From DeepSense to Open RAN: AI/ML Advancements in Dynamic Spectrum Sensing and Their Applications

Ryan Barker

REVIEW 5 major objections 3 minor 11 references

From DeepSense to Open RAN: AI/ML Advancements in Dynamic Spectrum Sensing and Their Applications

T0 review · 5 major / 3 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read AI spectrum sensing plus Open RAN could make 5G/6G self-optimizing

desk verdict A readable but citation-sloppy reading report; useful as a first orientation, not as a reliable map to the literature. read the letter →

arxiv 2502.02889 v2 pith:OUBQP5MZ submitted 2025-02-05 cs.NI eess.SP

Ryan Barker This is my paper · ORCID

classification cs.NIeess.SP

keywords dynamicspectrumsensingDeepSenseSweepwidebandsignalstitchingOpenRANxAppsdigitaltwins5G/6Gnetworks

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 reading report argues that two research lines—AI/ML dynamic spectrum sensing and Open RAN—are converging into a single software-defined infrastructure for 5G/6G. It presents DeepSense as the foundational step of feeding raw I/Q radio samples into convolutional neural networks, and reads DeepSweep and Wideband Signal Stitching as proof that the approach can scale to real-time, crowded spectrum environments. It then claims ORAN's digital twins and xApps give operators a safe way to simulate, deploy, and self-heal AI-driven network decisions. A sympathetic reader would care because these technologies target exactly the latency, scalability, and resilience constraints that autonomous vehicles, UAVs, and industrial automation will impose on next-generation wireless.

What carries the argument

The carrying mechanism is the radio-to-decision loop that starts with raw in-phase/quadrature (I/Q) samples and turns them into spectrogram-like images that CNNs classify, instead of passing signals through protocol-specific feature extractors. DeepSense establishes this loop; DeepSweep parallelizes it by chopping the spectrum into chunks processed by lightweight CNNs; Wideband Signal Stitching adds semantic segmentation at the I/Q-sample level, with non-local blocks and a synthetic-plus-over-the-air data augmentation pipeline that helps the network separate overlapping signals. On the ORAN side, the analogous mechanism is the closed control loop formed by digital twins, which simulate network configurations before deployment, and xApps, which run on the RAN Intelligent Controller to make real-time resource, scheduling, and healing decisions. The report's argument is that these two loops can be coupled into one intelligent infrastructure.

What would settle it

Reproduce DeepSweep on an independent wideband over-the-air dataset and measure latency and accuracy; if narrowband interference detection falls far below 98% or inference exceeds one millisecond at the claimed complexity, the report's central case for AI-native real-time spectrum sensing is weakened. On the ORAN side, a decisive negative result would be digital-twin predictions of VNF splitting or misconfiguration effects diverging from physical testbed measurements under the same traffic load.

Watch

Extended reading notes

Core claim

The report's central claim is that the combination of AI/ML dynamic spectrum sensing and ORAN's open, vendor-neutral architecture is the enabling condition for efficient, resilient, self-optimizing 5G/6G networks. It treats DeepSense as the foundational demonstration that CNNs can classify wideband spectrum directly from raw I/Q samples, then reads DeepSweep and Wideband Signal Stitching as evidence that the approach scales: DeepSweep's parallel 'chunk and process' CNN design reaches sub-millisecond inference with 10x lower complexity and 98% narrowband interference accuracy, while Stitching uses semantic segmentation with non-local blocks to handle fragmented and overlapping signals. On the network side, it argues that ORAN digital twins allow operators to simulate VNF splitting and misconfigurations before deployment, and that AI/ML xApps developed under frameworks like xDevSM and tested on AERPAW bring real-time resource scheduling and self-healing to live RANs. The synthesis claim is that these two lines are converging into a single software-defined intelligent infrastructure.

Load-bearing premise

Everything the report concludes depends on the assumption that the performance numbers it quotes from the underlying studies—sub-millisecond latency, 10x complexity reduction, 98% narrowband interference accuracy, and dependable digital-twin predictions—hold outside the specific experimental setups in which they were measured.

Editorial extensions

If this is right

If DeepSense-style sensing reaches sub-millisecond latency in deployed systems, autonomous vehicle and industrial automation links can act on spectrum changes fast enough to prevent missed collision alerts or coordination failures.
If semantic segmentation separates overlapping wideband signals as claimed, operators can exploit spectral holes more aggressively and detect narrowband interference that bounding-box methods miss.
If ORAN digital twins predict the effects of VNF splitting and AI/ML misconfigurations before deployment, operators can test self-optimizing changes without risking live network stability.
If xApps developed under frameworks like xDevSM and validated on AERPAW generalize beyond testbeds, UAV and emergency-response networks gain real-time resource scheduling and self-healing on vendor-neutral hardware.
Taken together, the convergence of DSS and ORAN implies that network intelligence can be updated in software rather than replaced in hardware, matching the service-based design goals of 5G/6G.

Reading between the lines

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

An implication the report leaves implicit is that DeepSweep-level sensing could turn spectrum awareness into a shared network service: a spectrum-sensing xApp that multiple tenants query, much like compute or storage today.
A testable extension is to benchmark the same semantic segmentation model on an independent public corpus of overlapping OFDM and narrowband emitters, checking whether the reported complexity and accuracy hold outside the original over-the-air setup.
The synthesis also suggests a stronger claim than the report states explicitly: the RAN Intelligent Controller could close the loop between sensing and action, using DeepSense-style predictions as direct inputs to xApp resource scheduling rather than as monitoring aids.
A concrete next experiment would compare an RL-trained xApp against the static CNN baseline on the AERPAW testbed, injecting misconfigurations to see whether autonomous optimization actually reduces downtime relative to the self-healing claims.

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

Colleague,

Quick take: this is a reading report, not a research preprint. It re-describes the DeepSense → DeepSweep → Wideband Signal Stitching line and then surveys ORAN digital twins, xApps, and AERPAW. There are no new equations, measurements, or frameworks. If you need a one-page orientation to that slice of the spectrum-sensing/ORAN literature, the prose is clear and the chosen papers are largely the right ones. The DeepSense-to-stitching arc is the actual trajectory in the literature, and the ORAN half hits the main themes: digital twins, xDevSM, AERPAW, and misconfiguration analysis.

The soft spots are real, and they matter for a report whose whole purpose is fidelity to sources. The reference mapping is internally inconsistent. Section II labels Stitching as [2] and DeepSweep as [3], but the reference list assigns [2] to DeepSweep and [3] to Stitching; Section III.B uses the opposite, correct mapping. Section II calls "Big Data Goes Small" [4] and "DeepLab" [5], but the references assign [4] to DeepLab and [5] to the VNF-splitting paper. Section III.B credits Wideband Signal Stitching's dataset pipeline to [4] (DeepLab), which is not the stitching paper. The ORAN attributions drift the same way: xDevSM is credited to "Melodia et al." [8] when the first author is Feraudo; AERPAW is credited to "Mandal et al." [9] when it is Moore et al. The headline performance numbers — sub-1 ms latency, 10× complexity reduction, 98% narrowband accuracy — are asserted without page or table pointers, so a reader cannot tell which experiment produced them or whether they generalize.

None of this makes the central narrative wrong: DeepSense-style CNN sensing is a real direction, and ORAN xApps are a real application. But the report's value is as a map, and a map with mislabeled streets is not safe. I would treat it as a student's reading report: useful for a first pass, not citable for specifics.

My recommendation: desk reject. There is no research claim to referee, and the citation errors need a full rewrite before it could serve as a dependable survey. If the authors fix the attributions and add exact table/figure references for the metrics, it could become a decent arXiv note or workshop paper.

Referee Report

5 major / 3 minor

Summary. The manuscript is a survey/reading report that synthesizes recent work in AI/ML-driven dynamic spectrum sensing (DeepSense, DeepSweep, Wideband Signal Stitching) and Open RAN (digital twins, xApps, AERPAW). It concludes that bridging these two areas can enable efficient, resilient, and self-optimizing 5G/6G networks. The paper contains no original measurements or derivations; its contribution is the synthesis and the reported performance figures drawn from the cited literature.

Significance. If the citations are reliable, the survey would be a useful compact entry point connecting the DeepSense line of work with recent ORAN xApp and digital-twin research. Strengths include its breadth of topics, the reproduced figures from the original papers, and the explicit focus on recent arXiv/conference literature. However, because the contribution is exclusively a synthesis, its value depends entirely on accurate representation of the cited sources; the attribution errors identified below undermine this value and make the paper, in its current form, an unreliable map to the literature.

major comments (5)

[Section II] The citation mapping in this section is reversed with respect to the reference list. The text attributes 'Stitching the Spectrum' to [2] and 'DeepSweep' to [3], but the reference list assigns [2] to DeepSweep and [3] to Stitching. Section III.B uses the opposite, correct mapping. This inconsistency prevents a reader from tracing the descriptions of semantic segmentation and parallelized spectrum sensing to the intended papers, and it must be corrected throughout.
[Section III.B] The dataset generation pipeline of Wideband Signal Stitching is attributed to reference [4], which is the DeepLab paper on semantic image segmentation. The described combination of over-the-air (OTA) signals with synthetic interference and noise is presented in the Stitching paper (reference [3]), not in DeepLab. This is a load-bearing misattribution because it is the only methodological detail given for the stitching framework.
[Section IV] The ORAN section contains multiple author-citation mismatches: 'Wang et al. [5]' (the first author of [5] is Amiri, not Wang), 'Allen et al. [6]' (the first author of [6] is Yungaicela-Naula), 'Hyodis et al. [7]' (the first author of [7] is Hoydis), 'Melodia et al. [8]' (the first author of [8] is Feraudo), and 'Mandal et al. [9]' (the first author of [9] is Moore). In addition, Section II cites AERPAW as [6] and xDevSM as [7], whereas [6] is the misconfiguration paper and [7] is Sionna. Since the persuasive force of this survey lies entirely in faithful reporting of the cited works, these errors are not cosmetic and must be fixed.
[Section III.B] The headline quantitative metrics for DeepSweep—sub-millisecond latency, 10x complexity reduction, and 98% narrowband interference accuracy—are asserted without any specific table, page, or figure reference to the source paper. Because the manuscript has no original experiments, a reader cannot determine which experimental setup produced these numbers, under what assumptions, or whether they generalize beyond that setup. The authors should either provide precise pointers into the DeepSweep paper or explicitly qualify the numbers as reported by that paper for a specific configuration.
[Section V] The abstract and conclusion claim that the report 'bridges' AI-based DSS methodologies with ORAN's open architecture, but the body does not support this bridge. Section III discusses only spectrum sensing and Section IV discusses only ORAN; there is no substantive discussion of how xApps, digital twins, or ORAN control loops consume or act on the spectrum-sensing outputs from DeepSense-family systems. If the synthesis is the central claim, a dedicated section or at least a bridging discussion is needed to make it credible.

minor comments (3)

[Abstract and passim] The abbreviation 'UA V' appears with a stray space in several places (e.g., the abstract and Section IV.B); it should be 'UAV'.
[Reference list] Reference [4] is listed as 'Deeplab' but the canonical title is 'DeepLab'; please correct the capitalization in the title.
[Figures 6 and 7] The captions for the segmentation model and pipeline figures are vague ('Semantic Spectrum Segmentation Model' and 'Scalable and Portable Pipeline') and would benefit from a one-sentence description of what is shown, especially since the figures appear to be adapted from the cited papers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the report is a reading survey whose claims summarize external cited works, and it performs no derivation or prediction that reduces to its own inputs.

full rationale

This manuscript is a reading report that synthesizes previously published results in dynamic spectrum sensing and Open RAN. It does not fit parameters, define quantities in terms of one another, or derive new predictions. The central claim—that AI/ML-driven DSS methods combined with ORAN architectures enable efficient, resilient, self-optimizing networks—is a qualitative synthesis of findings attributed to external papers, not a derivation whose conclusion is equivalent to its premise by construction. There are no equations, no fitted inputs renamed as predictions, and no invocation of an author-specific uniqueness theorem. Although the report contains citation-to-claim mismatches (for example, Section II labels 'Stitching the Spectrum' as [2] while the reference list assigns [2] to DeepSweep, and Section III.B attributes DeepLab [4] to Wideband Signal Stitching's pipeline), these are accuracy and traceability issues, not circularity: they affect whether the report faithfully represents its sources, but they do not mean the report's conclusions are forced by self-definition or self-citation. The authors of the report are not the authors of the cited DeepSense, DeepSweep, Stitching, or ORAN papers, so the self-citation patterns that would indicate imported uniqueness or ansatz-smuggling are absent. Per the review rules, the absence of internal derivation and reliance on independent external sources yields a circularity score of 0.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The paper is a review and introduces no fitted parameters, axioms, or new entities. Its only implicit assumption is that the summarized literature is reliable.

assumptions (1)

domain assumption The cited papers' reported performance metrics (latencies, accuracy, complexity reductions) are accurate and representative.
The review's favorable conclusions about DeepSense, DeepSweep, and ORAN rely on numbers taken from the cited works without independent verification (Section III.B).

how reviews work

0 comments

Cite this review

Pith. "Pith review of From DeepSense to Open RAN: AI/ML Advancements in Dynamic Spectrum Sensing and Their Applications." pith.science (2026). https://pith.science/paper/OUBQP5MZ

@misc{pith2026250202889,
  author       = {Pith},
  title        = {Pith review of: From DeepSense to Open RAN: AI/ML Advancements in Dynamic Spectrum Sensing and Their Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OUBQP5MZ}},
  note         = {Machine review of arXiv:2502.02889}
}

read the original abstract

The integration of Artificial Intelligence (AI) and Machine Learning (ML) in next-generation wireless communication systems has become a cornerstone for advancing intelligent, adaptive, and scalable networks. This reading report examines key innovations in dynamic spectrum sensing (DSS), beginning with the foundational DeepSense framework, which uses convolutional neural networks (CNNs) and spectrogram-based analysis for real-time wideband spectrum monitoring. Building on this groundwork, it highlights advancements such as DeepSweep and Wideband Signal Stitching, which address the challenges of scalability, latency, and dataset diversity through parallel processing, semantic segmentation, and robust data augmentation strategies. The report then explores Open Radio Access Networks (ORAN), focusing on AI/ML-driven enhancements for UAV experimentation, digital twin-based optimization, network slicing, and self-healing xApp development. By bridging AI-based DSS methodologies with ORAN's open, vendor-neutral architecture, these studies underscore the potential of software-defined, intelligent infrastructures in enabling efficient, resilient, and self-optimizing networks for 5G/6G ecosystems. Through this synthesis, the report highlights AI's transformative role in shaping the future of wireless communication and autonomous systems.

Figures

Figures reproduced from arXiv: 2502.02889 by the authors.

**Figure 2.** Spectrograms: A visual representation of signal frequency over time, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 5.** DeepSweep Spectrum Sensing CNN: The architecture utilizes [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗

Figures from the paper (4 more)

**Figure 6.** Figure 6: Semantic Spectrum Segmentation Model: Wideband Signal Stitching [PITH_FULL_IMAGE:figures/full_fig_p003_6.png]

**Figure 7.** Figure 7: Scalable and Portable Pipeline: Wideband Signal Stitching’s ar [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]

**Figure 8.** Figure 8: Challenges in Labeling Spectrum: Demonstrating the difficulty of [PITH_FULL_IMAGE:figures/full_fig_p004_8.png]

**Figure 9.** Figure 9: ORAN Architecture Overview: Integrating the 5G Core with the OSC [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 9 canonical work pages

[2]

C. P. Robinson, D. Uvaydov, S. D'Oro, and T. Melodia, ``Deepsweep: Parallel and scalable spectrum sensing via convolutional neural networks,'' 2024. [Online]. Available: https://arxiv.org/abs/2401.04805

work page Pith review arXiv 2024
[3]

Stitching the Spectrum: Semantic Spectrum Segmentation with Wideband Signal Stitching

D. Uvaydov, M. Zhang, C. P. Robinson, S. D'Oro, T. Melodia, and F. Restuccia, ``Stitching the spectrum: Semantic spectrum segmentation with wideband signal stitching,'' 2024. [Online]. Available: https://arxiv.org/abs/2402.03465

work page Pith review arXiv 2024
[4]

L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, ``Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,'' 2017. [Online]. Available: https://arxiv.org/abs/1606.00915

arXiv 2017
[5]

Amiri, N

E. Amiri, N. Wang, M. Shojafar, and R. Tafazolli, `` Optimizing Virtual Network Function Splitting in Open-RAN Environments ,'' in 2022 IEEE 47th Conference on Local Computer Networks (LCN) , Edmonton, AB, Canada, 2022, pp. 422--429

work page 2022
[8]

Feraudo, S

A. Feraudo, S. Maxenti, A. Lacava, P. Bellavista, M. Polese, and T. Melodia, ``xdevsm: Streamlining xapp development with a flexible framework for o-ran e2 service models,'' in Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, ser. ACM MobiCom ’24. 1em plus 0.5em minus 0.4em ACM, Dec. 2024, p. 1954–1961. [Online]....

work page arXiv 2024
[9]

Prototyping O-RAN Enabled UAV Experimentation for the AERPAW Testbed

J. Moore, A. S. Abdalla, C. Ueltschey, and V. Marojevic, ``Prototyping o-ran enabled uav experimentation for the aerpaw testbed,'' 2024. [Online]. Available: https://arxiv.org/abs/2411.04027

work page Pith review arXiv 2024
[6]

N. M. Yungaicela-Naula, V. Sharma, and S. Scott-Hayward, ``Misconfiguration in o-ran: Analysis of the impact of ai/ml,'' Computer Networks, vol. 247, p. 110455, Jun. 2024. [Online]. Available: http://dx.doi.org/10.1016/j.comnet.2024.110455

work page arXiv 2024
[7]

Hoydis, S

J. Hoydis, S. Cammerer, F. A. Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, ``Sionna: An open-source library for next-generation physical layer research,'' 2023. [Online]. Available: https://arxiv.org/abs/2203.11854

arXiv 2023

Show all 11 references

[1]

Uvaydov, S

D. Uvaydov, S. D’Oro, F. Restuccia, and T. Melodia, `` DeepSense: Fast Wideband Spectrum Sensing Through Real-Time In-the-Loop Deep Learning ,'' in IEEE INFOCOM 2021 - IEEE Conference on Computer Communications , Vancouver, BC, Canada, 2021, pp. 1--10

2021
[10]

Restuccia and T

F. Restuccia and T. Melodia, ``Big data goes small: Real-time spectrum-driven embedded wireless networking through deep learning in the rf loop,'' 2019. [Online]. Available: https://arxiv.org/abs/1903.05460

2019 arXiv
[11]

adobe:ns:meta/

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

1999

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [2]

C. P. Robinson, D. Uvaydov, S. D'Oro, and T. Melodia, ``Deepsweep: Parallel and scalable spectrum sensing via convolutional neural networks,'' 2024. [Online]. Available: https://arxiv.org/abs/2401.04805

work page Pith review arXiv 2024

[2] [3]

Stitching the Spectrum: Semantic Spectrum Segmentation with Wideband Signal Stitching

D. Uvaydov, M. Zhang, C. P. Robinson, S. D'Oro, T. Melodia, and F. Restuccia, ``Stitching the spectrum: Semantic spectrum segmentation with wideband signal stitching,'' 2024. [Online]. Available: https://arxiv.org/abs/2402.03465

work page Pith review arXiv 2024

[3] [4]

L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, ``Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,'' 2017. [Online]. Available: https://arxiv.org/abs/1606.00915

arXiv 2017

[4] [5]

Amiri, N

E. Amiri, N. Wang, M. Shojafar, and R. Tafazolli, `` Optimizing Virtual Network Function Splitting in Open-RAN Environments ,'' in 2022 IEEE 47th Conference on Local Computer Networks (LCN) , Edmonton, AB, Canada, 2022, pp. 422--429

work page 2022

[5] [8]

Feraudo, S

A. Feraudo, S. Maxenti, A. Lacava, P. Bellavista, M. Polese, and T. Melodia, ``xdevsm: Streamlining xapp development with a flexible framework for o-ran e2 service models,'' in Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, ser. ACM MobiCom ’24. 1em plus 0.5em minus 0.4em ACM, Dec. 2024, p. 1954–1961. [Online]....

work page arXiv 2024

[6] [9]

Prototyping O-RAN Enabled UAV Experimentation for the AERPAW Testbed

J. Moore, A. S. Abdalla, C. Ueltschey, and V. Marojevic, ``Prototyping o-ran enabled uav experimentation for the aerpaw testbed,'' 2024. [Online]. Available: https://arxiv.org/abs/2411.04027

work page Pith review arXiv 2024

[7] [6]

N. M. Yungaicela-Naula, V. Sharma, and S. Scott-Hayward, ``Misconfiguration in o-ran: Analysis of the impact of ai/ml,'' Computer Networks, vol. 247, p. 110455, Jun. 2024. [Online]. Available: http://dx.doi.org/10.1016/j.comnet.2024.110455

work page arXiv 2024

[8] [7]

Hoydis, S

J. Hoydis, S. Cammerer, F. A. Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, ``Sionna: An open-source library for next-generation physical layer research,'' 2023. [Online]. Available: https://arxiv.org/abs/2203.11854

arXiv 2023