Bayesian Combination of Weighted Principal-Component Analysis for Diagnosing Sensor Faults in Structural Monitoring Systems

Abstract

It is essential to diagnose, i.e., detect and isolate, potential sensor faults for structural health monitoring to guarantee reliable condition evaluations. This paper proposes an innovative method called weighted principal-component analysis for sensor-fault detection and isolation. It is first illustrated that the fault sensitivity of each principal direction of traditional principal-component analysis is different from others for the same fault occurring in a certain sensor. Then, a fault-sensitive factor is theoretically derived to quantify the fault sensitivities. Based on that, a weighted fault-detection statistic determined according to the difference in fault sensitivities is developed and shown to have enhanced fault-detection ability. Bayesian inference is used to integrate all the weighted statistics corresponding to all the sensors to quickly judge whether a sensor fault occurred. Meanwhile, contribution analysis is used to establish a fault isolation index to identify the specific faulty sensor. Case studies using numerical simulation and a benchmark model demonstrate that the new proposed method is excellent and superior to the traditional approach.