Ali Anaissi, Nguyen Lu Dang Khoa, Yang Wang
International Journal of Data Science and Analytics
Machine learning algorithms have been employed extensively in the area of structural health monitoring to compare new measurements with baselines to detect any structural change. One-class support vector machine (OCSVM) with Gaussian kernel function is a promising machine learning method which can learn only from one-class data and then classify any new query samples. However, generalization performance of OCSVM is profoundly influenced by its Gaussian model parameter 𝜎 . This paper proposes a new algorithm named appropriate distance to the enclosing surface (ADES) for tuning the Gaussian model parameter. The semantic idea of this algorithm is based on inspecting the spatial locations of the edge and interior samples, and their distances to the enclosing surface of OCSVM. The algorithm selects the optimal value of 𝜎 which generates a hyperplane that is maximally distant from the interior samples but close to the edge samples. The sets of interior and edge samples are identified using a hard margin linear support vector machine. The algorithm was successfully validated using sensing data collected from the Sydney Harbour Bridge and vehicle-mounted sensors for damage detection, in addition to five public data sets. The results obtained by ADES are compared to those of variance–mean, maximum distance and MIES methods. The results of the ADES approach outperform these state-of-the-art methods, especially on the bridge and road data sets. Experiments on these data sets show that an average 3% better accuracy is achieved by ADES over these state-of-the-art methods. The designed ADES algorithm is an appropriate choice to identify the optimal value of 𝜎 for OCSVM, especially in high-dimensional data sets.