Fully parallel learning neural network chip for real-time control.

摘要

A fully parallel learning neural network chip was applied to perform real-time output feedback control on a nonlinear dynamic plant. A hardware-friendly learning algorithm, the RWC algorithm was used. The original RWC chip was modified to be more suitable for real-time control applications. Software simulations indicated that the RWC algorithm was able to control an induction motor on-line to generate desired output stator current, despite the analog circuit nonlinearity. Another real-time application considered in the research was the combustion instability control in a jet or rocket engine. This is a dynamic nonlinear system, which can be very hard to control using traditional control methods. Extensive software simulations were carried out, using the RWC algorithm to control the combustion instability. The simulation results proved that the RWC algorithm worked with this application. This was the first time that the algorithm was proved to function with a real-time problem in simulation. The modified RWC chip was then fabricated. A series of preliminary hardware tests was carried out. They proved that the chip could perform on-chip learning, operating in a fully parallel manner. The RWC chip was applied to control a computer-simulated combustion to successfully suppress the oscillation. This was the first time that an analog neural network chip was tested to control a simulated dynamic, nonlinear system successfully.