Improving weak magnetic detection of ferromagnetic material defects diagnostics via transfer learning-enhanced residual networks
Year: 2025
Type: article
Abstract: The inspection of defect size in ferromagnetic materials poses significant challenges due to the massive volume of data and the inadequacy of conventional prediction techniques in terms of accuracy. This paper presents an improved residual network model incorporating channel attention mechanisms to predict the defect dimensions in ferromagnetic materials more precisely. The model enhances the ResNet18 architecture by introducing channel-wise attention mechanisms that assess the significance of each channel. Utilizing adaptive average pooling and fully connected layers, channel importance is computed and the Sigmoid activation function is applied to the input feature maps. The calculated weights are then multiplied by the input feature maps to emphasize feature importance on the channel dimension. The study transforms the obtained one-dimensional signals into two-dimensional grayscale images, subsequently extracting features from these images to mitigate the effects of manually selected features. Experimental results demonstrate that the modified ResNet18 network significantly increases prediction accuracy on a dataset of 10,000 generated grayscale images representing defects of 50mm length and 2mm depth with widths of 1, 2, 3, 4, and 5mm, achieving accuracies of 100%, 100%, 98.84%, 99.58%, and 100%, respectively. Similarly, for defects of 50mm length and 2mm width at depths of 1, 2, 3, 4, and 5mm, the model achieved prediction accuracies of 99.68%, 100%, 99.63%, 100%, and 99.60%. These results indicate a marked improvement in prediction accuracy when compared with the original ResNet18 model, confirming the effectiveness of the proposed channel attention-enhanced residual network in defect size prediction in ferromagnetic materials.
Authors: Yu Chen, Liangliang Li, Zhengxiang Ma, Xinling Wen, Jiabao Pang, Weitao Yuan
Language: English
Subfield: Mechanical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,200
OpenAlex ID: W4408918692