Análise, avaliação e validação do uso de técnicas de inteligência artificial para o diagnóstico de falhas em motores de indução

Abstract

Predictive maintenance uses several methods to monitor the conditions of electric motors applied in industrial plants. Among these methods, vibration analysis stands out as a widely used method due to the possibility of identifying a wide possibility of failures. This process still depends a lot on the analysis and manual processing of the informa tion collected, which can lead to a longer time to perform the analysis of the condition of a machine, which consequently generates more costs for the company. Depending on the number of equipment, this task can take a long time, requiring more manpower; It is also possible that errors in the interpretation of the data occur. With the result of this research, and with the algorithm tested and validated, interpretation errors must be minimized. This work presents the development of a computational model equipped with a classifier algorithm to receive vibration readings from electric motors and determine if the machine presents a failure behavior and, if so, which failure. Different classification algorithms will be compared in order to determine the most appropriate methodology to solve this problem. In this research, the MaFaulDa dataset was used since it contains vibration readings under different failure conditions simulated in the laboratory, which allows training the model to detect a wide range of defects. The computational model performs the pre-processing of the data through the Fourier Transform (FFT), then the algorithms are trained with the vibration signals in the frequency spectrum. After vali dation, the results of the algorithms were compared to determine the most suitable for the solution of the problem proposed in this work. After validation, the results of the algorithms were compared to determine the most suitable for the solution of the problem proposed in this work. After evaluating the accuracy of all four models, Random Forest showed the best accuracy for the dataset, with 97.18% accuracy. The second algorithm with the best hit rate was Support Vector Machines, with 94.05% accuracy with the fil tered dataset. The K-Nearest Neighbor model presented 91.39% of accuracy, making it the least assertive among those involved in the study, while the Artificial Neural Network presented a relevant advantage when trained with the complete dataset, without attribute filters.