Reinforcement Learning-based Task Aware Denoising for Ambient and Instrument Noise in Das Seismic Data

Abstract

Distributed acoustic sensing (DAS) offers advantages such as low cost, high sampling density, and minor geological structures identification. However, while highly sensitive fiber-optic sensors effectively capture valuable data, they also introduce complex noise fields containing simultaneous interference from heterogeneous sources, such as ambient vibrations, instrument system noise, and optical artifacts. This complex mixture is termed ‘multitype noise’, which significantly degrades data quality and intensifies spatiotemporal variation features, posing great challenges for data denoising. Although existing machine learning (ML)-based denoising networks enhance performance by increasing depth or complexity, insufficient attention to multitype noise features limits interpretability. Such architectures are inherently rigid, failing to dynamically adapt to the evolving statistics of new noise types. To handle complex multitype noise in DAS seismic data, we propose a reinforcement learning-based task aware denoising (RL-TAD) model. This model decomposes multitype noise suppression into adaptive dominant-noise suppression by region. Under RL-TAD, the Asynchronous Advantage Actor-Critic (A3C) agent is aware of denoising tasks in each data region and formulates visualizable and interpretable filtering policies. These policies guide filtering action selection from the action space for targeted dominant noise suppression. The region aware filtering policy aligns with the regional consistency of DAS seismic data in time and reduces computation. The action space takes multiple lightweight task-specific denoising networks as efficient filtering actions. Filtering actions can be flexibly adjusted for different noise types. Synthetic and field applications demonstrate RL-TAD effectively suppresses multitype noise, accurately recovers signal details, and easily adapts to new noise types via fine-tuning.