Fault-Tolerant Parallel Distributed Neural Network Processing System for Adaptive Control and System Identification

S&K Electronics (SKE) proposes a fault-tolerant, parallel distributed processing (FTPDP) system capable of implementing a variety of neural network paradigms for system identification and control applications. The proposed FTPDP system offers a scalable and modular hardware architecture that interconnects small, low-cost neural processors with fault-tolerant high-speed serial buses. The neural processors will have onboard analog or digital I/Os for local data collection and control. Firmware onboard the processors will handle the dynamic resource allocation and configuration, neural network algorithm execution, including on-line learning, inter-processor communication and fault isolation. A host software package will be developed to support application development and provide flexibility ti implement various neural network paradigms. The host software will also enable users to perform on-line load balancing by reassigning neurons to different physical neural processors as well as monitoring the status of the network.The proposed FTPDP system is well suited for current and future spacecraft missions requiring accurate pointing and structural stability, e.g. the Hubble Space Telescope, the Orbiting Stellar Interferometer, and the Precision Orbiting Interferometer in Space (POINTS).