밀도기반 군집화와 딥러닝을 활용한공정 주기 신호의 이상 탐지 및 분류

Abstract

Process Fault Detection and Classification (FDC) distinguishes between normal and abnormal process cyclesignals. In the case of process cycle signals, quality control is difficult due to lack of information on patterns anddata imbalance. In this paper, We proposed a method to extract key features of cycle signal data by using LSTMAutoencoder and performed DBSCAN clustering to obtain information on patterns when there is no informationon process cycle signals. We used data augmentation especially when cluster with low density to eliminate thedata imbalance of the process signal. Through the above process, We finally constructed a bidirectional LSTMmodel for real-time process cycle signal classification. This provides a basis for smart factories by suggestingways to actively respond without relying on domain knowledge.