Research On DE-BP Neural Network Spectrum Prediction For Cognitive Radio

Spectrum prediction is one of the key technologies in cognitive radio (CR) systems. Spectrum prediction technology can minimize the interference that unlicensed users or second user (SU) impose on licensed users (LU) or primary user (PU) so that unlicensed user or SU can find more available spectrum holes, thereby improving the overall system’ spectrum utilization. Therefore, spectrum prediction technology has atracted an extensive research in the academic community. In practical, it is difficult to obtain the utilization characteristics of licensed frequency bands, and the spectrum prediction contains complex non-linear characteristics. As neural network does not need inherent priori information distribution, neural network based method becomes an attractable spectrum prediction algorithm. Among the existing neural network based spectrum prediction algorithms, the network training cannot satisfy the desired results in most circumstances for its slow convergence and tendency to local minimum. To solve this problem, DE-BP based neural network spectrum prediction method is proposed in this thesis.Firstly, in this thesis, the current research situation of cognitive radio spectrum prediction is overviewed, and some existing spectrum prediction algorithms are analyzed and compared. Then, cognitive radio systems’channel status modeling methods are researched, especially two common spectrum modeling methods:Hidden Markov model and queuing model. These two models are analyzed through both theoretical and experimental ways. Meanwhile, M/Geo/1queuing model is utilized to generate channel status data of licensed users, and the data is considered as spectrum data of spectrum prediction method involved in this thesis.Secondly, the BP neural network is used to research and simulate cognitive radio spectrum prediction. Then, BP neural network’s threshold update formula is derived. The generated spectrum data in previous simulation study is regarded as the experimental data. Futhermore, the spectrum prediction results is applied to spectrum sensing. Theoretical analyses and simulation results show that this approach can save sensing energy and improve spectrum efficiency compared with single spectrum sensingThirdly, to solve the problem of slow convergence and tendency to local minimum existed in BP neural network based spectrum prediction methods, a DE-BP based neural network spectrum prediction algorithm which combine standard differential evolution (DE) algorithm with BP algorithm is designed. And the generated spectrum data is used to verify its prediction performance. Experimental results indicate that the algorithm can improve the spectrum prediction accuracy both in single-channel prediction and multi-channel combining prediction to some extent. Meanwhile, the algorithm’s spectrum prediction results are futher applied to spectrum sensing. Theoretical analyses and simulation results show that, compared with single spectrum sensing, the proposed approach in this thesis can reduce the interference to licensed users caused by unlicensed users as well as save sensing time.Finally, as standard DE algorithm’s slow convergence, tendency of local minimum and prematurity, a self-adaptive differential evolution algorithm (ADE) which combines a self-adaptive crossover probability computation method with existing chaotic sequence mutation factors is designed. Then the proposed ADE algorithm is integrated with BP algorithm to generate an ADE-BP algorithm. And it is futher applied to spectrum prediction. Simulation and experimental results indicate that the spectrum prediction accurancy of the proposed ADE-BP based neural network spectrum prediction algorithm in this thesis is superior to the DE-BP based algorithm without increasing the complexity.