Micromagnetic and Quantitative Prediction of Yield and Tensile Strength of Carbon Steels Using Transfer Learning Method

Abstract

This study investigates the correlation between various micromagnetic signature patterns and the yield and tensile strengths of carbon steel (Cr12MoV steel as per Chinese standards). For this purpose, back-propagation neural network (BP-NN) models are established to quantitatively predict the yield and tensile strengths of carbon steels. The accuracy of prediction models is significantly affected by the presence of redundant micromagnetic signature patterns. By carefully screening the input parameters, it is able to effectively mitigate prediction errors arising from unreasonable model inputs. In the field of micromagnetic nondestructive testing (NDT), prediction models calibrated for a specific instrument or sensor cannot be directly applied to another instrument or sensor. In the study, a joint distribution adaptation transfer learning strategy based on auxiliary data is proposed to enhance the generalization of prediction models for cross-instrument applications. When auxiliary data accounts for 30% of the source domain data, the joint distribution adaptation transfer learning method based on auxiliary data improves the robustness of the model. The accuracy of the yield strength and tensile strength calibration models witnesses remarkable improvements of approximately 91.4% and 93.5%, respectively.