Using ROC curves to assess the efficacy of several detectors of damage-induced nonlinearities in a bolted composite structure

Abstract We offer a comparison of several different detectors of damage-induced nonlinearities in assessing the connectivity of a composite-to-metal bolted joint. Each detector compares the structure's measured vibrational response to surrogate data, conforming to a general model for the healthy structure. The strength of this approach to detection is that it works in the presence of certain types of varying ambient conditions and is valid for structures excited with any stationary process. Here we employ several such detectors using dynamic strain response data collected near the joint as the structure was driven using simulated wave forcing (taken from the Pierson–Moskowitz frequency distribution for wave height). In an effort to simulate in situ monitoring conditions the experiments were carried out in the presence of strongly varying temperatures. The performance of the detectors was assessed using receiver operating characteristic (ROC) curves, a well known method for displaying detection characteristics. The ROC curve is well suited to the problem of vibration-based structural health monitoring applications where quantifying false positive and false negative errors is essential. The results of this work indicate that using the estimated auto-bicoherence of the systems response produced the best overall detection performance when compared to features based on a nonlinear prediction scheme and another based on information theory. For roughly 10% false alarms, the bicoherence detector gives nearly 90% probability of detection (POD). Conversely, for several of the other detectors 5–10% false alarms leads to ∼ 70 % POD. While the bicoherence (and bispectrum) have been used previously in the context of damage detection, this work represents the first attempt at using them in a surrogate-based detection scheme.