Pith. sign in

REVIEW 4 major objections 6 minor 35 references

A Cost-effective, Stand-alone, and Real-time TinyML-Based Gait Diagnosis Unit Aimed at Lower-limb Robotic Prostheses and Exoskeletons

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A wearable gait diagnosis unit built from two IMUs and a $7.30 ESP32 microcontroller classifies five gait scenarios with 92% accuracy and returns an anomaly score within 96 ms.

desk verdict A credible low-cost TinyML gait diagnosis prototype whose 92% accuracy claim is not yet shown to generalize beyond five healthy volunteers. read the letter →

arxiv 2411.08474 v1 pith:25AQHU2V submitted 2024-11-13 cs.RO

Zarin Anjum Madhiha , Antar Mazumder , Sohani Munteha Hiam This is my paper · ORCID
classification cs.RO
keywords TinyMLgaitclassificationanomalydetectionaffordableprosthesesexoskeletonsinertialmeasurementunitquantizedneuralnetworksembeddedAI
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

The paper argues that the expensive diagnosis-and-control computing unit is a major reason robotic prostheses and exoskeletons are inaccessible, and that it can be replaced by a tiny standalone sensor-and-microcontroller package. As a proof of concept, it builds a wearable gait diagnosis unit from two six-axis IMUs and an ESP32 NodeMCU that runs two parallel quantized int8 models: one neural network that classifies idle, walk, run, hopping, and skip, and one K-means model that assigns an anomaly score. The reported result is 92% overall classification accuracy and 95–96 ms of inference time after a 3-second window of acceleration data, at a component cost of about $7.30. A sympathetic reader would take the contribution to be the demonstration that near-PC-level gait analysis can run entirely on a board costing a few dollars, with no external computer or costly controller.

What carries the argument

The load-bearing mechanism is a parallel two-model pipeline designed for a resource-limited microcontroller. The classifier is a shallow artificial neural network fed by 222 features per 3-second window: for each of six acceleration axes, mean-subtracted and low-pass filtered data yield RMS, kurtosis, and skewness, while a 64-bin FFT yields spectral skewness, spectral kurtosis, and spectral power. The anomaly detector is a K-means model that takes four spectral-power features selected by feature importance, computes a 32-point cluster centroid, and uses the Euclidean distance from the test point to that centroid as the anomaly score, with a threshold of 3.0. The whole computation is quantized to 8-bit integers so both models fit and run on the ESP32.

What would settle it

Run the same GDU on a held-out group—particularly lower-limb prosthesis users or adults outside the original five—and compute the five-class confusion matrix and anomaly scores; if overall accuracy falls substantially below 92% or normal gait routinely produces anomaly scores above 3.0, the generalization claim does not hold.

Watch

Extended reading notes

Core claim

The central claim is that the cost bottleneck in lower-limb robotic devices—the diagnosis unit—can be moved onto a microcontroller without losing practical performance. Concretely, the paper reports that a unit made from two MPU6050 accelerometers and an ESP32 NodeMCU, running two parallel quantized int8 models, classifies five gait scenarios with 92% overall accuracy and produces an anomaly score within 95–96 ms after a 3-second window. The authors emphasize that the device is stand-alone and modular, so it can be strapped to any exoskeleton or prosthesis. They also report that the float32 version of the same model reaches 97.14% accuracy, indicating that most of the gap to PC-based systems comes from int8 quantization rather than from the feature set.

Load-bearing premise

The load-bearing premise is that models trained on five healthy volunteers, with three minutes of each activity per person, transfer to the larger population of prosthesis and exoskeleton users, combined with the assumption that the single anomaly threshold of 3.0 works across users and devices.

Editorial extensions

If this is right

  • A gait diagnosis unit with the same architecture can be fitted to any exoskeleton or prosthesis because the sensor module is modular and strapped rather than integrated.
  • The diagnosis cost drops to roughly $7.30 in components, compared with the 500–1000+ USD controllers and PCs used in the systems the paper compares against.
  • A controller can receive a gait classification and anomaly score within about 96 ms after collecting 3 seconds of data, which is fast enough for use in real-time control loops.
  • Quantizing the model from float32 to int8 costs about five percentage points of accuracy (97.14% to 92%) but enables deployment on the cheap microcontroller.
  • Anomaly detection does not require segmentation into gait phases; it is a distance from a spectral-feature centroid, so the same logic works across different movement styles.

Reading between the lines

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

  • Editorial inference: the 3.0 anomaly threshold was set by trial and error on five healthy volunteers, so per-user calibration would likely be needed before the unit is used clinically.
  • Editorial inference: because the classifier was trained only on healthy gait, its 92% accuracy is untested for amputee or pathological gait; a direct evaluation on prosthesis users would be the natural next experiment.
  • Editorial inference: the 3-second collection window means a control decision is delayed by at least 3 seconds; for real-time prosthesis control, a sliding or shorter window would be needed.
  • Editorial inference: the latency breakdown shows 92 ms for feature generation, so using a faster board or reducing FFT bins could shrink the window without changing the model.
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

4 major / 6 minor

Summary. The paper proposes a low-cost wearable gait diagnosis unit (GDU) built around an ESP32 NodeMCU and two MPU6050 IMUs, running a quantized int8 ANN to classify five gait scenarios (idle, walk, run, hopping, skip) and a parallel K-means model to generate an anomaly score. The authors report an overall classification accuracy of 92%, an inference latency of 95–96 ms after a 3-second data collection window, and a hardware cost of 7.30 USD. The work is positioned as a proof of concept for affordable, stand-alone, real-time gait diagnosis for lower-limb robotic prostheses and exoskeletons, with a comparison to prior systems and a discussion of limitations.

Significance. If substantiated, the system would be a valuable low-cost contribution to wearable gait monitoring for assistive devices. The strengths are the real embedded implementation with measured latency, the explicit cost comparison, the use of TinyML/int8 quantization, and the modular hardware design. However, the evidence base is very small: five healthy volunteers, 39 test samples, no cross-validation or confidence intervals, and an anomaly threshold chosen by trial and error. The strong claims of generalizability to prosthesis and exoskeleton users go well beyond what the presented data can support.

major comments (4)
  1. [Section 3.2 and Table 1] The evaluation uses only five healthy volunteers (ages 21–55) with no reported train/test split. Table 1 reports 39 test samples, and the paper never states that the test subjects were disjoint from the training subjects. If the test samples are drawn from the same five volunteers, the 92% accuracy likely reflects person-specific gait characteristics (sensor placement, cadence, execution style) rather than generalizable patterns. Moreover, the intended population—lower-limb prosthesis and exoskeleton users—was not included. This directly undermines the abstract's claim that the unit 'could be fitted to any prosthesis or exoskeleton.' Please provide the exact train/test split methodology, per-subject performance, and ideally an evaluation on the target population.
  2. [Section 4.1, Figure 7, and Section 5] The anomaly threshold of 3.0 is acknowledged in Section 4.1 and Section 5 to be determined by trial and error and to vary across age groups. It is then applied to the same users whose data were used to set it, and no true anomalous gait samples are tested. The anomaly-score utility is therefore not validated; the reported 'average anomaly scores' in Figure 7 are all below 3.0, but this is an artifact of the post hoc threshold choice. To support the anomaly-detection claim, the authors need an independent test set with known anomalies and either a fixed, pre-specified threshold or a principled threshold-selection method with separate validation.
  3. [Section 4.1, Table 1] The overall accuracy of 92% rests on only 39 test samples, with per-class support between 7 and 10. No confidence intervals, cross-validation, or statistical significance tests are reported. The macro and weighted averages (0.93–0.95) are computed on this same tiny set. At this sample size, the difference between 92% and lower accuracies is not statistically meaningful, so the claim 'effectively classify the gait scenarios with an overall accuracy of 92%' is stronger than the evidence warrants. The authors should report confidence intervals and ideally use leave-one-subject-out or a larger held-out set.
  4. [Section 3.1 and 3.4 and Abstract] The deployment was on an aluminum-alloy exoskeleton prototype with IMUs strapped to the mid-thigh and ankle of healthy volunteers. This does not support the broad claim that the unit 'could be fitted to any prosthesis or exoskeleton.' In a real prosthesis, the IMU would be mounted on a pylon or socket, and the gait kinematics of amputees differ substantially from those of healthy volunteers. Even for exoskeletons, the interface and load transfer differ from the prototype used. Please either restrict the claim to a proof-of-concept on healthy subjects or include additional tests on representative prosthetic and exoskeleton hardware with impaired gaits.
minor comments (6)
  1. [Section 3.3] The text reads 'Ege Impulse studio' but should be 'Edge Impulse studio'; the same typo appears in Section 3.2.
  2. [Section 4] In the list of result-analysis criteria, 'accuracy' is duplicated: 'precision, recall, accuracy, accuracy, f1-score' should be 'precision, recall, accuracy, F1-score.'
  3. [Section 5] The phrase 'rapidly device custom GDUs' should be 'rapidly devise custom GDUs.'
  4. [Table 2] The float32 accuracy of 97.14% is listed in Table 2 but is not supported by any confusion matrix, table, or description in Section 4; please clarify where and how this value was measured.
  5. [Figure 7] The caption says 'average anomaly scores' while the text refers to 'average highest anomaly scores'; please align these terms for clarity.
  6. [References] Reference [4] is cited for the WHO estimate on prostheses access but appears to be the 'Global vector control response 2017-2030'; this citation seems unrelated and should be replaced with the correct WHO source.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity in the K-means anomaly-threshold calibration; the 92% ANN classification claim is an independent metric, though its external validity is not established.

  1. fitted input called prediction [Section 3.3 (anomaly threshold), Section 4 (no performance parameter), Section 4.1 (Figure 7), Section 5 (trial-and-error admission)]
    "In this study, a tolerance of 3 was initially considered as the threshold. Therefore, any gait sample producing a higher score was considered as an anomaly. ... There was no specific performance parameter for the k-means-based anomaly detection as the anomaly threshold was arbitrary and subject to a variable for different users. Nevertheless, the max value of this threshold was kept below 3.0. ... the highest average maximum anomaly value was found to be around 2.38, which was well below the set threshold of 3. ..."

    The 3.0 anomaly threshold is an input selected by trial and error on the same normal gait data used in the evaluation, as the paper explicitly concedes. Reporting that normal samples produce scores below 3.0 (max 2.38) is therefore a consequence of the calibration choice, not an independent validation of anomaly detection. The paper also states there is no specific performance parameter for the K-means anomaly detector and that the threshold varies across users, so no independent anomalous test set or algorithmic threshold determination supports the anomaly-score utility. This circularity is confined to the anomaly-score component; the ANN classification accuracy is a separate, independent performance metric.

full rationale

The central classification claim, 92% overall accuracy, is an independent performance metric based on the confusion matrix in Figure 6 and Table 1; nothing in the paper defines the ANN's weights in terms of that reported accuracy. The paper does not specify a train/test split, so the accuracy's generalizability to new users is not established, but that is a data-leakage/external-validity concern rather than a circular derivation. The only identifiable circular step is the K-means anomaly threshold: the 3.0 tolerance was set by trial and error (Section 5), and then normal samples being below that threshold is reported as a finding. This is a fitted input used as if it were an independent check, but the paper explicitly disclaims any anomaly-detection performance parameter, limiting its impact. There are no load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling citations. Overall, the classification result is self-contained, with only minor circularity in the anomaly-threshold calibration, so a score of 2 is appropriate.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claims rest on several unvalidated assumptions, most importantly the representativeness of the 5-volunteer dataset and the sufficiency of the chosen features. The anomaly threshold is a free parameter fitted to the data rather than derived. No new entities are postulated.

free parameters (6)
  • Anomaly threshold = 3.0
    Threshold for flagging a sample as anomalous; set by trial and error (Section 4.1, Section 5).
  • K-means cluster size (Nc) = 32
    Number of points used to define each cluster in the anomaly detector (Section 3.3).
  • Dropout rate = 0.5
    Dropout applied to the second hidden layer of the ANN (Section 3.3).
  • Learning rate = 0.0005
    Adam optimizer learning rate (Section 3.3).
  • Low-pass filter cutoff = 3 Hz
    6th-order Butterworth filter cutoff used in preprocessing (Section 3.2).
  • FFT length = 64
    Number of FFT bins for spectral features (Section 3.2).
assumptions (4)
  • domain assumption The selected IMU positions (mid-thigh and ankle) are suitable for classifying the five gait scenarios.
    Authors rely on [20] for optimal sensor placement (Section 3.1).
  • domain assumption A 3-second window of 6-axis acceleration at 59 Hz contains enough information to distinguish idle, walk, run, hopping, and skip.
    The feature extraction uses one 3-second window per sample (Section 3.2).
  • domain assumption Data from five healthy volunteers (ages 21-55) is representative of the target prosthesis/exoskeleton user population.
    Model is trained exclusively on these subjects (Section 3.2) and no external validation is performed.
  • domain assumption The parallel K-means model trained on normal gait samples can identify deviations by distance from cluster centroids.
    The anomaly score is the average distance to the centroid (Section 3.3, Equations 12-14).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Cost-effective, Stand-alone, and Real-time TinyML-Based Gait Diagnosis Unit Aimed at Lower-limb Robotic Prostheses and Exoskeletons." pith.science (2026). https://pith.science/paper/25AQHU2V

@misc{pith2026241108474,
  author       = {Pith},
  title        = {Pith review of: A Cost-effective, Stand-alone, and Real-time TinyML-Based Gait Diagnosis Unit Aimed at Lower-limb Robotic Prostheses and Exoskeletons},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/25AQHU2V}},
  note         = {Machine review of arXiv:2411.08474}
}
read the original abstract

Robotic prostheses and exoskeletons can do wonders compared to their non-robotic counterpart. However, in a cost-soaring world where 1 in every 10 patients has access to normal medical prostheses, access to advanced ones is, unfortunately, extremely limited especially due to their high cost, a significant portion of which is contributed to by the diagnosis and controlling units. However, affordability is often not a major concern for developing such devices as with cost reduction, performance is also found to be deducted due to the cost vs. performance trade-off. Considering the gravity of such circumstances, the goal of this research was to propose an affordable wearable real-time gait diagnosis unit (GDU) aimed at robotic prostheses and exoskeletons. As a proof of concept, it has also developed the GDU prototype which leveraged TinyML to run two parallel quantized int8 models into an ESP32 NodeMCU development board (7.30 USD) to effectively classify five gait scenarios (idle, walk, run, hopping, and skip) and generate an anomaly score based on acceleration data received from two attached IMUs. The developed wearable gait diagnosis stand-alone unit could be fitted to any prosthesis or exoskeleton and could effectively classify the gait scenarios with an overall accuracy of 92% and provide anomaly scores within 95-96 ms with only 3 seconds of gait data in real-time.

Figures

Figures reproduced from arXiv: 2411.08474 by the authors.

Figure 1
Figure 1. Disparity in accessibility to prostheses and assistive devices: Demand vs. Access and the major catalyst. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Disparity in research focus: Robotic Prostheses vs. Affordable Robotic Prostheses based on the yearly number [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The GDU physical form and schematics and raw acceleration data from A1 and A2 for the five gait scenarios [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Gait data preprocessing and feature generation. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The parallel models leverage ANN to classify gait scenarios and K-means clustering to throw an anomaly [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Confusion matrix for the Neural Network-based classifier. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Average anomaly scores for users from different age groups: Higher margin visible among the youth. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Classification results and corresponding latency breakdown after data collection derived from the Arduino [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 32 canonical work pages

  1. [1]

    Advancements in limb prosthetics,

    Sitn, “Advancements in limb prosthetics,” Science in the News, 2013

    work page 2013

  2. [2]

    Speed-adaptation mechanism: Robotic prostheses can actively regulate joint torque,

    T. Lenzi, L. Hargrove, and J. Sensinger, “Speed-adaptation mechanism: Robotic prostheses can actively regulate joint torque,” IEEE Robotics & Automation Magazine, vol. 21, no. 4, pp. 94–107, 2014

    2014

  3. [3]

    Robotic exoskeletons: The current pros and cons,

    A. S. Gorgey, “Robotic exoskeletons: The current pros and cons,” World journal of orthopedics, vol. 9, no. 9, p. 112, 2018

    work page 2018

  4. [4]

    W. H. Organization, Global vector control response 2017-2030. Geneva: World Health Organization, 2017

    work page 2017

  5. [5]

    Technology for monitoring everyday prosthesis use: a systematic review,

    A. Chadwell, L. Diment, M. Micó-Amigo, D. Z. Morgado Ramírez, A. Dickinson, M. Granat, L. Kenney, S. Kheng, M. Sobuh, R. Ssekitoleko, et al., “Technology for monitoring everyday prosthesis use: a systematic review,” Journal of neuroengineering and rehabilitation, vol. 17, pp. 1–26, 2020

    work page 2020

  6. [6]

    Advances in active knee brace technology: A review of gait analysis, actuation, and control applications,

    U. Trivedi and A. Y . Joshi, “Advances in active knee brace technology: A review of gait analysis, actuation, and control applications,” Heliyon, 2024

    work page 2024

  7. [7]

    Time series classification using a modified lstm approach from accelerometer-based data: A comparative study for gait cycle detection,

    H. X. Tan, N. N. Aung, J. Tian, M. C. H. Chua, and Y . O. Yang, “Time series classification using a modified lstm approach from accelerometer-based data: A comparative study for gait cycle detection,” Gait & posture, vol. 74, pp. 128–134, 2019

    work page 2019

  8. [8]

    Walking gait phase detection based on acceleration signals using lstm-dnn algorithm,

    T. Zhen, L. Yan, and P. Yuan, “Walking gait phase detection based on acceleration signals using lstm-dnn algorithm,” Algorithms, vol. 12, no. 12, p. 253, 2019

    work page 2019

Show all 35 references
  1. [9]

    Continuous gait phase estimation using lstm for robotic transfemoral prosthesis across walking speeds,

    J. Lee, W. Hong, and P. Hur, “Continuous gait phase estimation using lstm for robotic transfemoral prosthesis across walking speeds,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 29, pp. 1470–1477, 2021

    2021

  2. [10]

    Piecewise linear labeling method for speed-adaptability enhancement in human gait phase estimation,

    W. Hong, J. Lee, and P. Hur, “Piecewise linear labeling method for speed-adaptability enhancement in human gait phase estimation,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 31, pp. 628–635, 2022

    2022

  3. [11]

    Effect of torso kinematics on gait phase estimation at different walking speeds,

    W. Hong, J. Lee, and P. Hur, “Effect of torso kinematics on gait phase estimation at different walking speeds,” Frontiers in Neurorobotics, vol. 16, p. 807826, 2022

    2022

  4. [12]

    Gait phase recognition using deep convolutional neural network with inertial measurement units,

    B. Su, C. Smith, and E. Gutierrez Farewik, “Gait phase recognition using deep convolutional neural network with inertial measurement units,” Biosensors, vol. 10, no. 9, p. 109, 2020

    2020

  5. [13]

    Bpnn-based real-time recognition of locomotion modes for an active pelvis orthosis with different assistive strategies,

    C. Gong, D. Xu, Z. Zhou, N. Vitiello, and Q. Wang, “Bpnn-based real-time recognition of locomotion modes for an active pelvis orthosis with different assistive strategies,”International Journal of Humanoid Robotics, vol. 17, no. 01, p. 2050004, 2020

    2020

  6. [14]

    Wearable inertial sensor system towards daily human kinematic gait analysis: benchmarking analysis to mvn biomech,

    J. Figueiredo, S. P. Carvalho, J. P. Vilas-Boas, L. M. Gonçalves, J. C. Moreno, and C. P. Santos, “Wearable inertial sensor system towards daily human kinematic gait analysis: benchmarking analysis to mvn biomech,” Sensors, vol. 20, no. 8, p. 2185, 2020

    2020

  7. [15]

    Real-time gait event detection for a lower extremity exoskeleton robot by infrared distance sensors,

    C. Li, Y . He, T. Chen, X. Chen, and S. Tian, “Real-time gait event detection for a lower extremity exoskeleton robot by infrared distance sensors,” IEEE Sensors Journal, vol. 21, no. 23, pp. 27116–27123, 2021

    2021

  8. [16]

    Gait phase classification for a lower limb exoskeleton system based on a graph convolutional network model,

    X. Wu, Y . Yuan, X. Zhang, C. Wang, T. Xu, and D. Tao, “Gait phase classification for a lower limb exoskeleton system based on a graph convolutional network model,” IEEE Transactions on Industrial Electronics, vol. 69, no. 5, pp. 4999–5008, 2021

    2021

  9. [17]

    Walking-speed-adaptive gait phase estimation for wearable robots,

    S. Choi, C. Ko, and K. Kong, “Walking-speed-adaptive gait phase estimation for wearable robots,” Sensors, vol. 23, no. 19, p. 8276, 2023

    2023

  10. [18]

    Real-time terrain recognition based on transformer and gait prediction based on cnn-lstm-attention for exosuit,

    F. Yang, C. Chen, Z. Wang, Y . Xu, Y . Liu, and X. Wu, “Real-time terrain recognition based on transformer and gait prediction based on cnn-lstm-attention for exosuit,” in 2023 IEEE International Conference on Real-time Computing and Robotics (RCAR), pp. 743–748, IEEE, 2023

    2023

  11. [19]

    Actuator optimization and deep learning-based control of pediatric knee exoskeleton for community-based mobility assistance,

    S. Zhang, J. Zhu, T.-H. Huang, S. Yu, J. S. Huang, I. Lopez-Sanchez, T. Devine, M. Abdelhady, M. Zheng, T. C. Bulea, et al., “Actuator optimization and deep learning-based control of pediatric knee exoskeleton for community-based mobility assistance,” Mechatronics, vol. 97, p....

    2024

  12. [20]

    Optimal locations and computational frameworks of fsr and imu sensors for measuring gait abnormalities,

    S. K. Manna, M. H. B. Azhar, and A. Greace, “Optimal locations and computational frameworks of fsr and imu sensors for measuring gait abnormalities,” Heliyon, 2023

    2023

  13. [21]

    Edge impulse cli - edge impulse documentation,

    E. Impulse, “Edge impulse cli - edge impulse documentation,” 2024

    2024

  14. [22]

    A waveform skewness index for measuring time series nonlinearity and its applications to the enso–indian monsoon relationship,

    J. Schulte, F. Policelli, and B. Zaitchik, “A waveform skewness index for measuring time series nonlinearity and its applications to the enso–indian monsoon relationship,” Nonlinear Processes in Geophysics, vol. 29, no. 1, pp. 1–15, 2022

    2022

  15. [23]

    An empirical study of the behaviour of the sample kurtosis in samples from symmetric stable distributions,

    J. M. Van Zyl and S. Van der Merwe, “An empirical study of the behaviour of the sample kurtosis in samples from symmetric stable distributions,” South African Statistical Journal, vol. 54, no. 2, pp. 255–265, 2020

    2020

  16. [24]

    Spectral kurtosis-based rfi mitigation for chime,

    J. Taylor, N. Denman, K. Bandura, P. Berger, K. Masui, A. Renard, I. Tretyakov, and K. Vanderlinde, “Spectral kurtosis-based rfi mitigation for chime,” Journal of Astronomical Instrumentation, vol. 8, no. 01, p. 1940004, 2019

    2019

  17. [25]

    Power spectrum and correlation,

    S. V . Vaseghi, “Power spectrum and correlation,” Advanced Digital Signal Processing and Noise Reduction, Saeed V . Vaseghi Copyright© 2000 John Wiley & Sons Ltd ISBNs: 0-471-62692-9 (Hardback): 0-470-84162-1 (Electronic), p. 263, 2000

    2000

  18. [26]

    Actuator and sensor fault classification for wind turbine systems based on fast fourier transform and uncorrelated multi-linear principal component analysis techniques,

    Y . Fu, Z. Gao, Y . Liu, A. Zhang, and X. Yin, “Actuator and sensor fault classification for wind turbine systems based on fast fourier transform and uncorrelated multi-linear principal component analysis techniques,” Processes, vol. 8, no. 9, p. 1066, 2020

    2020

  19. [27]

    Deep learning using rectified linear units (relu),

    A. F. Agarap, “Deep learning using rectified linear units (relu),” arXiv preprint arXiv:1803.08375, 2018

    2018 arXiv

  20. [28]

    Activation functions in neural networks,

    S. Sharma, S. Sharma, and A. Athaiya, “Activation functions in neural networks,”Towards Data Sci, vol. 6, no. 12, pp. 310–316, 2017

    2017

  21. [29]

    Taming the cross entropy loss,

    M. Martinez and R. Stiefelhagen, “Taming the cross entropy loss,” in German Conference on Pattern Recognition, pp. 628–637, Springer, 2018

    2018

  22. [30]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  23. [31]

    Anomaly detection (k-means) edge impulse documentation

    “Anomaly detection (k-means) edge impulse documentation.” https://docs.edgeimpulse.com/docs/ edge-impulse-studio/learning-blocks/anomaly-detection , 2024. Accessed: March 12, 2024

    2024

  24. [32]

    Implementation of k-means algorithm for clustering corn planting feasibility area in south lampung regency,

    A. Aldino, D. Darwis, A. Prastowo, and C. Sujana, “Implementation of k-means algorithm for clustering corn planting feasibility area in south lampung regency,” inJournal of Physics: Conference Series, vol. 1751, p. 012038, IOP Publishing, 2021

    2021

  25. [33]

    Tinyml for ultra-low power ai and large scale iot deployments: A systematic review,

    N. Schizas, A. Karras, C. Karras, and S. Sioutas, “Tinyml for ultra-low power ai and large scale iot deployments: A systematic review,”Future Internet, vol. 14, no. 12, p. 363, 2022

    2022

  26. [34]

    Real-time clinical gait analysis and foot anomalies detection using pressure sensors and convolutional neural network,

    M. Islam, M. Tabassum, M. M. Nishat, F. Faisal, and M. S. Hasan, “Real-time clinical gait analysis and foot anomalies detection using pressure sensors and convolutional neural network,” in 2022 7th International Conference on Business and Industrial Research (ICBIR), pp. 717–7...

    2022

  27. [35]

    A standalone computing system to classify human foot movements using machine learning techniques for ankle-foot prosthesis control,

    S. Negi and N. Sharma, “A standalone computing system to classify human foot movements using machine learning techniques for ankle-foot prosthesis control,” Computer Methods in Biomechanics and Biomedical Engineering, vol. 25, no. 12, pp. 1370–1380, 2022. 13

    2022

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.