【作者】 梁波；

【导师】 王国兴； 钱智量；

【作者基本信息】 上海交通大学 ， 集成电路工程（专业学位）， 2017， 硕士

【摘要】 随着嵌入式系统的发展,越来越多的研究人员将支持向量机(Support Vector Machine,SVM)算法实现在嵌入式系统中。然而嵌入式系统在硬件资源、储存容量、能效等方面有着严格的限制,因此资源和功耗的优化成为SVM在嵌入式系统中实现时主要的难点和挑战。SVM算法包含训练和决策两个过程,本文主要面向支持向量机决策过程的硬件实现。对此本文设计实现了基于现场可编程门阵列(FPGA)的决策函数的运算平台。我们首先完成了SVM硬件决策平台的架构设计,使用SPI(Serial Peripheral Interface)与同步器模块保证平台与外部能够进行稳定的数据通信。接着我们详细介绍了平台中各个模块的设计。在SVM决策函数模块的设计中,为了减少资源消耗和降低功耗,我们进行了数据精度表示的优化并提出了指数函数可变精度计算的设计策略。我们使用Xilinx Artix7系列的NexysVedio FPGA开发平台完成了SVM决策平台的原型实现和验证。与优化之前的设计相比,在不影响计算精度的前提下平台能够降低13%的功耗,结合延时的优化,在计算单个测试向量能够减少50%的能量。此外,本文还基于数字信号处理器(DSP)设计实现了用于血压预测的SVM决策过程。我们首先从人体的光电容积脉搏波(PPG)中提取出两个特征量,之后将其应用到采用浮点数计算实现的SVM模型中完成血压预测,SVM模型则由离线训练得到。该系统对于5名志愿者血压预测误差的均值±标准差为:8.42±10.30 mmHg(Systolic Blood Pressure,SBP)和6.03±7.64 mmHg(Diastolic Blood Pressure,DBP)。更多还原

【Abstract】 With the development of embedded systems,more and more researchers have implemented the Support Vector Machine(SVM)based algorithm using embedded systems for portable and wearable applications.However,one challenge is that embedded system has critical restrictions in hardware resources,storage capacity and energy efficiency.Therefore,how to optimize the hardware resource consumption and energy efficiency becomes the main difficulty and challenge.SVM algorithm includes two processes,namely training and inference.In this thesis,we focused on the hardware implementation of SVM inference model.First,we implemented an SVM based inference platform based on FPGA.The system architecture of the platform is presented firstly.The SPI and synchronizer module in the platform are designed to implement the data communication between the platform and the peripheral system(e.g.,CPU).Then the design of each module in the platform is elaborated,respectively.In order to reduce resource and power consumption,data precision representations were optimized and a strategy of variable precision exponential function calculation was proposed in SVM inference function module.Using the NexysVedio evaluation board with a Xilinx Artix7 FPGA chip,a prototype of the SVM inference platform was implemented and the system performance was evaluated.Compared to the design without the optimizations mentioned before,the platform can reduce 13% power and 50% energy per test vector calculation,with only a little impact on the final calculation accuracy.We also implemented a system for SVM-based blood pressure prediction based on Digital Signal Processor(DSP),including feature extraction and SVM inference model.Two features are extracted firstly from photoplethysmography(PPG)signal.Based on the input features,the SVM inference function,which is implemented using floating-point algorithmic routines,can perform blood pressure prediction accordingly.We tested 5 volunteers using our proposed system,the MAEs(Mean Absolute Error)±STD(Standard Derivation)of predicting error for systolic blood pressure(SBP)and diastolic blood pressure(DBP)are 8.42±10.30 mmHg and 6.03±7.64 mmHg,respectively.更多还原