Fault propagation analysis by combining data-driven causal analysis and plant connectivity

Causal Discovery in Manufacturing: A Structured Literature Review

“…Application: Due to their ability to detect causal direction, causal discovery methods are especially useful in manufacturing applications of root cause analysis. Several approaches have focused on combining process topology and connectivity information to improve accuracy and reduce the computational load of root cause analysis in industrial board and board machine case studies [59][60][61][62][63]. Other studies demonstrated the application of causal discovery methods in the detection of disturbance propagation paths in fluid catalytic cracking units [64], mineral concentrator plants [65], and semiconductor production facilities [66].…”

“…Methods and concepts: From the perspective of methods and approaches, the literature on the application of causal discovery methods in the context of root cause analysis for process improvement focused on established methods such as transfer entropy [59,60,65,[70][71][72], Granger causality [13,32,61,62,67], or Bayesian networks [34]. Aside from the application of existing methods, several studies have focused on the improvement of current approaches for data-driven root cause analysis.…”

“…Algorithmic improvements are focused on applications of principal component analysis [64], extension to multivariate scenarios [32], or improvements in the causality analysis models [33,69,71]. On the contrary, the latter approaches focus on the integration of Granger causality and transfer entropy approaches with process schematic data in adjacency (connectivity) matrix form [59][60][61][62][63]. Modern approaches, such as deep learning combined with graph processing techniques, have been applied in the context of root cause analysis in [66].…”

“…This indicates that causal discovery methods found usage in root cause analysis, yet current efforts are not focusing on the evaluation of these methods in production scenarios but are addressing the recognized challenges of existing approaches. Some of the existing challenges found in the literature are data stationarity [32,[59][60][61][62]69], distribution assumptions [34,71], small datasets [71], nonlinearity [33,69], computational time [59,60,64], scaling the analysis process [60], or inclusion of prior knowledge [59,67]. The lack of application of root cause analysis in real production scenarios can be associated with the mentioned challenges along with the overall complexity in the application of these methods.…”

“…However, these approaches often come with a set of assumptions that do not match the reallife complexity in the industry. Some of the common assumptions found in the literature are data stationarity [32,52,[59][60][61][62]69], normal distribution [34,71], linear relationships [33,69], causal sufficiency [30], and faithfulness [9].…”