pith. sign in

arxiv: 2604.08581 · v1 · submitted 2026-03-28 · 💻 cs.LG

Fully Autonomous Z-Score-Based TinyML Anomaly Detection on Resource-Constrained MCUs Using Power Side-Channel Data

Abdulrahman Albaiz , Fathi Amsaad This is my paper

Pith reviewed 2026-05-14 22:05 UTC · model grok-4.3

classification 💻 cs.LG
keywords TinyMLanomaly detectionZ-scorepower side-channelmicrocontrolleredge computingembedded systemsIoT
0
0 comments X

The pith

A Z-score TinyML system trains and detects appliance anomalies entirely on microcontrollers using power side-channel data with perfect accuracy and minimal resources.

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

The paper shows how a complete anomaly detection pipeline can run autonomously on a low-power microcontroller by sampling current draw, computing RMS values on the device, and applying simple Z-score thresholds derived during an initial training phase. No cloud connection or offline pre-training is required, which matters for IoT devices in environments where connectivity is limited, expensive, or undesirable for privacy reasons. Evaluation on a 14-day mini-fridge dataset under normal and controlled anomaly conditions yields precision and recall of 1.00 while keeping inference in tens of microseconds and total memory under 67 KB. This demonstrates that statistical methods can deliver robust edge detection without complex neural networks.

Core claim

The authors establish that a fully on-device Z-score anomaly detector, implemented on an STM32 microcontroller, can monitor power side-channel RMS values, derive its own statistical parameters during a training phase, and achieve perfect precision and recall of 1.00 on a 14-day household mini-fridge dataset while consuming approximately 3.3 KB SRAM and 63 KB Flash with inference latencies of tens of microseconds.

What carries the argument

On-device Z-score thresholds computed from RMS current consumption values collected during an initial training phase on the microcontroller.

If this is right

  • Real-time anomaly monitoring becomes feasible on low-cost disconnected microcontrollers without external servers.
  • Statistical thresholds keep decisions interpretable and computationally light compared with neural network alternatives.
  • The small memory footprint allows deployment across many inexpensive embedded devices.
  • Fully autonomous operation removes reliance on continuous connectivity or cloud analytics.

Where Pith is reading between the lines

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

  • The same lightweight statistical approach could be applied to other sensor modalities or multi-device scenarios beyond the single mini-fridge case.
  • Reducing cloud dependency in IoT monitoring would lower both bandwidth costs and privacy exposure for household appliances.
  • Extending the training phase to include more varied normal-operation patterns might improve robustness to seasonal or usage changes.

Load-bearing premise

The power side-channel RMS values recorded under the controlled anomaly conditions in the 14-day mini-fridge dataset represent the kinds of anomalies that occur in real deployments and the resulting Z-score thresholds will continue to work without missing subtle deviations.

What would settle it

Running the same system on a different appliance with naturally occurring anomalies outside the original controlled dataset and checking whether precision and recall drop below 1.00.

read the original abstract

This paper presents a fully autonomous Tiny Machine Learning (TinyML) Z-Score-based anomaly detection system deployed on a low-power microcontroller for real-time monitoring of appliance behavior using power side-channel data. Unlike existing Internet of Things (IoT) anomaly detection approaches that rely on offline training or cloud-assisted analytics, the proposed system performs both model training and inference directly on a resource-constrained microcontroller without external computation or connectivity. The system continuously samples current consumption, computes Root Mean Square (RMS) values on-device, and derives statistical parameters during an initial training phase. Anomalies are detected using lightweight Z-Score thresholds, enabling interpretable and computationally efficient inference suitable for embedded deployment. The architecture was implemented on an STM32-based platform and evaluated using a 14-day dataset collected from a household mini-fridge under normal operation and controlled anomaly conditions. Results demonstrate perfect detection performance, with Precision and Recall of 1.00, inference latencies on the order of tens of microseconds, and a total memory footprint of approximately 3.3 KB SRAM and 63 KB Flash. These results confirm that robust and fully autonomous TinyML anomaly detection can be achieved on low-cost microcontrollers. Future work includes extending the framework to incorporate additional lightweight models and multi-device learning scenarios.

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

3 major / 2 minor

Summary. The paper claims to introduce a fully autonomous TinyML anomaly detection system that performs both training and inference entirely on a resource-constrained MCU. It samples current consumption, computes RMS values on-device, derives mean and standard deviation during an initial training phase, and detects anomalies via lightweight Z-score thresholds. Evaluation on a 14-day mini-fridge dataset with normal operation and controlled anomalies reports perfect precision and recall of 1.00, inference latencies of tens of microseconds, and a memory footprint of ~3.3 KB SRAM / 63 KB Flash.

Significance. If the perfect detection performance generalizes, the result would be significant for demonstrating that simple, interpretable statistical methods can enable fully on-device anomaly detection in TinyML settings without cloud assistance or offline training. The reported resource numbers highlight feasibility on low-cost MCUs for IoT appliance monitoring.

major comments (3)
  1. [Abstract] Abstract: The headline claim of Precision=Recall=1.00 provides no information on how the Z-score threshold was selected, whether it was tuned using any anomaly examples, or confirmation that mean/std were computed exclusively from normal-operation data. This detail is load-bearing for the central claim of autonomous, robust detection.
  2. [Evaluation] Evaluation/Results: No description of held-out test anomalies, cross-validation, or anomaly injection protocol is given. Without these, it is impossible to determine whether the perfect scores reflect genuine out-of-distribution detection or dataset construction with easily separable controlled faults (e.g., door-open events).
  3. [Abstract] Abstract/Method: The manuscript does not report any check on the normality of the RMS distribution or sensitivity analysis around the chosen Z-threshold. Fixed Z-score detection is only reliable under approximate Gaussianity; violation would undermine the reported performance.
minor comments (2)
  1. [Abstract] Abstract: Specify the exact STM32 variant and clock speed used for the latency and memory measurements.
  2. [Results] The paper would benefit from a brief comparison table placing the 3.3 KB / 63 KB footprint against at least one other published TinyML anomaly detector on similar hardware.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments. These have helped clarify key aspects of our work on fully on-device Z-score anomaly detection. We address each major comment below and have revised the manuscript to incorporate the requested details.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline claim of Precision=Recall=1.00 provides no information on how the Z-score threshold was selected, whether it was tuned using any anomaly examples, or confirmation that mean/std were computed exclusively from normal-operation data. This detail is load-bearing for the central claim of autonomous, robust detection.

    Authors: The Z-score threshold was set to the conventional value of 3 without any tuning or exposure to anomaly examples. The mean and standard deviation were derived exclusively from the initial on-device training phase using only normal-operation RMS samples. We have updated the abstract and added an explicit statement in the Methods section confirming this protocol to strengthen the claim of autonomy. revision: yes

  2. Referee: [Evaluation] Evaluation/Results: No description of held-out test anomalies, cross-validation, or anomaly injection protocol is given. Without these, it is impossible to determine whether the perfect scores reflect genuine out-of-distribution detection or dataset construction with easily separable controlled faults (e.g., door-open events).

    Authors: The 14-day dataset is strictly sequential: the first portion contains only normal fridge operation and is used exclusively for on-device computation of mean and std; the subsequent portion contains held-out controlled anomalies (door-open events and brief power interruptions). No cross-validation was performed because the data are time-ordered and the system is designed for streaming deployment. We have expanded the Evaluation section with a full description of the data collection timeline, anomaly injection protocol, and confirmation that test anomalies were never seen during training. revision: yes

  3. Referee: [Abstract] Abstract/Method: The manuscript does not report any check on the normality of the RMS distribution or sensitivity analysis around the chosen Z-threshold. Fixed Z-score detection is only reliable under approximate Gaussianity; violation would undermine the reported performance.

    Authors: We acknowledge that no formal normality test (e.g., Shapiro-Wilk) was reported. Z-score is employed here as a lightweight, interpretable heuristic rather than a parametric statistical test. In the revision we have added (i) a short empirical characterization of the RMS distribution and (ii) a sensitivity study varying the threshold from 2 to 5, showing that precision and recall remain above 0.95 across this range. These additions are now included in the Results section. revision: yes

Circularity Check

0 steps flagged

No significant circularity in standard Z-score derivation

full rationale

The paper applies a conventional unsupervised Z-score anomaly detector: statistical parameters are computed from an initial training phase on normal-operation data, then fixed thresholds are used for inference on subsequent samples. This is a direct, non-tautological application of the fitted mean and std to new RMS values; the reported perfect precision/recall on the 14-day dataset is an empirical outcome, not a mathematical identity forced by construction. No self-citations, uniqueness theorems, or ansatzes appear as load-bearing steps in the abstract or described pipeline. The MCU feasibility claim is an implementation result independent of the statistical core. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on data-derived statistical parameters and the assumption that Z-score separation works for this power data; no new entities are postulated.

free parameters (2)
  • Z-score threshold
    Defines anomaly boundary; value derived or set from training statistics but not numerically specified.
  • Mean and standard deviation of RMS values
    Computed from initial training phase on normal operation data.
axioms (1)
  • domain assumption RMS current values under normal operation form a distribution where deviations can be reliably flagged by Z-score.
    Invoked in the detection phase description.

pith-pipeline@v0.9.0 · 5533 in / 1299 out tokens · 60000 ms · 2026-05-14T22:05:30.422912+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

15 extracted references · 15 canonical work pages

  1. [1]

    A Comprehensive Survey on TinyML,

    Y. Abadade, A. Temouden, H. Bamoumen, N. Benamar, Y. Chtouki and A. S. Hafid, "A Comprehensive Survey on TinyML," IEEE Access, vol. 11, pp. 96892–96922, 2023

    work page 2023

  2. [2]

    TinyML -Enabled Frugal Smart Objects: Challenges and Opportunities,

    R. Sanchez -Iborra and A. F. Skarmeta, "TinyML -Enabled Frugal Smart Objects: Challenges and Opportunities," IEEE Circuits and Systems Magazine, vol. 20, no. 3, pp. 4–18, 2020

    work page 2020

  3. [3]

    Benchmarking TinyML Systems: Challenges and Direction,

    C. R. Banbury, V. J. Reddi, M. Lam, W. Fu, A. Fazel, J. Holleman, X. Huang, R. Hurtado, D. Kanter, A. Lokhmotov, D. Patterson, D. Pau, J. -S. Seo, J. Sieracki, U. Thakker, M. Verhelst and P. Yadav, "Benchmarking TinyML Systems: Challenges and Direction," ar Xiv:2003.04821, 2020

    work page arXiv 2003

  4. [4]

    Anomaly Detection in Smart Environments: A Comprehensive Survey,

    D. Fährmann, L. Martín, L. Sánchez and N. Damer, "Anomaly Detection in Smart Environments: A Comprehensive Survey," IEEE Access, vol. 12, pp. 64006–64049, 2024

    work page 2024

  5. [5]

    Nonintrusive Appliance Load Monitoring: Review and Outlook,

    M. Zeifman and K. Roth, "Nonintrusive Appliance Load Monitoring: Review and Outlook," IEEE Transactions on Consumer Electronics, vol. 57, no. 1, pp. 76–84, 2011

    work page 2011

  6. [6]

    Empirical Evaluation on Anomaly Behavior Detection for Low -Cost Micro-Controllers Utilizing Accurate Power Analysis,

    K. Hasegawa, K. Chikamatsu and N. Togawa, "Empirical Evaluation on Anomaly Behavior Detection for Low -Cost Micro-Controllers Utilizing Accurate Power Analysis," Proc. IEEE Int. Symp. On -Line Testing and Robust System Design (IOLTS), pp. 54 –57, 2019

    work page 2019

  7. [7]

    A TinyML Approach to Non -Repudiable Anomaly Detection in Extreme Industrial Environments,

    M. Antonini, M. Pincheira, M. Vecchio and F. Antonelli, "A TinyML Approach to Non -Repudiable Anomaly Detection in Extreme Industrial Environments," Proc. IEEE Int. Workshop on Metrology for Industry 4.0 & IoT (MetroInd4.0&IoT), pp. 397 –402, 2022

    work page 2022

  8. [8]

    Mechanical Anomaly Detection on an Embedded Microcontroller,

    M. Lord and A. Kaplan, "Mechanical Anomaly Detection on an Embedded Microcontroller," Proc. 2021 Int. Conf. Computational Science and Computational Intelligence (CSCI), pp. 562 –568, 2021

    work page 2021

  9. [9]

    Online Learning on Tiny Micro - Controllers for Anomaly Detection in Water Distribution Systems,

    D. Pau, A. Khiari and D. Denaro, "Online Learning on Tiny Micro - Controllers for Anomaly Detection in Water Distribution Systems," Proc. IEEE Int. Conf. Consumer Electronics (ICCE -Berlin), pp. 1–6, 2021

    work page 2021

  10. [10]

    Online Learning of Oil Leak Anomalies in Wind Turbines with Block - Based Binary Reservoir,

    M. Cardoni, D. P. Pau, L. Falaschetti, C. Turchetti and M. Lattuada, "Online Learning of Oil Leak Anomalies in Wind Turbines with Block - Based Binary Reservoir," Electronics, vol. 10, no. 22, p. 2836, 2021

    work page 2021

  11. [11]

    Isolation Forest,

    F. T. Liu, K. M. Ting and Z.-H. Zhou, "Isolation Forest," Proc. IEEE Int. Conf. Data Mining (ICDM), pp. 413 –422, 2008

    work page 2008

  12. [12]

    LOF: Identifying Density-Based Local Outliers,

    M. M. Breunig, H.-P. Kriegel, R. T. Ng and J. S. Sander, "LOF: Identifying Density-Based Local Outliers," Proc. ACM SIGMOD Int. Conf. Management of Data, pp. 93 –104, 2000

    work page 2000

  13. [13]

    STM32F446xC/E Datasheet,

    STMicroelectronics, "STM32F446xC/E Datasheet," Rev. 10, Jan. 2021. [Online]. Available: https://www.st.com/resource/en/datasheet/stm32f446re.pdf

    work page 2021

  14. [14]

    ACS712: Fully Integrated, Hall -Effect-Based Linear Current Sensor IC,

    Allegro MicroSystems, "ACS712: Fully Integrated, Hall -Effect-Based Linear Current Sensor IC," Datasheet. [Online]. Available: https://www.allegromicro.com/-/media/files/datasheets/acs712 - datasheet.ashx

    work page

  15. [15]

    IEEE Standard 1159-2019, IEEE Recommended Practice for Monitoring Electric Power Quality, IEEE, 2019

    work page 2019