Sparse Causality Analysis Approach with Time-varying Parameters for Root Cause Localization of Nonstationary Process

Root cause diagnosis (RCD) is an important technique for maintaining the safe operation of industrial processes. Traditional RCD methods usually require stationarity assumptions. However, the process inevitably shows nonstationarity due to factors such as switching of operating conditions. Although there have been some previous studies trying to overcome the challenge of nonstationarity, these methods fail to guarantee the significance of the extracted causalities and lead to redundant relationships. To address the above issues, a sparse causal analysis model with time-varying parameters is extracted in this study. First, we propose an end-to-end information fusion and prediction task to characterize predictive relationships between variables and avoid repeated modeling. Second, we design time-varying parameters for the information fusion mechanism to cope with nonstationarity and automatically identify significant causality through sparse parameter updates. We design an update strategy that constrains the gradient information to guarantee sparsity. Finally, a causal metric is constructed for the time-varying predictive relationship to comprehensively obtain the overall causal relationship, which further guarantees causal significance. The validity of the proposed method is illustrated through a real industrial example collected from a thermal power plant.