Operation of a neural network controlled crystal oscillator

Abstract

The neural network control of a quartz oscillator has been demonstrated. We have shown that a single neural network can correct an oscillator's output frequency while several environmental sources of frequency shift act on the oscillator. The advantage that a neural network offers over microprocessor controlled oscillators or Kalman filter control techniques is that a neural network does not need an a priori model of the physical system to arrive at the correct control algorithm. The results of this demonstration indicate that a neural network can have beneficial applications in a variety of frequency standard devices. We believe that the neural controller will work best as a control system supervisor, rather than as the main controller of the frequency standard system. Specifically, the neural network would learn the non-linear effects associated with the several servos that control the main physical parameters of a frequency standard (temperature, magnetic field, etc.), and correct the oscillator's frequency based on the state of these servos and any additional sensed environmental parameters. The dominant remaining technical issue is the training of the neural network. During this demonstration program, maintaining a stable, reproducible environment that could be varied quickly and randomly over the whole training parameter space proved to be a significant technological challenge. Future efforts will focus on methods to more efficiently train the neural networks and the identification of specific devices. >