Distributed Robust Process Monitoring Based on Optimized Denoising Autoencoder With Reinforcement Learning

Abstract

Global monitoring for complex large-scale chemical processes is often challenging because of complex correlations among variables. This article proposes an optimized denoising autoencoder (DAE)-based distributed monitoring method to achieve efficient and robust monitoring of multiunit, nonlinear processes. First, a process is decomposed into multiple units, and then stacked DAE is used to extract the robust features of each unit and represent variable correlations within each unit. Second, deep regression neural networks are established between a local unit and its neighboring units to represent the correlations among units. A reinforcement learning-based neural architecture search method is proposed to avoid the tedious manual tuning process and obtain a high-performance neural network. Finally, a numerical simulation example, the Tennessee–Eastman benchmark process, and a laboratory-scale distillation process are used to verify the effectiveness of the proposed method.