Industrial Automatic Control Systems and Controllers Annotation << Back A Method for Synthesizing Specialized Neural Network Models for Predicting Roughness in end Milling based on the Vibroacoustic Response of the Technological System Dergunov A.A., Makarov M.A., Kholopov V.A. The paper discusses a method for synthesizing specialized neural network models for predicting surface roughness during end milling based on the vibroacoustic response of the process system. It is proposed to use a combination of the vibroacoustic response recorded during the machining of a test part using industrial vibration diagnostics, telemetry data from a numerical control device, and static process parameters as a source of indirect informative features reflecting the dynamic state of the process system. The method involves pre-training the neural network model on an extended dataset, including various combinations of cutting modes and process conditions, followed by further training on a local, automatically collected experimental dataset obtained for a target combination of equipment, cutting tool, workpiece material, and clamping arrangement. As a result of further training, a specialized roughness prediction model is synthesized for specific process conditions. Experimental verification was performed during shoulder milling with end mills on a CNC vertical milling machining center. To evaluate the method's performance, the predictive performance of the pre-trained model M0 and the specialized model MS was compared. The results showed that after retraining, the average absolute forecast error decreased from 1.12 to 0.30 μm, and the maximum deviation decreased from 1.96 to 0.72 μm. It was also found that increasing the local dataset size for retraining improves the model's accuracy up to the saturation point, after which further expansion of the sample does not lead to a significant reduction in error. Keywords: end milling, roughness, roughness prediction, specialized neural network model, retraining, vibroacoustic response, vibration acceleration, acoustic emission. DOI: 10.25791/asu.7.2026.1673 Pp. 40-48.