Abstract
Hammering inspection is one of the most widely used methods for concrete structure condition assessment. This paper proposes effective hammering sound analysis scheme, which delivers favourable defect detection performance under various noise situations. This work is inspired by the facts that hammering echo waveforms are noisy and redundant, while the underlying defect-induced patterns are anticipated to be sparse with respect to certain frequency bands. Grounded on such hypothesis, we introduce sparse coding approaches for hammering sound analysis so as to characterize representative patterns from corrupted echo signal. Particularly, there are two major steps in sparse coding: dictionary learning and codes generation. We investigated the two parts separately, allowing us to exploit contribution of each. Proposed approach is demonstrated with real-world data; furthermore, In order to validate noise robustness, echo condition assessment is performed under mismatch acoustic environments with varying noise intensities. Experimental results verified the effectiveness of our approach and even under extremely noisy scenario, i.e. -10dB, the system achieved 96% echo analysis accuracy. In addition, this paper also acts as a practical guide to select efficient sparse coding algorithms for real applications, not limited to hammering sound analysis.