Resource-Aware Model Comparison for DASHlink Multi-Class Flight-Anomaly Screening

Abstract

Flight-data prognostics and health management models have to do more than rank anomalous windows correctly. For future edge or onboard validation, a screening model also has to be compact, fast to score, false-alarm aware, and interpretable enough that its decision statistic can be trusted. Autoencoders are attractive in this setting because they can be trained from nominal data, but a single scalar reconstruction error may be too coarse to separate operationally different non-nominal approach behaviors. We tested that tradeoff on the NASA DASHlink multi-class flight-anomaly window benchmark using a fixed split, nominal-only threshold calibration, and a held-out test set. Each sample is a 160-second, 20-variable window labeled as Nominal, Speed High, Path High, or Flaps Late Setting. We treat the primary task as nominal-versus-non-nominal screening, while preserving class-wise recall so that the rare classes are not hidden by aggregate recall. The comparison includes raw-summary classifiers, classical unsupervised baselines, scalar convolutional variational autoencoder (CVAE) reconstruction scores, CVAE residual and latent representations, and raw-plus-CVAE hybrid models. Thresholded metrics use a nominal-calibrated p95 operating point; resource proxies include total scoring time and serialized pipeline size. Scalar CVAE reconstruction error was weak as a standalone detector, with AP 0.179 and worst-class recall 0.051. Residual and latent CVAE features retained substantially more class-discriminative information, and the strongest raw-plus-CVAE hybrid reached AP 0.935, recall 0.945, and worst-class recall 0.894 at 0.158 ms/window. Full-20 raw-summary HGB remained the simpler high-performing baseline. The practical result is that compact feature-based models are strong screening baselines, while reconstruction models are most useful here as auxiliary representation generators rather than standalone scalar anomaly detectors.