涡扇发动机剩余使用寿命估计：一种基于深度学习的滑动时间窗方法

摘要

系统退化是航空航天领域中普遍且不可避免的过程。为避免复杂工业系统发生非预期停机，人们采用预测与健康管理技术，借助大量传感器评估系统的健康状态及其剩余使用寿命（RUL）。已有众多研究者利用深度学习方法，依据传感器数据估计 RUL，其中多数工作只用单一深度神经网络（DNN）模型求解该问题。本文基于多个 DNN 模型构建了一种新的涡扇发动机 RUL 预测器。该方法采用时间窗技术准备样本，增强了 DNN 提取特征与学习涡扇发动机退化规律的能力。此外，本文用公认的模型评价指标验证了所提方法的有效性。实验结果表明，在四种不同的 DNN 中，基于长短期记忆网络（LSTM）的预测器在独立测试集上取得了更好的成绩：均方根误差为 15.30，平均绝对误差为 2.03，决定系数为 0.4354，优于此前报道的涡扇发动机 RUL 估计方法。

Abstract

System degradation is a common and unavoidable process that frequently occurs in aerospace sector. Thus, prognostics is employed to avoid unforeseen breakdowns in intricate industrial systems. In prognostics, the system health status, and its remaining useful life (RUL) are evaluated using numerous sensors. Numerous researchers have utilized deep-learning techniques to estimate RUL based on sensor data. Most of the studies proposed solving this problem with a single deep neural network (DNN) model. This paper developed a novel turbofan engine RUL predictor based on several DNN models. The method includes a time window technique for sample preparation, enhancing DNN's ability to extract features and learn the pattern of turbofan engine degradation. Furthermore, the effectiveness of the proposed approach was confirmed using well-known model evaluation metrics. The experimental results demonstrated that among four different DNNs, the long short-term memory (LSTM)-based predictor achieved the better scores on an independent testing dataset with a root-mean-square error of 15.30, mean absolute error score of 2.03, and R-squared score of 0.4354, which outperformed the previously reported results of turbofan RUL estimation methods.