Deep Learning-based Pilot Adaptation and Channel Estimation in OFDM Systems

Clear communication over wireless channels demands overcoming their disruptive effects. Doubly-selective fading channels, with rapidly changing parameters, pose a particular challenge for accurate channel estimation. Traditional models often falter, lacking solutions or becoming overly complex. This is where deep learning takes center stage. In OFDM systems, pilots assist in estimating the channel response, but they also come at the cost of reduced data throughput. Adaptive adjustment of pilot patterns based on the channel state offers a promising solution. This paper introduces a deep learning-based framework that leverages adaptive pilots for fast-varying channels. We employ two deep neural networks. First, the pilot adaptation network dynamically selects the pilot pattern, reacting to the channel coherence bandwidth. Second, the channel estimation network extracts features from the channel frequency response using a 1D convolutional neural network. It then harnesses the power of long short-term memory layers to learn the channel behavior and estimate the response across all pilots and data subcarriers. Training and testing datasets are generated using WINNER II. The entire communication link, equipped with our proposed method, undergoes rigorous simulations, evaluated by both bit error rate and pilot overhead. The simulation results illustrate that the proposed scheme outperforms the previous methods, because it yields the same or better errors with less pilot overhead. This translates to a substantial data rate boost, paving the way for faster wireless communication.