Prediction Of Spectrum Entropy Based On Feature Vectors

With the development and emergence of wireless communication services, the demand for the spectrum resource has drawn increasingly attention. However, the spectrum resource is usually allocated by a fixed assignment strategy, which is inefficient and causes wastage of spectrum resource. The cognitive radio is proposed to address the issue of spectrum efficiency and has been receiving an increasing attention in recent years.In cognitive radio networks, secondary users(SUs) use the unused licensed spectrum to communicate. In such a way, the efficiency of spectrum utilization could be improved. In order to enable SUs to quickly find unused licensed spectrum to communicate, researchers mainly studied how to predict the vacation portions of the licensed spectrum band as accurately as possible. Unfortunately, the idle spectrum bands change in the communication environment constantly. Therefore, the SU has to correspondingly track the idle spectrum band. If the unused licensed spectrum switches too frequently, it may cause severe delay and high packet loss, which may degrades the quality of communication seriously. Therefore, a suitable frequency band for SUs communication not only needs to be idle, but also needs to be idle for a long period of time. In other words, the spectrum should be stable: the spectrum should not change raptly between idle state and busy state. This thesis contributes to study how to measure and predict the stability of spectrum utilization. The main contributions are summarized as follows:First, this thesis proposes to use spectrum entropy to measure the stability of spectrum utilization.The utilization of the spectrum can be regarded as random variables, and the uncertainty of random variables can be measured by information entropy. Based on this idea, this thesis tries to measure the stability of spectrum utilization with the spectrum entropy. Furthermore, this thesis makes a quantitative analysis on the correlation of spectrum entropy. The results show that the spectrum entropy of the same channel has a high correlation, which indicates that predicting the spectrum entropy is possible.Second, this thesis proposed a new method, called mRMR-SVR, for the prediction of spectrum entropy.The traditional prediction methods for time series usually directly use continuous time delays to construct the predictor model, without excluding the time delays which are uncorrelated and may have a negative influence on prediction. This thesis proposed the maximum relevance minimum redundancy feature selection method(mRMR) and support vector regression(SVR) method for predicting the spectrum entropy. mRMR is able to analyze the correlation of time series and selects those delays which have the largest contribution to the prediction. While SVR is expected to train the selected features in order to select the optimal feature and predict the spectrum entropy.Third, this thesis carries out considerable experiments to evaluate the proposed method.Based on the spectrum data collected in Shenzhen, this thesis verifies the proposed prediction method, i.e., mRMR-SVR, and compares this method with the traditional ARIMA method. According to the experimental results, mRMR-SVR outperforms ARIMA by more than 21% in prediction accuracy.