Efeito da seleção de métricas de desempenho em modelos de aprendizado supervisionado para detecção de amaciamento em compressores

Abstract

This work presents an analysis on the effect of performance metrics on the selection of supervised methods for running-in detection in reciprocating hermetic compressors. Since the running-in phenomenon takes place only in the first few hours of the lifespan of such devices, datasets extracted from the compressor operation during this period and in its normal operation are usually more representative of its post running-in behavior, and, even after balancing the dataset, an adequate choice of the performance metric might reduce the undesired effects of this situation in classification models. In order to reach the proposed goal, parameters of random forest models and of data preprocessing methods were optimized for the following performance metrics: the area under the receiver operating characteristic curve, the F -score with three different configurations, and the Matthews correlation coefficient. When applied to experimental data, the results obtained with such models showed the Matthews correlation coefficient and the F -score with more weight attributed to precision (β = 0,5) as the most appropriate ones to the given case study.