Detección de fallas en termogramas utilizando aprendizaje profundo

Abstract

This research carried out the detection of failures in thermograms using Deep Learning. This field of Artificial Intelligence bases its operation on artificial neural networks which try to simulate the behavior of the human brain. For that reason, we got information from scientific articles, books, research, magazines, and news from Artificial Intelligence web pages and bookstores. A database containing thermograms with failure and thermograms without failure was needed to develop the algorithm. This database was obtained from a previous thermography study carried out in the Industrial Maintenance Career of ESPOCH. For the programming of the algorithm, the free software called Python was used. Initially, a pre-processing of the images and classification of the thermograms for training, validation, and testing were carried out. Because the database was relatively small, a tool called Image Data Generator was used. It serves to increase the number of images, making certain modifications to the original thermograms. Four models of neural networks were generated to observe which of them presented the best results and later selected as the best classification algorithm. Model 4 was the one that showed the best results, obtaining a precision of 97.29% and an accuracy of 97.94%. In addition, results of other metrics, such as the F1-score, were acquired, giving a mark of 97.95%. Finally, predictions were made with the test thermograms, and the algorithm classified them correctly, thus corroborating that it is a successful model for detecting failures in thermograms. For this reason, it is recommended to carry out more research on thermography with artificial intelligence within industrial maintenance.