WTSEConv1d-Net: A Robust 1-D Signal Recognition Framework for Distributed Acoustic Sensing

Abstract

Phase-sensitive optical time-domain reflectometry ( $\varphi $ -OTDR)-based distributed optical fiber acoustic sensing systems have been widely applied in large-scale security monitoring due to their excellent spatial resolution, long-distance sensing capability, and immunity to electromagnetic interference. Currently, the recognition of distributed acoustic sensing (DAS) signals primarily relies on convolutional neural networks (CNNs). However, CNNs are susceptible to random noise interference and the diversity of vibration signals in complex environments. As a result, they may capture spurious features that are statistically correlated with the training data but unrelated to the target events. When such correlations fail in the testing environment, recognition accuracy and robustness degrade significantly. To address this problem, this article proposes an integrated 1-D convolutional structure, WTSEConv1d, which integrates wavelet convolution (WTConv) with the squeeze-and-excitation (SE) attention mechanism. WTConv effectively expands the model’s receptive field with low parameter overhead, improving its ability to model long-term dependencies and thus increasing event classification accuracy. Nevertheless, when processing $\varphi $ -OTDR signals, WTConv may exhibit excessive responses to common high-frequency noise. To address this issue, the proposed WTSEConv1d-Net incorporates the SE attention mechanism into WTConv to dynamically suppress redundant channel activations and guide the model to focus on critical vibration patterns, thereby enhancing feature representation capability. Experimental results show that the proposed method achieves high classification accuracy in recognizing typical vibration events in the public dataset, including 01_background, 02_dig, 03_knock, 04_water, 05_shake, and 06_walk, and exhibits strong capability in focusing on effective signal regions. Furthermore, to evaluate the generalization capability and recognition stability of the model under different application scenarios, an external validation was conducted on a self-collected buried fiber-optic dataset. The results demonstrate that the proposed method still maintains favorable recognition performance under the tested scenarios. These improvements enhance the intelligent sensing performance and engineering applicability of $\varphi $ -OTDR systems under complex operating conditions.