Avaliação de redes neurais aplicadas na classificação e identificação de falhas de produtos em linha produtiva

Abstract

Artificial Intelligence (AI) is an emerging area that has applications in several sectors, such as engineering, industrial automation, health, agriculture, among others. In the last six years alone, the number of AI-related publications on arXiv has grown more than sixfold, from 5,478 in 2015 to 34,736 in 2020. Electrical engineering also benefits greatly from AI, with applications such as electricity demand forecasting, control of processes, selection of damaged products in the production process, among others. This area is extremely complex, as it covers a wide variety of techniques to solve different problems, where each method can have an infinity of variations. Among the various branches of AI, Deep Learning stands out due to its ability to solve complex problems with greater precision and speed. In this work, an image recognition system using neural networks for automated identification of damaged objects in industrial processes is proposed, in addition to the automated classification of these objects by type. Neural networks were trained with the purpose of classifying geometric figures, using parametric and structural variations, as well as the use of different data sets and different pre-processing strategies. The results obtained show the effectiveness of neural networks in the image classification task, with an accuracy higher than that estimated for human performance. The best model obtained in this study was compared to similar works, proving to be similar or superior in certain cases.