EfficientViT-KD for Early Warning of Rocket-Engine Combustion Instability from High-Speed Pressure Signals

Abstract

High-frequency combustion instability can increase by a few acoustic cycles, and the conventional amplitude threshold will be reached too late. EfficientViT-KD is an edge-oriented early-warning model in this paper that converts 50 ms multichannel pressure windows into compact time-frequency token maps and transfers spectral, feature, and uncertainty knowledge from a high-capacity teacher. A total of 1,248 firing segments were collected at 200 kHz from a laboratory liquid-propellant combustor, and among them, 217 instability events and operating shifts in mixture ratio, throttle command, and injector condition were identified. A campaign-separated protocol avoided windows from the same firing being included in both the training and test sets. EfficientViT-KD reached an event-level F1 score of 94.1%, a receiver operating characteristic area (ROC) of 0.971, and had a median warning lead time of 118ms. At 3.8ms per window and 4.1 million parameters, the student achieved 99.1% of the teacher's F1 score and reduced latency by 80.7%. Calibration error decreased from 0.058 to 0.021, and the mean time between false alarms reached 15.8 firing-equivalent hours. According to the above results, frequency-aware distillation can maintain the structure of the precursor in a small transformer and offer a practical decision margin for closed-loop mitigation without needing engine-specific fixed thresholds.