A Data-Driven Causality Analysis Tool for Fault Diagnosis in Industrial Processes

Causal Discovery in Manufacturing: A Structured Literature Review

“…In the literature, online and offline fault diagnosis types are discussed in terms of the temporal perspective and process operator engagement. In online applications, the point of being knowledgeable of the causal structure is to reduce the reaction time by (1) providing narrowed and focused information about the causes or (2) providing the reasoning for the underlying system conditions [52]. The former finds application in providing causes for rapidly degrading system components [53] or sub-systems where a fault has occurred [54], among others.…”

“…Rodrigo et al (2016) employed transfer entropy in the alarm log, process, and connectivity data in one of the steps of their general approach for alarm flood reduction [50]. Furthermore, in [51], the authors based their data similarity analysis approach on Granger causality, which was also used in [52]. Bayesian networks saw application in reliability management [53] and were one of the methods evaluated by Yadav et al (2017) alongside association rule mining, direct estimation, and propensity score-matching methods for the analysis of rare events [55].…”

“…Rather than using smaller use case examples and numerical studies, most of the literature focusing on fault detection, analysis, and management featured case studies that focused on a deep understanding of process behavior or the development of causality-based or causality-enhanced tools to improve process operation. Out of the reviewed literature, only two articles used industrial benchmark datasets in their evaluation [47,52]. In this regard, it can be concluded that in the transition from model-driven methods to data-driven methods, the domain experts in manufacturing are the ones primarily utilizing causal discovery methods.…”

“…Some of these disadvantages were noted in the literature: inability to deduce the cause-and-effect relationship from the correlation [47,49,55], poor performance in imbalanced data scenarios [49,55], or relying too much on the domain input [49,51]. On the other hand, some authors [47,52] focused their efforts on the intrinsic challenges of industrial data, such as nonstationarity and multicollinearity. Overall, from the analyzed literature in this application area, it is evident that causal discovery applications are scattered, with no systematic application or integration into continuous process improvement.…”

“…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].…”