Machine learning based adaptive soft sensor for flash point inference in a refinery realtime process

In industrial control processes, certain characteristics are sometimes difficult to measure by a physical sensor due to technical and/or economic limitations. This fact is especially true in the petrochemical industry. Some of those quantities are especially crucial for operators and process safety. This is the case for the automotive diesel Flash Point Temperature (FT). Traditional methods for FT estimation are based on the study of the empirical inference between flammability properties and the denoted target magnitude. The necessary measures are taken indirectly by samples from the process and analyzing them in the laboratory, this process implies time (can take hours from collection to flash temperature measurement) and thus make it very difficult for real-time monitorization, which in fact results in security and economical losses. This study defines a procedure based on Machine Learning modules that demonstrate the power of real-time monitorization over real data from an important international refinery. As input, easily measured values provided in real-time, such as temperature, pressure, and hydraulic flow are used and a benchmark of different regressive algorithms for FT estimation is presented. The study highlights the importance of sequencing preprocessing techniques for the correct inference of values. The implementation of adaptive learning strategies achieves considerable economic benefits in the productization of this soft sensor. The validity of the method is tested in the reality of a refinery. In addition, real-world industrial data sets tend to be unstable and volatile, and the data is often affected by noise, outliers, irrelevant or unnecessary features, and missing data. This contribution demonstrates with the inclusion of a new concept, called an adaptive soft sensor, the importance of the dynamic adaptation of the conformed schemes based on Machine Learning through their combination with feature selection, dimensional reduction, and signal processing techniques. The economic benefits of applying this soft sensor in the refinery's production plant and presented as potential semi-annual savings.