Temporal Convolution-Based Long-Short Term Memory Network With Attention Mechanism for Remaining Useful Life Prediction

Abstract

Predictive maintenance (PdM) is useful for engineers to schedule maintenance flexibly, to operate equipment efficiently, and also to avoid unexpected downtime. Remaining useful life (RUL) prediction is critical for PdM before the need for component replacement. Data-driven approaches have attracted more attention for RUL prediction in flexible production. Compared with statistical-based and conventional machine learning approaches, deep learning-based approaches can extract critical features from raw equipment sensor data without prior knowledge from a domain expert. This study proposes a temporal convolution-based long-short term memory (TCLSTM) network with attention mechanism to extract features from equipment sensor data and to build a regression model for RUL prediction. Temporal convolutional networks (TCN) are used for feature extraction. LSTM and attention layers are used to learn temporal dependencies among the extracted features. The fully connected network consists of dense layers and is used to build the RUL prediction model. To evaluate the effectiveness of the proposed method, an empirical dataset from a semiconductor ion mill etching process was used. According to the comparison results, the proposed TCLSTM network with attention mechanism has the lowest prediction error and outperforms TCN, LSTM, and other machine learning approaches. The experimental results demonstrate the practical viability of the proposed approach.