Pneuma released 920 million time steps of real in-service commercial aircraft data, and even strong models still struggle with the subtle fault signatures
AspireAI community · arXiv
*Predictive maintenance for commercial aviation has moved forward mostly on proprietary data, which means the headline results cannot be reproduced and the hard cases cannot be shared.* Pneuma is the first publicly available large-scale benchmark built from onboard time series of in-service commercial aircraft, centred on bleed air systems, with over 920 million multivariate time steps across four aircraft types and nine fault scenarios verified against actual maintenance records. It is structured around three tasks - multivariate forecasting, supervised fault classification and unsupervised anomaly detection - and under unified settings more than 25 models from four methodological families are evaluated on nine curated subsets, with the finding that even strong contemporary models struggle with subtle fault signatures rather than the obvious ones. The two practical results worth carrying are that in-context examples help consistently and multivariate inference helps consistently, and that forecasting quality is strongly associated with anomaly detection performance, so a forecasting gain should not be read as a maintenance gain. The limitation is narrow by design: one system, four aircraft types and heavy class imbalance, which the benchmark surfaces rather than hides by setting a 10 percent recall floor on the warning class.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.