Mitigation of environmental variabilities in damage detection: A comparative study of two semi-supervised approaches

Vibration-based structural health monitoring (VSHM) employs vibration signals as observables from which inferences are made concerning the integrity of structural systems. More specifically, the premise of this work is to detect damage through changes in a set of features extracted from the vibration signals. A major challenge in this regard is that false positives may arise due to the influence of environmental and operational variabilities (EOVs). Environmental variabilities, e.g. shifts in temperature and humidity, introduce changes in mechanical properties. These changes are reflected in the vibration response and can reduce the probability of detecting damage in a structure. This paper conducts a comparative study between a novel semi-supervised damage detection approach and a well known cointegration-based scheme to deal with EOVs. The novel approach uses the pattern recognition capability of an artificial
neural network (ANN) to learn how EOVs affect a Mahalanobis distance-based damage index in a reference state. The cointegration-based scheme seeks to mitigate the EOVs by computing stationary linear combinations of non-stationary output response signals. The merits of the damage detection methods are examined in the context of a mass-spring system, which is exposed to a simulated temperature field that renders the output response non-stationary. The
system is analysed in a reference state and a perturbed state in which damage is emulated by
reducing a single spring stiffness by 2%. Both methods are evaluated with the area under the
curve (AUC) for receiver operating characteristic (ROC) and the false alarm rate. The results show that the ANN-based damage detection approach outperforms the cointegration-based one in this particular example.