作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (4): 81-88,98. doi: 10.19678/j.issn.1000-3428.0061110

• 人工智能与模式识别 • 上一篇    下一篇

基于自适应小波分解的时间序列分类方法

梁小慧, 郭晟楠, 万怀宇   

  1. 北京交通大学 计算机与信息技术学院 交通数据分析与挖掘北京市重点实验室, 北京 100044
  • 收稿日期:2021-03-12 修回日期:2021-05-16 发布日期:2021-05-18
  • 作者简介:梁小慧(1995—),女,硕士研究生,主研方向为时空数据挖掘、深度学习;郭晟楠,博士研究生;万怀宇(通信作者),副教授、博士。
  • 基金资助:
    教育部-中国移动科研基金(MCM20180202)。

Time Series Classification Method Based on Adaptive Wavelet Decomposition

LIANG Xiaohui, GUO Shengnan, WAN Huaiyu   

  1. Beijing Key Laboratory of Traffic Data Analysis and Mining, School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China
  • Received:2021-03-12 Revised:2021-05-16 Published:2021-05-18

摘要: 时间序列分类即通过构建分类模型建模时间序列中的特征来实现对该时间序列的归类,是时间序列挖掘的重要研究分支。现有的时间序列分类方法多数从时域的角度对时间序列进行建模,忽视了时间序列中隐含的频域信息,而时间序列往往同时蕴含着多种不同变化速率的变化模式,这些变化模式在时域上相互叠加,使得时间序列的变化规律变得比较复杂,因此仅从时域的角度进行建模,难以有效地从复杂的规律中捕获其蕴含的多种相对简单的规律。提出一种基于自适应多级小波分解的神经网络方法AMWDNet,使用自适应小波分解建模时间序列中的多级时频信息,自适应小波分解模块能够同时从时域和频域的角度出发,对时间序列中蕴含的多种变化模式进行有效分解,通过使用长短期时间模式提取模块分别建模时间序列中的长期和短期时间模式。选取时间序列分类任务中8个主流的方法作为基准方法,在UCR数据集仓库中的8个数据集上进行对比实验,结果表明,AMWDNet在其中的7个数据集上取得了最高的分类准确率,相比于次优的基准方法提升了0.1~2.2个百分点,整体分类性能优于MLP和FCN等基准方法。

关键词: 时间序列分类, 自适应, 小波分解, 多分辨率分析, 时频分析

Abstract: Time series classification is an important research branch of time series mining, which involves constructing a classification model and modeling the characteristics of the time series.However, most of the existing classification methods only model the time series from the time domain perspective, ignoring its important frequency information.However, a time series usually contains various changing patterns with different changing speeds, which are combined in the time domain.It is difficult to capture the hidden patterns under the combined and complex patterns in the time domain.Considering these issues, this study proposes an Adaptive Multilevel Wavelet Decomposition-based neural Network (AMWDNet) to model the time-frequency information.First, it employs adaptive multilevel wavelet decomposition to extract multilevel time-frequency information, which can decompose a time series into several components from the perspective of both time and frequency domains.Then, it designs a long-term temporal pattern extraction module and a short-term temporal pattern extraction module to model the long-term and short-term temporal patterns, respectively.To evaluate the performance of AMWDNet, this study conducts experiments on eight real-world datasets from UCR Time Series Classification Archive.Based on the results, AMWDNet achieves the highest classification accuracy on seven datasets, exceeding that of the sub-optimal benchmark by 0.1~2.2 percentage point.Its overall classification performance surpasses those of benchmark methods such as MLP and FCN.

Key words: time series classification, adaptive, Wavelet decomposition, multi-resolution analysis, time-frequency analysis

中图分类号: