Off-Line Analysis of Dynamic Causal Dependencies in Evolving Industrial Alarm Floods

Off-line alarm flood root-cause analysis methods are frequently used as a primary step for online operator support and have been promoted in several research activities in recent years. Addressing causal relationships in nonlinear and nonstationary time series, however, remains a challenging task. Moreover, evolving alarm floods that can be characteristic of an escalating abnormal situation have been disregarded to date. To address and solve these limitations, this paper presents a novel hybrid off-line alarm flood root-cause analysis method that uses a data-driven nonlinear causality estimator based on "nonparametric multiplicative regression," a novel and efficient sliding window technique, and additional process information. The method proposed here allows for the analysis of the most critical alarms in a historical alarm flood by analyzing causally dominant and highly affected process variables based on their dynamic causal dependencies. The performance of our method is assessed using an openly accessible dataset based on the "Tennessee-Eastman-Process." It is demonstrated that propagating abnormal situations in industrial processes can demonstrate complex time-varying causalities, where the effect of different dominant process variables can significantly vary over time.