CoAtNet-EMA Model for Stall-Onset Prediction of Airfoils Using Unsteady Surface Pressure Data
Authors
Files
Abstract
Airfoil stall onset is an unsteady aerodynamic transition caused by pressure redistribution, boundary-layer separation and vortex growth, and it can produce a rapid loss of lift before a conventional force-based alarm becomes decisive. This paper introduces a CoAtNet-EMA model for stall-onset prediction from unsteady surface pressure data. Pressure taps on the suction and pressure sides are organized into short pressure windows together with angle-of-attack information, Reynolds-number bands, pitching-rate tags and sensor-quality flags. Convolutional blocks extract local pressure-gradient changes in leading-edge and mid-chord regions, while attention blocks connect early pressure-recovery loss with subsequent fluctuation growth along the airfoil surface. An efficient multi-scale attention module recalibrates pressure channels according to the strength and reliability of stall evidence. A simulated multi-airfoil dataset containing 18,600 pressure windows is constructed under static-ramp, dynamic-pitching and noisy-sensing conditions. The proposed model achieves a stall-onset classification accuracy of 93.4%, increases early-warning lead time by 18.7% relative to the temporal convolution baseline and reduces false alarms under unseen Reynolds-number conditions from 10.8% to 6.1%. The output includes stall probability, confidence, risk level, dominant pressure region and suggested aerodynamic review action. These results indicate that pressure-window learning can provide earlier stall-onset warning than threshold-based lift monitoring while preserving interpretable evidence for wind-tunnel analysis, flight-test review and control-system evaluation.
