Damage diagnosis using a kernel-based method

This paper presents the use of a kernel-based machine learning technique, popular in the field of pattern recognition, to detect and classify various forms of damage states in both isotropic and anisotropic structures. A classification algorithm based on one-class Support Vector Machines (SVMs) is used for damage detection. The SVMs use a Gaussian kernel to map the input attributes to the high dimensional feature space and the transformed features are linearly separated by a decision plane. A procedure for obtaining the optimal value of the hyperparameter that controls the smoothness of the kernel is described. The type of damage addressed in this paper includes a combination of loose bolt and fatigue crack damage in a single lap, Al 6061-T651, bolted joints. Graphite/epoxy composite laminates with different types of damage are also studied, taking into account uncertainties in the measurement and material properties. The results show that the algorithm is able to accurately distinguish between different torque states and changes in crack length in the bolted joint sample. In anisotropic media, the algorithm was able to detect and classify various damage signatures with significant accuracy, using mutual information of two sensors. The algorithm was able to produce similar levels of accuracy when variability due to material properties was introduced.