Geotechnically Informed Monitoring-Data Fusion and Gradient-Boosted Regression for Instrumented Driven Piles
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript presents a case-study-based framework for integrating long-term monitoring data from one instrumented test pile and three production piles at the Ramsayville Road Overpass. Geological information, field-vane and consolidation test results, static and high-strain dynamic testing, and long-term VWSG, VWP, EXT, and MPPE records are organized within a common elevation-time database. The study further reconstructs axial-force and bending-moment profiles from opposed strain gauges and compares persistence, a seven-day moving average, LightGBM, and CatBoost at several forecasting horizons.
The topic is relevant to geotechnical engineering practice because long-term pile-monitoring datasets often contain failed sensors, irregular sampling intervals, repeated derived variables, and complex relationships among different sensor systems. The data-cleaning, quality-flagging, leakage-control, and mechanically informed interpretation procedures developed in the manuscript are therefore of practical interest. It is also commendable that the authors report transparently that the boosted-tree models do not outperform the simpler temporal benchmarks.
Overall, the manuscript has a sound engineering basis and useful methodological potential. However, several aspects of the target definition, temporal validation, state-classification benchmarks, measurement uncertainty, and mechanical sign conventions require further clarification and analysis before the manuscript can be considered for publication. These issues appear addressable through additional analyses, clearer reporting, and more carefully bounded conclusions. A major revision is therefore recommended.
Abstract
- The abstract is generally complete, but it is information-dense. Consider reducing repeated numerical details and emphasizing the research question, the principal methodology, the benchmark-led result, and the main limitation.
- The statement that the framework supports targeted verification should be moderated to indicate that it provides a data basis for such verification unless a prospective decision rule is demonstrated.
- The macro F1-score should be interpreted together with the simple classification benchmarks requested above or presented more cautiously until that comparison is available.
- Clarify that the 3,072,338 entries are repeated time-stamped records rather than independent samples.
Introduction
- The literature review should be expanded to include long-term instrumented-pile monitoring, field observations of negative shaft resistance, sensor drift and quality control, structural-monitoring time series, and grouped temporal validation.
- Most of the current machine-learning references concern pile-capacity prediction, which differs from the long-term forecasting problem addressed here. A clearer comparison with the closest monitoring and time-series studies would strengthen the research gap.
- The end of the Introduction should state the primary research hypothesis more explicitly: whether the models can outperform persistence and whether data fusion offers additional value for anomaly screening or mechanical interpretation.
- The six novelty statements should be condensed and should distinguish established methods, project-specific implementation, and transferable methodological contributions.
Materials and Methods
- Explain explicitly that the independent physical sample consists of four piles and discuss temporal and within-pile dependence among the daily records.
- State the selection rule for the 20 axial-force and bending-moment profile dates to avoid concerns regarding selective presentation.
Results and Discussion
- In addition to pooled RMSE, MAE, and R², report stratified results and uncertainty by pile, gauge level, elevation range, and construction or monitoring stage where possible.
- Because RMSE is much larger than MAE, include error quantiles, boxplots, stable-period versus event-period results, or another robust characterization of extreme errors.
Conclusions
- State clearly that the forecast variable is the change in load relative to the initial monitoring baseline, not the total pile axial force or total dragload.
- Revise the state-classification conclusion to indicate that performance is high under empirical thresholds but that incremental value over simple benchmarks remains to be established.
- Emphasize that the results apply to the present project, sensor configuration, and monitoring stage and cannot yet be transferred directly to a new site or pile type.
- Retain the important boundary statement that the machine-learning models do not replace static-load testing, dynamic testing, resistance calculations, or neutral-plane analysis.
Author Response
Please see the attached file
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis study investigates one test pile and three production piles from the Ramsayville Road Overpass project in Ottawa, Canada. The authors integrate borehole data, field vane shear tests, laboratory tests, static load tests, high-strain dynamic tests, and long-term monitoring data from vibrating-wire strain gauges (VWSGs), vibrating-wire piezometers (VWPs), borehole extensometers (EXTs), and multi-point pile extensometers (MPPEs). A unified elevation-time data structure is established, and persistence, a 7-day moving average, LightGBM, and CatBoost are used to forecast pile axial force at 1-, 7-, and 30-day horizons. The main strengths of the manuscript are the richness of the engineering data, the substantial effort devoted to data processing, and the authors' attention to time-series validation and data-leakage control. It is particularly commendable that the authors report, without overstating the performance of the machine-learning models, that the simple persistence and 7-day moving-average models outperform LightGBM and CatBoost. The study has potential engineering value in the standardization, quality control, and multisource fusion of long-term pile-monitoring data. The following comments are offered for the authors' consideration in revising the manuscript:
1. The current results show that both persistence and the 7-day moving average outperform LightGBM and CatBoost. The manuscript should therefore emphasize the contributions of multisource monitoring-data fusion, data-leakage control, and the rigorous validation framework, while reducing the emphasis on the predictive superiority of gradient-boosted models. The title, abstract, and conclusions should be revised accordingly.
2. A uniform calendar-time cutoff should be adopted either for the entire site or separately for each pile. Blocked time-series validation or rolling-origin validation should also be added to prevent measurements from different sensor locations on the same date, together with VWP, EXT, and MPPE information, from appearing simultaneously in the training and validation sets.
3. Precision, recall, and F1-score should also be reported for the persistence and 7-day moving-average benchmarks. Particular attention should be given to samples involving state transitions. The LightGBM macro F1-score of 0.953 alone is insufficient to demonstrate additional value for state screening.
4. LightGBM does not outperform persistence for any of the held-out piles. The results should therefore not be described as showing "strong transfer" or "moderate transfer." Skill scores relative to both persistence and the moving-average benchmark should be reported, and the manuscript should clearly state that the method still requires historical monitoring data from the target pile.
5. Reducing VWP, EXT, and MPPE channels at different elevations to common means, extrema, and standard deviations may discard information related to stratigraphic position, sensor spacing, and local deformation compatibility. The authors should address this limitation appropriately.
6. Different baseline windows should be compared, and the number and proportion of records removed by each screening criterion should be reported. The sensitivity of the results to load thresholds, daily-change thresholds, sensor drift, and errors between paired strain gauges should also be evaluated.
7. The 82229 records are essentially pile-gauge-days rather than statistically independent pile-level samples. Because the study contains only four independent pile objects, this limitation should be stated clearly in the abstract, methodology, and discussio, and pile-specific uncertainty should be quantified.
8. The basis for selecting the 20 profile dates should be explained. The load-transfer zone should be interpreted jointly using the axial-force gradient, EXT settlement, MPPE compression, and VWP pore pressure. The effective perimeter adopted when back-calculating shaft resistance for the H-piles should also be clearly defined.
9. CatBoost is trained on only a 30,000-record subsample, whereas LightGBM is trained on the full dataset; therefore, the current comparison is not fully equitable. The models should be compared using consistent training datasets, hyperparameter-tuning procedures, and random seeds. A data dictionary, detailed data-processing workflow, and sufficient information to reproduce the key results should also be provided.
10. The static load test and high-strain dynamic test conducted in 2019 should not cite the 2026 editions of ASTM D1143/D1143M-26 and ASTM D4945-26. The standards in effect at the time of testing should be cited; in particular, the authors should verify the applicability of ASTM D1143/D1143M-07(2013)e1 and ASTM D4945-17. In addition, the literature review provides insufficient coverage of long-term pile monitoring, sensor-data fusion, time-series cross-validation, and data leakage, and should be supplemented with recent relevant studies. References [14]-[19] are internal project reports; their availability, repository or custodian, and means of access should be stated.
Author Response
See the attached file
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsI think the paper is well written and is of interest to the readers of this journal. Here are some comments for the authors:
Comment 1: The authors are requested to add one section “research objective” between introduction and materials and method to illustrate the research objectives of the paper. Also, the authors are requested to add one flowchart illustrating the overall flow of the research.
Comment 2: The authors are requested to add one figure to show the stratigraphy of the soil. In that figure, the authors are requested to include different information (unit weight, water content, plastic limit and other parameters) from table 1.
Comment 3: In section 2.3, the authors are requested to add one figure to show the different sensors study in this study. The authors are requested to provide a table showing the instrumentation details (name of the sensors, depth of installation for different piles).
Comment 4: The authors are requested to provide more information about field vane shear testing. The authors may one table showing the depth of testing, size of vane used for testing and shear strength.
Comment 5: The authors are requested to add one section “Data” presenting the details about different data used in this study.
Comment 6: I think this research is relevant in the field of engineering and mathematics. However, proper description of data and instrumentation hinders the understanding of this manuscript. Therefore, the authors are requested to provide proper description of the data collected and field instrumentation. Moreover, the authors used a few machine learning algorithm, which are basic for an esteemed journal. Therefore, the authors are requested to include other machine learning algorithms and deep learning to increase the novelty of this work.
Comment 7: The authors are requested to add one section “practical application” before conclusion to illustrate the practical application of the work.
Comment 8: The authors are requested to improve the resolution of the figures. The authors are requested to improve figure 1 and figure 4.
Comment 9: This research showed a case-study of geotechnical pile monitoring using sensors and different mathematical techniques. I would recommend converting the title of the paper into a case study because it is showing data of pile foundation from one region. The models may not be valid across the world.
Author Response
See attched file
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe author has made the required revisions, so this paper can be accepted.
Author Response
We thank the reviewer for constructive comments, which helped in improving the quality of the manuscript.
Reviewer 2 Report
Comments and Suggestions for Authorswell done!
The manuscript has been systematically revised and improved, and I believe it can be accepted for publication.
Author Response
Thank you very much for your comments; they helped us in shaping the manuscript.
Reviewer 3 Report
Comments and Suggestions for AuthorsI think the authors successfully revised the reviewer's comment in the paper.
Author Response
Thank you very much for your comments; they helped us improve the quality of the manuscript.
