Multi-index probabilistic anomaly detection for large span bridges using Bayesian estimation and evidential reasoning

To measure uncertainties within anomaly detection and distinguish sensor faults from anomalous events, a multi-index probabilistic anomaly detection approach is proposed for large span bridges based on Bayesian estimation and evidential reasoning. To avoid false detection raised by signal spikes, an energy index is first defined and extracted from pre-processed measurements, including missing data recovery and thermal response separation. Then, a probabilistic index, namely, certainty degree, is derived from probability density functions of detection triggers – extreme values predicted by using Bayesian estimation of the generalized Pareto distribution. To distinguish sensor faults from anomalous scenarios, evidential reasoning is applied to incorporate multiple certainty degrees into a joint one under the assumption that the probability of multi-sensor failing simultaneously is extremely low. Specifically, a large joint certainty degree indicates a high occurrence probability of anomalous scenarios, while a small one together with a large individual certainty degree depicts a high probability of sensor faults. Finally, the effectiveness of the proposed anomaly detection method is validated through structural health monitoring data from the Nanjing Dashengguan Yangtze River Bridge. Measurements from four sensors, that is, three cable forces and one deflection, are selected to detect anomalies based on their high pair-wise correlations. Two case studies are presented, namely, sensor fault detection and snow disaster detection. The sensor fault is detected through a certainty degree of almost 100% for the individual index and a joint certainty degree of nearly 0. The snowstorm is detected by a joint certainty degree of 36.82%