A noise-robust structural health monitoring framework using modified MFCC and deep autoencoders with Bhattacharyya distance

Abstract

This study proposes a novel structural damage detection method that integrates Mel-frequency cepstral coefficients (MFCCs) and deep autoencoder (DAE) networks to enhance robustness against measurement noise and uncertainties. MFCCs are extracted from power spectrum ratios derived from vibration signals to serve as noise-resilient features representing the dynamic characteristics of structures. A DAE is trained using healthy-state MFCCs to learn their intrinsic patterns, and reconstruction errors on testing data are subsequently analyzed. To account for uncertainties, multiple measurements are performed, and the resulting mean absolute error (MAE) distributions are modeled using Gaussian processes. The Bhattacharyya distance is then employed to quantify the differences between MAE distributions under healthy and potentially damaged states, leading to the definition of a damage indicator. Two case studies, including laboratory-controlled experiments on simply supported beams and field testing on a steel bridge, are conducted to validate the method. The results demonstrate that the proposed approach effectively identifies structural damage and exhibits strong resilience to varying noise levels, outperforming conventional MFCC-based techniques. This method shows significant potential for practical applications in structural health monitoring under uncertain environments.