本文为土耳其中东技术大学（作者：POYRAZ UMUT HATIPOGLU）的硕士论文，共139页。

由于需要以更复杂和抽象的方式来表示统计数据，深度学习是一个快速增长和有趣的领域。处理器和图形处理单元技术的发展不可否认地影响了深度网络的普及。

本文的主要目的是研究一种鲁棒的全功能时间序列分类方法。为此，提出了一种基于深度学习的新方法。由于时间序列数据具有复杂多变的结构，因此采用能够处理非线性复杂运算的算法比采用浅层结构的方法更为合适。虽然浅层结构的方法需要手工制作的特征和有关数据的专家知识，但基于深度学习的算法能够处理原始特征。针对来自不同研究领域的数据集，构建并训练了基于深度信念网络和堆叠式自动编码器的体系结构。在时间序列分类中，尽管动态时间规整和基于最近邻的分类方法很难被击败，但最近又提出了了很多的分类方法。为了检验所提出方法的性能，与流行的基准方法进行了比较分析。尽管分类结果的准确性更高，但基于深度学习的方法并不比其他方法高明。

Deep learning is a fast-growing and interesting field due to the need to represent statistical data in a more complex and abstract way. Development in the processors and graphics processing unit technology effects undeniably that the deep networks get that popularity. The main purpose of this work is to develop robust and full functional time series classification method. To achieve this intent a deep learning based novel methods are proposed. Because time series data can have complex and variable structure, it may be more suitable to use algorithms that can handle the nonlinear sophisticated operations rather than shallow-structured methods. While shallow structured methods need handcrafted features and expert knowledge about data, deep learning based algorithms are capable of working with raw features. Both deep belief network and stacked autoencoders based architectures are constructed and trained for the dataset gathered from different researches areas. In time series classification, even though dynamic time warping and nearest neighbor based methods are hard to beat, many classification methods have been studied recently. To examine the performance of proposed method comparative analysis is conducted with popular benchmark methods. Despite higher accuracy in the results, the deep learning based methods cannot outperform superiorly.

1.       引言

2. 时间序列与学习

3. 时间序列分类中的基准方法

4. 神经网络、深度学习与框架

5. 实验与结果

6. 结论与展望

附录A 混淆矩阵及其详细性能

附录B 训练集误差和验证集误差图

更多精彩文章请关注公众号：