Predicting Surface Roughness in Cnc Milling Using Machine Learning With Three-Axis Vibration Data

İşci M.

9TH INTERNATIONAL PALESTRA SCIENTIFIC RESEARCH CONGRESS, Skopje, Makedonya, 29 - 31 Mayıs 2026, ss.1-5, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: Skopje
  • Basıldığı Ülke: Makedonya
  • Sayfa Sayıları: ss.1-5
  • Kayseri Üniversitesi Adresli: Evet

Özet

This study aims to estimate workpiece surface roughness using three-axis vibration data obtained from CNC milling operations. The dataset consists of vibration signals along the X, Y, and Z axes from an accelerometer mounted on the spindle, and Ra surface roughness values for each workpiece. In the experimental setup, machining conditions were kept constant; workpiece material was FDAC, spindle speed was 3000 rpm, feed rate was 800 mm/min, tool diameter was 6 mm, and the number of teeth was 4. The data used in this study were obtained from the dataset entitled “Roughness of Milling Process,” published by Yang on IEEE DataPort. The raw vibration signals were windowed, and features in the time and frequency domains were extracted for each window. Support Vector Regression (SVR) and Random Forest models were created to estimate surface roughness using the obtained features. A workpiece-based leave-one-out validation approach was used during model evaluation to prevent data leakage. The results showed that specific information regarding surface roughness could be extracted from the vibration data. The Random Forest model yielded the best results in terms of average workpiece-level error. When examined on a fold basis, the lowest workpiece-level RMSE value was 0.0447 in the 4th fold, and the highest RMSE value was 1.2704 in the 1st fold. In the SVR model, the lowest RMSE was 0.0635 in the 4th fold, and the highest was 1.0918 in the 8th fold. The results show that lower errors are produced, especially at medium Ra values, while the prediction deviation increases at low and high Ra values. Furthermore, feature selection revealed that features based on band energies of the X and Y axes and on combined vibration magnitude were more effective for prediction. In conclusion, it has been shown that three-axis vibration signals can be used to predict surface roughness in CNC milling. However, the limited number of independent workpieces made the model's generalization difficult. It is believed that studies conducted with larger datasets and different processing conditions could further improve prediction performance.