摘要
时间序列分类是流式数据事件分析和数据挖掘的基础.针对现有方法损失时间属性、分类准确率低、效率低等问题,提出基于Gram矩阵的T-CNN时间序列分类方法.该方法对时间序列进行小波阈值去噪,过滤正态曲线噪声,提出基于Gram矩阵的无损时间域图像转换方法,保留事件全部信息.改进时间序列CNN分类方法,在卷积层计算引入Toeplitz卷积核矩阵,实现矩阵乘积替换卷积运算.引入Triplet网络思想,构建T-CNN分类模型,通过计算同类事件与不同类事件间的相似度优化CNN的平方损失函数,提高T-CNN模型梯度下降的收敛速率及分类准确性.实验表明,相比现有方法,T-CNN时间序列分类方法能够提高35%的分类准确率、35%的分类精确率及40%的分类效率.
Time series classification is the basis of streaming data event analysis and data mining. A T-CNN time series classification method based on Gram matrix was proposed, aiming at the problems of loss of time attribute,low classification accuracy and low efficiency of the existing methods. Specifically, the time series was denoised by wavelet threshold to filter out normal curve noise, and a lossless transformation method based on the Gram matrix was proposed to convert the time series into time-domain images and retain all event information. Then the CNN classification method of time series was improved, and the Toeplitz convolution kernel matrix was introduced into the convolutional layer calculation to realize the replacement of convolution operation with matrix product. The Triplet network was introduced to construct the T-CNN classification model, and the square loss function of CNN was optimized by calculating the similarities between similar events and different kinds of events, so as to improve the convergence rate of gradient descent and the classification accuracy of the T-CNN model. Experimental results show that compared with the existing methods, the proposed T-CNN time series classification method can improve the classification accuracy by 35%, the classification precision by 35% and the classification efficiency by 40%.
作者
王俊陆
李素
纪婉婷
姜天
宋宝燕
WANG Jun-lu;LI Su;JI Wan-ting;JIANG Tian;SONG Bao-yan(School of Information,Liaoning University,Shenyang 110036,China)
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2023年第2期267-276,共10页
Journal of Zhejiang University：Engineering Science
基金
国家重点研发计划资助项目。
