Enhancing CNN-based Channel Estimation using Transfer Learning in OFDM Systems
Lingrui Zhu, Carsten Bockelmann, Thorsten Schier, Salah Eddine Hajri, Armin Dekorsy · 2023
In this paper, we investigate the application of deep transfer learning to channel estimation for an orthogonal frequency division multiplexing (OFDM) system. Recently, deep learning has been applied for channel estimation and shown its promising performance, while it suffers in the presence of a mismatch between the training phase channel model and real-world channel conditions. In the following, we deploy two different convolutional neural network (CNN) models from the literature and highlight their performance degradation caused by mismatch problems of power delay profile (PDP) and Doppler spreads. Transfer learning then is deployed to resolve the mismatch problem without the need to completely retrain CNN models. Our results show that we can alleviate the degradation using smaller efforts with transfer learning, especially for CNN with a deeper structure. In the end, when comparing transfer learning and data augmentation, our study also indicates that transfer learning is the better choice when coping with channel mismatch problems.