ISCA Archive Interspeech 2025
ISCA Archive Interspeech 2025

PredTrAD – Prediction-based Transformer for Anomaly Detection in Multivariate Time Series Data

Jan Schuster, Alexander Wölfel, Fabian Brunner, Christian Bergler

Anomaly detection in multivariate time series data is an extensive field of research with significant impact on a broad spectrum of real-world applications. Building a reliable anomaly detection system is extremely challenging due to data imbalance, lack of labels, next to the actual definition of anomalies. The current study proposes a new transformer-based approach, together with a publicly available time series dataset – the TIKI data corpus – comprising various performance metrics of a Kubernetes cluster distributed over roughly 500 million timestamps. The proposed network, entitled PredTrAD, is compared with the state-of-the-art anomaly detection models – TranAD and DTAAD – in several experiments: (1) verification on three well-known benchmark corpora, (2) evaluation on the TIKI data, and (3) cross-entity anomaly detection. Across all experimental constellations, the proposed model has shown a significant performance growth, proven by F1 score enhancements of 2.41% up to 16.81%.