DEU-Net: A Self-Supervised Deep Learning-Based Denoising Model for Distributed Acoustic Sensing

Distributed Acoustic Sensing (DAS) is an emerging vibration measurement technology with wide applications in seismic signal analysis, characterized by its high sensitivity and resolution. However, the acquired signals contain significant amounts of noise. This paper proposes a U-Net-like model based on morphological filtering operations of dilation and erosion, replacing the standard convolution blocks in the structure with dilated-eroded convolution blocks. Furthermore, inception blocks are added to the first and last layers of the model to effectively remove noise from DAS signals while preserving signal amplitude and reconstructing weak P-wave signals. The effectiveness and practicality of this method are validated through comparative studies with Non-Local Means (NLM), BM3D, and JDAS methods.