CARE: A Cascaded Framework for Efficient and Reliable Time Series Anomaly Detection

Deep learning models have achieved state-of-the-art performance in time series anomaly detection, yet their complex architectures incur substantial inference overhead. Existing methods uniformly apply costly inference to all inputs, ignoring that anomalies are inherently scarce and most temporal data exhibit predictable patterns. To address this inefficiency, we propose CARE, a model-agnostic cascaded inference framework that pairs a Lightweight Pre-filter Model (LPM) with a high-capacity Complex Detection Model (CDM). The LPM rapidly filters high-confidence normal samples using a Residual MLP-based AutoEncoder and a Normality-Conditioned Gating mechanism. The gating mechanism incorporates Structure Attention and Risk-State Memory to capture channel-wise anomaly contributions and temporal risk evolution. A budget-aware gating objective learns normality confidence rankings while aligning CDM allocation with the available computational budget. Extensive experiments across eight real-world benchmarks demonstrate that CARE effectively isolates high-confidence normal samples. By prioritizing low-confidence samples for CDM processing, our framework achieves inference speedups of up to $4.1\times$ compared to the most accurate SOTA approaches, while maintaining competitive detection quality.