Applications of Neural Networks for Filtering

This paper investigates the applicability of feedforward neural networks for filtering purposes. A biochemical process was simulated to generate the data for training and testing the networks. Delayed measurements of one or more state variables were used as inputs to the networks, which were trained to provide filtered values of the state variables as outputs. The results of filtering were quite accurate. In most cases, linear models i.e., networks with linear activation functions, were found to be adequate. This is due to the short sampling time and the fact that the non-linearities of the process are not very strong in the region of state space considered.