Property prediction for high-chromium high-vanadium steel based on transfer component analysis with few-shot guided

Abstract

The transfer learning model improves accuracy by reducing the marginal and conditional probability distribution discrepancy between source and target domains. Based on the hypothesis of the ideal carbides of high-chromium high-vanadium steel, a mass fraction ratio written as (Cr + V)/C is deduced as vital feature to narrow the marginal probability distribution discrepancy. To align the conditional probability distribution of the source domain with the target domain, a few-shot guided transfer component analysis (TCA) method is proposed that a limited number of labeled samples taken from the target domain are used to guide the mapping. Then, the V/Cr combines with the optimal (Cr + V)/C is proposed to predict the composition of sample with the best wear resistance. Experimental results show that the proposed few-shot guided TCA method can considerably improve the prediction accuracy (R is higher than 0.99, RMSE is lower than 0.63HRC). The constructed (Cr + V)/C is the most critical feature. In addition, the predicted sample consisting of 2.1%C, 4%Cr, 4%V and 1.5%Mo has the best wear resistance with minimal abrasion weight loss in the test.

Keywords

• TCA
• Feature engineering
• Few-shot
• Wear resistance