This paper presents the development and testing of a new algorithm to identify and reconstruct faulty sensors, based on a statistical model using quantitative statistical process history. The process description is assured by a database containing the measurements selected under steady state condition and without faults during the operating life of the plant. An initial analysis of the historical data allows the algorithm to choose the most relevant signals to be used, through a variable selection process, optimizing the subsequent diagnostic phase. Multivariate regression models were used to perform a supervised feature selection and different state of the art selection criteria plus a new strategy were implemented and tested on experimental data comparing the different results. A statistical model of the system was then developed using Principal Component Analysis (PCA) and the impact of feature selection was estimated. Finally, a case study was developed based on a real micro gas turbine facility, at the University of Genoa, using historical data. The efficiency of the feature selection algorithm versus the knowledge of the experts in thermodynamics was evaluated.
