Dynamic Machining Force Signal-Based Surface Quality Prediction in MAF Using XGBoost

Why It Matters

Accurate, in‑process surface‑quality prediction enables manufacturers to adjust cutting conditions on the fly, reducing scrap and downtime. The demonstrated superiority of force‑signal‑driven AI models signals a shift toward smarter, data‑rich machining environments.

Key Takeaways

• XGBoost model predicts surface roughness using real-time force data.
• Wavelet analysis isolates 0.5–1.5 kHz frequency band for force features.
• Model achieved R² of 0.96 and RMSE of 0.0138, outperforming parameter‑only model.
• Dataset comprised 54 full‑factorial experiments with three process variables.
• Band‑pass filtered force signal improves prediction accuracy over raw data.

Pulse Analysis

Machining forces have long been recognized as a primary driver of part quality, yet most production lines still rely on static process settings or post‑run inspections. By capturing force data at the spindle and translating it into a frequency‑domain representation, manufacturers gain a live window into the interaction between tool and workpiece. Continuous wavelet transform is particularly suited for this task because it preserves both temporal and spectral information, allowing engineers to isolate the 0.5–1.5 kHz band where the most informative force fluctuations occur.

The research team paired this signal‑processing pipeline with XGBoost, a machine‑learning technique celebrated for handling heterogeneous data and delivering high predictive power with limited tuning. Using a full factorial matrix of current, cutting gap, and silicone gel viscosity, they generated 54 distinct experimental runs, each providing synchronized force signatures and measured surface roughness. After band‑pass filtering the signals, the extracted wavelet power features served as inputs to the XGBoost model, which learned complex, non‑linear relationships that traditional parameter‑only regressions miss. The resulting model achieved an R² of 0.96 and an RMSE of just 0.0138 µm, indicating near‑perfect alignment with actual surface outcomes.

For industry, these findings illustrate a practical pathway to embed AI‑driven quality control directly on the shop floor. Real‑time predictions mean operators can tweak feed rates, currents, or coolant properties before defects manifest, dramatically cutting waste and rework costs. Moreover, the methodology scales: the same wavelet‑XGBoost framework can be adapted to other machining processes, materials, or even additive manufacturing, positioning it as a cornerstone of the next generation of smart factories. As manufacturers pursue tighter tolerances and faster cycle times, integrating dynamic force analytics will likely become a competitive differentiator.