PhyG-MoE: A Physics-Guided Mixture-of-Experts Framework for Energy-Efficient GNSS Interference Recognition

TL;DR

This work tackles GNSS interference recognition under energy constraints, a scenario where static deep learning models waste computation on simple signals and struggle with highly entangled spectra. It introduces PhyG-MoE, a spectrum-aware mixture-of-experts that gates computation using a PSD-derived router to activate a hierarchy of specialized experts (MobileNetV4, SK-GhostNet, and TransNeXt) only as needed. The approach achieves 97.58% overall accuracy across 21 jamming categories while reducing average FLOPs and maintaining low latency, with a demonstrably interpretable routing policy that aligns with signal entropy. Practically, PhyG-MoE enables energy-efficient, cognitively adaptive GNSS receivers for SAGIN by balancing recognition performance and resource usage in varying interference conditions.

Abstract

Complex electromagnetic interference increasingly compromises Global Navigation Satellite Systems (GNSS), threatening the reliability of Space-Air-Ground Integrated Networks (SAGIN). Although deep learning has advanced interference recognition, current static models suffer from a \textbf{fundamental limitation}: they impose a fixed computational topology regardless of the input's physical entropy. This rigidity leads to severe resource mismatch, where simple primitives consume the same processing cost as chaotic, saturated mixtures. To resolve this, this paper introduces PhyG-MoE (Physics-Guided Mixture-of-Experts), a framework designed to \textbf{dynamically align model capacity with signal complexity}. Unlike static architectures, the proposed system employs a spectrum-based gating mechanism that routes signals based on their spectral feature entanglement. A high-capacity TransNeXt expert is activated on-demand to disentangle complex features in saturated scenarios, while lightweight experts handle fundamental signals to minimize latency. Evaluations on 21 jamming categories demonstrate that PhyG-MoE achieves an overall accuracy of 97.58\%. By resolving the intrinsic conflict between static computing and dynamic electromagnetic environments, the proposed framework significantly reduces computational overhead without performance degradation, offering a viable solution for resource-constrained cognitive receivers.

Figures (9)

• Figure 1: Illustration of the GNSS interference scenario. The receiver captures the superposition of authentic satellite signals, jamming signals, and noise.
• Figure 2: Hierarchical taxonomy of the generated jamming dataset. The model expands from single primitives to dual-component and multi-component compound interference defined by the active set $\mathbb{K}$.
• Figure 3: STFT Spectrograms of different jamming scenarios. Top row: Single jamming primitives. Middle row: Dual compound jamming. Bottom row: Triple compound jamming. The TF representation effectively reveals the complex entanglement of features, such as the vertical bursts of Pulse jamming intersecting with the continuous sweep of LFM, which challenges static classifiers.
• Figure 4: Power Spectral Density (PSD) estimates of the corresponding jamming scenarios. The MTM-based PSD provides a clear view of the energy distribution across frequencies, aiding in the differentiation of spectrally overlapping signals.
• Figure 5: Overview of the Proposed Cognitive Dual-Stream MoE Framework. The architecture integrates three specialized expert networks: (1) A TransNeXt-based expert with Coordinate-GLU for directional feature modeling in saturated environments, (2) An SK-GhostNet expert for multi-scale adaptive reception under varying bandwidths, and (3) A MobileNetV4 expert utilizing Universal Inverted Bottlenecks (UIB) for hardware-aware universal feature extraction. A spectrum-aware router mechanism dynamically aggregates expert outputs for robust classification.