Extended Kalman Filter and Neural Network Filter: Who Outperforms Whom in System identification?

Abstract

Many researchers believe that neural network filter which employs backpropagation learning technique (an approximation of a gradient descent procedure) is a degenerate form of Extended Kalman filter. The argument that the gradient computed from the neural network approach for the error surface is defined by the immediate training vector only and not the ensemble of training vectors, thereby producing less accurate results as compared to extended Kalman filter, is revisited in this paper. The complexity of Space Station Freedom dynamics is used to determine who outperforms whom in system identification. Control moment gyros that are part of the Station's basic momentum management system are chosen to provide input excitation in the form of applied torques. These torques together with the measured angular body rate responses are supplied to the filters. From these data, both filters are shown to accurately identify the station mass properties when excitation levels are high and well balanced between axes. The neural network filter, however, is shown to be more robust and to perform well even with weakly persistent, unbalanced signals contaminated with noise.