【作者】 吴玺；

【导师】 李廉； 安宁；

【作者基本信息】 合肥工业大学 ， 计算机应用技术， 2018， 博士

【摘要】 随着物联网和各类传感技术的发展,基于人工智能的人体活动感知识别技术正广泛地应用在健康医疗、智慧养老和智能安防等诸多领域。通过采集传感器数据进行人体活动感知识别从而实现日常健康监测与疾病预警正成为研究热点。而其中,帕金森病作为困扰诸多中老年人的神经退行性疾病,成为许多应用研究的靶点。当前的研究集中在帕金森病的诊断及康复训练,但相关的数据分析多是基于原始传感器数据,存在准确性差和效率低的问题。因此,本文面向帕金森病的诊断及运动功能康复两个方面,首先研究基于柔性阵列压力传感数据的帕金森病步态分析,用于帕金森病的诊断模型构建,然后研究基于三轴加速度计与陀螺仪传感数据的人体日常活动识别,用于帕金森病人运动功能康复过程中的日常活动监测与评估。本文主要的研究工作分为以下几个方面:(1)使用基于柔性阵列压力传感器的U-Gait步态感知系统,提取了帕金森病人组和无运动障碍疾病的正常对照组的足底压力数据。为验证足底压力数据对疾病诊断的有效性,本文对足底压力数据进行消除趋势波动分析(DFA)和消除趋势互相关分析(DCCA),实验结果表明正常对照组足底压力时间序列的自相关性和互相关性均高于帕金森病人组,相符于医学研究结果。在此基础上,本文对U-Gait步态感知系统采集的原始足底压力数据通过脚印还原和行走轨迹图生成、去噪、膨胀优化和分割等处理提取了 18个步态时空特征参数,并特别提取了受试者在U-Gait电子步道上两个弯道处的转弯步态参数来更好地反映帕金森病人转弯困难的步态特征。然后使用T-test方法验证了所提取的步态特征在帕金森病人组和正常对照组之间存在显著差异性,并使用ICC方法检验了步态特征数据的可靠性,为后续使用步态特征建立机器学习模型进行帕金森病识别诊断提供可行性和有效性的依据。(2)针对各种步态特征之间因为存在冗余或不相关性可能对帕金森病识别诊断模型的性能产生负面影响,本文提出了一种将Filter方法和Wrapper方法相融合的IWFS特征选择算法,该算法的核心思想是先用ReliefF方法将原始步态特征集按照相关性降序排列,然后使用Wrapper方法递增选择使得分类器准确率提高的特征。通过与K-近邻、朴素贝叶斯、支持向量机等经典分类器集成,验证了所提出的IWFS特征选择算法的有效性,以及所构建的基于步态特征选择的帕金森病识别诊断模型的良好分类性能,为实际临床医生对帕金森病的辅助诊断以及患者的康复治疗提供了有力依据。(3)针对帕金森病人运动功能康复过程中日常活动监测和评估的需求,本文通过三轴加速度计和陀螺仪传感数据实现了人体日常活动识别,从而帮助帕金森病人的运动功能恢复训练进行监测,并为康复治疗效果的评估提供客观依据。针对传统人体活动识别研究主要采用平面型分类器忽略了人体活动类别之间层次性的问题,本文提出了一种层次化人体活动识别模型,将多类活动识别问题转化为多层次的活动识别问题,分层多阶段地识别不同的活动子集以及每个子集中的相似活动。并且针对人体活动特征的研究主要集中在如何提取大量的时域和频域特征来更好地刻画原始传感器信号,而忽略了特征之间可能存在的冗余性甚至与活动识别任务不相关的问题,本文提出了两种不同的特征选择方法:基于类独立的方法和基于类依赖的方法。其中基于类依赖的特征选择方法为每个活动类别单独选择一个特征子集能够有效提高活动识别的准确率。实验结果表明,所提出的基于层次化结构和类依赖特征选择的人体活动识别方法与其他方法相比不仅提高了活动识别的准确率,而且具有良好的泛化能力。更多还原

【Abstract】 With the development of internet of things and various kinds of sensor technology,human activity perception recognition technology based on artificial intelligence is widely used in many fields,such as health care,intelligent pension and intelligent security.More and more applications have been made to realize daily health monitoring and disease early warning by collecting sensor data for human activity perception and recognition.Parkinson’s disease,as a neurodegenerative disease that troubles many middle-aged and elderly people,has become a target of many applied research.Current research focuses on the diagnosis of Parkinson’s disease and rehabilitation training.The relevant data analysis is mostly based on the original sensor data,and the accuracy is poor and the efficiency is low.First,we use U-Gait electronic walkway system to collect gait features of Parkinson’s disease patients and normal persons and study how to use these gait features for Parkinson’s disease detection.Then we use human activities recognition method based on wearable sensor data to monitor and evaluate daily activities of Parkinson’s patients during motor rehabilitation.The main research work of this dissertation includes the following aspects:(1)The original plantar pressure data of Parkinson’s disease patients and normal persons is extracted from U-Gait electronic walkway based on flexible array pressure sensor.Detrended fluctuation analysis(DFA)and detrended cross-correlation analysis(DCCA)are used to validate the validity of plantar pressure data for disease diagnosis.The experiment results show that the autocorrelation and cross-correlation of plantar pressure time series in normal persons are both higher than those in Parkinson’s disease patients,which is consistent with medical research results.Then,we design a method to calculate 18 gait spatio-temporal characteristics through footprint reduction,track generation,denoising,expansion optimization and segmentation processing.In order to reflect the gait characteristics of Parkinson’s patients with turning difficulty,the parameters of turning gait at the two corners of U-Gait are also extracted.We preprocess the gait data by eliminating individual height difference and normalization.In addition,the T-test result certifies the significant differences of the gait features between Parkinson’s disease patients and normal persons.And the reliability of gait features is verified by ICC method.The results provide a feasible and effective basis for the subsequent use of gait features to establish machine learning model for the identification of Parkinson’s disease.(2)Aiming at the negative effect of redundancy or irrelevance between gait features on the performance of Parkinson’s disease recognition model,this paper presents a new feature selection algorithm-IWFS,which combines Filter method with Wrapper method.First,IWFS algorithm uses ReliefF method to arrange the original features set according to the descending order of correlation.Then it uses Wrapper method to incrementally select the features which can improve the classification accuracy.The experimental results validate the good classification performance of this model,and provide a strong basis for the practical diagnosis of Parkinson’s disease.(3)To meet the needs of daily activities monitoring and evaluation in the rehabilitation process of Parkinson’s patients,this paper realizes the recognition of human daily activities through the data of triaxial accelerometer and gyroscope sensors.Due to the diversity and complexity of human activities,activity recognition methods often can not obtain good performance.Planar classifiers are mainly used in current research and only one classifier is constructed to identify all predefined activity sets.The hierarchy between human activity categories is ignored.And the research on feature using mainly focuses on extracting a large number of various time and frequency domain features.In fact,there may be some redundancy features and irrelevant features affecting the performance of the activity recognition model.We proposed a new human activity recognition model based on hierarchical structure and class-dependent feature selection.This model transforms the multi-class activity recognition problem into a multi-level activity recognition problem.And the feature selection method based on class-dependent selects an optimal feature subset for each activity class.The feature subsets of different classes may be different.Experimental results show that the proposed human activity recognition method has better recognition performance and generalization ability than other methods.更多还原