Fuzzy Clustering of Spatially Relevant Acoustic Data for Defect Detection

Efficient diagnosis of concrete structures is a growing issue in modern societies where concrete is an omnipresent material. The hammering test is a traditional nondestructive testing method that has been employed in the field for a long time and for which automation is highly desirable. The problem consists in determining from the sound returned after a hammer strike on a structure's surface if there is a defect beneath or not. In this letter, we present an unsupervised learning approach for the automation of hammering test. First, mean shift is used with Mel-frequency cepstrum coefficients at various parameter values in order to find a stable mode configuration. Then, the corresponding peaks are used to obtain seeds for spatial fuzzy c-means to cluster hammering samples while combining audio and position information. Experiments have been conducted in indoor artificial environment on concrete blocks containing man-made defects. Results showed the effectiveness and robustness of the proposed solution on detecting different types and number of defects. Our approach showed promising performance also in tests performed in outdoor environment using a fully automated hammering system.