The difference in the distribution of the extracted higher-order features in the domain and the target domain; the domain confusion layer is connected with the MK-MMD domain adaptation layer and includes a binary domain classifier whose output is the domain label value.

Optionally, the regression module is connected to a domain confusion layer, and mainly realizes the mapping between the extracted domain invariant features and their corresponding gas path parameter deviation values, and finally realizes deviation value mining.

The deep domain adaptive air path parameter deviation value regression model has the following optimization objectives:

Minimize the regression error of the gas path parameter deviation value on the source domain data set and the target domain data set;

Minimize the distribution difference between the source and target domains;

and maximizing the domain confusion loss on the source and target domains.

Compared with the prior art, the present invention achieves the following beneficial effects.

The invention uses the Res-BPNN model to deeply mine the high-order features of the air path state of the aero-engine, stacks multi-layer multi-core maximum mean difference MK-MMD adaptation layers, and then maps the extracted high-order features to the RKHS for difference measurement. In order to further reduce the difference in the probability distribution of air path state features of various models, a maximizing domain confusion method based on adversarial mechanism is introduced. The learned feature distributions are as close as possible, so as to confuse the source of the learned feature distribution in the current state of the target engine, and then deeply mine the domain invariant features to achieve the optimal depth domain self-adaptation effect. The invention successfully solves the problems that the air path parameter deviation value model lacks universality due to the complex working conditions and diverse models of civil aviation engines and the lack of usable information for new models. The related knowledge between them can be transferred to each other to realize the establishment of a gas path parameter deviation value model under cross working conditions and cross models, and then obtain the autonomy of engine monitoring.

Description of drawings

Fig. 1 is the overall flow chart of the first embodiment of the application;

2 is a schematic diagram of maximizing field confusion in Embodiment 1 of the present application;

Fig. 3 is the overall framework of the deep domain adaptive regression model based on Res-BPNN in the first embodiment of the application;

FIG. 4 is a graph showing the variation of MK-MMD and loss value with the number of iterations in the network training process of Embodiment 2 of the present application;

Fig. 5 is the DEGT regression effect diagram (unsupervised method) of the transfer task A of the second embodiment of the application;

FIG. 6 is a DEGT regression effect diagram (supervised method) of the transfer task A in the second embodiment of the present application.

Detailed ways

In the solutions provided in the embodiments of the present application, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

In order to better understand the above technical solutions, the technical solutions of the present application will be described in detail below through the accompanying drawings and specific embodiments. It is not a limitation on the technical solutions of the present application, and the embodiments of the present application and the technical features in the embodiments may be combined with each other under the condition of no conflict.

Example 1

A flow chart of a method for obtaining a deviation value of a gas path parameter provided by this embodiment is shown in FIG. 1 ,

The method includes:

Step 1: Collect the source domain and target domain aero-engine ACARS data, construct an engine sample data set, and divide the engine sample data set into a training set and a test set; usually 80% of the training set and 20% of the test set are divided;

In step 2, normalization preprocessing is performed on the data of the training set and the test set of the engine in the source domain and the target domain; in this embodiment, the ACARS data obtained from the civil aviation company can be used to directly perform sample normalization;

Step 3 constructs a regression model of the deviation value of the air path parameter of the depth field adaptation, and the regression model of the deviation value of the air path parameter of the depth field adaptation is composed of three parts: a feature extraction module, a field adaptation module and a regression module; it specifically includes:

Step 3.1 By stacking multi-layer Res-BP residual learning modules, build a gas path high-order feature extraction module, model architecture and hyperparameters refer to the gas path parameter deviation value regression model established under the condition of a large sample;

Step 3.2 Build a deep domain adaptation module, stacking multi-layer multi-core mean difference adaptation layers to measure the distribution difference between the source domain and target domain engines;

Step 3.3 build a regression module of the gas path parameter deviation value, and connect the regression module after the depth domain adaptive module to realize the gas path parameter deviation value regression;

Step 4 uses the target domain engine and the source domain engine training set to train the above-mentioned deep domain adaptive gas path parameter deviation value regression model;

Step 5: Test the engine samples extracted from the target domain engine test set by using the trained deep domain adaptive air path parameter deviation value regression model;

Step 6: Analyze the regression effect of the regression model of the self-adaptive air path parameter deviation value in the depth field, and evaluate it. If the accuracy meets the requirements, save the current depth field self-adaptive air path parameter deviation value regression model, and return if it is unqualified. step 1;

Step 7: Acquire the ACARS data of a small sample of the aero-engine of the new model, and obtain the deviation value of the gas path parameter by using the saved regression model of the deviation value of the gas path parameter, and obtain the regression result.

In step 2, Z-score normalization preprocessing is used.

In step 3.2, in order to fully exploit the domain invariant features and achieve the optimal depth domain adaptation effect, domain adversarial values are introduced, and domain confusion is achieved by connecting the domain discriminator and gradient inversion layer after the multi-layer multi-core mean difference adaptation layer. maximize.

The feature extraction module realizes the feature extraction of the air path state by stacking three Res-BP residual learning blocks. The feature extraction module has ten layers in total, including one input layer; each Res-BP residual learning block is mainly composed of three fully connected layers and It consists of a residual identity shortcut; the output of each fully connected layer is first batch regularized, and then each neuron is activated through the SELU activation function.

The deep domain adaptation module consists of a domain confusion layer and three MK-MMD domain adaptation layers; the three-layer MK-MMD domain adaptation layer is connected with the feature extraction module, and the MK-MMD is used to measure the range of the source domain and the target domain. Differences in the distribution of extracted higher-order features; the domain confusion layer is concatenated with the MK-MMD domain adaptation layer, including a binary domain classifier whose output is the domain label value.

The regression module is connected with the domain confusion layer, which mainly realizes the mapping between the extracted domain invariant features and their corresponding gas path parameter deviation values, and finally realizes deviation value mining.

The idea of deep domain adaptation methods is to match the feature distribution between the source domain and the target domain to learn domain-invariant features. The so-called learning domain invariant features means that the features learned from the source domain data or the target domain data should obey the same or almost the same feature distribution. If the feature has domain invariant characteristics, the feature can be used to effectively realize the task of mining the deviation value of air path parameters in the target domain data. Therefore, learning domain invariant features is the key to realize knowledge transfer and reuse. The following is a further description of the deep adaptation problem from the method of measuring the dissimilarity between domains and the method of domain confusion.

The Maximum Mean Discrepancy (MMD) mainly measures the difference between the two distributions by mapping two different distributions into the Reproducing Kernel Hilbert Space (RKHS).

以及属于t分布的样本集合

定义样本空间的连续函数集F，其连续映射f:χ→R，则两个分布之间的MMD，如公式(3-1)所示。First, define the set of samples belonging to the s-distribution

and the set of samples belonging to the t-distribution

Define the continuous function set F of the sample space, and its continuous mapping f:χ→R, then the MMD between the two distributions is shown in formula (3-1).

In the formula: E _s [ ]——represents the mathematical expectation of the distribution s;

E _t [ ] - represents the mathematical expectation of the distribution t.

Assume that the sample sets X _S and X _T are two sample sets obtained by independent and identical distribution sampling from the distributions s and t, respectively, and the sample sizes are m and n, respectively. An empirical estimate of MMD can be obtained based on X _S and X _T , as shown in formula (3-2).

Second, define H to represent the RHKS space, and F to be constrained to be a unit sphere within the RHKS space. Continuous mapping f:χ→R, for any x∈χ, the function f(x) is shown in formula (3-3).

——代表映射：χ→H，并且f与

的内积定义为核函数，如高斯核函数：

where:

——represents the mapping: χ→H, and f is equal to

The inner product of is defined as a kernel function, such as a Gaussian kernel function:

Therefore, through the above two definitions, it can be known that the MMD of the s distribution and the t distribution in the regenerated kernel Hilbert space is shown in formula (3-4).

In the formula: μ _s =E _s [f(X _S )], μ _t =E _t [f(X _T )].

Finally, in order to facilitate the calculation, the square form of MMD is usually adopted, such as formula (3-5).

In the formula: m, n——represent the sample size of the source domain dataset X _S and the target domain dataset X _T , respectively. When the distribution difference between X _S and X _S is smaller, the MMD distance is smaller, if and only if s and t obey the uniform distribution, MMD=0.

Deep Adaptation Networks can use the MMD variant algorithm, namely Multi-Kernel Maximum Mean Discrepancy (MK-MMD), which can better achieve domain adaptation and enhance the feature representation capability of neural networks. .

First, the definition of MK-MMD is given, and the empirical estimation of the MK-MMD distance between distribution s and distribution t is shown in formula (3-6).

where: H _k ——represents the RKHS with characteristic kernel k;

f( )——represents a continuous mapping function;

E _s [ ]——represents the mathematical expectation of the distribution s;

E _t [ ] - represents the mathematical expectation of the distribution t.

The feature kernels f( ), _k (x, y)=<f(x), f(y)> in MK-MMD, the multi-kernel is specifically defined as a convex combination with m positive semi-definite kernels ku, such as formula ( 3-7).

——为约束系数，以保证不同分布之间进行领域适配时存在多内核k独有特性。where:

- is a constraint coefficient to ensure that there are unique characteristics of multi-kernel k when performing domain adaptation between different distributions.

Since different feature distributions will change when the deep neural network is learning, it is impossible to determine which kernel function can show stronger mapping ability, and the multi-kernel k based on MK-MMD can enhance the feature distribution through different kernel functions. The adaptability to achieve the optimal and most reasonable kernel function selection.

In this embodiment, in order to further reduce the marginal distribution difference between features learned from different domains, a domain confusion method based on an adversarial mechanism is introduced, the schematic diagram of which is shown in FIG. 2 .

In this embodiment, in order to maximize the domain confusion, a domain confusion layer is added after the extracted depth feature layer, that is, it is determined whether the feature distribution of a certain sample after training comes from the source domain or the target domain. The more the extracted features can reflect the commonality between domains, the better the domain confusion effect will be. If the deep features extracted by the Res-BP neural network cannot be discriminated by the trained domain classifier whether it is a sample from the source domain or the target domain, the deep feature is said to have domain-invariant characteristics. The optimal effect of the domain classifier D can be optimized by formula (3-8).

where:

m——represents the batch size of training samples;

gi ——represents the true domain label of the _ith sample, _gi = 0 represents that the sample _xi comes from the source domain, and _gi =1 represents that the sample _xi comes from the target domain;

D( _xi )——represents the label value output by the sample _xi after passing through the domain classifier.

The deep domain adaptive regression model based on Res-BPNN provided in this embodiment is composed of three parts: a feature extraction module, a domain adaptation module and a regression module. The high-order feature extraction module adopts the Res-BPNN network structure to extract the air path state features by stacking N Res-BP residual learning blocks. The deep domain adaptation module is connected with the feature extraction module to learn domain invariant features. Finally, a regression module is added to form a deep domain adaptive regression model framework based on Res-BPNN, as shown in Figure 3.

(1) Feature extraction module: The feature extraction of air path state is realized by stacking three Res-BP residual learning blocks. The feature extraction module has ten layers in total, including one input layer. Each Res-BP residual learning block mainly consists of three fully connected layers and a residual identity shortcut. The output of each fully connected layer is first batch regularized, and then each neuron is activated through the SELU activation function.

(2) Deep domain adaptation module: The deep domain adaptation module consists of one domain confusion layer and three MK-MMD domain adaptation layers. The three-layer MK-MMD domain adaptation layer is connected with the feature extraction module, and the MK-MMD is used to measure the difference of the extracted higher-order feature distributions in the source and target domains. The domain confusion layer is connected with the MK-MMD domain adaptation layer, including a binary domain classifier whose output is the domain label value.

(3) Regression module: The regression module is connected with the domain confusion layer, and mainly realizes the mapping between the extracted domain invariant features and their corresponding gas path parameter deviation values, and finally realizes deviation value mining.

At present, the airline has introduced a new model of civil aviation engine, and only has the labelled data of this model in small batches. In order to establish a model of the gas path parameter deviation value of this model of civil aviation engine in a short period of time, it is necessary to make full use of the labelled data resources. This embodiment The regression error loss of the labeled data in the mini-batch target domain will be introduced into the final optimization objective of adaptive transfer learning in the traditional unsupervised domain. In order to facilitate comparison with other transfer learning algorithms, the proposed methods are divided into proposed methods (supervised) and proposed methods (unsupervised) according to whether the regression loss of labeled data in the target domain is introduced into the final optimization objective.

Therefore, the deep domain adaptive regression model based on Res-BPNN proposed in this embodiment finally has the following three optimization objectives:

(1) In order to realize the deviation value mining of aero-engine gas path parameters, the deep domain adaptive regression model based on Res-BPNN must realize the feature extraction of the gas path and the deviation value mining. Therefore, the first optimization goal of the deep domain adaptive regression model is to minimize the regression error of the deviation value of the gas path parameters on the source domain data set and the target domain data set. The loss function of the regression error is defined as the standard MSE loss function, such as the formula (3-9).

where:

n _S , n _T ——represent the number of source domain data and target domain data in a batch training set;

y _S , y _T ——represent the true value of the gas path parameter deviation between the source domain sample and the target domain sample;

——代表源域样本与目标域样本气路参数偏差值的回归值；

——Regression value representing the deviation of the gas path parameters between the source domain sample and the target domain sample;

β—represents the training weights for labeled data in the target domain. If and only if β = 0, no labeled datasets in the target domain are introduced into the final optimization objective.

(2) The deep domain adaptation module is used to learn domain invariant features, as shown in Figure 3-2, mainly including domain confusion layer and multi-layer MK-MMD domain adaptation layer. The MK-MMD domain adaptation layer is mainly to measure the distribution difference between the learned features in different domains. Therefore, the second optimization goal of the deep domain adaptive regression model is to minimize the distribution difference between the source and target domains. MK-MMD is used to measure the distribution difference of high-order features between the source domain and the target domain. The MK-MMD loss function is shown in formula (3-10).

where:

——分别代表第i层适配层的源域与目标域特征表示；

——represent the source domain and target domain feature representation of the i-th adaptation layer;

X _S and X _T - represent the source domain dataset and the target domain dataset, respectively;

l _i ——represents the i-th adaptation layer;

H _k - represents the RKHS with feature kernel k.

(3) In order to further reduce the difference in the probability distribution of the state characteristics of the gas path of each model, the maximum domain confusion method of the confrontation mechanism is adopted to make the learned feature distributions between different domains as close as possible, so as to confuse the current state of the target engine. source of feature distribution, and then deeply mine domain-invariant features. Therefore, the third optimization objective of the deep domain adaptive regression model is to maximize the domain confusion loss on the source and target domains, and the domain classification loss function is shown in Equation (3-11).

——代表域混淆层上已学习的特征表示。where:

— represents the learned feature representation on the domain confusion layer.

By combining the MSE regression loss, the MK-MMD loss and the domain confusion loss, the final optimization objective of the deep domain adaptive regression model based on Res-BPNN can be obtained as shown in formula (3-12).

L=L _reg + λL _MK-MMD + μL _domain (3-12)

where:

L _MK-MMD - represents the domain adaptation loss;

L _reg - represents regression loss;

L _domain — represents the domain confusion loss;

λ, μ——represent network hyperparameters, which are used to control the strength of adaptability in the depth domain;

When the model training is completed, if the learned high-order features have ambiguous domain categories and small differences between domains, the airway parameter deviation value regression module can accurately mine the deviation value of the target domain samples.

Embodiment 2

In the second embodiment, the technical solution of the first embodiment is experimentally verified by using the historical cruise data of the civil aviation engine. The following is a detailed description of data sampling and preprocessing, hyperparameter settings, and model performance comparison.

In order to fully verify the application of the deep domain adaptive regression model based on Res-BPNN proposed in the first embodiment in the field of civil aviation engine gas path parameter deviation value mining and the general applicability of the model, this embodiment is from CFM56-5B2 produced by GE. /3 and CFM56-7B26 two different types of civil aviation engines, respectively, two sets of data sets were obtained and applied to two gas path parameter deviation value mining experiments. The two sets of transfer learning tasks are transfer task A: CFM56-5B2/3→CFM56-7B26 and transfer task B: CFM56-7B26→CFM56-5B2/3.

Migration task A (CFM56-5B2/3→CFM56-7B26) means: firstly carry out supervised training from the collected labeled data of the CFM56-5B2 civil aviation engine, and assist the CFM56-7B26 civil aviation engine at the same time. Domain-adaptive knowledge transfer with small batches of data. Migration task B (CFM56-7B26→CFM56-5B2/3) is to exchange the source domain dataset and target domain dataset in the migration task CFM56-5B2/3→CFM56-7B26 and re-run the migration experiment to verify the model effectiveness. Table 3-1 and Table 3-2 show some data and data set distribution ratios of the two groups of transfer learning experiments.

As shown in Table 3-1, the historical cruise data of two different types of civil aviation engines, CFM56-5B2/3 and CFM56-7B26 used in this example also need to eliminate the dimensional influence between parameters, so it needs to be standardized and preprocessed , using the Z-score standardized preprocessing method, so that the sample parameters are comparable, and can improve the network training rate, the conversion formula is shown in formula (2-9).

Table 3-1 Partial data of two different types of engines

Table 3-2 Dataset assignments for two groups of different transfer learning tasks

The corresponding relationship between the input and output of the regression model of the gas path parameter deviation value is shown in formula (3-13).

where:

I _D- [ ]——represents the input corresponding to the deviation value of each gas path parameter;

O _{D- ——represents} the output of the deviation value of each gas path parameter;

x _i — represents the measured value of the i-th flight parameter.

The optimization method adopts the stochastic gradient descent algorithm SGD, and defines θ _f , θ _d , θ _reg as the optimization parameters of the feature extractor, the domain classifier and the final regressor, respectively. The formula (3-12) can be rewritten as formula (3-14) ).

Based on the stochastic gradient descent algorithm SGD and formula (3-14), the updating process of parameters θ _f , θ _d , and θ _reg can be written as formula (3-15).

In the formula: α——represents the learning rate of the neural network, which can be self-adjusted by formula (3-16).

where:

epoch, epochs - represent the number of times the network has completed training and set the number of training times to be completed;

α ₀ , β, δ——represent constants, here are 0.01, 0.75 and 10;

t——represents the network training progress, which varies linearly from 0 to 1.

The self-adjustment process of the hyperparameters λ and μ in Eq. (3-13) is similar to the self-adjustment process of the learning rate. This method makes the domain confusion layer and the domain adaptation layer more sensitive to noise signals in the early stage of the training process. low, its self-adjusting formula is shown in formula (3-17).

In the formula: γ——represents a constant, which is taken as 10 here.

The model optimizer uses SGD, the model training batch size (Batch-size) is set to 100, the number of iterations (Epochs) is set to 1000, the momentum ratio (Momentum) is set to 0.9, the activation function is SELU, and the initial weight and bias are Defaults. In this embodiment, the feature extraction module is composed of three Res-BP residual learning modules stacked, and the depth domain adaptation module is composed of three MK-MMD domain adaptation layers and one domain confusion layer, so as to fully mine the source domain and the Domain-invariant features between target domains.

In order to further prove the effectiveness of the deep domain adaptive regression model based on Res-BPNN proposed in Example 1, nine different transfer learning algorithms are used to establish three key gas path parameter deviation value regressions for transfer task A and transfer task B respectively. model and conduct experimental comparisons. The comparison methods used are shown in Table 3-3. Transfer Component Analysis (TCA), Joint Distribution Adaptation (JDA), Balanced Distribution Adaptation (BDA), Deep Domain Confusion Network ( Deep Domain Confusion (DDC), Deep Adaptation Networks (DAN), Domain-Adversarial Neural Network (DANN) and the method (unsupervised) proposed in the first embodiment of the present application are all unsupervised transfer learning algorithm. Considering the actual situation of the introduction of new models of civil aviation engines by civil aviation companies, in order to make full use of the existing small batch labeled target domain dataset resources, the regression error can be integrated into the final optimization goal, and fine-tuned with the Res-BPNN model and layer transfer. (Fine-tuning) and other supervised transfer learning for experimental comparison.

Table 3-3 Transfer learning methods

^a. The target domain data is not introduced into the final optimization objective, ^b. The target domain data is introduced into the final optimization objective.

The above-mentioned supervised and unsupervised learning methods were used to establish three key gas path parameter deviation value regression models of DEGT, DFF and DN2 for experimental verification, and compared with the gas path parameter deviation values provided by OEM. The performance comparison results of migration task B are shown in Table 3-4 to Table 3-6 and Table 3-7 to Table 3-9 respectively.

Table 3-4 Comparison of DEGT regression effects of migration task A

Table 3-5 Comparison of DFF regression effects of migration task A

Table 3-6 Comparison of DN2 regression effects of migration task A

Table 3-7 Comparison of DEGT regression effects of migration task B

Table 3-8 Comparison of DFF regression effects of migration task B

Table 3-9 Comparison of DN2 regression effects of migration task B

As shown in Table 3-4 to Table 3-9, the deep domain adaptive regression model based on Res-BPNN proposed in this application predicts three key performance deviation values DEGT, DFF and DN2 in transfer task A and transfer task B The fitting results obtained when , are the most prominent among all methods.

(1) By analyzing and comparing the experimental results of the two groups of unsupervised algorithms of TCA, JDA and BDA, as well as DDC, DAN, DANN and the proposed method (unsupervised), it can be found that the experimental results of the second group are compared with the migration of the first group. The effect is better. The three migration algorithms in the first group need to select the first m eigenvalues in the process of feature dimensionality reduction, and the useful information between flight parameters will be lost in the intermediate process, resulting in poor migration effect. The four migration algorithms in the second group all use deep networks to extract deep features, so they can better characterize the precise mapping relationship between air path parameters and their deviations, and fully reduce the differences between fields.

(2) By analyzing and comparing the four unsupervised algorithms DDC, DAN, DANN and the proposed algorithm (unsupervised), it can be found that the regression effect of the deep adaptive regression model based on Res-BPNN proposed in this application is the best. DDC learns domain-invariant features by introducing a layer of MMD domain adaptation layer into the Res-BPNN regression model architecture. On the basis of DDC, DAN can better learn domain-invariant features without increasing the additional training time of the network by introducing three-layer MK-MMD domain adaptation layers. DANN can deeply extract the common features between the source domain and the target domain by introducing the adversarial idea into the Res-BPNN regression model architecture. The method proposed in this application, by combining the advantages of DANN and DAN, and on the basis of the confrontation mechanism, additionally introduces three MK-MMD domain adaptation layers, which can more effectively reduce the difference in feature distribution between domains.

(3) By analyzing and comparing the three supervised learning algorithms Res-BPNN regression model, layer transfer fine-tuning and proposed method (supervised), it is found that the regression effect of the deep adaptive regression model based on Res-BPNN proposed in this application is also the best excellent. Since the new model of aero-engine has few labeled samples, the Res-BPNN regression model is directly used for training, which will inevitably lead to poor training effect of the deep network, and problems of model overfitting and poor regression accuracy. The layer migration fine-tuning method is used to pre-train the source domain regression model, and freeze the weights of the first n layers of network parameters, and use a small batch of labeled target domain data to fine-tune the parameters of the latter layers, learn high-order features, and realize the establishment of the bias value regression model. In this application, the deep domain adaptive regression model framework based on Res-BPNN, by introducing and adjusting the regression loss of the target domain mini-batch labeled dataset in the final optimization goal, transforms the unsupervised learning problem into a supervised learning problem, and fully exploits the The domain-invariant features between the source domain and the target domain further reduce the difference in domain feature distribution, and the transfer effect is significantly improved.

In order to directly reflect the superiority of the deep domain adaptive regression model based on Res-BPNN proposed in this application, the following is an explanation by selecting the regression experimental results of DEGT in the migration task A. The changes of MK-MMD and training loss with the number of iterations are as follows: shown in Figure 4. Since there are many unsupervised comparison experimental methods, in order to clearly reflect the fitting results of the method proposed in this application (unsupervised), only 50 test points are drawn here, as shown in FIG. 5 . The supervised comparison experiment results are plotted with 100 test points, as shown in Figure 6.

It can be seen from Figure 4 that with the continuous increase of the number of network iterations, the MK-MMD distribution difference measurement value and the network training loss value gradually decrease, and tend to converge to a certain extent. It can also be clearly seen from Figure 5 and Figure 6 that the deep domain adaptive regression model based on Res-BPNN proposed in this application can achieve better regression effects in both unsupervised learning algorithms or supervised learning algorithms, and further The effectiveness of the method proposed in this application is proved.

To sum up, the present invention provides a regression model of the deviation value of the air path parameters of a civil aviation engine under cross-working conditions and cross-models. According to the actual mapping relationship between the air path parameters of civil aviation engine and their deviation values, N Res-BP residual learning modules are used to extract air path state features, and multiple MK-MMD domain adaptation layers and domain confusion layers are used to deeply mine the source. The common features between the domain and the target domain are finally realized, and finally the deviation value regression is realized to form a deep domain adaptive regression model based on Res-BPNN. Taking into account the actual situation of the introduction of new aircraft models by civil aviation companies, in order to make full use of the small batch labeled data resources in the target domain, the regression loss of the target domain is introduced into the final optimization goal of the regression model. In order to verify the effectiveness of this model, whether the regression model is introduced with labeled target domain data, it is divided into two methods: supervised and unsupervised, and compared with other algorithms in experiments. The results show that the joint optimization method of domain adaptation and domain confusion can fully mine the domain invariant features between the source domain and the target domain, further reduce the difference in domain feature distribution, and improve the regression effect.

In the method steps adopted in the present invention, it is clarified that the processed data are all aero-engine ACARS data and how to process the ACARS data in each step, which shows that the deep domain adaptive regression model training method is closely related to aero-engine ACARS data processing. The solution proposed by the present invention solves the domain self-adaptation problem in the field of aero-engine health management, that is, how to transfer the associated knowledge between flight parameters learned under different aircraft types to each other, so as to realize cross-working conditions and cross-aircraft A model of the gas path parameter deviation value is established under the model, and then the technical problem of engine monitoring autonomy is obtained. The N Res-BP residual learning modules are used to extract the gas path state features, and multiple MK-MMD domain adaptation layers are used. The domain confusion layer deeply mines the common features between the source domain and the target domain, and finally realizes the technical means of deviation value regression. Invariant features, further reduce the difference in field feature distribution, and realize the method of obtaining the deviation value of gas path parameters under cross-working conditions and cross-models. Compared with the traditional modeling method, the regression effect is significantly improved.

As will be appreciated by those skilled in the art, the embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media having computer-usable program code embodied therein, including but not limited to disk storage, optical storage, and the like.

Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims (8)

1. A method for obtaining a gas circuit parameter deviation value under a small sample condition is characterized by comprising the following steps:

step 1, collecting source domain and target domain aircraft engine ACARS data, constructing an engine sample data set, and dividing the engine sample data set into a training set and a test set;

step 2, carrying out normalization pretreatment on the data of the training set and the test set of the engine in the source domain and the target domain;

step 3, constructing a depth-domain self-adaptive gas circuit parameter deviation value regression model, wherein the depth-domain self-adaptive gas circuit parameter deviation value regression model consists of a feature extraction module, a domain self-adaptive module and a regression module;

the step 3 specifically comprises:

step 3.1, a gas circuit high-order feature extraction module, a model framework and a gas circuit parameter deviation value regression model established under the condition of a super-parameter reference large sample are established by stacking a plurality of layers of Res-BP residual error learning modules;

step 3.2, a depth field self-adaptive module is constructed, and a plurality of layers of multi-core mean difference adaptation layer measurement source domains and distribution differences between target domain engines are stacked;

3.3, constructing a gas path parameter deviation value regression module, and connecting the regression module behind the depth field self-adaptive module to realize gas path parameter deviation value regression;

the depth field self-adaptive module consists of a field confusion layer and three MK-MMD field adaptation layers;

the three MK-MMD domain adaptation layers are connected with the feature extraction module and are used for measuring the difference of the extracted high-order feature distribution in the source domain and the target domain;

the domain confusion layer is connected with the MK-MMD domain adaptation layer and comprises a binary domain classifier, and the output of the binary domain classifier is a domain label value;

step 4, training the depth-domain self-adaptive gas circuit parameter deviation value regression model by using a target domain engine and a source domain engine training set;

step 5, testing the engine samples extracted from the target domain engine test set by using the trained depth domain self-adaptive gas circuit parameter deviation value regression model;

step 6, analyzing the regression effect of the depth field self-adaptive gas circuit parameter deviation value regression model, evaluating, if the precision meets the requirement, storing the current depth field self-adaptive gas circuit parameter deviation value regression model, and if the precision is not qualified, returning to the step 1;

and 7, acquiring ACARS data of the small-sample new aircraft engine, and acquiring a gas circuit parameter deviation value by using the stored gas circuit parameter deviation value regression model to obtain a regression result.

2. The method of claim 1, comprising:

and in the step 2, Z-score normalization pretreatment is adopted.

3. The method of claim 1, comprising:

in step 3.2, a domain confrontation value is introduced, and a domain confusion maximization is realized by connecting a domain discriminator and a gradient inversion layer after a plurality of layers of multi-core mean difference adaptation layers.

4. The method of claim 1, comprising:

the characteristic extraction module realizes gas circuit state characteristic extraction by stacking three Res-BP residual error learning blocks, and the characteristic extraction module comprises ten layers in total and comprises an input layer;

each Res-BP residual error learning block mainly comprises three full connection layers and a residual error constant shortcut;

the output of each fully-connected layer is firstly normalized in batch, and then each neuron is activated through a SELU activation function.

5. The method of claim 1, comprising:

and the regression module is connected with the domain confusion layer to realize mapping between the extracted domain invariant features and the corresponding gas circuit parameter deviation values, and finally realize deviation value mining.

6. The method according to one of claims 1 to 5, characterized in that:

the depth field self-adaptive gas circuit parameter deviation value regression model has the following optimization targets: and minimizing the regression error of the gas path parameter deviation values on the source domain data set and the target domain data set.

7. The method according to one of claims 1 to 5, characterized in that:

the depth field self-adaptive gas circuit parameter deviation value regression model has the following optimization targets: distribution differences between the source domain and the target domain are minimized.

8. The method according to any one of claims 1 to 5, wherein:

the depth field self-adaptive gas circuit parameter deviation value regression model has the following optimization targets: maximizing the domain confusion loss over the source domain and the target domain.

Images