Título: Anomaly detection in cable placement in the power meter production process using Xception convolutional neural networks
Autores: Jonathan Cavalcante de Oliveira, Marcelo Chamy Machado, Emmerson Santa Rita da Silva & Marcio Aurélio dos Santos Alencar
Resumo: In this study, we explore the use of computer vision and deep learning techniques for anomaly detection in cable positioning in the power meters’ manufacturing process. The dataset used has been generated directly in the manufacturing process, which consists of 62 RGB images with dimension of 4000×1800 pixels, 45 images classified as normal and 17 classified as anomalies. The approach used in this work consists of using transfer learning from the Xception convolutional neural network (CNN) architecture together with data augmentation techniques and proposing a convolutional neural network model based on CNN Xception. To decide the best optimization, experiments were performed with different optimizer algorithms: Adam, RMSProp, SGD and AdamW, using 50 epochs to train and test the CNN architectures. A fully connected layer with 256 neurons was also added in order to improve anomaly detection. The results of the experiments showed that the best optimization algorithm of the Xception architecture was Adam, achieving an accuracy and recall of 99.84% with an additional layer of 256 neurons.
Palavras-chave: computer vision; deep learning; anomaly detection; manufacturing; transfer learning; data augmentation; convolutional neural network; Xception.
Páginas: 7
Código DOI: 10.21528/CBIC2025-1174203
Artigo em PDF: CBIC_2025_paper1174203.pdf
Arquivo BibTeX:
CBIC_2025_1174203.bib