Bayesian Hierarchical Model for Change Point Detection in Multivariate Sequences

Abstract

Motivated by the wind turbine anomaly detection, we propose a Bayesian hierarchical model (BHM) for the mean-change detection in multivariate sequences. By combining the exchange random order distribution induced from the Poisson–Dirichlet process and nonlocal priors, BHM exhibits satisfactory performance for mean-shift detection with multivariate sequences under different error distributions. In particular, BHM yields the smallest detection error compared with other competitive methods considered in the article. We use a local scan procedure to accelerate the computation, while the anomaly locations are determined by maximizing the posterior probability through dynamic programming. We establish consistency of the estimated number and locations of the change points and conduct extensive simulations to evaluate the BHM approach. Among the popular change point detection algorithms, BHM yields the best performance for most of the datasets in terms of the F1 score for the wind turbine anomaly detection.

KEYWORDS:

• Change points
• Multivariate data
• Nonlocal prior
• Nonmaximum suppression
• Poisson–Dirichlet process

Supplementary Materials

PDF file containing additional simulation results, the dynamic programming algorithm as well as the proofs of Lemma 1 and Theorem 1. RCode_and_data: Zip file containing the R code for implementing the BHM method as well as the real data used in the article in RData form.