Reordered short-term autocorrelation-driven long-range discriminative convolutional autoencoder for dynamic process monitoring

Abstract

Deep neural networks (DNNs) can result in suboptimal monitoring performance due to nonlinearity, dynamics, and local characteristics in modern complex industrial processes. To surmount these limitations, this paper first proposes a novel data construction method to model the short-term autocorrelation and spatial correlations as a three-dimensional matrix and then reorder the elements of it to better encode the local and temporal structures. Subsequently, we design a new structure called Long-range Discriminative Attention (LDA) based on the self-attention mechanism to enlarge the receptive field of the original convolutional neural networks (CNNs) to extract global features. Finally, we propose a monitoring model named Long-range Discriminative Attention Autoencoder (LDCA) based on LDA to extract structural features between long-range and local variables from the constructed matrix. The effectiveness of the method in fault detection is verified by numerical examples and a three-phase flow process.