信道质量预测在认知无线电频谱感知方面具有重要的作用，预测可靠与否直接影响用户的正常通信。然而，目前国内外对信道质量的预测是基于时隙的预测，对硬件的要求很高，需要预测时间短；另外，预测模型多采用统计学方法，性能不佳，容易造成用户信道选择上的冲突。因此，本文在传统的信道质量预测模型基础上，利用改进的k-means信道聚类算法，提出了一种综合考虑能量检测值、时间和频率的概率选频模型。一方面从能量检测法出发，预测最佳能量检测门限；另一方面从非授权用户角度出发，估计时域和频域对信道质量的影响。

本文首先研究了二值时间序列预测模型、信道质量相关性预测模型和信道可用性预测模型这三种模型，分析这三种模型的缺点，提出概率选频模型。其次，为了建立概率选频模型，需要对信道质量进行聚类，在分析了原有k-means算法的缺点后，结合最大最小距离法自适应设置初始聚类中心，利用轮廓系数获取最优k值，提出了改进的自适应k-means信道聚类算法。再次，针对影响信道质量的能量检测门限设置困难问题，提出了多门限能量检测法，即基于能量检测值的概率选频法。最后，设计基于单变量、双变量和多变量影响信道质量的因素实验，划分信道，选出高概率、高质量信道。

实验结果显示，信道质量不受频段的影响，而受时间的影响，即授权用户使用授权频段不存在频段的选择性，却有时间的选择性。而基于能量检测值的概率选频方法，即多门限能量检测法，相比于传统的能量检测法，频谱感知效率提高了66%。

Channel prediction plays an important role in spectrum sensing of cognitive radio. Prediction accuracy directly affects the users' normal communication. However, most research of the channel prediction models is based on time slot which needs high-performance hardware and fast prediction time. Most channel prediction approaches use statistics methods, which leads to the conflict of user channel selection. Therefore, inspired by the traditional channel prediction model, using an improved k-means clustering algorithm, this paper proposes a novel channel prediction model with probability which considers the energy detection value, time and frequency. On one hand, our model can predict the optimal threshold of energy detection method. On the other hand, our model can evaluate the influence of time and frequency to the channel quality for the secondary users.

In order to reduce the cost of spectrum sensing in cognitive radio, this paper firstly studies three channel prediction models, binary time series model, spectral correlation model and channel usage prediction model. The concept of a probability based model for channel prediction is proposed. This model considers the factors influencing channel quality which include energy detection value, time and frequency. Secondly, in order to establish the prediction model, this paper takes actual experimental data to obtain the empirical relations between influencing factors and channel quality by improved k-means clustering algorithm. Furthermore, the probability distribution of channel quality over influencing factors with which corresponding channels are selected for further finer spectrum sensing is obtained. It can also help to predict the threshold in energy detection based on actual requirements of cognitive users for channel quality.

The results show that the channel quality is influenced by time instead of frequency range, which means the primary users have a preference for time rather than frequency. Compared with the traditional energy detection method, the spectrum sensing efficiency of our energy detection value based probability method which also called the multi-threshold energy detection method increases 66%.