A Comparative Study of Approaches to Damage Detection R.J. Barthorpe 1 , K.Worden 1 , C.Surace 2 , & G. Demarie 2 1 Dynamics Research Group, Department of Mechanical Engineering, University of Sheffield, UK E-mail: r.j.barthorpe@sheffield.ac.uk, k.worden@sheffield.ac.uk 2 Department of Structural and Geotechnical Engineering, Politecnico di Torino, Italy E-mail: cecilia.surace@polito.it Keywords: Structural Health Monitoring, damage detection, diagnostic methodologies. SUMMARY. The objective of the current paper is to discuss the relative strengths and weaknesses of the two main approaches to the diagnostic element of Structural Health Monitoring (SHM) and to try to indicate where they are best used in practice. Given the degree of commonality that is apparent, the paper will also discuss how the two approaches can support each other in the development of best practice and some speculation will be made as to how the approaches might be combined in order to exploit the strengths of both. 1 INTRODUCTION Current approaches to the diagnostic problem which is central to the field of Structural Health Monitoring (SHM) are usually based on two main possibilities: an inverse problem formulation and a machine learning approach. The first of these approaches, often called the model-based approach is usually applied by constructing a physics-based model of the structure of interest (e.g. a Finite Element (FE) model) and correlating it with experimental data. Once the model is established, it can be used in a monitoring phase by periodically updating the parameters of the model, usually by linear-algebraic methods. The nature of the problem means that the linear- algebraic formulation is often ill-posed and requires careful regularisation [1,2]. The alternative approach to diagnostics in SHM, often called the data-based approach, also involves the construction of a model, but this model is usually statistical. The model is established by means of machine learning or pattern recognition and may involve the use of classifiers or novelty (outlier) detectors [3,4]. It must be recognised that the problem of implementing a credible SHM strategy in any real-world context is much more wide-ranging than the choice of a diagnostic methodology. The broader aspects of SHM are however, not discussed here; the reader may consult [5] and [6] for more background; reference [7] must be considered as the current definitive guide to the subject. Both of the approaches discussed above have substantial support in the literature of SHM; however, they arguably have different strengths and weaknesses, which potentially make their domains of application problem-dependent. The methods also show a degree of commonality which is sometimes overlooked. In the first case, as observed above, both approaches can be said to be model-based. The distinction is in the type of model. If one classifies models into white, grey and black-box models according to their degree of a priori physical content; one would observe that the inverse problem approach seeks to establish a white-box model, while the machine learning approach uses a grey or black-box model. The advantage of the former is precisely that it exploits any available physical knowledge of the system of interest; the advantage of the latter is