A Hybrid Hidden Markov Model and Time-Frequency Approach to Impact Echo Signal Classification

Impact echo (IE) is a popular nondestructive evaluation technique, which can detect and locate delaminations and subsurface defects within concrete bridge decks from specific spectral signatures of the recorded signals. The traditional approach to IE data analysis is transforming the signals into the frequency domain to identify echoes corresponding to slab thickness, delaminations, or other defects. More recent studies have highlighted the advantages of a joint time-frequency (TF) analysis. Classification of IE signals using TF features has shown great promise in identifying the deck deterioration state at a given test point. However, as shown in this study, TF models learn features specific to the training dataset when limited training examples are used, thereby limiting their performances when the characteristics of the test data differ from those of the training data. For example, if training and test data pertain to two bridges with different thicknesses, TF models have shown poor generalizability. Hence, we investigate an alternative machine learning (ML)-based approach using hidden Markov model (HMM) for IE signal classification. Unlike TF models, HMM attempts to learn the generic patterns in signals corresponding to different damage states, rather than data-specific features and, hence, performs consistently well even with limited data availability. However, when data from multiple bridges are used for training, the TF model outperforms HMM in classifying IE signals. A hybrid HMM-TF model that can increase the reliability of the classification outcome is thereby explored. Results suggest that this hybrid approach has comparable performance to the HMM when limited data is available, and outperforms TF when multiple bridge data are used.

Automated summary

Impact echo (IE) is a popular nondestructive evaluation technique for detecting defects in concrete bridge decks. Traditional frequency domain analysis is limited in performance when dealing with limited training data, while time-frequency analysis and machine learning methods offer improved classification accuracy. This study explores a hybrid model combining hidden Markov model (HMM) and time-frequency analysis, which shows improved reliability in classification under different data conditions.