Application of artificial neural networks for filtering, smoothing and prediction for a biochemical process

The applicability of feed-forward neural networks for filtering, smoothing and prediction for processes is illustrated in this paper, exemplified by a simulated biochemical process.

Abstract

Abstract: The applicability of feed‐forward neural networks for filtering, smoothing and prediction for processes is illustrated in this paper, exemplified by a simulated biochemical process. Filtering, smoothing and prediction are valuable for noisy processes and most of the standard techniques need some knowledge about the process or a model of the process. Artificial neural networks provide a good alternative to the conventional filters and do not need a model of the process. The inputs to the networks were one or more state variables with delayed measurements; the networks calculate the noise‐free estimates of the variables. The Levenberg‐Marquardt method was used to train the neural networks by minimising the sum of squares of the residuals. The results of filtering, smoothing and prediction were reasonably accurate. Linear activation functions were found to be adequate for most purposes. This is due to the fact that the non‐linearity of the process is not very strong in the region of state space considered.