CN115375026A

CN115375026A - Method for predicting service life of aircraft engine in multiple fault modes

Info

Publication number: CN115375026A
Application number: CN202211022629.7A
Authority: CN
Inventors: 马剑; 张子博; 刘学; 张妍; 王超; 索明亮; 吕琛
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2022-06-09
Filing date: 2022-08-25
Publication date: 2022-11-22
Also published as: CN115859777A

Abstract

The invention discloses a method for predicting the service life of an aircraft engine in a multi-fault mode, which comprises the following steps: analyzing a training set of a plurality of aeroengines in a multi-fault mode, and determining a plurality of clustering categories of performance degradation paths of the aeroengines in the training set; training the life prediction model of the aero-engine of each cluster type according to the training set data of the aero-engine of each cluster type to obtain the trained life prediction model of each cluster type; testing the trained life prediction model of each clustering class according to the test set data of the aircraft engine of each clustering class and the corresponding real RUL, and taking the life prediction model passing the test as a practical life prediction model; and determining the attributive clustering category by using the performance degradation path of the aero-engine to be tested, and further determining the corresponding practical life prediction model so as to predict the RUL of the aero-engine to be tested by using the corresponding practical life prediction model.

Description

Method for predicting service life of aircraft engine in multiple fault modes

Technical Field

The invention relates to the field of aircraft engines, in particular to a life prediction method for an aircraft engine in a multi-fault mode.

Background

The aircraft engine is a core system of an aircraft, once a fault occurs, the fault of an aircraft system can be directly caused, so that serious consequences are caused, and at present, the improvement of the power performance and the guarantee of the long-term stable availability of the aircraft engine are important points and difficulties in the field of research. The fault prediction and health management technology can track and predict the running health state of the engine, and can give timely and accurate maintenance suggestions before faults occur, so that the long-term safe and reliable running of the engine can be effectively ensured.

The existing research mainly aims at the individual difference and the degradation characteristics of the aeroengine to carry out life prediction, and the research on the influence of different fault modes on the prediction result is less. Specifically, the engines of the same model inevitably have individual differences, including the difference of the degree of the fault of the factory, the difference of the degradation rate, the difference of the degradation characteristics and the like, so that the service life prediction of the aircraft engine has difficulty. The prediction work of the remaining service life (RUL) of the aircraft engine mainly considers the condition that the aircraft engine runs in a single fault mode or does not consider the influence of the fault mode on the RUL of the aircraft engine, the running condition is complex in the actual work of the aircraft, the condition of the single fault mode is ideal and difficult to realize, and experimental data acquired by a sensor often comprises a plurality of fault modes in the actual situation.

In addition, different characteristics and tracks exist in the performance degradation of the engine caused by different fault modes, the fault degree of the aircraft engine is continuously deepened along with the operation process, and the fault mode is weak at the initial operation stage, so that the performance degradation path of the aircraft engine is difficult to identify, and the difficulty in predicting the service life of the aircraft engine is further improved.

The effective prediction of the RUL of the engine can ensure that maintenance activities can be timely and effectively carried out, so that casualties and losses caused by faults of the engine are reduced, the economy and the safety of the aircraft engine are comprehensively considered, a diagnosis prediction technology aiming at the aircraft engine is urgently needed, and the residual service life of the aircraft engine can be accurately predicted.

Disclosure of Invention

The invention provides a method for predicting the service life of an aircraft engine in a multi-fault mode, which aims to solve the technical problem of accurately predicting the remaining service life of the aircraft engine.

The embodiment of the invention provides a method for predicting the service life of an aircraft engine in a multi-fault mode, which comprises the following steps: analyzing complete data of a plurality of aero-engines from operation to failure in a multi-failure mode as a training set, and determining a plurality of clustering categories of performance degradation paths of the aero-engines of the training set; training the life prediction model of the aero-engine of each cluster type according to the training set data of the aero-engine of each cluster type to obtain the trained life prediction model of each cluster type; taking data of a plurality of aeroengines which are finished before failure in a multi-fault mode as a test set, testing a trained life prediction model of each cluster type according to the test set data of the aeroengine of each cluster type and the corresponding real residual service life RUL, and taking the life prediction model which passes the test as a practical life prediction model; determining the attributive clustering category by using the performance degradation path of the aero-engine to be tested, and determining a corresponding practical life prediction model by using the attributive clustering category; and predicting the RUL of the aero-engine to be tested by using the corresponding practical service life prediction model.

Preferably, the analyzing the complete data of the plurality of aircraft engines from operation to fault in the multi-fault mode as a training set, and the determining the plurality of cluster categories of the performance degradation paths of the aircraft engines of the training set comprises: preprocessing the training set data to obtain the training set data with dimension influence removed and noise influence reduced; extracting degradation features related to an aeroengine performance degradation path from the training set data from which the dimensional influence and the noise influence have been removed; and clustering the time sequence of the degradation characteristics of the training set, which are related to the performance degradation path of the aero-engine, to obtain a plurality of cluster categories of the performance degradation path of the aero-engine of the training set.

Preferably, the preprocessing the training set data to obtain the training set data from which the dimensional influence has been removed and the noise influence has been reduced includes: normalizing the training set data to obtain the training set data without dimension influence; and carrying out smooth noise reduction treatment on the training set data with the dimension influence removed to obtain the training set data with the dimension influence removed and the noise influence reduced.

Preferably, the extracting degradation features related to the aircraft engine performance degradation path from the training set data from which the dimensional influence has been removed and the noise influence has been reduced includes: carrying out Principal Component Analysis (PCA) processing on the training set data with the dimension influence removed and the noise influence reduced to obtain a plurality of PCA characteristics of the training set data; and selecting a specified number of PCA features from the plurality of PCA features as degradation features related to the aircraft engine performance degradation path according to the sequence of the PCA feature variance importance degrees from large to small.

Preferably, the clustering the time series of degradation features of the training set related to the aircraft engine performance degradation path to obtain a plurality of cluster categories of the aircraft engine performance degradation path of the training set comprises: determining similarity between time sequences of degradation characteristics of different aero-engines of the training set by using a Dynamic Time Warping (DTW) algorithm; and clustering the time sequences of the degradation features of different aero-engines according to the similarity between the time sequences of the degradation features of different aero-engines of the training set to obtain a plurality of clustering categories of the performance degradation paths of the aero-engines of the training set.

Preferably, the method further comprises: and analyzing the test set, and determining the cluster category to which the performance degradation path of the aero-engine of the test set belongs.

Preferably, the analyzing the test set and the determining the cluster category to which the performance degradation path of the aircraft engine of the test set belongs includes: preprocessing the test set data to obtain the test set data with dimension influence removed and noise influence reduced; extracting a time series of degradation features of the aircraft engine of the test set from the test set data from which the dimensional effects have been removed and from which the noise effects have been reduced; and comparing the distance between the time series of the degradation characteristics of the aero-engine of the test set and the time series of the cluster center of each cluster category, and determining the cluster category of the cluster center with the minimum distance as the cluster category to which the performance degradation path of the aero-engine of the test set belongs.

Preferably, the determining the cluster category to which the aircraft engine to be tested belongs by using the performance degradation path of the aircraft engine to be tested includes: preprocessing all currently obtained data of the aero-engine to be tested to obtain data with dimension influences removed and noise influences reduced; extracting a time sequence of degradation characteristics of the aero-engine to be tested from the data of the aero-engine to be tested, wherein the data of the aero-engine to be tested are subjected to dimensional influence removal and noise influence reduction; comparing the distance between the time sequence of the degradation characteristics of the aero-engine to be tested and the time sequence of the clustering center of each clustering category, and determining the clustering category of the clustering center with the minimum distance as the clustering category to which the performance degradation path of the aero-engine to be tested belongs;

preferably, the service life prediction model is an aircraft engine residual service life prediction model under multiple faults based on an Informer neural network.

The method has the advantages that the residual service life of the aircraft engine can be accurately predicted, the normal operation of the aircraft engine is ensured to a great extent, and the maintenance and diagnosis efficiency is improved.

Drawings

FIG. 1 is a flow chart of a method for predicting the life of an aircraft engine in multiple failure modes provided by the present invention;

FIG. 2 is a detailed flow chart of the prediction of aircraft engine life under multiple failure modes provided by the present invention;

FIGS. 3a and 3b are graphs of raw data for 21 air path parameters of training engine Nos. 1 and 7, respectively;

FIGS. 4a, 4b, and 4c are a graph of 21 pre-processed parameters, a comparison graph of total LPC outlet temperature parameter, and a comparison graph of total HPC outlet pressure parameter, respectively, for Engine # 1;

FIG. 5 is a graph of pre-processed data for 21 gas path parameters for a fully trained engine;

FIG. 6 is a performance degradation path identification flow diagram;

FIG. 7 is a histogram of aircraft engine gas path parameters PCA feature variance importance;

8a, 8b and 8c are the relationship diagrams of feature 1, feature 2 and time after dimension reduction and the relationship diagrams of feature 1 and feature 2 respectively;

9a, 9b and 9c are the relationship diagrams of feature 1, feature 2 and time after clustering and the relationship diagrams of feature 1 and feature 2 respectively;

FIG. 10 is a diagram of a model for predicting the remaining useful life of an aircraft engine in a multiple failure mode;

FIG. 11 is a diagram of a position encoding step;

FIG. 12 is an attention calculation specific flow chart;

FIG. 13a is a comparison of predicted results and true values after path identification;

FIG. 13b is a comparison of predicted results and true values without path identification;

Detailed Description

It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, suffixes such as "module", "part", or "unit" used to represent elements are used only for facilitating the explanation of the present invention, and have no peculiar meaning by themselves. Thus, "module", "component" or "unit" may be used mixedly.

FIG. 1 is a flow chart of a method for predicting the life of an aircraft engine in multiple failure modes, as shown in FIG. 1, which may include: step S101: analyzing complete data from operation to failure of a plurality of aero-engines in a multi-failure mode as a training set, and determining a plurality of cluster categories of performance degradation paths of the aero-engines of the training set; step S102: training the life prediction model of the aero-engine of each cluster type according to the training set data of the aero-engine of each cluster type to obtain the trained life prediction model of each cluster type; step S103: taking data of a plurality of aero-engines which are ended before failure in a multi-fault mode as a test set, testing a trained life prediction model of each cluster type according to the test set data of the aero-engines of each cluster type and the corresponding real remaining service life RUL, and taking the life prediction model which passes the test as a practical life prediction model; step S104: determining the attributive clustering category by using the performance degradation path of the aero-engine to be tested, and determining a corresponding practical life prediction model by using the attributive clustering category; and predicting the RUL of the aero-engine to be tested by using the corresponding practical service life prediction model.

According to the method, the accurate prediction of the remaining service life of the aero-engine is realized through recognition of the performance degradation path of the aero-engine based on track clustering and prediction of the remaining service life of the aero-engine based on deep learning.

Wherein, step S101 may include: firstly, preprocessing training set data to obtain training set data with dimension influence removed and noise influence reduced, for example, firstly performing normalization processing on the training set data to obtain training set data with dimension influence removed, and then performing smooth noise reduction processing to obtain training set data with dimension influence removed and noise influence reduced, namely preprocessed training set data; secondly, extracting degradation features relevant to the aircraft engine performance degradation path from the training set data with the dimension influence removed and the noise influence reduced, for example, carrying out Principal Component Analysis (PCA) processing on the preprocessed training set data to obtain a plurality of PCA features and corresponding time sequences of the training set data, and selecting a specified number (for example, 2, 3, 4 and the like) of PCA features from the plurality of PCA features as the degradation features relevant to the aircraft engine performance degradation path according to the sequence of the variance importance degree of the PCA features from large to small; then, clustering the time series of the degradation features of the training set, which are related to the aircraft engine performance degradation path, to obtain a plurality of cluster categories of the performance degradation path of the aircraft engine of the training set, for example, determining the similarity between the time series of the degradation features of different aircraft engines of the training set by using a Dynamic Time Warping (DTW) algorithm, and clustering the time series of the degradation features of different aircraft engines according to the similarity between the time series of the degradation features of different aircraft engines of the training set to obtain a plurality of cluster categories of the performance degradation path of the aircraft engine of the training set.

Considering the influence of the performance degradation path difference on the prediction of the residual service life of the aero-engine, the method trains a life prediction model according to the cluster type of the performance degradation path of the training set aero-engine. For example, if there are two cluster categories a and B, then the life prediction model is trained using the training set data for the aircraft engine of cluster category a to obtain a trained life prediction model for cluster category a, and the life prediction model is trained using the training set data for the aircraft engine of cluster category B to obtain a trained life prediction model for cluster category B. It should be noted that the training set data used for training the model is the pre-processed data described above.

After the trained life prediction model for each cluster class is obtained, the trained life prediction model for each cluster class needs to be tested. In one embodiment, if the cluster type of the aircraft engine performance degradation path of the test set is known, the trained life prediction model of the corresponding cluster type can be directly tested according to the test set data of the aircraft engine of the test set and the corresponding real remaining service life RUL, and if the test is passed, the life prediction model is used as a life prediction model which can be actually used, namely a practical life prediction model. In another embodiment, if the cluster type of the aircraft engine performance degradation path of the test set is unknown, the test set needs to be analyzed first to determine the cluster type to which the performance degradation path of the aircraft engine of the test set belongs, which may include preprocessing the data of the test set to obtain the test set data from which the dimensional influence has been removed and the noise influence has been reduced, and the preprocessing method may refer to step S101; extracting a time sequence of degradation characteristics of the aircraft engine of the test set from the test set data from which the dimensional influence and the noise influence have been removed, wherein the specific processing method can refer to the step S101; determining the cluster category to which the performance degradation path of the aero-engine of the test set belongs according to the time sequence of the degradation feature of the aero-engine of the test set and the time sequence of the cluster center of each cluster category, for example, calculating the distance between the time sequence of the degradation feature of the aero-engine of the test set and the time sequence of the cluster center of each cluster category, determining the cluster category of the cluster center with the smallest distance as the cluster category to which the performance degradation path of the aero-engine of the test set belongs, namely, which cluster center is closer to, and then attributing the performance degradation path of the aero-engine of the test set to the cluster category in which the cluster center is located. And then testing the trained life prediction model of the determined corresponding clustering class according to the test set data of the test set of the aircraft engine and the corresponding real residual service life RUL, and if the test is passed, taking the life prediction model as a life prediction model which can be actually used, namely a practical life prediction model. It should be noted that, in the above two embodiments, the test set data adopted by the test model is preprocessed test set data, and the preprocessing method may refer to step S101. By comparing the test result of the step S103 with the prediction result of the prediction method for the unidentified performance degradation path, it can also be obviously found that the prediction accuracy of the present invention is significantly improved, so that the model can be applied to the RUL prediction of the aircraft engine to be tested.

The method is used for predicting the residual service life of the aircraft engine based on the identified performance degradation path, when the method is applied to actual prediction, namely RUL prediction is carried out on the aircraft engine to be detected, all data obtained at present of the aircraft engine to be detected, such as the data of the aircraft engine from operation to the current moment, are preprocessed to obtain data from which dimensional influences are removed and noise influences are reduced, and the preprocessing method can refer to the step S101; extracting a time sequence of degradation characteristics of the aero-engine to be detected from the data of the aero-engine to be detected, wherein the data of the aero-engine to be detected is subjected to dimension influence removal and noise influence reduction, comparing the distance between the time sequence of the degradation characteristics of the aero-engine to be detected and the time sequence of the clustering center of each clustering category, and determining the clustering category of the clustering center with the minimum distance as the clustering category to which the performance degradation path of the aero-engine to be detected belongs; after the cluster type of the aero-engine to be tested is determined, a practical life prediction model corresponding to the cluster type can be determined, and data of the aero-engine to be tested, with dimension influences removed and noise influences reduced, is input into the determined practical life prediction model to predict the RUL of the aero-engine to be tested.

Fig. 2 is a detailed flowchart of prediction of the life of an aircraft engine in multiple failure modes, as shown in fig. 2, the detailed contents are as follows, step 1: the aircraft engine parameter data preprocessing can comprise the following steps: all the engine data, i.e., the raw data, is subjected to normalization processing (e.g., maximum-minimum normalization processing) and noise reduction processing (e.g., local weighted regression processing) in sequence to obtain preprocessed data. Step 2: the method comprises the following steps of identifying the performance degradation path of the aeroengine based on track clustering, specifically, after dimensionality reduction is carried out on preprocessed data to extract degradation features (for example, based on PCA), converting the problem of identifying the performance degradation path into a track clustering problem, and performing the following steps: and calculating the similarity among the paths and clustering the paths. Because the path length is different and the change frequency is different, the engine has different service lives and different degradation rates, and the similarity between sequences with different lengths needs to be calculated firstly. The present invention may use a DTW method to calculate the similarity between engine performance degradation paths. In order to improve the calculation efficiency and enable the clustering center to correspond to a specific track, a K-Medoids algorithm can be selected from the clustering algorithm to perform clustering and identification on the performance degradation path of the aero-engine. And step 3: and predicting the residual service life of the aircraft engine based on deep learning. Considering that the invention takes the aircraft engine operation simulation data as a research basis, as the number of the engine samples is large, the sample data amount of each engine sample is large, and each sample data contains more parameters, the sample data set is very large and complex, and the accuracy requirement on the final residual service life prediction result is high, a deep learning method can be selected as a method for predicting the residual service life of the aircraft engine (or simply called as life prediction) in data driving, for example, an Informer is adopted as a life prediction model network, and the existing sample training network is utilized to complete the residual life prediction of the engine.

The gas circuit component is an important part of an aircraft engine, and the performance degradation change of the aircraft engine is described mainly by monitoring data of parameters such as temperature, pressure, rotating speed and the like of the gas circuit component, including core rotating speed, HPC outlet temperature, fan inlet pressure and the like. Since the overall performance of the gas path component and the aircraft engine is closely related and the probability of occurrence of a fault is higher than that of other components of the aircraft engine, the following embodiments take the aircraft engine gas path parameter data as an example and describe in detail.

1. And preprocessing gas path parameter data of the aircraft engine.

The aeroengine gas circuit parameter data are more, and multidimensional noise interference exists in the running process of the aeroengine, so that the gas circuit parameters need to be normalized and subjected to noise reduction treatment, the influence of parameter dimension, data mutation points and noise is eliminated, and the overall quality of the data is improved.

In the embodiment, a Data Challenge Data set in a 2008 PHM international conference is adopted, the Data set is acquired by a C-MAPSS model, and is a set of commercial turbofan engine operation simulation model developed by the United states national aerospace administration, and the model simulates degradation processes of five components (including a Fan (Fan), a high-pressure compressor (HPC) and the like) of the aero-engine under different operation conditions and failure modes, so that gas path parameter Data of each time in the operation of the aero-engine are acquired. The data set for the C-MAPSS simulation included a total of 4 sub-data sets, as shown in Table 1. In the embodiment, for the case of multiple failure modes under a single operating condition of the engine, the third sub data set FD003 is selected. In the data set, all the engines are of the same type, all the engines run under the same working condition, and the engines comprise two fault modes of HPC degradation and fan degradation, but do not mark corresponding degradation paths. For the engine degradation process in this data set, there is degradation for each component, and it is only the HPC or fan that eventually suffers degradation-type failure, and there is no case where both components suffer degradation-type failure at the same time. Sub data set FD003 includes three text files:

a. training set: containing 100 aero-engine samples, the complete data instance from run to fault of the survey record can be used to train the predictive model.

b. And (3) test set: comprising 100 samples of an aircraft engine, measurements recording incomplete instances of data ending before failure are intended to be used to predict the RUL of the aircraft engine.

c. Test set remaining life: the real RUL of 100 test aeroengine samples is included and is used for comparing with the predicted value to evaluate the predicted effect.

Each instance in the training set and test set samples consists of a time series of 26 variables. The 1 st variable represents an engine number, the 2 nd variable represents the number of flight cycles of the engine, and the 3 rd to 26 th variables represent engine operating parameters and comprise 3 working condition parameters and 21 gas path parameters, wherein the names of the 21 gas path parameters are shown in the table 2.

TABLE 1.C-MAPSS data set specification Table

Properties	FD-001	FD-002	FD-003	FD-004
					Number of operating modes of aircraft engine	1	6	1	6
Number of aircraft engine faults	1	1	2	2
					Number of aeroengines in training set	100	260	100	249
Number of test set aircraft engines	100	150	100	248

TABLE 2 Engine gas path parameter concrete meaning table

The method comprises the steps of firstly carrying out preliminary analysis on engine data to grasp the general trend of gas path parameters in the process of engine performance degradation. Fig. 3a and 3b are raw data graphs of 21 air path parameters of training engines No. 1 and No. 7, with the abscissa being the flight cycle and the ordinate being the corresponding operating values. It can be seen from the visualized data that some parameters indeed show a trend of variation with the increase of the number of flight cycles, which reflects the degradation of the performance of the aircraft engine. However, for the same aircraft engine, part of the gas path parameters show an overall ascending trend, part of the gas path parameters show an overall descending trend, and part of the gas path parameters are not changed all the time. For different aircraft engines, the variation trends of some parameters are the same, the variation trends of some parameters are different or even completely opposite, and the variation amplitudes with the same variation trend may be different. Moreover, the dimensional units of the parameters in the original data are not uniform, the numerical values are not in the same range, and large noise exists, which affects the accuracy of subsequent performance degradation path identification and residual life prediction. It is therefore necessary to preprocess the data to improve the quality of the data.

In this embodiment, data normalization processing based on a maximum and minimum value method and data smoothing and noise reduction processing based on a local weighted regression method are adopted.

1) And (4) carrying out data normalization processing based on a maximum and minimum value method.

All 21 gas path parameters are normalized, and the embodiment selects the data normalization method with the maximum and minimum values to normalize the parameters to the range of [0,1], so that the method is favorable for drawing images in parameter screening, and a specific calculation formula is shown in formula 1. Because the subsequent method is sensitive to the variance of the initial variable, the data is normalized to ensure that the contributions of all parameters under all operating conditions are equal, so that the life prediction is more accurate, reasonable and accurate. For a constant index parameter, such as the total fan inlet temperature, since its value does not change with time, i.e., contains no degradation information, it is uniformly normalized to a constant of 0.

Wherein,

is the maximum value of the jth parameter,

is the minimum value of the jth parameter, x ^i，j Is the value of the parameter before normalization,

is the parameter normalized value.

2) And carrying out data smoothing and noise reduction processing based on a local weighted regression method.

Although part of the gas path parameters are sensitive to the recession process of the aircraft engine, the performance degradation path identification and the residual service life prediction accuracy of the aircraft engine can be influenced if the parameters are not processed due to the fact that a large number of multi-dimensional noise signal interferences exist in the operation process. For the normalized data, the embodiment performs smoothing processing on the gas path parameters by using a local weighted regression (Lowess) smoothing method, so as to filter out random noise in gas path parameter signals acquired by the sensors, retain gas path parameter degradation characteristics, effectively reduce the influence of data mutation points on parameter degradation trends, and provide high-quality input data for subsequently establishing an effective health assessment and residual life prediction model. The local weighted regression algorithm is popularized and applied to the condition of a plurality of independent variables, and the core idea is to carry out weighted fitting on local data to finally obtain an estimated value of a fitting point. The method comprises the following specific steps:

(1) And determining a window range, wherein the window range is used for controlling the smooth scale of the local weighted regression data.

(2) Within a certain window range n, for all points q therein _k K =1,2, …, n, by a weighting function ω _k (q _i ) For g _i And developing d-order polynomial fitting.

(3) Q is obtained through calculation _i Fitting value p of _i Is used in place of q _i 。

Function omega _k (q _i ) The distribution of the weights is determined, and a more common weight function is selected in this embodiment, which is specifically shown in formula 2.

The function is an exponential decay function similar to a gaussian distribution, i.e. points further from the regression value are weighted less heavily. In the formula, lambda is a wavelength parameter, the rate of weight reduction along with distance is controlled, and the weight increases along with the increase of lambda along with the distance reduction rate. The local weighted regression has the characteristic that parameters in a linear regression model can be adaptively changed along with the change of an independent variable value, the parameters in the model can be changed along with the change of different independent variable values, and the model automatically gives an estimated value of a function after the local weighted regression in the range of an independent variable space.

Fig. 4a is a graph of the number 1 engine after the 21 parameters are preprocessed, fig. 4b is a graph of the number 1 training engine LPC outlet total temperature parameters before and after the preprocessing, fig. 4c is a graph of the number 1 training engine HPC outlet total pressure parameters before and after the preprocessing, the abscissa is the flight cycle, and the ordinate is the corresponding value before and after the processing. The visual data shows that the preprocessed data are normalized to [0,1], dimensional influence is removed, meanwhile, the data subjected to noise reduction processing extracts the decline trend in the original data and filters most of noise information, influence of random noise fluctuation in the engine data is effectively reduced no matter the data subjected to the rise trend or the data subjected to the fall trend, the preprocessed data have good signal-to-noise ratio, the data with higher quality is obtained through purification, and a good foundation is laid for the performance degradation path identification and the residual service life prediction of the follow-up aero-engine.

And (3) carrying out preliminary analysis on the preprocessed data, wherein fig. 5 is a preprocessed data graph of 21 gas circuit parameters of all the training engines, the abscissa is the residual service life, and the ordinate is a corresponding running value. The following conclusions can be drawn from the visualization data: (1) The change trend of part of parameters along with the increase of the number of the operation cycles proves that the performance of the aircraft engine is really degraded along with the operation process, so that the gas circuit parameters are reflected to change, and the residual service life of the aircraft engine can be predicted according to the parameter change in the operation of the aircraft engine. (2) The method is characterized in that the change trends of partial parameters such as HPC outlet total pressure, physical core speed, fuel flow, corrected core speed and the like visually represent two clusters, and the two clusters correspond to two fault modes contained in an FD003 data set aeroengine, so that different performance degradation paths can be proved to be different due to the fact that the different fault modes reflect different change tracks on gas path parameters. (3) For the parameters in the step (2), the parameters already present two distributions in the early stage of the operation, which proves that the factory random wear of the aircraft engine determines the fault mode of the aircraft engine under the influence of the condition without interference such as external collision, and therefore, the method for identifying the future performance degradation path of the aircraft engine through the gas circuit parameter range interval in the early operation of the aircraft engine is feasible. (4) Some parameters, such as LPC outlet temperature, HPC outlet temperature, LPT outlet temperature, etc., show a single trend of change with increasing number of operating cycles, demonstrating that different performance degradation paths have little or no effect on the parameters, and that the change is mainly related to the number of operating cycles.

In summary, since the air path parameters of the aircraft engine have noise interference and are not processed, the accuracy of identifying the performance degradation path and predicting the remaining service life of the aircraft engine is affected to a certain extent, and therefore, the parameter data preprocessing operation is required. The method mainly performs preliminary analysis on the data of the aircraft engine, and utilizes the maximum and minimum normalization and local weighted regression smoothing method to preprocess the parameters to improve the data quality, thereby establishing a foundation for subsequent accurate performance degradation path identification and residual service life prediction work.

2. And clustering and identifying the performance degradation paths of the aircraft engine in the multi-fault mode.

Due to the fact that different aero-engines have different factory wear degrees and component degradation speeds in the operation process of the engines are different, the difference exists between fault components when different engines are in fault, namely the difference exists between fault modes, and the direction and the degree of performance degradation of different engines are different, namely different performance degradation paths exist. The failure process of the aircraft engine is complex, the related factors are more, the characteristics in the degradation process are related to a plurality of factors, the characteristics are possibly caused by the self state difference of the engine, the characteristics are possibly caused by different failure modes, and the interference of other factors also exists, so that the prediction of the residual service life is more unfavorable to the development. And the influence of the performance degradation path can be mixed in the individual difference of the engine, if the engine can be identified and distinguished according to the performance degradation path, the influence of the performance degradation path difference on the prediction of the remaining service life of the engine is reduced, so that the relation between the performance degradation of the engine and the remaining service life can be more accurately grasped, and the accuracy of the prediction can be further improved. It can be seen from the preliminary analysis on the preprocessed data, that the aero-engine gas path parameters actually present different performance degradation paths, and if the performance degradation paths of the aero-engine can be accurately identified, prediction is performed on the remaining service life of the aero-engine according to the different performance degradation paths, so that the accuracy of the prediction is greatly improved, therefore, the performance degradation paths of the aero-engine are identified and developed in the embodiment, and a specific flow chart is shown in fig. 6.

1) A method for reducing the dimension of the gas path parameters of an engine based on principal component analysis.

The data driving method collects a large amount of gas path component sensor data in the operation process of the aero-engine through a large amount of observation on a plurality of gas path parameters reflecting the performance of the aero-engine so as to analyze the gas path parameter data and find the rule among the parameters. More aero-engine samples and more gas path parameter variables provide rich sample information for researching the performance degradation trend of the aero-engine, but the workload of calculating and analyzing gas path parameter data is increased to a certain extent. In most cases, strong correlation and certain redundancy may exist among a plurality of gas path parameters, so that the complexity of performance degradation path analysis and identification is increased, and if each gas path parameter is solely analyzed, the obtained result is often not in accordance with the actual situation and does not have integrity. However, the blind reduction of the gas path parameter data can lose much information reflecting the performance degradation of the aero-engine in the original data, so that the result is inconsistent with the actual situation. Reducing the number of variables in a data set undoubtedly reduces the accuracy of the overall data to some extent, however, dimensionality reduction is to retain most of the information of the data while retaining a small amount of data to facilitate calculation, thereby ensuring information screening within an acceptable accuracy reduction range. Compared with a complete and huge original data set, the data set subjected to the dimension reduction processing contains most of degradation information by using fewer degradation features, exploration and visualization of performance degradation path identification are easier, and a machine learning algorithm can analyze data more quickly and efficiently without processing variables irrelevant to a subject or having a small relationship.

Because certain correlation and redundancy exist among the variables of each gas path parameter, relatively few fading characteristics can be used for representing fading information existing in each gas path parameter. Principal Component Analysis (PCA) not only reduces the need for subsequent modelsThe number of the analyzed gas circuit parameters can also reduce data information loss to the maximum extent, so that the collected gas circuit parameter data of the aircraft engine can be comprehensively and efficiently analyzed. The PCA method can effectively reduce the dimensionality of the data set, and can extract most of the available information from a relatively complex and chaotic data set by converting the multi-variables into fewer features containing most of the information of the original data set. The mathematical principle of the method is that the direction of the maximum variance of an input data matrix is taken as a main characteristic axis of data, descending order is carried out according to the importance degree of the variance from large to small, the larger the variance is, the more important the information of the data in the dimension is, and finally a new matrix containing data characteristics is output. For an aircraft engine data vector X = [ X = ₁ ，X ₂ ，...，X _m ]，X _i ＝ [x _i1 ，x _i2 ，...，x _in ] ^T (m is a parameter serial number, and n is the total number of data of corresponding parameters), the PCA method mainly comprises the following steps:

(1) Calculating a covariance matrix of X, and calculating an eigenvalue and an eigenvector;

(2) Sorting the feature values in a descending order, and reserving a feature vector xi corresponding to the maximum feature value;

(3) Obtaining the characteristic Y = X ξ of the gas circuit parameter after dimensionality reduction;

(4) And integrating the features subjected to the dimension reduction into a data vector of the engine, and repeating the steps to obtain the features Z subjected to the dimension reduction of the engine.

The extracted engine characteristic quality determines the performance of the engine recession process, the better the engine characteristic extraction effect is, the more obvious the characteristics of the engine recession process are, and the subsequent fault path identification and the residual working life prediction can be more accurately carried out. And carrying out PCA (principal component analysis) processing on the 21 gas circuit parameters to reconstruct a vector space to obtain 21-dimensional characteristics, wherein FIG. 7 is a histogram of PCA characteristic variance importance of the gas circuit parameters of the aeroengine, the abscissa is the 21-dimensional characteristics arranged according to the descending order of importance, and the ordinate is the corresponding variance importance. As can be seen from the histogram, the importance of the first two features is 85.9%, the importance of the first three features is 93.6%, and the importance of the first four features is 99.4%. Thus, the top 2 or top 3 or top 4 features may be taken for analysis and failure mode clustering.

The feasibility of the embodiment is verified by taking the first 2 characteristics as an example, a relation graph of each characteristic and time is respectively made, and the relation between each characteristic and the degradation mode is analyzed. Fig. 8a is a graph of the characteristic 1 versus time, fig. 8b is a graph of the characteristic 2 versus time, with the abscissa representing the remaining useful life and the ordinate representing the corresponding characteristic value. It can be seen that feature 1 and feature 2 both cluster well in the time dimension. It can be seen from the relationship between the feature 1 and the feature 2 and time that, for different failure modes of the aircraft engine, the trend and the range of the change of the feature parameters are different to a certain extent, in order to investigate whether the cooperative relationship between the feature 1 and the feature 2 has further correlation with the failure mode, a relationship diagram of the feature 1 and the feature 2 is made to analyze the relationship between the whole feature 1 and the feature 2 and the degradation mode, as shown in fig. 8c, the abscissa is the corresponding value of the feature 1, the ordinate is the corresponding value of the feature 2, and the trajectory represents the evolution of the state of the aircraft engine from factory delivery to failure and reaches the final failure state point. It can be seen that the feature 1 and the feature 2 present two clusters as a whole, and both the feature trend and the final failure point have better clustering results, which corresponds to the fact that the aero-engine in the FD003 data set includes two failure modes, so that it is considered that two different performance degradation paths correspond to performance degradation caused by two different aero-engine failure modes, and two different final failure point clusters correspond to two different engine failure modes. Therefore, a track clustering method can be used for clustering the recession characteristic curves of the engine subsequently, so that the same fault mode can be clustered, and different fault modes can be classified, so that the residual working life of different fault modes can be predicted by adopting a proper method respectively, and the prediction precision is improved.

2) A performance degradation path similarity evaluation method based on dynamic time warping.

For the time series tracks of the characteristic 1 and the characteristic 2 containing most performance degradation information of the aircraft engine, the aim is to perform track clustering on the time series tracks so as to cluster the aircraft engine according to the performance degradation path, which needs to calculate and compare the similarity between the sequences. However, the lengths of two time sequences for comparing similarity are likely to have a certain difference, even the difference degree is large, the individual difference and the fault mode difference between different aero-engines are shown on the aero-engines, so that the service lives of the aero-engines are different, and the test set data does not belong to the full life cycle data like the training set data, and is cut off in the middle of operation, so that the test set data only comprises the time sequences before the operation is cut off, and the points are not in one-to-one correspondence, so that the euclidean distance cannot be simply used for measurement.

The idea of the Dynamic Time Warping (DTW) algorithm is that the time axis of an unknown sequence is scaled to be consistent with the length of a template sequence, so that the similarity between two time sequences is calculated and compared, and the DTW algorithm is very suitable for the time sequences with different rhythm frequencies and sequence lengths.

For an aircraft engine data vector X _a ＝[x _a1 ，x _a2 ，...，x _am ] ^T ，X _b ＝[x _b1 ，x _b2 ，...，x _bn ] ^T (a, b are aircraft engine numbers, m, n are corresponding operating cycle numbers), the DTW method mainly comprises the following steps:

(1) Constructing an m x n matrix grid with matrix elements (i, j) representing x _ai And x _bj Distance d (x) between two points _ai ，x _bj ) Each matrix element (i, j) represents x _ai And x _bj And (4) aligning.

(2) Find a regular path through this trellis and represent by W: w = [ W = ₁ ，w ₂ ，...，w _k ]，max{m，n}≤ k＜m+n-1。

(3) Path boundary condition constraints, i.e. w ₁ ＝(1，1)，w _k = (m, n), ensuring that the selected path must start from the common start point of the sequence and end at the common end point of the sequence.

(4) Path continuity constraint if w _p-1 = (r ', s'), then for the next point w of the path _p-1 = (r, s) needs to meet the requirements that (r-r ') is less than or equal to 1 and (s-s') is less than or equal to 1, and X is ensured _a And X _b Each coordinate in (a) appears in W.

(4) Path monotonicity constraint if w _p-1 = (r ', s'), then for the next point w of the path _p-1 = r, s is required to satisfy (r-r ') > 0 and (s-s') > 0, the point above the constraint W must be monotonic over time.

(5) Repeating (2), (3) and (4) to obtain the path with the minimum regular cost

K in the denominator is used to compensate for the different lengths of the warping paths. The similarity (i.e., the value of the path end point indication) can be obtained by the path with the minimum warping cost.

The similarity of the time series tracks of the characteristic 1 and the characteristic 2 among different aero-engines is calculated through a DTW method, and then clustering of the time series tracks can be achieved through a clustering method, so that the aero-engines are clustered according to the similarity of the performance degradation tracks, clustering results are visually embodied, and recognition of performance degradation paths of the aero-engines is achieved.

3) And (4) performing a performance degradation track class method based on K-Medoids.

The clustering algorithm can automatically assign similar samples to the same category, and eventually data points in the same group should have similar attributes and/or characteristics, while data points in different groups should have highly different attributes and/or characteristics. The K-Means is a simple and rapid clustering technology, the K-Means does not have the inclusion of noise data, and for track clustering, the similarity mean value has no practical meaning, namely, no track corresponds to the similarity mean value, which causes troubles to the identification and classification of the track with the performance degradation of the data of a subsequent test set. The biggest difference between K-Medoids and K-Means is that: the K-Means is to find the average value in the cluster as the centroid, and after finding the average value, the K-Medoids selects an actual point closest to the average value as the centroid, that is, the K-Medoids uses the actually existing optimal point in the data set as its centroid, so that the track corresponding to the centroid can be found, and therefore, the embodiment selects to use the K-Medoids algorithm to complete the subsequent clustering task. The K-medoids algorithm comprises the following steps:

(1) Randomly selecting k samples as initial clustering centers a = [ a ] ₁ ，a ₂ ，...，a _k ]。

(2) For each sample x _i Distances to the k cluster centers (i.e., similarities calculated by the DTW method) are calculated, respectively, and the distances are assigned to the cluster centers having the smallest distances.

(3) Recalculating cluster mean points

Cluster center a _j ＝min(a，a _j ') i.e. the new cluster center is the closest sample point to the cluster mean point.

(4) And (4) circularly repeating the steps (2) and (3) until the conditions of iteration times, minimum error change and the like meet the requirements.

4) And analyzing the clustering result of the performance degradation path of the engine.

The method comprises the steps of reducing dimensions of 21 air path parameters of the aircraft engine by a PCA method to obtain a feature 1 and a feature 2 which have larger variance importance and are related to a performance degradation path, calculating similarity by a DTW method and clustering by a K-Medoids method on time sequence track data of the feature 1 and the feature 2, wherein fig. 9a is a relation graph of the feature 1 and time, fig. 9b is a relation graph of the feature 2 and time, the abscissa is the residual service life, the ordinate is a corresponding feature value, fig. 9c is a relation graph of the feature 1 and the feature 2, the abscissa is a corresponding feature 1 value, and the ordinate is a corresponding feature 2 value. It can be seen that, in the track clustering algorithm used in this embodiment, both in the relationship between the characteristic and the time and in the relationship between the characteristic and the characteristic, the two clusters are presented well without a crossover phenomenon, and both in the characteristic trend and the final failure point, the two clusters have a better clustering result, which corresponds to the FD003 data set of the aero-engine including two failure modes, and the aero-engine is identified and distinguished according to different performance degradation paths, so as to improve the accuracy of the subsequent remaining life prediction.

In summary, since the parameters of the aircraft engine are numerous and there is a large redundancy between the parameters, it is necessary to perform parameter data dimension reduction to extract the degradation feature and then identify the performance degradation path. In the embodiment, the PCA method is mainly used for carrying out dimensionality reduction on 21 gas path parameters of the aircraft engine, the DTW method is used for calculating the similarity between processed characteristic time sequences, and the K-Medoids method is used for clustering performance degradation paths of the aircraft engine to establish a foundation for the subsequent accurate residual service life prediction work.

3. And predicting the residual service life of the aircraft engine in a multi-fault mode.

In view of the strong processing capability of the transform neural network on the sequence data, compared with other network structures, the method has significant advantages in accuracy and efficiency, so the transform neural network is selected on the prediction model, and due to the calculation burden and potential precision influence caused by long time sequence, the embodiment finally selects an inform neural network improved on the basis of the transform neural network as the prediction model of the residual service life of the aircraft engine in the multi-fault mode.

Because the data volume of the aero-engine is large, 100 engines are provided in total, each engine measures a plurality of groups of time sequence data, each group of time sequence data comprises 21 air path parameters, and therefore the information density contained in the data is relatively low. In order to improve the information utilization rate, the accuracy of life prediction is prevented from being influenced due to insufficient information grasp. The method comprises the steps of extracting characteristic data with high information quality and strong performance degradation expression capacity by performing attention extraction on parameters, and then purposefully removing the characteristic data of a low-attention path to obtain the characteristic data with high information density for service life prediction. By the method, the influence of low-quality data on the prediction result can be avoided while the calculation amount is reduced.

1) An Informmer neural network-based residual service life prediction model of the engine under multiple failure modes.

The Informmer model uses a ProbSparse self-attention mechanism, highlights leading attention by extracting a cascade layer, has excellent sequence dependence comparison performance, greatly improves time complexity and memory utilization rate, and greatly improves the calculation speed of long sequence prediction. As shown in fig. 10, the model structure diagram for predicting the remaining service life of the aircraft engine in the multiple failure modes includes a model main body framework divided into four parts, namely a position code, an encoder, a decoder and a full connection layer, and includes the following specific steps:

(1) And establishing two service life prediction models corresponding to the two performance degradation paths. And (3) dividing all the aero-engines into two types according to the performance degradation path, and respectively taking all the time sequence data of the preprocessed aero-engines as model input.

(2) The model input is converted into the input of the encoder after position coding, and the input is transmitted into the encoder, and the encoder is attached with position information.

(3) The encoder part comprises 3 encoder layers, each encoder layer comprises four steps of a multi-head ProbSparse self-attention mechanism, residual error connection and normalization, a feed-forward network, residual error connection and normalization, and finally a connection characteristic diagram (comprising a Query matrix and a Key matrix) is transmitted into a decoder.

(4) And converting a plurality of time sequence data of the preprocessed aircraft engine into the input of a decoder after position coding, and transmitting the input into the decoder to attach position information to the decoder.

(5) Through a Masked multi-head ProbSparse self-attention mechanism, the influence of the post time sequence on the pre time sequence is eliminated, namely, the model only knows data at each moment and before the moment, but not data after the moment, and residual error connection and normalization are carried out.

(6) And performing residual error connection and normalization through feedforward network processing.

(7) And (3) performing residual error connection and normalization by using an attention mechanism of an encoder-decoder, wherein a Value matrix is transmitted in the step (4) in the process, and a Query matrix and a Key matrix are transmitted in the step (2).

(8) And performing residual error connection and normalization through feedforward network processing.

(9) The decoded output is used as the input of the full link layer and is finally converted into the output of the final model, namely the predicted value of the RUL.

(10) And calculating loss by taking the error between the RUL predicted value and the RUL actual value as a loss function, updating the related parameters in the network through back propagation, and repeating training until the loss meets the requirement.

For position coding, in order to increase the operation efficiency, the elements in the input sequence are processed together, rather than one processing as RNN, but this ignores the precedence relationship of the elements in the sequence, so that it is necessary to perform position coding on each element in the sequence, so that the model has the capability of learning the position information of each element in the sequence, and in this embodiment, absolute position coding is used to label the elements in the sequence. For aeroengine gas circuit parameter time series W = { W = { (W) ₁ ，W ₂ ，W ₃ ，...，W _n }， W _i ＝{w ₁ ，w ₂ ，w ₃ ，...，w ₂₁ N is the length of the time series, each W _i The system comprises 21 gas path parameters, namely 21-dimensional vectors. Position-coding W by PE such that the input of the model is X = { W = ₁ +PE ₁ ，W ₂ +PE ₂ ，W ₃ +PE ₃ ，...，W _n + PE _n And (5) calculating the formula of PE, see

formulas

3 and 4.

Where pos is the position of w in the sequence, i is the position of the parameter in w, d _model The dimension for w encoding is 21, and the position encoding step diagram is shown in fig. 11.

The sine function is chosen because for any fixed offset k, PE _pos+k Can be expressed as PE _pos Is a linear function of (a). By carrying out position coding on the time sequence of the gas circuit parameters of the aircraft engine, the model obtains specific data information and obtains position information corresponding to the sequence, thereby improving the accuracy of model prediction.

For the ProbSparse attention mechanism, the core of the transform model is the attention mechanism, the main process is to calculate Q (Query), K (Key), and V (Value) of data, Q corresponds to a sequence to be expressed (called sequence a), K and V correspond to a sequence to express a (called sequence B), the attention calculation formula is shown in formula 5, dk is the square root of the K dimension, and the specific flow chart of the attention calculation is shown in fig. 12.

K and Q must satisfy the preconditions (otherwise no computation is possible) in the same high-dimensional space, whereas V does not necessarily need to be in the same high-dimensional space as K, Q, just as the output of the final model and V need to be in the same high-dimensional space. After each Q and all K dot products in the sequence, the discrete distributions obtained by Softmax are different, and if a Q-derived distribution is similar to a uniform distribution, each probability value approaches the reciprocal of the mean value and has a small value, and belongs to a tail in a long-tail distribution, such Q will not provide any value. Conversely, if a Q results in a distribution that is very different from the uniform distribution, then there must be several probability values that dominate the probability distribution, which is important in the self-attention mechanism, so that such a Q is of value. If the dot product results in the self-attention mechanism are subjected to visualization analysis, it is found that a long tail distribution is obeyed, that is, the dot product calculation results of a few Q and K dominate the distribution after Softmax, and the sparsity distribution means that an element in the sequence generally has higher similarity/relevance with only a few elements.

The core idea of the ProbSparse self-Attention mechanism is to find the important sparse queries, so that only the Attention values of the queries are calculated to optimize the calculation efficiency, and for Q ∈ R ^m×d ，K∈R ^n×d ，V∈R ^n×d The method comprises the following specific steps:

(1) Sampling K to obtain K _ sample and sampling length L _k = mlnn for each q _i e.Q, solving the value M about K _ sample, and calculating a formula shown in formula 6.

(2) Calculating u q with maximum M value _i Form a new matrix as

(3) Calculating out

By means of attention extraction of the parameters, the feature data with high information quality and strong performance degradation expression capability are extracted, and the feature data of the low-attention path is purposefully extracted to obtain the feature data with high information density, so that the accuracy of model prediction is improved.

2) And evaluating the prediction result of the residual service life prediction model.

And inputting the aeroengine gas circuit parameter time sequence data after the performance degradation track identification into an aeroengine residual service life prediction model under the multiple fault modes based on the inform, thereby obtaining the prediction result of the model.

The embodiment selects two indexes of Root Mean Square Error (RMSE) and scoring function (Score) to evaluate and measure the performance of the prediction model, and the lower the evaluation index is, the better the prediction effect is.

RMSE gives the same weight to the prediction errors in early and late stages of fading, and amplifies larger deviations by squaring the errors, and is therefore more sensitive to predicted values that deviate more from the true values,

is a prediction value of RUL, y _i The actual value is n, the total number of the predicted values is n, and the specific formula is shown in formula 7.

Score gives a larger penalty Score to the predicted residual life value exceeding the real residual life value, and inspects the capability of advanced prediction, so that the Score has a larger significance to predictive maintenance, and a specific formula is shown in formula 8.

Fig. 13a is a comparison graph of the predicted result and the true value after the path identification, fig. 13b is a comparison graph of the predicted result and the true value without the path identification, the abscissa is the engine number, the ordinate is the corresponding remaining service life, the solid line is the true value, and the dotted line is the model predicted value. It can be seen that after the performance degradation path is identified, the service life prediction accuracy of the engine is greatly improved compared with that of an unidentified performance degradation path, and table 3 is a prediction result comparison table before and after path identification, and is used for comparing RMSE and Score of prediction results before and after path identification. After the performance degradation path is identified, compared with the service life prediction accuracy before the performance degradation path is not identified, the service life prediction accuracy of the engine is improved by 26.4% (RMSE is used as a standard for measuring precision), and the correlation between the residual service life of the aircraft engine and the performance degradation path of the aircraft engine is proved, so that the identification of the performance degradation path of the aircraft engine plays an important role in prediction of the residual service life, and the accuracy of service life prediction can be effectively improved.

TABLE 3 comparison of predicted results before and after path identification

Table 4 shows the comparison results of the different prediction models in the prediction capability, which are the prediction results after identifying the performance degradation path. As shown in the table, the Informmer model obtains the best Score result and RMSE result, and has obvious advantages compared with other models, and has better prediction capability of the remaining service life of the aircraft engine under multiple failure modes.

TABLE 4 comparison of prediction results for different prediction models

In the embodiment, because the gas circuit parameter data of the aircraft engine has obvious time sequence, and the Informer neural network can effectively learn the development trend of the time sequence data, the model for predicting the remaining service life of the aircraft engine based on the Informer neural network model is established in the embodiment. Experiments show that the residual life prediction model of the engine based on the Informer neural network model can better predict the residual life of the aircraft engine in a multi-fault mode, has obvious advantages compared with other models, and can effectively improve the accuracy of life prediction through data after performance degradation paths are identified.

Thus, when the RUL prediction is carried out on the actual aircraft engine in the multi-fault mode, the model of the embodiment can be adopted to carry out the RUL prediction more accurately and effectively.

In summary, the invention has the following advantages:

(1) The invention provides a high-efficiency autonomous clustering identification method for performance degradation paths of an aircraft engine in a multi-fault mode, which is characterized in that path similarity among a plurality of engines with different degradation rates and different fault initial degrees is evaluated by a DTW (delay tolerant wavelet) method, and a K-Medoids performance degradation path clustering method is introduced on the basis to realize rapid and accurate classification of a large number of aircraft engine degradation paths;

(2) The invention provides an aircraft engine remaining service life prediction model based on an Informer deep neural network, which realizes high-precision service life prediction by comprehensively considering the influence of a performance degradation path on engine remaining service life prediction, and the prediction precision is improved by 26.4% after the performance degradation path is adopted for identification by comparison.

(3) The method takes the fault mode as a consideration factor, identifies the performance degradation path of the aircraft engine, brings the aircraft engines with the same performance degradation path into the same model for training and prediction, and is favorable for improving the residual life prediction precision.

The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, and are not to be construed as limiting the scope of the invention. Any modifications, equivalents and improvements which may occur to those skilled in the art without departing from the scope and spirit of the present invention are intended to be within the scope of the claims.

Claims

1. A method for predicting the service life of an aircraft engine in multiple failure modes is characterized by comprising the following steps:

analyzing complete data of a plurality of aero-engines from operation to failure in a multi-failure mode as a training set, and determining a plurality of clustering categories of performance degradation paths of the aero-engines of the training set;

training the life prediction model of the aero-engine of each cluster type according to the training set data of the aero-engine of each cluster type to obtain the trained life prediction model of each cluster type;

taking data of a plurality of aeroengines which are finished before failure in a multi-fault mode as a test set, testing a trained life prediction model of each cluster type according to the test set data of the aeroengine of each cluster type and the corresponding real residual service life RUL, and taking the life prediction model which passes the test as a practical life prediction model;

determining the attributive clustering category by using the performance degradation path of the aero-engine to be tested, and determining a corresponding practical life prediction model by using the attributive clustering category;

and predicting the RUL of the aero-engine to be tested by using the corresponding practical service life prediction model.

2. The method of claim 1, wherein the analyzing the complete data from run to fault of the plurality of aircraft engines under multiple fault modes as a training set, the determining a plurality of cluster categories for performance degradation paths for aircraft engines of the training set comprises:

preprocessing the training set data to obtain the training set data with dimension influence removed and noise influence reduced;

extracting degradation features related to an aeroengine performance degradation path from the training set data from which the dimensional influence and the noise influence have been removed;

and clustering the time sequence of the degradation characteristics of the training set, which are related to the performance degradation path of the aircraft engine, to obtain a plurality of cluster categories of the performance degradation path of the aircraft engine of the training set.

3. The method of claim 2, wherein preprocessing the training set data to obtain training set data with the dimensional effect removed and the noise effect reduced comprises:

carrying out normalization processing on the training set data to obtain the training set data with dimension influence removed;

and carrying out smooth noise reduction treatment on the training set data with the dimension influence removed to obtain the training set data with the dimension influence removed and the noise influence reduced.

4. The method of claim 2, wherein extracting degradation features associated with an aircraft engine performance degradation path from the training set data with the dimensional effects removed and the noise effects reduced comprises:

carrying out Principal Component Analysis (PCA) processing on the training set data with the dimension influence removed and the noise influence reduced to obtain a plurality of PCA characteristics of the training set data;

and selecting a specified number of PCA features from the plurality of PCA features as degradation features related to the aircraft engine performance degradation path according to the sequence of the PCA feature variance importance degrees from large to small.

5. The method of claim 2, wherein the clustering the time series of degradation features of the training set that are associated with aircraft engine performance degradation paths to obtain a plurality of cluster categories for the aircraft engine performance degradation paths of the training set comprises:

determining similarity between time sequences of degradation characteristics of different aero-engines of the training set by using a Dynamic Time Warping (DTW) algorithm;

and clustering the time sequences of the degradation features of different aero-engines according to the similarity between the time sequences of the degradation features of different aero-engines of the training set to obtain a plurality of clustering categories of the performance degradation paths of the aero-engines of the training set.

6. The method according to any one of claims 2-5, further comprising:

and analyzing the test set, and determining the cluster category to which the performance degradation path of the aero-engine of the test set belongs.

7. The method of claim 6, wherein analyzing the test set to determine a cluster class to which a performance degradation path for an aircraft engine of the test set belongs comprises:

preprocessing the test set data to obtain the test set data with dimension influence removed and noise influence reduced;

extracting a time series of degradation features of the aircraft engine of the test set from the test set data from which dimensional effects have been removed and from which noise effects have been reduced;

and comparing the distance between the time series of the degradation characteristics of the aero-engine of the test set and the time series of the cluster center of each cluster category, and determining the cluster category of the cluster center with the minimum distance as the cluster category to which the performance degradation path of the aero-engine of the test set belongs.

8. The method according to any one of claims 2 to 5, wherein the determining the cluster category to which the aircraft engine to be tested belongs by using the performance degradation path of the aircraft engine to be tested comprises:

preprocessing all currently obtained data of the aero-engine to be tested to obtain data with dimension influences removed and noise influences reduced;

extracting a time sequence of degradation characteristics of the aero-engine to be tested from the data of the aero-engine to be tested, wherein the data are subjected to dimensional influence removal and noise influence reduction;

and comparing the distance between the time sequence of the degradation characteristics of the aero-engine to be tested and the time sequence of the clustering center of each clustering category, and determining the clustering category of the clustering center with the minimum distance as the clustering category to which the performance degradation path of the aero-engine to be tested belongs.

9. The method as claimed in claim 1, wherein the life prediction model is an aircraft engine remaining service life prediction model under multiple faults based on an Informer neural network.