CN111597760B - Method for obtaining gas path parameter deviation value under small sample condition - Google Patents
Method for obtaining gas path parameter deviation value under small sample condition Download PDFInfo
- Publication number
- CN111597760B CN111597760B CN202010418279.0A CN202010418279A CN111597760B CN 111597760 B CN111597760 B CN 111597760B CN 202010418279 A CN202010418279 A CN 202010418279A CN 111597760 B CN111597760 B CN 111597760B
- Authority
- CN
- China
- Prior art keywords
- domain
- deviation value
- gas circuit
- parameter deviation
- circuit parameter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Operations Research (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Feedback Control In General (AREA)
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
Technical Field
The application belongs to the technical field of aeroengine monitoring and health management, relates to a method for acquiring a gas circuit parameter deviation value, and particularly relates to a method for acquiring a gas circuit parameter deviation value under a small sample condition.
Background
The gas circuit parameter monitoring is an important technical means for the health management of the aircraft engine, particularly the civil aircraft engine. An aircraft engine is one of heat engines, and the core components of the aircraft engine are gas path system components such as a compressor, a combustion chamber, a turbine and the like. The thermodynamic parameters of the gas circuit component reflect the performance state of the engine, and the commonly used gas circuit parameters include: EGT, FF, N1, N2, and the like. The parameters are collected by airborne equipment and transmitted to an aircraft monitoring base in the form of ACARS (aircraft communication addressing and Reporting System) messages. The aircraft engine gas path analysis method is characterized in that state parameters acquired by an onboard monitoring system are respectively sent to a monitoring base in an ACARS message form in a Take-off stage (Take off) and a Cruise stage (Cruise) of an aircraft, different messages are analyzed by using a message format to obtain original state parameters, then a deviation value of the gas path parameters is calculated by using a specific model, and the deviation value is subjected to smoothing processing to conveniently observe the deviation change trend of the gas path parameters, so that the performance of the engine is known. Finally, in order to carry out 'pre-diagnosis' on the engine, the deviation value of the gas path parameters of the engine needs to be predicted. Therefore, the accurate model for the deviation value of the engine gas path parameters is established on the premise of engine state monitoring and fault diagnosis.
However, the working condition environment of the civil aviation engine is complex, and the models are various, so that the gas path parameter deviation value model is lack of universality, and the problem that the available information quantity of a new model is insufficient is caused. The method is characterized in that a regression model of the gas circuit parameter deviation value of the civil aviation engine and a corresponding mining method of the gas circuit parameter deviation value of the engine, which realizes knowledge transfer and reuse, are established under the condition of cross-working condition and cross-machine type.
Disclosure of Invention
The technical problem that this application was solved is field self-adaptation problem in the aeroengine health control field. The invention aims to provide a method for acquiring a gas path parameter deviation value under a small sample condition. The method has the advantages that association knowledge between learned flight parameters under different models is migrated mutually, so that a gas circuit parameter deviation value model is built under the conditions of crossing working conditions and the models, and further engine monitoring autonomy is obtained.
In order to achieve the above purpose, the invention adopts the following technical scheme.
A method for obtaining a gas path parameter deviation value under a small sample condition comprises the following steps:
step 1, collecting aeroengine ACARS data of a source domain and a target domain, 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 source domain and the target domain engine;
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 method specifically comprises the following steps:
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;
3.2, constructing a depth field self-adaptive module, and stacking a plurality of layers of multi-core mean difference adaptation layer measurement source domains and distribution differences between target domain engines;
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;
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 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.
Optionally, Z-score normalization pre-treatment is used in step 2.
Optionally, in step 3.2, in order to fully mine the domain invariant features and achieve the optimal depth domain adaptive effect, a domain countermeasure mechanism is introduced, and a domain confusion maximization is achieved by connecting a domain discriminator and a gradient inversion layer after a plurality of layers of multi-core mean difference adaptation layers.
Optionally, the feature extraction module realizes gas circuit state feature extraction by stacking three Res-BP residual error learning blocks, and the feature extraction module has ten layers including one 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.
Optionally, the depth domain adaptation module consists of one domain confusion layer and three MK-MMD domain adaptation layers; connecting the three MK-MMD domain adaptation layers with a feature extraction module, and measuring the difference of the extracted high-order feature distribution in the source domain and the target domain by using MK-MMD; the domain obfuscation layer is connected to the MK-MMD domain adaptation layer and includes a binary domain classifier, the output of which is a domain tag value.
Optionally, the regression module is connected to the domain confusion layer, and mainly implements mapping between the extracted domain invariant features and the corresponding gas circuit parameter deviation values, and finally implements deviation value mining.
The depth field self-adaptive gas circuit parameter deviation value regression model has the following optimization targets:
minimizing the regression error of the gas path parameter deviation value on the source domain data set and the target domain data set;
minimizing a distribution difference between the source domain and the target domain;
and maximizing domain confusion loss over the source domain and the target domain.
Compared with the prior art, the invention has the following beneficial effects.
According to the method, the Res-BPNN model is adopted to deeply mine the high-order characteristics of the air path state of the aero-engine, a plurality of multi-core maximum mean difference MK-MMD adaptation layers are stacked, and the extracted high-order characteristics are mapped into the RKHS for difference measurement. In order to further reduce the difference of the state feature probability distribution of each machine type gas circuit, a maximized field confusion method based on a countermeasure mechanism is introduced, and the maximum field confusion loss is realized by connecting a field confusion layer behind a plurality of field feature adaptation layers, so that the feature distributions learned among different fields are as close as possible, the learned feature distribution sources under the current state of the target engine are confused, the field invariant features are deeply excavated, and the optimal depth field self-adaption effect is achieved. The method successfully solves the problems that the gas circuit parameter deviation value model is lack of universality and the usable information amount is deficient due to the fact that the working condition environment of the civil aviation engine is complex and the models are various, and the gas circuit parameter deviation value model is built under the conditions of different types and different types by mutually transferring the association knowledge among the learned flight parameters under different types, so that the monitoring autonomy of the engine is obtained.
Drawings
FIG. 1 is an overall flowchart of a first embodiment of the present application;
FIG. 2 is a schematic diagram of maximum area confusion according to an embodiment of the present application;
FIG. 3 is a general framework of a Res-BPNN-based depth-domain adaptive regression model according to an embodiment of the present invention;
FIG. 4 is a diagram of the variation of MK-MMD and loss values with iteration times in the network training process according to the second embodiment of the present application;
FIG. 5 is a diagram of the effect of the DEGT regression (unsupervised method) in migration task A according to the second embodiment of the present application;
fig. 6 is a diagram of the DEGT regression effect of the migration task a according to the second embodiment of the present application (supervised method).
Detailed Description
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, and not all of the embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present application without making any creative effort belong to the protection scope of the present application.
In order to better understand the technical solutions, the technical solutions of the present application are described in detail below with reference to the drawings and specific embodiments, and it should be understood that the specific features in the embodiments and examples of the present application are detailed descriptions of the technical solutions of the present application, and are not limitations of the technical solutions of the present application, and the technical features in the embodiments and examples of the present application may be combined with each other without conflict.
Example one
A flowchart of a method for obtaining a deviation value of a gas path parameter according to this embodiment is shown in fig. 1,
the method comprises 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; usually divided by 80% training set and 20% testing set;
step 2, carrying out normalization pretreatment on the data of the training set and the test set of the source domain and the target domain engine; in the embodiment, the ACARS data acquired from the civil aviation company can be adopted to directly carry out sample normalization;
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 model and a regression module; the method specifically comprises the following steps:
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;
3.2, constructing a depth field self-adaptive module, and stacking a plurality of layers of multi-core mean difference adaptation layer measurement source domains and distribution differences between target domain engines;
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;
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 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.
In step 2, Z-score normalization pretreatment is adopted.
In step 3.2, in order to fully mine the domain invariant features and achieve the optimal depth domain self-adaptive effect, a domain confrontation value is introduced, and the domain confusion maximization is achieved by connecting a domain discriminator and a gradient inversion layer after the multi-core mean difference adaptation layer.
The characteristic extraction module realizes gas circuit state characteristic extraction by stacking three Res-BP residual error learning blocks, and the characteristic extraction module has 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.
The depth field self-adaptive module consists of a field confusion layer and three MK-MMD field adaptation layers; connecting the three MK-MMD domain adaptation layers with a feature extraction module, and measuring the difference of the extracted high-order feature distribution in the source domain and the target domain by using MK-MMD; the domain obfuscation layer is connected to the MK-MMD domain adaptation layer and includes a binary domain classifier, the output of which is a domain tag value.
The regression module is connected with the domain confusion layer, mapping between the extracted domain invariant features and corresponding gas circuit parameter deviation values is mainly achieved, and deviation value mining is finally achieved.
The idea of the deep domain adaptive method is to match the feature distribution between the source domain and the target domain to learn the domain invariant features. By learning domain-invariant features, it is meant that features, whether learned from source domain data or target domain data, should follow the same or nearly the same feature distribution. If the characteristic has the domain invariant characteristic, the characteristic can be used for effectively realizing the gas path parameter deviation value mining task of the target domain data, so that the learning of the domain invariant characteristic is the key point for realizing knowledge migration and reuse. The depth adaptive problem is further described below from the inter-domain difference metric method and the domain confusion method, respectively.
The Maximum Mean Difference (MMD) is mainly determined by mapping two different distributions into the regenerative nuclear Hilbert Space (RKHS) and measuring the difference between the two distributions.
First, a set of samples belonging to an s-distribution is definedAnd a set of samples belonging to a t distributionA set of continuous functions F in sample space is defined, whose continuous mapping F χ → R, the MMD between the two distributions, as shown in equation (3-1).
In the formula: es[·]-a mathematical expectation representing the distribution s;
Et[·]the mathematical expectation representing the distribution t.
Assume a sample set XSAnd XTThe two sample sets are obtained by independent equal distribution sampling from the distributions s and t, and the sample capacities are m and n, respectively. Based on XSAnd XTAn empirical estimate of MMD can be obtained as shown in equation (3-2).
Second, definition H denotes the RKS space, and F is constrained to be a unit sphere within the RKS space. The continuous mapping f χ → R, for any x ∈ χ, the function f (x) is shown in formula (3-3).
In the formula:-representing the mapping: χ → H, and f andthe inner product of (b) is defined as a kernel function, such as a gaussian kernel function:
therefore, by the above two definitions, it can be seen that the MMD of the s distribution and the t distribution in the regenerative nuclear hilbert space is as shown in the formula (3-4).
In the formula: mu.ss=Es[f(XS)],μt=Et[f(XT)]。
Finally, for ease of calculation, the square form of MMD is usually used, as in equation (3-5).
In the formula: m, n-respectively represent the source domain data set XSAnd target domain dataset XTSample size of (2). When X isSAnd XSThe smaller the distribution difference of (c), the smaller the MMD distance, and if and only if s and t follow a uniform distribution, MMD is equal to 0.
The Deep Adaptation network (Deep Adaptation Networks) can adopt an MMD variant algorithm, namely Multi-Kernel Maximum Mean difference (MK-MMD), and can better realize field Adaptation and enhance the feature representation capability of the neural network.
First, given the definition of MK-MMD, the empirical estimate of the MK-MMD distance between distribution s and distribution t is shown in equation (3-6).
In the formula: hk-represents RKHS with a characteristic nucleus k;
f (-) represents a continuous mapping function;
Es[·]-a mathematical expectation representing the distribution s;
Et[·]the mathematical expectation representing the distribution t.
Characteristic kernel f (·), k (x, y) ═ in MK-MMD<f(x),f(y)>Multi-core is specifically defined as having m semi-positive cores kuThe convex combination of (3) is shown in the formula (3-7).
In the formula:the constraint coefficient is used to ensure that the unique characteristic of the multi-kernel k exists when the domain adaptation is carried out among different distributions.
Different feature distributions change during learning of the deep neural network, so that the kernel function which can show stronger mapping capability cannot be determined, and the multi-kernel k based on the MK-MMD can enhance the adaptability among the feature distributions through the different kernel functions to achieve optimal and most reasonable kernel function selection.
In order to further reduce the edge distribution difference between the features learned from different domains, the embodiment introduces a domain confusion method based on a countermeasure mechanism, and a schematic diagram of the domain confusion method is shown in fig. 2.
In order to achieve maximum domain confusion, a domain confusion layer is added after the extracted depth feature layer, that is, whether a sample feature distribution after training is from a source domain or a target domain is determined. The more the extracted features can reflect the commonality among the fields, the better the field confusion effect is. If the depth feature extracted by adopting the Res-BP neural network cannot be distinguished by the trained domain classifier as a sample from a source domain or a target domain, the depth feature is called to have a domain invariant characteristic. The best-effort domain classifier D can be obtained by optimizing equation (3-8).
In the formula:
m represents the training sample batch size;
gi-a real domain label, g, representing the ith sample i0 stands for sample xiFrom the source domain, gi1 stands for sample xiIs derived from the target domain;
D(xi) -representative sample xiAnd outputting the label value after passing through the domain classifier.
The Res-BPNN-based depth-domain adaptive regression model provided by the embodiment is composed of a feature extraction module, a domain adaptive module and a regression module. And the high-order characteristic extraction module adopts a Res-BPNN network structure and performs gas circuit state characteristic extraction by stacking N Res-BP residual error learning blocks. And connecting the depth field self-adaptive module with the feature extraction module so as to learn the domain invariant features. And finally, adding a regression module to form a Res-BPNN-based depth domain adaptive regression model framework, as shown in FIG. 3.
(1) A feature extraction module: and the gas circuit state feature extraction is realized by stacking three Res-BP residual error learning blocks, and the feature extraction modules comprise ten layers and comprise an 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 firstly normalized in batch, and then each neuron is activated through a SELU activation function.
(2) Depth domain adaptive module: the depth domain adaptive module consists of a domain confusion layer and three MK-MMD domain adaptation layers. And connecting the three MK-MMD domain adaptation layers with the feature extraction module, and measuring the difference of the extracted high-order feature distribution in the source domain and the target domain by using the MK-MMD. The domain obfuscation 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.
(3) A regression module: the regression module is connected with the domain confusion layer, mapping between the extracted domain invariant features and corresponding gas circuit parameter deviation values is mainly achieved, and deviation value mining is finally achieved.
At present, an airline company introduces a new model of civil aircraft engine, only has a small batch of labeled data of the model, and in order to establish a gas circuit parameter deviation value model of the model of civil aircraft engine in a short time, labeled data resources need to be fully utilized. In order to facilitate comparison with other transfer learning algorithms, the proposed method is divided into a proposed method (supervised) and a proposed method (unsupervised) according to whether regression loss of labeled data of the target domain is introduced into the final optimization target.
Therefore, the Res-BPNN-based depth-domain adaptive regression model proposed in this embodiment finally has the following three optimization objectives:
(1) in order to realize the excavation of the deviation value of the gas circuit parameter of the aeroengine, the gas circuit state feature extraction and the excavation of the deviation value must be realized on the basis of the Res-BPNN depth field self-adaptive regression model. Therefore, the first optimization goal of the depth-domain adaptive regression model is to minimize the regression error of the gas path parameter deviation values on the source domain data set and the target domain data set, and the loss function of the regression error is defined as the standard MSE loss function, as shown in formula (3-9).
In the formula:
nS、nT-representing the amount of source domain data and target domain data in a batch of training sets;
yS、yTthe real value of the gas path parameter deviation value of the representative source domain sample and the target domain sample;
the regression value of the gas circuit parameter deviation value of the representative source domain sample and the target domain sample;
beta-represents the target domain labeled data training weight. If and only if β is 0, no tagged set of data is introduced into the target domain in the final optimization target.
(2) The deep domain adaptation module is used for learning domain invariant features, and mainly comprises a domain confusion layer and a multi-layer MK-MMD domain adaptation layer as shown in FIG. 3-2. The MK-MMD domain adaptation layer is mainly used for measuring distribution difference between characteristics learned by different fields. Thus, a second optimization goal of the depth-domain adaptive regression model is to minimize the distribution difference between the source and target domains. And measuring the distribution difference of high-order characteristics between the source domain and the target domain by using MK-MMD, wherein an MK-MMD loss function is shown in a formula (3-10).
In the formula:
-source domain and target domain feature representations representing the i-th adaptation layer, respectively;
XS、XT-representing the source domain dataset and the target domain dataset, respectively;
li-represents the i-th adaptation layer;
Hk-represents RKHS with characteristic core k.
(3) In order to further reduce the difference of the characteristic probability distribution of the gas circuit state of each machine type, a maximum domain confusion method of a countermeasure mechanism is adopted, so that the learned characteristic distribution among different domains is close as possible, the learned characteristic distribution source of the target engine in the current state is confused, and the domain invariant characteristics are deeply excavated. Therefore, the third optimization goal of the depth-domain adaptive regression model is to maximize the domain confusion penalty on the source and target domains, the domain classification penalty function is shown in equations (3-11).
By combining MSE regression loss, MK-MMD loss and domain confusion loss, the final optimization objective of the Res-BPNN-based deep domain adaptive regression model can be obtained as shown in the formulas (3-12).
L=Lreg+λLMK-MMD+μLdomain (3-12)
In the formula:
LMK-MMD-representing a domain adaptation loss;
Lreg-represents a regression loss;
Ldomain-representing a domain confusion loss;
lambda and mu represent network hyper-parameters and are used for controlling the strength of the adaptivity of the depth field;
when the model training is finished, if the learned high-order features have fuzzy domain types and small inter-domain differences, the gas circuit parameter deviation value regression module can accurately mine the deviation value of the target domain sample.
Example two
In the second embodiment, the technical scheme of the first embodiment is verified through experiments by using historical cruise data of the civil aircraft engine. The following is a detailed description of data sampling and preprocessing, hyper-parameter setting, and model performance comparison.
In order to fully verify the example application of the Res-BPNN-based depth-domain adaptive regression model in the civil aviation engine gas path parameter deviation value mining field and the general applicability of the model, two groups of data sets are respectively obtained from two civil aviation engines of different models, namely CFM56-5B2/3 and CFM56-7B26, which are produced by GE company, and are respectively applied to two gas path parameter deviation value mining experiments. The two groups of migration learning tasks are migration task A: CFM56-5B2/3 → CFM56-7B26 and migration task B: CFM56-7B26 → CFM56-5B 2/3.
Migration task A (CFM56-5B2/3 → CFM56-7B26) represents: the method comprises the steps of firstly carrying out supervised training through collected labeled data of a CFM56-5B2 model civil aircraft engine, and simultaneously assisting small-batch data acquired from the CFM56-7B26 model civil aircraft engine to carry out domain adaptive knowledge transfer. The migration task B (CFM56-7B26 → CFM56-5B2/3) exchanges the source domain data set and the target domain data set in the migration task CFM56-5B2/3 → CFM56-7B26 and carries out the migration experiment again, thereby verifying the effectiveness of the model. The partial data and data set assignment ratios for the two sets of transfer learning experiments are shown in tables 3-1 and 3-2, for example.
As shown in Table 3-1, the historical cruise data of the CFM56-5B2/3 and the CFM56-7B26 of two different types of civil aircraft engines used in the embodiment also need to eliminate the dimensional influence between the parameters, so that the historical cruise data needs to be subjected to standardization preprocessing, and a Z-score standardization preprocessing method is adopted so as to enable the sample parameters to be comparable and improve the network training rate, wherein the conversion formula is shown as a formula (2-9).
TABLE 3-1 data for two different engine models
TABLE 3-2 data set Allocation for two different sets of migratory learning tasks
And (3) corresponding relation between the input and the output of the gas circuit parameter deviation value regression model, as shown in the formula (3-13).
In the formula:
ID-[·]representing the corresponding input of each gas circuit parameter deviation value;
OD--representing the output of the deviation values of the gas path parameters;
xi-a measurement value representing the ith flight parameter.
The optimization method adopts a random gradient descent algorithm SGD to define thetaf,θd,θregWhich are the optimization parameters of the feature extractor, the domain classifier, and the final regressor, respectively, the equations (3-12) can be rewritten to the equations (3-14).
Based on the SGD algorithm and the formula (3-14), the parameter thetaf,θd,θregThe update process can be written as equation (3-15).
In the formula: alpha-represents the neural network learning rate, which can be self-adjusted by the formula (3-16).
In the formula:
the epoch and the epochs respectively represent the number of times that the network finishes training and the number of times that the network needs to finish training;
α0β, δ — represent constants, taken here as 0.01, 0.75 and 10, respectively;
t-represents the network training progress, and linearly changes from 0 to 1.
The self-tuning process of the hyper-parameters λ and μ in the formula (3-13) is similar to the self-tuning process of the learning rate, and the method makes the domain aliasing layer and the domain adaptation layer less sensitive to noise signals in the early stage of the training process, and the self-tuning formula is shown in the formula (3-17).
In the formula: γ -represents a constant, here taken to be 10.
The model optimizer adopts SGD, the size of a model training Batch (Batch-size) is set to be 100, the number of iterations (epoch) is set to be 1000, the Momentum ratio (Momentum) is set to be 0.9, an SELU is adopted as an activation function, and initial weight and bias are default values. In this embodiment, the feature extraction module is formed by stacking three Res-BP residual error learning modules, and the depth field adaptive module is formed by three MK-MMD field adaptation layers and one field confusion layer, so as to fully mine the field invariant features between the source field and the target field.
In order to further prove the effectiveness of the Res-BPNN-based depth-domain adaptive regression model provided in the first embodiment, nine different migration learning algorithms are adopted to respectively establish three key gas circuit parameter deviation value regression models for the migration task a and the migration task B, and the three key gas circuit parameter deviation value regression models are compared in experiments. The comparison methods used are shown in table 3-3, and migration Component Analysis (TCA), Joint Distribution Adaptation (JDA), Balanced Distribution Adaptation (BDA), Deep Domain Confusion Network (DDC), Deep Adaptation Network (DAN), Domain-adaptive Neural Network (DANN), and the method proposed in the first embodiment of the present application (unsupervised) are all unsupervised migration learning algorithms. In consideration of the actual situation that a new type of civil aviation engine is introduced into the current civil aviation company, in order to fully utilize the existing small batch of labeled target domain data set resources, regression errors of the labeled target domain data set resources can be merged into a final optimization target, and the regression errors are compared with experiments of supervised transfer learning such as a Res-BPNN model, Fine-tuning layer transfer (Fine-tuning) and the like.
TABLE 3-3 migration learning method
a.No target domain data is introduced into the final optimization target,b.the final optimization objective incorporates objective domain data.
And (3) respectively establishing three key gas circuit parameter deviation value regression models of DEGT, DFF and DN2 by adopting the supervised and unsupervised learning methods for experimental verification, comparing the regression models with the gas circuit parameter deviation value provided by the OEM, and respectively showing the performance comparison results of the migration task A and the migration task B in tables 3-4-3-6 and tables 3-7-3-9.
TABLE 3-4 comparison of DEGT regression effect for migration task A
TABLE 3-5 comparison of DFF regression effects for migration task A
Table 3-6 comparison of DN2 regression effects for migration task A
Tables 3-7 comparison of the effects of the DEGT regression for migration task B
Tables 3-8 comparison of DFF regression effects for migration task B
Tables 3-9 DN2 regression effectiveness comparison of migration task B
As shown in tables 3-4 to 3-9, the fitting results obtained by the Res-BPNN-based deep domain adaptive regression model in the prediction of the three key performance deviation values DEGT, DFF and DN2 in the migration task A and the migration task B are the most prominent among all the methods.
(1) By analyzing and comparing the experimental results of the two groups of unsupervised algorithms of TCA, JDA and BDA, DDC, DAN, DANN and the proposed method (unsupervised), the migration effect of the experimental results of the second group is found to be better than that of the first group. The first group of three migration algorithms need to select the first m characteristic values in the characteristic dimension reduction process, and useful information among flight parameters can be lost in the middle process, so that the migration effect is poor. And the four migration algorithms of the second group adopt a deep network to extract depth features, so that the accurate mapping relation between the gas circuit parameters and the deviation values thereof can be represented better, and the difference between the fields is reduced fully.
(2) Through analysis and comparison of four unsupervised algorithms DDC, DAN and DANN and the proposed algorithm (unsupervised), the optimal regression effect of the Res-BPNN-based depth adaptive regression model provided by the application can be found. The DDC learns the domain invariant features by introducing an MMD domain adaptation layer into a Res-BPNN regression model architecture. The DAN can better learn the domain invariant features on the basis of DDC by introducing three MK-MMD domain adaptation layers without increasing additional network training time. DANN can deeply extract common characteristics between a source domain and a target domain by introducing countermeasure thought in a Res-BPNN regression model architecture. The method provided by the application can effectively reduce the difference of feature distribution among fields by combining the advantages of DANN and DAN and additionally introducing three MK-MMD domain adaptation layers on the basis of an antagonistic mechanism.
(3) Through analysis and comparison of the Res-BPNN regression model and the layer migration micro-tuning and extraction method (with supervision) of the three supervised learning algorithms, the result shows that the depth self-adaptive regression model based on Res-BPNN has the same optimal regression effect. Because the label samples of the aero-engine with the new model are few, the Res-BPNN regression model is directly adopted for training, and the problems of poor deep network training effect, overfitting of the model and poor regression precision are inevitably caused. And (3) learning high-order characteristics by adopting a layer migration fine tuning method, pre-training a source domain regression model, freezing the parameter weights of the former n layers of networks, fine tuning the latter several layers of parameters by adopting small batches of labeled target domain data, and realizing the establishment of an offset value regression model. According to the depth domain self-adaptive regression model framework based on Res-BPNN, the target domain small-batch labeled data set regression loss is introduced and adjusted in the final optimization target, the unsupervised learning problem is converted into the supervised learning problem, domain invariant features between a source domain and a target domain are fully mined, the domain feature distribution difference is further reduced, and the migration effect is remarkably improved.
In order to intuitively embody the superiority of the Res-BPNN-based depth-domain adaptive regression model provided by the present application, the regression experiment result of the DEGT in the migration task a is selected as follows, and the changes of MK-MMD and the training loss with the iteration number are shown in fig. 4. Since there are many unsupervised comparative experimental methods, only 50 test points are plotted here, as shown in fig. 5, in order to clearly embody the fitting result of the method proposed in the present application (unsupervised). The results of the supervised comparative experiment were plotted for 100 test points, as shown in fig. 6.
As can be seen from fig. 4, as the number of network iterations increases, the MK-MMD distribution difference metric value and the network training loss value decrease gradually and tend to converge to some extent. It is also apparent from fig. 5 and fig. 6 that the Res-BPNN-based deep-field adaptive regression model provided by the present application can obtain a better regression effect in an unsupervised learning algorithm or a supervised learning algorithm, thereby proving the effectiveness of the method provided by the present application.
In conclusion, the invention provides a cross-working condition and cross-machine type regression model for establishing the gas circuit parameter deviation value of the civil aviation engine. Aiming at the actual mapping relation between the gas circuit parameters of the civil aviation engine and the deviation values thereof, the gas circuit state characteristics are extracted through N Res-BP residual error learning modules, the common characteristics between the source domain and the target domain are deeply excavated by adopting a plurality of MK-MMD domain adaptation layers and domain confusion layers, and finally the deviation value regression is realized to form a Res-BPNN-based depth domain adaptive regression model. In consideration of the actual condition that a civil aviation company introduces a new model, in order to fully utilize the small-batch labeled data resources of the target domain, the regression loss of the target domain is introduced into the final optimization target of the regression model. In order to verify the effectiveness of the model, whether the regression model introduces labeled target domain data is divided into two methods, namely a supervised method and an unsupervised method, and the two methods are compared with other algorithms in an experiment. The result shows that the domain invariant feature between the source domain and the target domain can be more fully mined by adopting the method of the domain adaptation and the domain confusion combined optimization, the distribution difference of the domain features is further reduced, and the regression effect is improved.
In the steps of the method, the processed data are determined to be the aircraft engine ACARS data and how the ACARS data are processed in each step, and the method shows that the deep field adaptive regression model training method is closely related to the processing of the aircraft engine ACARS data. The invention provides a solution to solve the field self-adaptive problem in the field of aeroengine health management, namely how to mutually transfer the correlation knowledge among the learned flight parameters under different types of engines, realize the cross-working condition and cross-machine type establishment of a gas circuit parameter deviation value model, and further obtain the technical problem of engine monitoring autonomy, adopts N Res-BP residual error learning modules to extract gas circuit state characteristics, adopts a plurality of MK-MMD (MK-MMD) domain adaptation layers and domain confusion layers to deeply mine the common characteristics between a deviation value source domain and a target domain, and finally realizes the technical means of regression, utilizes the technical means following the natural law, can more fully mine the domain invariant characteristics between a source domain and a target domain civil aeroengine, further reduces the distribution difference of the domain characteristics, and realizes the method for obtaining the gas circuit parameter deviation value under the cross-working condition and cross-machine type, compared with the traditional modeling method, the regression effect is obviously improved.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or 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 (including, but not limited to, disk storage, optical storage, and the like) having computer-usable program code embodied therein.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present application without departing from the spirit and scope of the application. Thus, if such 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 intended to include such modifications and variations as well.
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.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010418279.0A CN111597760B (en) | 2020-05-18 | 2020-05-18 | Method for obtaining gas path parameter deviation value under small sample condition |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010418279.0A CN111597760B (en) | 2020-05-18 | 2020-05-18 | Method for obtaining gas path parameter deviation value under small sample condition |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111597760A CN111597760A (en) | 2020-08-28 |
CN111597760B true CN111597760B (en) | 2022-07-22 |
Family
ID=72191345
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010418279.0A Active CN111597760B (en) | 2020-05-18 | 2020-05-18 | Method for obtaining gas path parameter deviation value under small sample condition |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111597760B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112183626A (en) * | 2020-09-28 | 2021-01-05 | 上海明略人工智能(集团)有限公司 | Power grid stability judging method and system and computer equipment |
CN113420775B (en) * | 2021-03-31 | 2024-03-29 | 中国矿业大学 | Picture classification method based on nonlinearity self-adaptive subdomain field adaptation under extremely small quantity of training samples |
CN112989712B (en) * | 2021-04-27 | 2021-08-03 | 浙大城市学院 | Aeroengine fault diagnosis method based on 5G edge calculation and deep learning |
CN113506253B (en) * | 2021-06-30 | 2023-04-11 | 东北石油大学 | Sandstone pore detection method based on multilayer multi-core learning and region combination |
CN116128049B (en) * | 2023-04-04 | 2023-06-30 | 厦门大学 | XGBoost model-based migration condition selection method for water quality prediction model |
CN117034732A (en) * | 2023-04-14 | 2023-11-10 | 北京百度网讯科技有限公司 | Automatic driving model training method based on true and simulated countermeasure learning |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109766921A (en) * | 2018-12-19 | 2019-05-17 | 合肥工业大学 | A kind of vibration data Fault Classification based on depth domain-adaptive |
CN110909783A (en) * | 2019-11-15 | 2020-03-24 | 重庆邮电大学 | Blind domain image classification and reconstruction method based on enhanced reconstruction classification network |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109710636B (en) * | 2018-11-13 | 2022-10-21 | 广东工业大学 | Unsupervised industrial system anomaly detection method based on deep transfer learning |
-
2020
- 2020-05-18 CN CN202010418279.0A patent/CN111597760B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109766921A (en) * | 2018-12-19 | 2019-05-17 | 合肥工业大学 | A kind of vibration data Fault Classification based on depth domain-adaptive |
CN110909783A (en) * | 2019-11-15 | 2020-03-24 | 重庆邮电大学 | Blind domain image classification and reconstruction method based on enhanced reconstruction classification network |
Non-Patent Citations (2)
Title |
---|
【深度域适配】一、DANN与梯度反转层(GRL)详解;AI那点小事;《腾讯云社区》;20200415;全文 * |
Regression Model for Civil Aero-engine Gas Path Parameter Deviations Based on Res-BP Neural Network;Xingjie Zhou .etc;《2019 International Conference on Sensing, Diagnostics, Prognostics, and Control (SDPC)》;20191231;P188-196 * |
Also Published As
Publication number | Publication date |
---|---|
CN111597760A (en) | 2020-08-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111597760B (en) | Method for obtaining gas path parameter deviation value under small sample condition | |
Wang et al. | Combination of hyperband and Bayesian optimization for hyperparameter optimization in deep learning | |
CN112784881B (en) | Network abnormal flow detection method, model and system | |
CN114997051A (en) | Aero-engine service life prediction and health assessment method based on transfer learning | |
Fan et al. | A novel machine learning method based approach for Li-ion battery prognostic and health management | |
CN110609524B (en) | Industrial equipment residual life prediction model and construction method and application thereof | |
CN108664690A (en) | Long-life electron device reliability lifetime estimation method under more stress based on depth belief network | |
CN112784920A (en) | Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part | |
CN111914094A (en) | Knowledge graph representation learning method based on ternary interaction | |
CN111222847A (en) | Open-source community developer recommendation method based on deep learning and unsupervised clustering | |
CN110082738B (en) | Radar target identification method based on Gaussian mixture and tensor recurrent neural network | |
Wang et al. | A remaining useful life prediction model based on hybrid long-short sequences for engines | |
CN111340069A (en) | Incomplete data fine modeling and missing value filling method based on alternate learning | |
Oliinyk et al. | Synthesis of neuro-fuzzy networks on the basis of association rules | |
CN117313795A (en) | Intelligent building energy consumption prediction method based on improved DBO-LSTM | |
CN112613227B (en) | Model for predicting remaining service life of aero-engine based on hybrid machine learning | |
Subbotin | The sample properties evaluation for pattern recognition and intelligent diagnosis | |
Voronov et al. | Lead-acid battery maintenance using multilayer perceptron models | |
CN115375038A (en) | Aircraft engine failure mode identification and service life prediction method | |
Feng et al. | Dme: an adaptive and just-in-time weighted ensemble learning method for classifying block-based concept drift steam | |
Jing | Neural Network-based Pattern Recognition in the Framework of Edge Computing | |
CN118070689B (en) | Unmanned ship cluster game algorithm efficiency evaluation method and system based on model and data hybrid driving | |
CN114881073B (en) | Sectional fault diagnosis method and system for variable working condition rolling bearing | |
CN118395363B (en) | Active power distribution network abnormal state sensing method and system based on data enhancement | |
CN118501648B (en) | Semiconductor test control method and system based on big data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |