极限学习机及其在无线频谱预测中的应用研究

发布时间：2018-05-20 19:13

本文选题：前馈神经网络 + 极限学习机　；参考：《兰州大学》2014年硕士论文

【摘要】：极限学习机(Extreme Learning Machine, ELM)是由黄广斌提出的一种新型单层前馈神经网络。与传统前馈神经网络相比,极限学习机结构简单,学习速度快且具有全局搜索能力和良好的泛化性能。但是在实际的应用过程中,极限学习机的隐含层的节点数的设置会对实际问题有所影响。过多的隐含层节点会产生过拟合,并且隐含层节点中不免会有对实际问题作用很小或者无用的节点。针对这一问题,一种基于极限学习机的最优裁剪极限学习机(Optimally Pruned Extreme Learning Machine, OP-ELM)被提出,通过对极限学习机中隐含层节点进行裁剪,提高了极限学习机的鲁棒性和泛化性。认知无线电(Cognitive Radio, CR)是目前解决无线频谱资源日益紧缺问题的关键技术。频谱分配的不合理是造成频谱资源紧缺的重要原因,认知无线电通过感知主用户(Primary User, PU)的频谱占用情况,使次用户(Second User,SU)充分利用主用户的频谱空洞,智能的对频谱资源动态分配,实现可靠的通信服务并提高频谱的利用率。频谱预测是认知无线电中的关键技术,传统的预测方法有很多,如马尔科夫链方法、回归分析方法和神经网络方法等。传统的预测方法预测所需的时间较长,不能满足频谱预测的实时性的要求。极限学习机的引入,不仅满足了频谱预测实时性的要求,且优化裁剪极限学习机的鲁棒性和泛化性优于传统的极限学习机,更适应于认知无线电频谱预测问题。本文主要工作如下。 (1)系统并深入地研究了极限学习机的原理和特点。介绍了极限学习机和几种改进的极限学习机的数学模型和训练算法,从理论上阐述了极限学习机的快速特性。通过仿真实验比较分析比较了极限学习机与传统前馈神经网络的预测性能。 (2)针对隐含层节点数目过多,会影响网络性能的问题,对经典的极限学习机进行隐含层节点的调整,构造最优裁剪极限学习机。并通过基准实验对其性能进行了分析。 (3)将极限学习机用于认知无线电的频谱预测问题。针对现有预测方法在预测精度和预测速度上存在的不足,利用极限学习机的简单、快速及全局最优等特点,对认知无线电系统中主用户的频谱状态持续时间进行预测。比较了极限学习机及其几种改进模型与传统的前馈神经网络,反馈神经网络在频谱预测问题上的性能,实验表明,与传统前馈神经网络和反馈神经网络相比,极限学习机,特别是最优裁剪极限学习机,无论是在预测精度上还是在预测速度上都获得了较好的性能,更适用于无线频谱预测问题。
[Abstract]:Extreme Learning Machine, ELM) is a new single-layer feedforward neural network proposed by Huang Guangbin. Compared with the traditional feedforward neural network, the LLM has the advantages of simple structure, fast learning speed, global searching ability and good generalization performance. However, in the practical application, the number of nodes in the hidden layer of the LLM will affect the practical problems. Too many hidden layer nodes will be over-fitted, and there will inevitably be nodes with little or no effect on practical problems. In order to solve this problem, an optimal Pruned Extreme Learning Machine, OP-ELM based on LLM is proposed. The robustness and generalization of LLM are improved by cutting the hidden layer nodes in LLM. Cognitive Radio Cognitive Radio (CRC) is the key technology to solve the problem of increasing shortage of wireless spectrum resources. The unreasonable spectrum allocation is an important reason for the shortage of spectrum resources. By sensing the spectrum occupation of primary user (PU), cognitive radio makes the secondary user make full use of the main user's spectrum hole. Intelligent dynamic allocation of spectrum resources to achieve reliable communication services and improve the spectrum efficiency. Spectrum prediction is a key technology in cognitive radio. There are many traditional prediction methods, such as Markov chain method, regression analysis method and neural network method. The traditional prediction method takes a long time and can not meet the real-time requirement of spectrum prediction. The introduction of LLM not only meets the requirement of real-time spectrum prediction, but also optimizes the robustness and generalization of LLMs, which is more suitable for cognitive radio spectrum prediction. The main work of this paper is as follows. The principle and characteristics of LLM are studied systematically and deeply. This paper introduces the mathematical models and training algorithms of the ultimate learning machine and several improved learning machines, and expounds the fast characteristics of the ultimate learning machine theoretically. The prediction performance of LLM and traditional feedforward neural network is compared by simulation experiments. 2) aiming at the problem that too many hidden layer nodes will affect the performance of the network, the classical ultimate learning machine is adjusted to construct the optimal clipping ultimate learning machine. Its performance is analyzed by benchmark experiment. In this paper, the extreme learning machine is applied to the spectrum prediction of cognitive radio. Aiming at the shortcomings of existing prediction methods in prediction accuracy and prediction speed, the duration of spectrum state of primary users in cognitive radio systems is predicted by using the characteristics of simple, fast and global optimization of extreme learning machines (LLMs). This paper compares the performance of LLM and its improved models with those of traditional feedforward neural networks and feed-back neural networks in spectrum prediction. The experimental results show that compared with traditional feedforward neural networks and feed-back neural networks, the performance of LLMs is better than that of traditional feedforward neural networks and feed-back neural networks. Especially, the optimal clipping extreme learning machine has better performance in both prediction accuracy and prediction speed, so it is more suitable for wireless spectrum prediction.
【学位授予单位】：兰州大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN925

【参考文献】