Aprendizado de máquina aplicado ao reconhecimento automático de falhas em máquinas rotativas

• Ano de defesa: 2018
• Autor(a) principal: Diego Silva Caldeira Rocha
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Diagnóstico de falhas em máquinas rotativas; XGBoost; Extração de características; KNN; SVM; Vibrações mecânicas; Engenharia elétrica; Máquinas rotativas Falhas
• Link de acesso: http://hdl.handle.net/1843/BUBD-B4PP45

Rotary machines such as motors, generators and pumps are commonly used in almost all industrial processes. The analysis of mechanical vibrations has been an important technique adopted in companies to evaluate the state of operation of industrial machines. This work uses a database of mechanical vibration signals to automati-cally classify faults in rotary machines. Three models of extraction of characteristics of mechanical vibration signals are presented: (i) RMS (Root Means Squares), (ii) Haar Wavelet and fractal dimension and (iii) FFT (Fast Fourier transform) with statistical data. Finally, the machine learning concept is used with the classifiers KNN (K-NearestNeighbors), SVM (Support Vector Machine) and XGBoost (Extreme Gradient Boosting) to diagnose faults. The results demonstrate the effi ciency of all the techniques, although wavelet approach and fractal dimension combined with XGBoost, presenting the best results. It was possible to reach an accuracy of 98 . 7% (MAUC (Multi-class Extension ofAUC)=0.9704) on rotating machine failures and 99 . 36% of accuracy (MAUC=0.9965) for bearing problems. In addition, it obtained remarkable intraclass results and was very promising for the subject of this dissertation.