A Bayesian jump model-based pathwise sampling approach for online anomaly detection

Moving vehicle-based sensors (MVSs) have received growing attention for real-time anomaly detection in various applications such as wildfire and oil spill detection. To tackle challenges due to the spatial covariance structure among observations, uncertainties under partial observations, as well as the physical MVS movements, we propose a Bayesian jump model-based pathwise sampling approach to detect abrupt changes in an area of interest in real time using MVSs. Specifically, a jump-model based Bayesian scheme is proposed to exploit spatial correlation and real-time partial observations for status update, and to quantify uncertainties arising from noisy observations, limited observability, and anomaly occurrences. Based on the updated status, a Bayesian upper confidence bound is constructed as the sampling statistic to balance exploration and exploitation under noisy and partial observations. Guided by the sampling statistic, a route optimization model is then formulated to adaptively coordinate the routes of multiple MVSs for quick anomaly detection. We perform theoretical investigations and conduct simulations to confirm the exceptional effectiveness of the proposed method. A case study for early wildfire detection demonstrates that our proposed method outperforms benchmark methods, which contributes to the reduction of the area of affected land and wildfire-related costs.