Oversampling Application of Identifying 3D Selective Laser Sintering Yield by Hybrid Mathematical Classification Models

Abstract

Selective laser sintering (SLS) is one of the most popular 3D molding technologies; however, the manufacturing steps of SLS machines are cumbersome, and the most important step is focused on molding testing because it requires a lot of direct labor and material costs. This research establishes advanced hybrid mathematical classification models, including random forest (RF), support vector machine (SVM), and artificial neural network (ANN), for effectively identifying the SLS yield of the sintering results from three sintered objects (boxes, cylinders, and flats) to achieve the key purpose of reducing the number of model verification and machine parameter adjustments, thereby saving a lot of manufacturing time and costs. In the experimental process, performance evaluation indicators, such as classification accuracy (CA), area under the ROC curve (AUC), and F1-score, are used to measure the proposed models’ experience with practical industry data. In the experimental results, the ANN gets the highest 0.6168 of CA, and it is found that each machine reduces the average sintering time by four hours when compared with the original manufacturing process. Moreover, we employ an oversampling method to expand the sample data to overcome the existing problems of class imbalance in the dataset collected. An important finding is that the RF algorithm is more suitable for predicting the sintering failure of objects, and its average sintering times per machine are 1.7, which is lower than the 1.95 times of ANN and 2.25 times of SVM. Conclusively, this research yields some valuable empirical conclusions and core research findings. In terms of research contributions, the research results can be provided to relevant academic circles and industry requirements for referential use in follow-up studies or industrial applications.