A method, device, and electronic equipment for predicting the remaining life of a multi-failure mode engine.

Abstract

This invention discloses a method, apparatus, and electronic device for predicting the remaining life of a multi-fault mode engine. The method includes: acquiring health assessment time-series curves of a training engine under different fault modes, and selecting health assessment time-series curves of the training engine 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 time-series curves of the training engine under the two fault modes to obtain a trained SN-GRU model; and performing remaining life prediction processing on a test engine based on the trained SN-GRU model.

Description

Technical Field

[0001] This invention relates to the field of life prediction technology, and in particular to a method, device and electronic equipment for predicting the remaining life of a multi-failure mode engine. Background Technology

[0002] Aero engines are a critical component of aircraft, typically operating in harsh environments such as high pressure and high-speed rotation. Their operating profiles and failure outcomes are complex and variable. Long-term service can lead to performance degradation of components and even the entire aircraft. If the health status of the engine cannot be assessed in a timely manner and the remaining lifespan of the engine cannot be accurately predicted in advance, it will affect the reliability and safety of the aircraft.

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

[0004] Based on the above summary, research on the remaining life prediction of aero-engines under multiple operating conditions and multiple fault modes has gradually attracted attention. However, a complete solution has not yet been formed for the problem of remaining life prediction under varying operating conditions and multiple fault mode combinations. Summary of the Invention

[0005] This invention provides a method, device, and electronic equipment for predicting the remaining life of an engine under multiple fault modes, in order to solve the technical problems of low prediction efficiency and low model generalization ability when training different life prediction models due to differences in the health assessment curves of engines under multiple fault modes.

[0006] This invention provides a method for predicting the remaining life of a multi-failure mode engine, comprising:

[0007] Obtain the health assessment time series curves of the training engine under different fault modes, and select the health assessment time series curves of the training engine under two fault modes.

[0008] An SN-GRU model consisting of two Siamese networks and one GRU network is constructed, and the SN-GRU model is trained using the health assessment time series curves of the training engine under the two failure modes to obtain a trained SN-GRU model.

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

[0010] This invention also provides a multi-failure-mode engine remaining life prediction device, comprising:

[0011] The acquisition module is used to acquire the health assessment time series curves of the training engine under different fault modes, and select the health assessment time series curves of the training engine under two fault modes.

[0012] The construction and training module is used to construct an SN-GRU model consisting of two Siamese networks and one GRU network, and to train the SN-GRU model using the health assessment time series curves of the training engine 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] This invention also provides an electronic device, including 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, it implements the steps of the multi-failure mode engine remaining life prediction method according to any one of claims 1-8.

[0015] The beneficial effect of this invention is that it introduces deep metric learning into the life prediction framework. By designing the correspondence between "distribution difference - remaining life difference", it learns the nonlinear relationship between the temporal distribution difference between health assessment indicators and the remaining life difference. By learning the "difference" pattern, it reduces the gap between multiple failure modes, thereby making the prediction model applicable to multiple failure mode situations. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-fault mode engine remaining life prediction method provided in the embodiments of the present invention;

[0017] Figure 2 This is the technical roadmap for predicting the lifespan of aero-engines under complex operating conditions based on metric learning, as described in this invention.

[0018] Figure 3 This is a flowchart of a method for adaptively identifying engine operating conditions by combining clustering algorithms with metric learning;

[0019] Figure 4 This is a flowchart of the method for equivalent transformation of variable operating condition data;

[0020] Figure 5 It is a schematic diagram of curves representing the degradation under varying operating conditions;

[0021] Figure 6 This is a flowchart of a fault mode identification method based on variable weight prediction;

[0022] Figure 7 This is a flowchart for constructing a single-moment fault mode classification model;

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

[0024] Figure 9 This is a detailed structural diagram of the encoder and decoder in the LSTMCNN_AE network;

[0025] Figure 10 This is a flowchart of engine temporal degradation feature extraction based on LSTMCNN_AE;

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

[0027] Figure 12 This is a flowchart of multi-fault mode and variable operating condition engine health assessment based on NCA metric learning;

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

[0029] Figure 14 This is a schematic diagram of the SN-GRU network model;

[0030] Figure 15 This is a schematic diagram of the SN-GRU network structure;

[0031] Figure 16 This is a flowchart of life prediction under multiple fault modes of an engine based on SN-GRU. Detailed Implementation

[0032] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no inherent meaning. Therefore, "module," "part," or "unit" may be used interchangeably.

[0033] This invention introduces deep metric learning into the life prediction framework. By designing a correspondence between "distribution difference" and "remaining life difference," it learns the nonlinear relationship between the temporal distribution differences and remaining life differences among health assessment indicators. Through learning the patterns of these "differences," it reduces the gaps between multiple failure modes, thus making the prediction model applicable to multiple failure modes. Simultaneously, addressing the problem of insufficient training samples and network susceptibility to overfitting, a combined "direct + indirect" life estimation method is designed. This method combines estimations from all training samples, achieving robust and accurate prediction of engine life and improving the generalization of the prediction model under multiple failure modes. A classic deep metric learning strategy combining the Siamese network and the GRU recurrent neural network is employed to predict remaining life based on the engine health assessment results obtained under different failure modes.

[0034] Figure 1 This is a flowchart of a multi-fault mode engine remaining life prediction method provided by the present invention, such as... Figure 1 As shown, the process includes: S101: obtaining the health assessment time-series curves of the training engine under different fault modes, and selecting the health assessment time-series curves of the training engine under two fault modes; S102: constructing an SN-GRU model consisting of two Siamese networks and one GRU network, and training the SN-GRU model using the health assessment time-series curves of the training engine under the two fault modes 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 this embodiment of the invention, the Siamese network in the SN-GRU model consists of convolutional layers, flattened layers, and fully connected layers, and the GRU network in the SN-GRU model consists of GRU layers, dropout layers, and fully connected layers.

[0036] In this embodiment of the invention, training the SN-GRU model using the health assessment time-series curves of the training engine under the two fault modes to obtain the trained SN-GRU model includes: performing sliding window cutting on the health assessment time-series curves of the training engine under the two fault modes to obtain health index samples under the first fault mode and health index samples under the second fault mode; obtaining the remaining lifetime corresponding to the health index samples under the first fault mode and the remaining lifetime corresponding to the health index samples under the second fault mode; in each iteration, using the health index samples under the first fault mode and the health index samples under the second fault mode as inputs to the Siamese network, and using the remaining lifetime corresponding to the health index samples under the first fault mode and the remaining lifetime corresponding to the health index samples under the second fault mode as outputs of the GRU prediction network, and updating the hyperparameters of the Siamese network and the GRU prediction network using the gradient descent algorithm to obtain the trained SN-GRU model.

[0037] In this embodiment of the invention, performing remaining life prediction processing on the test engine based on the trained SN-GRU model includes: obtaining health index samples of the test engine under k failure modes. The health indicator samples are analyzed using the Siamese network and GRU prediction network in the trained SN-GRU model. The process is performed to obtain the direct predicted RUL of the test engine; the formula is:

[0038]

[0039] in, Let RUL be the RUL directly estimated based on the SN-GRU model, where G is the Siamese network in the SN-GRU model and P is the GRU prediction network in the SN-GRU model.

[0040] In this embodiment of the invention, the remaining life prediction processing of the test engine based on the trained SN-GRU model further includes: obtaining health index samples of the test engine under k failure modes. Second health indicator sample and its corresponding life expectancy (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.

[0041] In this embodiment of the invention, 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, including:

[0042]

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

[0044] In this embodiment of the invention, the remaining service 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 as follows:

[0045]

[0046] Where α and β are the fusion parameters.

[0047] In this embodiment of the invention, the life prediction processing of the test engine based on the trained SN-GRU model further includes: combining the predicted RUL and the last flight cycle of each test engine. k Adding the results, we obtain the predicted engine life L under k failure modes. k Its formula is:

[0048] L k =RUL k +cycle k

[0049] This invention also provides a multi-fault mode engine remaining life prediction device, comprising: an acquisition module, used to acquire health assessment time-series curves of a training engine under different fault modes, and select health assessment time-series 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 time-series 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 a test engine based on the trained SN-GRU model.

[0050] This 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, it implements the steps of the multi-failure mode engine remaining life prediction method.

[0051] There are several bottleneck technical problems to be solved in the life prediction of aero-engines under varying 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 varying operating conditions, difficulty in quantifying engine degradation trends under varying operating conditions, and the problem of decreased accuracy in engine life prediction under multiple fault modes. To address these problems, a life prediction method is proposed. For the overall life prediction process and technical route, please refer to [reference needed]. Figure 2 Specifically, it includes three parts: engine operating condition classification and fault mode identification based on improved clustering methods, variable operating condition engine health assessment based on improved autoencoders and metric learning, and multi-fault mode engine remaining life prediction based on time-series distribution metrics and domain generalization.

[0052] Before completing the prediction of the remaining service life of a multi-fault mode engine, this invention also includes: a variable-condition engine health assessment based on an improved autoencoder and metric learning, and an engine condition classification and fault mode identification process based on an improved clustering method, as detailed below:

[0053] I. Engine Operating Condition Classification and Fault Mode Recognition Based on Improved Clustering Method

[0054] An engine group fault mode identification method includes: performing dimensionality reduction on the performance parameters of training engines and test engines based on Principal Component Analysis (PCA) to obtain low-dimensional performance parameters for both training and test engines; in predicting single-moment fault modes, performing fault mode clustering on the low-dimensional performance parameters of the training engines using the Density-Based Clustering (DBSCAN) algorithm to obtain fault mode classification labels for each training engine; establishing a single-moment fault mode classification model based on Support Vector Machines (SVM), and training the SVM-based model using the low-dimensional performance parameters of the training engines and the corresponding fault mode classification labels for each training engine to obtain a trained SVM-based single-moment fault mode classification model; inputting the low-dimensional performance parameters of the test engines into the trained SVM-based single-moment fault mode classification model to obtain single-moment fault mode classification labels for each test engine; and obtaining the final fault mode classification labels for each test engine based on multi-moment fault mode weighted prediction and the single-moment fault mode classification labels for each test engine.

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

[0056] Based on the trained SVM-based single-time fault mode classification model, the prediction results of each test engine under the fault mode at a single time point are obtained, including:

[0057]

[0058] The weighted prediction results for each test engine under different fault modes, obtained by using the fault state of each test engine at a single moment and the combined fault state at multiple moments, include:

[0059]

[0060] Where m represents the failure mode; j represents the test engine; τ represents the recent weighting coefficient of the identification proportion of recent data; λ represents the comprehensive weighting coefficient of the number of flight cycles considered; k represents the value coefficient affecting the weight of recent far-term factors, and represents the flight cycle from the most recent moment. This represents the predicted state of test engine j in fault mode m at the time furthest from t; This represents the predicted state of engine j under fault mode m at time tk. This represents the prediction result of engine j under fault mode m obtained through weighted prediction.

[0061] The final fault mode classification label for each test engine is obtained by using the weighted prediction results of each test engine under different fault modes, including:

[0062]

[0063] Among them, Y j This indicates the final failure mode classification label for test engine j.

[0064] Before performing dimensionality reduction on the performance parameters of the training engine and the test engine based on principal component analysis (PCA), the process includes: screening out the performance parameters of the engine that are sensitive to degradation under varying operating conditions; extracting engine data under different operating conditions based on the individual engine operating condition classification results to obtain degradation-sensitive performance parameters under each operating condition; normalizing the degradation-sensitive performance parameters under each operating condition using a normalization method, and reconstructing the normalized degradation-sensitive performance parameters under each operating condition according to the engine's flight cycle sequence to obtain the fused multi-operating-condition degradation-sensitive performance parameters; and dividing the fused multi-operating-condition degradation-sensitive performance parameters into training engine performance parameters and test engine performance parameters.

[0065] The process of identifying performance parameters of an engine that are sensitive to degradation under varying operating conditions also includes: obtaining 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 initially identify the operating condition categories of the training engine and obtaining the cluster centers for each operating condition; using the Multidimensional Scaling Analysis (MDS) algorithm to scale the multidimensional operating parameters of the test engine and the corresponding cluster centers for each operating condition to one dimension, and using the one-dimensional operating condition cluster centers as the operating condition baseline; and adaptively identifying the operating conditions corresponding to each test engine in each flight cycle based on the principle of being closest to the operating condition baseline.

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

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

[0068] To accurately determine the operating condition label at a single moment, this invention proposes an improved clustering method to identify engine operating conditions, namely, introducing 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 a low-dimensional space, i.e., reconstructing the low-dimensional coordinates of the data, thereby efficiently measuring the true distance between data points. The MDS algorithm can project multi-dimensional operating parameters into a low-dimensional space while maintaining relative positions. The test engine can automatically identify and be assigned to the nearest operating condition category, thus adaptively achieving single-moment operating condition category classification for the engine. 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 centers, overcoming the problem of inconsistent Euclidean distance scales, realizing the classification of test engine operating conditions, and solving the problem of inaccurate clustering of test set operating conditions.

[0069] In engine operating condition identification, the multidimensional operating parameters of the training engine are used as the data foundation. The K-Means clustering algorithm is used to initially identify the operating condition categories of the training engine and obtain the cluster centers for each condition. Then, the MDS algorithm is used to scale the test engine operating parameters and the cluster centers to one dimension. Based on the principle of closest to the operating condition baseline, the operating conditions corresponding to each flight cycle of the test engine are adaptively identified. The specific process of adaptive engine operating condition identification based on 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 classify the engine's operating condition categories at a single moment includes the following steps: S111, obtaining the multidimensional operating parameters of the training engine under different operating conditions; S112, based on the multidimensional operating parameters obtained in step S111, using the K-Means clustering algorithm to initially identify the operating condition categories of the training engine and obtain the cluster centers of each operating condition; S113, using the MDS algorithm to scale the multidimensional operating parameters of the test engine and the cluster centers of the corresponding operating conditions to one dimension; S114, adaptively identifying the operating conditions of the test engine corresponding to each flight cycle based on the principle of being closest to the operating condition baseline.

[0070] The pseudocode for adaptive engine condition recognition based on a combination of metric learning and clustering is as follows:

[0071]

[0072]

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

[0074] (101) Engine operating parameter extraction

[0075] Data from aero-engines operating under complex conditions includes operational parameters and gas path parameters. The operational parameter data is extracted and used as the foundation for engine condition identification. The continuous changes in engine operational parameters indicate that the engine's operating conditions are constantly changing during operation. Combinations of operational parameters bring different operating conditions to the aircraft. When the values ​​of the three operational parameters are basically the same and within a very small fluctuation range, it can be indicated that the engine operating conditions are similar. Aero-engine operating conditions are usually random and their patterns are difficult to identify, which also brings certain difficulties to subsequent engine health assessments under varying operating conditions.

[0076] (102) Determination of cluster centers and division of operating condition labels for training engines based on K-Means

[0077] The operational parameters of the trained engines are used as the multidimensional input for K-Means clustering. Based on the number of aero-engine operating condition categories k, k combinations of operating condition parameters are randomly selected as initial cluster centers. During the training of the K-Means clustering model, the number of cluster centers and the maximum number of iterations per training session are set. After training, the model returns the optimal cluster centers and outputs the operating condition label and final cluster center results for each training engine flight cycle.

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

[0079] The operating parameters of the test engine are merged with the operating condition cluster center. The multidimensional data is scaled down to one dimension using the multidimensional scaling (MDS) algorithm. The one-dimensional operating condition center is used as the baseline. The scaled operating data at a single moment 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 Operating Condition Transformation

[0081] Air path parameters of aero-engines serve as crucial data for health assessment and lifespan prediction. However, these parameters fluctuate under varying operating conditions, making it difficult to accurately represent the true degradation trend of the engine. Typically, not all air path parameters are useful for studying engine condition; some parameters provide constant or discrete values ​​throughout the engine's lifespan. To accurately depict the true degradation trend of an engine under varying operating conditions, it is necessary to standardize performance parameters from different operating conditions to a comparable parameter standard. This invention proposes an equivalent transformation method for engine data under varying operating conditions. This method mainly consists of four parts: screening of degradation performance parameters, decomposition of multi-condition time-series data, standardization of performance parameters for each operating condition, and performance parameter reconstruction. The process of equivalent transformation of variable operating condition data is as follows: Figure 4 As shown.

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

[0083] In the screening of degradation performance parameters, by plotting engine life curves under different operating conditions, constant and unchanging airflow parameters with no obvious degradation trend are initially eliminated. This allows for the preliminary screening of performance parameters that show a clear degradation trend with operating conditions, exhibit smaller fluctuations, and are more sensitive to engine degradation, thus forming a preliminary framework. Figure 4 The performance parameter sequence under varying operating conditions is shown in the figure.

[0084] (1002) Decomposition of Multi-condition Time Series Data

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

[0086] (1003) Standardization of performance parameters under various operating conditions

[0087] Aero engines face frequent changes in operating conditions during operation. Under the same operating conditions, the degradation parameters of engine performance typically exhibit a gradual trend with relatively small fluctuations. However, when operating conditions change frequently, the degradation trend becomes difficult to detect.

[0088] To achieve the goals of parameter unification and degradation trend integration, two operations are required: First, for engine parameters, they need to be transformed into unified indicators that reflect degradation trends, and the data under different operating conditions need to be scaled and unified to a standard space; second, for degradation trends, the standardized data under different operating conditions need to be used to form curves that can clearly characterize the degradation under varying operating conditions, such as... Figure 5 As shown.

[0089] A normalization method is used to normalize the degradation sensitivity parameters under different operating conditions, thereby achieving parameter normalization for each operating condition. Let the aero-engine operate under flight cycle t... i At time t, the decay sensitivity parameter is x. i The corresponding working condition is m i Its state parameters can be expressed as (t i ,x i ,m i The engine operating conditions include k types, and the set of these operating condition categories can be represented as M = {M1, M2, ..., Mn}. k Therefore, the total set of performance parameters under each operating condition can be represented as {X}. (K) ={x i |m i ∈M}.

[0090] z-score normalization establishes a standardization 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 parameters under different operating conditions can be expressed as follows: and s kThen the normalization formula for data under operating condition k is:

[0091]

[0092] The z-score method uses the mean and standard deviation as normalization benchmarks, which have less impact on local fluctuations and noise. Given that aero-engine parameters usually have significant noise interference, it can more effectively address this issue.

[0093] (1004) Reconstruction of gas path performance parameters

[0094] Finally, the normalized performance data is reconstructed according to the engine's flight cycle sequence, thereby merging the normalized characteristic parameters under different operating conditions to reconstruct a fused multi-condition degradation time series, thus characterizing the engine's degradation under varying operating conditions. The reconstructed parameter data can be represented as follows:

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

[0096] Where X' represents the aerodynamic performance degradation parameter of the engine under the reconstructed variable operating conditions, x′ t Let N be the N-dimensional performance parameters under flight cycle t, where t = 1, 2, ..., n.

[0097] 1.3 Engine Failure Mode Recognition Based on Classification and Variable Weight Prediction

[0098] In fault mode identification, fault modes need to be classified according to the engine's degradation path. Considering the fluctuation of engine parameters, and the fact that K-Means lacks inclusiveness for noisy data, and that the mean similarity value has no practical meaning for trajectory clustering, this can cause problems for subsequent performance degradation trajectory identification and classification on the test set. Density-based clustering algorithms are not affected by parameter fluctuations and are suitable for fault mode identification in this research. Therefore, this invention designs a fault mode discrimination method based on density clustering-support vector machine (DBSCAN-SVM) and multi-time prediction weighting. The density clustering method is used to identify engine fault modes, and SVM is used to classify the fault state of the test unit at a single time. Finally, the fault states at multiple time steps are weighted to determine the fault mode of the test unit.

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

[0100] Engine fault mode recognition based on variable weight prediction mainly consists of two parts: establishing a single-time fault mode classification model and multi-time fault mode weighted prediction. The main process is as follows: Figure 6 As shown.

[0101] The pseudocode for engine fault mode recognition based on variable weight prediction is as follows:

[0102]

[0103]

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

[0105] The construction of a single-moment fault mode classification model mainly consists of three parts: parameter dimensionality 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 As shown.

[0106] (301) Parameter dimensionality reduction based on principal component analysis

[0107] Even after equivalent transformation of operating conditions, the gas path performance parameters remain multidimensional, and there may be correlations between the parameters, making the trend of failure modes somewhat difficult and complex. Analyzing the performance parameters separately would greatly reduce efficiency, leading to a problem-solving approach that isolates rather than comprehensively considers the issue. Therefore, considering the correlations between performance parameters, the PCA method from unsupervised metric learning is used for dimensionality reduction, obtaining uncorrelated 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 preserving the data's key information. Its core principle is to linearly transform the data in uncorrelated directions by calculating the eigenvectors of the covariance matrix. These new dimensions are then ordered according to their contribution to the total variance to identify the most important dimensions that retain the data's essential information. This helps reduce the dimensionality of the data while maintaining its key structural features.

[0109] (302) Degenerate trajectory clustering based on DBSCAN

[0110] The training engine monitors data from its healthy state to its decline. Considering that degradation characteristics are often not obvious in the early stages of engine operation, later decline data is selected to determine the fault mode category. After extracting low-dimensional gas path parameter features of the engine using the PCA method, the DBSCAN method is used to classify the later gas path parameter features of the training engine and obtain the corresponding fault mode for each training engine.

[0111] After fault mode clustering based on the DBSCAN algorithm, assuming there are M types of fault modes, the feature matrix of the m-th fault mode after clustering is:

[0112]

[0113] Among them, Z m For N m ×l matrix, N m Z represents the number of feature samples for the m-th type of fault mode, and l represents the dimension of the fault features. m The corresponding failure mode classification label can be represented as Y. m =[mm … m] T .

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

[0115] The feature matrices of various failure modes are used as input to the failure mode classifier, and the corresponding failure mode labels are used as the classifier output to train the failure mode classification model. Unlike neural networks, which 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 aero-engine failure modes is relatively small, this invention utilizes SVM to construct a single-moment failure mode classification model, laying the foundation for testing engine failure mode classification.

[0116] SVM is a supervised learning classification algorithm. Its principle is based on finding a hyperplane with the maximum margin to separate data points of different classes. The margin refers to the distance from the data point closest to the hyperplane to that hyperplane, and the support vectors are those data points that are closest to it. This invention designs a fault mode classification model based on SVM. If the training engine samples are X = {(x1,Y1),(x2,Y2),…,(x...}... N ,Y N The fault mode label is Y. k For each element ∈{1,2,…,M′}, k=1,2,…,N, the optimization problem of the SVM model for different types of fault modes can be expressed as:

[0117]

[0118] Wherein, φ(x) γ This is a mapping function that maps input engine data to category labels. The loss function of the model is defined by the penalty coefficient C and the hyperplane normal vector ω, which are obtained through model optimization of hyperparameters. ij Hyperplane intercept b ij This allows us to obtain the hyperplane partitioning for the i-th and j-th fault modes. For the test sample x... test The category to which a single SVM model belongs. The judgment is:

[0119]

[0120] Among them, if A value of 1 indicates that the test sample is the i-th type of fault mode; otherwise, it is the j-th type of fault mode.

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

[0122] Since engine performance degradation is a gradual process, the fault characteristics of test engines that have not fully degraded may not be significant enough to be accurately identified using 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 fault mode identification process, the dynamic changes during engine degradation are considered more comprehensively, overcoming the shortcomings of fault mode classification at a single moment.

[0123] According to the SVM fault mode classification model, the fault state of test engine j at a single moment can be represented as follows: Its state under fault mode m (m=1,2,…,M) can be calculated as follows:

[0124]

[0125] Subsequently, based on the weighted prediction of the fault states at multiple time points, the weighted prediction results of engine j under different fault modes can be calculated as follows:

[0126]

[0127] In this invention, a recent weighting coefficient τ (0 < τ < 1) and a comprehensive weighting coefficient λ (λ, t) are introduced to analyze the engine degradation trajectory. τ represents the proportion of recent data in the identification, and λ represents the number of flight cycles considered comprehensively. The engine's final failure mode can be calculated as follows: The fault mode with the highest probability is selected as the engine's final fault mode category.

[0128] The calculation method for τ is as follows: based on the recent data requirements, the number between (0,1) is taken as the weight. In this case, 0.7 is taken as the recent weight coefficient.

[0129] The calculation method for λ is: take 70% of the flight cycle number of the j-th engine.

[0130] II. Variable-condition engine health assessment based on improved autoencoder and metric learning

[0131] A variable-condition engine health assessment method includes: 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 temporal degradation feature data; constructing and training an NCA metric learning model for each fault mode; mapping the low-dimensional temporal degradation feature data using the trained NCA metric learning model for each fault mode to obtain mapped feature data; and performing engine health assessment processing based on an optimized Mahalanobis distance feature matrix and the mapped feature data.

[0132] The LSTMCNN_AE feature extraction model is trained using engine training set data to obtain the trained LSTMCNN_AE feature extraction model. This process includes: preprocessing multi-dimensional engine parameter data under different operating conditions using a sliding window with a set length and step size, and dividing the multi-dimensional engine parameter data into training samples and test samples according to the sliding window method; using the training samples as model input and output, reconstructing the input sequence through the LSTMCNN_AE feature extraction model; capturing feature information based on the training dataset, determining the model parameters of the feature extraction model, and optimizing the model parameters through a residual learning method to obtain the trained LSTMCNN_AE feature extraction model.

[0133] The construction and training of NCA metric learning models under various fault modes includes: dividing low-dimensional temporal degradation feature data into temporal degradation feature data under different fault modes based on the identified engine fault mode labels; dividing the temporal degradation feature data into healthy set feature data and degraded set feature data under each fault mode, and labeling the corresponding healthy set state label and degraded set state label respectively; constructing NCA metric learning models under each fault mode, and using the healthy set features and degraded set features under each fault mode as model inputs, while using the state labels corresponding to the features as model outputs for training, to obtain the trained NCA metric learning models and optimized Mahalanobis distance feature matrices under each fault mode.

[0134] The engine health assessment process based on the optimized Mahalanobis distance feature matrix and the mapped feature data includes: calculating the distance between the engine's health space and degradation space under each operating condition based on the optimized Mahalanobis distance feature matrix and the mapped feature data; normalizing the distance under each operating condition using a linear normalization method to obtain the engine's health assessment index under each operating condition; reconstructing the engine's health assessment index under each operating condition into a time-series health assessment curve based on flight cycle rearrangement; and traversing the assessment under all failure modes to form the health assessment results under each failure mode.

[0135] In engine health assessment under multiple fault modes and varying operating conditions, two problems arise: First, extracting temporal degradation features under varying operating conditions is difficult. Actual engine monitoring parameters are high-dimensional, temporal, and noisy. Some feature extraction methods focus primarily on dimensionality reduction, neglecting the temporal and noisy nature of engine data, making it difficult to obtain sufficient degradation information for health assessment. To address this issue, this invention proposes a novel autoencoder feature extraction method that utilizes the dimensionality reduction characteristics of autoencoders while capturing time-series information, reducing the impact of data fluctuations. Second, selecting the space for constructing health indicators is difficult. Due to the differences in the initial health state of individual engines and the inconsistency of operating conditions during operation, existing methods that rely only on the absolute amplitude of degradation features without considering the differences in features between the initial and current times are not suitable for health assessment under varying operating conditions. Furthermore, for engine data with unknown specific distributions, it is difficult to establish degradation indicators using distribution measurement methods. To address this problem, this invention introduces metric learning methods into health assessment. It employs the nearest neighbor component analysis (NCA) algorithm, which measures the difference between degradation characteristics and initial health characteristics under different operating conditions through nearest neighbor category classification and Mahalanobis distance transformation. Furthermore, it utilizes the concept of operating condition transformation to obtain health assessment values ​​under varying operating conditions, thereby completing the characterization of the degradation trend of an individual engine.

[0136] 2.1 Introduction to the NCA Algorithm Principle

[0137] The KNN method is a supervised nonparametric machine learning method that assigns a sample to the class that appears most frequently among its K nearest neighbors. Therefore, choosing a suitable distance metric is a key issue in the KNN method. The NCA algorithm is a metric learning method related to KNN. In this algorithm, nearest neighbors refer to similar samples. The concept of "neighbor" is an important idea in metric learning. The NCA algorithm randomly calculates nearest neighbors and uses leave-one-out cross-validation to obtain the transformation matrix in the Mahalanobis distance. This method mainly includes three key points: NCA is supervised learning based on the KNN classifier; NCA uses a distance metric to measure the similarity of neighboring samples; and NCA uses cross-validation to adjust parameters during training.

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

[0139]

[0140] Here, the positive semi-definite matrix Q is a learnable parameter, called the metric. The goal of metric learning is to optimize matrix Q under certain constraints, thereby learning a distance metric.

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

[0142]

[0143] Where, p ij Represents data point x i Instead of using random nearest neighbors, another data point x is selected. j The probability of the corresponding category. The probability of a sample point being correctly classified is:

[0144]

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

[0146]

[0147] 2.2 Engine Temporal Degradation Feature Extraction Based on LSTMCNN_AE

[0148] Autoencoders are a representative feature extraction method. Compared to other feature extraction methods, autoencoders can extract high-level abstract features from high-dimensional data in an unsupervised manner, making them suitable for feature extraction problems involving multi-dimensional performance monitoring parameters, such as those for aero-engines. An autoencoder consists of an encoder and a decoder. The former completes the abstract feature representation of the samples, while the latter reconstructs the input samples, using a gradient descent algorithm for iterative fine-tuning. However, given the temporal and high-noise nature of aero-engine monitoring data, a simple autoencoder is insufficient to fully extract degenerate features. LSTM networks are commonly used to process temporal features, suitable for mining complex relationships between redundant engine sequences and extracting features that reflect the essence of operation. However, the essential operational features are noisy, which can be further removed using CNN units. Therefore, this invention proposes a novel autoencoder feature extraction method, namely Long Short-Term Memory Convolutional Autoencoder (LSTMCNN_AE), to generate low-dimensional, low-noise engine time-series feature information, providing a data foundation for engine health assessment under varying operating conditions.

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

[0150] To construct a feature extraction model suitable for time-series and high-noise aero-engines, this invention proposes a feature extraction method based on the LSTMCNN_AE algorithm, the network structure of which is as follows: Figure 8 As shown, LSTM and convolutional units are embedded in this specific network to capture sequence information while reducing feature dimensionality and learning engine degradation features. The LSTM units in LSTMCNN_AE are used to capture sequence information from multi-sensor time-series data. Convolutional and deconvolutional units are further embedded after the LSTM units to filter noise and extract features corresponding to engine degradation. Residual learning is then employed to reduce training difficulty and improve the feature learning performance of LSTMCNN_AE.

[0151] LSTMCNN_AE consists of two parts: an encoder and a decoder, the details of which are as follows: Figure 9 As shown. Using the multidimensional decay-sensitive gas path parameters after operating condition transformation as input, an encoder composed of LSTM and convolutional units encodes the time-series input and extracts effective features from the original signal. A decoder composed of LSTM and deconvolutional units decodes the learned features and reconstructs the input data. LSTMCNN_AE mainly includes the following unit modules:

[0152] (21) LSTM coding unit

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

[0154] In LSTM, the update process of the hidden state ht at each time point t 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 These are the input gate, forget gate, output gate, and storage unit, respectively. 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 features h generated by LSTM are still noisy and high-dimensional, and cannot effectively represent the engine health status. Therefore, the output h is fed into a convolutional layer for further dimensionality reduction and feature learning. The features F learned by the convolutional layer are as follows:

[0163]

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

[0165] (23) Deconvolution layer

[0166] The learned features are then fed into the decoder for further decoding. The decoder consists of LSTM units and deconvolution layers, which generate reconstructed vectors. As shown in the figure, the deconvolution operation can transform features into their original shape. The deconvolution operation is the reverse of the convolution operation.

[0167]

[0168] in, This represents the deconvolution operation. For deconvolution kernel, Let be the deviation, and i be the number of channels.

[0169] (24) LSTM decoding unit

[0170] The LSTM unit in the decoder is used to capture temporal information from the features generated in the previous layer. The output of the deconvolution layer... The input is fed into the LSTM decoding unit. The output of the LSTM can be calculated as follows:

[0171]

[0172] (25) Residual Learning Block

[0173] Residual networks are commonly used to address the vanishing gradient problem as the network deepens. Residual learning can improve the feature learning performance of LSTMCNN_AE. The output of a residual learning block can be represented as follows:

[0174]

[0175] (26) Fully connected layer

[0176] To ensure the input data has the same shape, a fully connected layer is placed at the top of LSTMCNN_AE. The final output of the deep network can be obtained using the following formula:

[0177]

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

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

[0180] In offline LSTMCNN_AE modeling, the engine's airflow parameters are used to train the model. When degraded data is input into the model, the original training engine's airflow parameter data is reconstructed by setting the model structure and adjusting parameters. This assumes a multidimensional time series... If X contains all the data from the start of training unit i until it becomes invalid, then X is used. i To train the model, the reconstruction error can be expressed as: The optimal model is trained by minimizing the reconstruction error.

[0181] In online degradation feature extraction, all data from the training engine and all data from the test engine are used as inputs, and the model input is... The output is The corresponding degradation features are the reconstructed features learned by the LSTMCNN_AE encoder. That is, to extract the time-series degradation characteristics of the engine.

[0182] The engine temporal degradation feature extraction process based on LSMTCNN_AE is as follows: Figure 10 As shown, it includes:

[0183] (221) Offline LSTMCNN_AE modeling

[0184] The offline LSTMCNN_AE modeling stage consists of three parts.

[0185] 201. Determine the training and test sets for the model. The airflow data of the training engine is selected as the training set, and all data from both the training and test engines are used as the test set. Considering the strong temporal sequence of aero-engine data, a sliding window approach is used to preprocess the multi-dimensional airflow parameters.

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

[0187] The preprocessed model training set is This consists of two-dimensional gas path parameter data of length L×K (sliding window length × number of gas path parameters). The model test set is X. test ={X trian_all ,X test_all},in, These represent the two-dimensional air path parameter data of all training engines and test engines obtained through the sliding window method, respectively.

[0188] 202. Determine the model's input and output. During model training, the input and output sequences are the same, both being the training set data X. train The input sequence is reconstructed using the LSTMCNN_AE model.

[0189] 203. Train the LSTMCNN_AE feature extraction model. Based on the training dataset, capture feature information, determine the model parameters such as the number of layers, number of nodes, learning rate, batch size, and number of iterations, and optimize the model parameters through residual learning 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 acquiring degenerate features, the input is the test set data X. test ={X trian_all ,X test_all}, thereby obtaining low-dimensional degenerate 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 feature. This will establish a data foundation for subsequent health assessments.

[0194] 2.3 Engine Health Assessment Based on NCA Algorithm under Multiple Fault Modes and Variable Operating Conditions

[0195] After obtaining low-dimensional features of engine performance degradation, it is necessary to construct engine health assessment indicators under multiple fault modes and varying operating conditions based on these degradation features, thereby characterizing the engine's degradation trend under constantly changing operating conditions. This invention focuses on the advantages of the NCA algorithm and further proposes a health assessment model based on multiple fault modes and varying operating conditions. Health assessment based on the NCA metric learning algorithm can be trained and learned according to sample characteristics, enabling adaptive feature recognition and mapping with strong adaptability. Only health and degradation states need to be set to perform test data mapping and state separation, allowing direct use on other datasets, resulting in greater versatility and scalability. For the multiple fault mode problem, multiple NCA metric learning models are used to conduct health assessment research under different fault modes; for the varying operating conditions problem, the extracted features are mapped to a new space, and a learnable Mahalanobis distance matrix is ​​used to calculate the degree of degradation under different operating conditions, achieving robust and efficient engine health assessment under varying operating conditions.

[0196] This invention employs the NCA algorithm to partition the health and fault states of training and testing engines, and constructs a condition-specific health assessment model based on the concept of operating condition transformation. This model measures the differences between health and degradation under various operating conditions, and constructs a health index using distance values. This health index reflects the health status of the engine sample under a specific flight cycle. Subsequently, the health index is reconstructed using the flight cycle as the transformation benchmark. Therefore, the engine health assessment steps under multiple fault modes and varying operating conditions are: partitioning and mapping health and degradation sets based on the NCA model; constructing condition-specific health indices based on optimization matrices; and fusing the health states under varying operating conditions based on operating condition transformation. The specific process is as follows: Figure 12 As shown.

[0197] The pseudocode for engine health assessment under varying operating conditions based on the 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 Degenerate Sets Based on NCA Model

[0202] Based on the identified engine fault mode labels, the extracted time-series features are divided into features for different fault modes. Each failure mode corresponds to an NCA metric learning model. The following explanation uses the construction of the NCA metric learning model under the i-th failure mode as an example.

[0203] In the i-th failure mode, the NCA metric learning model is trained in a supervised manner using the healthy set and the degraded set of the training engine degradation features. Then, the engine degradation features are used as input to the NCA algorithm to divide the test samples into the healthy set and the degraded set.

[0204] (2011) Setting Status Labels for Healthy and Degenerate Sets

[0205] In actual engine operation, its lifespan does not decrease linearly with degradation. Instead, after reaching its lifespan limit after a period of operation, the remaining lifespan gradually decreases. Therefore, the initial degradation of the engine can be considered negligible, and the remaining lifespan is a constant value. After reaching the critical value, the remaining lifespan decreases linearly.

[0206] To address the degradation features of the training engine, a healthy set and a degradation set are partitioned. Based on the piecewise linear degradation principle, the healthy state features of the training engine are considered as the original input to the healthy set of the NCA model, while the degradation state features are considered as the original input to the degradation set. Health and fault labels are used as the model output. The degradation features of the training engine can be represented as X. train ={X1,X2,...X n_train}, where X n =X ij Let i represent the engine flight cycle and j represent the dimension of the degradation features. Using the first 10% of engine features as the healthy set and the last 90% as the degradation set, then the healthy set H... train and degenerate set D train It can be represented 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 It is the health trait corresponding to each flight cycle, D train H represents the fault characteristics corresponding to each flight cycle. i and D i These are the dimensionality-reduced features after processing by LSTMCNN_AE.

[0210] Based on the division into healthy and degenerate sets, labels are set for them. State labels can be represented as... in These are the status labels corresponding to the health status. Each engine degradation feature corresponds to a state label, which is a state label representing the degradation state. Therefore, the state labels for the healthy set and the degradation set can be represented as follows:

[0211]

[0212] (2012) NCA Metric Learning Model Training

[0213] In the training of the NCA metric learning model, the core objective is to train the model to effectively separate the features of engine health and degradation states, and to obtain a well-learned Mahalanobis distance matrix. This ensures sufficient separation between health and degradation features while relatively clustering similar features, laying a stable and efficient foundation for subsequent distance calculations. The training uses the engine health set and degradation set features X... input ={H train D train The state labels corresponding to the features are used as model inputs. As the model output, by setting the maximum number of model iterations max_iter, the NCA algorithm parameter random_state is optimized, thereby effectively classifying and identifying the engine health and degradation states, and obtaining the learned optimized Mahalanobis distance matrix Q.

[0214] (2013) Adaptive Mapping of Engine Degradation Features

[0215] In adaptive mapping of degradation features, the core is to adaptively map input features to a new space that conforms to the feature state based on the NCA metric learning model. This reduces the aliasing of health and fault features caused by data fluctuations or changes in operating conditions, making the metric calculation between health and degradation more stable, robust, and the distance representation more accurate. The training engine X will be used... train ={X1,X2,...X n_train} and test engine x test ={x1,x2,...x n_test The features are used for feature mapping, and the one-dimensional mapped features of the engine can be represented 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 represented as a set F1 = {H} of health mapping features and degradation mapping features. train1 D train1}, f1={H test1 D test1}

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

[0220] The core of constructing operational health indicators is to divide the mapping features based on operating conditions and build a health assessment scheme for each operating condition to describe the degree of engine degradation during that flight cycle. Unlike traditional distance measurement methods, which require manual setting of measurement data and length for each measurement and exhibit differences across different measurement spaces, the NCA-based metric learning method can utilize the learned optimized distance matrix for universal measurement, reducing uncertainty and improving 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 individual engines, 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 aero-engines with individual differences. Therefore, this paper adopts a measurement based on individual engines, using a learned covariance matrix to measure the difference between the degraded and healthy states under the same operating condition, and constructs an engine health status index (Confidence Value, CV) for each operating condition. The larger the distance value, the higher the degree of degradation, thus characterizing the engine's health status under that operating condition.

[0223] (2022) Construction of Health Indicators by Working Condition

[0224] Taking the i-th test engine as an example, its mapping feature can be represented as f i ={H i D i Based on the operating condition labels obtained in the previous chapter, the mapped features are divided into health and degradation features under different operating conditions, which can be represented as:

[0225]

[0226] Where k is the number of operating conditions. and These are the health mapping characteristics and degradation mapping characteristics of the engine under 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 through optimization learning. The calculation formula is as follows:

[0228]

[0229] The calculated distance set for the m-th operating condition is shown in the following formula. This distance can characterize the difference between the engine's current state and its healthy state under this operating condition:

[0230]

[0231] (2023) Construction of Health Assessment Indicators Based on Distance Normalization under Different Working Conditions

[0232] To standardize health indicators to a consistent health assessment metric, the calculated distance differences need to be normalized. In constructing health indicators, a healthy state is typically represented by CV=0, and a complete failure state by CV=1. Since the calculation standards differ for each operating condition, each condition is independently standardized. Linear normalization conforms to the unified 0-1 standard; therefore, it is chosen to normalize the distances for each operating condition, thus obtaining the engine health assessment index. Under the m-th (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 using the following formula:

[0233]

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

[0235] 203) Fusion of Health Status Based on Variable Operating Conditions

[0236] After obtaining the health indicators for each operating condition, we use the concept of operating condition change to conduct research on the generation of health indicators based on the flight cycle, thereby obtaining the health status assessment curves for each engine.

[0237] (2031) Reconstruction of Health Indicators under Variable Operating Conditions

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

[0239]

[0240] Based on the concept of operating condition transformation, the CV index is rearranged based on the flight cycle to reconstruct the engine health assessment curve under time-series conditions. This curve can quantitatively reflect the changes in engine health degradation in specific values ​​and lay a data foundation for subsequent prediction of remaining life under multiple failure modes.

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

[0242] The evaluation is iterated through all failure modes to generate health assessment results for each failure mode. (For training engines...) and test engine Generate a set of variable operating condition health assessments under multiple failure modes:

[0243]

[0244] (2033) Evaluation of the effectiveness of health assessment

[0245] This invention 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 demonstrate the effectiveness of the proposed health assessment method.

[0246] CV Curve Trend Indicator

[0247] The CV degradation curve generally exhibits a monotonic trend, which can characterize the engine degradation trend. For training engines, a significant degradation trend corresponds to the operational status of a training engine from healthy to complete failure. This paper proposes the Trend index to quantify the CV curve degradation trend, which is the sum of the increase or decrease in CV value within each flight cycle. The larger the index, the more significant the degradation trend. The specific calculation formula is as follows:

[0248]

[0249] in, Let n be the health index of the j-th engine in the i-th flight cycle under fault mode m, len be the operating time of that engine, and n be the health index of the j-th engine in the i-th flight cycle. m Let m be the number of engines in fault mode m, k be the number of fault modes, and n be the total number of engines.

[0250] CV curve volatility index Fluc

[0251] The CV degradation curve is relatively smooth with low variability, indicating a stable health assessment process and appropriate method. This paper proposes the Fluc index to quantify the CV curve variability, which is the sum of the absolute values ​​of the original generated CV value and the smoothed CV value for each engine. The smaller the index, the lower the CV curve variability. The specific calculation formula is as follows:

[0252]

[0253] in, and These are the raw calculated health index and the smoothed health index for the j-th engine in the i-th flight cycle under fault mode m, respectively, where len is the operating time of that engine, and n is the number of cycles. m Let m be the number of engines in fault mode m, k be the number of fault modes, and n be the total number of engines.

[0254] In summary, this invention extracts low-dimensional degradation information from multi-dimensional engine data based on an improved autoencoder model, which is then used as input data for health assessment under varying operating conditions. Engine health assessment under varying operating conditions is performed based on the NCA algorithm, yielding engine health assessment curves for different fault modes.

[0255] III. Multi-failure mode engine remaining life prediction based on metric learning and domain generalization

[0256] After determining the engine's health status trend under varying operating conditions, a suitable life prediction model under multiple fault modes needs to be established based on this. Addressing the differences in engine health assessment curves under multiple fault modes, training different life prediction models suffers from low prediction efficiency and low model generalization ability. This invention introduces deep metric learning into the life prediction framework. By designing a correspondence between "distribution difference - remaining life difference," it learns the nonlinear relationship between the temporal distribution differences and remaining life differences among health assessment indicators. Through learning the "difference" pattern, it reduces the gaps between multiple fault modes, thus making the prediction model applicable to multiple fault mode scenarios. Simultaneously, to address the problem of insufficient training samples and network susceptibility to overfitting, this invention designs a "direct + indirect" combined life estimation method. This method can estimate using all training samples, thereby achieving robust and accurate engine life prediction and improving the generalization ability of the prediction model under multiple fault modes. In this invention, a combination strategy of the classic deep metric learning Siamese network and the recurrent neural network GRU is used to predict life based on the engine health assessment results obtained under different fault modes.

[0257] Siamese neural networks are a classic method in deep metric learning. A Siamese network consists of two embedded networks that share weights, such as... Figure 13 As shown. Wherein, for the input sample (X) i X j The Siamese network is trained by minimizing the contrastive loss function.

[0258]

[0259] Where, if sample pair (X) i X j If both data points belong to the same category and have a value of 0, then Y ij It is a label with a value of 1; otherwise And α is the minimum margin of sample pairs of different classes.

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

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

[0262] 31) Construction of SN-GRU Multi-Fault Mode Remaining Life Prediction Model

[0263] (31) Design of SN-GRU model and training objectives

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

[0265] The SN-GRU model employs a structure connecting Siamese and GRU networks. The Siamese network serves as the time-series health metric measurement network, while the GRU network is the RUL prediction network. Here, G and P are the mapping functions of the Siamese embedding network and prediction network, respectively. By introducing a multi-task learning scheme, the training of the SN-GRU model primarily aims to achieve two objectives:

[0266] 31.1 First, by inputting the degradation curves (CVs) of different time periods into the Siamese network, the network is able to effectively calculate and learn the differences between them.

[0267] 31.2 Secondly, the degradation curve and its differences can be mapped to its RUL and RUL differences through the GRU prediction network, thereby learning the mapping relationship from the network and reducing the problem of inconsistent health indicator degradation caused by different failure modes.

[0268] Depending on network settings, θ G ,θ P The network hyperparameters are G and P, respectively. The first training objective of the proposed network is to minimize the individual prediction loss, i.e.:

[0269]

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

[0271] At the same time, the SN-GRU model can learn the non-linear relationship between differences in health indicators and differences in RUL, that is:

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

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

[0274]

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

[0276] The SN-GRU network comprises two shared embedding networks and a lifetime prediction network. For the embedding networks, a CNN architecture is employed, which outperforms other architectures for the lifetime prediction task within a reasonable training time. In the CNN architecture, 1-D convolutions are used to process the health assessment sequence through two convolutional layers, one flattened layer, and one fully connected layer, with appropriate dimensions and kernel sizes designed. The GRU prediction network consists of one GRU layer, two dropout layers, and two fully connected layers. Figure 15 This is a diagram of the SN-GRU network architecture.

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

[0278] This invention introduces Siamese networks into the Remaining Life (RUL) prediction problem, establishes an SN-GRU prediction model, and constructs an engine RUL prediction model suitable for multiple failure modes by designing the correspondence between "distribution difference" and "remaining life difference". The RUL prediction process for engines under multiple failure modes based on SN-GRU is as follows: Figure 16 As shown.

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

[0280]

[0281]

[0282] By designing the correspondence between "distribution difference and remaining life difference", the nonlinear relationship between the temporal distribution difference and remaining life difference among health assessment indicators is learned and measured. By learning the "difference", the gap between multiple failure modes is reduced, thereby making the prediction model applicable to multiple failure mode situations.

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

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

[0285] The engine CV results trained under different fault modes obtained in Chapter 3 are used as input to the SN-GRU model. Since the trained SN-GRU model is applicable to multiple fault modes, the CV indices under all fault modes are used as input for model training. This invention selects X1=X2={1,2,...,|CV|} from the training engine health index sequence and uses a sliding window for random sampling. and This serves as input for training in each iteration, thereby maximizing the use of shared knowledge across time series data. After similarity quantization of the sequences, if the prediction model can map randomly sampled CV data to the corresponding RUL, then the prediction model has stronger generalization ability.

[0286] (302) Design of SN-GRU life prediction model output under multiple failure 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 input samples of the set prediction model. and Then, by training the SN-GRU model, the mapping relationship between CV sequences and sequence differences and RUL sequences and RUL differences can be learned.

[0288] (303) SN-GRU model training

[0289] In each iteration, respectively with and As the input to the Siamese network, its corresponding and This is used as the output of the GRU prediction network. Then, it is processed using the gradient descent algorithm, based on the loss function designed in the previous stage. Update hyperparameter θ G ,θ P The trained SN-GRU model is obtained, which has strong generalization ability and can handle lifetime prediction problems under failure modes.

[0290] 33) Integrated lifetime estimation method based on SN-GRU

[0291] This invention designs a life estimation method that combines "direct and indirect" approaches, which can estimate the life of all training samples to achieve robust and accurate prediction of engine life and improve the generalization of the prediction model under multiple fault modes.

[0292] For a given test engine health index Two methods for predicting RUL are proposed.

[0293] (331) Direct RUL estimation

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

[0295]

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

[0297] (332) Indirect RUL estimation

[0298] Besides direct estimation, the RUL of the test engine can also be inferred based on the differences 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 it can be calculated by RUL j +P(G(CV i )-G(CV j To estimate RUL i Here, (CV) can be used. j ,RULj ) is considered as an estimate of CV i Reference. To increase the accuracy of RUL predictions on the test set, an estimate can be obtained as many times as the number of available training samples for each test sample, and these estimates can be 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, Defined as the indirectly estimated RUL, where |D| is the total number of estimated references.

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

[0302]

[0303] Where α and β are the fusion parameters.

[0304] Finally, the predicted RUL of each test engine is added to its last flight cycle 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, this invention presents a method for predicting the remaining service life (RUL) of engines under multiple failure modes based on deep metric learning and domain generalization. Building upon RUL prediction based on recurrent neural networks, deep metric learning is incorporated into the prediction strategy, providing a theoretical foundation for robust and accurate RUL prediction under multiple failure modes. To address the engine RUL prediction problem under multiple failure modes, this invention proposes an SN-GRU prediction model combining Siamese and GRU networks. By designing a correspondence between "distribution differences and remaining service life differences," the differences in health assessment indicators under different failure modes are reduced, while mitigating the impact of overfitting on the network, thus constructing an engine RUL prediction model applicable to multiple failure modes. To obtain more accurate RUL prediction results, this invention employs a "direct + indirect" RUL estimation method. This method combines estimation based on test set samples with auxiliary estimation using all training samples, increasing the sample size for estimation and achieving more reliable and accurate prediction results.

[0307] The following example illustrates the process of predicting the remaining life of a multi-failure mode engine.

[0308] Using the engine health assessment curves obtained under the two failure modes as the input basis for the life prediction model, a sliding window is performed with a sliding window length TW = 10 and a step size d = 1. Each CV i The input sample size is designed to be (None, 10, 2) to facilitate subsequent embedding network processing. Correspondingly, the RUL value of the last loop corresponding to the CV sample sliding window is used as the RUL. i .

[0309] In model training, the training engine data 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 for model optimization. The ratio of the training set to the validation set is 7:3. Based on the "maximum entropy principle," input sample pairs (CVs) are randomly selected from 249 training engine health assessment curve samples. 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 processes the health assessment sequence using 1-D convolutions, employing two convolutional layers, one flattened layer, and one fully connected layer, with corresponding dimensions and convolutional kernels. The embedding network shares the same structure and weights within the Siamese network. In the GRU prediction network, one GRU layer, one dropout layer, and two fully connected layers are designed. The convolutional layers, fully connected layers, and GRU layer use the ReLU activation function, while the output layer uses the Softmax activation function. The model structure design is shown in Table 31 below.

[0311] Table 31 SN-GRU Model Structural Design Table

[0312]

[0313]

[0314] The parameter settings for the SN-GRU remaining lifetime prediction model are shown in Table 32 below:

[0315] Table 32 Parameter Settings for SN-GRU Remaining Life Prediction Model

[0316]

[0317] As shown in the table above, in the SN-GRU remaining lifetime prediction model, the Adam optimization algorithm is used to optimize the hyperparameter θ in the Siamese network and the GRU prediction network. G ,θ PThe model is updated 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, with a maximum of 15,000 iterations. In the loss function calculation, the balancing parameter ω is set to 1, meaning that individual RUL predictions and difference predictions have equal weight in the loss function calculation. In RUL estimation, the fusion factors α and β are set to 0.6 and 0.4 respectively, resulting in a 3:2 ratio of direct RUL estimation to indirect RUL estimation.

[0318] In the SN-GRU remaining lifetime prediction model, the balance factor ω in the loss function is an important parameter during the training process, affecting the model's optimized training and final prediction accuracy. ω takes values ​​from (0,1). If ω is set to 1, individual RUL predictions and differential predictions have equal weight in the loss function calculation; if ω is set less than 1, differential predictions have a relatively smaller weight in the prediction. This section divides the range of the balance factor ω into four equal intervals: [0.01, 0.25], [0.25, 0.50], [0.50, 0.75], and [0.75, 1.00]. The endpoints of each interval are used as experimental values. The accuracy of the model on the training and validation sets is compared and analyzed, and the value with the highest accuracy on both sets is taken as the final value of ω. The balance factor parameter selection for training engine data is shown in Table 33 below.

[0319] Table 33 Model parameter balancing factors on training and validation sets ω Settings and prediction results

[0320]

[0321] Based on the accuracy training results of the SN-GRU model with different balance factor values, when the balance factor is at its lowest value (0.01), the model exhibits low accuracy on both the training and validation sets. This is because the loss function pays almost no attention to difference prediction, resulting in insufficient input data, overfitting, and low accuracy. As the balance factor value gradually increases, the accuracy on both the training and validation sets generally shows an upward trend. When the balance factor reaches 1, the model achieves the highest prediction accuracy on both sets. This is because the loss function gives equal importance to individual RUL prediction and difference prediction, ensuring sufficient training data and allowing the model to obtain balanced health assessment data from different failure modes from the difference metric. This reduces data discrepancies caused by multiple failure modes and improves the generalization ability of the SN-GRU model. Therefore, the model uses a balance factor ω of 1 during training, demonstrating the superiority of the proposed model in handling prediction problems under multiple failure modes.

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

[0323] When the fusion parameters are set to (0.6, 0.4), the engine evaluation results under multiple fault modes have the lowest RMSE, and the predicted RUL is more consistent with the actual RUL under different fault modes. Furthermore, when direct or indirect prediction is performed alone, the RUL prediction accuracy under multiple fault modes is low. This indicates that RUL estimation requires a combination of direct and indirect strategies to obtain prediction information from more reference data, thereby achieving better prediction results.

[0324] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the present invention.

Claims

1. A method for predicting the remaining life of a multi-failure mode engine, characterized in that, include: Obtain the health assessment time series curves of the training engine under different fault modes, and select the health assessment time series curves of the training engine under two fault modes. An SN-GRU model consisting of two Siamese networks and one GRU network is constructed, and the SN-GRU model is trained using the health assessment time series curves of the training engine under the two failure modes to obtain a trained SN-GRU model. Based on the trained SN-GRU model, the remaining life of the test engine is predicted, which includes: obtaining health index samples of the test engine under k failure modes. The health indicator samples are analyzed using the Siamese network and GRU prediction network in the trained SN-GRU model. After processing, the direct predicted RUL of the test engine is obtained, and its formula is: ； in, Here, RUL is directly estimated based on the SN-GRU model, and G is the Siamese network in the SN-GRU model. This refers to the GRU prediction network in the SN-GRU model. Obtain health indicator samples of the test engine under k failure modes. Second health indicator sample and its corresponding remaining lifespan Based on the trained SN-GRU model, and 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, and its formula includes: ； in, For RUL indirectly estimated based on the SN-GRU model, To estimate the total number of references; The direct and indirect predicted RUL of the test engine are fused to obtain the predicted RUL of the test engine under k failure modes, and the formula is as follows: ； in, and These are the fusion parameters.

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

3. The method according to claim 2, characterized in that, The SN-GRU model is trained using the health assessment time-series curves of the training engine under the two fault modes, resulting in a trained SN-GRU model including: By performing sliding window cutting on the health assessment time-series curves of the training engine under the two fault modes, health index samples under the first fault mode and health index samples under the second fault mode are obtained. Obtain the remaining lifetime corresponding to the health indicator sample under the first failure mode and the remaining lifetime corresponding to the health indicator sample under the second failure mode. In each iteration, the health index samples under the first failure mode and the health index samples under the second failure mode are used as the input to the Siamese network, while the remaining lifetimes corresponding to the health index samples under the first failure mode and the remaining lifetimes corresponding to the health index samples under the second failure 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 for the test engine based on the trained SN-GRU model also includes: The predicted RUL and final flight cycle of each test engine By adding them together, we can obtain the predicted engine life under k failure modes. Its formula is: 。 5. A multi-failure mode engine remaining life prediction device, characterized in that, include: The acquisition module is used to acquire the health assessment time series curves of the training engine under different fault modes, and select the health assessment time series curves of the training engine under two fault modes. The construction and training module is used to construct an SN-GRU model consisting of two Siamese networks and one GRU network, and to train the SN-GRU model using the health assessment time series curves of the training engine under the two failure modes to obtain a trained SN-GRU model. The prediction module, used to predict the remaining life of the test engine based on the trained SN-GRU model, includes: Obtain health indicator samples of the test engine under k failure modes. The health indicator samples are analyzed using the Siamese network and GRU prediction network in the trained SN-GRU model. After processing, the direct predicted RUL of the test engine is obtained, and its formula is: ； in, Here, RUL is directly estimated based on the SN-GRU model, and G is the Siamese network in the SN-GRU model. This refers to the GRU prediction network in the SN-GRU model. Obtain health indicator samples of the test engine under k failure modes. Second health indicator sample and its corresponding remaining lifespan Based on the trained SN-GRU model, and 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, and its formula includes: ； in, For RUL indirectly estimated based on the SN-GRU model, To estimate the total number of references; The direct and indirect predicted RUL of the test engine are fused to obtain the predicted RUL of the test engine under k failure modes, and the formula is as follows: ； in, and These are the fusion parameters.

6. 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, it implements the steps of the multi-failure mode engine remaining life prediction method according to any one of claims 1-4.

Images