Automated structural damage detection using one class machine learning

Abstract

Measuring and analysing the vibration of structures using sensors can help identify and detect damage, potentially prolonging the life of structures and preventing disasters. Wireless sensor systems promise to make this technology more affordable and more widely applicable. Data driven structural health monitoring methodologies take raw signals obtained from sensor networks, and process them to obtain damage sensitive features. New measurements are then compared with baselines to detect damage. Because damage-sensitive features also exhibit variation due to environmental and operational changes, these comparisons are not always straightforward and sophisticated statistical analysis is necessary in order to detect abnormal changes in the damage sensitive features. In this thesis, an automated methodology which uses the one-class support vector machine (OCSVM) for damage detection and localisation is proposed. The OCSVM is a nonparametric machine learning method which can accurately classify new data points based only on data from the baseline condition of the structure. This methodology combines feature extraction, by means of autoregressive modeling, and wavelet analysis, with statistical pattern recognition using the OCSVM. The potential for embedding this damage detection methodology at the sensor level is also discussed. Efficacy is demonstrated using real experimental data from a steel frame laboratory structure, for various damage locations and scenarios.