When estimating the First Arrival Path (FAP) delay with a Convolutional Neural Network (CNN), a large set of labeled samples is required to cover the diversity of multipath fading. However, collecting such data in real-world scenarios is expensive. To address this scarcity, we propose a Generative Diffusion Model (GDM)-based augmentation method to enlarge the labeled training dataset for the CNN. Specifically, we employ the cross-correlation algorithm to extract delay features from time domain, thereby generating Cross-Correlation Function (CCF) sequences. To exploit the image processing strengths of GDM, we encode CCF sequences into images for GDM training and subsequently decode the generated images back into sequences to augment the CNN dataset. The results of the simulation experiments demonstrate that this pipeline effectively enlarges and diversifies the dataset for CNN training, improving the FAP delay estimation accuracy while maintaining compatibility with various channel sequences.
Quan et al. (Thu,) studied this question.