Outlier Resilient Online Multivariate Change Point Detection Using Subsequence Divergence Estimation in Sensor Data Streams

Outliers in multivariate sensor signals affect the performance of online change point detection (CPD) algorithms. Data pre-processing to eliminate outliers is a standard solution, but it may be challenging for real-time applications. In this work, we propose an unsupervised change detection framework that consists of three stages. The first stage rapidly detects changes in multivariate sensor data streams in real-time using the KL-divergence estimator. The idea is to capture the abrupt and gradual variations using two rolling windows distant by a few subsequences. The second stage deals with the sensor stream redundancy issue by making small perturbations to ensure the positive definiteness of the covariance matrix. The third stage uses an optimal ensemble winsorization strategy to make the framework robust against outliers. In the absence of labeled datasets, the article presents two algorithms to generate synthetic data with abrupt and gradual changes. The proposed framework is compared with a few popular and contemporary change detection methods using a number of synthetic, benchmark, and experimental datasets. Results show that our framework outperforms the state-of-the-art methods in almost every aspect. The addition of perturbation to solve the issue of the singular covariance matrix brings more stability to the framework. Extensive experiments demonstrate that the proposed framework maintains impressive performance even in the presence of outliers. The framework’s robustness against outliers, combined with minimal latency time, makes it suitable for real-time decision-making applications dealing with sensor streams.