Time Series Anomaly Detection Using Contrastive Learning based One-Class Classification
Yeseul Lee, Yunseon Byun, Jun‐Geol Baek · 2023
Time series anomaly detection in industrial processes has recently attracted attention. However, since there are no labels in the manufacturing process data collected in real time, there are limitations in using supervised learning-based classification models. Therefore, the proposed method newly defines an objective function that simultaneously learns the OCC model and contrastive learning. In addition, by applying data augmentation techniques suitable for periodic time series data in contrastive learning, feature extraction that preserves temporal characteristics is possible. The effectiveness of representation extraction was verified by showing high anomaly detection performance even in datasets with similar normal and anomaly data forms.