An online calibration tool for soft sensors: development and experimental tests in a semi-industrial boiler plant

• Published in: Brazilian journal of chemical engineering 2020-03, Vol.37 (1), p.189-199
• Main Authors: Parente, Andréa Pereira, Valdman, Andrea, Folly, Rossana Odette M., de Souza, Maurício Bezerra, Fileti, Ana Maria Frattini
• Format: Article
• Language: English
• Subjects: Chemistry; Chemistry and Materials Science; Industrial Chemistry/Chemical Engineering; Original Paper
• ISSN: 0104-6632; 1678-4383
• DOI: 10.1007/s43153-019-00005-w

Soft sensors with real time prediction capabilities appear as a profitable solution for hard-to-measure variables whenever hard sensors are difficult to apply or subjected to high operational costs. Nonetheless, the use of soft sensors within industrial applications is still not widespread because of the systematic accuracy issues that can be introduced with process plant deviations from nominal operation states. Soft sensor models need to be constantly updated to avoid degradation of their prediction potential. This study presents an innovative view on a well-known artificial neural network (ANN) calibration method by developing a generic online calibration tool that can be used in independent data-driven soft sensors based on ANN multi-layer perceptron (MLP) models. The maintenance framework has been fully tested in a semi-industrial boiler plant to predict real time pollutant emission levels, presenting recalibration time responses up to 1 min, overall r 2 performance above 80% and an intuitive human–machine-interface.