Anomaly Ranking of Failure Causes in Manufacturing Process Using Causal Model

Identifying causes of unknown failures during the manufacturing process is quite difficult because of the uncertainty nature of the failures. As the manufacturing process becomes more complicated, failure-cause identification becomes more time consuming due to the increased number of causal candidates to be checked when addressing anomalies, and the complicatedly intertwined factors are difficult to discern. To solve these problems, various methods of anomaly detection and ranking have been developed using sensor data gathered during the manufacturing process. However, once an anomaly occurs at any point in the process, it propagates to other steps. Conventional anomaly detection methods are not sufficient against the complicated and unpredictable anomalies. In this study, we developed a novel anomaly ranking method based on a causal model. It employs a fusion model of human knowledge and sensor data to acquire a suitable causal model for accurate ranking. Experiments were conducted on a real packaging machine, the results of which showed a roughly 30% reduction of the ranking average compared with that of the conventional methods.