A deep learning channel estimation scheme in orthogonal frequency division multiplexing for X-ray communication (XCOM) is studied. The scheme uses simulated and detected data obtained with different working parameters and numbers of pilots as training and testing data, respectively, for the deep neural network (DNN) model. The bit-error-rate performance of the DNN model under various system operating parameters, numbers of pilot sequences, and channel obstructions is investigated separately. Experiment results showed that the deep-learning-based approach can address the distortion of the air-scintillator channel for XCOM, giving a performance comparable to those of least-squares and minimum-mean-square error estimation methods.
