EM-ADSNet: Joint Detection and Separation of Aliasing Electromagnetic Signals With an Unknown Number of Sources

Abstract

The rapid growth of electronic systems, including communication and radar platforms, has made electromagnetic environments increasingly complex. Signals with different modulation formats can overlap in space, time and frequency, creating severe aliasing in electromagnetic signal monitoring. Conventional separation methods typically require prior knowledge of the number of sources. This requirement limits their use in open electromagnetic environments, where the number of active signals is often unknown and difficult to estimate. To address this limitation, we propose EM-ADSNet (Electromagnetic Adaptive Detection-Separation Network), a single-channel deep-learning framework for joint signal-number detection and blind separation. Specifically, EM-ADSNet first employs a high-dimensional feature processing network to enhance the representation capacity for complex mixed electromagnetic signals. A Multi-Kernel Attention (MKA) encoder is then proposed to capture multi-scale signal characteristics through multi-kernel convolutions while enhancing the discriminative representation of mixed components via a lightweight attention mechanism. To handle mixtures with an unknown number of constituent signals, a dedicated detection module and an adaptive mask generation module are jointly designed. The detection module estimates the signal number by formulating it as a multi-class classification problem, and the estimated count further guides the mask generation module to activate the corresponding number of mask outputs for per-signal separation. Experiments on simulated datasets and the real-world MIT RF Challenge benchmark show that EM-ADSNet effectively separates complex mixed signals under a blind separation setting. The model remains robust under varying noise conditions and consistently outperforms state-of-the-art baseline methods.