2023NO.2探索与争鸣SCIENCE&TECHNOLOGYINFORMATION科技资讯科技资讯SCIENCE&TECHNOLOGYINFORMATION基于MLP神经网络的优化跌倒检测模型研究与实现王鹏宇1*邵卓雅2马传辉2王志恒2汤佳杰2(1.南通大学体育科学学院;2.南通大学信息科学技术学院江苏南通226019)摘要：该文以多层感知器（MultilayerPerceptron，MLP）神经网络模型为基础，提出了一种高效快速的跌倒检测多层感知模型（FallDetectionMultilayerPerceptron，FDMLP）。该模型基于公开的姿态数据集进行特征强化和标签分类，得到特征值数据集进行训练；通过学习率范围测试确定初始学习率的上下边界值，设置学习率随迭代数次进行指数衰减优化训练过程采用全整数量化的优化策略，将量化后的模型部署到嵌入式设备的Flash中，实现低功耗、高准确率的边缘计算。实验结果显示该文提出的FDMLP神经网络跌倒检测模型在使用特征数据集时候的准确率达99.99%，优于其他同类模型，且结构简单，适合部署在边缘设备上。关键词：人工智能边缘计算跌倒检测多层感知机嵌入式设备中图分类号：TP391文献标识码：A文章编号：1672-3791(2023)02-0248-04ResearchandImplementationofanOptimizedFallDetectionModelBasedonMLPNeuralNetworkWANGPengyu1*SHAOZhuoya2MAChuanhui2WANGZhiheng2TANGJiajie2(1.SchoolofSportsScience,NantongUniversity;2.SchoolofInformationScienceandTechnology,NantongUniversity,Nantong,JiangsuProvince,226019China)Abstract:ThispaperproposedanefficientandfastFDMLP(FallDetectionMultilayerPerceptron)basedonMLP(MultilayerPerceptron).Themodelisbasedonthepublicattitudedatasetforfeatureenhancementandlabelclassi‐fication,andtheeigenvaluedatasetisobtainedfortraining;Theupperandlowerboundaryvaluesoftheinitiallearningratearedeterminedthroughthelearningraterangetest,andthelearningrateissettoexponentiallydecaywithiterations.Theoptimizationstrategyoffullintegerquantizationisusedinthetrainingprocess.ThequantizedmodelisdeployedtotheFlashoftheembeddeddevicetoachieveedgecomputingwithlowpowerconsumptionandhighaccuracy.TheexperimentalresultsshowthattheFDMLPneuralnetworkfalldetectionmodelproposedinthispaperhasanaccuracyrateof99.99%whenusingthefeaturedataset,whichissuperiortoothersimilarmodels,andhasasimplestructure,andissuitable...
