The application of the robust inferential sensors was considered in 2004 when a vendor installed system was to be replaced and in-house solution was determined to be much more cost effective. Two modeling technologies were concurrently evaluated – Partial Least Squares (PLS) which is a linear approach and Genetic Programming (GP) which is a nonlinear approach. As a consequence GP models were proven to be superior and were implemented.
The robust implementation is achieved by sensor validation. Sensor validation provides redundant backup for the inputs of the inferential sensor, so the sensor continues to operate even if some of the inputs fails. This robustness is also an EPA requirement. Due to this implementation and the quality of the models, four years of EPA compliance are recorded as of 2008. Certain details on the inferential sensors maintenance and standardization will be discussed.