Abstract: The surface hardness is important index for the manufacturing quality evaluation of ferromagnetic materials and the non-destructively quantitative testing of surface hardness is one of the hot topics in the field. Magnetic Barkhausen noise (MBN) measurement technique is applied for quantitative evaluation of surface hardness in S136 steel. A self-developed device is used to conduct MBN measurements in sixty specimens of different surface hardness. The coefficient of variation of the repeated measured magnetic parameters is analyzed to show that the device has good repeated accuracy for MBN measurements. To achieve quantitative characterization of the surface hardness using magnetic parameters, the dependency of the surface hardness on six parameters extracted from the measured MBN and tangential magnetic field signals is individually analyzed. More importantly, comparative study is performed among the simple linear regression model, multiple linear regression model and BP neural network model to evaluate their accuracy in surface hardness prediction. The results show that the BP neural network model employing six magnetic parameters as inputs has highest accuracy in surface hardness prediction. A total of 300 cases are investigated and it is found that the average prediction error is only around 2.14% and the maximum error is about 11.74%, and the prediction errors out of 274 of 300 cases are less than 5.00%. A reference is provided for realizing the nondestructive quantitative detection of steel plate surface hardness based on magnetic Buckhausen noise method.

Key words: Barkhausen noise, surface hardness, BP neural network, quantitative prediction