SBIR Phase I: Very Large Scale Integrated (VLSI) Implementations of Neuromorphics Virtual Sensors for Intelligent Diagnostics and Control

SBIR 第一阶段：用于智能诊断和控制的神经形态虚拟传感器的超大规模集成 (VLSI) 实现

• 批准号: 9660637
• 负责人: Alexander Moopenn
• 金额: $ 7.48万
• 依托单位: Mosaix, LLC
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 1997
• 资助国家: 美国
• 起止时间: 1997-01-01 至 1997-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9660637&HistoricalAwards=false
• 关键词: SBIR; Phase; Very; Large; Scale

*** ABSTRACT 9660637 Moopenn This Small Business Innovation Research Phase I project will develop a novel, compact, low-cost, and single chip adaptive neuroprocessor. This bit-serial based digital CMOS VLSI electronic neural network device will combine on-chip integration of a fully reconfigurable feedforward/time-lagged recurrent neuroprocessor with a backpropagation weight training module. Specifically, the technical objective of this research is to develop a neuroprocessor chip suitable for direct insertion into Ford's advanced concept vehicles. In operation, this stand-alone electronic neural network will act as a co-processor to the engine computer's (EEC) central processing unit (CPU). The neuroprocessor will be software programmable, enabling it to execute multiple different neural network applications on-the-fly; be capable of event rate computational throughput ( 100 microseconds) per application; be of a single-chip design (neuroprocessor with on-chip weight training); and cost effective ( $20/chip). The importance of on-chip adaptation is to address the problems of fixed weight neural networks--namely that an applications synaptic weights (as optimized at the factory for a generic model) be allowed to tweak/self-calibrate themselves for optimal diagnostic and control performance on the vehicle in order to handle most accurately all conditions under which the particular system is deployed. This research is in direct collaboration with Ford Motor Company. The end product of this research and development is particularly relevant to all diagnostics and control applications in the automotive industry, in aerospace, as well as process control in the electronics commercial industry. ***

* 摘要 小行星9660637 这个小企业创新研究第一阶段项目将开发一种新颖，紧凑，低成本和单芯片自适应神经处理器。这种基于位串行的数字CMOS超大规模集成电路电子神经网络设备将完全可重配置的前馈/滞后循环神经处理器与反向传播权重训练模块的片上集成联合收割机结合起来。具体来说，这项研究的技术目标是开发一种适合直接插入福特先进概念车的神经处理器芯片。在操作中，这个独立的电子神经网络将充当发动机计算机（EEC）中央处理单元（CPU）的协处理器。神经处理器将是软件可编程的，使其能够在运行中执行多个不同的神经网络应用程序;能够为每个应用程序提供事件速率计算吞吐量（100微秒）;是单芯片设计（具有片上权重训练的神经处理器）;并且具有成本效益（20美元/芯片）。片上自适应的重要性在于解决固定权重神经网络的问题，即允许应用程序突触权重（在工厂针对通用模型进行优化）调整/自校准自身，以获得车辆上的最佳诊断和控制性能，以便最准确地处理部署特定系统的所有条件。 这项研究是与福特汽车公司直接合作进行的。 这项研究和开发的最终产品与汽车工业、航空航天以及电子商业行业的过程控制中的所有诊断和控制应用特别相关。***