Neural network modeling of surface aeration in mechanically agitated bioreactors

摘要

The aim is to train a neural network to predict the intensity of surface aeration through the free liquid surface in stirred bioreactors. The primary network for prediction is called a feedforward network with non-linear elements. The network consists of an input layer, an output layer and a hidden layer, and the non-linear transfer function in each processing element is typically a sigmoid. The network requires supervised learning and the learning algorithm is the back-propagation or generalized delta rule. The system considered for predicting the surface aeration is designed to take an input pattern of six real-valued variables in the range 0-1, defining the overall size, speed and physical properties, and to assign it one real-valued output, the surface aeration intensity. The data set is divided into 65 training patters, 20 validation patterns and 10 test or evaluation patterns. The training set is used during the learning process in order to minimize the global error E of the system by modifying the weights. The validation set is used to avoid overtraining and the test set evaluates the generalization performance of the network. This study discusses the ability of neural networks to generalize surface aeration in stirred bioreactors within the parameters spanned experimentally.