Abstract Fault diagnosis in chemical processes remains a challenging task due to strong variable coupling, complex nonlinear dynamics, and distributed fault‐related variations across the processes. To address these issues, this paper proposes an attention‐enhanced graph isomorphism network (AGIN) for chemical process fault diagnosis. Compared with conventional graph neural networks (GNNs), the proposed method alleviates the issues of over‐smoothed neighbourhood aggregation, enables adaptive interaction weighting, and enhances nonlinear representation capability. Specifically, a sum‐based graph isomorphism aggregation branch is introduced to preserve the structural diversity of neighbourhood feature distributions and improve sensitivity to subtle fault patterns. In parallel, a multi‐head graph attention branch is employed to adaptively capture heterogeneous interaction strengths among neighbouring samples. The outputs of the two branches are fused via residual addition and further refined by a channel attention mechanism, while multilayer perceptron‐based nonlinear transformation is adopted to enhance representation capability for complex process dynamics. Moreover, instead of introducing an additional sequence encoder, temporal dependencies are implicitly captured through message passing on temporally constructed neighbourhood graphs. Experimental results on two representative chemical process benchmarks, namely the continuous stirred tank reactor (CSTR) system and the three‐phase flow separation (TFF) process, demonstrate that AGIN consistently achieves superior diagnostic performance and learns more discriminative fault representations than competing methods.
Chen et al. (Sun,) studied this question.