Contrastive Learning for Anomaly Detection in Industrial Production Settings

Anomaly detection in multivariate time series is a critical task in industrial production environments, where timely identification of abnormal behavior can prevent costly failures. Recent advances in contrastive representation learning promise robust and generalizable embeddings for such data. This thesis integrates TS2Vec, a hierarchical contrastive learning model, together with additional contrastive learning variantsinto the Exathlon benchmark framework for anomaly detection and explanation discovery. To assess their effectiveness, the models are compared against an Autoencoder baseline using Exathlon’s anomaly detection metrics. A reference-based nearest-neighbor distance scoring method on the learned embeddings was developed and incorporated into the pipeline to generate anomaly scores. Beyond detection accuracy, the study examines model-agnostic explanation techniques and explores the potential of model-intrinsic mechanisms to improve interpretability. Results highlight the strengths and limitations of contrastive approaches compared to reconstruction-based methods and provide insights into the trade-offs between accuracy and explainability in anomaly detection.