A comparative study on machine learning regression algorithms aplied to modeling gas centrifuge / Um estudo comparativo sobre algoritmos de regressão de aprendizagem de máquinas aplicado à modelagem de centrífugas a gás

• Autor(a) principal: Andrade, Delvonei Alves de
• Data de Publicação: 2022
• Outros Autores: Mesquita, Roberto Navarro de, Nascimento, Natan Patussi
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Revista Veras
• Texto Completo: https://ojs.brazilianjournals.com.br/ojs/index.php/BRJD/article/view/50490
• topic: gas centrifuge; gas centrifuge; Uranium enrichment; Uranium enrichment; machine learning; machine learning; multivariate regression; multivariate regression; xgboost; xgboost; artificial neural network; artificial neural network; support vector machine; support vector machine; spline; spline; k- nearest neighbors; k- nearest neighbors; multivariate adaptive regression splines; multivariate adaptive regression splines

The gas Centrifuge is a very hard equipment to model, because it involves a gas dynamic with many complications, such as hypersonic waves and rarefied regions combined with continuous flow areas. Therefore, data analysis regressions remain currently a very important technique to understand and describe the problem in a practical way. This paper intends to apply and compare several regression techniques using machine learning, to obtain a hydraulic and a separative power model of gas centrifuge used in enrichment plants. For this purpose, a set of normalized data composed of 134 experimental lines was used, observing the variables of interest, the separation power (dU), and the waste pressure (Pw), through the following explanatory variables: feed flow (F), cut (q), and product pressure (Pp). The comparisons were presented between the results obtained for the models generated by the following: algorithms, multivariate regression, multivariate adaptive regression splines – MARS, bootstrap aggregating multivariate adaptive regression splines – Bagging MARS, artificial neural network – ANN, extreme gradient boosting – XGBoost, support vector regression– Poly SVR, radial basis Function support vector regression – RBF SVR, K-nearest neighbors – KNN and Stacked Ensemble. That way, to avoid overfitting and provide insights about generalization of the models in unseen data, during the training phase, the k-fold cross validation approach was used.  Subsequently, the residuals were analyzed, and the models were compared by the following metrics: Root mean square error – RMSE; Mean squared error – MSE; Mean absolute error – MAE; and Coefficient of determination – R2.