Normally, creating a neural network model requires at least the following tasks:
Data collection - This operation is essential since process data are the only base for building a neural net. The quality of the data determines the quality of the model.
Data preprocessing - This operation is necessary because real process data often contain missing values, outliers (data values outside the control limits you set), and possibly undesired data from different sources. These data have to be conditioned or pre-processed before they are used for network training.
Variable and time delay selection - This operation determines which of the available process variables are important ones that significantly affect the variables to be predicted. Inclusion of irrelevant variables can degrade the prediction accuracy.
Network training - This operation determines the number of hidden neurons and adjusts the weights based on a well-conditioned set of training data. The network's ability to change the weights allows the network to modify its neurons' behavior in response to their inputs, or to learn.
Network verification - This operation uses a separate set of data to test how well the neural network predicts the output.
DeltaV Neural automates most of these operations. The DeltaV implementation of these operations is explained in the Books Online section Using DeltaV Neural.