两类神经网络的CMOS模拟电路设计与研究

发布时间：2018-04-24 11:57

本文选题：神经网络 + 模拟电路　；参考：《湘潭大学》2015年硕士论文

【摘要】：神经网络是模拟人脑基本特性的智能系统,也是一门信息处理的科学。神经网络具有自适应学习、非线性映射、分布并行处理等特点。神经网络从单个神经元的模拟,到最终模拟大脑的信息处理功能。神经网络应用非常广泛,目前主要运用于非线性系统、网络故障、航空航天、智能机器人等领域。对于神经网络的研究主要分为三部分:理论研究、应用研究和实现技术研究。而实现技术上,主要有两种实现方法:软件实现和硬件实现。用软件实现神经网络,具有处理速度、并行程度低等缺点,这很难满足神经网络信息处理的实时性的要求。用硬件实现神经网络能体现网络的快速性、并行计算,且能实现大规模的信号处理,这在复杂的数据处理场合中是非常有利的。因此,硬件实现是神经网络发展的必然趋势。硬件实现方法中,基于模拟CMOS电路实现神经网络电路具有结构简单、集成速度快、占用芯片面积小、集成度高、功耗低等特点,因此本文研究采用模拟CMOS集成电路设计神经网络。神经网络模型中具有代表性的有:误差反向传播BP网络、径向基函数RBF网络、自组织网络、感知器、反馈Hopfield网络、小脑模型CMAC网络、模糊神经网络等。目前,已经用硬件实现了的神经网络有:BP网络、RBF网络、感知器等,而在其他的网络模型的硬件实现方案甚少。基于研究神经网络的全面性,本文主要研究用模拟CMOS电路实现自组织竞争神经和模糊神经网络,围绕这两种神经网络,做了如下相关工作:(1)针对神经网络的神经元模型中权值不可调的缺点,设计了线性可调运算跨导放大器和电流乘法器电路作为突触电路,通过改变外部电流实现权值可调功能,且设计的电路结构简单,线性度高。以此能作为基本单元应用于神经元电路中。(2)基于自组织竞争神经网络中竞争层算法难以实现的问题,设计了一种电流模式的最值电路模拟实现竞争算法,通过比较电流的大小达到竞争目的。该电路实现简单、模拟程度高、便于集成,与输入层结合能实现自组织竞争神经网络。(3)针对模糊神经网络的单元电路结构复杂且精度低的问题,本文对高斯函数电路、求小电路、去模糊电路的结构进行优化设计,从整体上提高模糊神经网络的精度和高速性。最后将设计的模糊神经网络用于实现一个非线性函数的逼近,并通过了仿真与验证。
[Abstract]:Neural network is an intelligent system which simulates the basic characteristics of human brain and is also a science of information processing. Neural network has the characteristics of adaptive learning, nonlinear mapping, distributed parallel processing and so on. Neural networks range from the simulation of a single neuron to the eventual simulation of the brain's information processing function. Neural network is widely used in nonlinear systems, network failures, aerospace, intelligent robots and other fields. The research of neural network is divided into three parts: theoretical research, application research and implementation technology research. On the other hand, there are two main methods: software implementation and hardware implementation. The realization of neural network by software has the disadvantages of low processing speed and low parallelism, which is difficult to meet the real-time requirement of neural network information processing. The implementation of neural network by hardware can reflect the rapidity of the network, parallel computing, and the realization of large-scale signal processing, which is very advantageous in the complex data processing situation. Therefore, hardware implementation is the inevitable trend of neural network development. In the hardware realization method, the neural network circuit based on analog CMOS circuit has the characteristics of simple structure, high integration speed, small chip area, high integration level and low power consumption. Therefore, this paper studies the use of analog CMOS integrated circuit to design neural networks. Some typical neural network models are error back-propagation BP network, radial basis function (RBF) network, self-organizing network, perceptron, feedback Hopfield network, cerebellar model CMAC network, fuzzy neural network and so on. At present, the neural networks that have been implemented by hardware include: BP network, RBF network, perceptron and so on, but there are few hardware implementation schemes in other network models. Based on the comprehensive study of neural networks, this paper mainly studies the implementation of self-organizing competitive neural networks and fuzzy neural networks using analog CMOS circuits, and focuses on these two kinds of neural networks. The following work is done: (1) aiming at the disadvantage of the unadjustable weight in the neural network model, a linear adjustable operational transconductance amplifier and a current multiplier circuit are designed as synaptic circuits to realize the adjustable weight function by changing the external current. The designed circuit has simple structure and high linearity. This method can be used as a basic unit in neuron circuits. (2) based on the problem that it is difficult to implement the competition layer algorithm in the self-organizing competitive neural network, a current-mode circuit simulation algorithm is designed to realize the competition. By comparing the magnitude of the current to achieve the goal of competition. The circuit is simple, high analog, easy to integrate, and can be combined with input layer to realize self-organizing competitive neural network. (3) aiming at the problem of complex structure and low precision of the cell circuit of fuzzy neural network, this paper deals with Gao Si function circuit. In order to improve the precision and high speed of the fuzzy neural network, the structure of the small circuit and the de-fuzzy circuit are optimized. Finally, the designed fuzzy neural network is used to realize the approximation of a nonlinear function, and the simulation and verification are carried out.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN710

【参考文献】