Multi-Head CNN-LSTM with Prediction Error Analysis for Remaining Useful Life Prediction

Abstract

Predicting accurate remaining useful life (RUL) of components plays a crucial role in making optimal decision for maintenance management. As sensor technology developed, multiple heterogeneous sensors are being used to collect information for monitoring the condition of components. Deep learning architectures, such as convolutional neural network (CNN) and long short term memory (LSTM), can be considered as a successful end-to-end framework to predict RUL from the multivariate time series collected by those sensors. For that, we employ an architecture combining the parallel branch of CNN in series with LSTM which is referred as multi-head CNN-LSTM. Furthermore, we propose a combination of the network with time series prediction error analysis (PEA). The prediction errors at entire time cycles of each time series are estimated by recursive least squares (RLS) and single exponential smoothing (SES) respectively. We analyze each of the two sequences of prediction errors with exponentially weighted moving average (EWMA) and combine them with Fisher's method. Finally, the output of the PEA is fed into the CNN-LSTM network as the additional input. We evaluate the performance of our methods on widely used C-MAPSS dataset. The experimental results suggest that using the PEA improves the performance of the deep learning-based RUL prediction model. Compared to other methods in recent literature, the proposed method achieves the state-of-the-art result on one sub-dataset and very competitive results on the others. In addition, the proposed method shows promising results in the consecutive RUL prediction following the degradation process of components.

Keywords

• predictive maintenance
• remaining useful life estimation
• deep learning
• time series analysis