Neural network feed forward compensation of nonlinear friction with a PID controller

A neural network is used to cancel nonlinear perturbations presented to a classical PID controller during trajectory following. Inputs are injected in parallel to both PID controller and neural network, the network's output is summed with that of the controller's in a feedforward fashion. The neural network is a modified backpropagation network, where the error signal propagated back through the network is not based upon least squared error, but rather the average of least squared error. This technique is shown to be less sensitive to noise and converges more smoothly to the minimum error. The network is trained "off-line" to learn the system dynamics and frictional disturbances. Then the tracking error generated by the PID controller with, and without the neural network feedforward is compared and evaluated. The results indicate that systems using a feedforward backpropagation neural network can significantly reduce tracking errors due to nonlinear disturbances.