摘要
Root cause diagnosis in large-scale industrial processes is challenging due to strong nonlinearity, complex variable coupling, and fault propagation effects. Classical Granger causality relies on linear autoregressive models and often fails to identify causal relationships in nonlinear dynamic systems. To address this issue, this paper proposes an echo state network (ESN)-based nonlinear Granger causality framework. By integrating ESN into the Granger causality framework, nonlinear temporal dependencies among process variables can be effectively captured with low computational cost. To ensure statistical reliability, the Wilcoxon signed-rank test is adopted to replace the conventional F-test. In addition, a block-diagonal reservoir structure is introduced to enhance interpretability and reduce spurious causality caused by strong variable coupling. Based on the inferred causal relationships, a direct causal graph is constructed to identify the fault root variable and propagation paths. The numerical simulation and the Tennessee Eastman Process case study demonstrate that the proposed ESN-based Granger method accurately identifies nonlinear causalities and root causes, outperforming the conventional Granger causality method. © 2026 IEEE.