一种基于CNN模型多元时间序列分类结构
首发时间:2017-11-27
摘要:多元时间序列分类问题是时间序列挖掘领域中的重要问题,目前的常规做法是使用基于欧氏距离或 DTW 距离的K近邻分类模型,或基于统计模型的特征提取方法,但仍然难以避免计算效率低、引入过多先验假设的问题,不具有普适意义。本文提出一种包含多通道的卷积神经网络结构,对不同的变量单独处理提取时间角度的特征,再通过后续的前馈全连接网络处理多维度之间的关系性特征。整体模型采取简单的梯度下降的方法进行训练。通过模型在UCR 数据集上的实验,表明该结构比常规的 K 近邻模型具有更好的多分类能力,且在泛化程度和训练速度上都胜于前馈全连接神经网络模型。
关键词: 计算机应用技术 多元时间序列 时间序列挖掘 卷积神经网络
For information in English, please click here
An Multivariate Time Series Classification Model Based on Convolutional Neural Network
Abstract:The classification for multivariate time series is an important issue in the field of time series mining, currently the common waysare using K nearest neighbor classification model based on Euclidean distance or Dynamic Time Warping distance or using the feature extraction based on statistical models, but it is still hard to avoid low computational efficiency and too much priori assumptions, which does not have universal significance. In this paper a multi-channel convolutional neural network (MC-CNN) structure is proposed. In the structure the feature is processed separately in different variables, then the relational features between multi-dimensions are dealt with through the subsequent feedforward fully connected network.Gradient descent method can be used to train the whole model.Experiments on the UCR dataset show that the proposed model has better multi-classification ability than the K-nearest neighbor model, and outperforms the feedforward fully connected neural network model in terms of generalization and training speed.
Keywords: computer application technology multivariate time series time series mining CNN
基金:
论文图表:
引用
No.****
同行评议
勘误表
一种基于CNN模型多元时间序列分类结构
评论
全部评论0/1000