Defect Recognition in Fiber-Optic Coil Winding Based on Improved MobileNetV1

An improved MobileNetV1 classification model is proposed in this study. An attention mechanism is introduced into the MobileNetV1. Flexibility and accuracy in identifying gap defect, stack defect, and normal state during the fiber-optic coil winding are demonstrated. Throughout the training process, a total of 480 images are used for training, 60 for validation and 60 for testing, respectively. The training and validation losses of the model, based on the improved MobileNetV1 by introducing the attention mechanism, are lower than those of other network models such as MobileNetV2, Resnet50, VGG16, Vision Transform, Swin Transform Base, Swin Transform Small, and Swin Transform Tiny. For gap defect, the precision reaches 95%, recall reaches 100%, and
${F}1$
-score reaches 97.43%. The precision for stack defect reaches 100%, recall reaches 100%, and
${F}1$
-score reaches 100%. In the normal state, the precision reaches 100%, recall reaches 95.24%, and
${F}1$
-score reaches 97.56%. The results indicate that the proposed algorithm effectively identifies gap defect, stack defect, and normal state during the winding process, enhancing the efficiency and accuracy of fiber-optic coil winding. This significantly reduces the cost of fiber-optic coil winding and holds crucial significance for achieving fully automated fiber winding machines. Contributions are made to advancing research in the field of artificial intelligence-based fully automated fiber winding machines, facilitating progress for other researchers in this domain.