Blind Estimation of the Mixed Source Number in uwDAS Single-Channel Vibration Signals Based on 2-D CNN

The sensing signal of the ultraweak fiber grating distributed acoustic sensors (uwDAS) system is a mixture of the superposition of multiple sources and noise due to high sensitivity. In the underdetermined case, especially of a single channel mixture, traditional source number estimation algorithms fail to accurately estimate the number of sources. To solve this problem, this article proposes a single-channel source number estimation method based on an attention mechanism and a 2-D convolutional neural network (CNN). The laboratory data collection system and urban road sensing platform for uwDAS are established. Various vibration signals are collected, and simulated multisource mixed datasets and real multisource mixed datasets are created. A 2-D CNN is constructed, and the optimization experiments of CNN layers and input features show that the three-layer CNN based on the Mel spectrum achieves good performance. Experimental results indicate that, under laboratory conditions, the accuracy of source number estimation on the test set reaches 0.86. Specifically, the accuracy of estimating the source number for the three-source real mixed vibration signals is as high as 0.95. Under buried conditions, the source number estimation accuracy for single-source and two-source vibration signals on road surfaces both exceeds 0.8.Index Terms—Attention mechanism, convolutional neural network (CNN), Mel spectrum, source number estimation, ultraweak fiber grating distributed acoustic sensors (uwDAS).