AME-TCN: Attention Mechanism Enhanced Temporal Convolutional Network for Fault Diagnosis in Industrial Processes

Abstract

As an indispensable part of process monitoring, the fault diagnosis has become a hot topic in both research and industry. Due to the large-scale monitoring data collected in industrial processes, data-driven methods based on deep learning have been widely used in fault diagnosis. Among these methods, Temporal Convolutional Networks (TCN), which has parallel architectures and larger receptive fields, does not suffer from gradient problems and has shown better performance than Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) in fault diagnosis. However, the generic TCN architecture pays equal attention to different monitoring variables and might degrade the fault diagnosis accuracy when some important parts of input data need to be emphasized. Hence, we propose a novel TCN-based fault diagnosis framework, called Attention Mechanism Enhanced Temporal Convolutional Network (AME-TCN). Attention mechanism is good at distinguishing the importance of different monitoring variables and could enhance the performance of TCN for fault diagnosis by weighting each variable to highlight more diagnosis-related parts. For performance validation, AME-TCN model was applied for Tennessee Eastman (TE) process. Experimental results indicated that AME-TCN method not only outperformed than traditional CNN and RNN models, but also enhanced the fault diagnosis ability of TCN.