부산대학교 도서관

Primary activity statistics contribute (PAS) significantly in increasing the efficiency of the dynamic spectrum access/cognitive radio system. PAS can be estimated from the spectrum sensing observations. To achieve a precise estimation of PAS, accurate spectrum sensing is required. However, it is difficult to maintain perfect spectrum sensing in a realistic scenario because of various hardware and sensing errors (false alarms and miss detections). In this work, Long-Short Term Memory autoencoder based deep learning framework is proposed to detect the sensing errors in imperfect spectrum sensing scenarios. Moreover, to correct the sensing errors, we propose a simple single iteration reconstruction algorithm and further estimate the PAS. The error in the estimated PAS is quantified through the Kolmogorov Smirnov distance. Finding suggests that relative error of estimated mean decreases from 80% to 9%. The proposed framework doesn’t require any prior knowledge of PU activity statistics to achieve this performance making it feasible in realistic scenarios.