Multivariate non-destructive evaluation for tensile strength of steel based on neural network

Abstract

Tensile strength (TS) is an important mechanical property of a material. The conventional mechanical measurement method destroys the object under investigation; hence, the non-destructive evaluation of tensile strength of materials has become a research hotspot in recent years. Currently, there are some accuracy problems associated with evaluating the tensile strength of materials on the basis of single non-destructive testing (NDT) methods such as ultrasonic or electromagnetic methods. In this study, 45 steel is used as an example to study various non-destructive testing methods. First, seven different heat treatment systems are used to prepare standard specimens with different tensile strengths, which are measured by tensile tests. Second, non-destructive testing signals for each specimen are obtained as ultrasonic signals, magnetic Barkhausen noise and magnetic hysteresis signals, and the characteristic parameters of the signals are extracted. Then, single-parameter non-destructive evaluation (SNE) models of tensile strength with three different non-destructive testing methods are developed. Furthermore, a multivariate non-destructive evaluation (MNE) method based on ultrasonic signals, magnetic Barkhausen noise and magnetic hysteresis is proposed to improve the accuracy of the tensile strength measurements obtained from non-destructive testing. A deep residual network (ResNet) is used to combine the features of the three non-destructive testing parameters and an MNE model of tensile strength is developed. Moreover, a data pretreatment method based on the fuzzy mapping relationship is applied to train the MNE model successfully and enhance the stability, accuracy and reliability of the obtained results. Finally, the accuracies of the above four tensile strength evaluation models are confirmed by verification using the specimens. The results show that the MNE model has higher accuracy than the SNE models.