Deep Adaptive Cross Domain Learning for Continuous Pressure Wave Signal Recovery

ICLR 2026 Conference Submission19603 Authors

Keywords: Signal recovery, episodic learning, adaptive noise learning, frequency-aware adversarial alignment.

TL;DR: An episodic learning framework incorporating a novel noise adaptation mechanism that progressively shifts from model-level to input-level adaptation

Abstract:

Mud pulse telemetry (MPT) enables real-time downhole data transmission by generating continuous pressure
wave signals in drilling fluid, supporting measurement and control during drilling. However, the signals are vulnerable to noise from the downhole environment, mud channel, and surface equipment, causing attenuation, distortion, and phase errors that hinder accurate reconstruction. Recent advances in deep learning offer promising solutions to these challenges but typically requires large amounts of labeled data, which are difficult and costly to obtain in field conditions. To address this issue, this paper proposes the Deep Adaptive Cross Domain Learning Network (**DACDL**), a framework featuring a novel noise adaptation mechanism that transitions from model-level to input-level adaptation. Our approach introduces three core innovations:
(1) An Episodic Learning Framework (**EL-Framework**) that simulates domain shift by alternately learning from simulated (gaussian) and real-world (real-world gaussian, pump or oilfield) domains, enhancing few-shot adaptation under label scarcity;
(2) A lightweight Adaptive Noise Learning Block (**ANL-Block**) that introduces sample-specific perturbations to align target input noise distributions with the source domain, alleviating generalization collapse caused by unseen noise types;
(3) A Frequency-aware Adversarial Alignment Block (**FAA-Block**) that explicitly aligns spectral characteristics between source and target domains, mitigating cross-domain frequency mismatches.
Moreover, the proposed ANL-Block is model-agnostic and can be plug-and-play into most existing methods.
Experimental results on three manually collected field datasets demonstrate the effectiveness of DACDL in practical field scenarios and highlight the model-agnostic adaptability of the ANL-Block.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 19603