Artificial Neural Networks of Improved Reliability for Industrial Process Supervision

An efficient model development strategy for bioprocesses based on neural networks in macroscopic balances: Part II

“…Other researchers who trained neural networks in a SGBM used only the integral mode, because their measurements were corrupted with a lot of noise (Psichogios and Ungar, 1992) or because the measurements were only available at large time intervals Simutis et al, 1995;Schubert et al, 1994). In the applied identification procedure in the present paper, the computational advantages of the differential mode could be used to train 20 neural networks with a varying number of hidden nodes and to select from these a proper number of hidden nodes.…”

“…The serial gray box modeling strategy, in which the neural network is placed in series with a white box (first principles) model, seems to be the most suitable to obtain reliable mathematical models of a process within a limited amount of time and with limited experimental effort. Various researchers Psichogios and Ungar, 1992;Schubert et al, 1994;Simutis et al, 1995;Tholudur and Ramirez, 1996;Thompsom and Kramer, 1994) showed the potential extrapolation properties of the serial gray box modeling strategy. However, these papers made no distinction between different types of extrapolation, which makes it very difficult to relate a priori the desired application domain of the model to the minimally required domain for the identification data.…”

Bioprocess Iterative Batch-to-Batch Optimization Based on Hybrid Parametric/Nonparametric Models

“…The trust region defines the state subspace where the nonparametric model has proven to be reliable. This region was determined by clustering the measured subspace in a fashion similar to that described by Leonard et al (1992) and Simutis et al (1995) to monitor the reliability of artificial neural networks in on‐line applications. The main idea behind this method is that if the network is properly trained, then it will perform well whenever the inputs are not far way from the “experience” acquired during the training phase.…”

“…The clusters centers m j were optimized by running the k ‐means algorithm over the experimental data used to identify 1c (Leonard et al, 1992; Simutis et al, 1995). The final set of clusters with widths given by eq forms a continuous density function f: c →ν by applying the maximum operator: …”

Estimation of optimal feeding strategies for fed-batch bioprocesses

“…Due to their capacity to learn, filter and generalize information through a training procedure, the artificial neural networks (ANN) have been utilized with a relatively high success for system modeling and identification in biotechnology [17][18][19][20][21]. In these applications in particular, the ANN are used commonly as black-box models of key variables.…”