Detecção automatizada de defeitos em impulsores fundidos utilizando inteligência artificial

Abstract

One of the problems faced by manufacturing industries is the casting defects present in their production. Traditionally, one of the most used methods in quality control is visual inspection, which is a non-destructive and inexpensive method of detecting nonconformities. However, their efficiency depends on several factors that are not always within the inspector’s control. Alternatively, some industries have adopted the use of machine vision systems in the quality sector as a way to replace human vision with more sophisticated computers and cameras. A technology capable of assisting such systems in their classification tasks are Neural Networks, a specific division of Artificial Intelligence. This work aims to develop a model in Convolutional Neural Network for the classification of cast impeller pumps through the analysis of their images, as a part of a machine vision system of a hypothetical manufacturing industry. The objective is for the algorithm to be able to accurately classify a component as conforming or defective, depending on the presence or absence of casting defects by analyzing its images. A problem that arises in this task is in the analysis of unstructured data by the computer. Using the CRISP-DM methodology and specific Python libraries, it was possible to develop a model with 99.44% accuracy in its predictions in the test data.