Prediction of micro-hardness in thread rolling of St37 by convolutional neural networks and transfer learning

Abstract

This study introduces a non-destructive method by applying convolutional neural networks (CNN) to predict the microhardness of the thread-rolled steel. Material microstructure images were collected for our research, and micro-hardness tests were conducted to label the extracted microstructure images. In recent years, researchers have used machine learning (ML) and deep learning (DL) models to predict material properties for forming, machining, additive manufacturing, and other processes. However, they encountered industrial limitations primarily because of the absence of historical information on new and unknown materials, which are necessary to predict material properties by DL models. These problems can be solved by employing CNN models. In our work, we used a CNN model with two convolutional layers and visual geometry group (VGG19) as transfer learning (TL). We predicted four classes of micro-hardness of the St37 rolled threads. The prediction results of the micro-hardness test images by our proposed CNN model and pre-trained VGG19 model are comparable. Our proposed model has produced the same precision and recall scores as VGG19 for class B and class C hardness. VGG19 performed slightly better than our model for precision in class A and recall in class D. We observed that the training time of our proposed model using the CPU (central processing unit) was approximately nine times faster than the VGG19 model. Our proposed CNN and VGG19 have direct applications in advanced manufacturing (AM). They can automatically predict the micro-hardness in the thread rolling of St37. Our proposed model requires less memory and computational power and can be deployed more efficiently than the VGG19 model.

CNN-based hot-rolled steel strip surface defects classification: a comparative study between different pre-trained CNN models

“…Transfer learning is a technique where knowledge acquired from previous tasks is leveraged to improve performance on a new task [30]. For instance, it involves utilizing the feature parameters learned from the ImageNet classification task and applying them to enhance the surface defects classification task.…”