Guided waves with machine learning for structural health monitoring: transparent features and Monte Carlo confidence

Abstract

Reliable discrimination of small damage states under operational variability requires uncertainty aware Structural Health Monitoring. A pipeline is presented that couples guided wave physics with supervised machine learning to classify damage severity in a metallic panel. The experimental platform is a 310 × 190 × 1 mm aluminum plate with one central piezoelectric actuator and three receivers, interrogated by five cycle tone bursts at 20 kHz and sampled at 250 kHz. Signals are reduced to vectors of 20 physics informed features including root mean square, peak measures, analytic envelope statistics in fundamental and second harmonic bands, band limited energies, a spectral peak near 20 kHz, inter channel correlation, and a second harmonic index that captures weak interface nonlinearity. Uncertainty is propagated with Monte Carlo waveform perturbations, 5 000 realizations per condition, using amplitude scaling around 5 percent, time of flight jitter around 20 µs, and broadband noise near 2 percent of peak. These perturbations yield prediction bands and calibrated decision scores. The method is benchmarked across four learners: random forest, support vector machine with radial basis function kernel, additive boosting, and a hierarchical screener that first detects any mass and then separates severities. A finite element model provides a physics baseline for feature design. The study is a laboratory proof of concept on one specimen and three conditions, practical implications for aerospace deployment are outlined, including transfer to composite skins and links to certification metrics such as probability of detection. Calibration against the pristine response verified timing and mode content.