An engine group fault mode recognition method and device based on density clustering-support vector machine and multi-time point prediction weighting, and electronic equipment

Abstract

The application discloses an engine group fault mode recognition method and device based on density clustering-support vector machine and multi-time point prediction weighting and electronic equipment, and the method comprises the following steps: performing dimension reduction processing on training engine performance parameters and test engine performance parameters respectively based on principal component analysis (PCA), to obtain training engine low-dimensional performance parameters and test engine low-dimensional performance parameters; in the prediction of engine single-time point fault mode, performing fault mode clustering processing on the training engine low-dimensional performance parameters based on a density-based spatial clustering of applications with noise (DBSCAN) algorithm, to obtain corresponding fault mode classification labels of each training engine; a single-time point fault mode classification model based on a support vector machine (SVM) is established, and the single-time point fault mode classification model based on the SVM is trained by using the training engine low-dimensional performance parameters and the corresponding fault mode classification labels of each training engine, to obtain a trained single-time point fault mode classification model based on the SVM; the test engine low-dimensional performance parameters are input into the trained single-time point fault mode classification model based on the SVM, to obtain single-time point fault mode classification labels of each test engine; and finally, based on multi-time point fault mode weighted prediction and the single-time point fault mode classification labels of each test engine, final fault mode classification labels of each test engine are obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault identification, and in particular relates to an engine group fault mode identification method and device based on density clustering-support vector machine and multi-time point prediction weighting and electronic equipment. BACKGROUND

[0002] An aero-engine is a key part of an airplane, usually working in harsh environmental conditions such as high pressure and high speed rotation, and the operation profile and failure result are complex and changeable. Long-term service process will cause performance degradation of components and even the whole machine. If the engine health state cannot be evaluated in time and the remaining life of the engine is accurately predicted in advance, the reliability and safety of the airplane will be affected.

[0003] Because the operation profile and the fault mode of the aero-engine are interlinked, the aero-engine will experience multiple different operation conditions before a fault mode occurs, and under the combination of different operation conditions, the engine degradation path and the fault mode will change. At the same time, the operation condition and the fault mode have a great influence on the error of the degradation path analysis and the life prediction, and the identification of the operation condition is beneficial to improve the accuracy and applicability of the prediction model. Therefore, the operation condition and the fault mode are difficult to identify under complex operation conditions. SUMMARY

[0004] The present application provides an engine group fault mode identification method and device based on density clustering-support vector machine and multi-time point prediction weighting, in order to solve the technical problem of difficulty in identifying the operation condition and the fault mode under complex operation conditions.

[0005] The present application provides an engine group fault mode identification method based on density clustering-support vector machine and multi-time point prediction weighting, comprising:

[0006] The principal component analysis (PCA) is used to reduce the dimension of the training engine performance parameters and the test engine performance parameters respectively, to obtain low-dimensional performance parameters of the training engine and low-dimensional performance parameters of the test engine;

[0007] In the prediction of the single-time point fault mode of the engine, the density-based spatial clustering of applications with noise (DBSCAN) algorithm is used to cluster the low-dimensional performance parameters of the training engine, to obtain the fault mode classification labels corresponding to each training engine;

[0008] A single-time point fault mode classification model based on support vector machine (SVM) is established, and the low-dimensional performance parameters of the training engine and the fault mode classification labels corresponding to each training engine are used to train the single-time point fault mode classification model based on SVM, to obtain the trained single-time point fault mode classification model based on SVM;

[0009] The low-dimensional performance parameter of the test engine is input into the trained single-time fault mode classification model based on SVM to obtain single-time fault mode classification labels of each test engine.

[0010] Based on the multi-time fault mode weighted prediction and the single-time fault mode classification labels of each test engine, final fault mode classification labels of each test engine are obtained.

[0011] The embodiment of the application further provides an engine group fault mode recognition device based on density clustering-support vector machine and multi-time prediction weighting, comprising:

[0012] The parameter dimension reduction module is used for performing dimension reduction processing on the performance parameters of the training engine and the performance parameters of the test engine based on principal component analysis (PCA) respectively to obtain low-dimensional performance parameters of the training engine and low-dimensional performance parameters of the test engine.

[0013] The fault mode classification module is used for performing fault mode clustering processing on the low-dimensional performance parameters of the training engine based on a density-based spatial clustering of applications with noise (DBSCAN) algorithm in the prediction of single-time fault modes of the engine to obtain fault mode classification labels corresponding to each training engine.

[0014] The model training module is used for establishing a single-time fault mode classification model based on support vector machine (SVM) and training the single-time fault mode classification model based on SVM by using the low-dimensional performance parameters of the training engine and the fault mode classification labels corresponding to each training engine to obtain a trained single-time fault mode classification model based on SVM.

[0015] The fault mode recognition module is used for inputting the low-dimensional performance parameters of the test engine into the trained single-time fault mode classification model based on SVM to obtain single-time fault mode classification labels of each test engine and obtaining final fault mode classification labels of each test engine based on multi-time fault mode weighted prediction and the single-time fault mode classification labels of each test engine.

[0016] The embodiment of the application further provides an electronic device comprising a memory, a processor and a computer program stored in the memory and capable of running on the processor; when the computer program is executed by the processor, the steps of the engine group fault mode recognition method based on density clustering-support vector machine and multi-time prediction weighting in any one of claims 1-8 are implemented.

[0017] The application has the beneficial effect that sensitive parameters related to engine degradation are extracted from multi-dimensional monitoring data, a mode combining metric learning and clustering algorithm is adopted, and operating conditions and fault modes are separated step by step, so that the accurate explanation of the engine operating state is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flow chart of an engine group fault mode recognition method based on density clustering-support vector machine and multi-time prediction weighting provided by the present application;

[0019] Figure 2 is a technical roadmap of an aero-engine life prediction method based on metric learning under complex operating conditions provided by the present application;

[0020] Figure 3 is a flow chart of a method for adaptively recognizing engine operating conditions by combining a clustering algorithm and metric learning;

[0021] Figure 4 is a flow chart of a method for equivalent transformation of variable operating condition data;

[0022] Figure 5 is a curve diagram illustrating a variable operating condition degradation condition;

[0023] Figure 6 is a flow chart of a fault mode recognition method based on variable weight prediction;

[0024] Figure 7 is a flow chart of a single-time fault mode classification model;

[0025] Figure 8 is a structure diagram of an LSTMCNN_AE network;

[0026] Figure 9 is a structure detail diagram of an encoder and a decoder in an LSTMCNN_AE network;

[0027] Figure 10 is a flow chart of an engine time series degradation feature extraction method based on LSTMCNN_AE;

[0028] Figure 11 is a schematic diagram of an LSTMCNN_AE model input sample generation process;

[0029] Figure 12 is a flow chart of a multi-fault mode and variable operating condition engine health assessment method based on NCA metric learning;

[0030] Figure 13 is a schematic diagram of a Siamese network structure;

[0031] Figure 14 is a structure schematic diagram of an SN-GRU network model;

[0032] Figure 15 is a schematic diagram of an SN-GRU network structure;

[0033] Figure 16is a flow chart of life prediction of an engine in a multi-fault mode based on an SN-GRU. DETAILED DESCRIPTION

[0034] It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application. In the following description, the suffixes "module", "part", or "unit" used for components are merely for convenience of description and have no specific meaning. Therefore, "module", "part", or "unit" can be mixedly used.

[0035] Figure 1 is a flow chart of an engine group fault mode recognition method provided by the present application based on density clustering-support vector machine and multi-time prediction weighting, as shown in the figure, comprising: S101: performing dimensionality reduction processing on training engine performance parameters and test engine performance parameters based on principal component analysis (PCA), to obtain training engine low-dimensional performance parameters and test engine low-dimensional performance parameters; S102: in the prediction of engine single-time fault mode, performing fault mode clustering processing on the training engine low-dimensional performance parameters based on a density clustering (DBSCAN) algorithm, to obtain fault mode classification labels corresponding to each training engine; S103: establishing a single-time fault mode classification model based on a support vector machine (SVM), and training the single-time fault mode classification model based on the SVM using the training engine low-dimensional performance parameters and the fault mode classification labels corresponding to each training engine, to obtain a trained single-time fault mode classification model based on the SVM; S104: inputting the test engine low-dimensional performance parameters into the trained single-time fault mode classification model based on the SVM, to obtain single-time fault mode classification labels of each test engine; and S105: obtaining final fault mode classification labels of each test engine based on multi-time fault mode weighted prediction and the single-time fault mode classification labels of each test engine. Figure 1

[0036] In the embodiment of the present application, the obtaining of the final fault mode classification labels of each test engine based on the multi-time fault mode weighted prediction and the single-time fault mode classification labels of each test engine comprises: obtaining, according to the trained single-time fault mode classification model based on the SVM, prediction results of each test engine in a fault mode at a single time; performing weighted prediction on the fault state of each test engine at a single time and the comprehensive multi-time fault state, to obtain weighted prediction results of each test engine in different fault modes; and obtaining the final fault mode classification labels of each test engine using the weighted prediction results of each test engine in different fault modes.

[0037] ​In the embodiment of the present application, the method comprises the following steps:

[0038]

[0039] In the embodiment of the present application, the method comprises the following steps:

[0040]

[0041] wherein, m represents a fault mode; j represents a test engine; τ represents a recent weighting coefficient of a recognition proportion of recent data; λ represents a comprehensive weighting coefficient of a flight cycle number considered comprehensively; k represents a value coefficient of an influence of a recent and long-term weight; and t-k represents a flight cycle distance from a most recent time point. represents a predicted state of the test engine j in the fault mode m at the time point t-k, represents a predicted state of the test engine j in the fault mode m at the time point t-k, represents a predicted result of the test engine j in the fault mode m obtained through the weighted prediction.

[0042] In the embodiment of the present application, the method comprises the following steps:

[0043]

[0044] wherein, Y j represents a final fault mode classification label of the test engine j.

[0045] In the embodiment of the present application, before the training engine performance parameters and the test engine performance parameters are respectively processed through the principal component analysis (PCA) for dimension reduction, the method further comprises the following steps: screening out performance parameters sensitive to degradation of the engine under variable operating conditions; extracting engine data under different operating conditions according to individual engine operating condition division results to obtain degradation-sensitive performance parameters under each operating condition; performing normalization processing on the degradation-sensitive performance parameters under each operating condition through a normalization method, and reconstructing the degradation-sensitive performance parameters under each operating condition after the normalization processing according to a flight cycle sequence of the engine to obtain degradation-sensitive performance parameters fused under multiple operating conditions; and dividing the degradation-sensitive performance parameters fused under multiple operating conditions into the training engine performance parameters and the test engine performance parameters.

[0046] In the embodiment of the present application, the performance parameters sensitive to degradation of the engine under variable operating conditions are screened, and the method further comprises: obtaining multi-dimensional operating parameters of the training engine under different operating conditions; preliminarily identifying the operating condition categories of the training engine by using a K-Means clustering algorithm according to the obtained multi-dimensional operating parameters, and obtaining the clustering centers of the operating conditions; scaling the multi-dimensional operating parameters of the test engine and the clustering centers of the corresponding operating conditions to one dimension by using a multi-dimensional scaling (MDS) algorithm, and taking the one-dimensional operating condition clustering centers as operating condition baselines; and adaptively identifying the operating conditions corresponding to each flight cycle of each test engine according to the principle of being closest to the operating condition baseline.

[0047] In the embodiment of the present application, the operating conditions corresponding to each flight cycle of each test engine are adaptively identified according to the principle of being closest to the operating condition baseline, which comprises: selecting the operating condition closest to the test engine as the operating condition of the test engine by measuring the distance between the operating parameters of the test engine at a single time and each operating condition baseline.

[0048] The present application also provides an engine fleet failure mode recognition device based on density clustering-support vector machine and multi-time prediction weighting, comprising: a parameter dimension reduction module, configured to perform dimension reduction processing on performance parameters of a training engine and performance parameters of a test engine respectively based on principal component analysis (PCA), to obtain low-dimensional performance parameters of the training engine and low-dimensional performance parameters of the test engine; a failure mode classification module, configured to perform failure mode clustering processing on the low-dimensional performance parameters of the training engine based on a density-based spatial clustering of applications with noise (DBSCAN) algorithm to obtain failure mode classification labels corresponding to each training engine in a single-time failure mode prediction of an engine; a model training module, configured to establish a single-time failure mode classification model based on a support vector machine (SVM), and train the single-time failure mode classification model based on the SVM by using the low-dimensional performance parameters of the training engine and the failure mode classification labels corresponding to each training engine, to obtain a trained single-time failure mode classification model based on the SVM; and a failure mode recognition module, configured to input the low-dimensional performance parameters of the test engine into the trained single-time failure mode classification model based on the SVM to obtain single-time failure mode classification labels of each test engine, and obtain final failure mode classification labels of each test engine based on multi-time failure mode weighted prediction and the single-time failure mode classification labels of each test engine.

[0049] The present application also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the engine fleet failure mode recognition method based on density clustering-support vector machine and multi-time prediction weighting are implemented.

[0050] There are many bottleneck technical problems to be solved for the life prediction of aero-engine under variable working conditions and multiple failure modes: difficulty in identifying running conditions and failure modes under complex operation conditions, difficulty in extracting time sequence degradation features under variable working conditions, difficulty in quantifying engine degradation trend under variable working conditions, and problem of reducing the accuracy of engine life prediction under multiple failure modes. In view of the above problems, a life prediction method is proposed, and the overall life prediction process and technical route please refer to Figure 2 , which specifically includes three parts: engine group failure mode recognition method based on density clustering-support vector machine and multi-time prediction weighting, variable working condition engine health assessment based on improved auto-encoder and metric learning, and multi-failure mode engine life prediction based on time sequence distribution metric and field generalization.

[0051] I. Engine group failure mode recognition method based on density clustering-support vector machine and multi-time prediction weighting

[0052] In view of the difficulty in identifying running conditions and failure modes under complex operation conditions, the present application starts from two aspects of identification order and identification method, and formulates the operation condition analysis process from individual engine to engine group. In terms of order, the flight condition in the running monitoring process of single engine is identified first, and then the failure mode of engine group is divided. In terms of method, since the running condition is the identification of single time of engine flight cycle, the combination of clustering method and metric learning can identify the category of operation parameters at a specific time, which provides a train of thought for the division of running condition. The performance degradation of engine is a gradual process. For the test engine which may be in the early stage of degradation, the fault feature is not obvious, and it is difficult to accurately distinguish through single time monitoring parameter, so the failure mode is the identification of multiple time evolution trend. For failure mode recognition, the density clustering method can identify different degradation trajectories, and the weighted fusion method provides a new train of thought for multi-time failure mode identification.

[0053] 1.1. Engine condition division based on combination of metric learning and clustering

[0054] In engine operating condition recognition, multi-dimensional operating parameters need to be clustered to obtain the operating condition classification result. In order to accurately determine the operating condition label at a single time, the application proposes an improved clustering method to identify the engine operating condition, that is, introducing the unsupervised metric learning MDS algorithm into the K-Means clustering method. The core of MDS is to calculate the distribution of data in low-dimensional space by using high-dimensional data, that is, to reconstruct the low-dimensional coordinates of the data, so as to efficiently measure the real distance between data. The MDS algorithm can project the multi-dimensional operating parameters into a low-dimensional space while keeping the relative position unchanged, and the test engine can automatically identify and belong to the nearest operating condition category, so as to adaptively realize the operating condition category division of the test engine at a single time. The MDS algorithm can be used to effectively calculate the relative distance between the high-dimensional operating condition data of the test set and the clustering center, and overcome the problem of inconsistent Euclidean distance scale, realize the operating condition category division of the test engine, and solve the problem of inaccurate operating condition clustering of the test set.

[0055] In engine operating condition recognition, the multi-dimensional operating parameters of the training engine are used as the data basis, and the K-Means clustering algorithm is used to preliminarily identify the operating condition categories of the training engine and obtain the clustering centers of each operating condition. Then, the MDS algorithm is used to scale the operating parameters of the test engine and the operating condition clustering centers to one dimension, and the operating conditions corresponding to each flight cycle of the test engine are adaptively identified by the principle of being closest to the operating condition baseline. The specific process of the engine operating condition adaptive identification based on the combination of metric learning and clustering is as shown in Figure 3

[0056] Briefly, the method of introducing the MDS algorithm into the K-Means clustering algorithm to adaptively realize the division of the engine operating condition category at a single time includes the following steps: S111, obtaining the multi-dimensional operating parameters of the training engine under different operating conditions; S112, preliminarily identifying the operating condition categories of the training engine by using the K-Means clustering algorithm according to the multi-dimensional operating parameters obtained in step S111, and obtaining the clustering centers of each operating condition; S113, scaling the multi-dimensional operating parameters of the test engine and the clustering centers of the corresponding operating conditions to one dimension by using the MDS algorithm; S114, adaptively identifying the operating conditions corresponding to each flight cycle of the test engine according to the principle of being closest to the operating condition baseline.

[0057] The pseudo code of the engine operating condition adaptive identification based on the combination of metric learning and clustering is as follows:

[0058]

[0059]

[0060] The process of the engine operating condition adaptive identification based on the combination of metric learning and clustering is as follows:

[0061] ​(101) Engine operating parameter extraction

[0062] Among the aero-engine data under complex operating conditions, including operating parameters and gas path parameters, the operating parameter data is extracted as the data basis for engine condition recognition. The continuous change of engine operating parameters indicates that the engine condition is constantly changing during operation. The combination of operating parameters brings different operating conditions to the aircraft. When the values of three operating parameters are basically the same and within a small fluctuation range, it can be indicated that the engine condition is the same. The operating condition of the aero-engine is usually randomly changed, and it is difficult to find its change rule, which also brings certain difficulty to the subsequent engine health assessment under variable operating conditions.

[0063] (102) K-Means-based training engine cluster center determination and condition label division

[0064] The divided training engine operating parameters are used as the multi-dimensional input of K-Means clustering, and k operating condition parameter combinations are randomly selected as the initial cluster centers according to the number k of aero-engine condition categories. In the training of the K-Means clustering model, the number of cluster centers and the maximum number of iterations for each training are set. After the training is completed, the model returns the optimal cluster center and outputs the operating condition label corresponding to each flight cycle of the training engine and the final cluster center result.

[0065] (103) Test engine condition adaptive recognition based on MDS metric

[0066] The operating parameters of the test engine are combined with the operating condition cluster centers, and the multi-dimensional scaling MDS algorithm is used to scale the multi-dimensional data to one dimension. With the one-dimensional operating condition center as the baseline, the single-time scaled operating data can automatically identify the baseline operating condition with the smallest distance, thereby adaptively selecting the operating condition and obtaining the operating condition of the test engine under each flight cycle.

[0067] 1.2, Engine data fusion based on operating condition equivalent transformation

[0068] The aero-engine gas path parameters can be used as an important basis for health assessment and life prediction, but the data of the gas path parameters fluctuate under the influence of the operating condition, making it difficult to represent the real degradation trend of the engine. Not all gas path parameters are useful for studying the engine state in the aero-engine gas path parameters, and some parameters provide constant or discrete values throughout the life cycle of the engine. In order to restore the real degradation trend of the engine under variable operating conditions, it is necessary to unify the performance parameters under different operating conditions to a certain comparable parameter standard. The present application proposes a variable operating condition engine data equivalent transformation method, which mainly consists of four parts: degradation performance parameter screening, multi-condition time series data decomposition, each operating condition performance parameter standardization, and performance parameter reconstruction. The process of variable operating condition data equivalent transformation is as follows Figure 4as shown.

[0069] (1001) Screening of sensitive parameters under variable conditions

[0070] In the screening of performance parameters, by drawing the engine-to-life curve under different conditions, the parameters that remain constant or have no obvious degradation trend are excluded, thus preliminarily screening the performance parameters that have obvious degradation trend, less fluctuation and are more sensitive to engine degradation under variable conditions, and forming the performance parameter sequence under variable conditions as shown in Figure 4

[0071] (1002) Decomposition of multi-condition time series data

[0072] In the decomposition of multi-condition time series data, based on the individual engine operating condition division result, engine data under different conditions is extracted to obtain the sensitive performance parameters under each condition, thus realizing the extraction of feature data under variable conditions. Based on the operating condition recognition label based on the improved clustering method, the engine time series data is decomposed based on the operating condition by matching the engine flight cycle with the operating condition. At this time, each gas path parameter of the engine can be divided into time discontinuous data under different conditions, but the gas path parameters show a relatively uniform degradation trend under the same condition, and the fluctuation is smaller than before the data decomposition, thus laying a foundation for parameter standardization and reconstruction.

[0073] (1003) Standardization of performance parameters under each condition

[0074] Aeroengines face frequent condition changes during operation. Under the same condition, the degradation parameters of engine performance usually show a gradual change trend, and the fluctuation is small. When the condition frequently switches, the degradation trend is difficult to show.

[0075] To realize the goal of parameter unification and degradation trend integration, two operations are needed: first, for engine parameters, convert them into a unified index that can reflect the degradation trend, scale the data under different conditions, and unify them to the standard space; second, for the degradation trend, the standardized data under different conditions are formed into a curve that can clearly characterize the degradation under variable conditions, as shown in Figure 5

[0076] The normalization method is used to normalize the sensitive performance parameters under different conditions, thus realizing the parameter normalization under different conditions. Let the sensitive performance parameter of the aeroengine at flight cycle t i be x i , the corresponding condition be m i , and the state parameter be (t i , x i , m i ​​), i = 1, 2,.., n. Engine operating conditions include k kinds, and the operating condition category set can be represented as M = {M1, M2,..., M k}, therefore, the total set of performance parameters under each condition can be represented as {X} (K) = {x i |m i ∈ M}.

[0077] z-score normalization establishes a standardized index by obtaining the mean and standard deviation in the data, which can transform the original data into data with a mean of 0 and a variance of 1. The mean and standard deviation of the parameters under different conditions can be represented as and s k , then the data normalization formula under condition k is:

[0078]

[0079] The z-score method takes the mean and standard deviation as the normalization reference, and has less influence on local fluctuations and noise. Given that the parameters of an aero-engine usually have large noise interference, this method can be used to deal with this problem more effectively.

[0080] (1004) Gas path performance parameter reconstruction

[0081] Finally, according to the flight cycle order of the engine, the normalized performance data is reconstructed to combine the normalized feature parameters under different operating conditions, and the degradation time series of multiple conditions are reconstructed to represent the degradation of the engine under variable conditions. The reconstructed parameter data can be represented as:

[0082] X' = {x1', x'2,..., x' n} (6)

[0083] Where X' is the reconstructed aero-engine aerodynamic performance degradation parameter under variable conditions, x t ' is the N-dimensional performance parameter under flight cycle t, t = 1, 2,..., n.

[0084] 1.3, Engine fault mode recognition based on classification and variable weight prediction

[0085] In the fault mode recognition, the fault mode needs to be divided according to the degradation path of the engine, considering that the engine parameters have certain volatility, K-Means does not have the inclusiveness of noise data, and the similarity mean does not have actual meaning for trajectory clustering, which will cause the problem of performance degradation trajectory recognition and classification of the subsequent test set data. The density-based clustering algorithm is not disturbed by the up and down floating of parameters, and is suitable for fault mode recognition. Therefore, a prediction weighted fault mode discrimination method based on density clustering-support vector machine (DBSCAN-SVM) and multi-time is designed, the engine fault mode is recognized by the density clustering method, the single-time fault state classification of the test unit is judged by using SVM, and finally the fault mode of the test unit is judged by weighting the multi-time fault state.

[0086] In the fault mode recognition method based on variable weight prediction, first, PCA dimension reduction is carried out based on the screened air path performance degradation parameters, and low-dimensional features are obtained for training the clustering of the accurate fault mode of the engine; then, a single-time fault mode classification model based on SVM is established based on density clustering, and a fault mode weighted prediction method is designed to complete the fault mode recognition of the test engine. The engine fault mode recognition based on variable weight prediction mainly includes two parts, which are single-time fault mode classification model establishment and multi-time fault mode weighted prediction, and the main process is as shown in Figure 6

[0087] The pseudo code of the engine fault mode recognition based on variable weight prediction is as follows:

[0088]

[0089]

[0090] 301) Constructing a single-time fault mode classification model

[0091] The construction of the single-time fault mode classification model mainly includes three parts: parameter dimension reduction based on principal component analysis, degradation trajectory clustering based on DBSCAN, and single-time fault mode classification model construction based on SVM, as shown in Figure 7

[0092] (301) Parameter dimension reduction based on principal component analysis

[0093] After the working condition equivalent transformation of the air path performance parameters, the parameters are still multi-dimensional, and there may be correlation between the parameters, so that the trend of the fault mode has certain difficulty and complexity. If the performance parameters are analyzed respectively, the efficiency will be greatly reduced, and the problem will be considered in isolation rather than comprehensively. Therefore, considering the correlation between the performance parameters, the PCA method in unsupervised metric learning is used for dimension reduction, which can obtain low-dimensional parameter information without losing original parameter information.​​

[0094] PCA is a non-parametric dimensionality reduction technique that aims to transform the numerous variables in a high-dimensional dataset into fewer new variables while preserving the key information of the data. Its core principle is to perform linear transformation of data in uncorrelated directions by computing the eigenvectors of the covariance matrix. These new dimensions are ordered according to their contribution to the total variance to identify the most important dimensions that retain important information of the data. This helps to reduce the dimensionality of the data while maintaining the key structural features of the data.

[0095] (302) Degradation trajectory clustering based on DBSCAN

[0096] The training engine monitors data from healthy to degraded, considering that the degradation features are not obvious at the beginning of engine operation, so the late degradation data is selected to determine the failure mode category. After extracting the low-dimensional gas path parameter features of the training engine by the PCA method, the DBSCAN method is used to divide the late gas path parameter features of the training engine, and the corresponding failure mode of each training engine is obtained.

[0097] After clustering the failure modes based on the DBSCAN algorithm, it is assumed that there are M failure modes, and the feature matrix of the mth failure mode after clustering is:

[0098]

[0099] where Z m is an N m x l matrix, N m is the number of feature samples of the mth failure mode, and l is the dimension of the failure feature. Z m The corresponding failure mode classification label can be represented as Y m = [m m … m] T .

[0100] (303) Construction of single-time failure mode classification model based on SVM

[0101] The feature matrix of each failure mode is used as the input of the failure mode classifier, and the corresponding failure mode label is used as the output of the classifier to train the failure mode classification model. Unlike neural networks that require a large amount of data for training, SVM as a classic classification algorithm can achieve excellent classification results with a small amount of samples. Since the sample size of the aero-engine failure mode is small, the present application uses SVM to construct a single-time failure mode classification model, which establishes a foundation for testing the failure mode classification of the engine.

[0102] SVM is a supervised learning classification algorithm. Its principle is based on finding a hyperplane with the largest margin to separate data points of different categories. The margin refers to the distance between the data point closest to the hyperplane and the plane, while the support vector is the data point with the closest distance. The present invention designs a fault mode classification model based on SVM. If the training engine sample is X = {(x1, Y1), (x2, Y2), ..., (x N ,Y N )}, the failure mode label is Y k ∈{1,2,…,M′},k=1,2,…,N, then the optimization problem of the SVM model for different types of failure modes can be expressed as:

[0103]

[0104] Among them, φ(x γ ) is the mapping function from input engine data to category labels, is the loss function of the model, and the penalty coefficient C of the hyperparameters and the hyperplane normal vector ω are optimized by the model. ij , hyperplane intercept b ij , thus obtaining the hyperplane partition of the i-th type of failure mode and the j-th type of failure mode. For the test sample x test , a single SVM model belongs to the category The judgment is:

[0105]

[0106] Among them, if If the value is 1, it means that the test sample is the i-th failure mode, otherwise it is the j-th failure mode.

[0107] 302) Multi-time Fault Mode Weighted Prediction

[0108] Because engine performance degradation is a gradual process, the fault characteristics of a test engine that has not fully degraded may not be distinct enough to accurately identify it using monitoring parameters at a single moment. Therefore, a weighted prediction method was introduced to comprehensively consider the fault mode characteristics at multiple moments to determine the engine's final failure mode. This fault mode identification process more comprehensively considers the dynamic changes during engine degradation, addressing the limitations of single-moment fault mode classification.

[0109] According to the SVM fault mode classification model, the fault state of the test engine j at a single moment can be expressed as Its state under fault mode m (m=1,2,…,M) can be calculated as:

[0110]

[0111] After that, the weighted prediction is made according to the comprehensive multi-time fault state, and the weighted prediction result of the test engine j under different fault modes can be calculated as:

[0112]

[0113] In the present application, a recent weighting coefficient τ (0 < τ < 1) and a comprehensive weighting coefficient λ (λ, t) are introduced to analyze the engine degradation trajectory, τ represents the recognition proportion of recent data, and λ represents the number of flight cycles considered comprehensively. The final failure mode of the engine can be calculated as The fault mode with the highest probability is selected as the final failure mode class of the engine.

[0114] The calculation method of τ is: according to the requirement of recent data, a number between 0 and 1 is taken as the weight, and 0.7 is taken as the recent weight coefficient in this case.

[0115] The calculation method of λ is: take 70% of the flight cycle number of the jth engine.

[0116] The process of engine group failure mode recognition based on density clustering-support vector machine and multi-time prediction weighting is illustrated by combining specific cases.

[0117] A, data set introduction

[0118] The present application selects the engine simulation data set provided by a company, which is generated by C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) simulation software. The model simulates the gradual degradation process of five key components, i.e. fan, low pressure compressor, high pressure compressor, high pressure turbine and low pressure turbine, under different operating conditions and failure modes, and generates degradation data of engine performance for subsequent analysis.

[0119] Based on the C-MAPSS simulation, four data sets are obtained. Due to the influence of failure modes and operating conditions on sensor data, the complexity of data sets 1 to 4 increases, and the situation of each data set is shown in Table 1. Since the present application mainly aims at the case of engine variable working condition and multiple failure modes, the fourth data set FD004 is selected, which includes 249 training engine data and 248 test engine data. Each engine runs under 6 different working condition changes, and the final failure mode can be divided into two kinds. The failure modes include fan degradation and high pressure compressor degradation, but the degradation path is not known, and each engine has only one failure mode, and there is no case of two failure modes occurring at the same time.

[0120] The engine contains both HPC degradation and fan degradation failure modes, but the corresponding degradation paths are not labeled.

[0121] Table 1C-MAPSS dataset details

[0122]

[0123] There are four main reasons for selecting this dataset. First, the dataset is authoritative, and the data provided by NASA is reliable and accurate. Second, the dataset is typical, and the monitoring data in the dataset is a mixture of degradation signals, invalid signals, and noise in a certain failure degradation mode, which indicates that it is challenging to make accurate assessments and predictions. Third, the dataset is scalable, and the study can be extended to more complex engine or other object studies after studying this dataset, such as the PHM08 challenge dataset, N-CMAPPS, etc. Fourth, the dataset is comparable, and many current studies focus on this dataset, making it easier to compare.

[0124] B. Engine operating condition division based on the combination of metric learning and clustering

[0125] B1) Extraction of aero-engine operating parameters

[0126] In the FD004 dataset, the sensor monitoring data includes 3 operating parameters and 21 gas path parameters, and the 3 operating parameters need to be extracted as the basis for engine operating condition recognition. According to the operating parameter changes of Engine 1, Engine 2, Engine 3, and Engine 4, it can be seen that the operating parameters of the same engine are in a constant change process, and the operating parameters of different engines differ greatly.

[0127] B2) Determination of training engine cluster centers and operating condition label division

[0128] The three-dimensional training engine operating parameters are used as the input of the K-Means clustering, and since there are 6 operating conditions for the engine samples, the clustering category k is set to 6, and the maximum number of model training iterations is set to 300. The K-Means clustering model is trained to obtain the engine operating condition cluster centers and the flight cycle clustering labels. The operating condition label records the operating condition of the flight cycle, and by comparing the operating condition labels, the flight altitude, Mach number, and throttle lever resolver angle range of the engine at that time can be determined.

[0129] In the aero-engine operating condition clustering results and clustering centers, the 6 points of the operating condition clustering results represent the center points of the 6 operating condition categories, and the 3 coordinate axes represent the 3 operating condition parameters of flight altitude, Mach number and throttle lever resolver angle, wherein the 6 categories represent 6 different operating conditions of the aero-engine. The clustering center coordinates and clustering radius of each operating condition are shown in Table 2. The operating condition clustering center is the basis for the test engine operating condition recognition based on the MDS algorithm.

[0130] Table 2 Clustering center and clustering radius of flight operating condition

[0131]

[0132] After identifying the operating conditions of the training engine, the operating condition label corresponding to each cycle is output, for example, the operating condition label corresponding to the first operating condition is 1, and the operating condition label corresponding to the sixth operating condition is 6.

[0133] B3) Test engine operating condition adaptive recognition based on MDS metric

[0134] The MDS metric learning method is used to scale the test engine operating parameters and operating condition centers to one-dimensional data, and the distance d between the operating data at a single time and each operating condition center is calculated. j The nearest distance is adaptively selected as the belonging operating condition.

[0135] C. Engine data fusion based on operating condition equivalent transformation

[0136] C1) Selection of operating condition-dependent sensitive gas path parameters

[0137] The 21 gas path parameters in FD004 are selected as the data basis for health assessment and life prediction, so the engine data fusion based on operating condition equivalent transformation is carried out for the 21 gas path parameters. Table 3 below shows the specific description and symbolic representation of the 21 gas path parameters in the C-MAPSS data.

[0138] Table 3 Chinese description and symbolic representation of gas path parameters

[0139]

[0140] In the variation trend of the original 21 gas path parameters of No. 1 training engine, the original gas path parameter data is disordered, and it is difficult to find the degradation rule of the engine gas path performance parameters without dividing the operating conditions. Therefore, it is necessary to select the operating condition-dependent sensitive gas path parameters that can accurately represent the degradation state of the engine, so as to more efficiently and accurately complete the subsequent assessment and prediction work.

[0141] Since the 5 parameters T2, P2, farB, Nf_dmd, PCNfR_dmd do not change with time under the same working condition, and the degradation trends of the 5 parameters T24, P15, epr, W31, W32 are not obvious, the 10 parameters that remain constant under various working conditions and are not sensitive to degradation can be initially excluded, and the remaining 11 parameters, i.e., T30, T50, P30, Nf, NC, Ps30, phi, NRf, NRc, BPR, and htBleed, are selected as index parameters for further evaluation and prediction.

[0142] C2) Multi-condition time series data decomposition

[0143] After screening the degradation-sensitive performance parameters of the aero-engine, the degradation trends of the degradation-sensitive performance parameters under 6 independent working conditions can be extracted according to the working condition division results. Under the same working condition, the change trends of the parameters are consistent as a whole, but under different working conditions, the degradation ranges of the same parameters differ greatly. Therefore, it is of great significance to perform data decomposition and reconstruction based on working conditions to restore the original degradation trends of the parameters.

[0144] C3) Standardization of performance parameters under each working condition

[0145] After the multi-condition time series data decomposition, the performance parameters under each working condition are standardized. For the normalization method, the applicability of two normalization methods for the standardization of performance parameters under each working condition of the No. 1 training engine is verified. Both linear normalization and z-score normalization methods can remove the interference of variable working conditions to some extent and integrate the degradation trends of the parameters under each working condition. By comparing the standardization results of Nf and NRf parameters in the two figures, it can be seen that the linear normalization method is easily affected by local outliers, and the performance of the z-score normalization method is more stable, which is completely consistent with the previous theoretical analysis.

[0146] C4) Reconstruction of gas path performance parameters

[0147] The gas path performance parameters standardized under each working condition are combined to reconstruct the degradation time series that integrates multiple working conditions. The engine data fusion method based on working condition equivalent transformation can achieve the effect of removing the interference of variable working conditions and restoring the trends of the gas path performance parameters.

[0148] D. Engine fault mode identification based on classification and variable weight prediction

[0149] (D1) Parameter dimension reduction based on principal component analysis

[0150] In the prediction of engine single-time fault mode, first, the parameter dimension reduction based on principal component analysis is carried out, which provides preliminary low-dimensional features for the clustering of gas path parameter data. For the 11 gas path parameter working condition fusion data obtained above, the PCA method is used for dimension reduction. Based on the proportion of parameter interpretable feature information of the PCA method, it can be known that the first principal component and the second principal component account for 96.2% of the total information, which can represent the degradation information of the engine gas path parameters. Therefore, the first principal component “PCA1” and the second principal component “PCA2” are selected as the subsequent fault mode analysis data.

[0151] (D2) Degradation clustering based on DBSCAN

[0152] The first principal component and the second principal component are selected as the degradation trajectory clustering data basis, and two types of degradation trends can be preliminarily identified from the data distribution, which is consistent with the two types of fault modes of the engine described by the data set.

[0153] Since the initial fault mode characteristics of the engine are not obvious, there is a situation of overlapping feature points, which makes it difficult to further divide the fault modes of each engine. Therefore, the late recession data is selected to determine the fault mode category. The training engine degradation corresponding to the last 30% flight cycles and the last 20% flight cycles is intercepted, respectively. The recession data in the late degradation can more clearly distinguish the fault mode without data overlap. Therefore, the last 20% degradation data is intercepted for fault mode division.

[0154] With the clustering radius e as 20 and the minimum number of regions MinPts as 1000, the DBSCAN clustering method is used to divide the above training engine degradation mode, so as to obtain the fault mode classification label of each training engine.

[0155] (D3) Construction of single-time fault mode classification model based on SVM

[0156] The PCA degradation features of the training engine are taken as the input, and the flight cycle fault mode label is taken as the output. The SVM binary classification model is trained, the single-time fault mode classification model is established, and the model basis for subsequent test engine multi-time weighted prediction is laid. Table 4 is the model parameters.

[0157] Table 4 Model parameters

[0158]

[0159] (D4) Multi-time fault mode weighted prediction

[0160] The trained fault mode classifier is used to classify the single-time fault mode of the test engine.

[0161] Since the engine only experiences one failure mode from operation to end, the weighted failure mode prediction method is adopted to predict the final failure mode of the engine, the near-term weight coefficient τ is set to 0.7, and the comprehensive consideration quantity λ is the number of predicted failure modes of the engine in the last 70%.

[0162] To sum up, the engine operating condition division based on the combination of metric learning and clustering, the engine data fusion based on operating condition equivalent transformation and the engine failure mode recognition method based on classification and edge weight prediction are established by analyzing the complex operation of the aero-engine and combining the metric learning method with the clustering method, and a complete process of engine variable operating condition and multi-failure mode separation and division is formed.

[0163] After completing the operating condition division of the engine, the data fusion of the engine under variable operating conditions and the engine failure mode recognition, the application further includes the operation processes of variable operating condition engine health assessment based on improved autoencoder and metric learning and multi-failure mode engine life prediction based on metric learning and field generalization, which are as follows:

[0164] II. Variable operating condition engine health assessment based on improved autoencoder and metric learning

[0165] The variable operating condition engine health assessment method comprises the following steps: constructing a long short-term memory convolutional autoencoder LSTMCNN_AE feature extraction model, training the LSTMCNN_AE feature extraction model using engine training set data to obtain a trained LSTMCNN_AE feature extraction model; inputting engine test set data into the trained LSTMCNN_AE feature extraction model to obtain low-dimensional time series degradation feature data; constructing and training an NCA metric learning model under each failure mode, and mapping the low-dimensional time series degradation feature data using the trained NCA metric learning model under each failure mode to obtain mapping feature data; and performing engine health assessment processing based on an optimized Mahalanobis distance feature matrix and the mapping feature data.

[0166] Training the LSTMCNN_AE feature extraction model using engine training set data to obtain a trained LSTMCNN_AE feature extraction model comprises the following steps: using a sliding window with a set length and a step to preprocess multi-dimensional engine parameter data under different operating conditions, and dividing the multi-dimensional engine parameter data into training samples and test samples according to the sliding window method; taking the training samples as the input and output of the model, reconstructing the input sequence through the LSTMCNN_AE feature extraction model; capturing feature information based on the training data set to determine the model parameters of the feature extraction model, and optimizing the model parameters through a residual learning method to obtain a trained LSTMCNN_AE feature extraction model.

[0167] The constructing and training of the NCA metric learning model under each failure mode comprises: dividing the low-dimensional time-series degradation feature data into time-series degradation feature data under different failure modes according to the identified engine failure mode labels; under each failure mode, dividing the time-series degradation feature data into health set feature data and degradation set feature data, and marking the corresponding health set state label and degradation set state label respectively; constructing the NCA metric learning model under each failure mode, and taking the health set feature and the degradation set feature under each failure mode as the model input, and taking the state label corresponding to the feature as the model output for training, to obtain the trained NCA metric learning model under each failure mode and the optimized Mahalanobis distance feature matrix.

[0168] The engine health assessment processing based on the optimized Mahalanobis distance feature matrix and the mapping feature data comprises: calculating the health space and degradation space distance of the engine under each working condition based on the optimized Mahalanobis distance feature matrix and the mapping feature data; performing working condition-specific normalization on the distance under each working condition by using a linear normalization method to obtain the health assessment index of the engine under each working condition; reconstructing the health assessment index of the engine under each working condition into a health assessment curve under the time sequence based on flight cycle rearrangement; traversing all the failure mode assessments to form the health assessment result under each failure mode.

[0169] In the engine health assessment under multiple failure modes and variable working conditions, the application proposes a new auto-encoder feature extraction method, introduces a metric learning method into the health assessment, adopts a nearest neighbor component analysis (NCA) algorithm, divides the neighboring classes and transforms the Mahalanobis distance to measure the difference between the degradation features and the initial health features under different working conditions, and then obtains the health assessment value under variable working conditions by using the working condition transformation idea, to complete the degradation trend representation of the individual engine.

[0170] 2.1, introduction of NCA algorithm principle

[0171] The KNN method is a supervised non-parametric machine learning method, which classifies the sample into the class that appears most frequently in the nearest neighbor by selecting K nearest neighbors of the sample. Assuming that the input sample is {x1, x2, …, x n}, and the corresponding class labels are c1, c2, …, c n , for the sample label to be learned, a distance measure is needed to make the nearest neighbor classification optimal. For the learnable Mahalanobis distance in metric learning, the transformation matrix Q=A T A is a symmetric positive semi-definite matrix, and the Mahalanobis distance between two samples can be expressed as:

[0172]

[0173] where Q is a learnable parameter, called metric, and the goal of metric learning is to optimize Q with some constraints to learn a distance metric.

[0174]

[0175] where p ij represents the data point x i In the random k-NN, another data point x j is chosen instead, and the probability of the corresponding class is given by:

[0176]

[0177] The optimization goal is to maximize the number of correctly classified points, and thus the optimization objective can be written as:

[0178]

[0179] 2.2 Engine time series degradation feature extraction based on LSTMCNN_AE

[0180] The application proposes a new auto-encoder feature extraction method, i.e., long short-term memory convolutional auto-encoder (LSTMCNN_AE), for generating low-dimensional and low-noise engine time series feature information, thereby providing a data basis for variable working condition engine health assessment.

[0181] 21) Construction of LSTMCNN_AE feature extraction network structure

[0182] In order to construct an aero-engine feature extraction model suitable for time series and high noise, the application proposes a feature extraction method based on LSTMCNN_AE algorithm, and the network structure is as shown in Figure 8 The LSTM and convolution units are embedded in the specific network to capture sequence information while reducing feature dimension and learning engine degradation features. The LSTM unit in LSTMCNN_AE is used to capture sequence information from multi-sensor time series data. The convolution and deconvolution units are further embedded after the LSTM unit to filter noise and extract corresponding engine degradation features. Then, residual learning is used to reduce the training difficulty and improve the feature learning performance of LSTMCNN_AE.

[0183] LSTMCNN_AE consists of two parts: an encoder and a decoder, and the details are as shown in Figure 9The multi-dimensional recession-sensitive gas path parameters after working condition transformation are taken as input. An encoder composed of LSTM and convolution units is used to encode the time series input and extract effective features from the original signal. A decoder composed of LSTM and deconvolution units is used to decode the learned features and reconstruct the input data. The LSTMCNN_AE mainly includes the following unit modules:

[0184] (21) LSTM encoding unit

[0185] The LSTM unit in the encoder is used to learn the time series information from the input signal. LSTM includes three gating units, namely input gate, forget gate and output gate. The special architecture of LSTM ensures that the model can avoid the problem of gradient disappearance and accurately capture the long-term relevance of time series information.

[0186] In LSTM, for each time point t, the update process of the hidden state ht can be described as follows:

[0187] i t =σ(W i x t +V i h t-1 +b i )

[0188] f t =σ(W f x t +V f h t-1 +b f )

[0189] O t =σ(W o x t +V o h t-1 +b o )

[0190] c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +W c h t-1 +b c )

[0191] h t =o t ⊙tanh(c t ) (16)

[0192] where i t , f t , Ot , c t are input gate, forget gate, output gate and memory cell respectively, W and V are weight matrices, b is bias vector, σ is activation function, and represents convolution unit.

[0193] (22) Convolutional layer

[0194] The feature h generated by LSTM is still noisy and high-dimensional, which cannot effectively represent the engine health state. Therefore, the output h is input to the convolutional layer for further dimension reduction and feature learning. The learned feature F of the convolutional layer is as follows:

[0195]

[0196] where represents convolution operation, w is convolution kernel, b is bias, and j is the number of channels.

[0197] (23) Deconvolutional layer

[0198] The learned feature is then transmitted to the decoder for further decoding. The decoder part is composed of LSTM units and deconvolutional layers, which can generate the reconstructed vector As shown in the figure, the deconvolution operation can convert the feature into the original shape. The operation of deconvolution is opposite to that of convolution:

[0199]

[0200] where, represents deconvolution operation, is deconvolution kernel, is bias, and i is the number of channels.

[0201] (24) LSTM decoding unit

[0202] The LSTM unit in the decoder is used to capture the time series information from the feature generated by the previous layer. The output of the deconvolutional layer is input into the LSTM decoding unit. The output of this LSTM can be calculated as follows:

[0203]

[0204] (25) Residual learning block

[0205] Residual networks are usually used to solve the problem of vanishing gradient with the deepening of the network. The adoption of residual learning can improve the feature learning performance of LSTMCNN_AE. The output of the residual learning block can be represented as follows:

[0206]

[0207] (26) Fully connected layer

[0208] To ensure the shape of input data is the same, a full connection layer is set at the top layer of LSTMCNN_AE. The final output of the deep network can be obtained by the following formula:

[0209]

[0210] 22) Engine time series degradation feature extraction process based on LSTMCNN_AE

[0211] The feature extraction based on LSTMCNN_AE is divided into offline LSTMCNN_AE modeling and online degradation feature extraction.

[0212] In offline LSTMCNN_AE modeling, the model is trained by using the training engine gas path parameters. When the degradation data is input into the model, the original training engine gas path parameter data is reconstructed by setting the model structure and adjusting the parameters. Assuming that the multi-dimensional time series contains all data from the start to failure of the training unit i, the model is trained by using X i , and the reconstruction error can be represented as The optimal model is trained by minimizing the reconstruction error.

[0213] In online degradation feature extraction, the entire data of the training engine and the entire data of the test engine are used as input. At this time, the model input is , and the output is The corresponding degradation feature is the reconstructed feature learned by the LSTMCNN_AE encoder , that is, the engine time series degradation feature is extracted.

[0214] The engine time series degradation feature extraction process based on LSMTCNN_AE is shown in Figure 10 , which includes:

[0215] (221) Offline LSTMCNN_AE modeling

[0216] The offline LSTMCNN_AE modeling stage consists of three parts

[0217] 201. Determine the training set and test set of the model. The gas path data of the training engine is selected as the training set of the model, and the entire data of the training engine and the test engine is selected as the test set. Considering the strong time series of the aero-engine data, the multi-dimensional gas path parameters are preprocessed by using the sliding window method.

[0218] In data preprocessing, the sliding window is L, and the step is d along the multi-dimensional gas path parameters to generate model input data. The sample generation process is shown in Figure 11 .

[0219] The preprocessed model training set is The length of the two-dimensional gas path parameter data is L×K (sliding window length×number of gas path parameters). The model test set is X test ={X trian_all ,X test_all},in, They respectively represent the two-dimensional gas path parameter data of all training engines and test engines obtained by sliding window method.

[0220] 202. Determine the input and output of the model. When training the model, the input and output sequences are the same, both of which are training set data X train , reconstruct the input sequence through the LSTMCNN_AE model.

[0221] 203. Train the LSTMCNN_AE feature extraction model. Capture feature information based on the training dataset, determine the model parameters such as the number of layers, number of nodes, learning rate, batch size, and number of iterations, and optimize the model parameters using residual learning to obtain the optimal feature extraction model.

[0222] (222) Online degradation feature extraction

[0223] Determine the input for feature extraction. When obtaining degenerate features, the input is the test set data X test ={X trian_all ,X test_all}, thereby obtaining low-dimensional degraded features through the constructed encoder.

[0224] Extract engine degradation features. Input the test set data into the trained LSTMCNN_AE model and obtain the low-dimensional data output by the encoder, which is the extracted low-dimensional degradation features. Establish a data foundation for subsequent health assessments.

[0225] 2.3 Engine Health Assessment under Multiple Fault Modes and Variable Operating Conditions Based on NCA Algorithm

[0226] The present invention uses the NCA algorithm to divide the health and fault states of the training and test engines, and constructs a health assessment model based on the idea of ​​working condition transformation, measures the difference between health and degradation under each working condition, and constructs a health index through the distance value. The health index can reflect the health state of the engine sample under the flight cycle. Afterwards, the health index is reconstructed with the flight cycle as the transformation benchmark. Therefore, the steps of engine health assessment under multiple fault modes and variable working conditions are divided into: division and mapping of health and degradation sets based on the NCA model, construction of health indicators based on the optimization matrix, and fusion of variable working condition health states based on working condition transformation. The specific process is as follows: Figure 12 shown.

[0227] The pseudo code for variable operating condition engine health assessment based on the NCA algorithm is as follows:

[0228]

[0229] The variable operating condition health assessment process based on NCA metric learning is mainly divided into three parts, which will be introduced below in turn.

[0230] 201) Partitioning and mapping of healthy and degraded sets based on NCA model

[0231] According to the identified engine failure mode labels, the extracted time series features are divided into features under different failure modes Each fault mode corresponds to an NCA metric learning model. The following takes the construction of the NCA metric learning model under the i-th fault mode as an example to illustrate.

[0232] Under the i-th fault mode, the NCA metric learning model is supervisedly trained using the healthy and degraded sets of the training engine degradation features. The degradation features of the test engine are then used as input to the NCA algorithm, dividing the test samples into healthy and degraded sets.

[0233] (2011) Healthy Set and Degraded Set State Labeling

[0234] In actual engine operation, its lifespan does not decrease linearly with degradation. Instead, after reaching its lifespan limit after a period of operation, the remaining lifespan gradually decreases. Therefore, it can be assumed that initial engine degradation is negligible, the remaining lifespan is constant, and after reaching a critical value, the remaining lifespan decreases linearly.

[0235] The healthy set and the degraded set are divided based on the degradation characteristics of the training engine. According to the principle of piecewise linear degradation, the healthy state characteristics of the training engine are regarded as the original input of the healthy set of the NCA model, and the engine degradation state characteristics are regarded as the original input of the degradation set, and the healthy and fault labels are used as the model output. The degradation characteristics of the training engine can be expressed as X train ={X1,X2,...X n_train}, where X n =X ij , i is the engine flight cycle, j is the degraded feature dimension. The first 10% of the engine features are taken as the healthy set and the last 90% of the features are taken as the degraded set. Then the healthy set H train and degenerate set D train It can be expressed as:

[0236] H train ={H1,H2,...H n_H} (twenty two)

[0237] D train ={D1,D2,...Dn_D}(23)

[0238] where H train is the health feature corresponding to each flight cycle, D train is the failure feature corresponding to each flight cycle, H i and D i are the reduced dimension features processed by LSTMCNN_AE,

[0239] On the basis of the division of the health set and the degradation set, the labels thereof are set. The state label can be represented as where H is the state label corresponding to the health state, and D is the state label corresponding to the degradation state. Each engine degradation feature corresponds to a state label. Therefore, the state labels corresponding to the health set and the degradation set can be represented as:

[0240]

[0241] (2012) NCA metric learning model training

[0242] In the NCA metric learning model training, the core is to make the model have the ability to effectively separate the engine health and degradation state features through training, and obtain a learned Mahalanobis distance matrix, so that the health and degradation features are sufficiently separated and the same class features are relatively aggregated, and a stable and efficient distance calculation basis is laid for the subsequent. Taking the training engine health set and degradation set features X input ={H train ,D train} as the model input, the state labels corresponding to the features as the model output, the model maximum iteration number max_iter is set, and the NCA algorithm parameter random_state is optimized, so as to effectively divide and identify the engine health and degradation state, and obtain the learned optimized Mahalanobis distance matrix Q.

[0243] (2013) Engine degradation feature adaptive mapping

[0244] In the adaptive mapping of the degradation feature, the core is to map the input feature to a new space that meets the state of the feature according to the NCA metric learning model, so as to reduce the health and failure feature aliasing caused by data fluctuations or working condition changes, make the metric calculation between the health and the degradation more stable, more robust, and the distance representation more accurate. Taking the training engine X train ={X1,X2,...X n_train} and the test engine x test ={x1,x2,...x n_testThe engine one-dimensional mapping features can be expressed as:

[0245] F train = {F1, F2,... F n_train} (26)

[0246] f test = {f1, f2,... f n_test} (27)

[0247] Each engine mapping feature can be expressed as a set of health mapping features and degradation feature mapping, F1= {H train1 , D train1}, f1= {H test1 , D test1}.

[0248] 202) Construction of sub-working condition health index based on optimization matrix

[0249] The core of the working condition health index construction is to divide the mapping features based on the operating condition, and to construct a health evaluation scheme under the sub-working condition to describe the degradation degree of the engine in this flight cycle. Unlike the traditional distance measurement method, the traditional distance measurement needs to set the measurement data and the measurement length manually, and there are differences between different measurement spaces. The method based on NCA distance measurement learning can use the learned optimal distance matrix for universal measurement, reducing the uncertainty of the measurement method and improving its robustness.

[0250] (2021) Design of sub-working condition measurement scheme

[0251] Considering that the engines are initially all in a healthy state, but there may be slight differences between individual engines, the method of measuring the distance between the degradation set and the health set composed of all engines has poor stability and is not suitable for aero-engines with individual differences. Therefore, based on individual engines, the learned covariance matrix is used to measure the difference between the degradation state and the health state under the same working condition, and the engine health state index value (Confidence Value, CV) under the sub-working condition is constructed. The greater the distance value, the higher the degree of recession, thereby representing the health status of the engine under the working condition.

[0252] (2022) Construction of sub-working condition health index

[0253] Taking the i-th test engine as an example, its mapping feature can be expressed as f i = {H i , D i}, and the mapping feature is divided into health and degradation features under different working conditions according to the working condition labels obtained in the previous chapter, which can be expressed as:

[0254]

[0255] where k is the number of operating conditions, and is the health mapping feature and the degradation mapping feature under the engine operating condition m.

[0256] Under each operating condition, the distance d between the fault state feature and the first health state feature is calculated based on the covariance matrix Q obtained by optimization learning, and the calculation formula is as follows:

[0257]

[0258] The distance set under the mth operating condition is shown in the following formula, which can represent the difference between the current state and the health state of the engine under this condition:

[0259]

[0260] (2023) Health evaluation index construction based on distance normalization under sub-condition

[0261] In order to unify the health index to a consistent health evaluation index measure, the distance difference obtained by calculation needs to be normalized. In the construction of health index, CV=0 usually represents the health state, and CV=1 represents the completely fault state. Since the calculation standards are different for each condition, independent standardization is performed for each condition, and the linear normalization method conforms to the unified standard of 0-1, so the linear normalization method is selected to normalize the distance of each condition, thereby obtaining the engine health evaluation index. In the mth(1≤m≤k) operating condition, the maximum distance value in all engine samples is The minimum distance value is Then in this condition, the standard distance value is calculated as shown in the following formula:

[0262]

[0263] where CV m is the normalized health evaluation index under the mth condition, d m is the distance value under the mth condition.

[0264] 203) Health state fusion under variable conditions based on condition transformation

[0265] After obtaining the health index of each condition, the health index generation research is carried out based on the flight cycle as the standard by means of condition transformation thinking, so as to obtain the health state evaluation curve of each engine.

[0266] (2031) Variable condition health index reconstruction

[0267] After the construction of the health indicators under different working conditions, the health evaluation indicator set of the training engine and the test engine under different working conditions can be obtained, and formula (33) is the set of health indicators of the test engine i under each working condition:

[0268]

[0269] Based on the working condition transformation idea, the CV index is rearranged based on the flight cycle, and the health evaluation curve under the timing condition of the engine is reconstructed. The curve can quantitatively reflect the change of the health degradation condition of the engine from the specific numerical value, and lay a data foundation for the subsequent residual life prediction under multiple fault modes.

[0270] (2032) Generation of health evaluation results under multiple fault modes and variable working conditions

[0271] Traverse all the evaluations under the fault modes to form the health evaluation results under each fault mode. For the training engine and the test engine Generate a variable working condition health evaluation set under multiple fault modes:

[0272]

[0273] (2033) Health evaluation effect evaluation

[0274] The present application proposes two evaluation indexes to quantify the results of the health evaluation method, and compares the proposed method with other mainstream health evaluation methods, thereby proving the effect of the health evaluation method proposed in the present application.

[0275] CV curve trend index Trend

[0276] The CV degradation curve shows a monotonous trend, which can represent the degradation trend of the engine. For the training engine, if the degradation trend is obvious, it is consistent with the operation condition of the training engine from health to complete failure. The present application proposes a Trend index to quantify the degradation trend of the CV curve, that is, the sum of the rising or falling of the CV value in each flight cycle. The larger the index, the more obvious the degradation trend, and the specific calculation formula is as follows:

[0277]

[0278] Wherein, is the health index of the jth engine under the ith flight cycle under the fault mode m, len is the running time of the engine, n m is the number of engines under the fault mode m, k is the number of fault modes, and n is the total number of engines.

[0279] CV curve fluctuation index Fluc

[0280] The CV degradation curve is relatively smooth, and the curve fluctuation is small, which can indicate that the health assessment process is relatively stable, and the selected method is appropriate. The Fluc index is proposed in this paper to quantify the size of the CV curve fluctuation, that is, the sum of the absolute values of the original CV value and the smoothed CV value of each engine, the smaller the index, the smaller the CV curve fluctuation, and the specific calculation formula is as follows:

[0281]

[0282] Wherein, and are the original calculated health index and the smoothed health index of the jth engine under the ith flight cycle under the fault mode m, len is the running time of the engine, n m is the number of engines under the fault mode m, k is the number of fault modes, and n is the total number of engines.

[0283] Three, multi-fault mode engine life prediction based on metric learning and domain generalization

[0284] After determining the health state trend of the engine under variable working conditions, a suitable multi-fault mode life prediction model needs to be established based on this. In this invention, the classical deep metric learning Siamese network and the recurrent neural network GRU combination strategy are selected, and the health evaluation results of the engine under different fault modes obtained above are used for life prediction.

[0285] A multi-fault mode engine life prediction method, comprising: obtaining health evaluation time series curves of training engines under different fault modes, and selecting health evaluation time series curves of training engines under two fault modes; constructing an SN-GRU model composed of two Siamese networks and a GRU network, and training the SN-GRU model using the health evaluation time series curves of the training engines under the two fault modes to obtain a trained SN-GRU model; and performing life prediction processing on a test engine based on the trained SN-GRU model.

[0286] The SN-GRU model is trained by using the health evaluation time curves of the training engines in the two failure modes, including: the health index samples in the first failure mode and the health index samples in the second failure mode are obtained by respectively performing sliding window cutting processing on the health evaluation time curves of the training engines in the two failure modes; the residual life corresponding to the health index samples in the first failure mode and the residual life corresponding to the health index samples in the second failure mode are obtained; in each iteration, the health index samples in the first failure mode and the health index samples in the second failure mode are respectively taken as the input of the Siamese network, the residual life corresponding to the health index samples in the first failure mode and the residual life corresponding to the health index samples in the second failure mode are simultaneously taken as the output of the GRU prediction network, and the hyperparameters of the Siamese network and the GRU prediction network are updated through a gradient descent algorithm, to obtain the trained SN-GRU model.

[0287] The life prediction processing of the test engine based on the trained SN-GRU model includes: obtaining health index samples of the test engine in k failure modes The health index samples are processed by the Siamese network and the GRU prediction network in the trained SN-GRU model to obtain the direct prediction RUL of the test engine; the formula is:

[0288]

[0289] Among them, is the RUL directly estimated based on the SN-GRU model, G is the Siamese network in the SN-GRU model, and P is the GRU prediction network in the SN-GRU model.

[0290] The life prediction processing of the test engine based on the trained SN-GRU model further includes: obtaining health index samples of the test engine in k failure modes The second health index sample and the residual life (CV j ,RUL j ) corresponding thereto; based on the trained SN-GRU model, the indirect prediction RUL of the test engine is obtained according to the difference between the health index sample and the second health index sample in the Siamese network.

[0291] The life prediction processing of the test engine based on the trained SN-GRU model further includes: adding the predicted RUL of each test engine and the last flight cycle cycle k to obtain the predicted life L of the test engine in k failure modes​k , whose formula is:

[0292] L k = RUL k + cycle k

[0293] Siamese network is a classic method in deep metric learning, which is composed of two embedding networks sharing weights, as shown in Figure 13 . Among them, for input samples (X i , X j ), the Siamese network is trained by minimizing the contrastive loss function:

[0294]

[0295] Where, if the two data of sample pair (X i , X j ) belong to the same class and the value is 0, Y ij is a label with a value of 1; otherwise and a is the minimum margin of sample pairs of different classes.

[0296] The present application focuses on the distance metric capability of Siamese network, embeds Siamese network structure into RUL prediction model, applies Siamese network to learn the difference between health indicators, and inputs GRU network to learn the mapping between indicator difference and RUL difference.

[0297] SN-GRU residual life prediction model under multiple failure modes

[0298] 31) SN-GRU residual life prediction model construction under multiple failure modes

[0299] (31) SN-GRU model and training target design

[0300] Regarding life prediction, the usual practice is to first perform RUL prediction, and then calculate the life of the equipment according to RUL. Therefore, based on the above engine health evaluation results and corresponding RUL of aero-engine under variable working condition and multiple failure modes, the SN-GRU model is constructed, and the structure of the network model is shown in Figure 14 .

[0301] The SN-GRU model adopts a structure in which a Siamese network and a GRU network are connected. The Siamese network is a time-series health index measurement network, and the GRU network is an RUL prediction network. G and P are respectively mapping functions of embedding networks and prediction networks of the Siamese. By introducing a multi-task learning scheme, the training of the SN-GRU model has two main purposes. First, by inputting the degradation curves CV of different time periods to the Siamese network, the network can effectively calculate and learn the differences. Second, the degradation curves and their differences can be mapped to their RUL and RUL differences through the GRU prediction network, so that the mapping relationship between them can be learned from the network, and the difference problem of inconsistent health index degradation caused by different failure modes can be reduced.

[0302] According to the network setting, θ G ,θ P are network hyperparameters of G and P respectively. The first training objective of the proposed network is to minimize the individual prediction loss, that is:

[0303]

[0304] Here, two networks are trained simultaneously to learn the mapping relationship between the health index CV i and the remaining useful life RUL i .

[0305] At the same time, the SN-GRU model can learn the nonlinear relationship between the health index difference and the RUL difference, that is:

[0306] P(z i -z j )=RUL i -RUL j (40)

[0307] Therefore, by learning the difference between the health evaluation indicators corresponding to the RUL, the difference between the RULs can be further estimated, so as to reduce the difference of CV curves caused by multiple failure modes, and improve the generalization ability of the prediction model. For a given health index pair (CV i ,CV j ) and RUL pair (RUL i ,RUL j ), the SN-GRU is trained by optimizing the loss function of formula (41) to meet the condition in formula (40). The regularization term helps to prevent overfitting. By combining the loss functions of formula (39) and formula (41), the total training objective shown in formula (42) is obtained, where ω is a balance parameter.

[0308]

[0309] (32) SN-GRU network structure design

[0310] The SN-GRU network includes two shared embedding networks and a life prediction network. For the embedding network, a CNN architecture is adopted, which can show better performance than other architectures for the RUL prediction task within a reasonable training time. In the CNN architecture, 1-D convolution is used to process the health assessment sequence, and 2 convolution layers, 1 flattening layer and 1 fully connected layer are designed with corresponding dimensions and convolution kernels. In the GRU prediction network, 1 GRU layer, 2 dropout layers and 2 fully connected layers are designed. Figure 15 SN-GRU network structure design diagram.

[0311] 32) Life prediction process under multiple failure modes based on SN-GRU model

[0312] The present application introduces the Siamese network into the RUL prediction problem, establishes the SN-GRU prediction model, designs the corresponding relationship of "distribution difference-residual life difference", and constructs an engine RUL prediction model suitable for multiple failure modes. The life prediction process of the engine under multiple failure modes based on SN-GRU is as shown in Figure 16

[0313] The model training pseudo code of the engine life prediction under multiple failure modes based on SN-GRU is as follows:

[0314]

[0315] By designing the corresponding relationship of "distribution difference-residual life difference", the nonlinear relationship between the time series distribution difference of health assessment indicators and the residual life difference is learned, and the gap between multiple failure modes is reduced through the learning of "difference", so that the prediction model is suitable for multiple failure modes. The model training steps of the engine life prediction under multiple failure modes based on SN-GRU mainly include three parts:

[0316] (301) SN-GRU life prediction model input design under multiple failure modes

[0317] According to the obtained training engine CV results under different failure modes as the input of the SN-GRU model, since the trained SN-GRU model is suitable for multiple failure modes, the CV indicators under all failure modes are used as the input of the model training.

[0318] The present application selects X1=X2={1,2,...,|CV|} in the training engine health indicator sequence, and adopts a sliding window random sampling and ​As the input of training in each iteration, so as to maximize the shared knowledge under the time series. After the similarity of the sequence is quantified, if the prediction model can map the CV data under random sampling to the corresponding RUL, the generalization ability of the prediction model is stronger.

[0319] (302) SN-GRU life prediction model output design under multi-fault mode

[0320] According to the flight cycle of the training engine, the current RUL value is calculated, and the corresponding RUL is obtained according to the set prediction model input sample, respectively And Then the SN-GRU model can be trained to learn the mapping relationship between the CV sequence and the sequence difference and the RUL sequence and the RUL difference.

[0321] (303) SN-GRU model training

[0322] In each round of iteration, respectively And As the input of Siamese network, its corresponding And As the output of the GRU prediction network. Then, through the gradient descent algorithm, and according to the loss function designed in the last stage Update the hyperparameters θ G ,θ P , obtain the trained SN-GRU model, which has strong generalization ability and can process life prediction problems under fault mode.

[0323] 33) SN-GRU based comprehensive life estimation method

[0324] The present application designs a "direct + indirect" combined life estimation method, which can combine the estimation of all training samples to realize the robust and accurate prediction of engine life and improve the generalization of the prediction model under multi-fault mode.

[0325] For a given test engine health indicator Two RUL prediction methods are proposed.

[0326] (331) Direct RUL estimation

[0327] The input health indicators are processed by the embedding network in Siamese and the GRU prediction network, and the RUL is directly predicted, as shown in the following formula:

[0328]

[0329] The The RUL is defined as the model-based RUL, and F and G are the embedding network and the prediction network in SN-GRU, respectively.

[0330] (332) Indirect RUL estimation

[0331] In addition to the direct estimation, the RUL of the test engine can also be inferred from the difference between the test sample and other samples in the Siamese network. According to the SN-GRU prediction model structure, if (CV j , RUL j ) and CV i are known, the RUL j can be estimated by calculating RUL i + P(G(CV j )-G(CV i )). Here, (CV j , RUL j ) can be regarded as a reference for estimating CV i . In order to increase the accuracy of the test set RUL prediction, as many estimates as the number of available training samples for each test sample can be obtained, and the estimates are combined to produce an integrated effect of reducing the estimation variance. The indirect RUL estimation formula based on the reference is as follows:

[0332]

[0333] wherein is defined as the RUL of the indirect estimation, and |D| is the total number of estimation references.

[0334] After completing the direct and indirect RUL estimation, the estimated data is fused to obtain the final RUL prediction result, and the fusion is shown in the following formula:

[0335]

[0336] wherein α and β are the fusion parameters, respectively.

[0337] Finally, the predicted RUL of each test engine and its last flight cycle are added to obtain the final test engine predicted life, and the calculation formula is as follows:

[0338] L k = RUL k + cycle k (46)

[0339] The preferred embodiments of the present application are described above with reference to the accompanying drawings, and are not limited to the scope of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the present application shall be within the scope of the present application.

Claims

1. A method for identifying engine group fault patterns based on density clustering-support vector machine and multi-time prediction weighting, characterized in that: include: Based on principal component analysis (PCA), the training engine performance parameters and the test engine performance parameters are respectively subjected to dimensionality reduction processing to obtain the training engine low-dimensional performance parameters and the test engine low-dimensional performance parameters; In predicting the single-moment fault mode of the engine, the low-dimensional performance parameters of the training engine are clustered based on the density clustering DBSCAN algorithm to obtain the fault mode classification label corresponding to each training engine; Establishing a single-moment fault mode classification model based on a support vector machine (SVM), and training the single-moment fault mode classification model based on the SVM using the low-dimensional performance parameters of the training engines and the fault mode classification labels corresponding to the training engines, to obtain a trained single-moment fault mode classification model based on the SVM; The single-moment failure mode classification label of each test engine is obtained by inputting the low-dimensional performance parameters of the test engine into the trained SVM-based single-moment failure mode classification model; Based on the multi-moment fault mode weighted prediction and the single-moment fault mode classification label of each test engine, a final fault mode classification label of each test engine is obtained.

2. The method according to claim 1, characterized in that The method of obtaining the final failure mode classification label of each test engine based on the multi-moment failure mode weighted prediction and the single-moment failure mode classification label of each test engine includes: According to the trained SVM-based single-moment fault mode classification model, the prediction results of each test engine under the fault mode at a single moment are obtained; The weighted prediction is performed using the fault state of each test engine at a single moment and the comprehensive multi-moment fault state to obtain the weighted prediction results of each test engine under different fault modes; The weighted prediction results of each test engine under different failure modes are used to obtain the final failure mode classification label of each test engine.

3. The method according to claim 2, characterized in that According to the trained SVM-based single-moment fault mode classification model, the prediction results of the fault mode of each test engine at a single moment are obtained, including: 。 4. The method according to claim 3, characterized in that The weighted prediction results of each test engine under different fault modes obtained by using the fault state of each test engine at a single moment and the comprehensive fault state at multiple moments are as follows: ； Where m represents the failure mode; j represents the test engine; The recent weighting coefficient represents the recognition weight of recent data; λ represents the comprehensive weighting coefficient of the number of flight cycles considered comprehensively; k represents the coefficient that affects the weight of the recent and long-term, and represents the flight cycle from the most recent moment; represents the predicted state of test engine j under fault mode m at the moment farthest from t; represents the predicted state of the test engine j under fault mode m at time tk, It represents the prediction result of engine j under fault mode m obtained through weighted prediction.

5. The method according to claim 4, characterized in that The method of obtaining the final failure mode classification label of each test engine by using the weighted prediction results of each test engine under different failure modes includes: ； in, represents the final failure mode classification label of test engine j.

6. The method according to claim 1, characterized in that Before performing dimensionality reduction processing on the training engine performance parameters and the test engine performance parameters based on principal component analysis (PCA), the following steps are also included: Screen out the performance parameters of the engine that are sensitive to degradation under variable operating conditions; Based on the results of the individual engine operating condition classification, the engine data under different operating conditions are extracted to obtain the degradation sensitivity performance parameters under each operating condition; The normalization method is used to normalize the fade sensitivity performance parameters under each working condition. The normalized fade sensitivity performance parameters under each working condition are reconstructed according to the flight cycle sequence of the engine to obtain the fade sensitivity performance parameters under multiple working conditions. The degradation-sensitive performance parameters under the fusion of multiple working conditions are divided into training engine performance parameters and test engine performance parameters.

7. The method according to claim 6, characterized in that The performance parameters that are sensitive to engine degradation under variable operating conditions include: Obtain multi-dimensional operating parameters of the training engine under different working conditions; Based on the obtained multi-dimensional operating parameters, the K-Means clustering algorithm is used to preliminarily identify the operating condition categories of the training engine and obtain the cluster center of each operating condition; The multidimensional operating parameters of the test engine and the cluster centers of the corresponding working conditions are scaled to one dimension using the multidimensional scaling analysis (MDS) algorithm, and the cluster center of the one-dimensional working condition is used as the working condition baseline. Based on the principle of being closest to the operating condition baseline, the operating conditions corresponding to each test engine in each flight cycle are adaptively identified.

8. The method according to claim 6, characterized in that Based on the principle of being closest to the operating condition baseline, the operating conditions corresponding to each test engine in each flight cycle are adaptively identified, including: By measuring the distance between the operating parameters of the test engine at a single moment and the baseline of each working condition, the working condition with the closest distance is selected as the operating condition of the test engine.

9. An engine group fault pattern recognition device based on density clustering-support vector machine and multi-time prediction weighting, characterized in that: include: A parameter dimensionality reduction module is used to perform dimensionality reduction processing on the training engine performance parameters and the test engine performance parameters based on principal component analysis (PCA) to obtain low-dimensional performance parameters of the training engine and low-dimensional performance parameters of the test engine; A fault mode classification module is used to perform fault mode clustering processing on the low-dimensional performance parameters of the training engine based on the density clustering DBSCAN algorithm in predicting the single-moment fault mode of the engine, and obtain a fault mode classification label corresponding to each training engine; a model training module, configured to establish a single-moment fault mode classification model based on a support vector machine (SVM), and train the single-moment fault mode classification model based on the SVM using the low-dimensional performance parameters of the training engines and the fault mode classification labels corresponding to the training engines, thereby obtaining a trained single-moment fault mode classification model based on the SVM; A fault mode recognition module is used to obtain a single-moment fault mode classification label of each test engine by inputting the low-dimensional performance parameters of the test engine into the trained SVM-based single-moment fault mode classification model; Based on the multi-moment fault mode weighted prediction and the single-moment fault mode classification label of each test engine, a final fault mode classification label of each test engine is obtained.

10. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the engine group fault pattern recognition method based on density clustering-support vector machine and multi-moment prediction weighting according to any one of claims 1 to 8 are implemented.

Images