基于并行计算的脑电信号分析方法研究

发布时间：2018-04-17 14:50

本文选题：脑电信号 + 高性能并行计算　；参考：《燕山大学》2012年硕士论文

【摘要】：人脑是复杂的非线性系统，脑电信号的研究是当今生命科学的重要前沿领域之一。脑电信号处理对于脑部相关疾病的检测、诊断和治疗至关重要，然而脑电信号的研究涉及到大量的脑电信号数据的采集和计算，大量神经数据的存储、管理和利用是一个巨大的挑战。尖端高性能计算技术应用于解决计算和数据密集型的神经问题仍处于起步阶段，本文在对不同脑电信号处理方法进行研究的基础上，将高性能并行计算技术引入到脑电信号的分析中，利用高性能并行计算技术中的基于中央处理器(Central Processing Unit，CPU)的线程池技术和基于图形处理器(GraphicProcessing Unit，GPU)的多线程技术对计算密集型脑电信号处理方法进行并行研究，开发了并行程序，提高了原有算法的计算精度和研究尺度，将并行设计后的算法用于脑电信号处理，辅助脑电信号的研究和分析，减少研究人员的工作量，缩短研究周期。首先，对并行计算的设计方法及目的进行了研究，选取适合本课题的并行解决方案和技术平台：基于CPU多线程的线程池技术和基于GPU的CUDA(Compute UnifiedDevice Architecture，统一设备架构)编程模型。分别对这两种技术进行原理和性能两方面的研究，开发相应程序。其次，对单通道脑电信号处理方法总体经验模态分解算法(Ensemble EmpiriealMode Decomposition，EEMD)进行并行化研究。EEMD非常适合用于非平稳、非线性的脑电信号分析，然而EEMD属于密集型计算，算法中存在大量并行的成份，本文从不同层次对该算法进行并行化研究，应用CPU多线程中的线程池技术和GPU多线程的CUDA并行技术分别对算法进行并行设计，相比于并行之前的算法，并行后算法的执行效率有了很大的提高，本文同时结合希尔伯特黄变换(Hilbert-HuangTransform，HHT)方法对癫痫信号进行分析，提取出更多物理特性，为癫痫的预测和诊断提供了依据。最后，并行研究由单通道脑电信号分析方法扩展到双通道、多通道脑电信号分析方法，对双通道脑电信号分析方法中的非线性相关性(Non-linear Interdependency，NLI)算法进行了并行化设计。NLI算法是一种非对称测量方法，用于判断两个序列间的耦合方向，可以得到两个序列的驱动和响应关系，但是NLI算法应用起来计算量大，不仅耗费时间而且难以向多通道脑电信号分析扩展，，本文主要应用基于GPU的CUDA技术，挖掘NLI算法的并行化成分，通过对多通道脑电信号进行NLI并行化设计，加速了算法并扩大了算法的应用尺度，使并行化NLI算法可以应用于多通道脑电信号的耦合方向分析中，并结合S估计器的方法，对多通道癫痫脑电信号进行了同步强度的分析，该方法为癫痫的发作和传播机制的研究提供了重要信息。
[Abstract]:Human brain is a complex nonlinear system, and the research of EEG is one of the important frontier fields of life science.EEG signal processing is very important for the detection, diagnosis and treatment of brain related diseases. However, the research of EEG signal involves the collection and calculation of a large amount of EEG data and the storage of a large amount of neural data.Management and utilization is a huge challenge.The application of cutting-edge high-performance computing technology in solving computational and data-intensive neural problems is still in its infancy.The high performance parallel computing technology is introduced into the analysis of EEG signals.Using the thread pool technology based on Central Processing Unit (CPU) and the multithread technology based on GraphicProcessing Unit (GPU) in high performance parallel computing technology, the computational intensive EEG signal processing method is studied in parallel, and a parallel program is developed.The calculation precision and research scale of the original algorithm are improved. The parallel designed algorithm is used in EEG signal processing to assist the research and analysis of EEG signal, to reduce the workload of researchers and shorten the research period.Firstly, the design method and purpose of parallel computing are studied, and the programming model of parallel solution and technology platform: thread pool technology based on CPU multithreading and CUDA(Compute UnifiedDevice architecture based on GPU are selected.The principle and performance of the two technologies are studied, and the corresponding programs are developed.Secondly, the parallelization of the general empirical mode decomposition algorithm Ensemble EmpiriealMode Decomposition.EEMD is very suitable for the analysis of non-stationary and nonlinear EEG signals. However, EEMD is a kind of intensive computation.There are a lot of parallel components in the algorithm. This paper studies the parallelization of the algorithm from different levels, applies the thread pool technology of CPU multithreading and the CUDA parallel technology of GPU multithreading to design the parallel algorithm respectively.Compared with the parallel algorithm, the efficiency of the parallel algorithm has been greatly improved. At the same time, we use Hilbert-Huang transform HHT method to analyze the epileptic signal and extract more physical properties.It provides a basis for the prediction and diagnosis of epilepsy.Finally, the parallel research is extended from single-channel EEG analysis to two-channel, multi-channel EEG analysis.In this paper, the parallel design of nonlinear dependence NLI (nonlinear dependence NLI) algorithm, which is used to determine the coupling direction between two sequences, is presented.The driving and response relationships of the two sequences can be obtained, but the application of NLI algorithm is time-consuming and difficult to extend to the multichannel EEG analysis. This paper mainly applies the CUDA technology based on GPU.The parallelization component of NLI algorithm is mined, and the NLI parallelization design of multichannel EEG signal is carried out, which accelerates the algorithm and expands the application scale of the algorithm. The parallelized NLI algorithm can be applied to the coupling direction analysis of multichannel EEG signals.The synchronous intensity of multi-channel epileptic EEG signal is analyzed by using S-estimator, which provides important information for the study of seizure and transmission mechanism.
【学位授予单位】：燕山大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TN911.6;R318.0

【相似文献】