Answers are grounded in this paper and cite their sources.
Demonstrates uncertainty-aware prognostics in aircraft engines, indicating improved fleet safety management.
Aircraft engine prognostics is critical for safety, cost reduction, and efficient fleet management. Despite the availability of rich multimodal sensor data, most current prognostic approaches are limited to single-modality analysis, lack uncertainty quantification, and rely on predefined sensor graphs. To address these limitations, this paper introduces the Multimodal Spatiotemporal Graph Attention Network with Uncertainty Quantification (MST-GAT-UQ), a novel deep learning framework for aircraft engine prognostics. The proposed model integrates temporal patterns via CNN-LSTM networks, spatial correlations through graph attention mechanisms, and frequency signatures using Short-Time Fourier Transform (STFT) with 2D CNN processing. Key innovations include a correlation-driven graph construction method that automatically learns sensor relationships from operational data, hierarchical attention fusion to adaptively weight different modalities, and Monte Carlo dropout augmented by temperature scaling for simultaneous Remaining Useful Life (RUL) and health state prediction with calibrated uncertainty intervals. The framework is comprehensively validated on a real-world operational dataset comprising 84 sensors and 43,492 samples from F100-PW-229 turbofan engines, achieving R 2 = 0.911, RMSE = 0.087, MAE = 0.072 on a normalized RUL scale within a synthetic-label verification framework (no ground-truth operational RUL is available for this dataset; reported metrics quantify the model’s fidelity in reproducing a synthetically imposed linear degradation trajectory and should not be interpreted as ground-truth remaining-life predictions), and a synthetic health-state replication accuracy of 99.5% (auxiliary task; health-state labels are a deterministic step-function of the synthetic RUL signal and carry no independent diagnostic information). Comparative analysis against the standard NASA C-MAPSS benchmark is presented as a zero-shot domain-transfer diagnostic, characterising the label-distribution shift between the two datasets, and confirms that native C-MAPSS training yields positive R 2 across all four subsets (FD001–FD004: R 2 = 0.607–0.863). This work represents a significant step toward reliable, uncertainty-aware, and multimodal prognostic systems for safety–critical aviation applications, with uncertainty intervals that are directionally informative but moderately overconfident (empirical coverage 88.7% at the 95% nominal level; ECE = 0.254), indicating that further calibration refinement is required before safety–critical operational deployment.
No takes yet. Share an insight, caveat, or question.
Sławomir Szrama (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: