Online monitoring and prediction for surface roughness in rotational electro-magnetic finishing using acoustic emission and vision-based neural network

The importance of surface finishing processes and accurate surface quality prediction models has increased in response to the growing demand for improved surface finish in ultra-precision applications. To enhance process efficiency and develop accurate predictive models, numerous studies have investigated the monitoring and prediction of surface roughness. However, existing mathematical approaches encounter challenges in establishing the correlation between input and output variables and providing real-time surface status monitoring. Therefore, this study aimed to monitor and predict surface roughness in real-time for the rotational electro-magnetic finishing (REMF) process using acoustic emission (AE) signals. First, a total of 72 fundamental experiments were conducted based on the mixed orthogonal array L-18( 2(1x) 3(4)) to determine the optimal configuration for achieving a high-quality surface. The results revealed that the best combination was achieved with an abrasive length of 3.0 mm, an abrasive diameter of 0.7 mm, a total abrasive weight of 2.0 kg, a rotational speed of 1800 rpm, and a working time of 10 min. To analyze signal features and develop an accurate surface prediction model, a convolutional neural network (CNN) was suggested, utilizing scalogram images as time-frequency characteristics of AE signals. The suggested model demonstrated outstanding quantitative results compared to those of the regression model, with training coefficient of determination (R2), mean squared error (MSE), and F- test of 0.986, 0.19x 10(-3), and 99%, and testing R-2, MSE, and F-test of 0.951, 2.23x 10(-3), and 99%, respectively. In addition, the suggested model showed good generalization ability with a relatively lower mean MSE of 0.003 through verification experiments. These results demonstrated that the sensory data and image-driven model were effective in real-time monitoring and surface roughness prediction in the REMF process with high accuracy and reliability.