【作者】 李宁；

【导师】 王友清；

【作者基本信息】 北京化工大学 ， 控制科学与工程， 2019， 硕士

【摘要】 随着生活质量的提升,糖尿病发病率也呈现出增长的趋势,如何治愈糖尿病已经成为重大全球性健康问题。目前,通过人工胰脏自动向患者体内注射胰岛素被认为是最有效管理血糖的手段之一。准确的血糖预测不仅能够为设计人工胰脏的控制算法提供必不可少的信息,还可以向患者或医生发出高低血糖预警,以便采取预前调节预防其发生。本文首先研究了血糖时间序列的混沌性:利用Wolf算法和小数据量法对血糖时间序列进行混沌性判别,由得到的最大Lyapunov指数大于零,表明血糖时间序列具有混沌特性。而后,在Takens理论的基础上,分别采用自相关函数(Autocorrelation Function,ACF)和虚假最近邻点法(False Nearest Neighbors,FNN)计算最佳嵌入维数和延迟时间并进行相空间重构。在相空间重构的基础上,利用回声状态网络(Echo State Networks,ESN)的快速非线性逼近能力建立起血糖混沌预报模型。考虑到ESN预测性能与中间层的选择有着密切的关系,本文提出了一种增量学习策略对最佳储蓄池规模进行在线寻优。为了挖掘前后序列数据的信息关联,增强模型的预测性,设计F-ESN结构将预测输出值实时反馈到输入端。最后,利用真实患者数据进行算法验证,结果表明本文所提的F-I-ESN具有优越的性能表现。将F-ESN预报模型与动态风险算法(Dynamic Risk,DR)相结合,提出了改进后的算法DR-FESN并应用到低血糖预警中。DR是利用当前血糖值和血糖的变化趋势做出低血糖风险估算的一种策略。验证实验数据源于12个有明显低血糖发生的虚拟病人,仿真结果显示DR-FESN在不损失低血糖预测敏感性和特异性的基础上,能够很大程度上提前发出预警信号,为病人和医生赢得宝贵的时间。更多还原

【Abstract】 With the rapid improvement of living quality,the morbidity of diabetes trends to increasing.It is major global health problem that how to cure the diabetes.Nowadays,continuous insulin injection through the artificial pancreas(AP)is considered as one of the most effective means to treat diabetes.Accurate glucose prediction not only can provide indispensable information for designing AP,but also can warn the upcoming the hyper/hypoglycemia events for patients or doctors.Frist,the chaos characteristic in glucose time series has been identified using the chaos theory,then the chaos characteristic is proved by using the Wolf algorithm and small-data method respectively.The obtained results of Lyapunov demonstrate that the glucose time series have chaotic.Next,based on the Takens theorem,the two parameters,named optimal embedded demission and delay time,can be obtained with the Autocorrelation Function and False Nearest Neighbors for reconstructing the phase spaceWhen the reconstruction completed,a fast-nonlinear dynamic network is conducted with intention of modeling the more precise prediction function.Considering the crucial relationship between the prediction performance and the lay of middle,the paper presents a novel strategy combing the increasing learning with ESN to optimize the optimal middle nodes online.Besides,in order to stress the connection data and improve the model predictability,this article designs output feedback structure,refer to F-ESN,linking the prediction output and input.Ultimately,the experiments have been enforced with the real patients’ data,which indicates the proposed method(F-I-ESN)is superior than others algorithms.A novel method combined the F-ESN with(Dynamic Risk,DR),named as DR-FESN,is proposed for predicting the hypoglycemia occurrence.DR is capability of evaluating the hypoglycemia risk by the fluctuation information of glucose and its variation tendency.The simulations are conducted with 12 virtual patients who have significant hypoglycemia symptom,which show that DR-FESN can greatly advance the early warning time without losing the sensitivity and specificity of hypoglycemia events.That will gain valuable cure time for patients and doctors.更多还原