Investigating Robustness in Detection, Localization and Compensation of Sensor Malfunctions in Degrading Structural Health Monitoring Systems

Abstract

Structural Health Monitoring (SHM) systems are essential for damage detection and maintenance planning in aging infrastructure. However, sensor degradation increases epistemic uncertainty and leads to incorrect SHM assessments, making it necessary to develop robust methods for handling sensor faults. This work investigates the robustness of a recently proposed framework for detecting, localizing, and compensating faulty sensors. The approach uses the Mahalanobis distance between sensor data frames from a current and a reference period, relying on cross-validation among sensor outputs. It includes a so-called α-level mapping, which links evaluated distances to corresponding levels of uncertainty relying on common statistical thresholds, as well as an algorithmic mechanism for excluding faulty sensors based on individual trust scores. The robustness study, using artificial FEM data from a steel lattice mast, identifies three key parameters that primarily govern the correct classification of faulty and healthy sensors over time. Results from the case study show that the method remains effective under multiple sensor faults, and identifies a favorable parameter domain, yielding the optimal performance. These findings support practical parameter recommendations to maintain confidence in SHM systems during long-term operation.