基于大数据与深度学习的生理信号分析

发布时间：2018-04-29 00:35

本文选题：Hadoop + 生理信号　；参考：《南京邮电大学》2017年硕士论文

【摘要】：随着信息化的快速发展,来自可穿戴设备、电子病历、便携式监护仪等医疗数据迅猛增长且存储结构多元化。传统的存储结构与计算模型不能够很好的解决这些数据的存储和计算分析的问题。幸运的是大数据量的增长能够很好的解决传统机器学习的方法中数据样本不足的问题,但是单机数据处理的计算能力达不到要求且需要专业的人员进行相关数据特征的人工提取。特征的提取过程麻烦且受到专家的主观因素影响将会导致分析的结果不准。为了解决上面的问题,本论文主要研究内容如下:(1)构建了一个生理大数据的集群分析平台,来解决数据大存储与计算能力不足的问题。该平台采用了Hadoop中的HDFS来解决非结构化数据的大规模存储问题。除此之外还采用了消息队列、流计算框架来提升平台的整体性能。(2)采用MapReduce计算框架解决了大数据集的计算分析问题。实现了基于MapReduce的BP神经网络的并行化,实验的结果表明神经网络并行化实现的可行性并且能够有效地提高分析的准确率、减少训练时间,加快研究速度及分析效率。(3)针对传统机器学习方法中需要人工的提取特征的不足,尝试采用了深度学习技术并将其应用到生理信号分析的方法中。目前国内基于深度学习的生理信号研究还很少,本文使用了深度学习中的DBN、CNN、SAE网络来获取抽象特征避免了人工干扰,同时结合了传统的分类器SVM及神经网络对相关生理特征进行分类。实验的结果表明本文提出的方法可以适用于生理信号的分类,并取得了不错的效果。
[Abstract]:With the rapid development of information technology, medical data from wearable devices, electronic medical records, portable monitors and other medical data are growing rapidly and the storage structure is diversified. The traditional storage structure and computing model can not solve the problem of data storage and analysis. Fortunately, the growth of large amounts of data can solve the problem of insufficient data samples in traditional machine learning methods. But the computing ability of single computer data processing is not up to the requirement and professional personnel are needed to carry out manual extraction of relevant data features. The feature extraction process is troublesome and influenced by the subjective factors of experts, which will lead to inaccurate analysis results. In order to solve the above problems, the main contents of this paper are as follows: 1) A cluster analysis platform of physiological big data is constructed to solve the problem of insufficient data storage and computing power. The platform uses HDFS in Hadoop to solve the problem of large-scale storage of unstructured data. In addition, the message queue and flow computing framework are used to improve the overall performance of the platform. (2) the MapReduce computing framework is used to solve the computing and analysis problem of big data set. The parallelization of BP neural network based on MapReduce is realized. The experimental results show that the parallelization of neural network is feasible and can effectively improve the accuracy of analysis and reduce the training time. To improve the research speed and analysis efficiency, aiming at the shortcomings of traditional machine learning methods, which need to extract features manually, this paper tries to adopt the deep learning technology and apply it to the analysis of physiological signals. At present, there are few researches on physiological signals based on deep learning in China. In this paper, the DBNNNSAE network is used to obtain abstract features to avoid artificial interference. At the same time, combining the traditional classifier SVM and neural network to classify the related physiological characteristics. The experimental results show that the proposed method can be applied to the classification of physiological signals, and good results have been obtained.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN911.6;TP18

【参考文献】