Civil aviation engine residual life prediction method considering covariable

Abstract

The invention relates to the technical field of civil aviation engine health state management, in particular to a civil aviation engine residual life prediction method considering covariables, which can effectively improve the accuracy of residual life prediction under various working conditions. The influence of covariables on target variables is captured through time step dynamic modeling, civil aviation engine washing data serves as future covariables, and a TFT time sequence prediction model with static covariables and past covariables removed is adopted as a residual life prediction model.

Description

Technical field:

[0001] The present invention relates to the technical field of civil aviation engine health status management, and in particular to a civil aviation engine remaining life prediction method taking covariates into consideration and capable of effectively improving the accuracy of remaining life prediction under various working conditions. Background technology:

[0002] Aircraft engine health management (EHM) plays an important role in the aviation industry. EGTM, as an important parameter for monitoring the health status of the engine, directly reflects the changes in the core performance of the engine. Nowadays, when studying the remaining life of aircraft engines, EGTM is usually used as the remaining life standard. When EGTM reaches the red line value, the aircraft engine is considered to have reached the end of its life. At present, the common methods for health trend prediction can be divided into three categories according to the basic technology, namely, methods based on physical models, methods based on statistical models, methods based on artificial intelligence, and hybrid methods. Among them, the methods based on physical models require prior knowledge of aircraft engine physics, which brings difficulties to the research on aircraft engine health trend prediction. The most commonly used methods are methods based on statistical models and methods based on artificial intelligence.

[0003] The RUL prediction method driven by physical models can achieve higher prediction accuracy, but due to the complex structure and working environment of aircraft engines, it is difficult to accurately construct its physical model. At the same time, the physical models of engines of different types and models are also different, which makes the models built with a lot of resources unable to be applied to other models or types of aircraft engines, and the versatility and scalability of the models are not strong. This also limits the application of RUL prediction methods driven by physical models in PHM research of aircraft engines. The RUL prediction method driven by statistical data does not need to consider the failure mechanism of the system, and can also output the uncertainty expression of RUL. However, the application of such methods requires certain prior knowledge, such as the failure threshold of aircraft engines. This is often difficult to determine in engineering practice. At the same time, in practical applications, the degradation of aircraft engines does not necessarily obey a certain distribution or a certain random process, which leads to the fact that the accuracy of the prediction cannot be guaranteed. The RUL prediction using artificial intelligence algorithms is simple and fast. Not only does it not need to consider the failure mechanism of the system, but it can also obtain higher prediction accuracy. However, the model training takes a long time, and only the point estimate of RUL can be given, but the uncertainty estimation result cannot be given.

[0004] Water washing, as a common measure of engine maintenance, will cause significant changes in EGTM parameters, but existing studies rarely explore the impact of water washing on EGTM prediction. Summary of the invention:

[0005] In order to more accurately predict the performance degradation trend and remaining service life of an aircraft engine, the present invention proposes a method for predicting the remaining service life of a civil aviation engine taking covariates into consideration.

[0006] The present invention is achieved by the following measures:

[0007] A method for predicting the remaining life of a civil aviation engine considering covariates is characterized in that future covariates are used as input features and the influence of covariates on target variables is captured through time-step dynamic modeling, wherein civil aviation engine water washing data is used as future covariates, and the remaining life prediction model adopts a TFT time series prediction model that removes static covariates and past covariates.

[0008] The water washing data record of the civil aviation engine described in the present invention is the water washing operation information of a certain engine at a specific time point, including the engine identification, the time point and whether the water washing operation is performed. Specifically, water washing is a regular maintenance measure, and its main purpose is to restore its performance by cleaning the internal deposits of the engine. The water washing data is represented as a discrete event record in the time series. In order to enable the water washing data to be used as a covariate in the prediction model, the water washing data needs to be encoded. The encoding rule is based on the cumulative number of water washing operations, which is converted into a numerical variable for model processing. If the engine is in After overhaul or in new state, if no water washing operation is performed, the code is 0; each time a water washing operation occurs, the code value increases by 1. For example, an engine that has not been water washed after overhaul is coded as 0, one water washing is coded as 1, two water washings are coded as 2, and so on. The coding process includes arranging the water washing records in chronological order, accumulating the occurrence of each water washing operation from the time of overhaul or delivery, and updating the corresponding code value for each time point; this coding method not only converts discrete water washing events into quantitative features, but also simplifies the expression of maintenance information, while enhancing the model's ability to capture the impact of maintenance events. Compared with the coding form that is only set to 1 at the water washing point and 0 at other places, this coding form can better capture the improvement effect of aircraft engine water washing on EGTM, and can also capture the information of the number of water washings, improving the effectiveness of water washing covariates. Through this process, water washing data is effectively constructed as future covariates, providing key support for the remaining life prediction model to describe the impact of maintenance.

[0009] The remaining life prediction model of the present invention comprises a variable selection network, an LSTM Encoder module, an LSTM Decoder module, an Add&Norm Gate module, a masked multi-head attention module, an Add&Norm Gate module, a gated residual network GRN, and a Dense module, wherein the feature data χ is sequentially sent to the masked multi-head attention module after passing through the variable selection network, the LSTM Encoder module, the Add&Norm Gate module, and the gated residual network GRN, and the future covariate γ is sequentially sent to the masked multi-head attention module after passing through the variable selection network, the LSTM Decoder module, the Add&Norm Gate module, and the gated residual network GRN, and then processed by the Add&Norm Gate module, the gated residual network GRN, the Add&Norm Gate module, and the Dense module respectively and then output.

[0010] The present invention establishes an aircraft engine remaining life prediction model based on a simplified Temporal Fusion Transformer (TFT) model. By introducing water washing data as future covariates, the influence of water washing operation on engine performance is effectively utilized through sequential coding. Compared with the traditional 0-1 coding and non-covariate models, sequential coding shows significant advantages in short-term prediction and can better adapt to engine data of different thrust levels. Experiments were conducted under different prediction window lengths, and the results showed that sequential coding can significantly reduce the MAE (Mean Absolute Error) error, especially in short-term prediction. The experiment verified the robustness of the proposed method and proved that it can effectively improve the accuracy of remaining life prediction under various working conditions. Description of the drawings:

[0011] Attached Figure 1 This is the EGTM decline diagram of a certain engine.

[0012] Attached Figure 2 It is a framework diagram of the remaining life prediction model in the present invention.

[0013] Attached Figure 3 It is a GRN unit framework diagram of the remaining life prediction model in the present invention.

[0014] Attached Figure 4 It is a schematic diagram of the variable selection network of the remaining life prediction model in the present invention. Specific implementation method:

[0015] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0016] Water washing is a common measure for engine maintenance, which will cause significant changes in EGTM parameters, but existing studies have rarely explored the impact of water washing on EGTM prediction. In order to more accurately predict the performance degradation trend and remaining service life of aircraft engines, the present invention proposes a prediction of the remaining service life of aircraft engines considering covariates.

[0017] During the actual operation of aircraft engines, the exhaust gas temperature margin (EGTM) is an important parameter for measuring the health status. However, EGTM does not always show a steady degradation trend, but is significantly affected by maintenance operations. For example, water washing operation is a common engine maintenance method that can improve combustion efficiency by removing internal engine deposits. During this process, EGTM usually increases suddenly, which is manifested as performance improvement in the short term. However, this sudden increase will interfere with the existing remaining life prediction methods based on degradation trends, reducing the reliability of the prediction results. At present, the remaining life prediction methods of aircraft engines are mainly divided into physical model methods and data-driven methods. The physical model method relies on complex physical process modeling and is suitable for describing the mechanism of engine degradation, but the model construction is complex and has limited adaptability to actual scenarios; the data-driven method uses historical operation data to make predictions through machine learning or statistical methods, which has strong flexibility and accuracy. However, these methods usually assume that the engine degradation process is continuous and stable, and fail to fully consider the nonlinear interference of operations such as water washing on EGTM. The limitation of existing methods lies in the insufficient consideration of the role of covariates. Covariates (such as water washing operation, environmental conditions, operating mode, etc.) can significantly affect the degradation behavior and remaining life of the engine, but they are often ignored or simplified in traditional prediction models. In order to improve the accuracy and practicality of the remaining life prediction, it is of great significance to incorporate covariates into the model. Specifically, the water washing operation data of aircraft engines as covariates has the following necessity and advantages: the sudden increase in the impact of water washing on EGTM is a significant factor that cannot be ignored, and it is difficult to accurately capture the degradation trend after the sudden increase by relying solely on historical degradation data. The introduction of water washing data can effectively make up for this deficiency. By modeling the impact of water washing operation on EGTM, the adaptability of the prediction model to the degradation trend after the sudden increase can be improved, and the accuracy of the algorithm prediction can be improved. Including water washing data as covariates helps to quantify the impact of maintenance operations on the health status of the engine, support more scientific maintenance decisions, and make the prediction process more interpretable. The introduction of covariates enables the prediction model to be dynamically adjusted to adapt to complex changes under different maintenance conditions and operating environments. In summary, using water washing operation data as a covariate is not only a key way to improve the existing aircraft engine remaining life prediction method, but also an important basis for promoting the development of prediction technology from static to dynamic and from a single degradation trend to multi-factor coupling.

[0018] Future covariates refer to external variables used in prediction models to describe the impact of specific events or operations on system states. These variables are usually dynamically related to time and can reflect the mechanism of action of key intervention operations. In the prediction of the remaining life of aircraft engines, the introduction of future covariates can help capture the nonlinear effects of maintenance operations such as water washing on engine performance, thereby improving the accuracy and robustness of the prediction model.

[0019] The form of water washing data is usually recorded as the water washing operation information of a certain engine at a specific time point, including engine identification, time point, and whether the water washing operation was performed. Specifically, water washing is a regular maintenance measure, and its main purpose is to restore its performance by cleaning the internal deposits of the engine. Therefore, this kind of data is presented as a discrete event record in the time series, which clearly identifies the water washing status of the engine during operation.

[0020] In order to enable the water washing data to be used as a covariate in the prediction model, it needs to be encoded. The coding rule is based on the cumulative number of water washing operations, which are converted into numerical variables for model processing. Specifically, if the engine is in a post-overhaul or new state and no water washing operation has been performed, the code is 0; each time a water washing operation occurs, the code value increases by 1. For example, an engine that has not been washed after an overhaul is coded as 0, a water washing once is coded as 1, a water washing twice is coded as 2, and so on. The coding process includes organizing the water washing records in chronological order, cumulatively recording the occurrence of each water washing operation from the time of overhaul or delivery, and updating the corresponding code value for each time point.

[0021] This encoding method not only converts discrete water washing events into quantitative features, but also simplifies the expression of maintenance information and enhances the model's ability to capture the impact of maintenance events. Compared with the encoding format that is set to 1 only at the water washing point and 0 at other locations, this encoding format can better capture the effect of aircraft engine water washing on EGTM, and can also capture the information of the number of water washings, thereby improving the effectiveness of water washing covariates. Through this process, water washing data is effectively constructed as future covariates, providing key support for the remaining life prediction model to describe the impact of maintenance.

[0022] Temporal Fusion Transformer (TFT) is an advanced time series forecasting algorithm that combines a multi-head attention mechanism with a recurrent neural network (RNN) to effectively process multivariate time series data. The core advantages of TFT lie in its flexibility and interpretability: by capturing the dependencies between different time steps and a dynamic weight assignment mechanism, TFT can accurately predict future trends in complex dynamic systems and provide a detailed explanation of feature importance. In addition, TFT combines static covariates, past covariates, and future covariates, making it perform well in processing multidimensional input data, especially for time series data with complex interactions.

[0023] In the present invention, a simplified version of the TFT algorithm is used, as shown in the attached Figure 2 As shown in the figure, the focus is on the future covariate part, while the modeling of static covariates and past covariates is removed. This simplified model design process includes: first, using future covariates as input features, capturing the impact of covariates on target variables through time-step dynamic modeling; second, optimizing the model structure to a lighter form, retaining the multi-head attention mechanism to improve the flexibility and accuracy of prediction. Through this simplification, the model is significantly reduced in complexity and is more suitable for the prediction needs of small-scale data sets and specific scenarios.

[0024] Combined with the actual needs of aircraft engine remaining life prediction, the introduction of future covariates can dynamically capture the impact of water washing operations on engine performance, thereby enhancing the model's adaptability to nonlinear changes; secondly, removing static and past covariates makes the model structure more compact and computationally efficient, and highly matches the time series data containing only water washing covariates. In addition, the simplified model has reduced dependence on data scale and feature dimension during training, and can more efficiently process aircraft engine water washing records and performance degradation data. Overall, the simplified TFT model achieves accurate modeling of key covariates with lower complexity, providing effective support for the remaining life prediction task of the present invention.

[0025] Example:

[0026] In order to verify the effectiveness of the model and method proposed in this paper, the data set used in this example comes from an airline company, covering the EGTM data of aircraft engines with three different thrust levels (22KIbf, 24KIbf, and 26KIbf) and the corresponding aircraft engine water washing operation data. This data set has rich actual operation information and maintenance records, providing a solid foundation for experimental verification.

[0027] The advantages of this dataset are mainly reflected in the following aspects: First, the data covers aircraft engines of various thrust levels, making the data representative and diverse, and able to reflect the performance degradation law of engines under different thrust conditions. Second, the data contains detailed water washing records, which provides necessary support for introducing water washing operations into covariate modeling and helps to reveal the dynamic impact of maintenance operations on engine performance. In addition, the time series characteristics of the data are clear, and the change trend of EGTM is significantly correlated with water washing events, which provides a natural experimental scenario for the application of the model in time series analysis and future covariate modeling.

[0028] By conducting experiments at different thrust levels, the robustness of the proposed algorithm can be effectively verified. Engines of different thrust levels have different operating conditions and degradation modes, which makes unified modeling of all thrust level data a challenge. The present invention can test the adaptability of the algorithm in a variety of scenarios by conducting experiments on engine data of three thrust levels respectively. The experimental results can not only verify the prediction accuracy of the algorithm under a single condition, but also further verify its generalization ability and stability under different thrust conditions. This experimental design of multiple thrust scenarios makes the research results more reliable and of practical application value, laying an important foundation for future algorithm promotion and engineering implementation.

[0029] In common time series prediction tasks, the length of the prediction window is usually set within a shorter time range so that the model can focus on capturing short-term change trends. Although this setting has certain advantages in capturing recent changes, it may be insufficient in predicting long-term trends. In the present invention, in order to better evaluate the performance of the model in predicting long-term changes, experiments with a variety of prediction window lengths are specially designed, which are set to 96, 192, 384, 768 and 1536 flight cycles respectively. Among them, the longest prediction window reaches 1536 flight cycles, which significantly exceeds the range of common time series predictions, providing an opportunity to explore the prediction capabilities of the model in more challenging scenarios.

[0030] On the basis of the above experimental design, experiments were conducted on aircraft engines of three thrust levels (22KIbf, 24KIbf, and 26KIbf) to fully verify the robustness of the method of the present invention. Engines of different thrust levels not only have different operating characteristics, but their degradation modes and the impact of water washing on performance may also differ significantly. Therefore, by conducting experiments under multiple thrust conditions, it is possible not only to test the performance of the model under a single thrust condition, but also to verify its adaptability and stability under different operating scenarios. This multi-scenario, multi-prediction window experimental design provides an important reference for evaluating the potential of the model in complex engineering applications.

[0031] Through the analysis of the experimental results in Table 1, we can clearly observe the difference in the effects of different coding forms in the prediction of the remaining life of aircraft engines. Compared with 0-1 coding and no covariate, sequential coding significantly reduces the prediction error and shows a lower MAE (Mean Absolute Error). Especially in short-term predictions (such as 96 flight cycles and 192 flight cycles), the advantages of sequential coding are particularly obvious, and its MAE value is much lower than other coding forms, which verifies the ability of sequential coding to capture the dynamic influence of covariates in short-term predictions. It can also be found from the results that the prediction effect of 0-1 coding is not much different from that of no covariate, and even in some cases (such as 192 flight cycles and 768 flight cycles) it fails to bring significant improvement. This shows that simple 0-1 coding cannot fully utilize the information contained in the washed covariate, and its effect on improving model performance is limited. In addition, as the prediction length increases, the effect of covariates on improving model performance gradually weakens. In longer prediction windows (such as 384 flight cycles and 768 flight cycles), although the MAE of sequential coding is still lower than that of no covariate, its relative advantage begins to shrink. This phenomenon may be related to the gradual attenuation of the effect of covariates on long-term forecasts, and also reflects the limitations of the model in dealing with long-term complex trends. In summary, sequential coding shows significant performance improvement in short-term forecasts, verifying the effectiveness of its introduction of washed covariates.

[0032] Table 1 Comparison of MAE errors under different prediction cycles and encoding forms

[0033]

[0034] The present invention proposes a method for predicting the remaining life of an aircraft engine based on a simplified version of the TFT (Temporal Fusion Transformer) algorithm. By introducing water washing data as future covariates, the accuracy and robustness of the model in long-term prediction are significantly improved. Experimental results show that compared with traditional 0-1 encoding and models without covariates, sequential encoding can effectively reduce prediction errors, especially in short-term predictions. With the increase of the prediction window, the improvement of the model performance by covariates gradually weakens. This phenomenon provides a valuable reference for further improving the long-term prediction ability of the model. The robustness of the proposed method is proved by experimental verification of aircraft engine data of different thrust levels. In the scenarios of three different thrust levels, the simplified TFT model shows strong adaptability and stability, and can better cope with prediction tasks under different working conditions. This study not only provides new ideas for the health monitoring and remaining life prediction of aircraft engines, but also provides a reference framework for the prediction tasks of other complex systems.

Claims

1. A method for predicting the remaining life of a civil aviation engine considering covariates, characterized in that: Future covariates are used as input features, and the impact of covariates on target variables is captured through time-step dynamic modeling. Civil aviation engine water washing data is used as future covariates, and the remaining life prediction model adopts the TFT time series prediction model that removes static covariates and past covariates.

2. A method for predicting the remaining life of a civil aviation engine considering covariates according to claim 1, characterized in that: The civil aviation engine water washing data record is the water washing operation information of a certain engine at a specific time point, including the engine identification, time point and whether the water washing operation is performed.

3. A method for predicting the remaining life of a civil aviation engine considering covariates according to claim 2, characterized in that: In order to use the water washing data as a covariate in the prediction model, the water washing data needs to be encoded. The encoding rule is based on the cumulative number of water washing operations, which is converted into a numerical variable for model processing. If the engine is in a state of overhaul or new engine and no water washing operation has been performed, the encoding is 0; Each time a water washing operation occurs, the code value increases by 1. For example, an engine that has not been water washed after overhaul is coded as 0, one water washing is coded as 1, two water washings are coded as 2, and so on. The coding process includes arranging the water washing records in chronological order, accumulating the occurrence of each water washing operation starting from the overhaul or delivery time point, and updating the corresponding code value for each time point.

4. The method for predicting the remaining life of a civil aviation engine considering covariates according to claim 1, characterized in that: The remaining life prediction model includes a variable selection network, an LSTM Encoder module, an LSTM Decoder module, an Add&Norm Gate module, a masked multi-head attention module, an Add&Norm Gate module, a gated residual network GRN, and a Dense module, wherein the feature data χ is sequentially sent to the masked multi-head attention module after passing through the variable selection network, the LSTM Encoder module, the Add&Norm Gate module, and the gated residual network GRN, and the future covariate γ is sequentially sent to the masked multi-head attention module after passing through the variable selection network, the LSTM Decoder module, the Add&Norm Gate module, and the gated residual network GRN, and then processed by the Add&Norm Gate module, the gated residual network GRN, the Add&Norm Gate module, and the Dense module respectively and then output.

Images