基于条件整流流的OFDM资源栅格生成式数据扩充技术研究

Generative Data Augmentation for OFDM Resource Grids Based on Conditional Rectified Flow

摘要:

针对正交频分复用（OFDM）小样本条件下训练数据覆盖不足、原始波形直接建模困难以及传统局部扰动增强难以刻画资源栅格整体分布的问题，对基于条件整流流的OFDM资源栅格生成式数据扩充方法进行了研究；构建了接收端OFDM资源栅格的I/Q双通道表示，设计了融合类别嵌入、FiLM特征调制、轻量U-Net骨干网络及物理机理约束的条件整流流生成模型；对BPSK、QPSK、16QAM和64QAM四类OFDM信号在加性高斯白噪声和瑞利衰落信道条件下进行了仿真实验；结果表明，在信噪比（SNR，signal-to-noise ratio）为5 dB条件下，生成样本的平均星座距离、90%分位星座距离和均方根误差向量幅度分别为0.510、0.613和0.929，均优于cVAE和cWGAN-GP；在瑞利衰落信道下，经生成样本扩充后，分类器准确率和宏平均F1值分别由50.55%和50.30%提高至67.98%和67.29%；该方法能够在保持OFDM资源栅格结构特征和通信物理一致性的同时提升小样本条件下的调制识别性能

Abstract:

To address insufficient training-data coverage in small-sample orthogonal frequency division multiplexing (OFDM) scenarios, the difficulty of directly modeling original waveforms, and the limited ability of conventional local-perturbation augmentation to characterize the overall distribution of resource grids, a generative data augmentation method for OFDM resource grids based on conditional rectified flow was investigated; an I/Q dual-channel representation of the received OFDM resource grid was constructed, and a conditional rectified flow generative model integrating class embedding, FiLM-based feature modulation, a lightweight U-Net backbone, and physical mechanism constraints was designed; simulation experiments were carried out on BPSK, QPSK, 16QAM, and 64QAM OFDM signals under additive white Gaussian noise and Rayleigh fading channels; the results show that, at a signal-to-noise ratio (SNR) of 5 dB, the generated samples achieve an average constellation distance of 0.510, a 90th-percentile constellation distance of 0.613, and a root mean square error vector magnitude (RMS-EVM) of 0.929, all better than those of cVAE and cWGAN-GP; under the Rayleigh fading channel, after augmentation with generated samples, the classifier accuracy and macro-F1 score increase from 50.55% and 50.30% to 67.98% and 67.29%, respectively; the proposed method improves modulation-recognition performance under small-sample conditions while preserving the structural characteristics and physical consistency of OFDM resource grids