Answers are grounded in this paper and cite their sources.
Convolutional recurrent networks are widely reported as being able to forecast process temperature in machining accurately. Many such evaluations, however, split overlapping sliding windows at random, which lets near-duplicate samples cross the train–test boundary. We re-examine process temperature forecasting and soft sensing on the AI4I 2020 benchmark under a leakage-controlled protocol: a chronological split with a purge gap, training-only scaling, five seeds, block-bootstrap confidence intervals and Diebold–Mariano tests with an autocorrelation-robust variance. LSTM, GRU, CNN-LSTM and an attention-augmented CNN-LSTM are compared with persistence, linear, random forest and multilayer-perceptron baselines that receive the same ten-record window. Under a random split the attention model reaches R2=0.899±0.006; under the chronological split it falls to 0.305±0.089. Rolling-origin cross-validation, a purge-gap sweep, a factorial ablation and sensitivity analyses support this picture. In soft sensing, the error is dominated by a drifting process–air temperature offset that the machine signals in this benchmark do not observe, and refitting on recent data matters more than architecture. When lagged process temperature is available, persistence (MAE 0.054 K) is not beaten at horizons of 1–30 records, and convolutional and attention stages add no benefit while costing up to 2.4 times the training time of a plain LSTM. On this benchmark, the evaluation protocol changed the conclusions more than did the choice of architecture. Claimed gains of deep sequence models for machine monitoring should therefore be reported against persistence and information-matched baselines under a time-ordered split.
No takes yet. Share an insight, caveat, or question.
Alsarhan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: