Fault-Tolerant Source Detection using Bayesian Sensor Reliability Models

Abstract

Mobile radiation detector systems are meant to effectively detect the presence of sources of dangerous radiation while minimizing the frequency of false alerts. Our Bayesian Aggregation (BA) framework allows for sensor fusion of a variety of sensing modalities to enable detection of radioactive sources as well as inference of their properties. In this study, we extend the BA framework to enable online monitoring of sensor reliability. Sensor failures can arise both from internal system failure (e.g. a broken detector crystal) and external environment changes (e.g. gain drift). Both can cause spurious false alarms and detection misses. Statistically modeling sensor failure data can help robustify detection performance against such occurrences. By augmenting BA with explicit sensor failure models and hypotheses, incorporating a Hidden Markov Model (HMM) to estimate and track sensor working/failure state at each time step, and using the Forward-Backward algorithm to retrospectively smooth out posterior probabilities given new data clues, we hope to enable more fault-tolerant and resilient threat detection.