Random Forest Classification Model Based Predcitive Analysis For Abort of Engine Test Based On Vibrational Data.

This paper presents a data-driven predictive health monitoring model for real-time abort detection during liquid-propellant rocket engine static fire testing. Rocket propulsion systems exhibit localized structural resonances, combustion instabilities, and fluid-dynamic fluctuations before a catastrophic component failure occurs. Threshold-based monitoring, which only watches slow-moving variables such as chamber pressure or turbopump temperature, cannot capture these fast, non-linear precursors, so shutdowns are often triggered late and cause avoidable hardware damage.To address this, the framework converts raw high-frequency vibration telemetry into a compact spectral feature space using the Fast Fourier Transform (FFT), extracting seven descriptors per time window: RMS acceleration, peak amplitude, crest factor, kurtosis, skewness, dominant peak frequency, and spectral entropy. These are combined with six core process variables (chamber pressure, oxidizer and fuel mass flow, turbopump RPM, thrust, and chamber temperature) into a single 13-dimensional feature vector per timestep. A Random Forest classifier is trained on this feature vector to predict whether a given moment of operation is Nominal or heading toward an Abort. On held-out telemetry, the model correctly separates nominal operation from developing anomalies, and its Gini feature-importance ranking shows that spectral entropy and peak frequency — both derived purely from the FFT of the vibration signal — are the strongest predictors, ahead of any single raw process variable. This confirms that failure precursors show up in the vibration spectrum earlier than in slow process measurements, and demonstrates a practical path toward automated, predictive (rather than purely reactive) test-stand abort systems.