Acoustic Detection of Lining Void: Realization with Light-Weight Algorithms and Edge Devices
Guan Wang, Mei Wang, Liyan Luo, Zhenghong Liu, Yingjie Zhong · 2024
This paper proposes a strategy for deploying lightweight and miniaturized deep-learning algorithms on edge devices, aiming to solve the problems of poor accuracy in manual operations and high cost of traditional detection devices in tunnel lining void detection without relying on cloud-service conditions. By analyzing different voiceprint characteristics of knocking sounds, the Mel-Frequency Cepstral Coefficients (MFCC) are used to extract key features of sound signals, and Principal Component Analysis (PCA) and Global Average Pooling (GAP) techniques are adopted for dimension-reduction processing to reduce data volume and computational resource requirements. On the GPU platform, multi-dimensional high-speed parallel training of the Artificial Neural Network (ANN) is achieved. Using the RK3568 platform, edge-device deployment testing without relying on Internet-of-Things communication is finally realized. Experimental results show that when the MFCC dimension reaches 16 and above 20, the accuracy of the test set exceeds 96%, verifying the effectiveness and high-efficiency of this scheme.