Zhang Xinghui, Xiao Lei, Kang Jianshe
Mechanical Engineering College, Shijiazhuang 050003, China.
The State Key Lab of Mechanical Transmission, Chongqing University, Chongqing 400030, China.
Sensors (Basel). 2015 Mar 23;15(3):6996-7015. doi: 10.3390/s150306996.
Tracking degradation of mechanical components is very critical for effective maintenance decision making. Remaining useful life (RUL) estimation is a widely used form of degradation prediction. RUL prediction methods when enough run-to-failure condition monitoring data can be used have been fully researched, but for some high reliability components, it is very difficult to collect run-to-failure condition monitoring data, i.e., from normal to failure. Only a certain number of condition indicators in certain period can be used to estimate RUL. In addition, some existing prediction methods have problems which block RUL estimation due to poor extrapolability. The predicted value converges to a certain constant or fluctuates in certain range. Moreover, the fluctuant condition features also have bad effects on prediction. In order to solve these dilemmas, this paper proposes a RUL prediction model based on neural network with dynamic windows. This model mainly consists of three steps: window size determination by increasing rate, change point detection and rolling prediction. The proposed method has two dominant strengths. One is that the proposed approach does not need to assume the degradation trajectory is subject to a certain distribution. The other is it can adapt to variation of degradation indicators which greatly benefits RUL prediction. Finally, the performance of the proposed RUL prediction model is validated by real field data and simulation data.
跟踪机械部件的退化对于有效的维护决策至关重要。剩余使用寿命（RUL）估计是一种广泛使用的退化预测形式。当有足够的运行至失效状态监测数据时，RUL预测方法已得到充分研究，但对于一些高可靠性部件，很难收集从正常到失效的运行至失效状态监测数据。只能使用特定时间段内的一定数量的状态指标来估计RUL。此外，一些现有的预测方法存在由于外推性差而阻碍RUL估计的问题。预测值收敛到某个常数或在一定范围内波动。而且，波动的状态特征也对预测有不良影响。为了解决这些难题，本文提出了一种基于动态窗口神经网络的RUL预测模型。该模型主要包括三个步骤：通过增长率确定窗口大小、变化点检测和滚动预测。所提出的方法有两个主要优点。一是所提出的方法不需要假设退化轨迹服从某种分布。另一个是它可以适应退化指标的变化，这对RUL预测非常有利。最后，通过实际现场数据和仿真数据验证了所提出的RUL预测模型的性能。
