Explainable Res2Net-SE Model for Quality Estimation of Robotic Screw Fastening
Authors
Recep UrfalıFiles
Abstract
Reliable quality estimation for robotic screw fastening is required to prevent loosening, thread damage, seal failure and late-stage product rejection caused by a single fastening error. This paper presents an explainable Res2Net-SE model for screw-fastening quality prediction based on torque-angle curves, axial pressure, spindle speed, depth displacement and visual inspection indicators. The fastening window is aligned at the torque-rise and seating stages, and multi-scale process features are used to represent peak torque, tightening slope, pressure stability, residual vibration and terminal depth. Res2Net blocks learn short-term impulses and gradual fastening trends within the same residual structure, while squeeze-and-excitation weighting enhances the signal channels that provide the most reliable evidence. The model outputs a quality score, confidence value, dominant evidence channel and review priority. A simulated robotic assembly dataset of 4,080 fastening windows was generated for normal fastening, under-tightening, over-tightening, thread slip, tilted screw entry and low-confidence mixed cases. The proposed model achieved 93.4% accuracy and 90.7% macro-F1, outperforming CNN, LSTM, Transformer and ResNet baselines. Torque noise and pressure drift still reduce recall for thread-slip and under-tightening cases. The results show that torque-angle data are more indicative of over-tightening, whereas pressure-depth data are closely associated with tilted entry and incomplete seating. Interpretable multi-scale feature learning can therefore support fastening quality assessment, rework prioritization and traceable robotic assembly inspection.
