Property Prediction Based On Machine Learning And Database Platform Of Cr-V Wear-resistant Steel

60%of mechanical equipment fails due to wear,and improving wear resistance to extend the life of components is of great significance to improve the economy of mechanical equipment.High chromium alloy steel is a widely used wear-resistant material,with the main wear-resistant phase being M₇C₃,and the addition of vanadium elements to precipitate higher hardness VC is an important way to further improve its wear resistance.Currently,there is no method to accurately calculate the optimal ratio of chromium to vanadium content in this new Cr-V wear-resistant alloy steel.In addition,there are differences in the requirements for the properties of wear-resistant materials under different working conditions.There are important engineering applications in the design and development of cost-effective wear-resistant materials that meet the requirements of use and are low-cost.The traditional trial-and-error method is difficult to extract the key factors affecting the properties,and can not achieve the"on-demand design"by accurately adjusting the corresponding compositions to obtain different property.The data-driven research method takes machine learning（ML）as means to establish structure-activity relationship from compositions to property,extract key features.Then,the optimal compositions that meet the target property based on the predicted results is verified,providing a new and effective way for efficient"on-demand"development.The limited number of high-quality homologous samples,the large differences in the feature distributions of homologous samples,and the lack of features at the microstructure scale present difficulties for ML based property prediction studies at this stage.Aiming at the above issues,the research background of this paper is the green and efficient design and development of Cr-V wear-resistant steel using data-driven methods.The feature distributions of the samples are analyzed and methods for direct modeling of homologous samples and transfer learning modeling of non-homologous samples are proposed.Reducing the discrepancy of distributions,extracting and constructing key features,predicting properties and matrix are then studied systematically in depth,respectively.The database platform is implemented and the efficient design of low vanadium wear-resistant steel can be guided.The main works and results of this paper are the following:1.A sample partitioning method based on weighted fuzzy C clustering was proposed to build a hardness prediction model for wear-resistant steel.In this approach,the Mahalanobis distance was weighted by the feature importance from the random forest,and the problem of amplifying the effect of trace alloying elements due to the inverse covariance matrix was addressed,which reduced the distribution discrepancy of important features for samples in the same cluster.The hardness prediction model was built on the set of post-clustered samples,and the accuracy of the model could be significantly improved by taking the Top10 features as input.Another set of samples not involved in the training was chosen as the validation set,and the corresponding model was used to predict hardness.The predicted R was above 0.99 and the RMSE was 1.06 HRC,which proved the generalization ability of the model.2.A few-shot guided transfer component analysis was proposed to address the modeling problem of Cr-V wear-resistant steel in the presence of scarce homologous samples.The samples of source and target domains had great distribution similarity on the constructed（Cr+V）/C,which could narrow the difference of marginal probability distribution.As the number of guided samples increased,the conditional probability distribution of source domain aligned with that of the target domain.The predicted hardness of the proposed method were 46%higher in R,and 91%lower in RMSE than those of the comparison model.（Cr+V）/C,V/Cr and hardness were the key features that affect the wear resistance,and the abrasion loss under the same abrasive wear condition could be predicted accurately.The analysis of the relationship between the features and the abrasion loss shown that the sample with larger V/Cr had better wear resistance,whose（Cr+V）/C fell in the design sweet zone[3,6].3.A multi-source domain transfer component analysis method was proposed.The fitness parameter was derived from the maximum mean discrepancy to distinguish the effect of each source domain on the hardness prediction model and synthesize the predicted results,which solved the start-up problem for the design of new wear-resistant steel without homologous samples.The proposed method achieved a higher prediction accuracy with a 39%reduction in MAE compared to the single source domain.A qualitative prediction method of abrasion loss based on（Cr+V）/C and V/Cr was proposed to rapidly screen candidate samples with the best wear resistance,and the wear-oriented alloy design criteria was obtained to provide guidance for the composition design of Cr-V wear-resistant steel.4.A database platform was developed to provide data collection,management,analysis and computing for wear-resistant steel.Based on the experiment data stored in the platform and the proposed property prediction model,a design process for the optimal composition of 3 % V wear-resistant steel was developed.It was found that samples generally had higher hardness in the search space defined by the wear-oriented alloy composition design criteria.Samples with Cr content in the range of 3 to 4 percent had the smallest abrasion loss.Experiments had shown that the measured properties of the samples with the optimal composition agreed with the predicted values,and the efficient design and property evaluation of Cr-V wear-resistant steel had been initially realized.