2020-10-272020-10-272020-02-03ROCHA, Rafael de Lima. Redes neurais convolucionais aplicadas à inspeção de componentes do vagão ferroviário. Orientador: Cleison Daniel Silva; Coorientadora: Ana Claudia da Silva Gomes. 2020. 78 f. Dissertação (Mestrado em Computação Aplicada) - Núcleo de Desenvolvimento Amazônico em Engenharia, Universidade Federal do Pará, Tucuruí, 2020. Disponível em: https://repositorio.ufpa.br/jspui/handle/2011/12790. Acesso em:.https://repositorio.ufpa.br/handle/2011/12790The railcar is one of the most important assets in a mining company, where tons of ore are transported daily by it, besides, the railcar can be used to transport people. Therefore, the inspection of defects in structural components of the railcar is a very important activity, making it possible to avoid problems in railway logistics, as well as to prevent accidents. The inspection task is performed visually by an operating technician who is exposed to accidents where the inspection is performed, in addition to the possibility of human error due to stress, fatigue, and others. The pad is a rail component analyzed in this work, where it is responsible for the primary suspension, a role that is important in the railcar dynamics. Thus, the purpose of this work is to use deep learning techniques, specifically convolutional neural networks (CNN) for the component inspection. CNN classifies the image of the structural component analyzed concerning the possible state it is in the railway, absent pad, undamaged pad, and damaged pad. Also, it intends to investigate the contribution of the component image in the frequency domain obtained through the magnitude and phase of the discrete Fourier transform (DFT) of the original image (spatial domain) in the CNN classification process. Histogram equalization and increasing the number of images through data augmentation techniques are also examined to evaluate their collaborations in improving classification performance. The results of CNN inspection of the pad prove to be quite inspiring, especially when the spatial component image is used together with the DFT magnitude image of the original image as CNN inputs, which are superior when only the original (spatial) image of the component is used, achieving a classification accuracy of 95.65%. In particular, the method that uses the increase in the number of training images by the data augmentation and the spatial domain and frequency (magnitude) images achieves the highest accuracy, with 97.47%, which represents approximately 385.5 correctly classified images from a total of 395.2 images.Acesso Abertohttp://creativecommons.org/licenses/by-nc-nd/3.0/br/Redes neurais (Computação)Aprendizado profundoFerrovias - Manutenção e reparosFerrovias - VagõesDissertaçãoCNPQ::ENGENHARIASDESENVOLVIMENTO DE SISTEMASCOMPUTAÇÃO APLICADA