Article Preview
TopIntroduction
System identification is a process of determining parameters of a dynamic system based on numerical analysis of measurement of input (excitation) and the corresponding output (response). In addition to updating numerical models for better response prediction, it is in principle possible to monitor the state of the structure and even detect damages based on changes in identified parameters (Koh et al., 2003). Damage detection is a challenging problem that is under vigorous investigation by numerous research groups using a variety of analytical and experimental techniques. Most of them are computationally expensive and tend to be numerically unstable for the case with a multi-degrees-of-freedom (MDOF) system. However, intelligent systems can address these issues. One such intelligent system is the Artificial Neural Network (ANN). Neural networks (NNs) are computing systems with the ability to learn from trainings, and are developed to imitate the way humans manage and process information. Based on trained neural networks, the behavior of complex systems may be modeled and predicted, even without a priori information about the structural or mathematical model (Wu et al., 1992; Elkordy et al., 1993; Hagan et al., 1996).
The multilayer feed forward neural networks with back propagation training algorithm commonly called the back propagation neural (BPN) networks are one of the most popularly used neural network architectures for damage prediction. The increase in the number of hidden layers and the network inputs from the experimental investigations showed the good accuracy of the network for the prediction of single and multiple member damage cases (Wu et al., 1992; Elkordy et al., 1993; Pandey & Barai, 1995; Adeli & Park, 1995). Another class of neural networks which have been used for structural damage assessment are counter propagation networks (CPNs). The CPN is marked by its quick training speed and ability to give good results for structural damage of beams and frames using various inputs like static displacements, natural frequencies, mode shapes and other parameters based on mode shapes (Szewczyk & Hajela, 1994; Zhao et al., 1998; Alok Madan, 2005). Both BPNs and CPNs have their own limitations and abilities for identification of structural damage. Most of these strategies involved identification of damage in a single stage scheme that required the network to be trained with large number of data for multiple member damage identification thereby affect the training efficiency and accuracy of the network. Very few studies have been conducted to improve training efficiency and accuracy of NNs by hybridizing (combining CPNs and BPNs) (Prashanth & Shankar, 2008), by selecting effective inputs (Ni et al., 2002; Prashanth & Shankar, 2008) and by generating efficient patterns (Yun & Bahng, 2000). Genetic algorithm (GA) is also used in detecting the structural damage by formulating the inverse problem into an optimization problem with error minimization between predicted and actual response of the structure (Mares & Surace, 1996; Friswell et al., 1998; Chou & Ghaboussi, 2001; Hao & Xia, 2002; Koh et al., 2003). In recent years a special class of artificial neural networks, the radial basis function (RBF) networks have received considerable attention because of their generalization ability. However, their use in applications for detecting structural damage has been limited. Some of the studies reported using RBFs on damage detection are structural identification of a helicopter rotor (Reddy & Ganguli, 2003), comparison of RBFs with multilayer perceptron (MLP) NNs for identifying creep deformation (EI-Shafie et al., 2010) and application of hybrid method of RBFs combined with genetic algorithm (GA) and fuzzy logic to locate the delamination in composite beams (Zheng et al., 2011). The minimum prediction error for damage detection by NNs reported is 3% in the reviewed literature, but because of the hybridization of NNs with other methods like GA and fuzzy systems more computation effort is consumed for good accuracy (Ni et al., 2002). Thus there is a need of further investigation on the improvement and simplification of RBF neural networks on the multi-member damage identification to get good accuracy with less effort.