TriScope-MTSAD: Progressive Tri-Granularity Evidence Decoupling with Redundancy-Aware Regularization for Multivariate Time-Series Anomaly Detection

Multivariate time-series anomaly detection (MTSAD) is challenging because abnormal events may appear as fine-grained deviations, local temporal distortions, or global semantic shifts. Existing methods often learn a unified representation or rely on one dominant evidence source, limiting their ability to describe heterogeneous anomalies. This paper proposes TriScope-MTSAD, a progressive tri-granularity evidence-decoupling framework with redundancy-aware regularization. It formulates detection as a sequential evidence-generation process comprising micro reconstruction, local patch-pattern modeling, and global semantic modeling. The micro operator captures reconstruction deviations and produces residual-aware cues; the local operator transforms these cues into short-segment evidence; and the global operator evaluates window-level semantic inconsistency conditioned on local-context representations. A training-only decoupling regularizer projects operator representations into a common evidence space and suppresses excessive cross-operator correlation. Cross-fusion anomaly scoring then combines the resulting evidence streams through robust scaling based on median absolute deviation (MAD), interquartile range (IQR), and standard deviation (Std), followed by weighted integration. Offline evaluations on Soil Moisture Active Passive (SMAP), Server Machine Dataset (SMD), Pooled Server Metrics (PSM), and Mars Science Laboratory (MSL) show competitive or superior performance against classical and recent baselines. Results on two real grain monitoring datasets further demonstrate applicability to retrospective analysis of storage and transportation records.