Aplicação de redes neurais para classificação de microinclusões de sulfeto de manganês em aços

Abstract

The factors that measure the influence of inclusions on steel properties are the distribution, morphology and size of inclusions. The standard used as a reference is ASTM E45, which explains the sample scanning method in order to analyze and classify inclusions in steels. Manual methods are the most used in laboratories and metallurgical companies because of their low cost, while automatic methods are characterized by high operating costs, which makes their use in industries difficult. Neural network models, on the other hand, are part of new technologies and are extremely advantageous for several applications. This Master's work was motivated by the use of a new methodology for classifying inclusions in steels using neural network models. The aim was to achieve the highest possible accuracy in classifying the severities of MnS inclusions using images captured from specimens processed in a laboratory of a metallurgical industry. The classification process by neural networks was validated through comparison with results obtained manually, showing higher accuracy and speed in decision making. The results also showed that the classification of severities with a smaller number of images in the database presented a lower accuracy. To understand the effectiveness of the neural network, the concept of confusion matrix was applied, which are comparative tables of the number of images that the neural network brought as a prediction of severities in relation to the actual manually classified values. The first test results were not positive, requiring a reclassification of all images in the database to reduce confusion in the neural network. After reclassification and application of the Dropout technique, the test results were superior to the previous ones. In conclusion, the training and validation accuracy results improved, making it possible to compare manual and automatic classification. In general, the neural network represented speed in decision making, proving to be a potential tool for the classification of inclusions.