This work presents a protocol-aware empirical assessment across three settings: a C-MAPSS degradation-risk proxy, normal-only training for anomalous-sound detection on MIMII, and BDG2 forecasting-residual diagnostics with synthetic target perturbations.
Abstract
Industrial monitoring models must detect operationally relevant deviations while satisfying target-specific data, calibration, and resource constraints. Time-series foundation models (TSFMs) promise reusable representations and zero-shot forecasts, yet evidence for their deployment value remains mixed when task definitions are heterogeneous and lightweight baselines are competitive. This work presents a protocol-aware empirical assessment across three settings: a C-MAPSS degradation-risk proxy, normal-only training for anomalous-sound detection on MIMII, and BDG2 forecasting-residual diagnostics with synthetic target perturbations. We assess classical one-class methods, compact neural autoencoders, residual forecasters, MOMENT-small, Chronos-T5, and TimesFM 2.5 in terms of anomaly-ranking performance, risk-horizon sensitivity, residual forecasting and perturbation sensitivity, and local implementation cost. Across 100 C-MAPSS engines evaluated out of fold, TCN-AE reaches fold-weighted AUROC/AUPRC 0.9570/0.8960, compared with 0.7310/0.3080 for MOMENT reconstruction; paired engine-cluster bootstrap confidence intervals exclude zero for both differences. Across five matched MIMII pump evaluations, OCSVM also exceeds MOMENT reconstruction in AUROC and AUPRC. On a fixed 12-meter BDG2 panel, TimesFM 2.5 has the lowest aligned forecast error and the highest synthetic AUROC point estimate, although synthetic AUPRC is similar across TSFM and fitted residual models. Same-device measurements show that MOMENT incurs higher latency, peak allocated VRAM, and serialized state-dictionary size than TCN-AE. Under the evaluated frozen and zero-shot settings, TSFMs are task-dependent deployment options rather than default replacements for fitted lightweight models.
Ship fuel consumption (SFC) prediction supports vessel operation optimisation, emissions estimation, and decision support systems (DSS) for sustainable maritime transportation. Numerous data-driven fuel models have been developed over the past two decades, but a critical and often overlooked limitation lies in their validation practices: most studies evaluate performance using random train--test splits, which, applied to high-frequency records, admit temporal leakage and yield optimistic results that do not reflect deployment conditions. This paper examines that gap using time-aware evaluation, specifically Time Series Cross-Validation (TSCV) and Blocked TSCV (BTSCV). Using the Canadian Coast Guard Ship (CCGS) \textit{Sir Wilfrid Laurier} as a case study, six regression models and a physics baseline are tuned under three time-aware schemes and three feature configurations, then evaluated on a common chronological hold-out set drawn from approximately 3.88 million steady-state 1\,Hz records.
Samarasimha Reddy Chittamuru, Ayhan Akinturk, A. Kennedy et al.· 0 citations
The Continuous-time Squared Error (CSE) is proposed, which employs importance weighting to eliminate the influence of the timestamp sampling distributions and theoretically proves that CSE's asymptotic estimation error with respect to continuous-time risk is no greater than that of MSE.
Rong Li, Haixin Xie, Xiao Wang et al.· 0 citations
Remaining Useful Life (RUL) prediction for turbofan engines is critical for balancing operational safety against maintenance costs and environmental impact from premature replacements. Models must reliably extrapolate beyond training data, yet no single method performs optimally across all operational contexts, making model selection fundamentally heuristic. Rather than demonstrating individual model superiority, this work recognizes that reliable prognostic performance emerges from adaptive model contribution. This work introduces an ensemble framework integrating three components: (1) legacy-representative data splitting that mirrors realistic deployment where models predict for newer assets using historical data, (2) median absolute deviation-based outlier filtering for stability, and (3) WTA³ weighting that dynamically adjusts model influence based on cycle-by-cycle performance. This treats model coordination as a time-varying optimization problem adapting to evolving degradation patterns. The framework is validated on NASA CMAPSS data using six diverse models spanning traditional machine learning (Random Forest, XGBoost, Support Vector Regression) and deep learning (LSTM, CNN, Transformer). We investigate how prediction robustness changes during temporal extrapolation, whether adaptive weighting provides more stable forecasts than individual models or fixed combinations, and how uncertainty quantification supports safer maintenance decisions. Initial validation demonstrates that the WTA³ meta-ensemble achieves approximately 3 cycles RMSE and under 3 cycles MAE, representing over 30% improvement over the best individual model, with particularly strong performance in the critical late-life phase. The ensemble maintains highly stable predictions with well-calibrated confidence intervals and substantially improved coverage compared to individual models. This work reframes prognostics from competitive model selection to adaptive coordination, demonstrating that ensemble stability under extrapolation can be systematically achieved. The approach provides actionable confidence bounds for aerospace maintenance programs, enabling cost-efficient scheduling while reducing environmental waste, directly supporting both economic and sustainability objectives where data scarcity and safety criticality are paramount.
Ayushi Bharti, N. Kim, Hee-Cheol Kim· e-Journal of Nondestructive...· 0 citations
Underprediction of high electrical demand can be more operationally consequential than an equally sized overprediction, yet standard point forecasting models are optimized primarily for average error. This case study evaluates a multivariate quantile Transformer as a safety-oriented next observation forecasting layer for a university laboratory. A timestamp-level audit identified 33,374 native measurements collected from 22 April to 12 December 2024 at a median interval of approximately 10 min. The final leakage-free pipeline uses only real observations, performs the chronological split before sequence generation, fits all scalers on training data only, and rejects windows containing gaps greater than 30 min. Persistence, fixed-order SARIMA, LSTM, GRU, CNN–LSTM, and an MSE-trained Transformer were evaluated on the same 6595-sample test period. GRU achieved the best deterministic accuracy (MAE 0.017744 kW; RMSE 0.023278 kW), whereas the proposed τ = 0.90 Transformer intentionally traded point accuracy (MAE 0.033830 ± 0.001133 kW) for asymmetric risk control. Across five independent runs, it achieved a pinball loss of 0.004453 ± 0.000069 kW, empirical coverage of 87.95 ± 1.01%, and a peak underprediction rate of 26.64 ± 5.32%, compared with 72.94–100% for the conventional benchmark outputs. Additional τ = 0.75 and τ = 0.95 experiments demonstrate the expected accuracy–safety trade-off. MAPE is not used as a primary metric because near-zero loads make percentage errors unstable. The results support the proposed model as a complementary upper quantile forecasting layer for this small, dynamic facility; they do not establish general performance at feeder or system scale.
Marwa O. Al Enany, Mazen Hesham Elnahal, Amira M. Gaber· Computers· 0 citations
This work explores the application of pre-trained time-series foundation models (FMs) for detecting anomalies in industrial processes and introduces a new time-series forecasting method that filters out suspicious data and uses previously predicted data as input, called Forecast Fallback (FF).
Short-term industrial energy forecasting supports load planning only when every predictor is available at forecast issuance. This study audits 15- and 60-minute forecasting with 35,040 real observations from a South Korean steel facility. We reconstruct a continuous 15-minute timeline, define a deployable 39-feature protocol using measurements available no later than the forecast origin, and contrast it with two diagnostic protocols that admit target-time sensors. Ridge regression, Random Forest, and XGBoost are compared with persistence and daily and weekly seasonal-naive baselines under chronological and random 70/15/15 train/calibration/test partitions. The clean chronological Random Forest achieved mean absolute error (MAE) of 3.829 kWh at 15 minutes and 7.484 kWh at 60 minutes, reducing MAE relative to persistence by 23.35% and 35.31%. Paired moving-block bootstrap intervals excluded zero for every comparison with the operational baselines. By contrast, target-time measurements reduced XGBoost MAE by 82.72% and 90.12%, showing that sub-1-kWh results describe contemporaneous estimation rather than deployable forecasting. Target-time CO2 was a strong proxy: its correlation with energy use was 0.988, and 97.69% of values equaled rounded usage multiplied by 0.00045. Split-conformal intervals achieved near-nominal marginal coverage, yet 90% peak-load coverage fell to 51.99% and 44.70%. The audit shows that feature-availability semantics and conditional uncertainty assessment affect interpretation more than the choice between strong tree learners.
Esam Miftah Abdulnabi, Nabeel Faraj Amhimmid, Ashraf Faraj Saed Albarki et al.· Libyan Journal of Applied an...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
