Multivariate time series anomaly detection: A framework of Hidden Markov Models
This paper addresses the challenge of multivariate time-series anomaly detection by proposing a unified probabilistic framework based on Hidden Markov Models (HMMs). Unlike conventional univariate approaches, the framework explicitly models both the dynamics of state transitions and cross-dimensional dependencies among variables, jointly learning normal behavioral patterns via a probabilistic graphical structure. Parameters are efficiently estimated using the Expectation-Maximization (EM) algorithm, and anomalies are scored via a likelihood-ratio-based mechanism. Extensive experiments on benchmark multivariate time-series datasets—including SMD, MSL, and SMAP—demonstrate that the method achieves an average 12.6% improvement in F1-score over strong baselines such as Isolation Forest, LSTM-VAE, and DeepAR. The approach delivers high accuracy, robustness to noise and distribution shifts, and inherent interpretability through its probabilistic formulation and explicit state-transition modeling.