- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper introduces FLOATSense, a dataset and benchmark for reconstructing the bending moment at eleven heights of floating offshore wind turbine towers from a tower-top accelerometer and SCADA signals, scored by the implied fatigue damage rather than by waveform error. The data come from 19,404 OpenFAST simulations of FLOATBench on three 22 MW towers, keeping the six turbulence realizations of each operating point in one split.
Twenty learned models, a physics method and a one-gain floor are evaluated within and across towers and under sensor ablation. Base R² saturates among the leading models while the top separates them, and the two rankings disagree. A condition-only control beats every entry in aggregate top R² but has zero within-condition correlation, so aggregate R² measures the damage level rather than realization tracking.
Critical strengths
- Field towers rarely have strain gauges beyond one section, whereas a tower-top accelerometer is standard, and no public benchmark scored this task by fatigue damage.
- The condition-only control reaches a top R² of 0.95 with ρ_wc = 0, and adding colored noise to the TCN output raises its top R² from 0.83/0.64/0.71 to 0.95/0.88/0.88 with ρ_wc essentially unchanged.
Critical weaknesses
- Nearly all models use 50 epochs at 10⁻³ with no schedule. The authors note that "a top score measures convergence as much as architecture," yet the abstract states that longer training recovers the top "for the two TCNs, not for the tested attention and state-space models."
Questions for authors
- Were the Hybrid results produced with the public implementation? Was base_correction constrained in any other way? What is the actual range of |1+c| on the test set?
3 comments
Expanded HPO, Hybrid Verification, and Corrections
We thank the reviewer for the careful assessment and for examining our code. We particularly appreciate their recognition of the practical relevance of reconstructing tower loads from available sensors, the novelty of providing the first public benchmark for evaluating this task through fatigue damage, and how the condition-only and colored-noise experiments demonstrate that aggregate accuracy can hide poor tracking of individual realizations.
We address these concerns by revising the abstract and expanding the evaluation to 2,340 training runs across all 20 models, distinguishing training effects from architectural conclusions. We also clarified the Hybrid's unconstrained and bounded components and verified that the public code reproduces all reported results.
Q1. Were the Hybrid results produced with the public code? Yes. We reran all nine Hybrid models with a fresh copy of the public code on the full test set, and every result matches the paper within 3×10⁻⁶. The models are included in the released data (see our comment "Complete Dataset and Evaluation Materials"), so these results can be checked directly.
Q2. Was base_correction constrained in any other way?
No. Our original description was incomplete: it mentioned only the bounded term, which implied that the entire correction was bounded by ±0.5. In fact, base_correction is a free per-frequency term, while only the condition-dependent component is bounded through 0.5 × tanh. We have clarified this in all four relevant locations: Section 4, Table 10, Appendix F, and Appendix G. The implementation and reported results remain unchanged.
Q3. What is the actual range of |1+c| on the test set? Across the nine Hybrid checkpoints, we computed |1+c| over the full test set and the scored frequency band (0–3 Hz). The observed values range from 0.00 to 3.39. Although 65% of frequency bins have gains outside [0.5, 1.5], these bins account for only 8% of the signal energy. In other words, the largest deviations from the physics gain occur mainly at frequencies with little signal energy, while the energy-weighted median gain is 0.97. This behavior is observed, not imposed by the model (Appendix G.1).
W1. The 50-epoch recipe and the abstract claim. We addressed the reviewer's concern in two ways: revising the abstract and expanding the evaluation to all 20 models. The abstract now limits its longer-training claim to the two convolutional models under a fixed training budget, without drawing conclusions about attention or state-space architectures.
We conducted validation-based hyperparameter optimization, varying learning rates, weight decay, schedules, and architecture-specific parameters. Each model received 30 trials per tower (1,800 trials). The two best configurations were confirmed over three seeds at 300 epochs, and the selected configuration was retrained for 300 epochs with three seeds. This required 2,340 training runs and at most 1,478 GPU-hours (Appendix K).
Top R² for the five best tuned models, median of three seeds, averaged over the three towers (all 20 in Table 34):
| Model | 50 epochs | Tuned |
|---|---|---|
| Prob-TCN | 0.86 | 0.98 |
| TCN | 0.76 | 0.95 |
| U-Net | 0.21 | 0.95 |
| LSTM | 0.42 | 0.91 |
| FNO | 0.05 | 0.75 |
Tuning substantially improves performance at the tower top: 18 of 20 models improve, with a median R² gain of 0.23, compared with 0.03 at the base. Prob-TCN remains first at the top, while U-Net and LSTM become competitive. Attention and state-space models also improve: S4D rises from 0.44 to 0.70, Mamba from −0.29 to 0.47, and iTransformer from −0.14 to 0.38.
These comparisons still use a common training loss. With a damage-aware loss, S4D reaches 0.88–0.92 at the top (seed 0; Appendix J.2), so we avoid broader claims about architectural superiority.
The validation-tuned track complements the fixed-budget analyses and confirms their main within-tower findings (Appendix K); rerunning the cross-tower, few-shot and sensor analyses with the tuned configurations is our next step. The central findings remain: performance varies substantially with tower height, and strong aggregate accuracy can still hide poor realization tracking.
Summary: We have addressed all three questions and corrected the Hybrid description without changing any result. We also tuned all 20 models (2,340 training runs), and the abstract no longer draws a conclusion about architectures from a single recipe. The central findings remain robust. We hope these improvements warrant reconsideration of your overall assessment.
Thank you for the detailed response and additional experiments. I maintain my weak-accept recommendation.
update
I have corrected the “Flag For Ethics Review” field, which I had accidentally marked as “Yes.”