Pith. sign in

REVIEW 3 major objections 5 minor 19 references

Data-Driven Energy Modeling of Industrial IoT Systems: A Benchmarking Approach

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Benchmark method predicts robot-arm energy to 4% error

desk verdict A solid, honest benchmarking study of an educational IIoT testbed that is worth a real referee round; the main caveat is uncalibrated 1 Hz smart-plug measurements. read the letter →

arxiv 2505.02543 v1 pith:W3FRXALD submitted 2025-05-05 cs.RO cs.SYeess.SY

Dimitris Kallis , Moysis Symeonides , Marios D. Dikaiakos This is my paper · ORCID
classification cs.ROcs.SYeess.SY
keywords industrialIoTcyber-physicalsystemsenergymodelingbenchmarkingmethodologyroboticarmmachinelearningpredictionpowerconsumptionsmartplugs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a benchmarking methodology built from component-level micro-benchmarks and an end-to-end sorting application can produce accurate, explainable predictors of power, energy, and duration for small industrial cyber-physical systems. It matters because current physics-based and black-box models are either too simplistic or too opaque to guide energy-efficient design of IIoT automation. On a testbed with a robotic arm, conveyor belt, suction end-effector, and smart camera, the authors show that arm velocity, acceleration, and payload barely move power consumption, that the suction pump dominates power, and that tree-based models predict power to about 3.6% and energy and duration to about 4.23% mean absolute percentage error. If right, the framework gives designers a low-cost way to profile and optimize energy before full deployment.

What carries the argument

The central mechanism is a control layer with adapter modules sitting between application workloads and the physical components: an instruction translator converts workload parameters such as belt speed, arm velocity, acceleration, and suction state into device commands, and a monitoring module records one-second power readings and system states into a dataset. Micro-benchmarks isolate each component's power profile, while an end-to-end sorting application provides round-level energy and duration targets. The same dataset feeds tree-based regressors, and feature-importance analysis explains the predictions. This object carries the argument because it ties configurable workload features to measured power and then to predicted energy.

What would settle it

Rerun the same micro-benchmarks and end-to-end rounds while measuring each component with a calibrated, high-frequency power meter in line with the supply; if the per-second plug readings miss the suction pump's inrush or the arm's motion transients, and total energy per round differs from the plug-based energy by more than the reported 4% margin, then the ground-truth assumption and the resulting rankings fail.

Watch

Extended reading notes

Core claim

The paper's central claim is that per-component benchmarking plus supervised learning yields a general, explainable route to IIoT energy models. Using adapters that translate workload configurations into actuator commands and log one-second power snapshots from smart plugs, the authors generated a dataset of 22,384 state rows from a sorting pipeline and trained 20 regression models. Their best models achieved MAPE of approximately 3.6% for instantaneous power and 4.23% for round energy and duration, with Random Forest leading on $R^2$. The analysis also concludes that the end-effector's pump is the most power-intensive component at about 22 watts, while arm velocity, acceleration, and payload have minimal impact on power; energy per round falls logarithmically as speed and acceleration rise because duration shrinks while power stays roughly constant.

Load-bearing premise

The load-bearing premise is that the smart plugs' once-per-second power readings, taken without calibration against a reference meter, are accurate enough to serve as ground truth for all component and system power conclusions, including the error rates and component rankings.

Editorial extensions

If this is right

  • Engineers can run micro-benchmarks through the control layer before field deployment to identify the dominant power consumer and target it for efficiency gains; in this testbed, the suction pump is that component.
  • Operating the sorting line at higher belt speeds and arm acceleration reduces energy per processed object, because round duration falls while average power stays roughly flat.
  • Tree-based regressors trained on 16 state features predict instantaneous power to about 3.6% MAPE and round energy and duration to about 4.23% MAPE on this setup.
  • Feature importance indicates that acceleration and velocity influence energy through task timing rather than through power draw, so latency-aware scheduling is an energy lever.
  • Because the control layer uses replaceable actuator and sensor adapters, the same benchmarking and modeling recipe can be extended to other IIoT cyber-physical systems with new components.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the smart-plug ground truth samples once per second and is not calibrated against a reference meter, the reported flat power profiles could miss sub-second transients; higher-frequency metering might revise the component rankings.
  • The exclusion of the smart camera, despite a measured 2.13 W detection difference, means the system-level model undercounts vision-heavy workloads; adding camera state as a feature would directly test the methodology's generality.
  • If round energy is dominated by duration, as the paper's feature-importance results suggest, then a duration-only regression might match the 4.23% MAPE; comparing such a model would isolate the value of the remaining features.
  • The logarithmic energy-per-object trend suggests a practical scheduling rule: push belt speed toward the knee of the curve before mechanical limits, and the same benchmark dataset could be used to locate that knee automatically.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a methodology and software framework for benchmarking and energy profiling of industrial IoT / cyber-physical systems, and demonstrates it on a testbed consisting of a Dobot Magician robotic arm, a conveyor belt, a suction end-effector, a smart camera, and a compute node. The authors define component-level micro-benchmarks and an end-to-end object-sorting application, collect a dataset of more than 22,000 1-second power samples from Meross smart plugs, perform exploratory analysis of component and application power, and train regression models to predict instantaneous power, per-round energy, and per-round duration. The headline claims are that the best models achieve approximately 3.6% MAPE for power prediction and 4.23% MAPE for energy and duration prediction, and that arm velocity, acceleration, and payload have negligible impact on power, while the suction pump is the most power-intensive component. The paper also reports that per-round energy decreases logarithmically with speed and acceleration.

Significance. If the measurement chain is trustworthy, the paper provides a useful, repeatable benchmarking methodology and a relatively complete empirical characterization of a small industrial IoT cell, including component-level analysis, end-to-end workload evaluation, multiple regression models, k-fold cross-validation, and feature-importance analysis. These contributions would be valuable to practitioners seeking to profile and optimize energy use in similar small-scale CPS. However, the scientific value is conditional on the validity of the smart-plug ground truth and on the independence of the energy model from the duration model; the manuscript does not yet establish either point, and the reported physical insights and accuracy numbers inherit that uncertainty.

major comments (3)
  1. [Section 3, Section 6.1.1] All power and energy conclusions rest on Meross smart plugs sampled once per second, yet the manuscript reports no calibration against a reference power meter and does not state whether the plugs report instantaneous samples or interval averages. This is load-bearing because the negative physical findings (arm velocity, acceleration, and payload have minimal power impact) and the reported MAPE values would be misleading if short-duration transients from servo movements, pump inrush, or gripper actuation are attenuated or aliased by 1 Hz sampling. I request a calibration comparison against a reference meter for both static and transient loads, a justification of the sampling rate, or a sensitivity analysis. Relatedly, the 2.13 W camera detection difference (Section 6.1.1) is dismissed as 'minimal' without a quantitative threshold; excluding the camera narrows the definition of system power and should be justified explicitly or the camera should be included in the analysis.
  2. [Section 6.2.2, Section 7.2.1] Energy per round is computed as mean power multiplied by duration (Section 6.2.2), and the authors acknowledge in Section 7.2.1 that mean power is relatively static compared with duration. Consequently, the energy model's metrics in Table 1 nearly coincide with the duration model's metrics, so the reported 4.23% MAPE for energy does not independently validate energy prediction; it largely reflects the duration model. The central claim of accurate energy modeling therefore needs either an independent energy target (e.g., integrated power readings) or an explicit treatment of energy as a composition of power and duration models with separate error propagation. In addition, the claim in Section 6.2.2 that energy drops 'logarithmically' with speed and acceleration is stated without a fitted model, coefficients, or goodness-of-fit; please fit and report the model or soften the claim to a monotonic decrease.
  3. [Section 7, Table 1] The cross-validation procedure is described only as 'k-fold cross-validation with stratified sampling' (Section 7). If the folds are split at the row level rather than grouped by trial or configuration, temporally correlated measurements from the same configuration can appear in both training and test folds, which would inflate R2 and deflate MAPE relative to generalization to new configurations. This is load-bearing for the claimed predictive accuracy. Please specify the grouping unit and report group-based cross-validation such as leave-one-configuration-out, or otherwise demonstrate that the reported metrics reflect generalization to unseen configurations.
minor comments (5)
  1. [Section 3] The setup description first states that the testbed includes 'three smart plugs' and later states that 'we use four Meross smart plugs'; please reconcile the count, since three actuators plus the camera-powered USB hub would require four.
  2. [Section 6.1.3] The subsection heading 'End-effector (suction cup)' is followed by an incomplete sentence fragment; please rephrase to a complete sentence.
  3. [Section 6.2.2, Figures 11-13] The 'logarithmic trend' language should be supported by fitted curves or a statistical test; without them, a reader cannot distinguish logarithmic from other monotonically decreasing trends.
  4. [Table 1] Random Forest and Extra Trees both report MAPE 0.036 for power; please clarify the tie-breaking criterion used to declare Random Forest the best model.
  5. [Section 4] Please state whether the dataset and framework code will be released; a data-availability statement would strengthen the benchmarking contribution and enable reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ML predictions are trained and evaluated on held-out measured data, and the energy–duration relationship is explicitly disclosed rather than hidden.

full rationale

The paper's derivation chain is a standard empirical pipeline: build a testbed, run micro-benchmarks and an end-to-end sorting workload, record power and timing measurements every second, assemble a dataset, train regression models with k-fold cross-validation, and report held-out error metrics. The targets (instantaneous power, round energy, round duration) are measured or computed directly from measurements, not generated by the models themselves. The energy per round is defined as mean power times duration, and the paper explicitly acknowledges that this makes the energy and duration models' errors nearly identical because average power is relatively static. That is a transparent physical relationship, not a disguised fitting of a parameter that is then relabeled as a prediction. The feature-importance finding that acceleration and velocity dominate energy mainly through their effect on timing is likewise explained by the same explicit formula, not smuggled in as an assumption. No load-bearing argument depends on a self-citation, no uniqueness theorem is imported from the authors' prior work, and no ansatz is adopted via citation. The main weaknesses of the paper, such as the lack of calibration of the Meross smart plugs against a reference meter and the reliance on 1 Hz sampling, are threats to measurement validity and generalizability, not instances of circular reasoning. Because the predictive claims are validated on held-out data and the energy–duration connection is openly stated, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim rests on measurement fidelity and on exclusion choices. The ML models themselves are fitted objects, not free parameters in a physical derivation. No new entities are introduced. The camera exclusion and the 1-second sampling are the most consequential assumptions.

assumptions (4)
  • domain assumption Smart plug power measurements are accurate ground truth for component power.
    All power and energy conclusions rely on uncalibrated Meross smart plug readings; no reference meter comparison is reported (Sec. 3, Sec. 4).
  • ad hoc to paper The smart camera's power consumption can be excluded as minimal and stable.
    Sec. 6.1.1 reports a 2.13 W difference between detection states yet the camera is excluded from analysis in Sec. 3 and Sec. 6.2.1; this choice affects the absolute energy totals.
  • domain assumption One-second sampling captures the power dynamics of the robotic arm.
    Monitoring snapshots are taken every second (Sec. 4), which may miss short transient peaks of the arm and pump.
  • domain assumption The normalized velocity and acceleration parameters map monotonically to physical motion speed.
    The Dobot API uses normalized 0-100 values (Sec. 3); the paper assumes these are sufficient features for ML models without validating against true joint velocities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data-Driven Energy Modeling of Industrial IoT Systems: A Benchmarking Approach." pith.science (2026). https://pith.science/paper/W3FRXALD

@misc{pith2026250502543,
  author       = {Pith},
  title        = {Pith review of: Data-Driven Energy Modeling of Industrial IoT Systems: A Benchmarking Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3FRXALD}},
  note         = {Machine review of arXiv:2505.02543}
}
read the original abstract

The widespread adoption of IoT has driven the development of cyber-physical systems (CPS) in industrial environments, leveraging Industrial IoTs (IIoTs) to automate manufacturing processes and enhance productivity. The transition to autonomous systems introduces significant operational costs, particularly in terms of energy consumption. Accurate modeling and prediction of IIoT energy requirements are critical, but traditional physics- and engineering-based approaches often fall short in addressing these challenges comprehensively. In this paper, we propose a novel methodology for benchmarking and analyzing IIoT devices and applications to uncover insights into their power demands, energy consumption, and performance. To demonstrate this methodology, we develop a comprehensive framework and apply it to study an industrial CPS comprising an educational robotic arm, a conveyor belt, a smart camera, and a compute node. By creating micro-benchmarks and an end-to-end application within this framework, we create an extensive performance and power consumption dataset, which we use to train and analyze ML models for predicting energy usage from features of the application and the CPS system. The proposed methodology and framework provide valuable insights into the energy dynamics of industrial CPS, offering practical implications for researchers and practitioners aiming to enhance the efficiency and sustainability of IIoT-driven automation.

Figures

Figures reproduced from arXiv: 2505.02543 by the authors.

Figure 1
Figure 1. Experimentation Setup & Robotic Arm a smart camera and three smart plugs. Moreover, a laptop is used to coordinate execution and retrieve the monitoring data. Dobot Magician Toolset: Dobot Magician2 , shown in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. High-level Overview of the Architecture (i) repeatably submit various workloads with different configurations, (ii) easily introduce new workloads and physical components, and (iii) extract monitoring datasets of each run. Our methodology starts with users submitting a set of experiment parameters to the control layer. These parameters include the selected workload and configuration preferences, which will be transl… view at source ↗
Figure 3
Figure 3. Power Demand (Watts) for Velocity Variations (in %) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Power Demand (Watts) for Acceleration Variations (in %) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Power Demand (Watts) for Payload Variations (in kgrams) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Power Demand (Watts) for Pump End-effector Status [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Power Demand (Watts) for Belt Speed Variations (in mm/s) [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Application Power Demand (Watts) for Velocity Levels (%) [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Application Power Demand (Watts) for Acceleration Levels (%) [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Application Power Demand (Watts) for Belt Speed (mm/s) [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 11
Figure 11. Figure 11: Energy (Joules) & Power Demand (Watts) for Velocity Levels [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: Energy (Joules) & Power Demand (Watts) for Acceleration Levels [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]
Figure 13
Figure 13. Figure 13: Energy (Joules) & Power Demand (Watts) for Belt Speed (mm/s) [PITH_FULL_IMAGE:figures/full_fig_p008_13.png]
Figure 14
Figure 14. Figure 14: Peak Power (Watts) and Energy (Joules) per object vs Throughput [PITH_FULL_IMAGE:figures/full_fig_p008_14.png]
Figure 15
Figure 15. Figure 15: Feature Importance for Duration Model & Energy Model [PITH_FULL_IMAGE:figures/full_fig_p010_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 18 canonical work pages

  1. [1]

    Industrial internet of things: Challenges, opportunities, and directions,

    E. Sisinni, A. Saifullah, S. Han, U. Jennehag, and M. Gidlund, “Industrial internet of things: Challenges, opportunities, and directions,” IEEE Transactions on Industrial Informatics, vol. 14, pp. 4724–4734, 2018

    work page 2018

  2. [2]

    5g-slicer: An emulator for mobile iot applications deployed over 5g network slices,

    M. Symeonides, D. Trihinas, G. Pallis, M. D. Dikaiakos, C. Psomas, and I. Krikidis, “5g-slicer: An emulator for mobile iot applications deployed over 5g network slices,” in IEEE/ACM IoTDI2022. IEEE, 2022

    work page 2022

  3. [3]

    Ecogreen: Electricity cost optimization for green datacenters in emerging power markets,

    A. Pahlevan, M. Zapater, A. K. Coskun, and D. Atienza, “Ecogreen: Electricity cost optimization for green datacenters in emerging power markets,” IEEE T-SUSC, vol. 6, no. 2, pp. 289–305, 2021

    work page 2021

  4. [4]

    Greenscale: Carbon-aware systems for edge computing,

    Y . G. Kim, U. Gupta, A. McCrabb, Y . Son, V . Bertacco, D. Brooks, and C.-J. Wu, “Greenscale: Carbon-aware systems for edge computing,” arXiv, Tech. Rep. arXiv:2304.00404v1, 2023

    arXiv 2023

  5. [5]

    Optimization of data and energy migrations in mini data centers for carbon-neutral computing,

    M. D. M. da Silva, A. Gamatié, G. Sassatelli, M. Poss, and M. Robert, “Optimization of data and energy migrations in mini data centers for carbon-neutral computing,” IEEE T-SUSC, vol. 8, no. 1, pp. 68–81, 2023

    work page 2023

  6. [6]

    A cyber- physical management system for medium-scale solar-powered data centers,

    M. D. Dikaiakos, N. G. Chatzigeorgiou, A. Tryfonos, A. Andreou, N. Loulloudes, G. Pallis, and G. E. Georgiou, “A cyber- physical management system for medium-scale solar-powered data centers,” Concurrency and Computation: Practice and Experience, p. e7658, 2023

    work page 2023

  7. [7]

    Ai index report,

    N. Maslej, L. Fattorini, R. Perrault, V . Parli, A. Reuel, E. Brynjolfsson, J. Etchemendy, K. Ligett, T. Lyons, J. Manyika, J. C. Niebles, Y . Shoham, R. Wald, and J. Clark, “Ai index report,” 2024

    work page 2024

  8. [8]

    Suleiman and M

    M. Suleiman and M. Baskar, The Coming Wave. BH, 2024

    work page 2024

Show all 19 references
  1. [9]

    Modeling and optimization of energy consumption in cooperative multi-robot systems,

    A. Vergnano, C. Thorstensson, B. Lennartson, P. Falkman, M. Pellicciari, F. Leali, and S. Biller, “Modeling and optimization of energy consumption in cooperative multi-robot systems,”IEEE Transactions on Automation Science and Engineering, vol. 9, no. 2, pp. 423–428, 2012

    2012

  2. [10]

    Energy consumption and dynamic behavior analysis of a six-axis industrial robot in an assembly system,

    Paryanto, M. Brossog, J. Kohl, J. Merhof, S. Spreng, and J. Franke, “Energy consumption and dynamic behavior analysis of a six-axis industrial robot in an assembly system,” Procedia CIRP, vol. 23, pp. 131–136, 2014, 5th CATS 2014

    2014

  3. [11]

    An efficient computation for energy optimization of robot trajectory,

    X. Li, Y . Lan, P. Jiang, H. Cao, and J. Zhou, “An efficient computation for energy optimization of robot trajectory,” IEEE Transactions on Industrial Electronics, vol. 69, no. 11, pp. 11 436–11 446, 2022

    2022

  4. [12]

    Data-driven energy estimation of individual instructions in user-defined robot programs for collaborative robots,

    J. Heredia, C. Schlette, and M. B. Kjærgaard, “Data-driven energy estimation of individual instructions in user-defined robot programs for collaborative robots,” IEEE Robotics and Automation Letters, vol. 6, no. 4, pp. 6836–6843, 2021

    2021

  5. [13]

    Going green with lightweight robots: Energy optimal programming of lightweight robots,

    ——, “Going green with lightweight robots: Energy optimal programming of lightweight robots,” in 2023 21st International Conference on Advanced Robotics (ICAR). IEEE, 2023, pp. 212–219

    2023

  6. [14]

    Ecdp: Energy consumption disaggregation pipeline for energy optimization in lightweight robots,

    J. Heredia, R. J. Kirschner, C. Schlette, S. Abdolshah, S. Haddadin, and M. B. Kjægaard, “Ecdp: Energy consumption disaggregation pipeline for energy optimization in lightweight robots,” IEEE Robotics and Automation Letters, vol. 8, no. 10, pp. 6107–6114, 2023. 11 A PREPRINT -...

    2023

  7. [15]

    Energy consumption modeling and parameter identification based on system decomposition of welding robots,

    W. Xiao, G. Han, A. S. Ally, and X. Chen, “Energy consumption modeling and parameter identification based on system decomposition of welding robots,” The International Journal of Advanced Manufacturing Technology, vol. 130, no. 3, pp. 1579–1594, 2024

    2024

  8. [16]

    Digital twin-based energy modeling of industrial robots,

    K. Yan, W. Xu, B. Yao, Z. Zhou, and D. T. Pham, “Digital twin-based energy modeling of industrial robots,” inMethods and Applications for Modeling and Simulation of Complex Systems, L. Li, K. Hasegawa, and S. Tanaka, Eds. Singapore: Springer Singapore, 2018, pp. 333–348

    2018

  9. [17]

    A transfer-learning based energy consumption modeling method for industrial robots,

    J. Yan and M. Zhang, “A transfer-learning based energy consumption modeling method for industrial robots,” Journal of Cleaner Production, vol. 325, p. 129299, 2021

    2021

  10. [18]

    Research on power modeling of the industrial robot based on resnet,

    M. Yao, Q. Zhao, Z. Shao, and Y . Zhao, “Research on power modeling of the industrial robot based on resnet,” inCACRE 2022. IEEE, 2022

    2022

  11. [19]

    [Online]

    Dobot Magician User Guide, Shenzhen Yuejiang Technology Co., Ltd, Shenzhen, China, August 2020. [Online]. Available: www.dobot.cc 12

    2020

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.