This paper emphasizes the importance of channel estimation in wireless communications, and points out the problem that existing neural network-based channel estimation methods are trained and tested only for specific or similar channels, resulting in poor generalization performance in various real channel environments. Considering the difficulty of online learning due to low latency and limited computational resources, we propose a design criterion for a neural network that shows robust performance on various wireless channels with only offline learning without prior channel information. Based on the proposed criterion, we propose a synthetic dataset generation method and a benchmark design that guarantees achieving a certain mean square error (MSE) in a new unknown channel, and show that the generalization performance of the proposed method is independent of the neural network structure through neural networks of various complexities. Experimental results confirm that the proposed method achieves robust generalization performance on wireless channels with fixed channel profiles and variable delay spreads.
