다중 식별자를 이용한 Adversarial Autoencoder 기반 제조 공정 이상 탐지

Abstract

When unexpected problems occur in manufacturing process, it is necessary to configure an anomaly detection system to monitor and control them. Abnormal data are critical because they cause a decrease in yield and poor quality. If abnormal data is not detected, the process continues and the loss becomes greater. Abnormal data have fewer numbers than normal data, resulting in class imbalance problems. Therefore, we solve the data imbalance problem by learning distribution of normal data only. Unlike conventional methods, adversarial autoencoder (AAE) is able to create distributions similar to the original data through competitive learning using discriminator. This paper proposes adversarial autoencoder with multiple discriminators, a method to learn the distribution of normal data more accurately by adding two discriminators to AAE. We use Long Short-Term Memory (LSTM) layer to fit the time series characteristics. Experiments confirm that the method proposed in this paper show great anomaly detection performance.