Abstract: Robust and quantitative prediction of surface hardened layer depth of 45 steel using micro-magnetic testing method was studied considering the repeatability of multi-functional micro-magnetic instrument in measuring multiple magnetic features. First, the repeatability of instrument in measuring 41 magnetic features was evaluated based on the statistical method of coefficient of variation (β) of the test data. By combining the index of β and the sensitivity (S) of magnetic feature to the variation in hardened layer depth, the magnetic features were filtered. Second, models of feedforward neural network (FNN) were established fusing multiple micro-magnetic features. Modeling strategy for improving the robustness of the model and model robustness evaluation method were proposed. Finally, the effect of rules of input nodes elimination and reservation on the robustness of the model were discussed. Compared with the traditional modeling method, when eight magneric features were eliminated from the input nodes of FNN model obeying the proposed rule, the mean value of MAE and the number of models with MAE value less than 5% decreased by about 68.8% and increased by 150%, respectively. This indicates that the proposed modeling strategy can effectively improve the robustness of the instrument, which is used for quantitative prediction of the surface hardened layer depth of 45 steel.