Cyclic Pattern-Based Anomaly Detection in Smart Manufacturing Systems using Contrastive Learning
Wonhwa Choi, Jun‐Geol Baek · 2025
In the manufacturing industry, the environment for collecting sensor data has expanded through Industry 4.0, but labeling is difficult, and there is relatively little faulty data, making it challenging to apply conventional supervised learning-based anomaly detection methods. Specifically, anomalies that occur in sensor data with cyclic patterns are difficult to detect with traditional methods. This study proposes an unsupervised anomaly detection approach using contrastive learning to address these challenges. The method segments sensor data based on cyclic patterns, which are identified through autocorrelation coefficients. Fast Fourier transform (FFT) is then applied to focus on low-frequency bands where anomalies frequently occur. The model learns normal and abnormal behaviors by comparing similar and dissimilar data segments without relying on labeled data. Experimental results show that cyclic pattern-based anomaly detection is effective for anomaly detection in manufacturing environments and could be useful for real-time monitoring.