Event Classification of Accelerometer Data for Industrial Package Monitoring with Embedded Deep Learning

Bastien Pasdeloup (IMT Atlantique); Hamoud Younes; Hugo Tessier; Manon Renault (IMT Atlantique); Mathieu L\'eonardon (IMT Atlantique); Ronan Le Roy

REVIEW 3 major objections 5 minor 21 references

Event Classification of Accelerometer Data for Industrial Package Monitoring with Embedded Deep Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 14 kB embedded neural network distinguishes forklift rides from truck rides on a reusable package with about 95% precision.

desk verdict A credible TinyML pipeline for package event monitoring whose reported 94.5/95.8% precision is an in-sample estimate because thresholds are tuned on the same 30% split used for the final numbers and the Dummy class is excluded from metrics. read the letter →

arxiv 2506.05435 v1 pith:364RON4T submitted 2025-06-05 cs.LG cs.AI

Manon Renault (IMT Atlantique) , Hamoud Younes , Hugo Tessier , Ronan Le Roy , Bastien Pasdeloup (IMT Atlantique) , Mathieu Léonardon (IMT Atlantique) This is my paper

classification cs.LGcs.AI

keywords embeddeddeeplearningTinyMLaccelerometereventclassificationpackagemonitoringreusablepackagingADASYNmodelcompressionESP32

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 an inexpensive, battery-powered device placed on a reusable shipping crate can tell, from accelerometer data alone, whether the crate is riding on a forklift or in a truck, and can do so accurately enough for real industrial logistics. The authors argue that this event-level awareness removes the need for always-on GPS, TDoA, or Bluetooth infrastructure and lets the device sleep until motion occurs, which is what makes multi-year battery life plausible. Their pipeline—a 1D convolutional neural network, ADASYN oversampling for an imbalanced time-series dataset, a confidence threshold that shunts uncertain predictions into a 'Dummy' class, and pruning plus 8-bit quantization—yields a 14 kB model with 94.54% precision on the forklift class and 95.83% on the truck class, running at 316 mW for 27 ms per inference on an ESP32. If the result holds outside the training facility where the data were collected, package tracking, loss attribution, and sustainable reusable-packaging logistics become feasible without dense wireless infrastructure.

What carries the argument

The load-bearing object is a one-dimensional convolutional neural network with convolutional layers, batch normalization, ReLU activations, and a softmax output over the classes Forklift, Truck, and Dummy. Three mechanisms carry the argument: ADASYN synthesizes minority-class samples near the decision boundary, a per-class confidence threshold maps low-confidence predictions into the Dummy class, and deep compression—L1-norm unstructured pruning with weight rewinding, full-int8 quantization through the TFLite converter, and Huffman coding—cuts the non-volatile memory footprint from 51 kB to 14 kB while preserving precision.

What would settle it

Take the same hardware and pipeline, collect a held-out dataset at two or more operating warehouses with different forklift and truck models, road surfaces, and handlers, label events by synchronized video, and measure precision and recall using the paper's fixed thresholds; if either per-class precision falls below about 90% or the Dummy rejection rate becomes so large that the device wakes too often, the central generalization claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a two-class accelerometer event classifier for package monitoring survives real embedded constraints. After training a baseline 1D CNN on 20 Hz tri-axial accelerometer data, applying ADASYN instead of SMOTE raises truck recall from 59.98% to 96.62% while keeping precision above threshold at 98.05% for forklift and 95.75% for truck. Pruning and full-integer quantization then shrink the deployed model from 51 kB to 14 kB, moving forklift precision to 94.54% and truck precision to 95.83%, with recall at 92.59% and 88.66%, respectively. The deployed pruned and quantized model runs on an ESP32 microcontroller at 316 mW during a 27 ms inference, barely above the 300 mW idle baseline, so event-triggered inference is cheap enough to support a device lifespan of several years. The 'Dummy' class is central to the design: predictions whose maximum softmax probability falls below a per-class threshold are assigned to Dummy rather than reported, letting the system trade recall for precision and absorb irrelevant events such as human handling, drops, or being opened.

Load-bearing premise

The load-bearing premise is that the data collected by four devices handled by trainees inside one logistics training facility represent the full range of real forklift and truck events a reusable package will encounter, and that the thresholds tuned on this dataset keep their precision when the device is deployed elsewhere.

Editorial extensions

If this is right

A reusable package can be monitored for two lifecycle events with no GPS, TDoA, or BLE infrastructure, because the 20 Hz accelerometer signal alone carries enough information.
Because the deployed model occupies only 14 kB, it fits entirely in the ESP32's SRAM, avoiding flash-memory access overhead and leaving room for other software modules.
With inference lasting 27 ms and drawing only slightly more than the 300 mW idle power, an event-triggered wake-up policy can support the multi-year lifespan required for reusable packaging.
Rejecting low-confidence predictions into a Dummy class lets the device deliberately sacrifice recall to avoid acting on uncertain events, reducing energy wasted on verification and communication.
ADASYN's boundary-focused oversampling substantially improves minority-class recall compared with SMOTE, which suggests it should be the default for imbalanced accelerometer event data.

Reading between the lines

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

If the generalization holds, the same thresholded-CNN recipe should transfer to other package-lifecycle events such as drops, tilts, or being opened, since the Dummy class already absorbs non-target motion; the testable extension is whether per-event precision stays high as the number of event classes grows.
The paper tunes thresholds to the precision-recall curve, but the economically optimal threshold also depends on the energy and bandwidth cost of a verification retry; a direct next step is to set thresholds by minimizing expected battery drain rather than maximizing precision alone.
Because the reported metrics exclude Dummy samples, a field deployment's true performance depends on the proportion of low-confidence events; an operational measure such as 'useful inferences per day' would combine per-class precision with the Dummy rejection rate.
The 14 kB footprint suggests the same ESP32 could host multiple specialized classifiers or a small ensemble, or fuse a second sensor, without leaving SRAM; this is a natural low-cost experiment consistent with the authors' stated multimodal future work.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Worth a look if you work on embedded ML or industrial IoT, with a caveat: the engineering story is coherent, but the headline precision is not a trustworthy estimate of field performance.

What is new and good: the application itself—distinguishing Forklift vs. Truck events on 20 Hz accelerometer data with a 14 kB pruned/quantized CNN on an ESP32—is not something I've seen in the two cited prior works. The compression pipeline (weight rewinding, unstructured pruning, INT8 quantization, Huffman coding) is standard but carefully executed, and the deployment details (ESP-IDF, SRAM fit, 27 ms inference, 316 mW measurement) are concrete and useful. The SMOTE/ADASYN comparison is also a reasonable contribution for imbalanced time series, though the results are only as solid as the evaluation they rest on.

The soft spots are real and load-bearing. Section III.D describes one split: 70% training, 30% for validation and testing. No separate held-out test set is mentioned. Section III.B says thresholds were adjusted for each class based on the model's performance, and Figure 2 shows precision-recall curves generated by varying thresholds. So the same 30% split is used to tune thresholds and to compute the final precision values in Tables I and II. The metric is selected on the test data. That alone makes the abstract's 94.54%/95.83% figures an in-sample estimate, likely optimistic.

Compounding this, the Dummy class is defined to include "any misclassified data" (Section III.A) and is then omitted from reported precision/recall. Low-confidence predictions get routed to Dummy, which is excluded from the denominators. Without a coverage figure (fraction of test events accepted as Forklift/Truck), the reader can't tell whether those precisions apply to all events or only a cherry-picked subset. The paper does not report error bars, repeated runs, or a baseline beyond SMOTE/ADASYN, and no code or data is released.

The claim that "no prior work on tracking using accelerometer signals has been addressed in the literature" is overstated; refs [5] and [6] are about activity and vibration classification, which is adjacent enough to require a more careful literature survey. The energy analysis is a minor weakness: the 300 mW baseline power when not inferring sounds high for a sleep state, and there is no battery-life model, despite lifespan being a stated goal.

Verdict: this is a plausible engineering system with a flawed evaluation protocol. It deserves a serious referee, but a revision should use a true held-out test set, tune thresholds on validation data only, report coverage and confidence intervals, add a baseline comparison, and honestly survey prior accelerometer-based tracking work. Not desk-reject, but major revision.

Referee Report

3 major / 5 minor

Summary. The paper describes a complete embedded deep-learning pipeline for classifying accelerometer events from reusable industrial packages into Forklift, Truck, and a Dummy rejection class. The pipeline uses a 1D CNN, SMOTE/ADASYN augmentation for class imbalance, per-class confidence thresholds, and pruning, quantization, and Huffman coding to compress the model to 14 kB. The authors report 94.54% precision for Forklift and 95.83% for Truck on the compressed model, with 316 mW inference power on an ESP32. The central claim is that this methodology enables accurate, low-power, long-lifetime package monitoring.

Significance. If the reported numbers are valid, the paper makes a useful applied contribution: it demonstrates that a very small (14 kB) CNN can run on an ESP32-class microcontroller and distinguish forklift from truck events from 20 Hz accelerometer data, with measured power consumption and a compression comparison. The deployment details, including the TFLite conversion and energy measurement, are valuable for TinyML practitioners. However, the evaluation protocol does not support the headline precision figures as unbiased estimates of field performance, and the Dummy-class exclusion makes the metrics difficult to interpret. The paper does not provide code or data, so reproducibility rests entirely on the textual description.

major comments (3)

[Section III.D and III.B] The evaluation uses a single 70/30 split with no separate held-out test set, and the same 30% partition is used to adjust per-class thresholds ('threshold values were adjusted for each class based on the model's performance') and to report the final precision/recall values in Tables I and II. Additionally, Section II.B states that the window length Vl was chosen based on the value that yielded the highest precision on this evaluation data. The reported metrics are therefore in-sample estimates after optimizing at least the thresholds and the window length on the test split. A proper evaluation would require a separate test set used only once, or nested cross-validation, with repeated runs and confidence intervals.
[Section III.A and II.C] The Dummy class is defined to include 'any misclassified data' and 'data that did not meet the threshold criteria,' and Section III.A states that Dummy 'was omitted from the performance results.' This creates a circular evaluation: samples that the model classifies with low confidence are relabeled as Dummy and then excluded from the precision/recall denominators, so the reported two-class metrics can be made arbitrarily high by raising the thresholds. The paper does not report coverage, i.e., the fraction of all events accepted as Forklift or Truck, nor a three-class confusion matrix that includes Dummy. Without this information, the headline precision values of 94.54% and 95.83% do not indicate performance on the full stream of events.
[Section III.A and III.D] All data were collected with four devices at a single logistics training facility, with trainees operating the trucks and forklifts, and the model is evaluated on a single split from that dataset. There are no repeated runs, no error bars, and no external validation set from a different facility or different handling conditions. Consequently, the conclusion in Section IV that the system 'demonstrating efficiency in real-life scenarios' is not supported by the evidence as presented. At minimum, the authors should report the number of independent recording sessions, the variability across devices, and results from a held-out subset of data collected at a different time or place.

minor comments (5)

[Section II.B] The variable name 'laxes' appears to be a typo for 'n_axes' or 'axes'; please rename for clarity and define it consistently with Section III.A's description of tri-axial accelerometer data.
[Section II.F.2] The text states that 'All weights, biases, and activation layers are converted from 32-byte floating-point to 8-byte integer format'; this should read '32-bit' and '8-bit'. The same issue appears in Section II.F.3 where 'encoding dictionary' is mentioned.
[Section III.E.3] The power measurement description would benefit from clarity on whether the 316 mW figure includes the sensor sampling and base microcontroller power, or only the additional inference power; the sentence 'the IoT device has a baseline power consumption of 300 mW' suggests the additional cost is 16 mW, which is a more meaningful quantity for battery-life calculations.
[Section I] The claim that 'no prior work on tracking using accelerometer signals has been addressed in the literature' is too strong given the cited references [5] and [6] on accelerometer-based activity classification; consider softening the statement to focus on package monitoring specifically.
[Section II.E.2] The ADASYN description is incomplete: the equation 'gi = ˆri × G' should specify how the synthetic sample vector is constructed from the original sample and its neighbors, and the text should clarify that the interpolation coefficient is drawn randomly per sample. As written, the formula does not define a time-series-consistent interpolation.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline precision figures are selected on the same 30% split used for final metrics, and a class defined partly as 'any misclassified data' is omitted from those metrics.

fitted input called prediction [Section III.B and III.D; Tables I and II; abstract precision values]
"The event detection dataset has been split into 70% training and 30% for validation and testing. ... Additionally, the threshold values were adjusted for each class based on the model's performance."

The abstract's 94.54% Forklift and 95.83% Truck precision values are computed on the same 30% split that was used to choose the per-class thresholds. Section III.D describes only a 70/30 split with no separate held-out test set, and Section III.B states the thresholds were 'adjusted for each class based on the model's performance.' Thus the reported precision is an in-sample optimum of a decision-rule parameter, not an unbiased estimate on unseen data. Selecting the threshold to maximize precision on the evaluation set and then reporting precision on that same set makes the headline metric partially a fitted quantity rather than an independent measurement.
self definitional [Section III.A and Section II.A; omission of Dummy from performance results]
"Dummy: this event describes all other movements that the package encountered (being moved by a person, opened, etc.), data that did not meet the threshold criteria (see Section II-D), and any misclassified data. ... Because the 'Dummy' class represents misclassified samples and insignificant events, it was omitted from the performance results."

The Dummy class is defined, in part, as 'any misclassified data' and as data that fails the model-dependent threshold, and then this class is omitted from the reported performance results. Because low-confidence or wrong predictions are routed into a class that is excluded from the Forklift/Truck precision and recall tables, errors can disappear from the metric by construction. The paper does not report the fraction of samples accepted as Forklift or Truck (coverage), so the reader cannot tell how much of the quoted precision is an artifact of excluding Dummy. This is a definitional circularity, though not a complete tautology since recall is still lowered when true positives are sent to Dummy.

full rationale

The core empirical pipeline—CNN training, data augmentation comparison, pruning, quantization, and energy measurement—is self-contained and does not reduce to its own inputs. However, the central headline numbers (94.54% and 95.83% precision) are weakened by two related protocol choices. First, the per-class thresholds are tuned on the same 30% split used for the final metrics, so the reported precision is an in-sample, threshold-optimized value rather than a held-out estimate. Second, the Dummy class is partly defined as misclassified or low-confidence data and then omitted from the performance results, allowing errors to be excluded by definition; without a coverage figure, the precision claims are not fully interpretable. The pruning ratio is taken from the authors' prior work [17], but this is a minor, non-load-bearing self-citation because the compression results are measured and reported in this paper independently. Overall, the methodology has independent content, but the headline evaluation is partially circular, so a moderate score is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central performance numbers depend on several hand-chosen hyperparameters (window length, thresholds, augmentation settings, pruning ratio, training schedule), a narrow single-site dataset, and a Dummy class that is designed and then excluded from the headline metrics. No external benchmark or public data is used to anchor the results.

free parameters (6)

Window length Vl = Not reported
Window length was chosen as the value that yielded the highest precision (Section II.B and III), but the value is never given.
Per-class confidence thresholds = Not reported, adjusted per configuration
Thresholds are adjusted for each class based on model performance (Section III.B), selecting the operating point used for reported precision.
SMOTE/ADASYN number of neighbors k = 5
Section III.C states 5 nearest neighbors were used, with no sensitivity analysis.
ADASYN beta = 0.4
Section III.C states beta = 0.4, a factor controlling the target balance between classes.
Pruning ratio = 50%
Section III.E.1 selects 50% pruning based on prior work [17] by the same group, not on the current dataset.
Training hyperparameters = LR 0.5, momentum 0.9, weight decay 2e-5, epochs 20, batch size 200
Section III.D lists these hand-chosen values; no tuning search or sensitivity is reported.

assumptions (5)

domain assumption 20 Hz tri-axial accelerometer signals are sufficient to discriminate Forklift from Truck events.
The dataset uses a 20 Hz sampling frequency (Section III.A) and the paper provides no analysis confirming this captures the relevant motion dynamics.
domain assumption The training-facility recordings are representative of real-world package lifecycles.
Section III.A says data were collected at a training facility for logistics technicians; no test data from operational sites is used.
domain assumption All relevant state changes are detectable as motion that can wake the device, and wake-on-motion is the dominant battery drain.
Section IV states wake time is key, but the paper does not measure sleep current or the duty cycle of the wake trigger.
ad hoc to paper The Dummy class can absorb low-confidence and ambiguous samples without biasing the two-class performance metrics.
Section II.A introduces Dummy as a catch-all and Section III.A removes it from reported results, changing the evaluation population.
domain assumption The TFLite model metrics computed before deployment equal the deployed model behavior on the ESP32.
Section II.F says metrics were computed on the TFLite model; the only on-device measurement is power, not classification accuracy.

invented entities (1)

The 'Dummy' class
purpose: A catch-all label for irrelevant events, low-confidence predictions, and misclassified samples, which is then excluded from reported metrics.
The Dummy class exists only inside the model's labeling scheme. It has no external falsifiable handle and directly affects the reported precision since all low-confidence predictions are routed into it and omitted from evaluation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Event Classification of Accelerometer Data for Industrial Package Monitoring with Embedded Deep Learning." pith.science (2026). https://pith.science/paper/364RON4T

@misc{pith2026250605435,
  author       = {Pith},
  title        = {Pith review of: Event Classification of Accelerometer Data for Industrial Package Monitoring with Embedded Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/364RON4T}},
  note         = {Machine review of arXiv:2506.05435}
}

read the original abstract

Package monitoring is an important topic in industrial applications, with significant implications for operational efficiency and ecological sustainability. In this study, we propose an approach that employs an embedded system, placed on reusable packages, to detect their state (on a Forklift, in a Truck, or in an undetermined location). We aim to design a system with a lifespan of several years, corresponding to the lifespan of reusable packages. Our analysis demonstrates that maximizing device lifespan requires minimizing wake time. We propose a pipeline that includes data processing, training, and evaluation of the deep learning model designed for imbalanced, multiclass time series data collected from an embedded sensor. The method uses a one-dimensional Convolutional Neural Network architecture to classify accelerometer data from the IoT device. Before training, two data augmentation techniques are tested to solve the imbalance problem of the dataset: the Synthetic Minority Oversampling TEchnique and the ADAptive SYNthetic sampling approach. After training, compression techniques are implemented to have a small model size. On the considered twoclass problem, the methodology yields a precision of 94.54% for the first class and 95.83% for the second class, while compression techniques reduce the model size by a factor of four. The trained model is deployed on the IoT device, where it operates with a power consumption of 316 mW during inference.

Figures

Figures reproduced from arXiv: 2506.05435 by the authors.

**Figure 1.** Detailed model architecture. ”Dummy” class. In case of a bad prediction, some algorithms can be implemented to perform corrections. However, this produces a relatively important number of operations, hence high energy consumption. To optimize classification performance, we select a threshold that maximizes precision [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Precision-recall curves with data augmentation techniques. Different [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages

[5]

Wearable sport activity classification based on deep convolutional neural network,

Y .-L. Hsu, H.-C. Chang, and Y .-J. Chiu, “Wearable sport activity classification based on deep convolutional neural network,” vol. 7, 2019, pp. 170 199–170 212

work page 2019
[6]

Machinery value estimation method based on iiot system utilizing 1d-cnn model for low sampling rate vibration signals from mems,

G. Lee, S.-W. Kim, and M. Jeon, “Machinery value estimation method based on iiot system utilizing 1d-cnn model for low sampling rate vibration signals from mems,” vol. 10, no. 14, 2023, pp. 12 261–12 275

work page 2023
[1]

Smart packaging: Opportunities and challenges,

D. Schaefer and W. M. Cheung, “Smart packaging: Opportunities and challenges,” Procedia CIRP , vol. 72, pp. 1022–1027, 2018, 51st CIRP Conference on Manufacturing Systems. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2212827118304104

work page 2018
[2]

Design and implementation of an accurate real time gps tracking system,

H. A. Abdallah Dafallah, “Design and implementation of an accurate real time gps tracking system,” in The Third International Conference on e-Technologies and Networks for Development (ICeND2014) , 2014, pp. 183–188

work page 2014
[3]

Indoor multiple sound source tracking using refined tdoa measurements,

L. Sun and Q. Cheng, “Indoor multiple sound source tracking using refined tdoa measurements,” in 2015 49th Annual Conference on Infor- mation Sciences and Systems (CISS) , 2015, pp. 1–5

work page 2015
[4]

Iot-td: Iot dataset for multiple model ble-based indoor localization/tracking,

M. Salimibeni, Z. Hajiakhondi-Meybodi, P. Malekzadeh, M. Atashi, K. N. Plataniotis, and A. Mohammadi, “Iot-td: Iot dataset for multiple model ble-based indoor localization/tracking,” in 2020 28th European Signal Processing Conference (EUSIPCO) , 2021, pp. 1697–1701

work page 2020
[7]

An empirical study on ai-powered edge computing architectures for real-time iot applications,

A. Yasmin, T. Mahmud, M. Debnath, and A. H. H. Ngu, “An empirical study on ai-powered edge computing architectures for real-time iot applications,” pp. 1422–1431, 2024

work page 2024
[8]

Tinyml smart sensor for energy saving in internet of things precision agriculture platform,

C. Nicolas, B. Naila, and R.-C. Amar, “Tinyml smart sensor for energy saving in internet of things precision agriculture platform,” in 2022 Thirteenth International Conference on Ubiquitous and Future Networks (ICUFN), 2022, pp. 256–259

work page 2022

Show all 21 references

[9]

tinycare: A tinyml-based low-cost continu- ous blood pressure estimation on the extreme edge,

K. Ahmed and M. Hassan, “tinycare: A tinyml-based low-cost continu- ous blood pressure estimation on the extreme edge,” in 2022 IEEE 10th International Conference on Healthcare Informatics (ICHI) , 2022, pp. 264–275

2022
[10]

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. 96 892–96 922, 2023

2023
[11]

SMOTE: Synthetic minority over-sampling technique,

N. V . Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: Synthetic minority over-sampling technique,” vol. 16, pp. 321–357, 2002. [Online]. Available: https://www.jair.org/index.php/jair/ article/view/10302

2002
[12]

ADASYN: Adaptive synthetic sampling approach for imbalanced learning,

H. He, Y . Bai, E. A. Garcia, and S. Li, “ADASYN: Adaptive synthetic sampling approach for imbalanced learning,” pp. 1322–1328, 2008, ISSN: 2161-4407. [Online]. Available: https://ieeexplore.ieee.org/ document/4633969/?arnumber=4633969

2008
[13]

Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,

S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,” 2016. [Online]. Available: http://arxiv.org/abs/1510.00149

2016 arXiv
[14]

Accessed: 2024-09-13

(2024) Introduction to onnx. Accessed: 2024-09-13. [Online]. Available: https://onnx.ai/onnx/intro/

2024
[15]

Accessed: 2024-09-13

(2024) Tensorflow lite micro. Accessed: 2024-09-13. [Online]. Available: https://www.tensorflow.org/lite/microcontrollers

2024
[16]

Optimal brain damage,

Y . LeCun, J. S. Denker, and S. A. Solla, “Optimal brain damage,” Advances in Neural Information Processing Systems 2 (NIPS 1989) , 1989

1989
[17]

Inter- operability of compression techniques for efficient deployment of CNNs on microcontrollers,

H. Younes, H. L. Blevec, M. L ´eonardon, and V . Gripon, “Inter- operability of compression techniques for efficient deployment of CNNs on microcontrollers,” in Advances in System-Integrated Intelligence . Springer International Publishing, 2023, pp. 543–552

2023
[18]

Learning both weights and con- nections for efficient neural network,

S. Han, J. Pool, J. Tran, and W. Dally, “Learning both weights and con- nections for efficient neural network,” inAdvances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, Eds., vol. 28. Curran Associates, Inc., 2015

2015
[19]

Quantized convolutional neural networks for mobile devices,

J. Wu, C. Leng, Y . Wang, Q. Hu, and J. Cheng, “Quantized convolutional neural networks for mobile devices,” 2016. [Online]. Available: https://arxiv.org/abs/1512.06473

2016 arXiv
[20]

A method for the construction of minimum-redundancy codes,

D. A. Huffman, “A method for the construction of minimum-redundancy codes,” Proceedings of the IRE , vol. 40, no. 9, pp. 1098–1101, 1952

1952
[21]

Accessed: 2024-11-27

(2024) Esp-idf programming guide. Accessed: 2024-11-27. [Online]. Available: https://docs.espressif.com/projects/esp-idf/en/latest/esp32s3/

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [5]

Wearable sport activity classification based on deep convolutional neural network,

Y .-L. Hsu, H.-C. Chang, and Y .-J. Chiu, “Wearable sport activity classification based on deep convolutional neural network,” vol. 7, 2019, pp. 170 199–170 212

work page 2019

[2] [6]

Machinery value estimation method based on iiot system utilizing 1d-cnn model for low sampling rate vibration signals from mems,

G. Lee, S.-W. Kim, and M. Jeon, “Machinery value estimation method based on iiot system utilizing 1d-cnn model for low sampling rate vibration signals from mems,” vol. 10, no. 14, 2023, pp. 12 261–12 275

work page 2023

[3] [1]

Smart packaging: Opportunities and challenges,

D. Schaefer and W. M. Cheung, “Smart packaging: Opportunities and challenges,” Procedia CIRP , vol. 72, pp. 1022–1027, 2018, 51st CIRP Conference on Manufacturing Systems. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2212827118304104

work page 2018

[4] [2]

Design and implementation of an accurate real time gps tracking system,

H. A. Abdallah Dafallah, “Design and implementation of an accurate real time gps tracking system,” in The Third International Conference on e-Technologies and Networks for Development (ICeND2014) , 2014, pp. 183–188

work page 2014

[5] [3]

Indoor multiple sound source tracking using refined tdoa measurements,

L. Sun and Q. Cheng, “Indoor multiple sound source tracking using refined tdoa measurements,” in 2015 49th Annual Conference on Infor- mation Sciences and Systems (CISS) , 2015, pp. 1–5

work page 2015

[6] [4]

Iot-td: Iot dataset for multiple model ble-based indoor localization/tracking,

M. Salimibeni, Z. Hajiakhondi-Meybodi, P. Malekzadeh, M. Atashi, K. N. Plataniotis, and A. Mohammadi, “Iot-td: Iot dataset for multiple model ble-based indoor localization/tracking,” in 2020 28th European Signal Processing Conference (EUSIPCO) , 2021, pp. 1697–1701

work page 2020

[7] [7]

An empirical study on ai-powered edge computing architectures for real-time iot applications,

A. Yasmin, T. Mahmud, M. Debnath, and A. H. H. Ngu, “An empirical study on ai-powered edge computing architectures for real-time iot applications,” pp. 1422–1431, 2024

work page 2024

[8] [8]

Tinyml smart sensor for energy saving in internet of things precision agriculture platform,

C. Nicolas, B. Naila, and R.-C. Amar, “Tinyml smart sensor for energy saving in internet of things precision agriculture platform,” in 2022 Thirteenth International Conference on Ubiquitous and Future Networks (ICUFN), 2022, pp. 256–259

work page 2022