Data-Driven Fault Detection in Turbojet Engines: Benchmarking Temporal SVM Against State-of-the-Art Models

Abstract

The detection of faults within turbojet engines represents an absolute necessity for operational safety and performance since failure to identify faults can lead to destructive system collapses. Real-time monitoring becomes possible through data-driven methods because they replace physical models which depend on system identification. The research presents T-SVM as a new framework for turbojet engine shaft speed fault detection through comparison with static SVM, XGBoost, Autoencoders and Long Short-Term Memory (LSTM) networks. The T-SVM algorithm uses Temporal Feature Enhancement (TFE) feature extraction technique with a 3-time-step delay to process experimental data containing an 18% fault rate. T-SVM demonstrates an accuracy level of 76.3% ± 23.5% and F1-score level of 0.51 ± 0.33 and AUC of 0.68 ± 0.27 during cross-validation with Bayesian hyperparameter optimization. This performance surpasses static SVM (59.4% ± 37.9% accuracy) by 16.9% but trails XGBoost (90.6% ± 1.3%), Autoencoders (90.7% ± 1.2%), and LSTM (90.7% ± 0.9%) due to their superior The T-SVM model shows high potential for real-time safety-critical usage because its true positive detection reaches 96.7% and its inference runs in under 0.01 seconds. Time-series analysis confirms how T-SVM identifies faulty sections when operating under varied conditions. The presented study demonstrates the trade-offs in fault detection based on data and recommends T-SVM as a powerful method which can be optimized by ensemble strategies to improve generalization capabilities.