Desarrollo de técnicas de minería de datos en procesos industrialesmodelización en líneas de producción de acero

Abstract

Data mining can be defined as the process of extracting useful, previously unknown, and comprehensible information from very large databases. In the industrial arena, one of the most interesting applications of data mining is the system modelling. Both the rapid growth of data storage capacities of modern industrial processes and the development of data processors, provide new possibilities to analyze their behaviour. Moreover, taking into account that in the most industrial processes, the relationships between variables are non-linear and the difficulty in obtaining explicit models to define their behaviour, we can understand the importance of models based on data against other analytic models based on explicit equations. Nowadays, neural networks are one of the most significant techniques to model manufacturing systems because of their efficiency and simplicity. Neural networks are the central issue around which this thesis is developed. Basically, this thesis proposes the use of neural networks, and other techniques from data mining, to model an industrial process: a steel hot dip galvanising line. In particular, manufacturing process data are used to improve the current control systems by means of models capable to on-line predict the mechanical properties of the galvanised steel strip and a model of the velocity of the steel strip inside de process furnace. Unfortunately, it is well known the occurrence of outliers in real industrial datasets (electromagnetic interferences, peak current values during motor start-up, the human factor, etc.). In order to manage the presence of outliers in neural networks training, a new robust learning algorithm is proposed. This approach, which is an innovation in the so-called robust neural networks, is based on the non-linear t-estimator and the backpropagation learning algorithm.