A transfer learning approach using improved copula subspace division for multi‐mode fault detection

Abstract

Multi‐mode characteristics of industrial processes are prominent in the area of chemical production due to a diversified market demand. Despite mounting interests in predictive modelling for the optimization of operating conditions in chemical production processes, particularly in the petrochemical industry with multiple feeds and a range of cracking furnaces, targeted solutions that could hold wider applicability are typically hindered by the lack of available data. To overcome the limitation posed by data scarcity, an inductive approach based on transfer learning for fault detection is proposed utilizing copula subspace division (CSD), named TrAdaBoost CSD (TCSD). The proposed TCSD method is based on the probability view to transfer different most similar source samples to target samples. To select the optimal number of source samples, two adaptive indices were proposed and designed to adaptively assign the optimal model training iterations and sample number per iteration. Imbalance in data samples between the target and source datasets was addressed via the adaptive active vine copula‐based probabilistic method. The effectiveness and superiority of the proposed TCSD approaches are validated via a numerical example (with the ranges of normalized fault detection and false alarm rates [FDR and FAR] between 0.82–0.94 and 0.01–0.04, respectively), the Tennessee Eastman process (~10% and 2.24% improvement for FDR and FAR, respectively), and the ethylene cracking furnace process (~15% and 8% improvement for FDR and FAR, respectively).

Joint alignment network preserving structural information for multimode process fault diagnosis

“…The two domains are aligned to draw closer, so that the classifier in the source domain can also successfully perform in the target domain. In general, these methods can be divided into four categories, [ 4 ] including instance‐based TL, [ 5,6 ] parameter‐based TL, [ 7,8 ] feature‐based TL, [ 9–11 ] and relational‐based TL. [ 12 ] The more mainstream method is feature‐level TL.…”