Time-Series Foundation Model Embeddings for Remaining Useful Life Estimation

This paper from Friedrich-Alexander-Universität Erlangen–Nürnberg and Nokia Solutions and Networks shows that a frozen time-series foundation model can become a strong Remaining Useful Life estimator with minimal task-specific training. The authors feed resampled multivariate industrial sensor windows into Chronos-2, extract context embeddings, and train only a lightweight MLP regression head, rather than fine-tuning the backbone. On two real-world Nokia device datasets, the approach outperforms linear regression, random forests, gradient boosting, GRU, LSTM, Temporal CNN, and Transformer baselines under the same preprocessing and chronological split protocol. For 5-step windows, it reaches MAE 44 and MSE 6513 on Device A, and MAE 64 and MSE 7212 on Device B, while the strongest non-TSFM baseline remains at MAE 88 on Device A. The study also shows that context length is critical: increasing the Chronos-2 lookback to 80 steps sharply improves accuracy, reducing MAE by about 2× versus a 5-step context before saturating. An ablation confirms that the pretrained embeddings carry most of the signal: a linear head still achieves MAE 60 on Device A, while the default 2-layer MLP is best at MAE 44, with little benefit from a deeper 4-layer head. The result is a data-efficient RUL pipeline that turns a forecasting foundation model into an industrial prognostics feature extractor.