Estudo e Comparação de algoritmos de classificação de imagens para identificação de cristais de alita (C3S)

Abstract

For the production of Portland cement, clinker quality is fundamental, representing the greatest challenge for cement industries. Clinker is a sintered and pelletized material, produced from the burning of a mixture of limestone and clay raw meal at a temperature of 1450°C. During its final production stage, an inspection is carried out to provide data on the material’s quality. This verification can be conducted using techniques such as X-ray diffraction and microscopy, with the latter being the most frequent. Because it is performed manually, microscopic analysis can suffer from human error, resulting in delays in detecting production problems and inconsistent analysis results. Therefore, a need arises for a solution based on the classification of Alite (C3S) crystals that contributes to a process that yields results and can benefit the cement industry. The morphology and distribution of these crystals in images are crucial indicators of the produced clinker’s quality. Through the effective classification of these crystals, it is possible to automate analyses that depend on the evaluation of their morphology. Such an advancement would expedite the quality assessment process and also increase its precision, significantly benefiting the cement production industry. In this context, this work focuses on three algorithms: KNN (K-Nearest Neighbors), SVM (Support Vector Machines), and Decision Trees, along with feature extraction techniques for the classification of C3S crystals in microscopic clinker images. This research concludes by presenting the comparative results of the three supervised models regarding their capacity to classify Alite crystals. The Decision Tree model showed the highest performance in the morphological classification of C3S crystals, achieving 97.19% accuracy and a 97.16% F1-Score, which were superior results to those obtained by the SVM and KNN algorithms. The learning curves indicated a high discriminative power and good model generalization, showing only slight overfitting. The study faced limitations, mainly the scarcity of literature on the Alite phase and a limited set of images for training. The research contributes to the advancement of automated morphological analysis. Future investigations should focus on expanding the database, exploring deep learning architectures, and incorporating new attributes to enhance the robustness of the predictive models.