AI regression model based on effective abrasive force in MAF process for high-precision surface roughness prediction

Surface roughness is one of the most critical quality indicators in precision machining, as it directly affects product functionality, durability, and assemblability. Therefore, this study proposes a quantification method based on frequency response analysis to predict surface roughness improvement using abrasive force signals generated during the magnetic abrasive finishing (MAF) process. Raw abrasive force signals measured by the tool dynamometer were analyzed using fast Fourier transform (FFT) to identify the effective frequency band between 500 and 1,200 Hz. Subsequently, the real-time effective abrasive force was quantitatively extracted through inverse FFT (IFFT) and root mean square (RMS) calculations. The extracted effective force was used as an input variable in the AI-based second-order regression model, and its predictive performance was compared with a model using only process variables based on a total of 27 experimental datasets. As a result, the model incorporating real-time effective abrasive force demonstrated superior prediction accuracy with a coefficient of determination (\(\:{R}^{2}\)) of 0.9571 and a root mean square error (RMSE) of 0.0204. In addition, further validation experiments conducted under arbitrary conditions confirmed its high generalization capability with an average of 0.0117.