기계 건강 지표 구축을 위한 재구성 기반 이상 탐지

Abstract

In the manufacturing process, maintenance is performed at a specific time point based on continuous monitoring of equipment and processes. However, accurate predictive maintenance of time-series data is difficult. This is because, due to the characteristics of the process equipment, a single equipment experiences various working conditions. It shows various outputs even under similar conditions. Therefore, we proposed a new reconstruction-based anomaly detection. Our method uses a property in which a reconstruction error is calculated through input values and reconstruction values. This builds a sophisticated health indicator (HI) by deferring model selection until the smallest reconstruction value is obtained when training the model. As a result, through advanced HI construction, it is possible to accurately identify and quantify the degree of degradation of machinery. Experiments confirm that the proposed method showed superior performance in terms of initial anomaly detection compared to other models.