Detecção de anomalias por meio de machine learning: estudo de caso de sistemas pneumáticos em caminhões

Abstract

The main idea of the this paper is the application of Machine Learning concepts and techniques, for the prediction of equipment failure related to the Mechanical Engineering area. Despite the vast possibility in choosing a software for the development of the work, we chose to use Orange Data Mining due to its friendly and intuitive interface promoted by the visual programming of the software. Real data collected daily from Scania trucks are used in the work. The guarantee of safety while driving heavy vehicles, a study in the prediction of failures in pneumatic systems is valuable, since it is responsible for the entire braking system of the vehicle. Measurement metrics such as Classification Accuracy, Precision, Recall and Specificity were used in the work. Algorithms such as: Logistic Regression, Naive Bayes and Random Forest were used in the Software for measurement. After performing two different methods in the treatment of the 60000 existing instances in the database, an accuracy of 99.4% was obtained in Random Forest, 85.6% in Naive Bayes and 99.1% in Logistic Regression. The high accuracy obtained in the algorithms demonstrates the real possibility of predicting failures in pneumatic truck equipment, with the possibility of developing equipment for real-time measurement in order to avoid possible accidents.