Training-Free Self-Supervised Zero-Shot Anomaly Detection for Multivariate Time-Series Data

This paper presents a fully training-free, self-supervised, zero-shot anomaly detection framework for multivariate time-series data. The proposed approach operates without labeled anomalies, backpropagation, or retraining, making it suitable for real-world deployment scenarios where anomaly labels are unavailable. The framework segments sensor streams into sliding windows and constructs deterministic statistical embeddings capturing marginal behavior and cross-sensor dependencies. Anomaly detection is achieved through a combination of semantic rarity estimation, robust statistical deviation using median-based normalization, and cross-sensor dependency shift analysis. Unlike learning-based approaches, the method does not require parameter optimization or threshold calibration, enabling stable, interpretable, and computationally efficient deployment. Experimental evaluation demonstrates consistent qualitative and quantitative performance against classical unsupervised baselines, particularly under distribution shift and non-stationary conditions. The proposed system provides a reproducible and deployment-ready alternative to deep learning-based anomaly detection models, emphasizing interpretability, robustness, and zero-shot generalization.