Lyapunov–Granger Prior-Driven Spatiotemporal Anomaly Detection for Civil Aircraft Engines

With the rapid growth of civil aviation fleets, Quick Access Recorder (QAR)–based engine health monitoring is increasingly important for flight safety and predictive maintenance. However, aircraft engine parameters are nonlinear, strongly coupled, and sensitive to operating conditions, making traditional linear or static correlation methods insufficient for capturing phase-dependent relationships. To address this issue, this paper proposes a Lyapunov–Granger prior-driven spatiotemporal anomaly detection framework for civil aircraft engines. Phasewise largest Lyapunov exponent analysis is first used to screen parameters with similar nonlinear dynamic complexity, followed by Granger causality analysis to identify directional predictive dependencies. The constructed Lyapunov–Granger graph serves as a weakly physical and statistically supported prior for Graph Attention Network–based spatial aggregation. Meanwhile, a self-attention-enhanced Temporal Convolutional Network extracts temporal features, and a Gaussian hidden Markov model calculates anomaly scores from the learned spatiotemporal feature sequence. Experiments on real QAR data from two representative engine fault cases show that the complete Temporal Convolutional Network–Graph Attention Network–hidden Markov model framework outperforms ablated variants, demonstrating the contributions of both temporal representation learning and prior-guided spatial aggregation.