| 摘要: |
| 时间序列数据广泛存在于我们的生活中, 吸引了越来越多的学者对其进行深入的研究. 时间序列分类是时间序列的一个重要研究领域, 目前已有上百种分类算法被提出. 这些方法大致分为基于距离的方法、基于特征的方法以及基于深度学习的方法. 前两类方法需要手动处理特征和人为选择分类器, 而大多数的深度学习方法属于端到端的方法, 并且在时间序列分类问题中表现出不错的分类效果. 但是, 目前基于深度学习的方法很少能够针对时间序列数据中时间尺度选择的问题对网络进行改进, 在网络结构方面, 很少将网络进行融合, 从而更好地发挥各自的优势.为了解决这两类问题, 提出一种多尺度残差全卷积神经网络(MRes-FCN)结构, 用来处理时间序列问题. 该结构主要分为数据预处理阶段、全卷积网络与残差网络结合的阶段. 为了评价该方法的性能, 在UCR的85个公开数据集上进行了实验, 与基于距离的方法、基于特征的方法和基于深度学习的方法分别进行了比较. 实验结果表明: 所提出的方法较其他方法而言具有很好的性能, 在多个数据集上优于大多数方法. |
| 关键词: 时间序列 时间序列分类 残差网络 全卷积神经网络 深度学习 |
| DOI：10.13328/j.cnki.jos.006142 |
| 分类号:TP18 |
| 基金项目:北京市自然科学基金(4182052); 国家自然科学基金(61672086, 61702030, 61771058) |
|
| Time Series Classification Algorithm Based on Multiscale Residual Full Convolutional Neural Network |
|
ZHANG Ya-Wen1,2, WANG Zhi-Hai1,2, LIU Hai-Yang1,2, ZENG Zhao-Bo1,2
|
|
1.School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China;2.Beijing Key Laboratory of Traffic Data Analysis and Mining (Beijing Jiaotong University), Beijing 100044, China
|
| Abstract: |
| Time series data widely exists in daily lives, attracting more and more scholars to conduct in-depth research on it. Time series classification is an important research field of time series, and hundreds of classification algorithms have been proposed. These methods are roughly divided into distance-based methods, feature-based methods, and deep learning-based methods. The first two types of methods require manual processing of features and artificial selection of classifiers, and most deep learning-based methods are end-to-end methods and show good classification results in time series classification problems. Nevertheless, the current deep learning-based methods are rarely able to improve the network for the problem of time scale selection in time series data, and rarely integrate the network in terms of network structure to better leverage their respective advantages. In order to solve these two kinds of problems, this study proposes a multi-scale residual full convolutional neural network (MRes-FCN) structure to deal with time series problems. The structure is mainly divided into the data preprocessing stage, the stage of combining the full convolutional network and the residual network. In order to evaluate the performance of this method, this study conducted experiments on 85 public data sets of UCR and compared them with distance-based methods, feature-based methods, and deep learning-based methods. Experiments show that the proposed method has better performance than other methods, and it is better than most methods on multiple data sets. |
| Key words: time series time series classification residual network full convolutional neural network deep learning |
