Deep residual temporal independent variable analysis for nonlinear dynamic process monitoring

Abstract

Modern industrial processes exhibit strong dynamic and nonlinear characteristics, which pose significant challenges to data‐driven process monitoring. Existing lightweight deep models (LDMs) alleviate the opacity of deep neural networks but inherit a dimension dilemma from their underlying single‐layer statistical decompositions. Moreover, the time‐lag shift method commonly used to incorporate temporal information further aggravates dimensional redundancy and multicollinearity. To address these issues, this paper proposes a deep residual temporal independent variable analysis (DR‐TIVA) network for nonlinear dynamic process monitoring. A temporal independent variable analysis (TIVA) module is developed, which transposes the joint‐space idea of independent variable analysis from process–KPI modelling to real‐time–history modelling and explicitly isolates temporal common, real‐time‐specific and history‐specific independent components. A deep residual stacking architecture with asymmetric residual treatment is then designed, in which a nonlinear activation is applied only to the real‐time residual to enrich the representation capacity of subsequent layers while the historical residual is preserved in its original space, so that the layer‐wise predictions remain directly accumulable in the original output space and the deep model retains the statistical interpretability of latent‐variable modelling. The proposed method is validated on a nonlinear dynamic numerical example and the multiphase flow facility (MFF) process. The experimental results analyze the impact of different activation functions on experimental performance, while also verifying that the proposed method has a significant advantage in detecting dynamic relationships within nonlinear data and is more sensitive to slugging faults in the MFF process.