The selection of pattern features for structural damage detection using an extended Bayesian ANN algorithm

Abstract

Pattern recognition is a promising approach for the detection of structural damage using measured dynamic data. Much research of pattern recognition has employed artificial neural networks (ANNs) as a systematic way of matching pattern features. When such methods are used, the ANN design becomes the most fundamental factor affecting performance and effectiveness of the pattern recognition process. The Bayesian ANN design algorithm is proposed in Lam et al. [Lam HF, Yuen KV, Beck JL. Structural health monitoring via measured Ritz vectors utilizing artificial neural networks. Computer-Aided Civil and Infrastructure Engineering 2006;21:232–41] provides a mathematically rigorous way of determining the number of hidden neurons for a single-hidden-layer feedforward ANN. The first objective of this paper is to extend this Bayesian ANN design algorithm to cover the selection of activation (transfer) functions for neurons in the hidden layer. The proposed algorithm is found to be computationally efficient and is suitable for real-time design of an ANN. As most existing ANN design techniques require the ANN model class to be known before the training process, a technique that can automatically select an “optimal” ANN model class is essential. As modal parameters and Ritz vectors are commonly used pattern features in the literature, the second objective of this paper is to compare the performance of these two pattern features in structural damage detection using pattern recognition. To make a fair judgment between the features, the IASC–ASCE benchmark structure is employed in a case study. The results show that the performance of ANNs trained by modal parameters is slightly better than that of ANNs trained by Ritz vectors.