Piezoelectric Sensor-Based Health Monitoring of Railroad Tracks Using a Two-Step Support Vector Machine Classifier

Abstract

A piezoelectric sensor-based health monitoring technique using a two-step support vector machine (SVM) classifier is developed for railroad track damage identification. A built-in active sensing system composed of two lead-zirconate-titanate patches was investigated in conjunction with both impedance and guided wave propagation methods to detect two kinds of damage in a railroad track (hole damage 0.5 cm in diameter at the web section and transverse cut damage 7.5 cm in length and 0.5 cm in depth at the head section). Two damage-sensitive features were separately extracted from each method: (1) Feature I: root-mean-square deviations of impedance signatures; and (2) Feature II: sum of square of wavelet coefficients for maximum energy mode of guided waves. By defining appropriate damage indices from these two damage-sensitive features, a two-dimensional damage feature (2D DF) space was made. In order to enhance the damage identification capability of the current active sensing system, a two-step SVM classifier was applied to the 2D DF space. As a result, optimal separable hyperplanes were successfully established by the two-step SVM classifier: damage detection was accomplished by the first step SVM, and damage classification was carried out by the second step SVM. Finally, the applicability of the proposed two-step SVM classifier has been verified by 30 test patterns obtained in advance from the experimental study.