A novel Mel-frequency cepstral analysis based damage diagnostic technique using ambient vibration data

A two-stage novel Mel-frequency cepstral analysis based damage diagnostic technique addressing detection, localization and characterization using ambient vibration data is proposed. In the first stage of damage detection, the measured acceleration time history responses are converted into damage sensitive feature vectors using the concept of Mel-Frequency Cepstral Coefficients (MFCC). The damage is then identified using these transformed cepstral feature vectors using the popular Mahalanobis distance measure. The proposed damage index for detection clearly distinguishes damage and other operational and environmental variations using a threshold established based on the time history data of the healthy structure. The proposed damage index for detection also works with limited measurements and fewer features. In the second stage, the difference in spatial curvature of the Mel-frequency coefficient energy between the reference and current state of the structure is proposed for damage localization. The proposed damage localization index is robust enough to identify multiple damages present in the structure and also found to be insensitive to noise. The effectiveness of the proposed approach is verified using the synthetic datasets of the benchmark simply supported beam simulating environmental variability and experimental datasets of the Wooden Truss Bridge, provided by Helsinki Metropolia University of Applied Sciences. Finally, experimental investigations have been carried out to demonstrate the practical viability of the proposed MFCC approach.

Automated summary

A two-stage novel Mel-frequency cepstral analysis based damage diagnostic technique is proposed for detection, localization, and characterization using ambient vibration data. The technique effectively distinguishes damage from other variations and is robust enough to identify multiple damages present in the structure. The proposed approach has been verified using synthetic datasets and experimental datasets, demonstrating its practical viability.