Multi-fault-mode engine residual life prediction method and device and electronic equipment

Abstract

The invention discloses a multi-fault-mode engine residual life prediction method and device and electronic equipment, and the method comprises the steps: obtaining health evaluation time sequence curves of a training engine in different fault modes, and selecting the health evaluation time sequence curves of the training engine in two fault modes; constructing an SN-GRU model composed of two Siamese networks and a GRU network, and training the SN-GRU model by using the health assessment time sequence curve of the training engine in the two fault modes to obtain a trained SN-GRU model; and performing residual life prediction processing on the tested engine based on the trained SN-GRU model.

Description

Technical Field

[0001] The present invention relates to the technical field of life prediction, and in particular to a method, device and electronic equipment for predicting the remaining life of a multi-fault mode engine. Background Art

[0002] Aircraft engines are a key part of aircraft. They usually operate under harsh environmental conditions such as high pressure and high-speed rotation, and their operating profiles and failure results are complex and changeable. Long-term service will lead to performance degradation of components and even the entire aircraft. If the engine health status cannot be evaluated in time and the remaining life of the engine cannot be accurately predicted in advance, the reliability and safety of the aircraft will be affected.

[0003] In recent years, the problem of remaining life prediction of equipment under multiple fault modes has gradually attracted the attention of researchers. In order to deal with the problem of RUL prediction under multiple fault modes of aircraft engines, a two-step approach is usually adopted. The monitoring data is classified into groups for each mode based on the diagnostic model, and then several corresponding prediction models are independently developed for each fault mode data. Yan et al. proposed a spectral correlation density combination method for fault type and scale diagnosis, and applied the semi-supervised co-training method to RUL prediction. Chehade et al. first estimated the fault mode based on the calculated posterior probability, then constructed a Bayesian degradation model, and determined the fault threshold in the RUL prediction of each mode. Li et al. constructed a deep branch network (DBNet) for RUL prediction under multiple fault modes based on the multi-task learning (MTL) method, and learned the fault mode classification and RUL prediction in DBNet in a sequential manner.

[0004] Based on the above summary, the research on the remaining life prediction of aircraft engines under multiple operating conditions and multiple fault modes has gradually attracted people's attention. However, a complete solution has not yet been formed for the problem of remaining life prediction under variable engine operating conditions and multiple fault mode combinations. Summary of the invention

[0005] The present invention provides a method, device and electronic equipment for predicting the remaining life of an engine with multiple fault modes, so as to solve the technical problems that the health assessment curves of the engine under multiple fault modes are different, and the prediction efficiency and model generalization ability of different life prediction models are low.

[0006] An embodiment of the present invention provides a method for predicting the remaining life of an engine with multiple fault modes, comprising:

[0007] Obtain health assessment timing curves of training engines under different fault modes, and select health assessment timing curves of training engines under two fault modes;

[0008] Constructing an SN-GRU model consisting of two Siamese networks and one GRU network, and training the SN-GRU model using the health assessment timing curves of the training engines under the two failure modes to obtain a trained SN-GRU model;

[0009] The remaining life prediction process of the test engine is performed based on the trained SN-GRU model.

[0010] The embodiment of the present invention also provides a device for predicting the remaining life of a multi-failure mode engine, comprising:

[0011] An acquisition module is used to acquire the health assessment timing curves of the training engine under different fault modes, and select the health assessment timing curves of the training engine under two fault modes;

[0012] A construction and training module, used to construct an SN-GRU model consisting of two Siamese networks and one GRU network, and train the SN-GRU model using the health assessment timing curves of the training engines under the two failure modes to obtain a trained SN-GRU model;

[0013] The prediction module is used to perform remaining life prediction processing on the test engine based on the trained SN-GRU model.

[0014] An embodiment of the present invention 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 method for predicting the remaining life of a multi-fault mode engine described in any one of claims 1 to 8 are implemented.

[0015] The beneficial effect of the present invention is that deep metric learning is introduced into the life prediction framework. By designing the corresponding relationship of "distribution difference-remaining life difference", the nonlinear relationship between the time series distribution difference and the remaining life difference between the metric health assessment indicators is learned, and the gap between multiple fault modes is reduced by learning the "difference" law, so that the prediction model is applicable to multiple fault mode situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of a method for predicting the remaining life of a multi-fault mode engine provided by an embodiment of the present invention;

[0017] Figure 2 It is the technical roadmap of the present invention for predicting the life of aircraft engines under complex operating conditions based on metric learning;

[0018] Figure 3 It is a flow chart of a method for adaptively identifying engine operating conditions by combining clustering algorithm with metric learning;

[0019] Figure 4 It is a flow chart of the method for equivalent transformation of variable working condition data;

[0020] Figure 5 It is a curve diagram representing the degradation of variable working conditions;

[0021] Figure 6 It is a flow chart of a fault mode identification method based on variable weight prediction;

[0022] Figure 7 It is a flow chart for constructing a single-moment fault mode classification model;

[0023] Figure 8 This is the structure diagram of the LSTMCNN_AE network;

[0024] Fig. 9 This is the structural details of the encoder and decoder in the LSTMCNN_AE network;

[0025] Fig.10 This is the engine timing degradation feature extraction flow chart based on LSTMCNN_AE;

[0026] Fig.11 This is a schematic diagram of the LSTMCNN_AE model input sample generation process;

[0027] Fig.12 It is a flow chart of engine health assessment for multiple fault modes and variable operating conditions based on NCA metric learning;

[0028] Fig.13 This is a schematic diagram of the Siamese network structure;

[0029] Fig.14 It is a structural diagram of the SN-GRU network model;

[0030] Fig.15 It is a schematic diagram of the SN-GRU network structure;

[0031] Fig.16 It is a flowchart of engine life prediction under multiple fault modes based on SN-GRU. DETAILED DESCRIPTION

[0032] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In the subsequent description, the use of suffixes such as "module", "component" or "unit" used to represent elements is only to facilitate the description of the present invention and has no specific meaning in itself. Therefore, "module", "component" or "unit" can be used in a mixed manner.

[0033] The present invention introduces deep metric learning into the life prediction framework, and by designing the corresponding relationship of "distribution difference-remaining life difference", learns the nonlinear relationship between the time series distribution difference and the remaining life difference between the metric health assessment indicators, and reduces the gap between multiple fault modes by learning the "difference" law, so that the prediction model can be applied to multiple fault mode situations. At the same time, in order to solve the problem of insufficient training samples and the network being susceptible to overfitting, a "direct + indirect" combined life estimation method is designed, which can combine all training samples for estimation, thereby achieving robust and accurate prediction of engine life and improving the generalization of the prediction model under multiple fault modes. The classic deep metric learning Siamese network and the recurrent neural network GRU combination strategy are used to predict the remaining life based on the engine health assessment results obtained under different fault modes.

[0034] Figure 1 is a flowchart of a method for predicting the remaining life of a multi-fault mode engine provided by the present invention, such as Figure 1 As shown, it includes: S101: obtaining the health assessment timing curves of the training engine under different fault modes, and selecting the health assessment timing curves of the training engine under two fault modes; S102: constructing an SN-GRU model composed of two Siamese networks and one GRU network, and using the health assessment timing curves of the training engine under the two fault modes to train the SN-GRU model to obtain a trained SN-GRU model; S103: performing life prediction processing on the test engine based on the trained SN-GRU model.

[0035] In an embodiment of the present invention, the Siamese network in the SN-GRU model consists of a convolutional layer, a flat layer and a fully connected layer, and the GRU network in the SN-GRU model consists of a GRU layer, a dropout layer and a fully connected layer.

[0036] In an embodiment of the present invention, the SN-GRU model is trained using the health assessment timing curves of the training engine under the two fault modes to obtain a trained SN-GRU model, including: obtaining health indicator samples under the first fault mode and health indicator samples under the second fault mode by respectively performing sliding window cutting processing on the health assessment timing curves of the training engine under the two fault modes; obtaining the remaining life corresponding to the health indicator samples under the first fault mode and the remaining life corresponding to the health indicator samples under the second fault mode; in each round of iteration, the health indicator samples under the first fault mode and the health indicator samples under the second fault mode are respectively used as inputs of the Siamese network, and the remaining life corresponding to the health indicator samples under the first fault mode and the remaining life corresponding to the health indicator samples under the second fault mode are used as outputs of the GRU prediction network, and the hyperparameters of the Siamese network and the GRU prediction network are updated by the gradient descent algorithm to obtain a trained SN-GRU model.

[0037] In an embodiment of the present invention, the remaining life prediction process of the test engine based on the trained SN-GRU model includes: obtaining health indicator samples of the test engine under k failure modes The health indicator samples are predicted by the Siamese network and the GRU prediction network in the trained SN-GRU model. After processing, the direct predicted RUL of the test engine is obtained; the formula is:

[0038]

[0039] in, 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.

[0040] In the embodiment of the present invention, the remaining life prediction processing of the test engine based on the trained SN-GRU model further includes: obtaining health indicator samples of the test engine under k failure modes The second health indicator sample and its corresponding remaining life span (CV j ,RUL j ); Based on the trained SN-GRU model, according to the difference between the health indicator sample and the second health indicator sample in the Siamese network, an indirect predicted RUL of the test engine is obtained.

[0041] In an embodiment of the present invention, based on the trained SN-GRU model, according to the difference between the health indicator sample and the second health indicator sample in the Siamese network, obtaining the indirect predicted RUL of the test engine includes:

[0042]

[0043] in, is the RUL estimated indirectly based on the SN-GRU model, and |D| is the total number of estimated references.

[0044] In an embodiment of the present invention, the remaining life prediction processing of the test engine based on the trained SN-GRU model further includes: fusing the direct predicted RUL and the indirect predicted RUL of the test engine to obtain the predicted RUL of the test engine under k failure modes; the formula is:

[0045]

[0046] Among them, α and β are fusion parameters respectively.

[0047] In the embodiment of the present invention, the life prediction processing of the test engine based on the trained SN-GRU model further includes: k Add them together to get the predicted life L of the engine tested under k failure modes k , the formula is:

[0048] L k =RUL k +cycle k

[0049] An embodiment of the present invention also provides a device for predicting the remaining life of an engine with multiple fault modes, comprising: an acquisition module, used to obtain the health assessment timing curves of the training engine under different fault modes, and select the health assessment timing curves of the training engine under two fault modes; a construction and training module, used to construct an SN-GRU model composed of two Siamese networks and one GRU network, and train the SN-GRU model using the health assessment timing curves of the training engine under the two fault modes to obtain a trained SN-GRU model; and a prediction module, used to perform remaining life prediction processing on the test engine based on the trained SN-GRU model.

[0050] An embodiment of the present invention also provides an electronic device, characterized in that it includes 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 method for predicting the remaining life of a multi-fault mode engine are implemented.

[0051] There are many technical bottlenecks that need to be solved in the life prediction of aircraft engines under variable operating conditions and multiple fault modes: difficulty in identifying operating conditions and fault modes under complex operating conditions, difficulty in extracting time series degradation features under variable operating conditions, difficulty in quantifying engine degradation trends under variable operating conditions, and decreased accuracy of engine life prediction under multiple fault modes. In response to the above problems, a life prediction method is proposed. Please refer to the overall life prediction process and technical route. Figure 2 , which specifically includes three parts: engine operating condition division and fault mode recognition based on improved clustering method, variable operating condition engine health assessment based on improved autoencoder and metric learning, and multi-fault mode engine remaining life prediction based on time series distribution metric and domain generalization.

[0052] Before completing the remaining life prediction of the multi-fault mode engine, the present invention also includes: variable working condition engine health assessment based on improved autoencoder and metric learning and engine working condition division and fault mode identification based on improved clustering method, which are specifically as follows:

[0053] 1. Engine operating condition division and fault mode identification based on improved clustering method

[0054] The invention discloses a method for identifying fault patterns of engine groups, comprising: performing dimensionality reduction processing on training engine performance parameters and 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; in predicting the single-moment fault pattern of the engine, performing fault pattern clustering processing on the low-dimensional performance parameters of the training engine based on a density clustering (DBSCAN) algorithm to obtain fault pattern classification labels corresponding to each training engine; establishing a single-moment fault pattern classification model based on a support vector machine (SVM), and training the single-moment fault pattern classification model based on the SVM using the low-dimensional performance parameters of the training engine and the fault pattern classification labels corresponding to each training engine to obtain a trained single-moment fault pattern classification model based on the SVM; obtaining single-moment fault pattern classification labels of each test engine by inputting the low-dimensional performance parameters of the test engine into the trained single-moment fault pattern classification model based on the SVM; and obtaining final fault pattern classification labels of each test engine based on weighted prediction of multi-moment fault patterns and the single-moment fault pattern classification labels of each test engine.

[0055] The method of obtaining the final fault mode classification label of each test engine based on the weighted prediction of multi-moment fault modes and the single-moment fault mode classification label of each test engine includes: obtaining the prediction result of the fault mode of each test engine at a single moment according to the trained SVM-based single-moment fault mode classification model; performing weighted prediction using the fault state of each test engine at a single moment and the comprehensive multi-moment fault state to obtain the weighted prediction result of each test engine under different fault modes; and obtaining the final fault mode classification label of each test engine using the weighted prediction results of each test engine under different fault modes.

[0056] 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, including:

[0057]

[0058] 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 include:

[0059]

[0060] Among them, m represents the fault mode; j represents the test engine; τ represents the recent weighting coefficient of the recognition proportion of recent data; λ represents the comprehensive weighting coefficient of the number of flight cycles considered comprehensively; k represents the value coefficient affecting the weight of the recent and long-term, and represents the flight cycle at the most recent moment; represents the predicted state of the 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 by weighted prediction.

[0061] The method of using the weighted prediction results of each test engine under different failure modes to obtain the final failure mode classification label of each test engine includes:

[0062]

[0063] Among them, Y j represents the final failure mode classification label of test engine j.

[0064] Before the training engine performance parameters and the test engine performance parameters are subjected to dimensionality reduction processing based on principal component analysis PCA, the method also includes: screening out the performance parameters of the engine that are sensitive to decay under variable conditions; extracting the engine data under different conditions based on the results of the division of the operating conditions of the individual engine, and obtaining the decay-sensitive performance parameters under each condition; normalizing the decay-sensitive performance parameters under each condition using a normalization method, and reconstructing the normalized decay-sensitive performance parameters under each condition according to the flight cycle sequence of the engine, and obtaining the decay-sensitive performance parameters under the fusion of multiple conditions; and dividing the decay-sensitive performance parameters under the fusion of multiple conditions into the training engine performance parameters and the test engine performance parameters.

[0065] Screening out the performance parameters of the engine that are sensitive to degradation under variable operating conditions also includes: obtaining the multidimensional operating parameters of the training engine under different operating conditions; based on the obtained multidimensional operating parameters, using the K-Means clustering algorithm to preliminarily identify the operating condition category of the training engine, and obtain the cluster center of each operating condition; using the multidimensional scaling analysis MDS algorithm to scale the multidimensional operating parameters of the test engine and the cluster center of the corresponding operating condition to one dimension, and using the one-dimensional operating condition cluster center as the operating condition baseline; based on the principle of being closest to the operating condition baseline, adaptively identifying the operating condition corresponding to each test engine in each flight cycle.

[0066] 1.1 Engine operating condition classification based on metric learning and clustering

[0067] In the identification of engine operating conditions, it is necessary to cluster the multi-dimensional operating parameters to obtain the operating condition classification results. Clustering methods are usually used to identify training sets and test sets, but since the test engine data cannot be predicted, it is difficult to ensure that the results of the direct use of clustering methods correspond to the classification categories of the training set.

[0068] In order to accurately determine the operating condition label at a single moment, the present invention proposes an improved clustering method to identify the engine operating condition, that is, to introduce the unsupervised metric learning MDS algorithm into the K-Means clustering method. The core of MDS is to use high-dimensional data to calculate the distribution of data in low dimensions, that is, to reconstruct the low-dimensional coordinates of the data, so as to efficiently measure the true distance between the data. The MDS algorithm can project multi-dimensional operating parameters into a low-dimensional space while keeping the relative position unchanged. The test engine can automatically identify and classify it into the nearest operating condition category, thereby adaptively realizing the engine's operating condition classification at a single moment. 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 cluster center, and overcome the problem of inconsistent Euclidean distance scales, realize the classification of test engine operating conditions, and solve the problem of inaccurate clustering of test set operating conditions.

[0069] In the engine operating condition identification, 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 training engine operating condition categories, and obtain the cluster center of each operating condition. Then, the MDS algorithm is used to scale the test engine operating parameters and the operating condition cluster center to one dimension, and the operating condition corresponding to each flight cycle of the test engine is adaptively identified based on the principle of the closest distance to the operating condition baseline. The specific process of adaptive identification of engine operating conditions based on the combination of metric learning and clustering is as follows: Figure 3 As shown. The method of introducing the MDS algorithm into the K-Means clustering algorithm to adaptively realize the classification of the engine single-time working condition category includes the following steps: S111, obtaining the multi-dimensional operating parameters of the training engine under different working conditions; S112, using the K-Means clustering algorithm to preliminarily identify the working condition category of the training engine according to the multi-dimensional operating parameters obtained in step S111, and obtaining the cluster center of each working condition; S113, using the MDS algorithm to scale the multi-dimensional operating parameters of the test engine and the cluster center of the corresponding working condition to one dimension; S114, adaptively identifying the operating condition corresponding to the test engine in each flight cycle based on the principle of being closest to the working condition baseline.

[0070] The pseudo code for adaptive recognition of engine operating conditions based on metric learning and clustering is as follows:

[0071]

[0072]

[0073] The process of adaptive recognition of engine operating conditions based on the combination of metric learning and clustering is as follows:

[0074] (101) Engine operating parameter extraction

[0075] The data of aircraft engines under complex operating conditions include operating parameters and gas path parameters. The operating parameter data is extracted as the data basis for engine operating condition identification. The operating parameters of the engine are constantly changing, indicating that the operating conditions of the engine are constantly changing during operation. The combination of operating parameters brings different operating conditions to the aircraft. When the values ​​of the three operating parameters are basically the same and within a very small fluctuation range, it can be indicated that the engine operating conditions are the same. The operating conditions of aircraft engines usually change randomly, and it is impossible to find their changing patterns, which also brings certain difficulties to the subsequent engine health assessment under variable conditions.

[0076] (102) K-Means-based training engine cluster center determination and operating condition label division

[0077] The divided training engine operating parameters are used as the multi-dimensional input of K-Means clustering. According to the number of aviation engine operating condition categories k, k operating condition parameter combination points are randomly selected as the initial cluster centers. In the training 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 training engine flight cycle and the final cluster center result.

[0078] (103) Adaptive identification of test engine operating conditions based on MDS metrics

[0079] The operating parameters of the test engine are merged with the operating condition clustering center, and the multidimensional data is scaled to one dimension using the multidimensional scaling MDS algorithm. The one-dimensional operating condition center is used as the baseline. The operating data after single-time scaling 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 in each flight cycle.

[0080] 1.2 Engine data fusion based on equivalent transformation of operating conditions

[0081] Aircraft engine gas path parameters can serve as an important basis for health assessment and life prediction, but the data of gas path parameters fluctuate up and down under the influence of operating conditions, making it difficult to characterize the true degradation trend of the engine. Usually among aircraft engine gas path parameters, not all gas path parameters are useful for studying the engine status, and some parameters provide constant or discrete values ​​throughout the life cycle of the engine. In order to restore the true degradation trend of the engine under variable operating conditions, it is necessary to unify the performance parameters under different operating conditions to a comparable parameter standard. The present invention proposes a method for equivalent transformation of variable operating condition engine data, which mainly consists of four parts: screening of decay performance parameters, decomposition of multi-operating condition time series data, standardization of performance parameters of each operating condition, and reconstruction of performance parameters. The process of equivalent transformation of variable operating condition data is as follows: Figure 4 shown.

[0082] (1001) Screening of decay-sensitive gas path parameters under variable operating conditions

[0083] In the screening of degradation performance parameters, by drawing the engine life curve under different working conditions, the gas path parameters that are constant and have no obvious degradation trend are initially excluded, so as to preliminarily screen out the performance parameters that have obvious degradation trend with working conditions, small fluctuation, and are more sensitive to engine degradation, and form a Figure 4 The performance parameter sequence under variable operating conditions shown in .

[0084] (1002) Multi-operating condition time series data decomposition

[0085] In the decomposition of multi-condition time series data, based on the results of the division of individual engine operating conditions, the engine data under different conditions are extracted to obtain the decay-sensitive performance parameters under each condition, thereby realizing the extraction of feature data under variable conditions. Based on the condition identification label based on the improved clustering method, the engine time series data is decomposed based on the condition by matching the engine flight cycle with the condition. At this time, the various gas path parameters of the engine can be divided into time-discontinuous data under different conditions, but under the same condition, the gas path parameters show a relatively uniform degradation trend, and the fluctuation is smaller than before the data decomposition, thus laying the foundation for parameter standardization and reconstruction.

[0086] (1003) Standardization of performance parameters for each operating condition

[0087] Aircraft engines face frequent changes in operating conditions during operation. Under the same operating conditions, the degradation parameters of engine performance usually show a gradual change trend with small fluctuations. However, when the operating conditions change frequently, the degradation trend is difficult to show.

[0088] To achieve the goal of parameter unification and degradation trend integration, two operations need to be taken: first, for engine parameters, convert them into unified indicators that can reflect degradation trends, scale the data under different working conditions, and unify them into the standard space; second, for degradation trends, form the standardized data under different working conditions into a curve that can clearly characterize the degradation of variable working conditions, such as Figure 5 shown.

[0089] The normalization method is used to normalize the degradation sensitivity performance parameters under different working conditions, so as to achieve the parameter normalization of different working conditions. i At the moment, the decay sensitivity performance parameter is x i , the corresponding working condition is m i , its state parameter can be expressed as (t i ,x i ,m i ),i=1,2,..,n. The engine operating conditions include k types, and its operating condition category set can be expressed as M={M1,M2,...,M k}, therefore, the total set of performance parameters under each working condition can be expressed as {X} (K) ={x i |m i ∈M}.

[0090] Z-score normalization establishes a standardized index by obtaining the mean and standard deviation of the data. This method 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 working conditions can be expressed as and k, then the data normalization formula under working condition k is:

[0091]

[0092] The z-score method uses the mean and standard deviation as the normalization benchmark and has less impact on local fluctuations and noise. Given that aircraft engine parameters usually have greater noise interference, it can more effectively deal with this problem.

[0093] (1004) Gas path performance parameter reconstruction

[0094] Finally, according to the flight cycle sequence of the engine, the normalized performance data is reconstructed to merge the normalized characteristic parameters under different operating conditions and reconstruct the degradation time series integrating multiple operating conditions to characterize the degradation of the engine under variable operating conditions. The reconstructed parameter data can be expressed as:

[0095] X'={x1',x'2,...,x' n} (6)

[0096] Among them, X' is the aerodynamic performance degradation parameter of the engine under the reconstructed variable working condition, x' t is the N-dimensional performance parameter under flight cycle t, t=1,2,…,n.

[0097] 1.3 Engine Fault Pattern Recognition Based on Classification and Variable Weight Prediction

[0098] In fault mode recognition, it is necessary to divide the fault mode according to the degradation path of the engine, taking into account the certain volatility of engine parameters, and K-Means does not have the inclusiveness of noise data, and for trajectory clustering, the similarity mean has no practical meaning, which will cause trouble for the subsequent test set data performance degradation trajectory recognition and classification. The density-based clustering algorithm is not affected by the up and down fluctuations of parameters and is suitable for fault mode recognition in research. Therefore, the present invention designs a weighted fault mode discrimination method based on density clustering-support vector machine (DBSCAN-SVM) and multi-time prediction, identifies the engine fault mode through the density clustering method, and adopts SVM to discriminate the single-time fault state classification of the test unit, and finally weightedly considers the multi-time fault state to discriminate the test unit fault mode.

[0099] In the fault mode recognition method based on variable weight prediction, PCA dimensionality reduction is first performed based on the screened gas path performance degradation parameters to preliminarily obtain low-dimensional features for training the clustering of precise fault modes of the engine; then, a single-moment fault mode classification model based on SVM is established on the basis of density clustering, and a fault mode weighted prediction method is designed to complete the fault mode recognition of the test engine.

[0100] Engine fault pattern recognition based on variable weight prediction is mainly divided into two parts, namely, the establishment of a single-time fault pattern classification model and the weighted prediction of multi-time fault patterns. The main process is as follows Figure 6 shown.

[0101] The pseudo code of engine failure mode recognition based on variable weight prediction is as follows:

[0102]

[0103]

[0104] 301) Constructing a single-moment fault mode classification model

[0105] The construction of the single-moment fault mode classification model is mainly divided into three parts: parameter dimension reduction based on principal component analysis, degradation trajectory clustering based on DBSCAN, and single-moment fault mode classification model construction based on SVM. The main process is as follows Figure 7 shown.

[0106] (301) Parameter Dimensionality Reduction Based on Principal Component Analysis

[0107] After the equivalent transformation of the working conditions, the gas path performance parameters are still multidimensional, and there may be correlation between the parameters, which makes the trend of the failure mode difficult and complex. If the performance parameters are analyzed separately, the efficiency will be greatly reduced, and the problem will be viewed in isolation rather than comprehensively. Therefore, considering the correlation between the performance parameters, the PCA method in unsupervised metric learning is used for dimensionality reduction, which can obtain unrelated low-dimensional parameter information without losing the original parameter information.

[0108] PCA is a nonparametric dimensionality reduction technique that aims to transform numerous variables in a high-dimensional dataset into fewer new variables while retaining the key information of the data. Its core principle is to linearly transform the data in unrelated directions by calculating the eigenvectors of the covariance matrix. These new dimensions are sorted by 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 characteristics of the data.

[0109] (302) Clustering of Degraded Trajectories Based on DBSCAN

[0110] The training engine monitors the data from healthy to decayed. Considering that the degradation characteristics are often not obvious in the early stage of engine operation, the late decay data is selected to identify its fault mode category. After extracting the low-dimensional gas path parameter characteristics of the engine by PCA method, the DBSCAN method is used to divide the late gas path parameter characteristics of the training engine, and the corresponding fault mode of each training engine is obtained.

[0111] After the fault mode clustering based on the DBSCAN algorithm, assuming that there are M types of fault modes, the characteristic matrix of the mth type of fault after clustering is:

[0112]

[0113] Among them, Z m N m ×l matrix, N m is the number of characteristic samples of the mth type of fault mode, and l is the dimension of the fault feature. m The corresponding failure mode classification label can be expressed as Y m =[mm … m] T .

[0114] (303) Construction of single moment fault mode classification model based on SVM

[0115] The characteristic matrix of each type of fault mode is used as the input of the fault mode classifier, and the corresponding fault mode label is used as the output of the classifier to train the fault 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 number of samples. Since the sample size of aircraft engine fault modes is small, the present invention uses SVM to construct a single-moment fault mode classification model to establish a foundation for testing engine fault mode classification.

[0116] SVM is a supervised learning classification algorithm, and 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 from the data point closest to the hyperplane to the plane, and 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 of different types of fault modes can be expressed as:

[0117]

[0118] 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 fault mode and the j-th fault mode. For the test sample x test , a single SVM model belongs to the category The judgment is:

[0119]

[0120] Among them, if If the value is 1, it means that the test sample is in the i-th failure mode, otherwise it is in the j-th failure mode.

[0121] 302) Multi-time Fault Mode Weighted Prediction

[0122] Since the performance degradation of the engine is a gradual process, the fault characteristics of the test engine that has not completely degraded may not be significant enough, and it is difficult to accurately identify it through the monitoring parameters at a single moment. Therefore, a weighted prediction method is introduced to determine the final fault mode of the engine by comprehensively considering the fault mode characteristics at multiple moments. In the process of fault mode identification, the dynamic changes in the engine degradation process are more comprehensively considered, which makes up for the defects caused by the fault mode classification at a single moment.

[0123] 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:

[0124]

[0125] Afterwards, weighted prediction is performed based on the comprehensive multi-time fault status, and the weighted prediction results of the test engine j under different fault modes can be calculated as:

[0126]

[0127] The present invention introduces the recent weighting coefficient τ (0 < τ < 1) and the comprehensive weighting coefficient λ (λ, t) to analyze the engine degradation trajectory, where τ represents the recognition weight of recent data and λ represents the number of flight cycles considered comprehensively. The final engine failure mode can be calculated as The failure mode with the highest probability is selected as the final failure mode category of the engine.

[0128] The calculation method of τ is: take a number between (0,1) as the weight according to the recent data requirements. In this case, 0.7 is taken as the recent weight coefficient.

[0129] The calculation method of λ is: take 70% of the flight cycles of the jth engine.

[0130] 2. Engine Health Assessment under Variable Operating Conditions Based on Improved Autoencoder and Metric Learning

[0131] The variable operating condition engine health assessment method comprises: 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 fault mode, and mapping the low-dimensional time series degradation feature data using the trained NCA metric learning model under each fault 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.

[0132] The LSTMCNN_AE feature extraction model is trained using engine training set data to obtain a trained LSTMCNN_AE feature extraction model, including: preprocessing multidimensional engine parameter data under different working conditions using a sliding window of set length and step size, and dividing the multidimensional engine parameter data into training samples and test samples according to a sliding window method; using the training samples as model input and output, and reconstructing the input sequence through the LSTMCNN_AE feature extraction model; capturing feature information based on the training data set, determining 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.

[0133] Constructing and training the NCA metric learning model under each fault mode includes: dividing the low-dimensional timing degradation feature data into timing degradation feature data under different fault modes according to the identified engine fault mode label; under each fault mode, dividing the timing degradation feature data into healthy set feature data and degraded set feature data, and marking the corresponding healthy set state label and degraded set state label respectively; constructing the NCA metric learning model under each fault mode, and taking the healthy set features and degraded set features under each fault mode as the model input, and taking the state label corresponding to the feature as the model output for training, so as to obtain the trained NCA metric learning model and the optimized Mahalanobis distance feature matrix under each fault mode.

[0134] The engine health assessment processing based on the optimized Mahalanobis distance feature matrix and the mapping feature data includes: calculating the distance between the healthy space and the degraded space of each operating condition of the engine based on the optimized Mahalanobis distance feature matrix and the mapping feature data; normalizing the distance under each operating condition by a linear normalization method to obtain the health assessment index of the engine under each operating condition; reconstructing the health assessment index of the engine under each operating condition into a health assessment curve under a time series condition based on the rearrangement of the flight cycle; traversing the assessment under all fault modes to form a health assessment result under each fault mode.

[0135] In the engine health assessment under multiple fault modes and variable operating conditions, two problems will be faced: First, it is difficult to extract time-series degradation features under variable operating conditions. The actual monitoring parameters of the engine are high-dimensional, time-series, and noisy. Some feature extraction methods mainly focus on dimensionality reduction, while ignoring the time-series and high noise of engine data, which makes it difficult to obtain sufficient degradation information for health assessment. In response to this problem, the present invention proposes a new autoencoder feature extraction method, which not only utilizes the characteristics of autoencoder dimensionality reduction, but also can capture time series information and reduce the impact of data fluctuations. Secondly, it is difficult to select the space for health indicator construction. Due to the differences in the initial health status of individual engines and the inconsistency of the operating conditions during the operation process, the existing methods that only rely on the absolute amplitude of the degradation feature without considering the difference between the initial and current moments are not applicable to the problem of variable operating condition health assessment, and for engine data with unknown specific distribution, it is difficult to use distribution measurement methods to establish degradation indicators. To address this problem, the present invention introduces the metric learning method into health assessment and adopts the neighbor component analysis (NCA) algorithm to measure the difference between the degradation characteristics and the initial health characteristics under different working conditions through neighbor category division and Mahalanobis distance transformation. Then, the working condition transformation idea is used to obtain the health assessment value under variable working conditions, thereby completing the degradation trend characterization of individual engines.

[0136] 2.1 Introduction to the Principle of NCA Algorithm

[0137] The KNN method is a supervised nonparametric machine learning method. By selecting the K nearest neighbors of a sample, the sample is classified into the class that appears most frequently among the neighbors. Therefore, choosing a suitable distance metric is a key issue in the KNN method. The NCA algorithm is a metric learning method associated with KNN. The neighbors in the algorithm are similar samples. The concept of "neighbor" is an important concept in metric learning. The NCA algorithm randomly calculates the nearest neighbors and obtains the transformation matrix in the Mahalanobis distance through the leave-one-out cross-validation. The method mainly includes three key points: NCA is supervised learning based on the KNN classifier; NCA measures the similarity of neighboring samples based on the distance metric method; NCA uses the cross-validation method to adjust parameters during training.

[0138] Assume that the input sample is {x1,x2,…,x n}, and its corresponding category labels are c1, c2, …, c n For the sample labels to be learned, we need to find a distance measure that optimizes the nearest neighbor classification. For the Mahalanobis distance that can be learned in metric learning, its transformation matrix Q = A T A is a symmetric semi-positive definite matrix, then the Mahalanobis distance between two samples can be expressed as:

[0139]

[0140] Among them, the semi-positive definite matrix Q is a learnable parameter, called a metric. The goal of metric learning is to optimize the matrix Q through some constraints to learn a distance metric.

[0141] When calculating the classification error by leaving one out, a differentiable softmax function is introduced because the error function is not continuous. Therefore, the neighbor distribution of the sample is defined as:

[0142]

[0143] Among them, p ij Represents data point x i In the random neighbor, choose another data point x instead j The probability of the corresponding category. The probability of the sample point selecting the correct classification is:

[0144]

[0145] The objective function of optimization is to maximize the number of correctly classified points, so the optimization objective can be expressed as:

[0146]

[0147] 2.2 Engine timing degradation feature extraction based on LSTMCNN_AE

[0148] Autoencoder is one of the representative feature extraction methods. Compared with other feature extraction methods, autoencoder can extract high-level abstract features from high-dimensional data in an unsupervised situation, which is suitable for feature extraction problems with multi-dimensional performance monitoring parameters such as aircraft engines. The autoencoder consists of an encoder and a decoder. The former completes the abstract feature representation of the sample, and the latter completes the reconstruction of the input sample, and iterative fine-tuning is completed by the gradient descent algorithm. However, in the face of the time series and high noise of aircraft engine monitoring data, it is difficult for a simple autoencoder to achieve the effect of fully extracting degraded features. LSTM networks are commonly used to process time series features, suitable for mining the complex relationship between redundant sequences of engines, and extracting features that can reflect the essence of operation. However, there is noise in the essential features of operation, and CNN units can be further used to remove the noise. Therefore, the present invention proposes a new autoencoder feature extraction method, namely, a long short-term memory convolutional autoencoder (LSTMCNN_AE), which is used to generate low-dimensional, low-noise engine time series feature information, providing a data basis for variable operating condition engine health assessment.

[0149] 21) Construction of LSTMCNN_AE feature extraction network structure

[0150] In order to construct a feature extraction model for aircraft engines with time series and high noise, the present invention proposes a feature extraction method based on the LSTMCNN_AE algorithm, whose network structure is as follows: Figure 8 As shown in the figure, LSTM and convolution units are embedded in this specific network to capture sequence information while reducing feature dimensionality and learning engine degradation features. The LSTM unit in LSTMCNN_AE is used to capture sequence information from multi-sensor time series data. Convolution and deconvolution units are further embedded after the LSTM unit to filter noise and extract features corresponding to engine degradation. Afterwards, residual learning is used to reduce the difficulty of training and improve the feature learning performance of LSTMCNN_AE.

[0151] LSTMCNN_AE consists of two parts: encoder and decoder, the details of which are as follows Fig. 9 As shown. Taking the multi-dimensional decay-sensitive gas path parameters after the working condition transformation as input, the encoder composed of LSTM and convolution units is used to encode the time series input and extract effective features from the original signal. The decoder composed of LSTM and deconvolution units is used to decode the learned features and reconstruct the input data. LSTMCNN_AE mainly includes the following unit modules:

[0152] (21) LSTM encoding unit

[0153] The LSTM unit in the encoder is used to learn the temporal information from the input signal. LSTM consists of three gating units, namely the input gate, the forget gate, and the output gate. The special architecture of LSTM ensures that the model can avoid the gradient vanishing problem and accurately capture the long-term correlation of temporal information.

[0154] In LSTM, for each time point t, the update process of the hidden state ht can be described as follows:

[0155] i t =σ(W i x t +V i h t-1 +b i )

[0156] f t =σ(W f x t +V f h t-1 +b f )

[0157] O t =σ(W o x t +V o h t-1 +bo )

[0158] c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +W c h t-1 +b c )

[0159] h t =o t ⊙tanh(c t ) (16)

[0160] Among them, i t , f t , O t , c t They are the input gate, forget gate, output gate and storage unit, W and V are weight matrices, b is the bias vector, σ is the activation function, and ⊙ represents the convolution unit.

[0161] (22) Convolutional layer

[0162] The feature h generated by LSTM is still noisy and high-dimensional, and cannot effectively represent the engine health status. Therefore, the output h is input to the convolutional layer for further dimensionality reduction and feature learning. The feature F learned by the convolutional layer is as follows:

[0163]

[0164] Where * represents the convolution operation, w is the convolution kernel, b is the bias, and j is the number of channels.

[0165] (23) Deconvolution layer

[0166] The learned features are then passed to the decoder for further decoding. The decoder part consists of LSTM units and deconvolution layers to generate the reconstruction vector As shown in the figure, the deconvolution operation can transform the features into the original shape. The deconvolution operation is the opposite of the convolution operation:

[0167]

[0168] in, represents the deconvolution operation, is the deconvolution kernel, is the deviation, and i is the number of channels.

[0169] (24) LSTM decoding unit

[0170] LSTM units in the decoder are used to capture temporal information from the features generated by the previous layer. Output of the deconvolution layer Input to the LSTM decoding unit. The output of the LSTM can be calculated as follows:

[0171]

[0172] (25) Residual Learning Block

[0173] Residual networks are usually used to solve the problem that the training gradient disappears as the network deepens. Using residual learning can improve the feature learning performance of LSTMCNN_AE. The output of the residual learning block can be expressed as follows:

[0174]

[0175] (26) Fully connected layer

[0176] In order to ensure that the input data has the same shape, the fully connected layer is set on the top layer of LSTMCNN_AE. The final output of the deep network can be obtained by the following formula:

[0177]

[0178] 22) Engine timing degradation feature extraction process based on LSTMCNN_AE

[0179] Feature extraction based on LSTMCNN_AE mainly separates the lines LSTMCNN_AE modeling and online degradation feature extraction.

[0180] In offline LSTMCNN_AE modeling, the training engine gas path parameters are used to train the model. When the degradation data is input into the model, the gas path parameter data of the original training engine is reconstructed by setting the model structure and adjusting the parameters. Assuming that the multidimensional time series Contains all the data of training unit i from the beginning to failure, then use X i To train the model, the reconstruction error can be expressed as The optimal model is trained by minimizing the reconstruction error.

[0181] In the online degradation feature extraction, all the data of the training engine and all the data of the test engine are used as input. The model input is The output is The corresponding degraded features are the reconstructed features learned by the LSTMCNN_AE encoder That is, extracting the engine timing degradation characteristics.

[0182] The engine timing degradation feature extraction process based on LSMTCNN_AE is as follows Fig.10 As shown, including:

[0183] (221) Offline LSTMCNN_AE modeling

[0184] The offline LSTMCNN_AE modeling phase consists of three parts:

[0185] 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 all the data of the training engine and the test engine are selected as the test set. Considering that the aerospace engine data has a strong time series, the sliding window method is used to preprocess the multidimensional gas path parameters.

[0186] In data preprocessing, the sliding window is L and the step size is d, sliding along the multidimensional gas path parameters to generate model input data. The sample generation process is as follows: Fig.11 shown.

[0187] 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.

[0188] 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.

[0189] 203. Train the LSTMCNN_AE feature extraction model. Capture feature information based on the training data set, determine the model parameters such as the number of model layers, number of nodes, learning rate, batch size, number of iterations, and optimize the model parameters through the residual learning method to obtain the optimal feature extraction model.

[0190] (222) Online degradation feature extraction

[0191] The online degradation feature extraction stage consists of two parts.

[0192] Determine the input for feature extraction. When obtaining degradation 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.

[0193] Extract engine degradation features. Input the test set data into the trained LSTMCNN_AE model to 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.

[0194] 2.3 Engine health assessment under multiple fault modes and variable operating conditions based on NCA algorithm

[0195] After obtaining the low-dimensional features of engine performance degradation, it is necessary to complete the construction of health assessment indicators of the engine under multiple fault modes and variable operating conditions based on the degradation features, so as to characterize the degradation trend of the engine under constantly changing operating conditions. The present invention focuses on the advantages of the NCA algorithm, and further proposes a health assessment model based on multiple fault modes and variable operating conditions. The health assessment based on the NCA metric learning algorithm can be trained and learned according to the characteristics of the samples, and can adaptively identify and map features, with strong adaptive ability; only the health and degradation states need to be set to map and separate the test data, and it can be directly used for other data sets, with stronger versatility and scalability. For the problem of multiple fault modes, multiple NCA metric learning models are used to carry out health assessment research under different fault modes; for the problem of variable operating conditions, the extracted features are mapped to a new space, and the learnable Mahalanobis distance matrix is ​​used to calculate the degree of degradation under different operating conditions, so as to achieve robust and efficient health assessment of the engine under variable operating conditions.

[0196] 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 for each operating condition based on the idea of ​​operating condition transformation, measures the differences between health and degradation under each operating condition, and constructs a health index through a distance value. The health index can reflect the health status of the engine sample under the flight cycle. Afterwards, the health index is reconstructed based on the flight cycle as the transformation benchmark. Therefore, the steps of engine health assessment under multiple fault modes and variable operating conditions are divided into: division and mapping of health and degradation sets based on the NCA model, construction of health indicators for each operating condition based on the optimization matrix, and fusion of variable operating condition health states based on operating condition transformation. The specific process is as follows: Fig.12 shown.

[0197] The pseudo code of variable operating condition engine health assessment based on NCA algorithm is as follows:

[0198]

[0199]

[0200] The variable operating condition health assessment process based on NCA metric learning is mainly divided into three parts, which will be introduced in turn below.

[0201] 201) Partitioning and mapping of healthy and degraded sets based on NCA model

[0202] 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.

[0203] Under the i-th fault mode, the NCA metric learning model is supervisedly trained with the healthy set and the degraded set of the training engine degradation characteristics. Then the test engine degradation characteristics are used as the input of the NCA algorithm to divide the test samples into the healthy set and the degraded set.

[0204] (2011) Healthy set and degraded set state label setting

[0205] In the actual operation of the engine, its life does not decrease linearly with the degradation process, but after running for a period of time and reaching the life limit, the remaining life gradually decreases. Therefore, it can be considered that the initial degradation of the engine can be ignored, the remaining life is a constant value, and the remaining life decreases linearly after reaching the critical value.

[0206] The healthy set and the degraded set are divided according to the degradation characteristics of the training engine. According to the piecewise linear degradation principle, 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:

[0207] H train ={H1,H2,...H n_H} (twenty two)

[0208] D train ={D1,D2,...D n_D}(twenty three)

[0209] Among them, H train is the health characteristic corresponding to each flight cycle, D train is the fault feature corresponding to each flight cycle, H i and D i is the dimension reduction feature after LSTMCNN_AE processing,

[0210] Based on the division of healthy set and degraded set, the labels are set. The state label can be expressed as in is the status label corresponding to the health status, is the state label corresponding to the degradation state, and each engine degradation feature corresponds to a state label. Therefore, the state labels corresponding to the healthy set and the degradation set can be expressed as:

[0211]

[0212] (2012) NCA metric learning model training

[0213] In the training of NCA metric learning model, the core is to enable the model to effectively separate the healthy and degraded state features of the engine through training, and obtain a well-learned Mahalanobis distance matrix, so that the healthy and degraded features are sufficiently separated and the same features are relatively aggregated, and lay a stable and efficient foundation for subsequent distance calculation. input ={H train ,D train} as model input, the state label corresponding to the feature As the model output, by setting the maximum number of model iterations max_iter and optimizing the NCA algorithm parameter random_state, the healthy and degraded states of the engine are effectively divided and identified, and the learned optimized Mahalanobis distance matrix Q is obtained.

[0214] (2013) Adaptive mapping of engine degradation characteristics

[0215] In the adaptive mapping of degradation features, the core is to adaptively map the input features to a new space that matches the feature state according to the NCA metric learning model, thereby reducing the aliasing of health and fault features caused by data fluctuations or operating condition changes, making the metric calculation between health and degradation more stable and robust, and the distance representation more accurate. train ={X1,X2,...X n_train} and test engine x test ={x1,x2,...x n_test} features for feature mapping, the one-dimensional mapping features of the engine can be expressed as:

[0216] F train ={F1,F2,...F n_train}(26)

[0217] f test ={f1,f2,...f n_test}(27)

[0218] Among them, each engine mapping feature can be expressed as a set of healthy mapping features and degraded feature mappings F1 = {H train1 ,D train1}, f1={H test1 ,D test1}.

[0219] 202) Construction of health indicators for different working conditions based on optimization matrix

[0220] The core of the construction of working condition health index is to divide the mapping features based on the operating conditions, and construct a health assessment scheme under different working conditions to describe the degree of degradation of the engine under this flight cycle. Unlike the traditional distance measurement method, the traditional distance measurement requires manual setting of the measurement data and measurement length each time, and there are differences in different measurement spaces. The method based on NCA metric learning can use the learned optimal distance matrix for universal measurement, reduce the uncertainty of the measurement method and improve its robustness.

[0221] (2021) Design of measurement scheme for different working conditions

[0222] Considering that although all engines are initially in a healthy state, there may be slight differences between the engines at the beginning. The method of measuring the distance between the degraded set and the healthy set composed of all engines has poor stability and is not suitable for aircraft engines with individual differences. Therefore, this paper adopts the measurement based on individual engines, measures the difference between the degraded state and the healthy state based on the covariance matrix of learning under the same working condition, and constructs the engine health status index value (Confidence Value, CV) under different working conditions. The larger the distance value, the higher the degree of degradation, thereby characterizing the health status of the engine under this working condition.

[0223] (2022) Construction of health indicators for different working conditions

[0224] Taking the i-th test engine as an example, its mapping feature can be expressed as f i ={H i ,D i}, according to the working condition labels obtained in the previous section, the mapping features are divided into healthy and degraded features under different working conditions, which can be expressed as:

[0225]

[0226] Where k is the number of operating conditions, and are the healthy mapping characteristics and the degraded mapping characteristics under the engine operating condition m.

[0227] Under each operating condition, the distance d between the fault state feature and the first healthy state feature is calculated based on the covariance matrix Q obtained by optimization learning. The calculation formula is as follows:

[0228]

[0229] The calculated distance set under the mth working condition is shown in the following formula. This distance can characterize the difference between the current state and the healthy state of the engine under this working condition:

[0230]

[0231] (2023) Construction of health assessment index based on normalized distance of different working conditions

[0232] In order to unify the health indicators into a consistent health assessment indicator measurement, the calculated distance differences need to be normalized. In constructing health indicators, CV = 0 usually represents a healthy state, and CV = 1 represents a completely faulty state. Since the calculation standards for each operating condition are different, each operating condition is independently standardized. The linear normalization method meets the unified standard of 0-1, so the linear normalization method is selected to normalize the distance of each operating condition to obtain the engine health assessment index. Under the mth (1≤m≤k) operating condition, the maximum distance value among all engine samples is The minimum distance value is Under this condition, the standard distance value is calculated as shown in the following formula:

[0233]

[0234] Among them, CV m is the normalized health assessment index under working condition m, d m is the distance value under working condition m.

[0235] 203) Fusion of health status based on operating condition transformation

[0236] After obtaining the health indicators of each operating condition, with the help of operating condition transformation thinking, we carried out research on the generation of health indicators based on the flight cycle as the standard, so as to obtain the health status assessment curve of each engine.

[0237] (2031) Reconstruction of health indicators under variable working conditions

[0238] After constructing the health indicators for each working condition, the health evaluation indicator set for the training engine and the test engine under each working condition can be obtained. Formula (33) is the set of health indicators for the test engine i under each working condition:

[0239]

[0240] Based on the idea of ​​operating condition transformation, the CV indicators are rearranged based on the flight cycle to reconstruct the health assessment curve of the engine under timing conditions. This curve can quantitatively reflect the changes in the health degradation of the engine from a specific numerical value, and lay a data foundation for the subsequent remaining life prediction under multiple fault modes.

[0241] (2032) Generation of health assessment results under multiple failure modes and variable operating conditions

[0242] Traverse the evaluation under all failure modes to form the health evaluation results under each failure mode. and test engines Generate a set of variable condition health assessments under multiple fault modes:

[0243]

[0244] (2033) Evaluation of health assessment results

[0245] This paper proposes two evaluation indicators to quantify the results of the health assessment method, and compares the proposed method with other mainstream health assessment methods to prove the effectiveness of the health assessment method proposed in this paper.

[0246] CV curve trend indicator Trend

[0247] The CV degradation curve generally shows a monotonic trend, which can characterize the engine degradation trend. For training engines, if the degradation trend is obvious, it is consistent with the operation of the training engine from healthy to completely failed. This paper proposes the Trend index to quantify the degradation trend of the CV curve, which is the sum of the increase or decrease of the CV value in each flight cycle. The larger the index, the more obvious the degradation trend. The specific calculation formula is as follows:

[0248]

[0249] in, is the health index of the jth engine in the ith flight cycle under fault mode m, len is the running time of the engine, n m is the number of engines under fault mode m, k is the number of fault modes, and n is the total number of engines.

[0250] CV curve volatility indicator Fluc

[0251] The CV degradation curve is relatively smooth and has low volatility, which indicates that the health assessment process is relatively stable and the selected method is appropriate. This paper proposes the Fluc index to quantify the volatility of the CV curve, which is the sum of the absolute values ​​of the original CV value of each engine and the smoothed CV value. The smaller the index, the smaller the volatility of the CV curve. The specific calculation formula is as follows:

[0252]

[0253] in, and are the original calculated health index and smoothed health index of the jth engine in the ith flight cycle under fault mode m, len is the running time of the engine, n m is the number of engines under fault mode m, k is the number of fault modes, and n is the total number of engines.

[0254] In summary, the present invention extracts low-dimensional degradation information from multi-dimensional engine data based on an improved autoencoder model, which is used as input data for health assessment under variable working conditions. Based on the NCA algorithm, engine health assessment under variable working conditions is performed to obtain health assessment curves of the engine under different fault modes.

[0255] 3. Prediction of Remaining Life of Multi-Failure Mode Engines Based on Metric Learning and Domain Generalization

[0256] After determining the health status trend of the engine under variable working conditions, it is necessary to establish a suitable life prediction model under multiple fault modes based on this. In view of the differences in the health assessment curves of the engine under multiple fault modes, the problems of low prediction efficiency and low model generalization ability in training different life prediction models, deep metric learning is introduced into the life prediction framework, and the nonlinear relationship between the time series distribution difference and the remaining life difference between the metric health assessment indicators is learned by designing the corresponding relationship of "distribution difference-remaining life difference". The gap between multiple fault modes is reduced by learning the "difference" law, so that the prediction model is applicable to multiple fault mode situations. At the same time, in view of the problem that the training samples are insufficient and the network is susceptible to overfitting, the present invention designs a life estimation method combining "direct + indirect", which can combine all training samples for estimation, thereby realizing robust and accurate prediction of engine life and improving the generalization of the prediction model under multiple fault modes. In the present invention, the classic deep metric learning Siamese network and the recurrent neural network GRU combination strategy are selected to predict the life based on the engine health assessment results obtained above under different fault modes.

[0257] Siamese neural network is a classic method in deep metric learning. Siamese network consists of two embedded networks with shared weights, such as Fig.13 As shown. Among them, for the input sample (X i , X j ), train the Siamese network by minimizing the contrastive loss function:

[0258]

[0259] Among them, if the sample pair (X i , X j ) belong to the same category and have a value of 0, then Y ij is a label with value 1; otherwise And α is the minimum margin between pairs of samples of different categories.

[0260] The present invention focuses on the distance measurement capability of Siamese network, embeds the Siamese network structure into the RUL prediction model, applies the Siamese network to learn the differences between health indicators, and inputs the GRU network to perform mapping learning between indicator differences and RUL differences.

[0261] SN-GRU Remaining Life Prediction Model under Multiple Fault Modes

[0262] 31) Construction of SN-GRU multi-fault mode remaining life prediction model

[0263] (31) SN-GRU model and training target design

[0264] Based on the above-mentioned engine health assessment results under the variable operating condition and multiple fault modes of the aircraft engine and the corresponding RUL, the SN-GRU model is constructed. The structure of the network model is as follows: Fig.14 shown.

[0265] The SN-GRU model uses a structure in which Siamese and GRU networks are connected. The Siamese network is a time series health indicator measurement network, and the GRU network is a RUL prediction network. G and P are the mapping functions of the Siamese embedding network and the prediction network, respectively. By introducing a multi-task learning scheme, the training of the SN-GRU model has two main purposes:

[0266] 31.1 First, by inputting the degradation curves CV of different time periods into the Siamese network, the network can effectively calculate and learn their differences.

[0267] 31.2 Secondly, 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 can be learned from the network and the problem of inconsistent health indicator degradation caused by different failure modes can be reduced.

[0268] According to the network settings, θ G ,θ P are the network hyperparameters of G and P respectively. The first training goal of the proposed network is to minimize the individual prediction loss, namely:

[0269]

[0270] Here, two networks are trained simultaneously to learn the health indicator CV i and remaining useful life RUL i The mapping relationship between them.

[0271] At the same time, the SN-GRU model is able to learn the nonlinear relationship between health indicator differences and RUL differences, namely:

[0272] P(z i -z j )=RUL i -RUL j (40)

[0273] Therefore, by learning the differences between the health assessment indicators corresponding to RUL, the differences between RULs can be further estimated, thereby reducing the differences in CV curves caused by multiple failure modes and improving the generalization ability of the prediction model. i ,CV j ) and RUL to (RUL i ,RUL j ), the SN-GRU is trained by optimizing the loss function of formula (41) to satisfy the condition in formula (40). The regularization term helps prevent overfitting. By combining the loss functions of formula (39) and formula (41), the overall training objective shown in formula (42) is obtained, where ω is a balancing parameter.

[0274]

[0275] (32) SN-GRU network structure design

[0276] The SN-GRU network consists of two shared embedding networks and a life prediction network. For the embedding network, the CNN architecture is used, which can perform better 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, through 2 convolutional layers, 1 flat layer and 1 fully connected layer, and the corresponding dimensions and convolution kernels are designed. In the GRU prediction network, 1 GRU layer, 2 dropout layers and 2 fully connected layers are designed. Fig.15 Design diagram of SN-GRU network structure.

[0277] 32) Life prediction process under multiple fault modes based on SN-GRU model

[0278] The present invention introduces Siamese network into RUL prediction problem, establishes SN-GRU prediction model, and constructs engine RUL prediction model suitable for multiple fault modes by designing the corresponding relationship of "distribution difference-remaining life difference". The life prediction process of engine under multiple fault modes based on SN-GRU is as follows Fig.16 shown.

[0279] The pseudo code for model training of engine life prediction under multiple fault modes based on SN-GRU is as follows:

[0280]

[0281]

[0282] By designing the correspondence between “distribution difference-remaining life difference”, the nonlinear relationship between the time series distribution difference and the remaining life difference between the health assessment indicators is learned, and the gap between multiple fault modes is reduced by learning the “difference”, so that the prediction model can be applied to multiple fault mode situations.

[0283] The model training steps for engine life prediction under multiple fault modes based on SN-GRU mainly include three parts:

[0284] (301) Input design of SN-GRU life prediction model under multiple fault modes

[0285] The CV results of the engine training under different fault modes obtained in Chapter 3 are used as the input of the SN-GRU model. Since the trained SN-GRU model is applicable to multiple fault modes, the CV indicators under all fault modes are used as the input of the model training. The present invention selects X1=X2={1,2,...,|CV|} in the training engine health indicator sequence and adopts sliding window random sampling and As the input of training in each iteration, the shared knowledge under the time series is maximized. After quantifying the similarity of the sequence, if the prediction model can map the CV data under random sampling to the corresponding RUL, the generalization ability of the prediction model will be stronger.

[0286] (302) Output design of SN-GRU life prediction model under multiple fault modes

[0287] The current RUL value is calculated based on the flight cycle of the training engine, and the corresponding RUL is obtained based on the set prediction model input sample, which are respectively and Then, the mapping relationship between CV sequence and sequence difference and RUL sequence and RUL difference can be learned by training the SN-GRU model.

[0288] (303) SN-GRU model training

[0289] In each iteration, and As the input of the Siamese network, its corresponding and As the output of the GRU prediction network. After that, through the gradient descent algorithm, and according to the loss function designed in the previous stage Update hyperparameter θ G ,θ P , and obtain the trained SN-GRU model, which has strong generalization ability and can handle the life prediction problem under failure mode.

[0290] 33) Comprehensive lifespan estimation method based on SN-GRU

[0291] The present invention designs a life estimation method combining "direct + indirect", which can combine all training samples for estimation, thereby achieving robust and accurate prediction of engine life and improving the generalization of the prediction model under multiple fault modes.

[0292] For a given test engine health indicator Two RUL prediction methods are proposed.

[0293] (331) Direct RUL estimation

[0294] The input health indicators are processed through the embedding network and GRU prediction network in Siamese to directly predict the RUL, as shown in the following formula:

[0295]

[0296] Will Defined as model-based RUL, F and G are the embedding network and prediction network in SN-GRU, respectively.

[0297] (332) Indirect RUL estimation

[0298] In addition to direct estimation, the RUL of the test engine can also be inferred based on 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 , then RUL can be calculated j +P(G(CV i )-G(CV j )) to estimate RUL i Here, (CV j ,RULj ) is considered as the estimated CV i To increase the accuracy of the RUL prediction of the test set, as many estimates as the number of available training samples for each test sample can be obtained and combined to produce an ensemble effect that reduces the variance of the estimates. The indirect RUL estimation formula based on the reference is as follows:

[0299]

[0300] Among them, is defined as the indirectly estimated RUL, and |D| is the total number of estimated references.

[0301] After completing the direct and indirect RUL estimation, the estimated data is fused to obtain the final RUL prediction result. The fusion is shown in the following formula:

[0302]

[0303] Among them, α and β are fusion parameters respectively.

[0304] Finally, the predicted RUL of each test engine and its final flight cycle are added together to obtain the final predicted life of the test engine. The calculation formula is as follows:

[0305] L k =RUL k +cycle k (46)

[0306] In summary, the present invention has developed a method for predicting the remaining life of an engine with multiple fault modes based on deep metric learning and domain generalization. On the basis of RUL prediction based on recurrent neural networks, deep metric learning is introduced into the prediction strategy to provide a theoretical basis for robust and accurate life prediction under multiple fault modes. In order to solve the problem of engine life prediction under multiple fault modes, the present invention proposes an SN-GRU prediction model that combines Siamese networks and GRU networks. By designing the corresponding relationship of "distribution difference-remaining life difference", while reducing the difference in health assessment indicators under different fault modes, the influence of the network on overfitting problems is reduced to construct an engine remaining life prediction model suitable for multiple fault modes. In order to obtain more accurate life prediction results, the present invention uses a "direct + indirect" life estimation method, which can perform auxiliary estimation based on all training samples while estimating based on test set samples, increase the sample size of the estimation, and obtain a more reliable and accurate prediction effect.

[0307] The following is an example of the process of predicting the remaining life of an engine with multiple fault modes using a specific case.

[0308] The engine health assessment curves under the two failure modes are used as the input basis of the life prediction model, and sliding window cutting is performed. The sliding window length TW = 10, the step length d = 1, and each CV i The input sample design size is (None, 10, 2) to facilitate the subsequent processing of the embedding network. Correspondingly, the RUL value of the last cycle corresponding to the CV sample sliding window is used as the RUL i .

[0309] In model training, the data of the training engine is divided into a training set and a validation set. The training set is used to train the model, and the validation set is used to optimize the model. The ratio of the training set to the validation set is 7:3. According to the "maximum entropy principle", the input sample pairs (CV i ,CV j ) and RUL to (RUL i ,RUL j ), in order to train the SN-GRU model.

[0310] The SN-GRU model consists of a Siamese network and a GRU network. In the Siamese network, the embedding network uses 1-D convolution to process the health assessment sequence, through 2 convolutional layers, 1 flat layer and 1 fully connected layer, and designs the corresponding dimensions and convolution kernels. In the Siamese network, the embedding network has the same structure and weights; in the GRU prediction network, 1 GRU layer, 1 dropout layer and 2 fully connected layers are designed. Among them, the convolutional layer, the fully connected layer, and the GRU layer use the relu activation function, and the output layer uses the Softmax activation function. The model structure design is shown in Table 31 below.

[0311] Table 31 SN-GRU model structure design table

[0312]

[0313]

[0314] The settings of the SN-GRU remaining life prediction model parameters are shown in Table 32 below:

[0315] Table 32 SN-GRU remaining life prediction model parameter setting table

[0316]

[0317] As shown in the table above, in the SN-GRU remaining life prediction model, the Adam optimization algorithm is used to optimize the hyperparameters θ in the Siamese network and the GRU prediction network. G ,θ PThe update is performed with a learning rate of 0.001 and a batch size of 10. The number of iterations is calculated based on the length of the training set, and the maximum number of model iterations is 15000. When calculating the loss function, the balance parameter ω is set to 1, that is, individual RUL prediction and difference prediction have the same proportion in the loss function calculation. When estimating RUL, the fusion factors α and β are taken as 0.6 and 0.4 respectively, that is, the ratio of direct RUL estimation to indirect RUL estimation is 3:2.

[0318] In the SN-GRU remaining life prediction model, the balance factor ω in the loss function is an important parameter in the training process, which affects the optimization training of the model and the final prediction accuracy. The value of ω is (0,1]. If ω is set to 1, the individual RUL prediction and difference prediction account for the same proportion in the loss function calculation; if ω is set to less than 1, the difference prediction accounts for a relatively small proportion in the prediction. This section divides the range of the balance factor ω into four intervals, namely: [0.01, 0.25], [0.25, 0.50], [0.50, 0.75], [0.75, 1.00], and takes the endpoints of the intervals as the value test, and compares and analyzes the accuracy of the model on the training set and the validation set, and takes the value with the highest accuracy in the training set and the validation set as the final value of ω. The balance factor parameters are selected for the training engine data, and the selection results are shown in Table 33 below:

[0319] Table 33 Model parameter balance factors for training set and validation set ω Settings and prediction results

[0320]

[0321] According to the accuracy training results of the SN-GRU model with different balance factor values, when the balance factor is the lowest, that is, 0.01, the model has low accuracy on both the training set and the validation set. This is because there is almost no attention to the difference prediction in the loss function, which makes the input data insufficient, the model overfits, and leads to low accuracy. When the balance factor value gradually increases, it can be seen that the accuracy of the training set and the validation set basically shows an upward trend. When the balance factor is 1, the prediction accuracy of the model on both the training set and the validation set is the highest. This is because in the loss function, the individual RUL prediction and the difference prediction are given the same attention, the model training data is sufficient, and the health assessment data under different fault modes can be obtained from the difference metric, thereby reducing the data difference caused by multiple fault modes and improving the generalization ability of the SN-GRU model. Therefore, the model takes the balance factor ω as 1 during the training process, which also reflects the superiority of the proposed model in dealing with prediction problems under multiple fault modes.

[0322] In the RUL estimation process, according to the proposed RUL estimation method, the fusion parameters α and β are important references in the RUL estimation process. The fusion parameter value is (0,1), and there is a mathematical relationship α+β=1. This section selects its value combination according to its value range and mathematical relationship, selects three combinations as experimental values, and uses the life prediction results of the test engine and the evaluation index RMSE as the selection criteria. The fusion parameter combination (α, β) is divided into: (0,1), (0.4,0.6), (0.5,0.5), (0.6,0.4), (1,0).

[0323] When the fusion parameters are matched to (0.6, 0.4), the RMSE of the test engine evaluation results under multiple fault modes is the lowest, and the predicted RUL and the actual RUL are more consistent under different fault modes; and when direct or indirect prediction is performed separately, the RUL prediction accuracy under multiple fault modes is low. It can be seen that RUL estimation requires a combination of direct + indirect strategies to obtain prediction information from more reference data, thereby obtaining better prediction results.

[0324] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the present invention is not limited thereby. Any modification, equivalent substitution and improvement made by those skilled in the art without departing from the scope and essence of the present invention shall be within the scope of the present invention.

Claims

1. A method for predicting the remaining life of a multi-fault mode engine, characterized in that: include: Obtain health assessment timing curves of training engines under different fault modes, and select health assessment timing curves of training engines under two fault modes; Constructing an SN-GRU model consisting of two Siamese networks and one GRU network, and training the SN-GRU model using the health assessment timing curves of the training engines under the two failure modes to obtain a trained SN-GRU model; The remaining life prediction process of the test engine is performed based on the trained SN-GRU model.

2. The method according to claim 1, characterized in that The Siamese network in the SN-GRU model consists of a convolutional layer, a flat layer, and a fully connected layer, and the GRU network in the SN-GRU model consists of a GRU layer, a dropout layer, and a fully connected layer.

3. The method according to claim 2, characterized in that The SN-GRU model is trained using the health assessment timing curves of the training engine under the two fault modes, and the trained SN-GRU model includes: By performing sliding window cutting processing on the health assessment time series curves of the training engine under the two fault modes, respectively, a health indicator sample under the first fault mode and a health indicator sample under the second fault mode are obtained; Obtain the remaining life corresponding to the health indicator sample in the first fault mode and the remaining life corresponding to the health indicator sample in the second fault mode; In each round of iteration, the health indicator samples under the first fault mode and the health indicator samples under the second fault mode are used as the input of the Siamese network respectively. At the same time, the remaining life corresponding to the health indicator samples under the first fault mode and the remaining life corresponding to the health indicator samples under the second fault mode are used as the output of the GRU prediction network. The hyperparameters of the Siamese network and the GRU prediction network are updated by the gradient descent algorithm to obtain the trained SN-GRU model.

4. The method according to claim 3, characterized in that The remaining life prediction process of the test engine based on the trained SN-GRU model includes: Get the health indicator samples of the test engine under k failure modes The health indicator samples are predicted by the Siamese network and the GRU prediction network in the trained SN-GRU model. After processing, the direct predicted RUL of the test engine is obtained; the formula is: in, 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.

5. The method according to claim 4, characterized in that The remaining life prediction process of the test engine based on the trained SN-GRU model further includes: Get the health indicator samples of the test engine under k failure modes The second health indicator sample and its corresponding remaining life span (CV j ,RUL j ); Based on the trained SN-GRU model, the indirect predicted RUL of the test engine is obtained according to the difference between the health indicator sample and the second health indicator sample in the Siamese network.

6. The method according to claim 5, characterized in that Based on the trained SN-GRU model, according to the difference between the health indicator sample and the second health indicator sample in the Siamese network, the indirect predicted RUL of the test engine is obtained, including: in, is the RUL estimated indirectly based on the SN-GRU model, and |D| is the total number of estimated references.

7. The method according to claim 6, characterized in that The remaining life prediction process of the test engine based on the trained SN-GRU model further includes: The direct predicted RUL and the indirect predicted RUL of the test engine are fused to obtain the predicted RUL of the test engine under k failure modes; the formula is: Among them, α and β are fusion parameters respectively.

8. The method according to claim 7, characterized in that The remaining life prediction process of the test engine based on the trained SN-GRU model further includes: The predicted RUL of each test engine and its final flight cycle k Add them together to get the predicted life L of the engine tested under k failure modes k , the formula is: L k =RUL k +cycle k 。 9. A device for predicting the remaining life of a multi-fault mode engine, characterized in that: include: An acquisition module is used to acquire the health assessment timing curves of the training engine under different fault modes, and select the health assessment timing curves of the training engine under two fault modes; A construction and training module, used to construct an SN-GRU model consisting of two Siamese networks and one GRU network, and train the SN-GRU model using the health assessment timing curves of the training engines under the two failure modes to obtain a trained SN-GRU model; The prediction module is used to perform remaining life prediction processing on the test engine based on the trained SN-GRU model.

10. An electronic device, characterized in that: It 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 method for predicting the remaining life of a multi-fault mode engine described in any one of claims 1 to 8 are implemented.

Images