Skip to Content
MachinesMachines
  • Article
  • Open Access

18 July 2023

15 Pages

A CNN-Based Methodology for Identifying Mechanical Faults in Induction Motors Using Thermography

,
,
,
,
,
and
1
C. A. Mecatrónica, Faculty of Engineering, Autonomus University of Queretaro, Campus San Juan del Río, Río Moctezuma 249, Col. San Cayetano, San Juan del Río 76807, Mexico
2
G. C. Sistemas de Inteligencia Artificial Aplicados a Modelos Biomédicos y Mecánicos, Faculty of Engineering, Autonomus University of Queretaro, Campus San Juan del Río, Río Moctezuma 249, Col. San Cayetano, San Juan del Río 76807, Mexico
3
Tequexquite: Centro de Investigación y Desarrollo Tecnológico para la Accesibilidad e Innovación Social, Faculty of Engineering, Autonomus University of Queretaro, Campus Tequisquiapan, Carretera San Juan del Río-Xilitla, Km 19+500, Tequisquiapan 76750, Mexico
4
G. C. Sistemas de Inteligencia Artificial Aplicados a Modelos Biomédicos y Mecánicos, Faculty of Engineering, Autonomus University of Queretaro, Campus Aeropuerto, Carretera a Chichimequillas S/N, Ejido Bolaños, Santiago de Querétaro 76140, Mexico

Abstract

Infrared thermography (IRT) has become an interesting alternative for performing condition assessments of different types of induction motor (IM)-based equipment when it operates under harsh conditions. The reported results from state-of-the-art articles that have analyzed thermal images do not consider (1): the presence of more than one fault, and (2) the inevitable noise-corruption the images suffer. Bearing in mind these reasons, this paper presents a convolutional neural network (CNN)-based methodology that is specifically designed to deal with noise-corrupted images for detecting the failures that have the highest incidence rate: bearing and broken bar failures; moreover, rotor misalignment failure is also considered, as it can cause a further increase in electricity consumption. The presented results show that the proposal is effective in detecting healthy and failure states, as well as identifying the failure nature, as a 95% accuracy is achieved. These results allow considering the proposal as an interesting alternative for using IRT images obtained in hostile environments.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.