A Hyper-solution SVM Classification Framework: Application to On-line Aircraft Structural Health Monitoring

Abstract

Support Vector Machines (SVMs) classification learning is a powerful paradigm to investigate inverse input-output relationship of a specific problem according to some available and representative dataset. In particular, SVMs are able to identify even non-linear relationship by mapping non-linearly separable data into potentially linearly separable one through families of spatial transformation (kernel trick). With respect to this, several parameters (i.e., kernel function and its internal parameters) have to be tuned; however the optimal configuration is usually difficult to be defined a-priori.In this paper, a hyper-solution framework for SVM classification is presented. The main idea is to perform, simultaneously, three different SVM-based classification learning tasks: Model Selection, Multiple Kernel Learning and Ensemble Learning. The meta-heuristic known as Genetic Algorithms (GA) has been proposed to search for the most reliable final hyper-classifier (an SVM with a basic kernel, an SVM with a combination of kernel, or an ensemble of different SVMs, respectively), and the corresponding optimal configuration.We have applied the proposed framework on a critical and quite complex problem: the on-line assessment of structural health of aircraft fuselage panels, a crucial task both in military and civilian settings. In particular, the framework has been used to implement a diagnosis task, that is detecting a possible damage and identifying the structural component involved, according to the strain field measured through a monitoring sensor network deployed on the helicopter fuselage panels.Finite Elements (FE) simulation has been configured to simulate the response of a real panel to different damages. The resulting simulated strain fields have been used to build a dataset. Results obtained through 3 folds-cross validation proved the framework is reliable. Finally, when compared to results obtained by the authors in a previous work based on Artificial Neural Networks (ANN) classification learning paradigm, the proposed SVM framework proved to be more effective and reliable.