Adaptive Moving-Window Dual-Test Granger Causality for Root Cause Diagnosis of Non-Stationary Industrial Processes

The presence of non-stationary features poses a major challenge to root cause diagnosis in industrial processes, as they can distort fault propagation paths inferred through causal testing. To address this issue, an adaptive moving-window dual-test Granger causality framework is proposed for non-stationary industrial processes. First, a dual non-stationary test mechanism, which integrates the Augmented Dickey–Fuller and Kwiatkowski–Phillips–Schmidt–Shin tests, is developed to assess the stationarity of process variables. Next, an adaptive moving-window strategy is designed to adjust window lengths based on the non-stationarity test results. Time series are then segmented according to the selected windows, and a vector error-correction model is fitted to provide a robust basis for causal testing. Subsequently, Granger causality tests are conducted within each window to capture the true causal relationships among variables. Finally, window-wise scores are aggregated to identify the root cause and infer the fault propagation path. The proposed framework is evaluated on the Tennessee Eastman Process, and the results demonstrate that it effectively improves the accuracy of root cause diagnosis.