Detección de fallos en mantenimiento predictivo utilizando el método de aprendizaje de máquina Random Forest

Abstract

The objective of this research was to create a predictive model of supervised machine learning, using the Random Forest method and the free software Python as a programming language, in order to detect faults in a machine. For this, the predictive maintenance data set ai4i2020 obtained from the machine learning repository of California University, Irvine (UCI) was obtained. The creation of the model was carried out in 6 steps: problem definition (identification of characteristics, task to be carried out and objective variable), data preparation (exploratory analysis, cleaning, oversampling due to the imbalance in the objective variable and division of data for training and test, where the data was split 75% and 25% respectively), feature engineering (extraction of statistical features in the time domain with the help of the Python library TSFEL, selection of features by recursive elimination of RFE attributes) , model training (hyperparameter optimization), model evaluation (using the original feature set and selected feature set), post hoc analysis (ranking of features that contributed most and least to failure prediction, statistical significance) . Obtaining as a result, the model showed a better performance when using the set of selected characteristics and optimized hyperparameters, with 99,26% accuracy and 98,95% precision. It is concluded that the model fails with a high performance for fault detection and generalizes correctly for new data. It is recommended that to obtain good results, you pay close attention to the data preparation step, since it largely depends on it.