pith. sign in

arxiv: 2606.20952 · v1 · pith:2LMRTCILnew · submitted 2026-06-18 · 💻 cs.SE

Tiny Machine-Learning Operations within Cyber-Physical Systems: a Field Study

Filippo Scaramuzza , Damian A. Tamburri This is my paper

Pith reviewed 2026-06-26 16:00 UTC · model grok-4.3

classification 💻 cs.SE
keywords TinyMLOpsCyber-Physical SystemsField StudyOffshore WindEmbedded Machine LearningKnowledge-Driven FeaturesLoad Peak PredictionConcept Drift
0
0 comments X

The pith

A knowledge-infused tiny ML pipeline on microcontrollers can predict industrial load peaks three minutes ahead while fitting in 32 kB.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper reports a field study of TinyMLOps in cyber-physical systems using data from offshore wind cable trenching. It shows how a pipeline combining domain knowledge with machine learning produces compact models that run on microcontrollers and predict problems in advance. The models achieve good accuracy while using very little memory, and injecting expert knowledge makes them more reliable and useful. When the predictions are used in practice, they lead to less wasted time in operations. Readers would care because this makes advanced analytics possible in places where big computers or cloud services are not available.

Core claim

The study establishes that an end-to-end TinyMLOps pipeline fusing domain physics, expert speculation, and sensor streams can deliver explainable, low-footprint models for on-device deployment in CPS. On 4.4 GB of data from two campaigns, the classifier anticipates harmful load peaks up to three minutes ahead at 0.84 AUC within a 32 kB footprint on an ARM Cortex-M4. Injecting prior knowledge halves false alarms and surfaces actionable rules, and replaying recommendations in dashboards indicates an 11% reduction in non-productive time.

What carries the argument

The knowledge-centered TinyMLOps pipeline that fuses domain physics, expert speculation, and sensor streams to deliver explainable low-footprint models.

If this is right

  • The pipeline enables real-time on-device predictions in resource-constrained CPS settings.
  • Prior knowledge integration improves model reliability and provides operational insights.
  • Rolling temporal cross-validation supports adaptation to changing conditions in field data.
  • The approach demonstrates measurable operational benefits like reduced non-productive time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This method could be tested in other CPS domains such as manufacturing or transportation for similar resource limits.
  • The released code and dataset allow independent verification and extension to new hardware.
  • If the knowledge features are domain-specific, broader application would require adapting the physics inputs accordingly.

Load-bearing premise

The data from the two offshore-wind cable-trenching campaigns is representative and the knowledge-driven features with temporal validation will generalize to other campaigns without bias.

What would settle it

A follow-up study on a new campaign showing AUC below 0.7 or no reduction in non-productive time would falsify the practical effectiveness claim.

Figures

Figures reproduced from arXiv: 2606.20952 by Damian A. Tamburri, Filippo Scaramuzza.

Figure 1
Figure 1. Figure 1: Offshore wind-farm overview [28] [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cable-burial design and Depth of Lowering (DoL). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Technical drawing of the trenching ROV, highlighting the FWD/AFT [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: ACG load over time for a representative cable: sudden surges (peaks), [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Soil-layer depths for a representative cable (dashboard screenshot); [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Soil-layer geometry and definitions. Single-tipped arrows denote [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Load-peak labelling (1200 s centred window) for the K25–K33 cable; [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 13
Figure 13. Figure 13: Fitted coefficients of the boulder-probability [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
Figure 12
Figure 12. Figure 12: Fitted coefficients of ∆EF SACG on peak probability for South Fork: the confirmed effect transfers out-of-sample. C. Predictive Performance and Operational Benefit With the confirmed knowledge encoded, the validated pipeline anticipates harmful ACG-load peaks up to three minutes ahead—the lead time implied by the trencher’s geom￾etry and speed—at 0.84 AUC under rolling temporal cross￾validation, within a … view at source ↗
read the original abstract

Machine-Learning Operations (MLOps) is maturing into a software-engineering discipline, yet its tiny-scale variant (TinyMLOps)-targeting the resource-constrained microcontrollers embedded in cyber-physical systems (CPS)-remains poorly understood in industrial practice. Opaque models, noisy heterogeneous data, and tight memory budgets hinder adoption in safety-critical settings, where most decisions still rely on human experts. We report a field study of an end-to-end, knowledge-centered TinyMLOps pipeline that fuses domain physics, expert speculation, and sensor streams to deliver explainable, low-footprint models deployable on-device. The pipeline spans automated collection and cleaning of heterogeneous time series, knowledge-driven feature construction, interpretable regularized models, and rolling temporal cross-validation under concept drift. We evaluate it on 4.4 GB of data from two offshore-wind cable-trenching campaigns. The classifier anticipates harmful load peaks up to three minutes ahead at 0.84 AUC within a 32 kB footprint on an ARM Cortex-M4; an ablation shows that injecting prior knowledge halves false alarms and surfaces actionable operational rules. Replaying recommendations in operational dashboards indicates an 11% reduction in non-productive time. We distill engineering lessons and validity threats for trustworthy TinyMLOps in CPS, and release code and an annotated dataset to support reproducibility.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper reports a field study of an end-to-end knowledge-centered TinyMLOps pipeline for cyber-physical systems. It fuses domain physics and sensor streams from two offshore-wind cable-trenching campaigns (4.4 GB total) to produce interpretable regularized models deployable on microcontrollers. The classifier predicts harmful load peaks up to three minutes ahead at 0.84 AUC within a 32 kB footprint on an ARM Cortex-M4; an ablation shows prior-knowledge injection halves false alarms; dashboard replay indicates an 11% reduction in non-productive time. The work releases code and an annotated dataset and distills engineering lessons plus validity threats.

Significance. If the empirical results hold, the study supplies concrete, reproducible evidence that knowledge-driven feature construction and rolling temporal cross-validation can yield explainable, tiny-footprint models suitable for safety-critical CPS. The explicit release of code and dataset is a notable strength that directly supports verification and extension by the community.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (Evaluation): the headline metrics (0.84 AUC under rolling temporal cross-validation, 11% non-productive-time reduction) are stated without describing the data-cleaning rules, the exact regularization formulation, or how the AUC was computed while accounting for operator variability; these details are load-bearing for the performance and ablation claims.
  2. [§5] §5 (Validity threats): the generalization argument rests on data from exactly two campaigns; the rolling CV addresses within-campaign drift but no cross-campaign hold-out or sensitivity test to changes in physical regime or sensor suite is reported, leaving the claim that the pipeline and surfaced rules will transfer to new CPS deployments as an untested extrapolation.
minor comments (1)
  1. [§3.2 and figures] Figure captions and §3.2 should explicitly state the time horizon used for the three-minute-ahead prediction and the memory footprint measurement protocol on the Cortex-M4.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point by point below, indicating where revisions will be made to improve clarity and strengthen the empirical claims.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Evaluation): the headline metrics (0.84 AUC under rolling temporal cross-validation, 11% non-productive-time reduction) are stated without describing the data-cleaning rules, the exact regularization formulation, or how the AUC was computed while accounting for operator variability; these details are load-bearing for the performance and ablation claims.

    Authors: We agree these details are essential for reproducibility. In the revised version we will add an explicit subsection in §4 describing the data-cleaning rules applied to the 4.4 GB dataset, the precise regularization formulation (including parameter selection), and the AUC computation procedure under rolling temporal cross-validation, with explicit discussion of how operator variability is handled via the temporal splits. The abstract will be updated to reference these additions where space allows. The released code already encodes these steps; the revision will make the paper self-contained. revision: yes

  2. Referee: [§5] §5 (Validity threats): the generalization argument rests on data from exactly two campaigns; the rolling CV addresses within-campaign drift but no cross-campaign hold-out or sensitivity test to changes in physical regime or sensor suite is reported, leaving the claim that the pipeline and surfaced rules will transfer to new CPS deployments as an untested extrapolation.

    Authors: We accept that cross-campaign generalization requires direct evidence. In revision we will add a leave-one-campaign-out validation (training on one campaign and testing on the other) together with a sensitivity analysis to regime and sensor-suite variations feasible with the released dataset. Section 5 will be updated to reflect the new results and any remaining limitations. This directly tests transferability rather than leaving it as extrapolation. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical field study with released code and data.

full rationale

The manuscript is a field study reporting classifier performance (0.84 AUC, 32 kB footprint) on 4.4 GB of sensor data from two offshore-wind campaigns. It employs knowledge-driven feature construction and rolling temporal cross-validation but presents no equations, fitted-parameter predictions, or derivation chains that reduce reported metrics to quantities defined inside the same paper. No self-citation load-bearing steps, uniqueness theorems, or ansatzes are invoked to justify core results. The explicit release of code and annotated dataset renders the pipeline externally verifiable against independent benchmarks, satisfying the self-contained criterion.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that domain physics and expert speculation can be reliably translated into features and regularization terms that improve out-of-sample performance on unseen campaigns; no free parameters or invented entities are named in the abstract.

axioms (1)
  • domain assumption Domain physics and expert speculation can be fused with heterogeneous sensor streams to produce features and constraints that improve model performance and interpretability under concept drift.
    Invoked as the core of the knowledge-centered pipeline.

pith-pipeline@v0.9.1-grok · 5770 in / 1357 out tokens · 39955 ms · 2026-06-26T16:00:10.336827+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 3 canonical work pages

  1. [1]

    Machine learning operations (mlops): Overview, definition, and architecture,

    D. Kreuzberger, N. Kühl, and S. Hirschl, “Machine learning operations (mlops): Overview, definition, and architecture,”IEEE Access, 2023

    2023

  2. [2]

    Sustainable mlops: Trends and chal- lenges,

    D. A. Tamburri, “Sustainable mlops: Trends and chal- lenges,” in2020 22nd international symposium on sym- bolic and numeric algorithms for scientific computing (SYNASC). IEEE, 2020, pp. 17–23

    2020

  3. [3]

    Tinyml platforms benchmarking

    A. Osman, U. Abid, L. Gemma, M. Perotto, and D. Brunelli, “Tinyml platforms benchmarking.” inApplePies, ser. Lecture Notes in Electrical Engineering, S. Saponara and A. D. Gloria, Eds., vol. 866. Springer, 2021, pp. 139–

    2021

  4. [4]

    Available: http://dblp.uni-trier.de/db/conf/ applepies/applepies2021.html#OsmanAGPB21

    [Online]. Available: http://dblp.uni-trier.de/db/conf/ applepies/applepies2021.html#OsmanAGPB21

  5. [5]

    Mlops challenges in industry 4.0,

    L. Faubel, K. Schmid, and H. Eichelberger, “Mlops challenges in industry 4.0,”SN Computer Science, vol. 4, no. 6, p. 828, 2023

    2023

  6. [6]

    Computing foundations and practice for cyber-physical systems: A preliminary report,

    E. A. Lee, “Computing foundations and practice for cyber-physical systems: A preliminary report,”University of Berkeley, Tech. Rep. UCB/EECS-2007-72, vol. 21, 2007

    2007

  7. [7]

    A review of the integration of cyber-physical system and internet of things,

    R. S. Nandhini and R. Lakshmanan, “A review of the integration of cyber-physical system and internet of things,”International Journal of Advanced Computer Science and Applications, 2022. [Online]. Available: https://pdfs.semanticscholar.org/ 364a/a30ad46b3a5e6fc2062bd630802c1ce7384d.pdf

    2022

  8. [8]

    Industry 4.0 market size, share and trends 2026 to 2035,

    Precedence Research, “Industry 4.0 market size, share and trends 2026 to 2035,” https: //www.precedenceresearch.com/industry-4-0-market, 2026, accessed: 2026-06-08

    2026

  9. [9]

    Cyber-physical system market: Global industry analysis 2015–2024 and opportunity as- sessment 2025–2035,

    Future Market Insights, “Cyber-physical system market: Global industry analysis 2015–2024 and opportunity as- sessment 2025–2035,” https://www.futuremarketinsights. com/reports/cyber-physical-systems-market, 2025, ac- cessed: 2026-06-08

    2015

  10. [10]

    Uncertainty handling in cyber–physical systems: State-of-the-art approaches, tools, causes, and future directions,

    M. N. Asmat, S. U. R. Khan, and S. Hussain, “Uncertainty handling in cyber–physical systems: State-of-the-art approaches, tools, causes, and future directions,”Journal of Software: Evolution and Process, p. e2428, 2022. [Online]. Available: https://onlinelibrary.wiley.com/doi/10.1002/smr.2428

    work page doi:10.1002/smr.2428 2022

  11. [11]

    Towards safe machine learning for cps: infer uncertainty from training data,

    X. Gu and A. Easwaran, “Towards safe machine learning for cps: infer uncertainty from training data,” inProceed- ings of the 10th ACM/IEEE International Conference on Cyber-Physical Systems, 2019, pp. 249–258

    2019

  12. [12]

    On the safety of automotive systems incorporating machine learning based compo- nents: a position paper,

    M. Gharib, P. Lollini, M. Botta, E. Amparore, S. Do- natelli, and A. Bondavalli, “On the safety of automotive systems incorporating machine learning based compo- nents: a position paper,” in2018 48th Annual IEEE/IFIP International Conference on Dependable Systems and Networks Workshops (DSN-W). IEEE, 2018, pp. 271– 274

    2018

  13. [13]

    Informed machine learning–a taxonomy and survey of integrating prior knowledge into learning systems,

    L. V on Rueden, S. Mayer, K. Beckh, B. Georgiev, S. Giesselbach, R. Heese, B. Kirsch, J. Pfrommer, A. Pick, R. Ramamurthyet al., “Informed machine learning–a taxonomy and survey of integrating prior knowledge into learning systems,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 1, pp. 614–633, 2021

    2021

  14. [14]

    Explainable machine learning with prior knowledge: an overview,

    K. Beckh, S. Müller, M. Jakobs, V . Toborek, H. Tan, R. Fischer, P. Welke, S. Houben, and L. von Rueden, “Explainable machine learning with prior knowledge: an overview,”arXiv preprint arXiv:2105.10172, 2021

    work page arXiv 2021

  15. [15]

    Physics-informed machine learn- ing,

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, IEEE TRANSACTIONS ON SOFTW ARE ENGINEERING, VOL. XX, NO. X, 2026 12 S. Wang, and L. Yang, “Physics-informed machine learn- ing,”Nature Reviews Physics, vol. 3, no. 6, pp. 422–440, 2021

    2026

  16. [16]

    Machine learning for naval architecture, ocean and marine engineering,

    J. Panda, “Machine learning for naval architecture, ocean and marine engineering,”Journal of Marine Science and Technology, vol. 28, no. 1, pp. 1–26, 2023

    2023

  17. [17]

    Review of the current status, technology and future trends of offshore wind farms,

    H. Díaz and C. Guedes Soares, “Review of the current status, technology and future trends of offshore wind farms,”Ocean Engineering, vol. 209, p. 107381,

  18. [18]

    Available: https://www.sciencedirect

    [Online]. Available: https://www.sciencedirect. com/science/article/pii/S002980182030411X

  19. [19]

    Toward data-driven optimal control: A systematic review of the landscape,

    K. Prag, M. Woolway, and T. Celik, “Toward data-driven optimal control: A systematic review of the landscape,” IEEE Access, vol. 10, pp. 32 190–32 212, 2022. [Online]. Available: https://link.springer.com/article/10. 1007/s00773-022-00914-5

    2022

  20. [20]

    Learning-based model predictive control: To- ward safe learning in control,

    L. Hewing, K. P. Wabersich, M. Menner, and M. N. Zeilinger, “Learning-based model predictive control: To- ward safe learning in control,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 3, pp. 269–296, 2020

    2020

  21. [21]

    Resilient model predictive control of cyber–physical systems under dos attacks,

    Q. Sun, K. Zhang, and Y . Shi, “Resilient model predictive control of cyber–physical systems under dos attacks,” IEEE Transactions on Industrial Informatics, vol. 16, no. 7, pp. 4920–4927, 2019

    2019

  22. [22]

    Model-free fault detection and isolation in large-scale cyber-physical systems,

    C. Alippi, S. Ntalampiras, and M. Roveri, “Model-free fault detection and isolation in large-scale cyber-physical systems,”IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 1, no. 1, pp. 61–71, 2016

    2016

  23. [23]

    Operationalizing machine learning models: a systematic literature review,

    A. B. Kolltveit and J. Li, “Operationalizing machine learning models: a systematic literature review,” inPro- ceedings of the 1st Workshop on Software Engineering for Responsible AI, 2022, pp. 1–8

    2022

  24. [24]

    Towards crisp-ml (q): a machine learning process model with quality assurance methodology,

    S. Studer, T. B. Bui, C. Drescher, A. Hanuschkin, L. Win- kler, S. Peters, and K.-R. Müller, “Towards crisp-ml (q): a machine learning process model with quality assurance methodology,”Machine learning and knowledge extrac- tion, vol. 3, no. 2, pp. 392–413, 2021

    2021

  25. [25]

    Continuous integration and delivery practices for cyber-physical systems: An interview-based study,

    F. Zampetti, D. A. Tamburri, S. Panichella, A. Panichella, G. Canfora, and M. D. Penta, “Continuous integration and delivery practices for cyber-physical systems: An interview-based study,”ACM Transactions on Software Engineering and Methodology, 2022. [Online]. Available: https://dl.acm.org/doi/pdf/10.1145/3571854

    work page doi:10.1145/3571854 2022

  26. [26]

    Challenges of machine learn- ing applied to safety-critical cyber-physical systems,

    A. Pereira and C. Thomas, “Challenges of machine learn- ing applied to safety-critical cyber-physical systems,” Machine Learning and Knowledge Extraction, vol. 2, no. 4, pp. 579–602, 2020

    2020

  27. [27]

    Modelling the waterjet cable trenching pro- cess on sand dunes,

    S. Warringa, V . Rhee, S. Miedema, C. Lupea, and C. Visser, “Modelling the waterjet cable trenching pro- cess on sand dunes,” inProceedings of the 22nd World Dredging Conference, Shanghai, China, 2019, pp. 22–26

    2019

  28. [28]

    An overview of case study,

    K. Harling, “An overview of case study,”Available at SSRN 2141476, 2012

    2012

  29. [29]

    Review of offshore wind farm cost components,

    A. G. Gonzalez-Rodriguez, “Review of offshore wind farm cost components,”Energy for Sustainable Development, vol. 37, pp. 10–19, 2017. [Online]. Available: https://www.sciencedirect.com/science/article/ pii/S0973082616303647

    2017

  30. [30]

    Offshore wind,

    Star of the South, “Offshore wind,” 2023, consulted at 31 October 2023. [Online]. Available: https://www. starofthesouth.com.au/offshore-wind

    2023

  31. [31]

    Seabed recovery following protective burial of subsea cables - observations from the continental margin,

    C. Kraus and L. Carter, “Seabed recovery following protective burial of subsea cables - observations from the continental margin,”Ocean Engineering, vol. 157, pp. 251–261, 2018. [Online]. Available: https://www.sciencedirect.com/science/article/ pii/S0029801818303020

    2018

  32. [32]

    Perspective review on subsea jet trenching technology and modeling,

    P. G. Atangana Njock, Q. Zheng, N. Zhang, and Y .-S. Xu, “Perspective review on subsea jet trenching technology and modeling,”Journal of Marine Science and Engineering, vol. 8, no. 6, 2020. [Online]. Available: https://www.mdpi.com/2077-1312/8/6/460

    2020

  33. [33]

    Cable installation and burial: Practical considerations,

    J. Pyrah, “Cable installation and burial: Practical considerations,”Marine Technology Society Journal, vol. 44, no. 1, pp. 52–56, 2010. [Online]. Available: https://www.ingentaconnect.com/content/mts/mtsj/2010/ 00000044/00000001/art00006?crawler=true&mimetype= application/pdf

    2010

  34. [34]

    Kaskasi offshore windfarm,

    4COffshore, “Kaskasi offshore windfarm,” 2023, consulted at 2023-08-10. [Online]. Avail- able: https://www.4coffshore.com/windfarms/germany/ kaskasi-germany-de33.html

    2023