ScholarWorks@성균관대학교

초록

Time series anomaly detection faces a fundamental trade-off between detection accuracy and computational efficiency. We present MAMBA-PAD, a selective state space framework that combines adaptive Δ t gating, channel-wise variable selection, and stabilized dual-objective training. The model achieves high detection accuracy with linear-time inference and remains robust across diverse anomaly regimes. Key stabilizers include annealed weighting between reconstruction and classification, validationcalibrated thresholds, and a fastpath for strongly periodic signals. Experiments on eight benchmark datasets (NAB, SMAP, MSL, NYC Taxi, ECG, Machine Temperature, Network Traffic, CPU Utilization; Yahoo S5 skipped for class imbalance) show F1=0.86 ± 0.07, AUC=0.93, and 14 ms inference latency; total training across datasets completes in 36-41s. Interpretability analysis yields Pearson r=0.67, Spearman ρ=0.68(p<0.001); channel-wise gates highlight domain-relevant variables.

키워드