Dimension-Aware NAS for Cryocooler Lifetime Prediction with Small-Data Models
A new paper that appeared on arXiv (2608.06993) presents the FSD-RM (Family of Small-Data Representation Models) framework, which offers an alternative to large-scale pretrained time-series models, particularly for industries and scientific areas with limited data. The authors believe that the benefits of large-scale pretraining are compromised due to insufficient diverse data in these fields. They recommend using capacity-controlled representation learning with popular encoder architectures like CNN1D, LSTM, GRU, and Transformer, as these work well with small data and are interpretable. The encoders are trained unsupervised on multivariate telemetry data and used in a two-stage process to forecast lifetime outcomes, focusing on cryocooler lifetimes, essential for space missions and industrial cooling. The paper is categorized as a cross-type submission on arXiv, suggesting it might have been presented elsewhere.
Key facts
- Paper arXiv:2608.06993 proposes FSD-RM paradigm for small-data representation learning.
- Focuses on cryocooler lifetime prediction using telemetry data.
- Uses CNN1D, LSTM, GRU, and Transformer encoders.
- Employs dimension-aware neural architecture search (NAS).
- Encoders are trained unsupervised on multivariate telemetry.
- Two-stage pipeline for downstream lifetime prediction.
- Alternative to large-scale pretrained time-series models.
- Aims for interpretability and suitability in small-data settings.
Entities
Institutions
- arXiv