Gross errors detection of industrial data by neural network and cluster techniques

• Autor(a) principal: Alves,R.M.B.
• Data de Publicação: 2002
• Outros Autores: Nascimento,C.A.O.
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Brazilian Journal of Chemical Engineering
• Texto Completo: http://old.scielo.br/scielo.php?script=sci_arttext&pid=S0104-66322002000400018
• topic: gross error; neural network; modeling; data analysis

This article describes the application of a three-layer feed-forward neural network to analyze industrial plant data. To adjust mathematical models (for control or optimization purposes) from plant data, it is necessary to analyze and detect outliers and systematic errors and to remove them. The system studied is the feed preparation of an isoprene production unit and represents a multivariable problem. To detect outliers in a multivariable system is not an easy task. The technique used in this paper is able to identify this kind of error. The methodology employed involves construction of a reliable neural network model to represent the process and its training with a few iterations (a few thousand). Thus, the points at which errors between the experimental and calculated data appear to be scattered far from the majority of the values are probably outliers. In some cases, outlier points can be easily detected, but in others, they are not so obvious. In these cases, they are separated and a cluster with other similar data is built. After analyzing these clusters based on the similarity principle or by hypothesis tests for means, it is then decided whether or not these points can be excluded. At the same time the process is checked for any abnormalities recorded during the specific period. Three year's worth of process data were analyzed and about 30% of the data were excluded.