A Prediction System Based on Neural Networks and Rough Sets in a Highly Automated Production Process

In this paper we describe results relating to intelligent data processing and the use of neural networks predictors in a highly automated factory. We consider the daily updated number of overdue orders as a data output from the production process (the production system can be viewed as processing data since information is generated by the process along with the products). For modelling and prediction purposes we analyse historical data on the number of overdue orders. These missed deliveries are caused by many interrelated reasons in a complex manufacturing system (such as the distribution of incoming orders, machine breakdown, personnel shortage, scheduling ineeciency, etc.), but we do not seek to establish causal links in our analysis. We demonstrate that rough sets analysis is an eeective preprocessing technique for reducing the number of data without reducing the useful information, before using the data to train a neural network. We divide (based on statistical and expert decisions) the expected range of the number of overdue orders into a small number of bins. Similarly we quantize and normalise the other data in the system and x time windows for the data and use this as a reeection of the internal system state. Then we use rough sets analysis as a tool (and front-end to neural networks) for reducing and choosing the most relevant sets of internal states for predicting the level of overdue orders ve days in the future. We use the selected data to train a feedforward backpropagation neural network and a recurrent neural network and compare them with the same networks trained on all the input data.