Research on fault diagnosis of industrial materials based on hybrid deep learning model

Bearing fault detection is becoming more and more important in industrial development, and deep learning image processing technology provides a new solution for this. In this study, ResNet50 is used to replace VGG-16 as the feature extraction network of Faster R-CNN, and feature pyramid network (FPN) and parallel attention module (PAM) are introduced to achieve higher detection accuracy and speed. The experimental validation was conducted with the Case Western Reserve University bearing dataset using a three-fold cross-validation and compared with Yolov5, FPN, and the original Faster R-CNN model. The experimental results show that the accuracy of the proposed bearing image fault detection method is 78.6%, the accuracy is 77.4%, and the recall rate is 76.9%, which can locate and identify bearing faults more accurately. Future work could focus on further optimizing the model structure to enhance detection performance, strengthening the model's generalization ability to meet the detection requirements of different types of bearing faults.