A flow-information density clustering method for detecting step faults with control-compensation mechanisms

In modern closed-loop industrial control systems, controllers generate compensatory actions immediately after a step disturbance to maintain steady-state operation.This compensation temporarily masks the observable symptoms of a fault and alters the underlying causal dependencies among process variables, making traditional deviation-based detection methods insufficient for timely identification.To address this issue, this paper proposes an unsupervised step-fault identification method based on information-flow density clustering.Liang-Kleeman information flow (LKIF) is first employed to quantify the strength and direction of causal interactions, forming causal features that reflect structural variations induced by faults.A flow-difference statistic is then used to construct an information-flow density score (IDS), which captures local abrupt changes along causal chains.Finally, DBSCAN is applied to the IDS distribution to achieve anomaly detection without requiring labeled samples.The proposed method is validated on the Tennessee Eastman Process (TEP).Experimental results show that it significantly outperforms PCA, KPCA, and recurrence-plot-based feature extraction and clustering methods in terms of ARI and V-measure, offering a more interpretable and effective approach for dynamic fault identification in closed-loop systems.