Damage detection in phase II structural health monitoring benchmark problem using Bayesian designed artificial neural network

Abstract

Pattern recognition using artificial neural network (ANN) is one of the promising approaches for detecting damages in structures. The basic idea of applying ANN in structural damage detection is to treat the calculated pattern features from a structural model as input and the corresponding damage scenarios as target in training an ANN. The trained ANN is then able to estimate the damage scenario by fitting the measured pattern features to the input of it. However, the design of the ANN is critical to the damage detection performance. This study presents a Bayesian model class selection method for optimal design of the ANN based on the given set of input-target training pairs, and hence, it avoids any subjective judgment and ad hoc assumption in the ANN design. The ANN designed by the Bayesian model class selection was applied to detect damages in the IASC-ASCE Structural Health Monitoring (SHM) Phase II Simulated Benchmark structure. In this study the damage induced changes in modal parameters were used as pattern features in the damage identification. Four damage cases were considered, in which single and multiples damages were considered in the IASC-ASCE SHM Phase II Benchmark structure. The results have shown that the ANN designed by the Bayesian model class selection method was able to accurately identify the damage locations and severities in all damage cases.