Probabilistic anomaly detection considering multi-level uncertainties for cable-stayed bridges

Abstract

To model multi-level uncertainties within anomaly detection process for large-span bridges, a probabilistic anomaly detection method is proposed considering uncertain models in the data collection, thermal response separation, and trigger estimation. The uncertain model in the data collection is established with the measured value and measuring errors. The uncertain model in the thermal response separation is built through the linear Bayesian estimation. The uncertain distribution of the anomaly detection trigger is obtained via Bayesian estimation of generalized Pareto distribution. Subsequently, measurements from multi-sensors are used to detect anomalies in a probabilistic way. Evidential reasoning, a decision-level fusion tool, is used to derive a collective detection rate to distinguish sensor malfunctions from anomalous scenarios. Specifically, anomalous scenarios deserve a large collective detection rate, whilst sensor malfunctions are subject to a small collective detection rate and a large individual detection rate. Two cases (i.e., sensor malfunction and snow disaster) are illustrated based on measurements from a large span cable-stayed bridge. As a result, the sensor malfunction is detected with an individual detection rate of 89.20% and a collective detection rate of 2.77%. The snowstorm is detected by a collective detection rate of almost 100%.