To address the issues of sample imbalance, unstable generation quality, and insufficient feature extraction in spectrum anomaly signal detection under complex electromagnetic environments, this paper proposes a VAE-GANGP identification model that integrates a Variational Autoencoder (VAE) with a Gradient Penalty-based Generative Adversarial Network (GAN-GP). First, the VAE is employed to encode the original spectrum, generating structured latent features that follow a standard normal distribution. This replaces the random noise input in traditional GANs, significantly enhancing the semantic consistency of generated samples and training stability. Second, an adversarial training mechanism based on Wasserstein distance with gradient penalty (WGAN-GP) is introduced, effectively mitigating mode collapse and gradient vanishing, thereby improving the model’s capability to fit complex signal distributions. Furthermore, a multi-objective optimization function combining reconstruction error and adversarial loss is constructed, establishing an end-to-end integrated framework for feature learning, signal reconstruction, and anomaly discrimination. Experiments are conducted using a synthetic dataset comprising various modulation types and simulated environments with different signal-to-noise ratios for systematic validation. The results demonstrate that the spectrum data generated by VAE-GANGP closely matches the distribution of real signals. Under AWGN-dominated synthetic test conditions, the model achieves an anomaly detection accuracy of 98.1%. When evaluated under more realistic channel impairments (phase noise, multipath, impulsive interference), the model maintains competitive performance, outperforming existing methods and demonstrating promising potential for practical electromagnetic spectrum monitoring. Its performance significantly surpasses traditional detection methods and single deep learning models, providing a highly reliable and adaptive solution for spatial electromagnetic spectrum anomaly detection.

UAV radio-frequency (RF) signal identification is an important technique for low-altitude security monitoring. However, background noise interference and distortions in amplitude and phase features can reduce the robustness of methods based on a single frequency-domain representation. A one-dimensional power spectrum alone provides limited capability for characterizing pairwise relationships among frequency components, whereas two-dimensional feature representations may weaken spectral-line continuity. To address these limitations, this paper proposes a Power Spectral Density-Frequency-domain Gramian Angular Field Dual-stream Fusion Network (PFG-DSFNet). First, each RF observation window is converted into a power spectral density (PSD) sequence and further mapped to a frequency-domain Gramian Angular Field (F-GAF) image. Then, the PSD sequence and F-GAF image are used to extract the spectral envelope and heterogeneous two-dimensional frequency-domain features, respectively. Furthermore, an Adaptive Gated Low-rank Bilinear Interaction (AGLRBI) module is designed for the dual-stream features to model second-order interactions and perform adaptive fusion. In this way, the dual-stream representation framework enhances high-accuracy UAV RF signal identification. Experiments on the DroneRF 10-class classification task show that PFG-DSFNet achieves an accuracy of 99.35% and a Macro-F1 score of 0.9934. Ablation experiments further show that the bilinear interaction and adaptive gating components each contribute to the final performance, and their combination achieves the best result. Additional experiments under mixed-SNR conditions and in measured scenarios further support the robustness of the proposed method.