A Multi-Channel Multi-Head CNN Framework for Fault Classification in Industrial Process

This paper proposes a novel fault classification method via convolutional neural network with multi-channel and multi-head along the time dimension, which is defined as MM-CNN. The MM-CNN extracts features of industrial process data by convolutional layers from local to global level. Unlike traditional methods, this method can capture the independent features of every process variable and the relevant characteristics from sensor data to learn more useful latent fault patterns. Besides, a data preprocessing approach is proposed to transform original data for convolutional neural network. Finally, for all the all 21 faults in Tennessee Eastman(TE) process, this paper compares the proposed method with the published methods and a architecture which combines the convolutions of multi-head and a bidirectional Long Short-Term Memory (Bi-LSTM) layer in fault classification. The simulation results show that the method has better classification performance than the state-of-the-art methods.