Enhancing the interpretability of anomaly detection methods for the prediction of foreign object contamination

抄録

Filters are often employed in fluid transport lines that produce beverages, serving as a countermeasure against foreign objects. However, when these filters become clogged during removal, this can lead to an enlargement of the openings or damage to the mesh, leading to secondary generation of foreign objects. Thus, a system that continuously monitors to prevent accidents before they occur is necessary. However, it can be difficult to learn due to the absence of data from actual abnormal situations, and various elements intertwine on the production floor. Therefore, it is necessary to verify which methods are optimal for responding to these situations, and analyze elements for problem identification and improvement. In this study, we leverage the pressure loss caused by filter clogs in the fluid conveyance lines that produce beverages to validate various anomaly detection methods through real simulations. We test the appropriateness of each method based on the nature of the problem and the characteristics of the data, thereby enhancing the interpretability of the model. We compare the detection capabilities of different anomaly detection methods and perform an analysis of feature importance to clarify the importance of each element. Since the results of anomaly detection methods can vary based on the nature of the problem and the characteristics of the data, making interpretability clear can make the results of the anomaly detection methods easier to understand and improve prediction accuracy. By understanding the causes of abnormalities and the contribution of characteristics, it is possible to suppress the occurrence of foreign objects.