Utilização de redes neurais artificiais para a classificação de corrosão por pite observadas em imagens obtidas com o ensaio não invasivo de queda de potencial

Abstract

Metallurgical Engineering is responsible for the transformation of metals and the development of alloys with special properties. However, these alloys are subject to various degradation mechanisms over time, among which corrosion stands out as one of the most significant. Corrosion is an electrochemical process that leads to the deterioration of metallic materials due to their environment and can occur either uniformly or in a localized manner. Among the types of localized corrosion, pitting corrosion is notable, characterized by the formation of small pits on the material's surface. Despite their small size, these pits can compromise the structural integrity of metallic components. Automatically assessing the severity of these pits and classifying them is important to ensure safety and prevent accidents. Moreover, some applications tolerate certain levels of corrosion, and recognizing these exceptions is also important for cost reduction. In this study, the Electrical Potential Drop nondestructive testing method was used to detect corrosion pits in metallic components. The analyzed images, representing the current density on the material's surface, were generated by finite element simulations. To increase the variability of the simulated images, parameter ranges were defined for pit diameter, depth, and position, ensuring the representation of different defect scenarios. Two artificial neural networks with distinct topologies were applied to these simulations: a Multilayer Perceptron (MLP) and a Convolutional Neural Network (CNN). The networks were trained with images at different resolutions (64×64, 36×36, and 28×28 pixels) to evaluate the impact on accuracy. The best-performing MLP was configured with two hidden layers of 16 neurons, the Elu activation function, and a resolution of 28×28 pixels, achieving 93.88% accuracy and an F1-macro score of 0.9375, with a training time of 2.96 seconds. On the other hand, the best CNN reached 95.92% accuracy, outperforming all MLP models. This architecture included two convolutional layers with 8 and 64 kernels, max pooling, and dense layers with 64 and 128 neurons. The CNN’s confusion matrix showed that the network correctly classified the “Critical” and “Severe” classes.