基于深度置信网络的时间序列预测方法及其应用研究

发布时间：2018-02-24 12:18

本文关键词： 时间序列预测深度置信网络融合预测轴承锂电池　出处：《哈尔滨工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着物联网和大数据处理技术的快速发展,时间序列数据的获取变得更加方便快捷,与此同时时间序列数据逐渐呈现出大数据量、高复杂度和高非线性的特点,因此传统的时间序列预测方法已经逐渐无法满足当前时代下的预测要求。为了从大量时间序列数据中快速的提取出有用信息,研究面向复杂时间序列的新预测技术变得愈加重要。深度置信网络(Deep Belief Network,DBN)作为深度学习的主要算法之一,以其强大的特征提取和函数表征能力以及在处理高复杂度的非线性数据等方面的优势,已经在图像分类、语音识别和故障诊断等领域得到了全面的推广,但是在时间序列领域,尤其是预测中的研究及应用还远远不够。因此本文以解决目前时间序列预测领域存在的问题为出发点,例如针对复杂问题难以建立准确物理模型以及时间序列函数关系难以表征等问题,对基于DBN的数据驱动预测技术展开了深入的研究。本文首先在分析DBN基本原理的基础上,提出了基于DBN的时间序列预测模型,并对模型的框架及流程进行了详细的说明。同时为了探究DBN网络参数对预测结果的影响,分别从输入节点数、网络层数和隐层节点数这三个角度进行了研究,并利用标准数据集对DBN与传统时间预测算法进行了性能对比。其次为了进一步剖析基础DBN算法在实际应用场合下的预测性能,研究中以预测与健康管理(Prognostics and Health Management,PHM)2012挑战赛(轴承健康状态预测)为背景,分别从单步预测、多步预测和剩余寿命预测三个层面对DBN预测性能进行深入的研究。最后针对基础DBN算法长期预测精度欠佳以及缺乏不确定度表达的问题,将相关向量机(Relevance Vector Machine,RVM)作为DBN模型的预测层,提出了基于DBN和RVM融合的时间序列预测方法,并将该融合算法应用到马里兰大学锂电池剩余寿命预测中,从而对融合算法的性能进行验证及分析。研究结果表明,与传统时间序列预测方法相比,基础DBN更适合预测高维度高复杂度的数据,并且具备优异的短期预测性能,但长期预测性能一般且预测结果无法给出不确定度表达。本文提出的基于DBN和RVM的融合算法不仅提高了其长期预测性能,而且使预测结果具备了不确定度表达能力,使其在实际应用中更具有科学价值。
[Abstract]:With the rapid development of Internet of things and big data processing technology, the acquisition of time series data becomes more convenient and faster. At the same time, time series data gradually show the characteristics of large amount of data, high complexity and high nonlinearity. Therefore, the traditional time series prediction method has been unable to meet the prediction requirements in the present era. In order to extract useful information quickly from a large number of time series data, It is becoming more and more important to study new prediction techniques for complex time series. Deep Belief Network (DBN) is one of the main algorithms for depth learning. Because of its strong ability of feature extraction and function representation, and its advantages in dealing with nonlinear data with high complexity, it has been widely used in image classification, speech recognition and fault diagnosis, but in the field of time series. Especially, the research and application of prediction is far from enough. Therefore, this paper takes the existing problems in the field of time series prediction as the starting point. For example, it is difficult to establish accurate physical model for complex problems and the relationship of time series functions is difficult to be represented. The data-driven prediction technology based on DBN is deeply studied in this paper. Firstly, based on the analysis of the basic principles of DBN, A time series prediction model based on DBN is proposed, and the framework and flow of the model are explained in detail. In order to explore the influence of DBN network parameters on the prediction results, the number of input nodes is analyzed respectively. The network layer number and hidden layer node number are studied, and the performance of DBN is compared with the traditional time prediction algorithm by using standard data set. Secondly, the prediction performance of the basic DBN algorithm in practical applications is analyzed. In this study, the prediction and health management and Health management challenge (bearing health state prediction) was used to predict the health status of the bearing in a single step. The performance of DBN prediction is studied in three aspects: multistep prediction and residual life prediction. Finally, the long term prediction accuracy of the basic DBN algorithm and the lack of uncertainty representation are discussed. The correlation vector machine (Relevance Vector Machine) is used as the prediction layer of the DBN model. A time series prediction method based on the fusion of DBN and RVM is proposed, and the fusion algorithm is applied to predict the residual life of the lithium battery at the University of Maryland. As a result, the performance of the fusion algorithm is verified and analyzed. The results show that compared with the traditional time series prediction method, the basic DBN is more suitable for predicting high-dimensional and high-complexity data, and has excellent short-term prediction performance. However, the long-term prediction performance is general and the prediction results can not be expressed with uncertainty. The fusion algorithm based on DBN and RVM not only improves the long-term prediction performance, but also makes the prediction results have the ability to express uncertainty. It has more scientific value in practical application.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：O211.61

【相似文献】