SKANet: A Cognitive Dual-Stream Framework with Adaptive Modality Fusion for Robust Compound GNSS Interference Classification

TL;DR

This work tackles the challenging problem of classifying compound GNSS interference by proposing SKANet, a cognitive dual‑stream network that jointly analyzes Time‑Frequency Images and Power Spectral Density. A Multi‑Branch Selective Kernel–Asymmetric Convolution Block enables adaptive receptive fields to capture both micro‑scale pulses and macro‑scale sweeps, while a Squeeze‑and‑Excitation fusion module harmonizes information from the TF and PSD streams. On a large synthetic dataset of nine compound interference classes, SKANet achieves an Overall Accuracy of 96.99% and shows strong robustness in low‑JNR conditions, significantly outperforming several CNN/Transformer baselines. The architecture offers a practical pathway for reliable, real‑time GNSS interference classification in dynamic electromagnetic environments, with potential benefits for cognitive radio and secure navigation systems.

Abstract

As the electromagnetic environment becomes increasingly complex, Global Navigation Satellite Systems (GNSS) face growing threats from sophisticated jamming interference. Although Deep Learning (DL) effectively identifies basic interference, classifying compound interference remains difficult due to the superposition of diverse jamming sources. Existing single-domain approaches often suffer from performance degradation because transient burst signals and continuous global signals require conflicting feature extraction scales. We propose the Selective Kernel and Asymmetric convolution Network(SKANet), a cognitive deep learning framework built upon a dual-stream architecture that integrates Time-Frequency Images (TFIs) and Power Spectral Density (PSD). Distinct from conventional fusion methods that rely on static receptive fields, the proposed architecture incorporates a Multi-Branch Selective Kernel (SK) module combined with Asymmetric Convolution Blocks (ACBs). This mechanism enables the network to dynamically adjust its receptive fields, acting as an adaptive filter that simultaneously captures micro-scale transient features and macro-scale spectral trends within entangled compound signals. To complement this spatial-temporal adaptation, a Squeeze-and-Excitation (SE) mechanism is integrated at the fusion stage to adaptively recalibrate the contribution of heterogeneous features from each modality. Evaluations on a dataset of 405,000 samples demonstrate that SKANet achieves an overall accuracy of 96.99\%, exhibiting superior robustness for compound jamming classification, particularly under low Jamming-to-Noise Ratio (JNR) regimes.

Figures (8)

• Figure 1: System model of the proposed GNSS interference scenario. (a) Physical layout of a GNSS receiver under compound jamming, where the source transmits superposed primitives (e.g., LFM and Pulse) mixed with a power ratio $\alpha$. (b) Signal superposition model at the receiver front-end. The received signal $y(t)$ is modeled as the summation of aggregate GNSS signals $\sum s_{\text{GNSS}}^{(m)}(t)$, compound jamming $j(t)$, and additive white Gaussian noise $n(t)$. Key parameters, including Power Ratio (PR) and Jamming-to-Noise Ratio (JNR), are defined based on the component power levels.
• Figure 2: The hierarchical taxonomy of GNSS interference types considered in the proposed framework. The model categorizes interference into five fundamental single primitives (STJ, MTJ, LFM, Pulse, PBNJ) and their compound variations formed by linear superposition. This hierarchy serves as the basis for the dataset generation and the nine-class classification task, distinguishing between isolated jamming sources and complex entangled signal mixtures.
• Figure 3: Time-Frequency Images (TFIs) of the 9 compound interference patterns generated via STFT. The 3$\times$3 grid layout illustrates the distinct spectral overlap characteristics of different jamming combinations within the limited GNSS bandwidth.
• Figure 4: Power Spectral Density (PSD) analysis of the 9 compound interference types. The frequency domain characteristics provide complementary energy distribution profiles, facilitating in the discrimination of spectrally similar compound signals.
• Figure 5: Schematic of the proposed SKANet architecture. The framework features a cognitive dual-stream design to process heterogeneous features. The STFT Stream utilizes a deep hierarchical backbone with Multi-Branch SK-ACB modules to extract high-level semantic representations from Time-Frequency Images. The PSD Stream employs a lightweight shallow network to capture global energy statistics from Power Spectral Density. A Squeeze-and-Excitation (SE) fusion block adaptively recalibrates and integrates the multi-modal features via concatenation and element-wise multiplication before the final classification head.