An aero-engine life prediction method under multiple failure modes

Abstract

The application discloses an aero-engine life prediction method under a multiple fault mode, comprising the following steps: analyzing a training set of multiple aero-engines under the multiple fault mode, and determining multiple clustering categories of a performance degradation path of the training set aero-engines; training a life prediction model of the aero-engine of each clustering category according to training set data of the aero-engine of each clustering category, obtaining a trained life prediction model of each clustering category; testing the trained life prediction model of each clustering category according to test set data of the aero-engine of each clustering category and corresponding real RUL, and taking the life prediction model passing the test as a practical life prediction model; determining the clustering category to which a performance degradation path of a to-be-tested aero-engine belongs, and then determining the corresponding practical life prediction model, so as to predict the RUL of the to-be-tested aero-engine by using the corresponding practical life prediction model.

Description

Technical Field

[0001] This invention relates to the field of aero-engines, and in particular to a method for predicting the lifespan of aero-engines under multiple failure modes. Background Technology

[0002] Aero engines are the core system of aircraft. A malfunction in an aero engine will directly lead to systemic failures in the aircraft, resulting in serious consequences. Currently, improving their power performance and ensuring their long-term stable availability are key research areas and challenges in this field. Fault prediction and health management technologies can track and predict the engine's operational health status and provide timely and accurate maintenance recommendations before a malfunction occurs, effectively ensuring the long-term safe and reliable operation of the engine.

[0003] Existing research primarily focuses on life prediction based on individual differences and degradation characteristics of aero-engines, with limited research on the impact of different failure modes on the prediction results. Specifically, engines of the same model inevitably exhibit individual differences, including variations in the degree of factory-delivered faults, degradation rates, and degradation characteristics, making life prediction challenging. Current research on remaining service life (RUL) prediction for aero-engines mainly considers single-fault mode operation or ignores the impact of failure modes on RUL. However, actual aircraft operating conditions are complex, and single-fault mode conditions are idealized and difficult to achieve. In reality, experimental data collected by sensors often includes multiple failure modes.

[0004] In addition, different failure modes lead to different characteristics and trajectories of engine performance degradation. The degree of failure of aero engines deepens continuously during operation. In the early stage of operation, the failure modes are relatively weak, which makes it difficult to identify the performance degradation path of aero engines and further increases the difficulty of predicting the life of aero engines.

[0005] Effective engine RUL prediction can ensure timely and effective maintenance activities, thereby reducing casualties and losses caused by engine failures. Therefore, considering the economy and safety of aero engines, there is an urgent need for a diagnostic prediction technology for aero engines that can accurately predict the remaining service life of aero engines. Summary of the Invention

[0006] This invention provides a method for predicting the lifespan of an aero-engine under multiple failure modes, in order to solve the technical problem of accurately predicting the remaining lifespan of an aero-engine.

[0007] This invention provides a method for predicting the lifespan of aero-engines under multiple failure modes. The method includes: analyzing complete data from operation to failure of multiple aero-engines under multiple failure modes as a training set to determine multiple cluster categories of the performance degradation paths of the aero-engines in the training set; training a lifespan prediction model for each cluster category of aero-engines based on the training set data of the aero-engines in each cluster category to obtain a trained lifespan prediction model for each cluster category; using the data of multiple aero-engines before failure under multiple failure modes as a test set, and testing the trained lifespan prediction model for each cluster category based on the test set data of the aero-engines in each cluster category and the corresponding Real Remaining Service (RUL), and using the lifespan prediction model that passes the test as a practical lifespan prediction model; determining the cluster category to which the aero-engine under test belongs based on its performance degradation path, and determining the corresponding practical lifespan prediction model based on its cluster category; and predicting the RUL of the aero-engine under test using the corresponding practical lifespan prediction model.

[0008] Preferably, the step of analyzing complete data from operation to failure of multiple aero-engines under multiple failure modes as a training set to determine multiple cluster categories of the aero-engine performance degradation path in the training set includes: preprocessing the training set data to obtain training set data with dimensional effects removed and noise effects reduced; extracting degradation features related to the aero-engine performance degradation path from the training set data with dimensional effects removed and noise effects reduced; and clustering the time series of degradation features related to the aero-engine performance degradation path in the training set to obtain multiple cluster categories of the aero-engine performance degradation path in the training set.

[0009] Preferably, the preprocessing of the training set data to obtain training set data with the influence of dimensions removed and the influence of noise reduced includes: normalizing the training set data to obtain training set data with the influence of dimensions removed; and smoothing and denoising the training set data with the influence of dimensions removed to obtain training set data with the influence of dimensions removed and the influence of noise reduced.

[0010] Preferably, the step of extracting degradation features related to the performance degradation path of aero-engines from the training set data after removing the influence of dimensions and reducing the influence of noise includes: performing principal component analysis (PCA) on the training set data after removing the influence of dimensions and reducing the influence of noise to obtain multiple PCA features of the training set data; and selecting a specified number of PCA features from the multiple PCA features as degradation features related to the performance degradation path of aero-engines in descending order of the importance of the PCA feature variance.

[0011] Preferably, the step of clustering the time series of degradation features related to the performance degradation path of the aero-engine in the training set to obtain multiple cluster categories of the performance degradation path of the aero-engine in the training set includes: using the Dynamic Time Warping (DTW) algorithm to determine the similarity between the time series of degradation features of different aero-engines in the training set; and clustering the time series of degradation features of different aero-engines according to the similarity between the time series of degradation features of different aero-engines in the training set to obtain multiple cluster categories of the performance degradation path of the aero-engine in the training set.

[0012] Preferably, the method further includes: analyzing the test set to determine the cluster category to which the performance degradation path of the aero-engine in the test set belongs.

[0013] Preferably, the step of analyzing the test set to determine the cluster category to which the performance degradation path of the aero-engine in the test set belongs includes: preprocessing the test set data to obtain test set data with dimensional effects removed and noise effects reduced; extracting the time series of degradation features of the aero-engine in the test set from the test set data with dimensional effects removed and noise effects reduced; comparing the distance between the time series of degradation features of the aero-engine in the test set and the time series of the cluster center of each cluster category, and determining the cluster category of the cluster center with the smallest distance as the cluster category to which the performance degradation path of the aero-engine in the test set belongs.

[0014] Preferably, determining the cluster category to which the performance degradation path of the aircraft-to-be under test belongs includes: preprocessing all currently obtained data of the aircraft-to-be under test to obtain data with removed dimension effects and reduced noise effects; extracting the time series of degradation features of the aircraft-to-be under test from the data with removed dimension effects and reduced noise effects; comparing the distance between the time series of degradation features of the aircraft-to-be under test and the time series of the cluster centers of each cluster category, and determining the cluster category to which the cluster center with the smallest distance belongs to the cluster category to which the performance degradation path of the aircraft-to-be under test belongs;

[0015] Preferably, the life prediction model is a prediction model for the remaining service life of an aero-engine under multiple faults based on the Informer neural network.

[0016] The beneficial effects of this invention are that it can accurately predict the remaining service life of an aircraft engine, greatly ensuring the normal operation of the aircraft engine and improving maintenance and diagnostic efficiency. Attached Figure Description

[0017] Figure 1This is a flowchart of the aero-engine life prediction method under multiple fault modes provided by the present invention.

[0018] Figure 2 This is a detailed flowchart of the aero-engine life prediction under multiple fault modes provided by the present invention.

[0019] Figure 3a and Figure 3b These are the raw data graphs for 21 air path parameters of training engines No. 1 and No. 7, respectively.

[0020] Figure 4a , Figure 4b , Figure 4c These are the images of the 21 parameters of Engine 1 after preprocessing, the comparison image of the LPC outlet total temperature parameter before and after preprocessing, and the comparison image of the HPC outlet total pressure parameter before and after preprocessing.

[0021] Figure 5 This is a preprocessed data graph of all 21 air path parameters of the training engine;

[0022] Figure 6 This is a flowchart for identifying performance degradation paths;

[0023] Figure 7 This is a bar chart of the importance of PCA characteristic variance of aero-engine air path parameters;

[0024] Figure 8a , Figure 8b , Figure 8c These are graphs showing the relationship between feature 1, feature 2 and time after dimensionality reduction, and graphs showing the relationship between feature 1 and feature 2.

[0025] Figure 9a , Figure 9b , Figure 9c These are graphs showing the relationship between clustered features 1, features 2, and time, and graphs showing the relationship between features 1 and features 2.

[0026] Figure 10 This is a structural diagram of the prediction model for the remaining service life of an aero-engine under multiple failure modes;

[0027] Figure 11 This is a diagram of the location coding steps;

[0028] Figure 12 This is a flowchart illustrating the specific process of attention calculation;

[0029] Figure 13a It is a comparison chart of the predicted results and the actual values ​​after path recognition;

[0030] Figure 13b This is a comparison chart of the predicted results without path identification and the actual values; Detailed Implementation

[0031] 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.

[0032] Figure 1 This is a flowchart of the aero-engine life prediction method under multiple failure modes provided by the present invention, such as... Figure 1 As shown, the method may include: Step S101: Analyzing complete data from operation to failure of multiple aero-engines under multiple failure modes as a training set to determine multiple cluster categories of the performance degradation path of the aero-engines in the training set; Step S102: Training the life prediction model of each cluster category of aero-engines based on the training set data of each cluster category to obtain a trained life prediction model for each cluster category; Step S103: Using the data of multiple aero-engines before failure under multiple failure modes as a test set, and testing the trained life prediction model of each cluster category based on the test set data of each cluster category of aero-engines and the corresponding real remaining service life (RUL), and using the life prediction model that passes the test as the practical life prediction model; Step S104: Determining the cluster category to which the aero-engine under test belongs using the performance degradation path of the aero-engine under test, and determining the corresponding practical life prediction model using the cluster category to which it belongs; and predicting the RUL of the aero-engine under test using the corresponding practical life prediction model.

[0033] This invention achieves accurate prediction of the remaining service life of aero-engines by identifying aero-engine performance degradation paths based on trajectory clustering and predicting the remaining service life of aero-engines based on deep learning.

[0034] Step S101 may include: First, preprocessing the training set data to obtain training set data with dimensional effects removed and noise effects reduced. For example, the training set data is first normalized to obtain training set data with dimensional effects removed, and then smoothed and denoised to obtain training set data with dimensional effects removed and noise effects reduced, i.e., preprocessed training set data; Second, extracting degradation features related to the performance degradation path of aero-engines from the training set data with dimensional effects removed and noise effects reduced. For example, principal component analysis (PCA) is performed on the preprocessed training set data to obtain multiple PCA features and corresponding time series of the training set data, and then... The system selects a specified number (e.g., 2, 3, 4, etc.) of PCA features from multiple PCA features, ranked from largest to smallest importance, as degradation features related to the performance degradation path of aero-engines. Then, it clusters the time series of these degradation features in the training set to obtain multiple cluster categories for the performance degradation path of aero-engines in the training set. For example, it uses the Dynamic Time Warping (DTW) algorithm to determine the similarity between the time series of degradation features of different aero-engines in the training set. Based on this similarity, it clusters the time series of degradation features of different aero-engines to obtain multiple cluster categories for the performance degradation path of aero-engines in the training set.

[0035] Considering the impact of different performance degradation paths on the prediction of the remaining service life of aero-engines, this invention trains the service life prediction model according to the clustering categories of the performance degradation paths of aero-engines in the training set. For example, if there are two clusters, A and B, then the service life prediction model is trained using the training set data of aero-engines in cluster A, resulting in a trained service life prediction model for cluster A; similarly, the service life prediction model is trained using the training set data of aero-engines in cluster B, resulting in a trained service life prediction model for cluster B. It should be noted that the training set data used for training the model is the aforementioned preprocessed data.

[0036] After obtaining the trained lifetime prediction model for each cluster category, it is necessary to test the trained lifetime prediction model for each cluster category. In one implementation, if the cluster categories of the performance degradation path of the aero-engines in the test set are known, the trained lifetime prediction model for the corresponding cluster category can be directly tested based on the test set data of the aero-engines in the test set and the corresponding real remaining useful life (RUL). If the test is passed, the lifetime prediction model is used as a practical lifetime prediction model, i.e., a practical lifetime prediction model. In another embodiment, if the cluster category of the performance degradation path of the aero-engines in the test set is unknown, it is necessary to first analyze the test set to determine the cluster category to which the performance degradation path of the aero-engines in the test set belongs. This may include: preprocessing the test set data to obtain test set data with the influence of dimensions removed and the influence of noise reduced, the preprocessing method can refer to step S101; extracting the time series of the degradation characteristics of the aero-engines in the test set from the test set data with the influence of dimensions removed and the influence of noise reduced, the specific processing method can refer to step S101; and determining the cluster category to which the performance degradation path of the aero-engines in the test set belongs based on the time series of the degradation characteristics of the aero-engines in the test set and the time series of the cluster centers of each cluster category. For example, the distance between the time series of the degradation characteristics of the aero-engines in the test set and the time series of the cluster centers of each cluster category is calculated. The cluster category of the cluster center with the smallest distance is determined as the cluster category to which the performance degradation path of the aero-engines in the test set belongs. That is, the performance degradation path of the aero-engines in the test set is assigned to the cluster category of the cluster center that is closer to it. Then, based on the test set data of the aero-engines in the test set and the corresponding real remaining service life (RUL), the trained life prediction model for the determined corresponding cluster category is tested. If the test is passed, the life prediction model is used as a practical life prediction model, i.e., a practical life prediction model. It should be noted that in the above two embodiments, the test set data used by the test model is preprocessed test set data, and the preprocessing method can refer to step S101. By comparing the test results of step S103 with the prediction results of the prediction method for unidentified performance degradation paths, it can be clearly found that the prediction accuracy of the present invention is significantly improved. Therefore, the model can be applied to the RUL prediction of the aero-engine under test.

[0037] This invention predicts the remaining service life (RUL) of an aero-engine based on identified performance degradation paths. When applied to practical prediction, i.e., predicting the RUL of the aero-engine under test, all currently available data of the aero-engine under test, such as data from operation to the current moment, are preprocessed to obtain data with dimensional effects removed and noise effects reduced. The preprocessing method can refer to step S101. The time series of degradation characteristics of the aero-engine under test is extracted from the data with dimensional effects removed and noise effects reduced. The distance between the time series of degradation characteristics and the time series of the cluster centers of each cluster category is compared. The cluster category with the smallest distance is determined as the cluster category to which the performance degradation path of the aero-engine under test belongs. After determining the cluster category of the aero-engine under test, the corresponding practical service life prediction model can be determined. The data with dimensional effects removed and noise effects reduced of the aero-engine under test is input into the determined practical service life prediction model to predict the RUL of the aero-engine under test.

[0038] Figure 2 This is a detailed flowchart of the aero-engine life prediction under multiple failure modes provided by the present invention, such as... Figure 2As shown, the specific content is divided into three steps: Step 1: Preprocessing of aero-engine parameter data, which may include: normalizing all engine data, i.e., the raw data, sequentially (e.g., maximum and minimum value normalization) and denoising (e.g., local weighted regression), to obtain preprocessed data. Step 2: Identification of aero-engine performance degradation paths based on trajectory clustering. Specifically, after dimensionality reduction of the preprocessed data to extract degradation features (e.g., based on PCA), the problem of identifying performance degradation paths can be transformed into a trajectory clustering problem, which is divided into two steps: calculating the similarity between paths and clustering the paths. Since the paths have different lengths and different frequencies of change, they manifest as different engine lifespans and different degradation rates on the engine, so it is necessary to first calculate the similarity between sequences of different lengths. This invention can use the DTW method to calculate the similarity between engine performance degradation paths. In order to improve computational efficiency and make the cluster centers correspond to specific trajectories, the K-Medoids algorithm can be selected for clustering and identifying aero-engine performance degradation paths. Step 3: Prediction of remaining service life of aero-engines based on deep learning. Considering that this invention uses aero-engine operation simulation data as the research basis, and given the large number of engine samples, the large amount of sample data for each engine sample, and the large number of parameters contained in each sample data, the sample dataset is very large and complex. At the same time, the accuracy requirements for the final remaining service life prediction results are high. Therefore, a data-driven method based on deep learning can be selected as the method for predicting the remaining service life of aero-engines (or simply life prediction). For example, Informer can be used as the life prediction model network, and the network can be trained using existing samples to complete the prediction of the remaining service life of the engine.

[0039] The gas path components are a crucial part of aero-engines. Currently, performance degradation is primarily described by monitoring parameters such as temperature, pressure, and rotational speed of these components, including core rotational speed, HPC outlet temperature, and fan inlet pressure. Because the gas path components are closely related to the overall performance of the aero-engine and have a higher probability of failure compared to other components, the following examples use aero-engine gas path parameter data as a case study for detailed explanation.

[0040] 1. Preprocessing of aero-engine gas path parameter data.

[0041] There is a large amount of air path parameter data for aero engines, and there is multi-dimensional noise interference during the operation of aero engines. Therefore, it is necessary to normalize and reduce the noise of the air path parameters to eliminate the influence of parameter dimensions, data mutation points and noise, and improve the overall quality of the data.

[0042] Since sensor data from actual aero-engine operation is difficult to obtain, this embodiment uses the Data Challenge dataset from the 2008 PHM International Conference. This dataset was collected by the C-MAPSS model, a commercial turbofan engine operation simulation model developed by NASA. This model simulates the degradation process of five components of an aero-engine (including the fan, high-pressure compressor (HPC), etc.) under different operating conditions and failure modes, thereby obtaining moment-by-moment air path parameter data during aero-engine operation. The C-MAPSS simulation dataset includes four sub-datasets, as shown in Table 1. This embodiment addresses the case of multiple failure modes under a single engine operating condition; therefore, the third sub-dataset, FD003, is selected. In this dataset, all engines are of the same model and operate under the same conditions. The engines include two failure modes: HPC degradation and fan degradation, but the corresponding degradation paths are not labeled. For the engine degradation process in this dataset, each component undergoes degradation, but only the HPC or fan ultimately experiences a degradation-type failure; there is no situation where both components experience degradation-type failures simultaneously. Subdataset FD003 includes three text files:

[0043] a. Training set: Contains 100 aero-engine samples, measuring and recording complete data instances from operation to failure, which can be used to train predictive models.

[0044] b. Test set: Contains 100 aero-engine samples, measuring incomplete data instances that ended before failure, with the aim of predicting the RUL of aero-engines.

[0045] c. Test set remaining lifetime: Contains real RUL values ​​from 100 test aero-engine samples, used to compare with predicted values ​​and evaluate the prediction performance.

[0046] Each instance in the training and test set samples consists of a time series containing 26 variables. The first variable represents the engine number, the second variable represents the number of flight cycles of the engine, and the third to 26th variables represent engine operating parameters, including 3 operating condition parameters and 21 gas path parameters. The names of the 21 gas path parameters are shown in Table 2.

[0047] Table 1. Detailed Explanation of the C-MAPSS Dataset

[0048] property FD-001 FD-002 FD-003 FD-004 Aircraft engine operating condition data 1 6 1 6 Number of aircraft engine failures 1 1 2 2 Training set number of aircraft engines 100 260 100 249 Test set number of aircraft engines 100 150 100 248

[0049] Table 2. Specific Meanings of Engine Gas Path Parameters

[0050]

[0051]

[0052] A preliminary analysis of the engine data was conducted to grasp the overall trend of the gas path parameters during engine performance degradation. Figures 3a and 3b are the raw data graphs of 21 gas path parameters for training engines No. 1 and No. 7, with the horizontal axis representing flight cycles and the vertical axis representing the corresponding operating values. The visualization shows that some parameters do indeed change with the increase of flight cycles, reflecting the performance degradation of the aero-engine. However, for the same aero-engine, some gas path parameters show an overall upward trend, some show an overall downward trend, and some remain unchanged. For different aero-engines, some parameters show the same trend, while others show different or even opposite trends, and the magnitude of changes for parameters with the same trend may vary. Furthermore, the raw data contains inconsistent units of measurement, values ​​within different ranges, and significant noise, which will affect the accuracy of subsequent performance degradation path identification and remaining life prediction. Therefore, data preprocessing is necessary to improve data quality.

[0053] This embodiment employs data normalization processing based on the maximum and minimum value method and data smoothing and noise reduction processing based on the local weighted regression method.

[0054] 1) Data normalization processing based on the maximum and minimum value method.

[0055] All 21 gas path parameters were normalized. In this embodiment, the maximum and minimum value data normalization method was selected to standardize the parameters to the range of [0, 1]. This is beneficial for image plotting during parameter selection. The specific calculation formula is shown in Equation 1. Since subsequent methods are sensitive to the variance of initial variables, data normalization ensures that the contribution of all parameters is equal under all operating conditions, thus making the lifetime prediction more accurate and reasonable. For constant exponential parameters, such as the total temperature at the fan inlet, since its value does not change over time and does not contain degradation information, it is uniformly normalized to a constant of 0.

[0056]

[0057] in, It is the maximum value of the j-th parameter. It is the minimum value of the j-th parameter, x i，j These are the values ​​of the parameters before normalization. It is the value after parameter normalization.

[0058] 2) Data smoothing and noise reduction based on local weighted regression method.

[0059] While some gas path parameters are sensitive to the degradation process of aero-engines, the presence of numerous multi-dimensional noise signals during operation can negatively impact the accuracy of identifying performance degradation paths and predicting remaining service life if left untreated. This embodiment employs Lowess local weighted regression to smooth the gas path parameters after normalization, filtering out random noise from the sensor-collected signals, preserving degradation characteristics, and effectively reducing the impact of data abrupt changes on parameter degradation trends. This provides high-quality input data for establishing effective health assessment and remaining service life prediction models. The Lowess local weighted regression algorithm extends the algorithm to cases with multiple independent variables. Its core idea is to perform weighted fitting on local data to obtain estimated values ​​for the fitted points. The specific steps of this method are as follows:

[0060] (1) Determine the window range, which is used to control the scale of smoothing of locally weighted regression data.

[0061] (2) Within a defined window range n, for all points q k k = 1, 2, ..., n, through the weighting function ω k (q i ) for g i Perform d-order polynomial fitting.

[0062] (3) q was obtained through calculation i The fitted value p i , used to replace q i .

[0063] function ω k (q i The distribution of weights is determined by the formula. In this embodiment, a commonly used weighting function is selected, as shown in Formula 2.

[0064]

[0065] This function is an exponentially decaying function similar to a Gaussian distribution, meaning that points farther from the regression value have lower weights. In the formula, λ is the wavelength parameter, controlling the rate at which the weight decreases with distance; the weight increases with increasing λ. A key characteristic of locally weighted regression is its ability to adaptively change the parameters in the linear regression model as the independent variable values ​​change. For different independent variable values, the parameters in the model will change accordingly. Within the space of independent variables, the model will automatically provide an estimate of the function after locally weighted regression.

[0066] Figure 4a This is a preprocessed image of 21 parameters for engine No. 1. Figure 4bThis is a comparison chart of the total temperature parameters at the LPC outlet of training engine No. 1 before and after preprocessing. Figure 4c This is a comparison chart of the total pressure parameters at the HPC outlet of training engine No. 1 before and after preprocessing. The horizontal axis represents the flight cycle, and the vertical axis represents the corresponding values ​​before and after processing. Visualization shows that the preprocessed data is standardized to within [0, 1], removing the influence of dimensions. Simultaneously, the noise reduction process extracts the degradation trend from the original data while filtering out most of the noise information. This effectively reduces the impact of random noise fluctuations in the engine data, whether for upward or downward trending data. The preprocessed data has a good signal-to-noise ratio, resulting in high-quality data that lays a solid foundation for subsequent identification of aero-engine performance degradation paths and prediction of remaining service life.

[0067] A preliminary analysis was conducted on the preprocessed data. Figure 5 This is a preprocessed data graph of all 21 air path parameters of the training engine. The horizontal axis represents the remaining service life, and the vertical axis represents the corresponding operating value. The following conclusions can be drawn from the visualized data: (1) Some parameters show a changing trend with the increase of the number of operating cycles, proving that the performance of the aero-engine does indeed degrade as the operation progresses, which leads to changes reflected in the air path parameters. Therefore, the remaining service life of the aero-engine can be predicted by the parameter changes during the operation of the aero-engine. (2) The changing trends of some parameters, such as HPC outlet total pressure, physical core velocity, fuel flow, and corrected core velocity, visually show two clusters. This corresponds to the two fault modes of the aero-engine in the FD003 dataset. This proves that different fault modes do lead to different performance degradation paths, which are reflected in different change trajectories in the air path parameters. (3) For the parameters in (2), the parameters have already shown two distributions in the early stage of operation. This proves that under the influence of no external collisions or other interference, the random wear of the aero-engine at the factory determines its fault mode. Therefore, it is feasible to identify the future performance degradation path by the range of air path parameters in the early operation of the aero-engine. (4) Some parameters, such as LPC outlet temperature, HPC outlet temperature, and LPT outlet temperature, show a single trend as the number of operating cycles increases, proving that different performance degradation paths have little or no impact on these parameters, and their changes are mainly related to the number of operating cycles.

[0068] In summary, noise interference in the aero-engine's airflow parameters, if left untreated, can negatively impact the accuracy of performance degradation path identification and remaining service life prediction. Therefore, parameter data preprocessing is necessary. This embodiment primarily performs a preliminary analysis of aero-engine data, employing maximum-minimum normalization and local weighted regression smoothing to preprocess the parameters and improve data quality, laying the foundation for subsequent accurate performance degradation path identification and remaining service life prediction.

[0069] 2. Clustering and identification of performance degradation paths in aero-engines under multiple fault modes.

[0070] Because different aero-engines have varying degrees of wear at the time of manufacture and different rates of component degradation during operation, the faulty components at the time of failure differ, resulting in different failure modes. This leads to different directions and degrees of performance degradation, i.e., different performance degradation paths. The failure process of aero-engines is complex and involves many factors, meaning the characteristics of the degradation process are related to numerous factors. These factors may be caused by differences in the engine's own condition, different failure modes, or interference from other factors, which is detrimental to predicting remaining service life. Furthermore, the influence of performance degradation paths is superimposed on individual engine differences. If engines could be identified and differentiated according to their performance degradation paths, reducing the impact of these differences on remaining service life prediction, and thus more accurately grasping the relationship between engine performance degradation and remaining service life, the accuracy of predictions could be further improved. The preliminary analysis of the preprocessed data shows that the aero-engine gas path parameters do indeed exhibit different characteristics for different performance degradation paths. Accurately identifying these degradation paths and predicting the remaining service life of the aero-engine based on these different paths would significantly improve prediction accuracy. Therefore, this embodiment focuses on identifying aero-engine performance degradation paths, as illustrated in the flowchart below. Figure 6 As shown.

[0071] 1) Dimensionality reduction method for engine air path parameters based on principal component analysis.

[0072] Data-driven methods collect extensive sensor data from various gas path parameters reflecting aero-engine performance during operation, enabling data analysis to identify patterns among these parameters. While a large sample size of aero-engines and numerous gas path parameter variables provide rich information for studying aero-engine performance degradation trends, they also increase the workload of data computation and analysis. In many cases, strong correlations and redundancies exist among many gas path parameters, increasing the complexity of performance degradation path analysis and identification. Analyzing each gas path parameter individually often yields results that do not reflect reality and lack a holistic perspective. However, blindly reducing the number of gas path parameter data can lead to the loss of much information reflecting aero-engine performance degradation, resulting in inconsistencies between the results and actual conditions. Reducing the number of variables in the dataset undoubtedly lowers the overall data accuracy to some extent; however, dimensionality reduction aims to retain a small amount of data for computational convenience while preserving the majority of information, ensuring information filtering within an acceptable range of accuracy degradation. Compared to the complete and massive original dataset, the dimensionality-reduced dataset contains most of the degradation information with fewer degradation features, making it easier to explore and visualize the performance degradation path. Furthermore, machine learning algorithms can analyze the data faster and more efficiently without having to deal with variables that are irrelevant to the subject or have a very small relationship with it.

[0073] Because there is some correlation and redundancy among the various gas path parameter variables, a relatively small number of decay features can be used to represent the decay information present in each gas path parameter. Principal Component Analysis (PCA) can not only reduce the number of gas path parameters that need to be analyzed by subsequent models, but also minimize data loss, thereby enabling comprehensive and efficient analysis of the collected aero-engine gas path parameter data. The PCA method can effectively reduce the dimensionality of the dataset. By converting multiple variables into fewer features containing most of the information in the original dataset, most of the effective information can be extracted from the relatively complex and chaotic dataset. The mathematical principle of this method is to take the direction of the maximum variance of the input data matrix as the principal feature axis of the data, and arrange them in descending order of variance importance. The larger the variance, the more important the information in that dimension. Finally, a new matrix containing the data features is output. For the aero-engine data vector X = [X1, X2, ..., X...], ... m ], X i = [x i1 x i2 , ..., x in ] T (where m is the parameter number and n is the total number of data for the corresponding parameter) After that, the main steps of the PCA method are as follows:

[0074] (1) Calculate the covariance matrix of X, and calculate the eigenvalues ​​and eigenvectors;

[0075] (2) Sort the eigenvalues ​​in descending order and retain the eigenvector ξ corresponding to the largest eigenvalue;

[0076] (3) Obtain the dimension-reduced feature Y = Xξ of the gas path parameters;

[0077] (4) Integrate the reduced features into the engine data vector, repeat the above steps, and obtain the reduced engine features Z.

[0078] The quality of the extracted engine features determines the performance of the engine degradation process. Better feature extraction results in more distinct characteristics of the degradation process, leading to more accurate subsequent fault path identification and remaining service life prediction. PCA was performed on 21 gas path parameters to reconstruct a 21-dimensional feature space. Figure 7 This is a bar chart showing the importance of PCA feature variances of aero-engine gas path parameters. The horizontal axis represents the 21-dimensional features arranged in descending order of importance, and the vertical axis represents the corresponding variance importance. The bar chart shows that the importance of the first two features is 85.9%, the first three features is 93.6%, and the first four features is 99.4%. Therefore, the first two, three, or four features can be used for analysis and failure mode clustering.

[0079] Taking the first two features as an example, we verify the feasibility of this embodiment by creating a relationship diagram between each feature and time, and analyzing the relationship between each feature and the degradation mode. Figure 8a This is a graph showing the relationship between feature 1 and time. Figure 8b This is a graph showing the relationship between Feature 2 and time. The horizontal axis represents remaining service life, and the vertical axis represents the corresponding feature value. It can be seen that both Feature 1 and Feature 2 exhibit good clustering in the time dimension. From the relationship between Feature 1 and Feature 2 and time, it can be seen that for different aero-engine failure modes, the trends and ranges of their feature parameter changes differ to some extent. To explore whether there is further correlation between the synergistic relationship of Feature 1 and Feature 2 and the failure mode, a graph showing the relationship between Feature 1 and Feature 2 as a whole and the degradation mode is analyzed. Figure 8cAs shown, the horizontal axis represents the value corresponding to Feature 1, and the vertical axis represents the value corresponding to Feature 2. The trajectory represents the evolution of the aero-engine's condition from manufacturing to failure, reaching the final failure state point. It can be seen that Features 1 and 2 exhibit two clusters, showing good clustering results both in terms of feature trends and the final failure point. This corresponds to the two failure modes of the aero-engines in the FD003 dataset. Therefore, it is believed that the two different performance degradation paths correspond to the performance degradation caused by two different aero-engine failure modes, and the two different final failure point clusters correspond to two different engine failure modes. Therefore, trajectory clustering methods can be used to cluster the engine degradation feature curves, enabling clustering of the same failure mode and classification of different failure modes. This allows for the application of appropriate methods to predict the remaining service life for different failure modes, improving prediction accuracy.

[0080] 2) A performance degradation path similarity evaluation method based on dynamic time warping.

[0081] For the time series trajectories of Features 1 and Features 2, which contain most of the performance degradation information of aero-engines, the goal is to cluster these trajectories to achieve clustering of aero-engines based on their performance degradation paths. This requires calculating and comparing the similarity between the sequences. However, the lengths of two time series being compared for similarity may differ, even significantly. In the case of aero-engines, this manifests as individual differences and failure mode differences between different aero-engines, leading to variations in their lifespans. Furthermore, unlike the training set data, the test set data is not full-lifecycle data; it is truncated midway through operation, resulting in only the time series before the truncation. Moreover, the points are not in a one-to-one correspondence, making it impossible to simply use Euclidean distance for measurement. Therefore, it is necessary to locally scale the aero-engine gas path parameter time series on the time axis. The purpose of this operation is to ensure that the shape of the two time series maintains maximum consistency after this operation, thereby obtaining the maximum possible similarity between the two time series and achieving better comparison results.

[0082] The idea behind the Dynamic Time Warping (DTW) algorithm is to scale the time axis of the unknown sequence to match the length of the template sequence, thereby calculating and comparing the similarity between the two time series. It is very suitable for time series with different rhythm frequencies and sequence lengths. This invention applies it to the problem of clustering trajectories of aero-engine performance degradation.

[0083] For the aero-engine data vector X a =[x a1 x a2 , ..., x am ] T Xb =[x b1 x b2 , ..., x bn ] T (where a and b are the aircraft engine numbers, and m and n are the corresponding run cycles) The main steps of the DTW method are as follows:

[0084] (1) Construct an m×n matrix grid, where the matrix element (i, j) represents x. ai and x bj The distance d(x) between two points ai x bj Each matrix element (i, j) represents x. ai With x bj Alignment.

[0085] (2) Find a regular path through this grid and denote it as W: W = [w1, w2, ..., w k ], max{m, n}≤ k<m+n-1.

[0086] (3) Path boundary condition constraints, i.e., w1 = (1, 1), w k = (m, n), ensuring that the selected path always starts from the common starting point of the sequence and ends at the common ending point of the sequence.

[0087] (4) Path continuity constraint, if w p-1 = (r′, s′), then for the next point w on the path p-1 =(r, s) needs to satisfy (rr′)≤1 and (ss′)≤1, guaranteeing X a and X b Each coordinate in the matrix appears in W.

[0088] (4) Path monotonicity constraint, if w p-1 = (r′, s′), then for the next point w on the path p-1 = (r, s) needs to satisfy (rr′)≥0 and (ss′)≥0, and the points on W must be monotonically increasing with time.

[0089] (5) Repeat (2)(3)(4) to obtain the path with the minimum regularization cost. The K in the denominator is used to compensate for normalized paths of different lengths. The similarity (i.e., the value indicated by the path's endpoint) can be obtained by finding the path with the minimum normalization cost.

[0090] The similarity of time-series trajectories of features 1 and features 2 between different aero-engines is calculated using the DTW method. Subsequently, clustering methods can be used to cluster the time-series trajectories, thereby clustering aero-engines according to the similarity of their performance degradation trajectories. This allows for a more intuitive representation of the clustering results and enables the identification of aero-engine performance degradation paths.

[0091] 3) Performance degradation trajectory class method based on K-Medoids.

[0092] Clustering algorithms automatically group similar samples into the same category. Ultimately, data points within the same group should have similar attributes and / or features, while data points in different groups should have highly different attributes and / or features. K-Means is a simple and fast clustering technique. However, because K-Means lacks inclusiveness for noisy data, and for trajectory clustering, the mean similarity has no practical meaning (i.e., no trajectory corresponds to it), it can negatively impact the performance of subsequent test set data, hindering trajectory recognition and classification. The K-Medoids algorithm effectively addresses this problem. Comparative results show that K-Medoids significantly outperforms K-Means in terms of time for cluster centroid selection and cluster overlap space complexity. The biggest difference between K-Medoids and K-Means is that K-Means finds the mean of the cluster as the centroid, while K-Medoids, after finding the mean, selects the actual point closest to that mean as the centroid. In other words, K-Medoids uses the truly most existing point in the dataset as its centroid, thus finding the trajectory corresponding to the centroid. Therefore, this embodiment chooses the K-Medoids algorithm to complete the subsequent clustering task. The steps of the K-Medoids algorithm are as follows:

[0093] (1) Randomly select k samples as initial cluster centers a = [a1, a2, ..., a3] k ].

[0094] (2) For each sample x i Calculate the distance to each of the k cluster centers (i.e., the similarity calculated by the DTW method), and assign it to the cluster center with the smallest distance.

[0095] (3) Recalculate the cluster mean. Cluster center a j =min(a, a) j The new cluster center is the sample point closest to the cluster mean.

[0096] (4) Repeat steps (2) and (3) above until the requirements such as the number of iterations and the minimum error change are met.

[0097] 4) Analyze the clustering results of engine performance degradation paths.

[0098] The 21 gas path parameters of the aero-engine were dimensionality reduced using the PCA method to obtain features 1 and 2 with high variance importance and correlation with performance degradation paths. The time series trajectory data of features 1 and 2 were then subjected to similarity calculation using the DTW method and K-Medoids clustering. Figure 9a This is a graph showing the relationship between feature 1 and time. Figure 9b This is a graph showing the relationship between feature 2 and time. The horizontal axis represents the remaining useful life, and the vertical axis represents the corresponding feature value. Figure 9c This is a graph showing the relationship between Feature 1 and Feature 2, with the horizontal axis representing the value of Feature 1 and the vertical axis representing the value of Feature 2. As can be seen, the trajectory clustering algorithm used in this embodiment effectively demonstrates that the two clusters do not overlap, both in terms of the relationship between features and time and the relationship between features themselves. Furthermore, it exhibits good clustering results in both feature trends and the final failure point. This aligns with the fact that the aero-engines in the FD003 dataset contain two failure modes, allowing for the identification and differentiation of aero-engines according to different performance degradation paths, thereby improving the accuracy of subsequent remaining life prediction.

[0099] In summary, due to the numerous parameters of aero-engines and the significant redundancy among them, it is necessary to perform dimensionality reduction of the parameter data to extract degradation features and then identify their performance degradation paths. This embodiment primarily uses the PCA method to reduce the dimensionality of 21 gas path parameters of the aero-engine, the DTW method to calculate the similarity between the processed feature time series, and the K-Medoids method to cluster the aero-engine performance degradation paths, laying the foundation for subsequent accurate prediction of remaining service life.

[0100] 3. Prediction of remaining service life of aircraft engines under multiple failure modes.

[0101] Given the powerful processing capabilities of Transformer neural networks for sequential data, which have significant advantages in accuracy and efficiency compared to other network structures, Transformer neural networks were chosen as the prediction model. However, due to the computational burden and potential impact on accuracy caused by long time series, this embodiment ultimately selected an Informer neural network, which is an improvement on Transformer neural networks, as the prediction model for the remaining service life of aero-engines under multiple fault modes.

[0102] Due to the large volume of data from aero-engines (100 engines in total), with several sets of time-series data measured for each engine, and each set containing 21 gas path parameters, the information density of the data is relatively low. To improve information utilization and prevent insufficient information from affecting the accuracy of lifespan prediction, attention extraction is performed on the parameters to extract feature data with high information quality and strong performance degradation representation capabilities. Then, feature data from low-attention-prone paths are selectively removed to obtain feature data with higher information density for lifespan prediction. This method can reduce computational load while avoiding the influence of low-quality data on prediction results.

[0103] 1) Engine remaining service life prediction model under multiple fault modes based on Informer neural network.

[0104] The Informer model uses the ProbSparse self-attention mechanism, extracting cascaded layers to highlight the dominant attention, resulting in excellent sequence dependency alignment performance and significant improvements in time complexity and memory utilization, greatly increasing the computational speed for long sequence prediction. Figure 10 The diagram shown illustrates the structure of an aero-engine remaining service life prediction model under multiple failure modes. The main framework of the model consists of four parts: position encoder, encoder, decoder, and fully connected layer. The specific steps are as follows:

[0105] (1) Establish two life prediction models, corresponding to two performance degradation paths. All aero engines are divided into two categories according to the performance degradation path, and the preprocessed time series data of all aero engines are used as the model input.

[0106] (2) The model input is converted into the encoder input after position encoding and then fed into the encoder to attach position information.

[0107] (3) The encoder part contains three encoder layers. Each encoder layer includes four steps: multi-head ProbSparse self-attention mechanism, residual connection and normalization, feedforward network, residual connection and normalization, and finally the connection feature map (including Query matrix and Key matrix) is passed into the decoder.

[0108] (4) The preprocessed time series data of the aero-engine is converted into the input of the decoder after being encoded by position and then fed into the decoder to attach position information.

[0109] (5) After passing through the Masked multi-head ProbSparse self-attention mechanism, the influence of the later time series on the earlier time series is eliminated. That is, the model only knows the data before each time and does not know the data after that time. Residual connection and normalization are then performed.

[0110] (6) After processing by the feedforward network, residual connections and normalization are performed.

[0111] (7) Through the encoder-decoder attention mechanism, the Value matrix is ​​passed in step (4), and the Query matrix and Key matrix are passed in step (2), and residual connection and normalization are performed.

[0112] (8) After processing by the feedforward network, residual connections and normalization are performed.

[0113] (9) The decoded output is used as the input of the fully connected layer and is eventually transformed into the output of the final model, namely the predicted value of RUL.

[0114] (10) The error between the predicted RUL value and the actual RUL value is used as the loss function to calculate the loss. The relevant parameters in the network are updated through backpropagation, and the training is repeated until the loss meets the requirements.

[0115] For positional encoding, to accelerate computation, elements in the input sequence are processed together, rather than one by one as in an RNN. However, this ignores the order of elements in the sequence. Therefore, positional encoding is needed for each element in the sequence, enabling the model to learn the positional information of each element. This embodiment uses absolute positional encoding to label the elements in the sequence. For the time series of aero-engine gas path parameters W = {W1, W2, W3, ..., W...} n}, W i ={w1, w2, w3, ..., w 21}, where n is the length of the time series, and each W i It contains 21 gas path parameters, which is a 21-dimensional vector. Position encoding PE is applied to W, resulting in the model input X = {W1 + PE1, W2 + PE2, W3 + PE3, ..., W...} n +PE n The formulas for calculating PE are shown in Equations 3 and 4.

[0116]

[0117]

[0118] Where pos is the position of w in the sequence, i is the position of the parameter in w, and d model The dimension for encoding w is 21. The positional encoding steps are shown in the diagram below. Figure 11 As shown.

[0119] The sine function was chosen because for any fixed offset k, PE pos+k It can be represented as PEpos The linear function. By encoding the time series of aero-engine gas path parameters by position, the model obtains not only specific data information but also the corresponding position information of the sequence, thereby improving the accuracy of model predictions.

[0120] Regarding the ProbSparse attention mechanism, the core of the Transformer model is the attention mechanism. The main process involves calculating the Q (Query), K (Key), and V (Value) of the data. Q corresponds to the sequence to be represented (called sequence A), and K and V correspond to the sequences used to represent A (called sequence B). The attention calculation formula is shown in Equation 5, where dk is the square root of the K dimension. The specific flowchart of the attention calculation is shown below. Figure 12 As shown.

[0121]

[0122] K and Q must satisfy the prerequisite that they exist in the same high-dimensional space (otherwise, computation is impossible), while V does not necessarily need to exist in the same high-dimensional space as K and Q; it only needs to satisfy that the output of the final model exists in the same high-dimensional space as V. The discrete distribution obtained after softmax for each Q dot product with all K in the sequence is different. If the distribution of a Q is similar to a uniform distribution, then each probability value approaches the reciprocal of the mean and is very small, belonging to the tail of a long-tailed distribution; such a Q provides little value. Conversely, if the distribution of a Q differs greatly from a uniform distribution, then there must be several high probability values. These high probability values ​​dominate the probability distribution and are very important in self-attention mechanisms; therefore, such a Q is valuable. Visualizing the dot product results in self-attention mechanisms reveals that they follow a long-tailed distribution, meaning that the dot product results of a few Qs and Ks dominate the distribution after softmax. This sparse distribution means that an element in the sequence generally only has high similarity / correlation with a few other elements.

[0123] The core idea of ​​the ProbSparse self-attention mechanism is to find these important sparse queries and then only compute the attention values ​​of these queries to optimize computational efficiency for Q∈R. m×d , K∈R n×d , V∈R n×d The specific steps are as follows:

[0124] (1) Sample K to obtain K_sample, with a sampling length L. k =mlnn, for each q i The value of M is calculated for ∈Q with respect to K_sample. The calculation formula is shown in Equation 6.

[0125]

[0126] (2) Calculate the q values ​​of u with the largest M value. i The new matrix is ​​denoted as

[0127] (3) Calculation

[0128] By extracting attention from the parameters, we can extract feature data with high information quality and strong expressive ability despite performance degradation. We can also purposefully extract feature data from low attention-attention paths to obtain feature data with high information density, which helps to improve the accuracy of model prediction.

[0129] 2) Evaluation of the prediction results of the remaining useful life prediction model.

[0130] The time series data of the aero-engine gas path parameters after performance degradation trajectory identification are input into the aero-engine remaining service life prediction model based on Informer multi-fault mode to obtain the model's prediction results.

[0131] In this embodiment, two metrics are selected: root mean square error (RMSE) and score function (Score) to evaluate and measure the performance of the prediction model. The lower the evaluation metric, the better the prediction effect.

[0132] RMSE assigns equal weight to prediction errors in the early and late stages of a recession, and amplifies larger deviations by squaring the errors. Therefore, it is more sensitive to predictions that deviate significantly from the actual values. For the predicted value of RUL, y i is the actual value, and n is the total number of predicted values. The specific formula is shown in Equation 7.

[0133]

[0134] The score assigns a larger penalty score to the predicted remaining lifetime value that exceeds the actual remaining lifetime value, which examines the ability to predict ahead and is of great significance for predictive maintenance. The specific formula is shown in Equation 8.

[0135]

[0136] Figure 13a This is a comparison chart of the predicted results and the actual values ​​after path recognition. Figure 13bThis is a comparison chart of the predicted results and actual values ​​before path identification. The horizontal axis represents the engine number, and the vertical axis represents the corresponding remaining service life. The solid line represents the actual value, and the dashed line represents the model's predicted value. It can be seen that after identifying the performance degradation path, the accuracy of engine life prediction is significantly improved compared to before the path identification. Table 3 compares the prediction results before and after path identification, using RMSE and Score to compare the prediction results. It can be seen that after identifying the performance degradation path, the accuracy of engine life prediction is improved by 26.4% compared to before the path identification (using RMSE as the standard for measuring accuracy). This proves that the remaining service life of the aero-engine is correlated with the performance degradation path of that aero-engine. Therefore, identifying the performance degradation path of the aero-engine plays an important role in the prediction of remaining service life and can effectively improve the accuracy of life prediction.

[0137] Table 3. Comparison of prediction results before and after path recognition

[0138]

[0139] Table 4 compares the predictive capabilities of different prediction models, all of which are predictions after identifying performance degradation paths. As shown in the table above, the Informer model achieved the best Score and RMSE results, and has a significant advantage over other models, demonstrating better predictive capability for the remaining service life of aero-engines under multiple failure modes.

[0140] Table 4. Comparison of Prediction Results of Different Prediction Models

[0141]

[0142] In this embodiment, since the air path parameter data of aero-engines exhibits significant temporal sequence characteristics, and the Informer neural network can effectively learn the development trend of temporal data, this embodiment establishes an aero-engine remaining service life prediction model based on the Informer neural network model. Experiments show that the engine remaining service life prediction model based on the Informer neural network model can predict the remaining service life of aero-engines under multiple failure modes quite well, exhibiting significant advantages compared to other models. Furthermore, data after identifying performance degradation paths can effectively improve the accuracy of service life prediction.

[0143] Thus, when performing RUL prediction on actual aero engines under multiple failure modes, the model in this embodiment can be used to perform more accurate and effective RUL prediction.

[0144] In summary, the present invention has the following advantages:

[0145] (1) This invention provides an efficient autonomous clustering identification method for performance degradation paths of aero engines under multiple fault modes. The method evaluates the path similarity between multiple engines with different decay rates and different fault initiation degrees through the DTW method. Based on this, the K-Medoids performance degradation path clustering method is introduced to achieve rapid and accurate classification of a large number of aero engine degradation paths.

[0146] (2) This invention proposes an aero-engine remaining service life prediction model based on Informer deep neural network. By comprehensively considering the impact of performance degradation path on engine remaining service life prediction, high-precision service life prediction is achieved. The prediction accuracy is improved by 26.4% compared with the one using performance degradation path identification.

[0147] (3) This invention takes failure modes as a consideration factor to identify the performance degradation path of aero engines. It brings aero engines with the same performance degradation path into the same model for training and prediction, which is beneficial to improve the accuracy of remaining life prediction.

[0148] 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 lifespan of an aero-engine under multiple fault modes, characterized in that, The method includes: The complete data of multiple aero-engines from operation to failure under multiple failure modes is used as a training set for analysis to determine multiple cluster categories of the aero-engine performance degradation path in the training set. The analysis includes: normalizing and denoising the training set data to obtain preprocessed data; performing Principal Component Analysis (PCA) on the preprocessed data to obtain multiple PCA features; selecting a specified number of PCA features as degradation features related to the aero-engine performance degradation path in descending order of PCA feature variance importance; and clustering the time series of degradation features related to the aero-engine performance degradation path to obtain multiple cluster categories of the aero-engine performance degradation path in the training set. Based on the training set data of aero-engines in each cluster, the life prediction model of aero-engines in each cluster is trained to obtain the trained life prediction model of each cluster; wherein, the life prediction model is an aero-engine remaining service life prediction model based on Informer deep neural network. The test set is used to collect data from multiple aero-engines before failure under multiple failure modes. Based on the test set data and the corresponding real remaining service life (RUL) of aero-engines in each cluster, the trained life prediction model for each cluster is tested. The life prediction model that passes the test is used as the practical life prediction model. By utilizing the performance degradation path of the aero-engine under test, its cluster category is determined, and the corresponding practical life prediction model is determined using its cluster category. The RUL of the aero-engine under test is predicted using the corresponding practical life prediction model.

2. The method according to claim 1, characterized in that, The clustering of time series of degradation features related to the performance degradation path of aero-engines yields multiple clustering categories for the performance degradation path of aero-engines in the training set, including: The similarity between time series of degradation features of different aero-engines in the training set was determined using the Dynamic Time Warping (DTW) algorithm. Based on the similarity between the time series of degradation features of different aero-engines in the training set, the time series of degradation features of different aero-engines are clustered to obtain multiple cluster categories of the performance degradation path of the aero-engines in the training set.

3. The method according to claim 1, characterized in that, The test set data for each cluster of aero-engines was obtained through the following steps: The test set data is preprocessed to obtain test set data with the influence of dimensions removed and the influence of noise reduced; Extract the time series of degradation characteristics of the aero-engines in the test set from the test set data after removing the influence of dimensions and reducing the influence of noise; By comparing the time series of degradation characteristics of the aero-engines in the test set with the time series of cluster centers of each cluster category, the cluster category of the cluster center with the smallest distance is determined as the cluster category to which the performance degradation path of the aero-engines in the test set belongs, thereby obtaining the test set data of the aero-engines in each cluster category.

4. The method according to claim 1, characterized in that, The method of determining the cluster category to which the test aero-engine belongs by utilizing its performance degradation path includes: All currently available data of the aircraft engine under test are preprocessed to obtain data with the influence of dimensions removed and the influence of noise reduced; Extract the time series of degradation characteristics of the aero-engine under test from the data of the aero-engine under test after removing the dimension effects and reducing the noise effects; By comparing the time series of the degradation characteristics of the aero-engine under test with the time series of the cluster centers of each cluster category, the cluster category of the cluster center with the smallest distance is determined as the cluster category to which the performance degradation path of the aero-engine under test belongs.

Images