Soft-sensor design for vacuum distillation bottom product penetration classification

Abstract

Petroleum oil refineries are complex systems that convert crude into subproducts of value. The profit of the refinery depends on the quality of the resultant subproducts, which are usually determined by a laboratory analysis called “Needle penetration”. Normally, this laboratory analysis is costly and time-consuming since entails around four hours for its accomplishment. In order to solve this limitation, this paper proposes a novel soft-sensor design for online vacuum distillation bottom product penetration classification. The design of the soft-sensor is based on a new approach, the two-stage methodology, that considers the joint effect of both Normalization and Supervised Filter Feature Weighting methods to transform the features. This methodology stands on the analysis of the real impact of applying normalization methods on the contribution of each feature, providing results that significantly differ from the traditional premises of the state-of-the-art. The analysis includes the impact of normalization on distance metrics such as the Euclidean. Also, a new adaptation of Pearson correlation for the estimation of the feature weights respect to categorical labels is proposed in this work. Once the features are transformed, five well-known Machine Learning (ML) algorithms (K-means, K-NN, RFc, SVC and MLP) are considered for the design of the soft-sensor. The final soft-sensor design is selected based on the feature space transformation strategy and the ML algorithm that achieves the best results in terms of: accuracy, precision, generalization and explicability. In order to validate the proposal, real monitored data from a petroleum refinery plant sited in The Basque Country is employed. Results show that the proposed two-stage methodology improves the results obtained by the Normalization methods.