Pith. sign in

REVIEW 3 major objections 6 minor 19 references

AI-Assisted NLOS Sensing for RIS-Based Indoor Localization in Smart Factories

T0 review · 3 major / 6 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 customized CNN trained on CIR spectrograms can classify LOS/NLOS for RIS-based factory localization with 95–99% accuracy, outperforming VGG-16.

desk verdict The measurement campaign is real, but the headline accuracy figures don't match the paper's own Table II and the evaluation protocol is too thin to support any of the performance claims. read the letter →

arxiv 2505.15989 v1 pith:YPMDH6IS submitted 2025-05-21 eess.SP

Taofeek A.O. Yusuf , Sigurd S. Petersen , Puchu Li , Jian Ren , Placido Mursia , Vincenzo Sciancalepore , Xavier Costa Pérez , Gilberto Berardinelli
show 1 more author
Ming Shen
This is my paper · ORCID
classification eess.SP
keywords reconfigurableintelligentsurfacesNLOSclassificationLOS/NLOSsensingindoorlocalizationsmartfactoryconvolutionalneuralnetworkchannelimpulseresponsespectrogram
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 sets out to show that a small, purpose-built convolutional neural network can reliably tell line-of-sight (LOS) from non-line-of-sight (NLOS) conditions in reconfigurable-intelligent-surface (RIS) measurements, a prerequisite for trustworthy indoor localization in smart factories. The authors collected channel impulse responses rendered as spectrogram images in three physical environments — an anechoic chamber, a meeting room, and a high-frequency lab — and augmented them to create synthetic, mixed-measured, and mixed-synthetic variants. They report that their customized CNN reaches 95.0–99.0% classification accuracy across these datasets, beating the standard pre-trained VGG-16 network (85.5–88.0%). The wider point is that AI-assisted NLOS sensing could make RIS-based factory localization safer and more reliable, and that dedicated small models can surpass large generic ones when trained on measured radio data.

What carries the argument

The central object is the customized CNN (cCNN), whose input is a 224×224 three-channel spectrogram image computed from the measured channel impulse response (CIR). The network consists of three convolutional blocks — each with a 3×3 convolution, batch normalization, ReLU activation, and max-pooling — followed by flattening, a fully connected layer with 256 units, dropout at $p=0.5$, and a softmax output over three classes (LOS, NLOS with a 1 m metal plate, NLOS with a 0.75 m metal plate). This architecture is expressed as the composite function $f(X)$ in Equation (1). The machinery works by learning spatial patterns in the spectrograms that distinguish direct-path from blocked or reflected propagation, and it is compared against VGG-16 on the same image inputs.

What would settle it

Run the trained cCNN and VGG-16 on a freshly measured set of CIR spectrograms from the same three rooms, with all synthetic and augmented data excluded, and compare the classification accuracy; if the custom CNN no longer tops VGG-16 by the reported margins, or if its accuracy drops below about 86% on the HF-lab set, the paper's core superiority claim would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that its custom CNN classifies LOS/NLOS conditions from CIR spectrograms with accuracy between 95.0% and 99.0% across three indoor environments, outperforming a standard pre-trained VGG-16, reported at 85.5–88.0%. In the per-environment tables the custom CNN's measured-data accuracies are 95.0% (meeting room), 86.0% (HF lab), and 99.9% (anechoic chamber); on synthetic and mixed datasets it stays above 86% in most cases. The authors argue this demonstrates that a small, purpose-built network can learn robust propagation features from real measured radio data, including under NLOS blockage by metal plates, and that the approach is useful when only limited or noisy measurement data are available.

Load-bearing premise

Every accuracy figure in the paper rests on the unstated premise that the test set contains no samples used for training and no augmented copies of those training samples.

Editorial extensions

If this is right

  • If the reported accuracies hold, RIS-based localization in factory settings can use this CNN as a low-cost NLOS detector to flag unreliable position estimates and switch to beam-steering mitigations.
  • The model's apparent robustness to mixed-synthetic (noisy) data suggests that facilities without extensive measurement campaigns can generate training data by augmenting a small set of real samples.
  • The fact that a custom 3-block CNN outperforms a much larger pre-trained VGG-16 on spectrogram inputs implies that domain-specific architectures can be smaller and faster, easing deployment on embedded RIS controllers.
  • The three-environment generalization supports the idea that the same network weights could be reused across rooms and factories with retraining only on a few new measurements.

Reading between the lines

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

  • We infer that the abstract's headline range (95.0–99.0%) is drawn from the meeting room and chamber results; the HF-lab measured case in Table II is 86.0%, so the '95–99' range should be read as the spread of the best environment, not the worst-case performance.
  • We infer that because the paper does not state a train/test split or cross-validation scheme, the accuracy figures are only as strong as the implicit premise that the test set is disjoint from the training samples and from augmented copies of them; a re-run with explicit disjoint splits would either confirm or weaken the claimed margins over VGG-16.
  • We infer that the mixed-synthetic results hovering around 88% in all three environments hint that the CNN's advantage comes largely from discriminating spectrogram texture rather than environment-specific artifacts; if so, the same model could be calibrated on one factory and applied to another with only modest fine-tuning.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes a customized convolutional neural network (cCNN) for classifying line-of-sight (LOS) and non-line-of-sight (NLOS) conditions in RIS-based indoor localization, using spectrogram images derived from channel impulse response (CIR) measurements. The measurement campaign covers three environments (meeting room, HF laboratory, and anechoic chamber) with two metal-plate blockage sizes, and the data are organized into four types: measured, synthetic (augmented), mixed measured, and mixed synthetic. The cCNN is compared against a pre-trained VGG-16 model, with accuracy results reported in Table II and Figure 6. The abstract and conclusion claim that the cCNN achieves 95.0%-99.0% accuracy and that VGG-16 achieves 85.5%-88.0%, while also positioning the framework as scalable and suitable for 6G-enabled smart factories.

Significance. If properly validated, the real-world RIS measurement dataset collected across three distinct environments would be a useful contribution, and the cCNN architecture is straightforward and potentially reproducible. The paper also demonstrates a systematic comparison across four data types, which is commendable. However, as written, the central performance claims are internally inconsistent with the paper's own results table, and the evaluation methodology is critically under-specified, making it impossible to assess whether the reported accuracies reflect generalization. These issues currently overshadow the potential value of the dataset and the classification framework.

major comments (3)
  1. [Abstract, Section VI, and Table II (Section V)] The abstract and conclusion state that the cCNN achieves 95.0%-99.0% accuracy and that VGG-16 achieves 85.5%-88.0%, but Table II reports cCNN accuracies as low as 86.0% (HF Lab measured data and mixed synthetic data in all environments) and VGG-16 accuracies as low as 29.0% (HF Lab measured data) and 38.0% (synthetic data). No passage in Section V explains how the abstract ranges were derived; if they are meant as ranges across scenarios and data types, they are contradicted by the table. This internal inconsistency invalidates the headline performance claim as written and must be corrected or explicitly reconciled.
  2. [Section V, Table II and Section V-B] The paper reports no train/test split, cross-validation scheme, number of independent runs, or hyperparameter values for either cCNN or VGG-16. Without this information, the accuracy figures in Table II and Figure 6 cannot be interpreted as estimates of generalization performance, and the comparison to VGG-16 is not statistically grounded. The authors must specify the data partitioning protocol and, ideally, provide cross-validated results with confidence intervals or error bars.
  3. [Section V-A.2 and V-A.4] The synthetic and mixed synthetic datasets are generated from the measured dataset using transformations such as flipping, rotation, resizing, cropping, and brightness/contrast adjustments. Because the paper never states how the augmented samples are partitioned between training and test sets, there is a risk that augmented versions of training samples appear in the test set, which would inflate the reported accuracies. The authors must describe how they ensured disjointness (for example, splitting by measurement angle or by original sample identifier) for all four data types to rule out data leakage.
minor comments (6)
  1. [Section IV, first paragraph] The text refers to "the scenario experiments in Section II" when it should refer to Section III, which contains the measurement campaign descriptions.
  2. [Figure 3 caption] The caption contains a typo: "Anaechoic" should be "Anechoic."
  3. [Figure 6] Figure 6 labels the third environment as "Big Chamber," while the text consistently uses "Anechoic Chamber"; the labels should be unified.
  4. [Section V-A.2] The phrase "using a sample of each of the LOS and NLOS depicted with a metal-plate of size 1m-by-1m and 0.75m-by-0.75m block, respectively" is grammatically unclear; please specify which plate size corresponds to which NLOS class and how the augmentation samples were selected.
  5. [Section I, final sentence] The sentence describing the SoftMax activation and the introduction of NLOS via metal plates is garbled; it should be rewritten to clearly state that the model outputs probabilities for three classes (LOS, NLOS with large plate, NLOS with small plate).
  6. [Section V-B, paragraph on VGG-16] The phrase "a state-of-the-art model that has been tested and trusted over time" is informal and does not constitute a technical justification for selecting VGG-16 as the baseline; a citation or a brief rationale based on architecture characteristics would be more appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: this is an empirical CNN benchmarking study whose accuracy claims are measured outcomes, not derivations from fitted inputs or self-cited theorems.

full rationale

The paper is an empirical classification study: the proposed cCNN is trained on measured, synthetic, mixed-measured, and mixed-synthetic spectrogram/CIR images, and compared against VGG-16, with accuracies reported in Table II. There is no analytical derivation in which an output is defined in terms of an input, no parameter fitted to one subset and then reported as a prediction on a closely related quantity, no load-bearing self-citation, and no imported uniqueness theorem. Equations (1)-(11) merely spell out a standard CNN forward pass (convolution, batch norm, ReLU, pooling, flattening, fully connected layers, softmax) and do not constrain the accuracy results. The accuracy numbers are empirical outcomes on datasets described in Section V-A. Concerns raised by the manuscript's own presentation, such as the abstract's 95.0%-99.0% and 85.5%-88.0% ranges not matching the values in Table II, and the absence of an explicit train/test split for the augmented and mixed datasets, are internal-consistency and evaluation-leakage risks rather than circularity. No equation is equal to another by construction, and no claim reduces to its own premise. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central accuracy claim rests on standard CNN training plus three unverified domain assumptions: the spectrogram image representation is sufficient, the three rooms represent factories, and the reported accuracies are from non-leaked held-out data. The only free parameters are the hand-chosen CNN hyperparameters and the unreported mixing and split choices. No new physical entities are introduced.

free parameters (2)
  • CNN hyperparameters = 32/64/128 filters, 3x3 kernels, dropout p=0.5, 224x224 input
    Chosen by hand without ablation; the reported accuracy values depend on these choices and no sensitivity analysis is provided.
  • Mixed dataset mixing ratios = unreported
    The paper says the mixed datasets contain more of one component than the other but does not give exact ratios; the accuracy results depend on these ratios.
assumptions (3)
  • domain assumption CIR-to-spectrogram PNG images retain the LOS/NLOS information needed for classification.
    The whole pipeline assumes the chosen image representation is sufficient; no baseline compares raw CIR features against spectrogram images.
  • domain assumption The three measurement environments are representative of smart factory conditions.
    The paper extrapolates to smart factories from an anechoic chamber, a meeting room, and an HF lab, but does not validate in an actual factory.
  • ad hoc to paper Reported accuracies come from held-out test data without overlap with augmented training samples.
    No train/test split or cross-validation is described; synthetic data are generated from original samples, creating leakage risk.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI-Assisted NLOS Sensing for RIS-Based Indoor Localization in Smart Factories." pith.science (2026). https://pith.science/paper/YPMDH6IS

@misc{pith2026250515989,
  author       = {Pith},
  title        = {Pith review of: AI-Assisted NLOS Sensing for RIS-Based Indoor Localization in Smart Factories},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YPMDH6IS}},
  note         = {Machine review of arXiv:2505.15989}
}
read the original abstract

In the era of Industry 4.0, precise indoor localization is vital for automation and efficiency in smart factories. Reconfigurable Intelligent Surfaces (RIS) are emerging as key enablers in 6G networks for joint sensing and communication. However, RIS faces significant challenges in Non-Line-of-Sight (NLOS) and multipath propagation, particularly in localization scenarios, where detecting NLOS conditions is crucial for ensuring not only reliable results and increased connectivity but also the safety of smart factory personnel. This study introduces an AI-assisted framework employing a Convolutional Neural Network (CNN) customized for accurate Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) classification to enhance RIS-based localization using measured, synthetic, mixed-measured, and mixed-synthetic experimental data, that is, original, augmented, slightly noisy, and highly noisy data, respectively. Validated through such data from three different environments, the proposed customized-CNN (cCNN) model achieves {95.0\%-99.0\%} accuracy, outperforming standard pre-trained models like Visual Geometry Group 16 (VGG-16) with an accuracy of {85.5\%-88.0\%}. By addressing RIS limitations in NLOS scenarios, this framework offers scalable and high-precision localization solutions for 6G-enabled smart factories.

Figures

Figures reproduced from arXiv: 2505.15989 by the authors.

Figure 1
Figure 1. Meeting Room’s LOS and NLOS experimen￾tal measurements with azimuth, RIS’s elevation, and the Rx-horn-antenna’s direction all set to zero [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. HF lab’s LOS and NLOS experimental measurements with azimuth, RIS’s elevation, and the Rx-horn-antenna’s direction all set to zero. 1) General setup: The measurement system com￾prises a Vector Network Analyzer (VNA) operating between 4.8 GHz and 5.2 GHz, a 16-by-16 RIS panel, a transmitter, and a receiver mounted on a turntable, as seen in Figures [1-4] where the VNA port-1 is connected to the transmitter in front o… view at source ↗
Figure 3
Figure 3. Anaechoic chamber’s LOS and NLOS exper￾imental measurements with azimuth, RIS’s elevation, and the Rx-horn-antenna’s direction all set to zero. In contrast, port-2 is connected to the receiver. The turntable rotates 360◦ with a 5◦ increment, allowing for spatial characterization of the environment. The RIS panel is fixed at 0◦ of azimuth and elevation for all scenarios. Channel Impulse Response (CIR) data were recor… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: CIR and spectrogram image data experimental measurement scenarios for LOS and NLOS using 5mm [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Deep learning architecture for AI-assisted LOS and NLOS sensing. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Model’s accuracy scores for different environmental scenarios and data genres. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 8 canonical work pages

  1. [1]

    N. Wu, X. Wang, Z. Fei, F. Xia, J. Huang and A. Nal- lanathan, ”RIS-Assisted Integrated Sensing and Backscatter Communications for Future IoT Networks,” in IEEE Internet of Things Magazine, vol. 7, no. 4, pp. 44-50, July 2024, doi: 10.1109/IOTM.001.2300184

    work page doi:10.1109/iotm.001.2300184 2024

  2. [2]

    Kubota, S

    A. Kubota, S. Sampei and T. Takahashi, ”A Study on Conver- sion of NLoS to LoS conditions using Sidelink in Smart Factory Environments,” 2021 IEEE VTS 17th Asia Pacific Wireless Communications Symposium (APWCS), Osaka, Japan, 2021, pp. 1-5, doi: 10.1109/APWCS50173.2021.9548715

    work page arXiv 2021

  3. [3]

    Rinch, A

    O. Rinch, A. Elzanaty and A. Alsharoa, ”Enhancing Near- Field Wireless Localization With LiDAR-Assisted RIS in Multipath Environments,” in IEEE Wireless Communications Letters, vol. 12, no. 12, pp. 2168-2172, Dec. 2023, doi: 10.1109/LWC.2023.3311730

    arXiv 2023

  4. [4]

    Rinchi, A

    O. Rinchi, A. Elzanaty and M. -S. Alouini, ”Compressive Near- Field Localization for Multipath RIS-Aided Environments,” in IEEE Communications Letters, vol. 26, no. 6, pp. 1268-1272, June 2022, doi: 10.1109/LCOMM.2022.3151036

    arXiv 2022

  5. [6]

    Rinchi, A

    O. Rinchi, A. Elzanaty and A. Alsharoa, ”Single-Snapshot Localization for Near-Field RIS Model Using Atomic Norm Minimization,” GLOBECOM 2022 - 2022 IEEE Global Com- munications Conference, Rio de Janeiro, Brazil, 2022, pp. 2432- 2437, doi: 10.1109/GLOBECOM48099.2022.10000689

    arXiv 2022

  6. [7]

    Z. Li, Z. Wan, K. Ying, Y . Mei, M. Ke and Z. Gao, ”Recon- figurable Intelligent Surface Assisted Localization Over Near- Field Beam Squint Effect,” 2022 International Symposium on Wireless Communication Systems (ISWCS), Hangzhou, China, 2022, pp. 1-6, doi: 10.1109/ISWCS56560.2022.9940428

    arXiv 2022

  7. [8]

    Gonz ´alez-Prelcic et al., ”The Integrated Sensing and Com- munication Revolution for 6G: Vision, Techniques, and Applica- tions,” in Proceedings of the IEEE, vol

    N. Gonz ´alez-Prelcic et al., ”The Integrated Sensing and Com- munication Revolution for 6G: Vision, Techniques, and Applica- tions,” in Proceedings of the IEEE, vol. 112, no. 7, pp. 676-723, July 2024, doi: 10.1109/JPROC.2024.3397609

    arXiv 2024

  8. [9]

    Z. Gong, L. Wu, Z. Zhang, J. Dang, Y . Wu and J. Wang, ”Asynchronous RIS-Assisted Localization: A Comprehensive Analysis of Fundamental Limits,” in IEEE Transactions on Wireless Communications, vol. 23, no. 9, pp. 10974-10989, Sept. 2024, doi: 10.1109/TWC.2024.3377367

    work page arXiv 2024

Show all 19 references
  1. [10]

    Z. Li, Z. Gao and T. Li, ”Sensing User’s Channel and Location With Terahertz Extra-Large Reconfigurable Intelligent Surface Under Hybrid-Field Beam Squint Effect,” in IEEE Journal of Selected Topics in Signal Processing, vol. 17, no. 4, pp. 893- 911, July 2023, doi: 10.1109/JS...

    2023

  2. [11]

    Y . Han, S. Jin, C. -K. Wen and T. Q. S. Quek, ”Localiza- tion and Channel Reconstruction for Extra Large RIS-Assisted Massive MIMO Systems,” in IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 5, pp. 1011-1025, Aug. 2022, doi: 10.1109/JSTSP.2022.3174654

    2022

  3. [12]

    M. Li, S. Zhang, Y . Ge, Z. Li, F. Gao and P. Fan, ”STAR-RIS Aided Integrated Sensing and Communication Over High Mobility Scenario,” in IEEE Transactions on Commu- nications, vol. 72, no. 8, pp. 4788-4802, Aug. 2024, doi: 10.1109/TCOMM.2024.3381725

    2024

  4. [13]

    Cheng, J

    Y . Cheng, J. Du, J. Liu, L. Jin, X. Li and D. B. da Costa, ”Nested Tensor-Based Framework for ISAC Assisted by Reconfigurable Intelligent Surface,” in IEEE Transactions on Vehicular Technology, vol. 73, no. 3, pp. 4412-4417, March 2024, doi: 10.1109/TVT.2023.3324448

    2024

  5. [14]

    Z. Wang, Z. Liu, Y . Shen, A. Conti and M. Z. Win, ”Lo- cation Awareness in Beyond 5G Networks via Reconfigurable Intelligent Surfaces,” in IEEE Journal on Selected Areas in Communications, vol. 40, no. 7, pp. 2011-2025, July 2022, doi: 10.1109/JSAC.2022.3155542

    2011

  6. [15]

    F. Wang, X. Houv, X. Wang, X. Li, L. Chen and T. Asai, ”Reconfigurable Intelligent Surface Aided Joint Communica- tion And Positioning,” 2023 IEEE 98th Vehicular Technology Conference (VTC2023-Fall), Hong Kong, Hong Kong, 2023, pp. 1-6, doi: 10.1109/VTC2023-Fall60731.2023.10333393

    2023

  7. [16]

    Hassouna, S., Jamshed, M.A., Ur-Rehman, M. et al. RIS- assisted near-field localization using practical phase shift model. Sci Rep 14, 4350 (2024). https://doi.org/10.1038/s41598-024- 54859-3

    2024 doi

  8. [17]

    Li et al., ”Design of Low-Cost Single-Layer 2-Bit Re- flective Programmable Metasurface Based on Folded Ground,” in IEEE Transactions on Microwave Theory and Tech- niques, vol

    P. Li et al., ”Design of Low-Cost Single-Layer 2-Bit Re- flective Programmable Metasurface Based on Folded Ground,” in IEEE Transactions on Microwave Theory and Tech- niques, vol. 71, no. 8, pp. 3455-3465, Aug. 2023, doi: 10.1109/TMTT.2023.3259526

    2023

  9. [18]

    E. C. Strinati et al., ”Distributed Intelligent Integrated Sensing and Communications: The 6G-DISAC Approach,” 2024 Joint European Conference on Networks and Communications and 6G Summit (EuCNC/6G Summit), Antwerp, Belgium, 2024, pp. 392-397, doi: 10.1109/EuCNC/6GSummit60053.2...

    2024

  10. [19]

    Albanese, P

    A. Albanese, P. Mursia, V . Sciancalepore and X. Costa- P´erez, ”PAPIR: Practical RIS-aided Localization via Statisti- cal User Information,” 2021 IEEE 22nd International Work- shop on Signal Processing Advances in Wireless Commu- nications (SPAWC), Lucca, Italy, 2021, pp. 531...

    2021

  11. [20]

    Encinas-Lago, F

    G. Encinas-Lago, F. Devoti, M. Rossanese, V . Sciancalepore, M. D. Renzo, X. Costa-P ´erez, ”COLoRIS: Localization- agnostic Smart Surfaces Enabling Opportunistic ISAC in 6G Networks”, Networking and Internet Architecture, doi:https://doi.org/10.48550/arXiv.2406.07377

    doi arXiv

Pith tools

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