신경회로망과 CSM을 이용한 Ni-Ti 합금의 회전연마특성 예측모델 개발

Abstract

For decades demands for small components requiring high precision have been steadily increasing. Among them, the use of Ni-Ti alloy materials increases, and the need to improve surface roughness that affects the performance and quality of products is emphasized. However, the conventional finishing processing methods were limited in the size and shape of the workpiece. Hence, this study presents a magnetic transporter rotational finishing(MTRF) process using a magnetic field that was not restricted by the shape of the workpiece. This study aims to explore the optimal condition of the process parameter in the MTRF to improve the surface roughness of Ni-Ti alloy. The MTRF process simultaneously processes a plurality of complex-shaped components using a flexible tool. However, it is not economical in terms of cost and time to find the optimal surface finishing processing conditions by performing numerous experiments with an expensive Ni-Ti alloy material to reduce surface roughness. Hence, finite element analysis was employed to predict and score the surface integrity by considering the movement of the magnetic transporter in proportion to the abrasive motion. Since the magnetic transporter flow was affected by magnetic force, the magnetic flux density distribution was simulated using the magnetostatic simulation. Surface integrity was scored based on impulse, contact time, and contact distribution from the transient structural simulation results. Also, image classification was conducted by using a convolutional neural network(CNN) and the EfficientNet with compound scaling method(CSM). This study is significant for applying a predictive model that does not undergo trial and error to the simulation result image. The prediction results of these networks were compared to evaluate the surface integrity of Ni-Ti alloy in the MTRF process. As a result, the CSM model achieved excellent performance for classification accuracy with 98.4% and 93.8% of train datasets and validation datasets compared to the CNN model, respectively.