Plenary lecture 1: statistical techniques for virtual sensors design using neural networks

This plenary speech covers the advantages of having statistical analysis to input data previous to training Artificial Neural Networks. It will be also presented some industrial applications including methodologies for designing virtual sensors for oil companies. Shorter training periods, simpler topologies and more reliable networks can be found. The presented techniques for variables and patterns selection allow reducing the data dimension, obtaining quicker training, simpler topologies and lower prediction errors. The pattern reduction techniques allow generating a data partition for training and validation based on statistical analysis. Additionally, these selection techniques can be used for reducing the patterns number in the data when it is very high. The Outliers detection techniques can be used when great volumes of data are used for neural networks training and it is possible to use them for developing algorithms that detect possible observations significantly different from the rest of the data. These techniques can depurate and select those data that provide a better training. It is very important the fusion of both disciplines: Artificial intelligence and Statistical Data Analysis. The work shows the advantages that it has for the practical Statistic the Artificial intelligence and vice versa.