多传感器人体运动类型识别集成学习方法研究

论文总字数：36841字

摘 要

现代社会快速的生活节奏和巨大的工作压力，严重影响着个人的身体健康，伴随着人们健康观念的普及、医疗资源不足、成本上涨以及人口老龄化等趋势，运动健身已从少数人的爱好逐步转变成大众追求健康的生活习惯和消费趋势。然而，由于缺乏有效的指导和反馈，运动健身往往未能达到预期效果，时常出现运动不足或过量、运动疲劳和伤害等问题。

为了减少这些问题，本文根据模式识别和信息融合的相关理论，用基于集成学习的分类模型，应用于人体运动类型的识别，从而让模型对人体运动类型有更好的理解与识别。本文受到Boosting算法和Bagging算法的启发，提出了基于多传感器融合的集成学习算法和基于多分类算法融合的集成学习算法，利用准确率或查准率来确定基分类器对不同运动类型的识别权重，设置阈值，剔除识别权重小于阈值的基分类器，提高集成学习算法的泛化能力，降低集成学习算法的偶然误差。

最后，选择 PAMAP数据库的相关数据作为运动类型识别的原始数据，根据数据特点，设计相应的特征提取与特征选择方式，利用weka软件和Java编写的上位机软件计算识别率，并通过与单一分类算法识别率比较、不同加权方式选择的比较、十倍交叉验证、两两相互验证、多人相互验证和稳定性分析等多种方式来验证本文所提出的两种集成学习算法的合理性，实验结果表明这两种算法具有良好的识别性能，可以有效地提高人体运动类型识别率，主要体现在以下四点：

1）所设计的集成学习算法比单一算法具有更高的识别率，能有效地降低识别的偶然误差；

2）在所设计的集成学习算法中，准确率加权略优于查准率加权；

3）基于多传感器融合的集成学习算法对同一个样本集和不同的样本集都有比较好的分类效果；

4）基于多分类算法融合的集成学习算法对同一个样本集有比较好的分类效果，而对不同的样本集识别效果比较差。

关键词：模式识别，信息融合，集成学习，识别权重，算法验证

Ensemble learning research on multi-sensor human movement type recognition

Abstract

In modern society, fast pace of life and the tremendous work pressure seriously affect the personal health. Along with the popularity of people’s health concept, lack of healthcare resources, rising costs and aging population, sports and fitness gradually transform into the living habits and consumption trends of public healthy pursuit from a few people’s hobby. However, due to the lack of effective guidance and feedback, sports and fitness frequently fail to achieve the desired results, and often appears inadequate or excessive exercise, sports fatigue, injuries and other problems.

In order to reduce these problems, this paper applies ensemble learning classification model based on pattern recognition and information fusion into the recognition of human movement type, which gives the model a better understanding and recognition performance for human movement type. With the inspiration of Boosting and Bagging, this paper proposes two algorithms: an ensemble learning algorithm based on multi-sensor fusion and an ensemble learning algorithm based on multi-classifier fusion. Both of them use accuracy rate or precision rate to determine the recognition weight of every base classifier towards different movement types. Then the two algorithms will get rid of those base classifiers whose recognition weights lower than the set threshold. In conclusion, the two algorithms can enhance generalization ability and lower accidental error.

Finally, this paper selects the PAMAP database data as raw data to recognize human movement type, designs the appropriate feature extraction and feature selection method according to the characteristics of the data, uses weka software and PC software written in Java to calculate the recognition rate, and verifies the reasonableness of the proposed two ensemble learning algorithms by a variety of verification methods, such as comparison of the single classification algorithm, comparison of different weighted ways, tenfold cross verification, pairwise mutual verification, stability analysis and so on. Experimental results show that the two ensemble learning algorithms have good recognition performance and can effectively improve the recognition rate of human movement type. Mainly reflect in the following four points:

1) Designed ensemble learning algorithms have higher recognition rate than the single algorithm, and can effectively reduce recognized accidental errors;

2) In designed ensemble learning algorithms, accuracy rate is better than precision rate for weight;

3) For the same sample set and different sample sets, the ensemble learning algorithm based on multi-sensor fusion has a better classification result;

4) The ensemble learning algorithm based on multi-classifier fusion has a better classification result for the same sample set, but for different sample sets, recognition effect is relatively poor.

KEY WORDS: pattern recognition, information fusion, ensemble learning, recognition weights, algorithm verification

目录

摘要 I

Abstract II

目录 III

第一章 绪论 1

1.1 引言 1

1.2 模式识别 1

1.3 信息融合 2

1.4 集成学习 2

1.5 本文的研究目的和主要研究内容 3

1.5.1 研究目的 3

1.5.2 研究内容 3

第二章 分类算法简介 4

2.1 贝叶斯分类 4

2.1.1 贝叶斯概率理论 4

2.1.2 贝叶斯分类模型 4

2.2 决策树 5

2.2.1 决策树生成 5

2.2.2 测试属性选择 5

2.2.3 决策树剪枝 6

2.3 K最近邻 6

2.3.1 基本思想 6

2.3.2 距离衡量 6

2.4 支持向量机 6

2.4.1 模型建立 6

2.4.2 核函数 7

2.5 人工神经网络 7

2.5.1 网络结构 7

2.5.2 算法流程 8

2.6 算法评价 8

2.6.1 贝叶斯分类 9

2.6.2 决策树 9

2.6.3 K最近邻 9

2.6.4 支持向量机 9

2.6.5 人工神经网络 9

第三章 集成学习算法简介 10

3.1 Boosting 10

3.1.1 AdaBoost 10

3.1.2 误差分析 11

3.2 Bagging 11

3.2.1 算法描述 12

3.2.2 误差分析 12

3.3 Boosting和Bagging区别 13

第四章 集成学习算法设计 14

4.1 基于多传感器融合的集成学习算法 14

4.1.1 输入 14

4.1.2 训练阶段 14

4.1.3 测试阶段 15

4.1.4 输出 15

4.2 基于多分类算法融合的集成学习算法 15

4.2.1 输入 16

4.2.2 训练阶段 16

4.2.3 测试阶段 16

4.2.4 输出 16

4.3 算法比较 16

4.3.1 相同点 17

4.3.2 不同点 17

第五章 软件设计 18

5.1 weka软件介绍 18

5.1.1 软件界面 18

5.1.2 数据格式 19

5.2 Java上位机软件编写 19

5.2.1 基于多传感器融合的集成学习算法界面 19

剩余内容已隐藏，请支付后下载全文，论文总字数：36841字