A semi-supervised feature contrast convolutional neural network for processes fault diagnosis

Abstract

BackgroundModern industrial processes involve multiple operating units, which perform their respective functions and are coupled with each other. Accurate extraction of complex non-linear relationship contained in process variables is the key to fault diagnosis. Most fault diagnosis methods based on deep learning heavily rely on labeled date, yet labelled samples are limited in the real industrial process. MethodsThis paper proposes a semi-supervised feature contrast convolutional neural network (SS-FCCNN) model for multi-unit, non-linear, label-deficient industrial process. Firstly, based on the physical structure of industrial process, local unit convolutional neural network (LU-CNN) and global unit convolutional neural network (GU-CNN) are proposed. LU-CNN extracts the deep features of the variables involved in each operating unit as the local intra-unit features. GU-CNN extracts global inter-unit features and models the complex relationship between operating units due to fault propagation. Subsequently, the unit feature contrast network (UFCN) and data corrosion strategy are applied to labeled data and unlabeled data respectively, which restricts the unit features and enhances the generalization of unit features extracted by the network. Case studies of Tennessee Eastman process demonstrate the effectiveness of SS-FCCNN.