Artificial Neural Network (ANN)-based Crack Identification in Aluminum Plates with Lamb Wave Signals YE LU, 1,2 LIN YE, 1, *ZHONGQING SU, 2 LIMIN ZHOU 2 AND LI CHENG 2 1 Laboratory of Smart Materials and Structures (LSMS), Centre for Advanced Materials Technology (CAMT) School of Aerospace, Mechanical and Mechatronic Engineering, The University of Sydney, NSW 2006, Australia 2 Department of Mechanical Engineering, The Hong Kong Polytechnic University, Hong Kong SAR, China ABSTRACT: An inverse analysis based on the artificial neural network technique is introduced for effective identification of crack damage in aluminum plates. The concepts of digital damage fingerprints and damage parameter database, which are prerequisites for neural network developing and training, are presented. Parameterized modeling for finite element analysis and an information mapping approach are applied to constitute the damage parameter database cost-effectively. The generalization performance of the neural network is examined by a process of ‘leave-one-out’ cross-validation and diverse factors are discussed, based on which the optimization of the neural network architecture is evaluated. The capability of this inverse approach is assessed by two crack cases from experiments, with good accuracy obtained in damage parameters (central position, size, and orientation). Key Words: Lamb waves, artificial neural network, damage detection, digital damage fingerprints. INTRODUCTION T HE artificial neural network (ANN) technique is a promising solution for effective damage identifica- tion as a typical nonlinear inverse problem. Various changes in the characteristics of structural dynamic/ static signals associated with damage have been employed for network training because they are easy to capture and sensitive to the existence of damage in some cases. These parameters include modal shapes and frequencies (Chang et al., 2000; Suh et al., 2000; Yun and Bahng, 2000; Ni et al., 2002; Yuan et al., 2003), displacement (Xu et al., 2001), velocity (Xu et al., 2004), and strain (Kudva et al., 1992; Shaw et al., 1995; Hwu and Liang, 2001). For identification of crack damage, Mahmoud and Kiefa (1999) input the first six natural frequencies into a neural network to estimate the location and size of surface cracks (15% deeper than the beam depth) in a steel cantilevered beam. Using natural frequencies and modal shapes as input patterns, Choubey et al. (2006) determined the size and location of through-thickness cracks which were 30% larger than the effective length of vessel structures. Liu et al. (2002) validated the feasibility of the ANN technique for crack detection using the responses of surface displacement. However, it is understood in general that initial or local damage leads to less detectable or undetectable changes in global structural dynamic parameters at low frequen- cies, which, as input vectors, compromise the perfor- mance of the ANN technique (Su, 2004). Lamb waves propagating through damaged plate-like structures carry characteristics that can be correlated with the location and severity of a defect, providing another set of information for effective damage evalua- tion. For example, forward estimation of crack para- meters (location, size, and orientation) and their effects on Lamb wave propagation have been extensively investigated in previous studies. It has been observed that, with an increase in excitation frequency, the value of the reflection coefficient of Lamb waves approaches a plateau of crack depth to plate thickness for surface- breaking cracks, and of crack length to plate width for through-thickness cracks (Lowe, 1998; Lowe and Diligent, 2002). Similar results have also been obtained in pipes for the reflection of the fundamental torsional mode by notches and cracks of different depths and circumferences in a frequency range of 0.01–0.3 MHz (Demma et al., 2003). However, a single actuator/sensor pair might be appropriate for evaluating only one crack parameter in simple cases, and crack size and orientation cannot be determined simultaneously with certainty because different cracks may show the same reflection/ transmission coefficient, resulting in an incorrect judg- ment (Lu et al., 2007, 2008). As an inverse problem, it is difficult to employ these characteristics directly from *Author to whom correspondence should be addressed. E-mail: l.ye@usyd.edu.au Figures 1 and 7 appear in color online: http://jim.sagepub.com JOURNAL OF INTELLIGENT MATERIAL SYSTEMS AND STRUCTURES, Vol. 20—January 2009 39 1045-389X/09/01 039–11 $10.00/0 DOI: 10.1177/1045389X07088782 ß SAGE Publications 2009 Los Angeles, London, New Delhi and Singapore at The Hong Kong Polytechnic University on February 4, 2016 jim.sagepub.com Downloaded from