Improving weak magnetic detection of ferromagnetic material defects diagnostics via transfer learning-enhanced residual networks

The use of weak magnetic detection technology to identify defect sizes in ferromagnetic materials poses significant challenges due to the large volume of data and the relatively low prediction accuracy of conventional methods. Therefore, this paper proposes an improved Residual Networks (ResNet18) model that integrates transfer learning and channel attention mechanisms. Compared to traditional methods, the following improvements have been made: First, transfer learning has significantly reduced the data volume and time cost required for training from scratch, enhancing the model's generalization capability. Second, we have added a channel attention mechanism to the ResNet18 model, which involves calculating the importance of each channel through adaptive average pooling and fully connected layers, and generating channel weights using a Sigmoid function. This improvement allows the model to more accurately focus on features with higher relevance to defect sizes. Experimental results demonstrate that for grayscale images with defect lengths of 50 mm, depths of 2 mm, and widths of 1, 2, 3, 4, and 5 mm, the prediction accuracies reached 100 %, 100 %, 98.84 %, 99.58 %, and 100 %, respectively. For grayscale images with defect lengths of 50 mm, widths of 2 mm, and depths of 1, 2, 3, 4, and 5 mm, the prediction accuracies were 99.68 %, 100 %, 99.63 %, 100 %, and 99.60 %, respectively. Compared to the traditional ResNet18 model, the improved model not only enhances the accuracy of defect size prediction but also exhibits greater robustness, providing a new and effective method for defect classification in weak magnetic detection of ferromagnetic materials.