The Effect of Freezing and Thresholding Selection in Transfer Autoencoders on Time Series Anomaly Detection Zaman Serisi Anomali Tespitinde Transfer Otokodlayicida Dondurma ve Eşikleme Seçiminin Etkisi
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11637004
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: anomaly detection, freezing strategy, industrial control systems, LSTM autoencoder, transfer learning
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Anomaly detection in multivariate time series within industrial control systems is challenging due to label scarcity, domain shift, and process differences. This study investigates the cross-domain transfer performance of an LSTM-based sequential autoencoder (SeqAE) on the SWaT, WADI, and TEP datasets. The model was pre-trained in the source domain and evaluated in the target domain with a zero-training base and four transfer/freeze strategies. Four training ratios (10%, 20%, 50%, 70%), four thresholding methods (quantile, MAD, adaptive_MAD, top_k), and five random seeds were used in the experiments. The results show relatively high F1 values in the SWaT objective, low but close values across strategies in the WADI objective, and high F1/PR-AUC values in the TEP objective. Overall findings indicate that the effect of the freeze strategy is dataset-dependent and that threshold selection plays a critical role in performance.