摘要
Most existing root cause diagnosis (RCD) algorithms implicitly assume that the sampling rate of process data satisfy the timescale need of capturing complete causal interactions, and the interactions within one sampling interval (instantaneous causalities) are overlooked, which is actually crucial for correct RCD analysis. This paper proposes a disturbance RCD strategy based on error correction instantaneous Granger causality (ECIGC) for nonstationary industrial processes, which can effectively capture instantaneous causal relationships among disturbance variables and obtain correct RCD results. Specifically, to diminish spurious causality induced by nonstationarity, a first-order differenced series was acquired by difference operation and served as the input to the structural equation model (SEM). Then, an error correction term constructed from the original series is incorporated into the SEM to model short-term fluctuations and better characterize long-term dependencies among variables. Subsequently, Renyi transfer entropy (RTE) was introduced into the variational inference to optimize the parameters of the edge probability matrix of the variational distribution, thereby improving convergence speed and estimation efficiency. Finally, by averaging across time steps and thresholding the derived edge probability matrix to obtain a binary adjacency matrix, the causal diagram over the disturbance variables was constructed, and the root cause could be identified. The effectiveness of the proposed method is demonstrated through numerical simulation, the benchmark Tennessee Eastman process, and the Three Phase Flow process, which are representative non-stationary industrial systems. The results show that incorporating instantaneous causal analysis can significantly improve the accuracy and interpretability of root cause diagnosis and disturbance propagation analysis. Furthermore, the proposed method achieves shorter computation time than the comparison methods in all case studies.