Dual-TCN-GRU: A Dual Path Temporal Convolutional Network with GRU Fusion for Aero-Engine Fault Detection

Abstract

Conventional aero-engine fault detection techniques tend to have problems simultaneously extracting local anomalies in sensor data and long-term temporal dependencies. To solve this problem, we propose a new fault detection scheme that only uses a Dual-Path Temporal Convolutional Network (Dual-TCN) and a Gated Recurrent Unit (GRU) module. The model, in turn, takes advantage of dual parallel branches of TCNs to extract local and global features and integrates these features with the GRU to model the progression of faults in time. Validated on the dataset of the National Aeronautics and Space Administration, called C-MAPSS, the proposed technique achieves a detection accuracy of 91.39%, which is better than CNN and LSTM baselines, demonstrating interesting improvements in the precision, recall, and F1-score. Experimental results further demonstrate the effectiveness of the dual-path feature extraction and GRU fusion strategy; this method is potentially useful to realize the real-time and accurate detection of faults in complex aero-engine systems.