An Application of Semi-Supervised Algorithms in Structural Health Monitoring

Structural health monitoring is improved by the application of statistical damage-detection. Classification of unlabeled data is one of major challenges in machine learning applications. Semi-Supervised methods are shown to enhance the efficiency of classification by using unlabeled data in classification process. In this study, a semi-supervised support vector machine is applied to classify between healthy and unhealthy stages of a sample mathematical model. A hybrid approach has been utilized to generate feature vectors from dynamic response of structure. Using acceleration responses, which are perturbed by noises to consider ambient and sensor noises, dynamic properties of model are obtained by stochastic subspace state-space system identification methods. Modal strain energy is used as damage sensitive feature (DSF). This DSF is then used in Transductive Support Vector Machine (TSVM) algorithm. Also, Support Vector Machines (SVM) algorithm is utilized to compare results. To evaluate the performance of classification procedure, Precision and Recall criteria for the mentioned algorithms are presented. It can be seen that the use of unlabeled data will enhance the effectiveness of the classification methods especially in the lack labeled data. When labeled dataset are large enough, results for both supervised and semi- supervised support vector machines are almost the same.

Keywords:

Statistical Pattern Recognition , Support Vector Machines Method , Semi-Supervised Learning Algorithms , System Identification