FG-DeformT: Grouped Deformable Temporal Modeling for Flight-Level Aircraft Anomaly Detection

Aircraft anomaly detection from multivariate telemetry time series is critical for flight safety and predictive maintenance. However, random window-level evaluation may overestimate performance because highly correlated windows from the same flight can appear in both the training and test sets. To obtain a stricter assessment of cross-flight generalization, this study adopts a leakage-aware flight-level validation protocol and proposes FG-DeformT, a grouped deformable Transformer for aircraft time-series anomaly detection. The model combines global temporal modeling with a grouped temporal offset mechanism to capture heterogeneous local temporal shifts across latent feature subspaces. Experiments on the ALFA dataset show that conventional baselines degrade sharply under flight-level evaluation, whereas FG-DeformT remains robust, achieving a Precision of 0.925, a Recall of 0.941, and an F1-score of 0.933 on the primary binary anomaly detection task. Ablation and downstream analyses further indicate that the grouped offset mechanism contributes to anomaly detection, while its benefit for fine-grained fault isolation is category-dependent rather than uniformly superior. External validation and computational analyses are included to further examine the framework’s applicability and practical feasibility. Overall, this study highlights the importance of leakage-aware flight-level evaluation and suggests that grouped temporal deformation offers a practical way to model heterogeneous temporal patterns in safety-critical telemetry data.