On the use of the cepstrum and artificial neural networks to identify structural mass changes from response-only measurements Ulrike Dackermann 1 , Wade A. Smith 2 , Jianchun Li 1 , Robert B. Randall 2 1 Centre for Built Infrastructure Research, Faculty of Engineering and Information Technology University of Technology, Sydney, Australia e-mail: ulrike.dackermann@uts.edu.au 2 School of Mechanical and Manufacturing Engineering, University of New South Wales, Sydney, Australia Abstract This paper presents a damage identification technique based on response-only data utilising cepstrum analysis and artificial neural networks (ANNs) for the identification of added mass in a two-storey framed structure. The proposed technique applies cepstrum-based operational modal analysis (OMA) for the regeneration of frequency response functions (FRFs), and added mass is detected through the combined use of principal component analysis (PCA) for data compression and ANNs for feature extraction and pattern recognition. In particular, different treatments of the zeros in the curve-fitting of the transfer function cepstrum are investigated to improve the automation potential of the method for application in continuous online structural health monitoring (SHM). The proposed technique is validated on a laboratory structure tested on a large-scale shake table with ambient base loading. The results of the investigation show that the method is effective in identifying added mass based on response-only measurements. 1 Introduction To ensure the safety and reliability of ageing civil structures and to prevent catastrophic failures, early and reliable damage detection and health assessment is critically important. To facilitate this need, the field of structural health monitoring (SHM) has evolved since its inception in the 1960s, and it is widely applied today to assess the health condition of a structure and to thereby prolong its lifespan. SHM most commonly employs local and offline assessment techniques, with visual inspection being the most widespread method, followed by local non-destructive testing (NDT) techniques such as stress wave, ultrasonic wave, X-ray, magnetic field, acoustic, radiography and thermal field methods [1]. Most of these established techniques, however, are very labour intensive and time consuming [2], and often disassembly of secondary parts is required to gain access to vital load-bearing structural elements. An alternative to these local techniques is found in global approaches, which monitor the behaviour of the entire structure. Such behaviour is influenced by local changes, and hence, by solving an inverse problem, they detect, locate and evaluate damage with minimal labour and cost, and without requiring access to the damaged sites. Due to these benefits a major focus of SHM research in the last two decades has been on developing global NDT techniques, particularly ones that can be applied remotely and without supervision, thus making them suitable for online SHM applications. As such, vibration-based techniques, which generally satisfy these criteria, have received considerable attention in recent years [3]. Typical vibration-based parameters used for SHM include both indirectly measured data – such as resonant frequencies, mode shapes, modal flexibilities, modal stiffness and modal strain energy – and directly measured data – such as time domain data, power spectral densities and frequency response functions (FRFs) [1].