A Deep Learning Approach for Wireless Spectrum Sensing in Communications-based Train Control: A Over-fitting Problem and Solution

This work introduces the spectrum sensing-based deep learning approach to overcome the wireless interference in the communication-based train control application. The four-level-SNR classification problem in this application is defined. Recently, several works applied the end-to-end learning approach using convolutional neural networks for spectrum sensing. However, the present work points out that the over-fitting problem easily occurs if only the limited frequency selective fading conditions of a data set are considered for the training process in the end-to-end learning approach on the multiple-SNR classification. This over-fitting problem cannot be solved simply by adding more frequency selective fading conditions into the training data set because there are many possible conditions in real communication transmission. This paper then proposes a new learning network, including a new input feature that has a strong relationship with the multiple-SNR classification problem. The evaluation results suggest that the proposed approach can solve such an over-fitting problem.