Support vector machines for automated modelling of nonlinear structures using health monitoring results

Structural health monitoring (SHM) is backwards analysis of past to current state of damage, but cannot create a structure-specific nonlinear model for forward analysis of future response and damage. This paper aims to develop an automated modelling approach to translate proven hysteresis loop analysis (HLA) SHM results into nonlinear foundation models for response forecasting in subsequent events, particularly for steel structures with post-yielding behaviors. Support vector machine (SVM) is employed to identify the proposed nonlinear baseline model. Stiffness features are extracted from HLA to train the SVM model incorporating the constraints of SHM identification. A proof-of-concept case study validates the ability of the proposed method to accurately identify 12 model parameters with average error of 2.8% for a nonlinear numerical structure in the presence of 10% RMS measurement noise. Experimental validation from a full-scale 3-storey real building shows the predicted nonlinear responses match the measured response well with cross correlation coefficients R-coeff = 0.94, 0.92 and 0.89 for the first, second and third floor, respectively. In addition, the predicted stiffness changes also match the SHM results very well with errors less than 2.1%. Finally, and most importantly, the identified model is able to predict the response of 2 further events with average of correlation coefficient R-coeff = 0.91 and average error of 1.9% for stiffness changes across all cases. The overall results validate the ability of the created predictive model to accurately capture the essential dynamics and structural degradation, as well as predicting future possible response and risk. (c) 2020 Elsevier Ltd. All rights reserved.

Automated summary

This paper proposes an automated modelling approach that translates hysteresis loop analysis SHM results into nonlinear foundation models for forecasting responses in subsequent events. Experimental validation shows the method accurately identifies model parameters and predicts structural responses effectively.