Multi-Source Res2Net-CBAM Model for Digital-Twin Calibration in Wind-Tunnel Testing

Source abstract

Wind-tunnel measurements remain essential for validating aerodynamic models, but digital-twin calibration must reconcile evidence that differs in spatial density, uncertainty, and failure mode. This study proposes a multi-source Res2Net-CBAM framework that jointly processes balance coefficients, pressure-tap arrays, flow-visualization images, and simulation-prior fields. Source-specific encoders project the four inputs into a shared representation; hierarchical Res2Net blocks capture local and wake-scale structures, while sequential channel and spatial attention suppress irrelevant features. A reliability estimator uses branch residuals, source-availability indicators, and measurement-quality descriptors to normalize the contribution of each source before a residual head updates aerodynamic coefficients and pressure-related field descriptors. Evaluation on 1,680 synchronized operating-condition windows shows a lift-coefficient mean absolute error of 0.014, a drag-coefficient mean absolute error of 0.0021, and a reduction in pressure-field root-mean-square error from 7.9% for the image-only baseline to 3.6%. The mean twin-alignment score is 0.931, and inference remains below 18 ms per operating point under the reported comparison protocol. Ablation and missing-source tests indicate that reliability weighting provides the largest error reduction, followed by attention refinement and multi-scale grouping. The method is therefore suitable for between-run calibration review and digital-twin updating when measurement completeness varies. Its present evidence is limited to the tested operating envelope and does not establish cross-facility transfer or safety-critical real-time certification.