An Improved LSTM Neural Network with Uncertainty to Predict Remaining Useful Life

Abstract

Data-driven Prognostic(DDP) has become one of the major method of component of prognostic and healthy management(PHM) systems in the industrial area. The fault prediction methods mainly include fault failure probability assessment and remaining useful life(RUL) prediction. As the basis for the development of equipment maintenance strategy, the remaining service life prediction is one of the important links of PHM. Accurately predicting the RUL can provide comprehensive, accurate and effective information for the development of equipment maintenance strategies, which helps to avoid equipment failure and reduce the loss caused by failure, thus ensuring the safe and reliable operation of the equipment. In recent years, the RUL prediction has received extensive attention in research and engineering fields and achieved certain results. Among them, the method based on degraded data modeling has become one of the mainstream methods in the field of life prediction because it does not require failure data and the convenience of characterizing the uncertainty of degradation. DDP about RUL method based on degradation data can be classified into the machine learning method and the mathematical statistics method. Prognostic techniques are designed to accurately estimate the RUL of subsystems or components using sensor data. However, mathematical statistics methods of estimating RUL use sensor data to make assumptions as to how the system degrades or fades (eg, exponential decay); As well as the current some machine learning methods ignore the uncertainty. Based on current problems, we propose a novel Long-Short Term Memory(LSTM) Neural Network complement with Uncertainty: automatically learn higher-level abstract representations from the underlying raw sensor data, and use these representations to estimate RUL from the sensor data; it does not rely on any degradation trend assumption, is robust to noise, and can handle missing values and uncertainty in sensor data. We compared several publicly available algorithms on a publicly available Turbofan engine dataset and found that several of the proposed metrics (Score, etc.) outperformed the previously proposed state-of-art techniques.