Structural Novelty Detection With Modified Mel-Frequency Cepstral Coefficients and Bhattacharyya Distance

A novel data-driven damage detection framework, based on probabilistic modeling of mel-frequency cepstral coefficient (MFCC) distributions, is proposed. This study replaces traditional autospectral densities with the power spectrum derived from cross-spectral density ratios in MFCC extraction, enhancing sensitivity to subtle dynamic changes across structural locations. The Bhattacharyya distance (D_B) is introduced to quantify dissimilarities between MFCC distributions under baseline and potential damage scenarios. Subsequently, a damage indicator (DI_B) based on the Bhattacharyya distance is proposed. To mitigate uncertainties caused by environmental noise and measurement variability, a statistically sound threshold is established through Bayesian resampling and Monte Carlo simulation. When the DI_B values of structural states exceed this threshold, it indicates the presence of damage. Additionally, a vectorization scheme is employed to improve computational efficiency, enabling faster processing of multichannel data. The accuracy and effectiveness of the proposed method are validated through a laboratory experiment involving four beams and a field test conducted on a steel truss bridge. The results demonstrate the proposed method’s ability to detect and classify damage states accurately under diverse conditions, highlighting its applicability for reliable structural health monitoring (SHM).