Anomaly detection of overhead hoist transport system considering multiple operational states with conditional recurrent autoencoder

Abstract

In contemporary semiconductor fabrication facilities (FABs), the overhead hoist transport (OHT) system is primarily used for automated transportation. This system consists of a track system and OHT vehicles where the vehicles travel on the track to transfer wafers between equipment. To ensure reliable transportation, it is necessary to check the abnormalities of the OHT system continuously. Therefore, we propose condition-based maintenance of the OHT system where the relevant sensor data are collected to determine whether the current operation is abnormal. Meanwhile, since the OHT system has various operational states as a dynamic system, not only sensor data but also this state context should be considered for anomaly detection. In this respect, we propose a conditional recurrent autoencoder (CRAE) for anomaly detection of the OHT system. Recurrent neural network (RNN) structure is employed to effectively process temporal sensor data and conditional input structure is employed to consider the context of the state. The proposed model is verified with data collected in a laboratory imitating the actual factories. Consequently, CRAE showed promising performance for detecting various abnormalities including actual sensor anomalies and state-dependent anomalies.