On the Design of a Light-Weight Deep Learning Framework for Embedding in 5G Software Modem

• Published in: IEEE access 2024, Vol.12, p.151008-151018
• Main Authors: Min, Woonggyu, Kang, Seungwoo, Kim, Juyeop, Jo, Ohyun
• Format: Article
• Language: English
• Subjects: 5G mobile communication; Adaptation models; Computational modeling; Deep learning; Embedding; Hardware; Image enhancement; Modems; Radio signals; Real-time systems; Reference signals; Signal classification; Signal processing algorithms; Software; software modem; Software radio; Source code; Weight reduction; Wireless communication
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3477427

In this paper, we present a light-weight deep learning framework specifically designed and implemented for embedding in 5G software modems. The framework is developed completely using the C language to operate in real time by being mounted on a software modem. The framework incorporates an imagification process proposed by the authors which can enhance efficient reference signal classification in constrained environments. Imagification is the proposed technique that converts radio signal data, which is sequence data, into image form, processing the data into a structure similar to the RGB color model used in traditional images. This enables the application of convolutional operations to reduce complexity and training time. The performance of the framework is validated using a spec-compliant 5G software modem testbed developed by the authors, achieving up to 99.7% accuracy even at a relatively low SNR of −2.74 dB. These results demonstrate the feasibility of integrating the deep learning framework into a practical 5G software modem. Additionally, we perform hyperparameter optimization to identify the most suitable learning structure for the system. The developed source code is available at Github for public use.