Cavity detection and localization based on pitch analyses and applications of multitask learning

Abstract

This study developed an anomaly detection method that utilizes a characteristic cavity-borne pitch and a nonparametric filter based on localized autocorrelation. Employing a simulated model cavity, microphone tests were conducted to obtain acoustic signals. Experimental findings revealed that the fundamental frequency of the cavity was approximately 174.3 Hz, and likelihood of cavity presence decreased with increasing cavity distance while free boundaries and looser sand density increased this likelihood. In addition, the proposed nonparametric filter enhanced accuracy of classification and cavity distance estimation of multitask learning models. This study suggests that the proposed methods can effectively detect and localize cavities.