Small-Dataset Machine Learning for Wear Prediction of Laser Powder Bed Fusion Fabricated Steel

Abstract

The wear performance of an additively manufactured part is crucial to ensure the component’s functionality and reliability. Nevertheless, wear prediction is arduous due to numerous influential factors in both the manufacturing procedure and contact conditions. Machine learning offers a facile path to predict mechanical properties if sufficient datasets are available, without which it is very challenging to attain a high prediction accuracy. In this work, high-accuracy wear prediction of 316L stainless steel parts fabricated using laser powder bed fusion and in situ surface modification is achieved based on only 54 sets of data using a combination of an improved machine learning algorithm and data augmentation. A new modification temperature ratio was introduced for data augmentation. Four common machine learning algorithms and sparrow search algorithm optimized back propagation neural network were conducted and compared. The results indicated that the prediction accuracy of all algorithms was improved after data augmentation, while the improved machine learning algorithm achieved the highest prediction accuracy (R2 = 0.978). Such an approach is applicable to predict other systematically complex properties of parts fabricated using other additive manufacturing technologies.