EMind: A Foundation Model for Multi-task Electromagnetic Signals Understanding

Luqing Luo, Wenjin Gui, Yunfei Liu, Fengxiang Wang, Ziyue Zhuang, Yunxi Zhang, Zonghao Guo, Qirui Zhao, Zizhi Ma, Hanxiang He, Mingxuan Liu, Zhou Cong, Xinzhu Liu, Jinhai Li, Xin Qiu, Wupeng Xie, Yangang Sun, Maosong Sun

Keywords: electromagnetic signals; foundation model; dataset construction; multitask learning

TL;DR: This paper presents an application of foundation models to the electromagnetic signal modality.

Abstract: Deep understanding of electromagnetic signals is a prerequisite for electromagnetic space intelligence applications. Electromagnetic signals, with their high heterogeneity and prominent time-frequency dynamics, pose challenges for extracting representative features. Moreover, the scarcity of high-quality datasets further hinders the performance of electromagnetic signal foundation models under cross-task and cross-scenario conditions. To address these issues, we present a pre-training pipeline for electromagnetic signals foundation model, featuring the creation of a large-scale electromagnetic signal dataset and a foundation model to extract generalizable representations across diverse signals. We curate EMdata-81M, a high-quality electromagnetic signal dataset integrating 14 public and in-house sources, which is cleaned, annotated, and formatted uniformly. EMdata-81M comprises 81 million samples, covering 4 scene types, with signal lengths ranging from 128 to 4,096. To enable efficient training of large-scale variable-length data, we propose EMind, a foundation model tailored for electromagnetic signals. Specifically, EMind leverages a low-redundancy length adaptive multi-signal packing method and a hardware-aware adjustable dataset weighting strategy, improving representative feature extraction and, in turn, enhancing performance across downstream tasks. Extensive experiments demonstrate that EMind achieves state-of-the-art in various tasks, including modulation classification, parameter regression, radio frequency fingerprinting, and interference recognition, under full fine-tuning, strict train-test splits, and few-shot scenarios. It further attains competitive results on generative tasks including blind source separation and signal denoising. This highlights the effectiveness and scalability of our pipeline in unified electromagnetic signal understanding.

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

Submission Number: 2734