Development ofaSoft Sensor foraThermal Cracking Unitusing asmall experimental data set
Anthony J. DiBella, Luigi Fortuna, S. Grazianil, Giuseppe Napoli, Maria Gabriella Xibilia, Universit, Viale A. Doria · 2007
Inthis paper wecompare anumber ofstrategies to Inthispaperwe propose andanalyze fourmethods copewiththeproblem ofsmall datasets intheidentification ofa concerning process identification withsmall datasets based on nonlinear process. Fourmethods areanalyzed: expansion ofthe neural networks. training setbyadding zero-mean fixed-variance gaussian noise, We provide awidesimulation-based comparative study of expansion ofthetraining setbyadding zero-mean gaussian noise theperformance ofneural networks trained after varying variance variable according withsignal amplitude, integrationtypes ofgaussian noise injection andintegrating bootstrap and between bootstrap methodandstacked neural networks, anda newmethod based ontheintegration ofbootstrap method, ofthe stacked neural networks method(16). noise injection method, andofstacked neural networks. Such Dataacquired onarealprocess regarding theThermal methods havebeenapplied todevelop a SoftSensor fora Cracking Unitworking inarefinery inSicily (Italy), areused Thermal Cracking Unit working inarefinery inSicily, Italy. toillustrate theimprovements thatproposed techniques produce.