Generic tool-performance-estimation scheme for anomaly detection of equipment

The status of production equipment in industries is prone to change due to various factors. Thus, a proper anomaly detection procedure for equipment is indispensable. Supervised learning is commonly used in equipment state monitoring. However, these approaches need much manpower and time to obtain sufficient labeled data. In contrast to the technique, unsupervised learning-based Autoencoder, which learns from unlabeled datasets to discover patterns without human supervision, has recently demonstrated its promising performance in anomaly detection. Based on this advantage, this paper proposes the tool performance estimation (TPE) scheme to clearly detect abnormalities of any machinery equipment in industries. Through the collaborative interaction among four well-designed modules: AE (Autoencoder), PO (parameter optimization), RE (reconstruction error), and DI (deviation indicator), TPE defines three systematic and generic troubleshooting scenarios: (i) equipment, (ii) parameters, and (iii) others, by analyzing the severity of RE and DI. TPE is proven to immediately remind users to troubleshoot when abnormalities occur during production. In example 1, TPE has the best accuracy among benchmark methods. In example 2, TPE saves 33.9 % of productivity and a 34-day earlier warning to prevent production from shutting down.