Variational autoencoder-based multipath clustering algorithm for terahertz channels

To address the shortcomings of multipath clustering algorithms in terahertz channel modeling, particularly in terms of multidimensional parameter adaptability and unsupervised feature separation, a variational autoencoder-based latent space multipath clustering (VAE-LMC) model was proposed. Firstly, the variational autoencoder (VAE) was utilized to learn latent representations of multipath delays and arrival angles, enhancing feature separability. Secondly, K-Means clustering was embedded into the VAE framework, with joint optimization of reconstruction loss, KL divergence, and clustering loss functions to resolve the challenges of feature separation in unsupervised learning. Finally, multipath clustering was performed in the latent space, and the results were mapped back to the real data space. Terahertz channel measurements at 129.5~135 GHz were conducted in a small factory scenario to construct training datasets and testing datasets. Experimental results demonstrate that the VAE-LMC model exhibits significant advantages in intra-cluster and inter-cluster characteristics, environmental consistency, and computational complexity, providing an efficient solution for terahertz channel multipath clustering in complex scenarios.