EVOLUTIONARY MODEL SELECTION FOR DATADRIVEN SOFT SENSORS
MARCHIORO, Matheus Henrique Reis1; MEZA, Gilberto Reynoso 3; RIBEIRO, Victor Henrique Alves2;
Introdução:Industrial processes must be well equipped with a diversity of sensor to be able to monitor their current state and take decisions. However, there are situations where not all variables from a process can be measured. Therefore, many researchers started using available information from the process to create predictive models as way of estimate the value of “hard-to-measure” variables. In the literature, these models are known as Soft Sensors. One possible solution to create such systems is using supervised machine learning.
Objetivo:The aim of this project consists in using evolutionary optimization to perform model selection on the creation of prediction models for soft sensor development.
Metodologia:Problems in which the designer must deal with the fulfillment of multiple objectives are known as multi-objective problems (MOPs). When dealing with an MOP, we usually seek a Pareto optimal solution. In order to approximate this Pareto set, the Genetic Algorithm (GA) will be applied in conjunction with the techniques Support Vector Regression (SVR) and Decision Trees (DT). The main goal is to generate many potentially desirable Pareto optimal solutions, and then select the most preferable alternative. The selection takes place in a multi-criteria decision-making (MCDM) step. Physical Programming, Promethee and TOPSIS have been used as MCDM techniques to deal with MOPs. The combination between MOP, optimization and MCDM is called as Multi-Objective Optimization Design (MOOD). Many combinations between different techniques will be applied to develop prediction models for two soft sensor problems. The first problem consists in a Debutanizer Column and second is a Sulfur Recover Unit (SRU). Critical Difference Diagrams will be used as a way to evaluate the performance of the combinations tested.
Resultados:TOPSIS proved to be the worst method for selecting possible solutions, while Physical Programming and Promethee, statistically, have the same performance. Among the strategies of multi-objective optimization implemented, Hyperparameter Selection (HP) and Feature Selection (FS) proved to be the best candidates for designing predictive models for soft sensor applications.
Conclusões:It was possible to implement a framework capable of evaluating several multi-objective optimization techniques. This project has great potential for future work, either for the validation of the performance of the combination of techniques, as well as for the development of models superior to those in the literature, since the developed framework can be submitted to more advanced techniques.
Palavras-chave: Soft Sensor. Multi-objective Optimization Design. Support Vector Regression. Decision Tree. Multicriteria Decision Making.