Data-driven deep representation of acoustic signals amplifies the accuracy of coin-tap test for non-destructive detection of defects

Coin-tap test is a simple way for non-destructive detection of defects, and has long been used in engineering structures. However, improving the accuracy of coin-tap test is challenging. In this work, we propose a data-driven deep representation method for acoustic signals to amplify the accuracy of coin-tap test. We design an incremental dense one-dimensional convolutional neural network (IDCNN) with two feature aggregation blocks to organize deep representations. We introduce six types of defects to three types of bi-layered structures, use coin-tap tests to obtain acoustic signals, and train the IDCNN. The results show that the IDCNN performs well for deep representation of acoustic signals and significantly amplifies the accuracy of coin-tap test. The accuracy rate for defect detection ranges from 98.42 % to 99.06 %. The rates of missing and false alarms for defects are extremely low, ranging from 0.94 % to 1.58 % and from 0.72 % to 1.32 %, respectively. The results show that the data-driven deep representation of acoustic signals results in an effective coin-tap test for non-destructive detection of defects. The proposed method has potential for broad applications in acoustic-based non-destructive tests.