Post-Inference Guided Transformer for Anomaly Interval Localization in Multivariate Time Series

Uzuntürk G., TAŞKAYA TEMİZEL T.

35th IEEE International Workshop on Machine Learning for Signal Processing, MLSP 2025, İstanbul, Türkiye, 31 Ağustos - 03 Eylül 2025, (Tam Metin Bildiri) identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/mlsp62443.2025.11204313
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: Anomaly interval detection, domain knowledge, multiclass classification, multivariate time series, post-inference strategies, transformer networks
  • Orta Doğu Teknik Üniversitesi Adresli: Evet

Özet

Range-based anomaly detection in multivariate time series is critical for real-world applications such as industrial monitoring, where anomalies often span intervals and exhibit distinct types. We propose a transformer-based framework designed for highly imbalanced, multiclass anomaly detection at the range level. To improve prediction stability and coherence, we introduce two post-inference strategies: (i) a majority voting mechanism to consolidate overlapping multi-step predictions, and (ii) a transition-aware masking scheme that applies domain-specific rules to constrain class transitions. We evaluate our method on the Exathlon benchmark, extending its original binary protocol to support multiclass assessment. Under binary evaluation, our approach outperforms standard forecasting- and reconstruction-based baselines, achieving up to a 24% improvement in F1 score.