Dynamic causal modeling for nonstationary industrial process performance degradation analysis and fault prognosis

Abstract

Performance degradation analysis and fault prognosis of industrial processes refer to distinguish different degradation stage and raise warnings before a failure occurs, which are crucial to ensure the operation efficiency and safety. However, processes usually have nonstationarity caused by operating condition switching, which may be confused with the underlying temporal variations in the degradation process. Considering that the degradation process is essentially the changes in short-term temporal information transmission between variables, we propose a dynamic causal graph-driven performance degradation analysis and fault prognosis framework (DCG-PDA), which can reveal the intrinsic connection between causality and degradation mechanism. First, a dynamic causal graph framework is constructed, in which the nonstationarity is removed by transfer entropy-guided symbolization, to extract significant short-term causal information transmission. Second, changes in causality are effectively captured and quantified, and the causal graphs corresponding to some time points with larger changes in causality are selected to characterize the representative causal patterns in the degradation process. Finally, a node anomaly score is designed to consider the accumulation of causal structural changes at anomaly time points and measure variable-wise degradation contributions, so as to realize accurate locations for key variables in the degradation process. The validity of the proposed method is illustrated through a real compressor fouling process.