NGAFID-Driven Whole-Airframe Health Analytics for Real-World Aircraft via Information-Propagation-Constrained Attention

Real-world flight profiles are high-noise, low-signal and strongly-coupled multi-task time series where weak fault signatures are easily drowned by long-range operational variations. We introduce an information-propagationconstrained attention framework that limits self-attention receptive fields to suppress cross-stage noise while retaining local fault cues. The architecture first tokenises each profile via a lightweight convolutional tokenizer (ConvTok) embedding local shape, statistics and positional codes; tokens are then processed by an AnyAttn backbone whose multi-head self-attention switches between global, sliding-window (SWLA), multi-window (MWLA) or log-parse local (LPLA) modes via learnable band masks. Experiments on the National General Aviation Flight Information Database (NGAFID) - 28,000 Cessna-172 flights covering 36 maintenance classesshow that global attention achieves 79.29 % anomaly-detection accuracy, while MWLA with 3-token windows raises fault classification F1 from 34.46 % to 51.76 %, verifying that restricted receptive fields enlarge effective signal-to-noise ratio under label noise and few-shot constraints.

Keywords: Aviation health monitoring, real-world aviation dataset, time-series classification, anomaly detection, fault classification, information-propagation constraint, local self-attention.