Static neural network process models: considerations and case studies

Neural networks are beginning to be used for the modeling of complex manufacturing processes, usually for process and quality control. Often these models are used to identify optimal process settings. Since a neural network is an empirical model, it is highly dependent on the data used in construction and validation. Using data directly from production ensures availability and fidelity, however the samples may not reflect the entire range of probable operation and, in particular, may not include the optimal process settings. Supplementing production data with observations gathered from designed experiments alleviates the problem of overly focused or incomplete production data sets. This paper considers practical aspects of building and validating neural network models of manufacturing processes, and illustrates the recommended approaches with two diverse case studies.