Frequency-based change detection via multi-output log-gaussian cox processes

Abstract

With the development of sensor technology and an avalanche of distributed sensors, the capability to describe patterns and detect change-points is a core skill in system monitoring and prognostics. When data takes the form of frequencies or the number of counts, counting processes such as Poisson processes have been extensively used for modeling. However, most of the existing frequency-based approaches rely on parametric models or deterministic frameworks, thus failing to consider the complex systems’ uncertainties with temporal and environmental contexts. Another challenge is analyzing interrelated multi-sensors simultaneously to detect change-points that cannot be found independently.

This paper presents a multi-output log-Gaussian Cox process (MOLGCP) approach as a frequency-based change-point detection algorithm for real-time monitoring of dynamic systems. MOLGCP models the time-varying intensities of focal events defined over multiple correlated channels in a flexible and interpretable way. Cross-spectral mixture (CSM) kernels are used for model construction to capture both negative and positive correlations as well as the phase difference between channels. Adaptive and scalable decision-making strategies are suggested to identify anomalies in real-time. We show that computational complexities can be reduced and the method can be implemented for online purposes. Finally, extreme value theory (EVT) is used to set up dynamic thresholds considering the correlation between channels. Our method is validated with two different types of datasets: synthetic data and vibration data.