A Model-Free Kullback-Leibler Divergence Filter for Anomaly Detection in Noisy Data Series

TL;DR

This paper introduces a model-free Kullback-Leibler Divergence filter that detects anomalies in noisy sensor data series, useful for non-destructive inspection in inaccessible environments, validated against industry benchmarks.

Contribution

It presents a novel, model-free KLD-based anomaly detection method applicable to noisy sensor data, with validation in industrial settings.

Findings

Effective in noisy environments

Outperforms existing algorithms

Applicable to various sensor types

Abstract

We propose a Kullback-Leibler Divergence (KLD) filter to extract anomalies within data series generated by a broad class of proximity sensors, along with the anomaly locations and their relative sizes. The technique applies to devices commonly used in engineering practice, such as those mounted on mobile robots for non-destructive inspection of hazardous or other environments that may not be directly accessible to humans. The raw data generated by this class of sensors can be challenging to analyze due to the prevalence of noise over the signal content. The proposed filter is built to detect the difference of information content between data series collected by the sensor and baseline data series. It is applicable in a model-based or model-free context. The performance of the KLD filter is validated in an industrial-norm setup and benchmarked against a peer industrially-adopted algorithm.