Neural network-based hybrid estimator for smoothly-varying measurement signals

This paper presents an adaptive estimator for predictive filtering of smoothly-varying signals. The hybrid estimator uses a multilayer backpropagation neural network as a derivative detector that identifies whether the first, second, and third derivative exist in the incoming signal. The detector is used to control the path multipliers of the estimator. Therefore, the estimator order can be changed dynamically when the piecewise polynomial signal changes its characteristics in different time segments. The use of a neural network requires to determine the optimal number of input and hidden nodes needed for successful derivative detection. In this paper, a method of Predictive Minimum Description Length (PMDL) has been used to solve this problem so that over-fitting and under-fitting can be penalized automatically. This results in a network that has both excellent generalization capability and moderate computational complexity