Granger causality analysis using error correction model for root cause diagnosis in non-stationary industrial processes

Background

Non-stationary characteristics commonly arise in multivariate time series after faults occur. However, existing Granger-based root cause diagnosis (RCD) methods struggle to address the challenges posed by such characteristics.

Methods

To overcome this limitation, a novel Granger causality-based method integrating an error correction model derived from cointegration analysis has been developed. The approach begins with Johansen cointegration analysis on the non-stationary multivariate time series to verify whether there is a cointegration relationship among them. To avoid spurious regression, each variable in the full and reduced Granger models is differenced according to its integration order. Because differencing can obscure long-run relationships, we recapture them with cointegration analysis and add an error-correction term that measures departures from equilibrium in the previous period. We then test the resulting prediction residuals for Granger causality significance, yielding a reliable causal diagram of the fault propagation.

Significant findings

The proposed method’s effectiveness is demonstrated through a numerical simulation, the benchmark Tennessee Eastman process, and a real-world case involving a coal conveyor motor fault. These examples illustrate its robustness and applicability in diagnosing faults in complex industrial processes.