Proceedings Article | 27 September 2024
Liangliang Li, Yu Chen, Weitao Yuan, Kun Liang, Zhengxiang Ma, Xinling Wen, Jiabao Pang
KEYWORDS: Magnetism, Ferromagnetics, Performance modeling, Data modeling, Data conversion, Signal processing, Neural networks, Matrices, Inspection
In response to the challenges of large data volumes and insufficient accuracy of traditional prediction methods in defect size detection of ferromagnetic materials, this study proposes an optimized residual network model for precise prediction of defect sizes in ferromagnetic materials. The model enhances the ResNet18 network by integrating a channel attention mechanism, utilizing adaptive average pooling and fully connected layers to determine channel weights, and applying the Sigmoid activation function to process the input feature maps. Furthermore, by multiplying the obtained weights with the input feature maps, the model dynamically weights the importance of features along the channel dimension. In the experiments, one-dimensional signals were converted into two-dimensional grayscale images, and the features of these images were extracted to eliminate biases in manual feature selection. The results demonstrate that the optimized network achieved prediction accuracies of 95.65%, 100.00%, 97.01%, 78.07%, and 93.22% for defect sizes with a length of 50mm, depth of 2mm, and varying widths of 1mm to 5mm, based on 12,500 grayscale images. Similarly, for defects measuring 50mm in length, 2mm in width, and varying depths of 1mm to 5mm, the network attained prediction accuracies of 98.50%, 99.22%, 98.31%, 100.00%, and 99.22%, respectively, on a set of 12,500 grayscale images. Compared to the original ResNet18 model, the optimized model in this study exhibits significant improvements in prediction accuracy.
