An interpretable paradigm for predictive sensing of combustion flow features in a scramjet combustor

This study develops a real-time sensing framework for closed-loop active combustion control in reusable scramjets, addressing the critical bottleneck where existing methods cannot simultaneously balance prediction accuracy, computational efficiency, operational robustness, and physical interpretability. Through systematic pulse combustion wind tunnel tests at Mach 7.0 cruise conditions with perpendicular/parallel fuel injection (2–4 MPa, equivalence ratio 0.31–0.42), we establish a 19,050-sample multi-source dataset and physically validate 14-channel sparse wall pressure signals as effective proxies for high-dimensional combustion characteristics. A cascaded framework is proposed comprising: (1) a Sensor Failure Compensation Model (Model A) that corrects 1–7 channel faults within 0.0115 ms (1.57% test error), and (2) a Cross-time Feature Capture Model (Model B) achieving 23.74 dB PSNR/0.80 SSIM for 5–20 ms flow predictions at 5 ms/frame inference speed (7.2 × faster than UNet, 70.13 MACs). The system maintains cascaded stability under sensor failures while SHAP-based interpretability analysis quantifies six critical pressure points strongly correlated with flow discontinuities at combustor cavity slopes—providing direct guidance for sensor-array optimization. This integrated approach enables adaptive signal compensation, enhanced feature extraction, and advanced combustion dynamics prediction, breaking the real-time sensing bottleneck for unsteady reacting flows and establishing foundation for active control in the scramjet combustor. • Cascaded architecture enabling real-time sensor fault compensation and combustion flow prediction. • Wall pressure signals as physical proxies for unsteady heat release characteristics. • Interpretable identification of optimal sensor placement for combustion instability monitoring. • Embedded-ready framework with fault-tolerant design for harsh propulsion environments.