电子学报 ›› 2022, Vol. 50 ›› Issue (1): 177-184.DOI: 10.12263/DZXB.20201003

• 学术论文 • 上一篇    下一篇

面向帕金森病语音诊断的非监督两步式卷积稀疏迁移学习算法

张小恒1,2, 张馨月1, 李勇明1, 王品1, 刘玉川1   

  1. 1.重庆大学微电子与通信工程学院,重庆 400030
    2.重庆广播电视大学,重庆 400052
  • 收稿日期:2020-09-10 修回日期:2021-02-24 出版日期:2022-01-25 发布日期:2022-01-25
  • 作者简介:张小恒 男,1980年生,四川达州人.博士研究生,副教授.主要研究领域为医学信号处理、机器学习. E-mail:7818320@qq.com
    张馨月 女,1996年生,四川泸县人.硕士研究生.主要研究领域为医学信号处理. E-mail:1029323666@qq.com
    李勇明(通信作者) 男,1976年生,四川绵阳人.博士,教授、博士生导师.主要研究领域为医学信号处理、机器学习. E-mail:yongmingli@cqu.edu.cn
  • 基金资助:
    国家自然科学基金(61771080);重庆市自然科学基金(cstc2020jcyj-msxmX0100);重庆市社会科学规划项目(2018YBYY133)

An Unsupervised Two-Step Convolution Sparse Transfer Learning Algorithm for Parkinson’s Disease Speech Diagnosis

ZHANG Xiao-heng1,2, ZHANG Xin-yue1, LI Yong-ming1, WANG Pin1, LIU Yu-chuan1   

  1. 1.College of Communication Engineering,Chongqing University,Chongqing 400030,China
    2.Chongqing Radio & TV University,Chongqing 400052,China
  • Received:2020-09-10 Revised:2021-02-24 Online:2022-01-25 Published:2022-01-25

摘要:

帕金森病(Parkinson’s Disease,PD)语音诊断存在小样本问题,如果借助相关语音数据集进行迁移学习,容易加重训练集和测试集之间的分布差异,影响分类准确率.为了解决上述矛盾问题,本文提出了两步式稀疏迁移学习算法.该算法分为两大步:第一步算法为语音段特征同时优选的快速卷积稀疏编码算法,构造卷积稀疏编码算子用于快速学习公共语音数据集的结构信息,然后将其迁移到PD语音目标集以弥补后者样本信息的不足,接着再同时对语音段和特征进行同时优选以获得更有价值的信息;第二步算法为联合局部结构信息分布对齐算法,对训练集和测试集进行域适应,在保持各自样本结构信息的同时,最小化分布误差.实验结果表明:本文算法中每一步迁移学习算法均有效;与相关算法相比,本文算法准确率显著较高,达97.5%.

关键词: 语音诊断, 帕金森症(PD), 两步式稀疏迁移学习, 卷积稀疏迁移学习, 域适应

Abstract:

Parkinson's disease(PD) speech diagnosis has a small sample problem. Although it is possible to transfer learning with the help of relevant speech datasets. The introduction of other samples will lead to the distribution difference between samples of different subjects, so the classification accuracy is greatly affected. Therefore, in this paper, to solve the problems above, we propose a novel unsupervised two-step convolutional sparse transfer leaning algorithm. The algorithm is divided into two steps: fast convolutional sparse coding with coordinate selection of samples and features(FCSC&SF), joint local structure distribution alignment(JLSDA). In the FCSC&SF, speech structure among public speech dataset is quickly learned by fast convolution sparse coding(FCSC), and transferred into the target dataset, after that, the more valuable information is obtained by coordinate selection of samples and features. JLSDA is designed to maintain the local structure information in the two domains, and reduce the distribution difference between the two domains at the same time. The experimental results showed that each step of the proposed algorithm has a positive effect on the classification results; compared with the representative relevant algorithms, the accuracy of the proposed method is significantly higher at 97.5%.

Key words: speech diagnosis, parkinson’s disease(PD), two-step sparse transfer learning, convolutional sparse coding transfer learning, domain adaptation

中图分类号: