Implementation and Evaluation of a Deep Learning-Based Wireless Decoder Using Real-World Over-the-Air Data

Wireless communication forms the backbone of our increasingly connected world. To address the growing demand for faster, more reliable, and adaptive wireless systems, researchers are exploring the potential of deep learning as a powerful alternative to classical model-based design approaches. Conventional receivers rely on predefined channel models for system optimization, which often fall short under real-world uncertainties. In contrast, deep learning-based receivers can learn complex channel characteristics directly from data, enabling enhanced adaptability and performance in dynamic environments. This paper presents the design, implementation, and evaluation of a deep learning-based decoder for an orthogonal frequency division multiplexing system using real-world over-the-air data captured via a software-defined radio testbed. Unlike most existing studies that rely solely on synthetic or simulation-based datasets, our setup collects real-world multi-path channel-impaired signals based on the IEEE 802.11n standard in presence of co-channel and/or adjacent channel interference effects in the experimental environment. The proposed deep learning decoder, implemented using a feedforward neural network architecture with linear and ReLU activation using stochastic gradient descent, reconstructs the original transmitted data from received baseband samples. Experimental results show that the deep learning-based decoder outperforms the conventional receiver, achieving up to a 3 dB improvement in block error rate for quadrature phase shift keying modulation even in the presence of interference. These findings demonstrate the feasibility and practical advantage of deep learning-based decoding under real channel conditions with interfering sources and pave the way for extending this approach to higher-order modulations and broader wireless communication standards.