No abstract

With the development of the Full Authority Digital Engine Controller (FADEC) technology, the aero-engine on-board model is widely used in Engine Health Management (EHM) and control. Due to the FADEC’s limited computational capability and storage capacity, the model should not be very intricate; consequently, the interpolation model is widely utilized. Although the interpolation model’s low precision precludes further development of on-board models for EHM and control. To address the trade-off between precision and complexity, a novel on-board modeling method is proposed based on the Nonlinear Autoregressive with Exogenous Inputs Backpropagation neural network (NARX-BPNN) trained using the mini-batch Levenberg–Marquardt (LM) algorithm on large Quick Access Recorder (QAR) data. The NARX model’s features and time delay are chosen by referring to the line interpolation model, which gives interpretability for feature selection. The combination of a shallow neural network and big data training can guarantee the on-board model’s real-time and storage requirements, as well as its generalizability. The mini-batch LM method can avoid both the local optimum problem in the shallow neural network and the storage difficulty associated with massive data while still achieving a rapid convergence rate due to the LM algorithm’s global view. The NARX-BPNN models are compared to an existing line interpolation model using 100 different aero-engines' QAR data. The results reveal that accuracy may be increased by approximately 30% while maintaining superior dynamic performance and anti-noise capacity compared to the line interpolation approach.

In modern industrial systems, condition-based maintenance (CBM) has been wildly adopted as an efficient maintenance strategy. Prognostics, as a key enabler of CBM, involves the kernel task of estimating the remaining useful life (RUL) for engineered systems. Much research in recent years has focused on developing new machine learning (ML) based approaches for RUL estimation. A variety of ML algorithms have been employed in these approaches. However, there was no research on applying deep reinforcement learning (DRL) to RUL estimation. To fill this research gap, a novel DRL based prognostic approach is proposed for RUL estimation in this paper. In the proposed approach, the conventional RUL estimation task is first formulated into a Markov decision process (MDP) model. Then an advanced DRL algorithm is employed to learn the optimal RUL estimation policy from this MDP environment. The effectiveness and superiority of the proposed approach are demonstrated through a case study on turbofan engines in C-MAPSS dataset. Compared to other approaches, the proposed approach obtains superior performance on all four sub-datasets of C-MAPSS dataset. What is more, on the most complicated subdatasets FD002 and FD004, the RMSE metric is improved by 14.4% and 7.81%, and the score metric is improved by 3.7% and 48.79%, respectively.

Remaining Useful Life (RUL) estimation is a crucial technology in prognostic and health management (PHM) for modern aero-engines, as it ensures the reliability and safety of aircraft. With advances in sensor technology, data-driven approaches for RUL estimation have gained significant interest in recent years, especially deep learning-based methods. To further contribute to the field and improve the accuracy of RUL estimation, this paper, proposes novel Transformer-based fusion models for aero-engine RUL estimation. The vanilla Transformer is adapted for RUL estimation by modifying its structure based on the characteristics of aero-engine sensor data. The modified Transformer is then fused with Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) to extract degradation features from multiple aspects. Specifically, the LSTM and CNN layers are incorporated into the decoder and encoder of the Transformer. The effectiveness and superiority of the proposed models are demonstrated through experiments on the C-MAPSS benchmark dataset. The experimental results show that the proposed LSTM-Transformer fusion model outperforms the existing state-of-the-art approaches, with up to 66.53% and 84.86% improvement in RMSE and score metrics, respectively.INDEX TERMS Aero-engine, prognostic and health management, remaining useful life estimation, transformer, long short-term memory neural network, convolution neural network.

Estimation of the aero-engine remaining useful life (RUL) is a significant part of prognostics and health management (PHM) and the basis of condition-based maintenance (CBM) which can improve the reliability and economy. Multiple operating conditions, nonlinear degradation, and early prediction are significant and distinctive issues compared with other prognostics problems. While these issues do not get enough attention and researches in aero-engine RUL estimation. In view of these points, three specific data preparation approaches and a novel loss function are introduced. The data preparation approaches can extract high-quality data for the long short-term memory (LSTM) neural network according to the characteristic of aero-engine degradation data. Among these approaches, operating condition normalization is an effective method to handle the multiple operating conditions problems, and RUL limitation identification is a novel method to identify the turning point of the nonlinear degradation process. The scoring function is an innovative loss function used to replace the mean square error (MSE) loss function which has a preference for early prediction. The comparisons with the original LSTM and some other approaches indicate that the combination of the data preparations and the scoring loss function is an effective solution for the above issues, and can achieve the best performance among the approaches.