A COMPARATIVE STUDY OF ANOMALIB AND RULE-BASED MODELS FOR DETECTING CHAIN JUMPING IN CAR ENGINE MANUFACTURING

Abstract

In the evolving landscape of Industry 4.0, the automotive manufacturing sector is increasingly relying on artificial intelligence (AI) to automate processes previously handled manually. This study addresses the need for automated detection of chain jumping in car engine production, a task traditionally performed by humans. The goal is to identify the best model for anomaly detection to enhance efficiency and reduce human intervention. We leverage Anomalib, a comprehensive library for anomaly detection, and evaluate four of its prominent models: DRÆM, Efficient AD, Padim, and Patchcore. Additionally, we compare these AI models against a traditional rule-based approach that uses color detection at specific positions through conventional image recognition techniques. The models were assessed based on two key evaluation metrics: recall and accuracy. The training dataset comprised 400 samples, split evenly between OK and NG (Not Good) conditions. Validation was conducted using a separate dataset, and the results were averaged for final comparison. Our findings reveal that the Patchcore model outperforms others, achieving a recall of 99.44% and an accuracy of 97.64%. Interestingly, while other AI models did not surpass the rule-based approach, the rule-based method achieved a recall of 98.75% and an accuracy of 96.31%, making it the second-best performer. This suggests that while AI offers significant advantages, it is not always superior to traditional methods. Nevertheless, Patchcore is selected as the optimal solution for detecting chain jumping due to its superior performance in both accuracy and recall.