Research on Surface Defect Detection in Submersible Pump Impellers Based on the Novel MXNet Deep Learning Model

In industrial production, surface defects of submersible pump impellers often have a significant impact on equipment performance and lifespan. However, the current defect detection is mainly performed manually, which is time-consuming, inefficient, and prone to human errors. Furthermore, the existing mainstream automatic detection methods have limitations, such as limited accuracy, complexity of the model, and low generalization, which make them difficult to promote widely. Therefore, an automated technology that can improve the efficiency and accuracy of the detection process is urgently needed. This study aims to develop an automatic defect detection method based on a Convolutional Neural Network (CNN) called MXNet to ensure superior product quality and reduce manual labor. This model, derived from the Xception architecture, has been specifically tailored for the task with custom convolutional and fully connected layers for precise classification, data augmentation techniques and an incorporated attention module to highlight vital regions in input images, coupled with the utilization of binary cross-entropy loss and custom callback functions, and a strategy for dynamically adjusting the learning rate. To demonstrate the superiority of this model, we compared it with four mainstream defect detection models. Experimental results show that the average accuracy and average recall rate of MXNet on the validation set reached 99.23% and 98.95% respectively, outperforming the comparative models in all performance indicators with shorter runtime and fewer parameters. This suggests that MXNet can not only achieve high-accuracy defect detection but also has high practicality, and is expected to have broad application potential in future in-dustrial production practices.