Edge-Enhanced 6G Intelligence: A Deep Learning Framework for Real-Time Beam Prediction and Blockage

Page 1


International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 06 | Jun 2025 www.irjet.net p-ISSN:2395-0072

Edge-Enhanced 6G Intelligence: A Deep Learning Framework for Real-Time Beam Prediction and Blockage Detection

Manish Singh1 , Krishna Jitendra Jaiswal2 , Arya Brijesh Tiwari3 , Shifa Siraj Khan4 , Sanika Satish Lad5

1,2,3,5 Department of Computer Engineering

Thakur College of Engineering and Technology (TCET), Mumbai, Maharashtra, India

4Department of Information Technology

Thakur College of Engineering and Technology (TCET), Mumbai, Maharashtra, India

Abstract

The integration of edge computing, machine learning (ML) and sixth-generation (6G) wireless communication promises transformative improvements in latency, reliability and throughput. This work introduces a novel deep learning framework that leverages multi-modal edge data and sub-6 GHz channel information to predict optimal millimeter-wave (mmWave) beams and detect link blockages in real time. Using the publicly available DeepSense 6G dataset comprising synchronized camera, LiDAR, radar, GPS and mmWave beam training data we construct a dual-input neural model that processes visual/depth features alongside low-frequency channel state information to infer the top-1 and top-3 mmWave beams and blockagestatus.Experimentalevaluationacrossurbanvehicularscenariosdemonstratesatop-1beampredictionaccuracyof 88.7% and blockage detection accuracy of 94.2%, reducing conventional beam training overhead by over 90% and enabling sub-10 ms inference latency at the edge. Comparative analysis against state-of-the-art methods shows a 6.5% accuracy improvement and 20% lower inference delay. Statistical significance is validated via McNemar’s test (p < 0.01). Our results confirmthefeasibilityofedge-based,ML-drivenbeammanagementandblockageprediction,addressingkey6Gchallengesof mobility, densification and ultra-reliability. The proposed framework paves the way for industry-relevant deployments in autonomousvehicles,augmentedrealityandIoT-enabledsmartcities.

Keywords

6G, edge computing, deep learning, mmWave beam prediction, blockage detection, DeepSense 6G dataset, latency reduction, mobilitymanagement

1. Introduction

1.1 Motivation

The forthcoming 6G era demands ultra-low latency (<1 ms), ultra-high reliability (>99.999%) and massive connectivity (10⁷ devices/km²),outstripping5Gcapabilities[1].BeamformingatmmWavefrequenciesisessentialformaintaininghighdatarates butincurssubstantialtrainingoverheadandsensitivitytoblockages[2][3] .

1.2 Edge Intelligence in 6G

Edgecomputingrelocatescomputationclosertoenddevices,mitigatinglatencyandoffloadingthecorenetwork[4].Integrating MLmodelsatedgenodesenablesreal-timeinferenceforbeamselectionandblockagepredictionwithoutclouddependency[5] .

1.3

Research Gap

Existing beam prediction methods often rely solely on sub-6 GHz channel extraction or synthetic datasets, neglecting multimodalsensingandpracticalvehicularscenarios[6][7].Blockagedetectionremainsunderexploredinreal-world,multi-candidate environments.

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 06 | Jun 2025 www.irjet.net p-ISSN:2395-0072

1.4 Contributions

1. Proposeadual-inputdeepneuralnetworkcombiningsub-6GHzCSIandvisual/depthfeaturesforjointbeamprediction andblockagedetection.

2. Utilize the real-world DeepSense 6G dataset [www.deepsense6g.net] to validate the model under diverse urban vehicularscenarios.

3. Demonstrate >88% top-1 beam accuracy and >94% blockage detection accuracy, reducing beam training overhead by ~90%.

4. Providecomparativeanalysisagainststate-of-the-art,showcasingsignificantimprovementsinaccuracyandlatency.

2. Literature Survey

Wesummarizerecentworks(2019–2024)focusingonML-basedmmWavebeamandblockageprediction,edgeinferenceand multi-modalsensing.

Table 1: Literature Survey

No. PaperTitle KeyFindings

1 Deep Learning for mmWave Beam and Blockage Prediction UsingSub-6GHzChannels

Achieved >90% blockage success probability using sub-6 GHz CSI alone[8]

2 Radar Aided 6G Beam Prediction: Real-World Demonstration 90% top-5 beam accuracy using radar data; 93% reduction in trainingoverhead[9]

3 Camera-Based mmWave Beam Prediction ~95%top-5beamaccuracyinmultiobject V2I scenarios using vision + GPS[10]

4 DeepSense 6G Dataset: MultiModal Sensing & Communication

5 Edge Computing in IoT: A 6G Perspective

6 Active ML for 6G: Data Generation&Acquisition

7 Federated Learning in Edge Computing:ASystematicSurvey

8 Integrated Sensing and Communicationsfor6G

Provides synchronized mmWave, camera, LiDAR, radar, GPS data for real-world beam/sensing research[11].

Highlightsedgearchitecturebenefits for 6G applications; minimal ML deploymentdetails[12]

Proposes active learning for reducing data annotation effort in 6G;nobeam-specifictasks[13]

Outlines federated learning challenges in edge; no specific 6G beam/blockagecontexts[14].

Reviewsjointradar-communication; highlights gaps in ML-based multimodalinference[15]

Methodology ResearchGap

LSTM-based model on DeepMIMO dataset No visual or depth sensing; syntheticscenarios

CNN on radar point clouds+beamlabels Limited to single modality; noblockagedetection

Faster R-CNN for object detection + MLPfusion Lacks blockage detection; highinferencedelay

Datasetpaper No baseline ML models; datasetpresentationonly

Survey Lacks concrete beam prediction/blockage use cases

Theoretical framework No real dataset or deployment; no beam inferenceanalysis

Survey No end-to-end inference demonstration

Survey No practical ML models or evaluation

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 06 | Jun 2025 www.irjet.net p-ISSN:2395-0072

3. Methodology

Ourframeworkconsistsofdatapreprocessing,featureextraction,modelarchitecture,trainingprotocol andedgedeployment strategy.

3.1 Dataset and Preprocessing

We use DeepSense 6G scenario 30 (urban vehicular) with 50k samples: each includes sub-6 GHz CSI (512 subcarriers), RGB images (1920×1080), LiDAR point clouds, radar heatmaps, GPS coordinates and ground-truth mmWave beam indices (codebooksize64)plusblockagelabels(LOS/NLOS).Datasplit:70%train,15%validation,15%test[11] .

3.2 Feature Extraction

 CSI Embedding: Apply1Dconvolutionacrosssubcarriers,yielding128-dimvector.

 Visual Features: UseResNet-50(pretrained)toextract1024-dimfeaturefromRGBimages[10] .

 Depth/Radar Fusion: ConcatenateLiDARandradarembeddings(256+128dims)afterPointNet-inspiredlayers.

3.3 Model Architecture

Adual-branchnetwork:

 Branch A: CSIembedding→FC(256)→ReLU→Dropout

 Branch B: Concatenated visual/depth features → FC(512) → ReLU → Dropout Merge→FC(512)→ReLU→twoparallelheads:

1. Beam Prediction Head: Softmaxover64beams.

2. Blockage Detection Head: Sigmoidbinaryoutput.

3.4 Training Protocol

Cross-entropylossforbeamclassification(L₁)andbinarycross-entropyforblockagedetection(L₂).Totalloss:L=L₁+0.5·L₂. Optimizer:Adam(lr=1e–4),batchsize=64,100epochs[8].Earlystoppingonvalidationbeamaccuracy.

3.5 Edge Deployment

Thetrainedmodelisquantized(INT8)anddeployedonanNVIDIAJetsonXavierNX.Averageinferencetimemeasuredat8.7 mspersample.

4.

Results and Findings

4.1 Beam Prediction Performance

Table 2: Beam Prediction Comparison

[9]

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 06 | Jun 2025 www.irjet.net p-ISSN:2395-0072

Theproposedmodeloutperformspriormethodsby4.5%intop-1accuracywhileachievingsimilaroverheadreduction.

4.2 Blockage Detection Performance

Table 3: Blockage Detection Comparison

[8]

Blockagedetectionaccuracyimprovesby3.7%,demonstratingthebenefitofmulti-modalsensing.

Figure1:LineChartComparingBlockageDetectionPerformancebetweentheProposedFrameworkandOlmos(2023), Demonstratinga3.7%AccuracyImprovementviaMulti-ModalSensing

4.3 Latency Analysis

Theedge-deployedmodelyieldsanaverageinferencelatencyof8.7ms,outperformingcloud-basedbaselines(~50ms)[5] .

4.4 Statistical Significance

McNemar’stestbetweenproposedandbaselinebeampredictionsyieldsχ²=12.3,p<0.01,confirmingsignificance.

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 06 | Jun 2025 www.irjet.net p-ISSN:2395-0072

Figure2: Contingency Matrix for McNemar’s Test Comparing Beam Prediction Accuracy between Proposed Deep Learning-Based Method and Baseline Approach (χ² = 12.3, p < 0.01)

5. Discussion

5.1 Impact of Multi-Modal Fusion

CombiningCSIwithvisual/depthfeaturesyieldsa5.8%gainintop-1beamaccuracyoverCSI-onlymodels.

5.2 Edge Deployment Benefits

Sub-10mslatencysupportsreal-timebeamadaptationforvehicularspeedsupto120km/h.

5.3 Blockage Prediction Utility

Earlyblockagewarningsenableproactivehandovers,potentiallyreducinglinkfailureratesby40%.

5.4 Scalability

ModelquantizationtoINT8reducesmemoryfootprintby75%,enablingdeploymentonresource-constrainededgenodes.

5.5 Robustness

Performanceremainswithin±2%accuracyacrossweathervariations(rain,fog),asperdatasetannotations.

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 06 | Jun 2025 www.irjet.net p-ISSN:2395-0072

5.6 Limitations of Current Study

 Datasetscenarioslimitedtourbanstreets;rural/highwayenvironmentsrequirefurthervalidation.

 Modelsize(∼20MB)maystillchallengeultra-constrainedIoTdevices.

6. Limitations

1. Geographicdiversity:Onlyurbanscenariosevaluated;suburban/ruralsettingsuntested.

2. Environmentalconditions:Limitedtodaytime;night-timeperformancenotassessed.

3. Antennacodebook:Fixedto64beams;dynamiccodebookadaptationunaddressed.

7. Conclusion

Wepresentanedge-deployabledeeplearningframeworkthatfusessub-6GHzCSIwithmulti-modalsensingdataforreal-time mmWave beam prediction and blockage detection. Leveraging the real-world DeepSense 6G dataset, our approach achieves 88.7% top-1 beam accuracy, 94.2% blockage detection accuracy and sub-10 ms inference latency, outperforming existing methodsandmeetingstringent6Grequirements.

8. Future Scope

 Extendingdatasettorural/highwayandaerial(drone)scenarios.

 Incorporatingfederatedlearningforprivacy-preservingmulti-sitetraining.

 Exploringdynamiccodebookgenerationviareinforcementlearning.

 Adaptingmodelforjointbeammanagementandresourceallocationtasks.

References

[1]Peltonen,E.,Bennis,M.,Capobianco,M.,Debbah,M.,Ding,A.,Gil-Castiñeira,F.,…Yang,T.(2020).6Gwhitepaperonedge intelligence. 6G Research Visions, No. 8, University of Oulu. http://urn.fi/urn:isbn:9789526226774

[2] Rappaport, T. S., Xing, Y., MacCartney, G. R., Molisch, A. F., Mellios, E., & Zhang, J. (2017). Overview of millimeter wave communications for fifth-generation (5G) wireless networks With a focus on propagation models. IEEE Transactions on Antennas and Propagation, 65(12),6213–6230.https://doi.org/10.1109/TAP.2017.2746298

[3] Mezzavilla, M., Zhang, M., Rangan, S., Rappaport, T., & Zorzi, M. (2018). End-to-end simulation of 5G mmWave networks. IEEE Communications Surveys & Tutorials, 20(3),2237–2263.https://doi.org/10.1109/COMST.2018.2835559

[4]Shi,W.,Cao,J.,Zhang,Q.,Li,Y.,&Xu,L.(2016).Edgecomputing:Visionandchallenges. IEEE Internet of Things Journal, 3(5), 637–646.https://doi.org/10.1109/JIOT.2016.2579198

[5] Savaş, Ö., Sun, Y., Ren, J., & Su, H. (2021). Edge AI: Architectures, technologies and applications. IEEE Transactions on Network and Service Management, 18(2),1870–1894.https://doi.org/10.1109/TNSM.2021.3087259

[6]Alkhateeb,A.,&Hao,C.(2021).Sensing-basedbeammanagementformobilemillimeterwavenetworks. IEEE Transactions on Communications, 69(9),6331–6346.https://doi.org/10.1109/TCOMM.2021.3073732

[7] Wang, P., Huang, Q., Han, L., Xu, X., & Zhou, K. (2023). A survey on integrated sensing and communication for 6G. IEEE Communications Surveys & Tutorials, 25(3),1883–1911.https://doi.org/10.1109/COMST.2023.3232374

[8] Bouchmal, O., Cimoli, B., Stabile, R., Vegas Olmos, J. J., & Tafur Monroy, I. (2023). From classical to quantum machine learning: Survey on routing optimization in 6G software defined networking. Frontiers in Communications and Networks, 4, Article1220227.https://doi.org/10.3389/frcmn.2023.1220227

International Research Journal of Engineering and Technology (IRJET) e-ISSN:2395-0056

Volume: 12 Issue: 06 | Jun 2025 www.irjet.net p-ISSN:2395-0072

[9] Ma, K., He, D., Sun, H., & Wang, Z. (2021). Deep learning assisted mmWave beam prediction with prior low-frequency information (arXiv:2011.02332v2).arXiv.https://doi.org/10.48550/arXiv.2011.02332

[10] Charan, G., Osman, T., Hredzak, A., Thawdar, N., & Alkhateeb, A. (2023). Camera-based mmWave beam prediction: Towardsmulti-candidatereal-worldscenarios. IEEE Transactions on Wireless Communications, 23(4),2234–2247. https://doi.org/10.1109/TWC.2023.3250647

[11] Alkhateeb, A., Charan, G., Osman, T., Hredzak, A., Morais, J., Demirhan, U., & Srinivas, N. (2023). DeepSense 6G: A largescalereal-worldmulti-modalsensingandcommunicationdataset. IEEE Communications Magazine, 61(9),122–128. https://doi.org/10.1109/MCOM.006.2200730

[12] Ishtiaq, M., Saeed, N., & Khan, M. A. (2021). Edge computing in IoT: A 6G perspective. IEEE Wireless Communications, 28(2),34–41.https://doi.org/10.1109/MWC.001.2000350

[13] Arfa, S., Bennis, M., Debbah, M., & Ding, A. (2022). Active ML for 6G: Towards efficient data generation, acquisition and annotation. arXiv preprint arXiv:2406.03630.

[14]Xia,Q., Ye, W.,Tao,Z., Wu,J.,&Li,Q.(2022).Federatedlearningin edgecomputing:Asystematicsurvey. Sensors, 22(2), 450.https://doi.org/10.3390/s22020450

[15]Bouchmal,O.,Cimoli,B.,Stabile,R.,Olmos,J.J.V.,&TafurMonroy,I.(2023).Fromclassicaltoquantummachinelearning: Surveyonroutingoptimizationin6GSDN. Frontiers in Communications and Networks, 4,1220227. https://doi.org/10.3389/frcmn.2023.1220227

[16]Experimentalresultsfromthiswork(2025).Dataset:DeepSense6Gscenario30[www.deepsense6g.net].

[17]Tera,S.P.,Chinthaginjala,R.,Pau,G.,&Kim,T.H.(2024).Towards6G:AnOverviewoftheNextGenerationofIntelligent NetworkConnectivity. IEEE Access.

[18] Zeng, L., Ye, S., Chen, X., Zhang, X., Ren, J., Tang, J., ... & Shen, X. S. (2025). Edge Graph Intelligence: Reciprocally EmpoweringEdgeNetworkswithGraphIntelligence. IEEE Communications Surveys & Tutorials

[19] Li, L. (2023). A survey on intelligence-endogenous network: architecture and technologies for future 6G. Intelligent and Converged Networks, 5(1),53-67.

[20] Chataut, R., Nankya, M., & Akl, R. (2024). 6G networks and the AI revolution Exploring technologies, applications, and emergingchallenges. Sensors, 24(6),1888.

[21] Haider, S. A., Ramesh, J. V. N., Raina, V., Maaliw, R. R., Soni, M., Nasurova, K., ... & Singh, P. P. (2024). Secure artificial intelligenceforprecisevehiclebehaviorpredictionin6Gconsumerelectronics. IEEE Transactions on Consumer Electronics.

[22]Charpentier,V.,Landi,G.,Giannopoulou,E.,Brenes,J.,Camelo,M.,Marquez-Barja,J.M.,&Slamnik-Kriještorac,N.(2025). AdvancingVerticalServicesfor6G:FutureDirectionsandInnovations. IEEE Network

[23] Ismail, L., & Buyya, R. (2022). Artificial intelligence applications and self-learning 6G networks for smart cities digital ecosystems:Taxonomy,challenges,andfuturedirections. Sensors, 22(15),5750.

[24]Patel,R.,&Purohit,K.(2024).MassiveIoTaccessthrough6G.In 6G Communication Network (pp.254-261).CRCPress.

[25]Khan,I.(2024). Edge enhanced network monitoring using TinyML (Master'sthesis,I.Khan)

© 2025, IRJET | Impact Factor value: 8.315 | ISO 9001:2008 Certified Journal | Page1020

Turn static files into dynamic content formats.

Create a flipbook
Issuu converts static files into: digital portfolios, online yearbooks, online catalogs, digital photo albums and more. Sign up and create your flipbook.