AISR :A Protocol-Closed Architecture for Interference-Aware Channel-Mask Control in Industrial LoRaWAN

  • Renas Rajab AsaadDepartment of Information Technology, Technical College of Informatics–Akre, Akre University for Applied Sciences, Duhok, Iraq.
  • Madasu HanmandluDepartment of Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, India.
  • Kharabela SwainNstitute of Applied Sciences, Mangalayatan University, Aligarh, Uttar Pradesh, India.

Abstract

Industrial Long Range Wide Area Network (LoRaWAN) telemetry operates under variable propagation, contention, external interference, regional spectrum rules, and limited downlink capacity. An adaptive channel policy is useful only when the network can observe alternatives, predict a benefit over the time required to deliver control, issue a legal command that the device accepts, and verify that the realized transmissions improve application outcomes after sensing and reconfiguration cost. This paper reconstructs Adaptive Interference-Aware Spectrum Reallocation (AISR) as a protocol-closed network-side controller. A gateway or dedicated monitor supplies calibrated channel observations; a causal predictor estimates packet success, energy, and deadline risk; a constrained optimizer selects device-specific legal channel masks; and a verification layer records command acceptance, actual channel use, delivery, latency, and energy. The proposed architecture introduces control age, abstention, fallback, hysteresis, and auditability as first-class design variables. It also separates the evidential roles of the available harbor RSSI/SNR localization dataset from those of a genuine intervention trial. The harbor record can characterize observed link variation, but it lacks simultaneous counterfactual channel measurements, a complete attempted-packet denominator, MAC-command traces, application delivery times, and device-current measurements. No performance improvement is therefore claimed. A staged evaluation protocol compares fixed plans, configured Adaptive Data Rate (ADR), a valid history-based policy, published allocation families, and AISR under identical traffic, sensing budgets, channel conditions, and control authority. The resulting framework makes interference-aware LoRaWAN channel management falsifiable at the level of radio observation, protocol execution, and application benefit.

Keywords:Industrial long range wide area networkChannel-mask managementInterference predictionAdaptive data ratePacket deliveryControl ageDevice energyReproducibility

1. Introduction

Low-power wide-area networks support periodic sensing and supervisory telemetry across ports, factories, utilities, and geographically distributed industrial assets. Long Range Wide Area Network (LoRaWAN) combines long-range LoRa modulation with a networked MAC architecture, but reliability is shaped by propagation, channel occupancy, spreading factor, transmit power, offered load, capture, gateway diversity, receive-window behavior, and regional duty or dwell restrictions. Surveys of LoRaWAN, comparative LPWAN design, and massive-IoT scalability have consistently identified collisions, interference, downlink scarcity, and resource allocation as coupled limitations [1–5].

The attraction of interference-aware adaptation is straightforward: when one legal channel becomes persistently unfavorable, a network may prefer another. The implementation is not instantaneous. LoRaWAN channel availability is represented through regional and device-specific rules, and network control may be conveyed through MAC commands such as LinkADRReq under the applicable specification [6], [7]. The command must wait for a downlink opportunity, respect gateway airtime constraints, reach the device, be accepted, and affect a later uplink. Consequently, an observation can be stale before the selected mask becomes active.

Prior research already covered multiple parts of this problem. Joint spreading-factor and channel assignment was optimized in multi-operator networks [8]; adaptive resource allocation was compared with optimization-based alternatives [9]; multi-armed bandit and learning-automata approaches selected channels and spreading factors [10], [11]; game-theoretic and decentralized reinforcement-learning methods considered channel competition [12], [13]; and channel allocation was studied under intra- and inter-system interference [14]. Adaptive Data Rate (ADR)research addressed convergence, mobility, fairness, reliability, and energy [15–19]. Thus, the scientific gap is not a lack of adaptive algorithms. It is the closure of the complete observation– prediction–actuation–verification loop under actual LoRaWAN authority and cost.

Adaptive Interference-Aware Spectrum Reallocation (AISR) is formulated around that closure. The contribution is fourfold. First, a control contract binds every decision to a named regional plan, device capability, protocol version, observation source, and command path. Second, a robust objective balances lower-confidence delivery, energy, deadline risk, and update cost while permitting abstention. Third, control age and acceptance are modeled explicitly rather than treating an intended mask as an executed action. Fourth, a staged evidence protocol distinguishes exploratory link variation from counterfactual channel benefit and industrial transfer. The work is a proposed design; it does not report a completed implementation or validate the numerical improvements stated in the source manuscript.

2. Related Work and the Closure Gap

2.1 Scalability, Interference, and Resource Allocation

LoRaWAN scalability studies showed that pure-ALOHA approximations can omit capture, multi-channel reception, multiple spreading factors, and constrained downlinks. Empirical and simulation studies found that inter-network interference, device density, and downstream traffic can materially change delivery [20–22]. Broad surveys organized solutions around spreading-factor assignment, power, frequency allocation, gateway deployment, scheduling, and interference cancellation [1–4]. This literature established that a channel decision cannot be isolated from the other radio and MAC variables.

2.2 Adaptive Data Rate and Adaptive Channel Selection

ADR studies primarily adjusted spreading factor and transmit power, with extensions for mobility, convergence, fairness, energy, capacity planning, and transmission-parameter selection [15–19], [23], [24]. Channel-specific work ranged from proportional-fair optimization and multi-operator games to reinforcement learning, bandits, and TinyML [8–14], [25]. These methods differed in who controlled the action, what observations were available, and whether they modeled LoRa devices or LoRaWAN network commands. A result for direct device-side selection cannot automatically serve as a network-server baseline; conversely, a server policy is ineffective when downlink control cannot be delivered.

2.3 Data-Driven Prediction and Industrial Evidence

Industrial deployments confirm that nominal radio range does not remove application-specific constraints on throughput, energy, maintainability, and operational integration [26]. Consequently, a harbor trace can motivate hypotheses about link variation, but it cannot by itself demonstrate interference attribution, legal mask execution, or transfer to another industrial topology.

Machine-learning surveys documented growing use of supervised learning, reinforcement learning, and lightweight inference for LoRaWAN resource management [27], [28]. Experimental work also predicted packet delivery from observable link and interference conditions [29]. Yet prediction accuracy is not policy value: a predictor can be well calibrated while the chosen action is illegal, unavailable, delayed, or too costly. Industrial evidence additionally requires stable definitions of generated messages, deduplicated delivery, deadlines, current consumption, and surveillance duration. The ACOGA routing work in [30] concerns complementary network-layer energy management; its outcome cannot substitute for a LoRaWAN channel-mask comparator.

Table 1. Literature strands and the control-loop element they leave outside the claim.

This table is shown in the PDF version of the article.

3. System Boundary and Control Contract

Consider N end devices, one or more gateways, a network server, and the region-valid channel set Bᵣ. At decision epoch t, device i has an enabled-channel mask mᵢ,ₜ, data-rate and power state, firmware-defined channel-selection behavior, and a next feasible downlink opportunity. The control contract K records the LoRaWAN version, regional parameters, device class, firmware, default and mandatory channels, supported mask commands, duty or dwell restrictions, gateway downlink capacity, monitoring hardware, clock synchronization, and logging boundary. An action outside K is infeasible even if it optimizes a radio score. Observation also needs a boundary. Gateway packet metadata describe channels on which packets were successfully demodulated; they do not necessarily measure occupancy or interference on unchosen channels. Full channel sensing requires a calibrated receiver or monitor whose bandwidth, sampling schedule, detection threshold, coverage, and energy are documented. The policy must distinguish missing observation from a clean channel and must not infer the counterfactual response of an unobserved channel from RSSI alone. Observation, forecasting, constraint screening, MAC actuation, and realized-outcome verification form one auditable cycle. A missing stage opens the loop and prevents a causal policy claim.

This figure is shown in the PDF version of the article.

Figure 1. Protocol-closed AISR loop.

The harbor dataset in [31], [32] contains timestamps, device identity, location, and RSSI/SNR observations from three gateways for one mobile and two fixed nodes. The campaign used a stated physical configuration and is valuable for exploring temporal, spatial, device, and gateway variation. It does not record simultaneous outcomes on alternative channels, all generated messages and losses, MAC updates and acceptance, application delivery time, or current traces. It can support observation-model development and split design, but not a direct estimate of AISR policy value.

Table 2. Evidence boundary for the harbor record and an intervention trial.

This table is shown in the PDF version of the article.

Architecture

4.1 Forecasts with Uncertainty

controller that must abstain under uncertainty [33].

4. Adaptive Interference-Aware Spectrum Reallocation Decision

For candidate channel c and device i, a causal model estimates a lower-confidence packet-success probability q̲ᵢᶜ,ₜ, expected routine device energy Êᵢᶜ,ₜ, and expected end-to-end latency L̂ ᵢᶜ,ₜ over a forecast horizon H. Features use only observations available before t and include channel occupancy, recent delivery, RSSI/SNR summaries, data rate, power, payload, traffic, gateway diversity, device state, and time context. Calibration and prediction intervals are evaluated on held-out acquisition periods; point forecasts alone are insufficient for a

4.2 Feasible Masks and Normalized Objective

Let πᵢᶜ(mᵢ) denote the implementation-specific probability that device i transmits on channel c under mask mᵢ. For joint mask m, one candidate robust objective is minₘ Σᵢ wᵢ{α[1 − Σᶜπᵢᶜ(mᵢ)q̲ᵢᶜ(m)] + βÊᵢ(m)/E₀ + γL̂ ᵢ(m)/L₀} + λΣᵢCᵤᵖᵈᵃᵗᵉ(mᵢ,mᵢ,ₚᵣᵉᵛ). The weights are prespecified; E₀ and L₀ provide units; Cupdate includes incremental downlink airtime, receive-window energy, processing, and switching risk. Predictors depend on the joint action because concentrating devices on one channel may move rather than remove congestion.

The feasible set Fᵣ contains only masks allowed by regional parameters, mandatory-channel rules, device capability, protocol state, and available downlink budget. Additional constraints can enforce minimum channel diversity, fairness, deadline-risk limits, and a maximum update rate. The controller sends an update only if its conservative expected advantage exceeds a hysteresis margin δ and the forecast remains valid through the expected control age.

4.3 Control Age, Abstention, and Fallback

Define control age Δₜᵗˡ as the elapsed time from the final observation used by the decision to the first uplink actually affected by an accepted mask. It includes sensing, feature construction, inference, optimization, wait for a legal downlink, command delivery and acceptance, and wait for the next uplink. A policy optimized for horizon H must abstain when Δₜᵗˡ > H or when acceptance is uncertain. The active state changes only after logged acceptance; until then, the previous legal mask remains authoritative. A fallback mask preserves mandatory channels and service continuity.

This figure is shown in the PDF version of the article.

Figure 2. Control age of a LoRaWAN mask update.

The decision cannot affect traffic until a valid downlink is received and accepted and a later uplink uses the new state.

The prospective algorithm is therefore: observe; timestamp and validate data; form causal features; forecast with uncertainty; enumerate legal masks; solve the robust objective; apply the hysteresis and control-age tests; transmit a valid command when capacity exists; verify acceptance; and score only subsequently affected traffic. Failed or stale updates are outcomes, not rows to be silently removed.

5. Reproducible Evaluation

5.1 Experimental Stages and Comparators

Stage A tests monitoring calibration and prediction without adaptation. Stage B conducts controlled counterfactual probing or replay in which alternative channels are observed under matched conditions. Stage C activates protocol-valid mask updates and logs their full lifecycle. Stage D repeats the study at an independent industrial site. Comparators are a fixed legal plan, the network server’s configured ADR behavior, a simple history-based legal-mask policy, and AISR. Published allocation families in [8–14] are added only when their observations and action authority can be made equivalent.

All policies receive the same offered traffic, channel conditions, sensing budget, gateway set, device firmware, retry behavior, and evaluation horizon. The ADR configuration, including margin and update fields, is archived. The experiment logs generated application IDs, actual uplink channel/data rate/power, every gateway receipt, network-server deduplication, acknowledgments, downlinks, mask versions, acceptance, deadlines, monitor samples, and synchronized current traces. Train, validation, and test periods are split by session or time block before feature scaling or model selection.

5.2 Measures and Statistical Analysis

Packet-delivery ratio is the number of distinct generated messages delivered within the declared window divided by all generated messages. End-to-end latency runs from application generation to application delivery; undelivered messages and deadline misses are reported separately. Device energy is integrated over a declared boundary that includes sensing, MCU, transmit, receive windows, retries, and control, consistent with explicit LoRaWAN energy-model boundaries [34]. Goodput is delivered payload bits per time. Control overhead records command count, downlink airtime, acceptance, time to effect, failed updates, and incremental energy.

Repeated experiments report absolute outcomes, paired policy differences, uncertainty intervals, and failure cases. Time blocks or independent runs, rather than packets, are the resampling unit when temporal dependence is material. Calibration is assessed with reliability curves and proper scores [33]. Multiple algorithms are compared through predeclared omnibus and corrected pairwise procedures rather than isolated significance tests [35], [36]. Results are stratified by offered load, device count, mobility, channel occupancy, gateway diversity, distance, payload, and interference regime.

Table 3. Measures and raw material required to support them.

This table is shown in the PDF version of the article.

This figure is shown in the PDF version of the article.

Figure 3. Evidence intersection for an AISR claim.

Radio variation, protocol execution, and complete control cost must be observed together; any one component alone is insufficient.

5.3 Ablation and Stress Testing

Ablations remove the predictor, lower-confidence bound, joint congestion term, update cost, hysteresis, and abstention rule separately. Stress tests vary external interference, burst duration, mobility, missing monitor samples, clock error, gateway downlink saturation, command loss, firmware heterogeneity, and sudden channel-regime shifts. An oracle with future knowledge may be shown only as an upper bound. The main comparison remains causal and must include monitoring and control latency.

Table 4. Prespecified failure modes and safe controller responses.

This table is shown in the PDF version of the article.

empirical question.

are necessary before deployment claims.

7. Limitations and Research Program

6. Interpretation, Security, and Deployment Boundaries

The source manuscript reported gains in packet delivery, latency, and energy without the missing result table, generated-message denominator, synchronized event traces, current measurements, simulator definition, or uncertainty intervals required to reproduce them. Those percentages are not retained as findings. Likewise, plots of unexplained scores under unrelated technical labels cannot stand in for the defined communication outcomes. The rebuilt manuscript preserves the architectural hypothesis while returning efficacy to an open

Security belongs inside the control boundary. A forged monitor stream can make a healthy channel appear occupied; a modified channel mask can deny service; replayed acceptance can create false state; and compromised model or threshold files can bias allocation. Production studies should authenticate management traffic, version and sign models/configurations, record actor and timestamp for each state transition, rate-limit changes, and maintain a recoverable legal fallback. The controller must not disable mandatory channels or convert a monitoring optimization into an availability attack.

Industrial generalization also remains bounded. A harbor localization campaign demonstrates harsh-environment link variability but does not represent every factory layout, interference source, traffic mix, or regulatory region. Periodic telemetry results should not be narrated as closed-loop safety control. Independent site replication, long negative-exposure periods, maintenance behavior, gateway failure, and firmware diversity

AISR is not yet an implemented predictor, optimizer, firmware feature, or field-tested system. The mask-to-channel probability πᵢᶜ depends on device behavior and must be measured. Counterfactual channel response is only partially identifiable without randomized probing or simultaneous monitoring. Joint action optimization can become combinatorial at scale. Downlink scarcity can dominate any radio advantage, while monitoring energy may lie outside the end-device budget but still inside the system boundary.

The strongest next program has four linked studies. First, reproduce the harbor observation analysis with session-wise splits and calibrated forecasts, making no policy claim. Second, construct a controlled multi-channel testbed with attempted-message, monitor, MAC, and current traces. Third, evaluate fixed, ADR, history-based, published channel-allocation, and AISR policies under identical authority and load. Fourth, repeat a frozen controller at an independent industrial site. Each stage should preserve raw events, configuration, code, seeds, firmware hashes, and table-generation scripts so later evidence extends rather than replaces the earlier record.

8. Conclusion

AISR is recast as a protocol-closed controller rather than an unconstrained channel-ranking algorithm. A valid decision requires observable alternatives, causal and calibrated forecasts, a region- and device-legal mask, a deliverable MAC command, explicit acceptance, actual use of the new state, and a net application benefit after sensing and control cost. Control age, abstention, hysteresis, fallback, and an event-level audit trail connect these requirements into a testable architecture. The cited harbor data can support link-variation analysis but cannot establish counterfactual delivery, latency, or energy gains. Those claims require a new intervention record or a calibrated simulator followed by independent field validation. The paper therefore offers a rigorous route to evidence while leaving performance superiority deliberately unresolved.

Acknowledgments

The authors would like to acknowledge the researchers and scholars whose previous contributions to LoRaWAN networks, interference management, adaptive spectrum allocation, and industrial Internet of Things have provided a valuable foundation for this study.

Author Contributions

R. R. A.: Conceptualization, methodology, investigation, and writing—original draft.

M. H.: Technical analysis, theoretical development, and critical review.

Kh. S.: Supervision, validation framework, and writing—review and editing.

All authors reviewed and approved the final manuscript.

Funding

This research received no specific grant from any funding agency in the public, commercial, or nonprofit sectors.

Data Availability

The data supporting this study include previously published research and an existing harbor RSSI/SNR localization dataset. No new experimental dataset was generated. The limitations of the available dataset and the requirements for further validation are discussed in the manuscript.

Conflicts of Interest

The authors declare that they have no financial, professional, or personal conflicts of interest that could have influenced the research, its interpretation, or the publication of this article.

References

  1. [1] Sundaram, J. P. S., Du, W., & Zhao, Z. (2019). A survey on LoRa networking: Research problems, current solutions, and open issues. IEEE Communications Surveys & Tutorials, 22(1), 371–388. https://doi.org/10.1109/COMST.2019.2949598
  2. [2] Jouhari, M., Saeed, N., Alouini, M.S., & Amhoud, E. M. (2023). A survey on scalable LoRaWAN for massive IoT: Recent advances, potentials, and challenges. IEEE Communications Surveys & Tutorials, 25(3), 1841–1876. https://doi.org/10.1109/COMST.2023.3274934
  3. [3] Haxhibeqiri, J., Poorter, E. De, Moerman, I., & Hoebeke, J. (2018). A survey of LoRaWAN for IoT: From technology to application. Sensors, 18(11), 3995. https://doi.org/10.3390/s18113995
  4. [4] Sun, Z., Yang, H., Liu, K., Yin, Z., Li, Z., & Xu, W. (2022). Recent advances in LoRa: A comprehensive survey. ACM Transactions on Sensor Networks, 18(4), 1–44. https://doi.org/10.1145/3543856
  5. [5] Ikpehai, A., Adebisi, B., Rabie, K. M., Anoh, K., Ande, R. E., Hammoudeh, M., ..., & Mbanaso, U. M. (2018). Low-power wide area network technologies for internet-of-things: A comparative review. IEEE Internet of Things Journal, 6(2), 2225-2240. https://doi.org/10.1109/JIOT.2018.2883728
  6. [6] Povalac, A., Kral, J., Arthaber, H., Kolar, O., & Novak, M. (2023). Exploring lorawan traffic: In-depth analysis of iot network communications. Sensors, 23(17), 7333. https://doi.org/10.3390/s23177333
  7. [7] Alliance, L. (2022). RP002-1.0. 4 Regional Parameters. https://resources.lora-alliance.org/technical-specifications/rp002-1-0-4-regional-parameters
  8. [8] Fawaz, H., Khawam, K., Lahoud, S., Adjih, C., & Martin, S. (2020). Joint spreading factor and channel assignment in multi-operator LoRaWAN deployments. Sensors, 21(1), 162. https://doi.org/10.3390/s21010162
  9. [9] Moraes, J., Oliveira, H., Cerqueira, E., Both, C., Zeadally, S., & Rosário, D. (2022). Evaluation of an adaptive resource allocation for lorawan. Journal of Signal Processing Systems, 94(1), 65–79. https://doi.org/10.1007/s11265-021-01678-8
  10. [10] Urabe, I., Li, A., Fujisawa, M., Kim, S.J., & Hasegawa, M. (2023). Combinatorial mab-based joint channel and spreading factor selection for lora devices. Sensors, 23(15), 6687. https://doi.org/10.3390/s23156687
  11. [11] Atadet, L. A., Musabe, R., Hitimana, E., & Gatera, O. (2025). Optimizing LoRaWAN performance through learning automata-based channel selection. Future internet, 17(12), 555. https://doi.org/10.3390/fi17120555
  12. [12] Khawam, K., Fawaz, H., Lahoud, S., Maillard, O.A., & Martin, S. (2022). A channel selection game for multi-operator LoRaWAN deployments. Computer Networks, 216, 109185. https://doi.org/10.1016/j.comnet.2022.109185
  13. [13] Li, A., Fujisawa, M., Urabe, I., Kitagawa, R., Kim, S. J., & Hasegawa, M. (2021). A lightweight decentralized reinforcement learning based channel selection approach for high-density LoRaWAN. 2021 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN) (pp. 9-14). IEEE. https://doi.org/10.1109/DySPAN53946.2021.9677146
  14. [14] Kakuda, S., Yamazaki, Y., Katagiri, K., Fujii, T., Takyu, O., Ohta, M., & Adachi, K. (2021). Channel allocation for lorawanconsidering intra-system and inter-system interferences. 2021 IEEE 94th Vehicular Technology Conference (VTC2021-Fall) (pp. 01-07). IEEE. 2021 IEEE 94th vehicular technology conference (vtc2021-fall) (pp. 1–7). https://doi.org/10.1109/VTC2021-Fall52928.2021.9625176
  15. [15] Kufakunesu, R., Hancke, G. P., & Abu-Mahfouz, A. M. (2020). A survey on adaptive data rate optimization in lorawan: Recent solutions and major challenges. Sensors, 20(18), 5044. https://doi.org/10.3390/s20185044
  16. [16] Sandoval, R. M., Garcia-Sanchez, A.J., & Garcia-Haro, J. (2019). Optimizing and updating lora communication parameters: A machine learning approach. IEEE Transactions on Network and Service Management, 16(3), 884–895. https://doi.org/10.1109/TNSM.2019.2927759
  17. [17] Reynders, B., Wang, Q., Tuset-Peiro, P., Vilajosana, X., & Pollin, S. (2018). Improving reliability and scalability of LoRaWANs through lightweight scheduling. IEEE Internet of Things Journal, 5(3), 1830–1842. https://doi.org/10.1109/JIOT.2018.2815150
  18. [18] Li, S., Raza, U., & Khan, A. (2018). How agile is the adaptive data rate mechanism of LoRaWAN?. In 2018 IEEE Global Communications Conference (GLOBECOM) (pp. 206-212). IEEE. https://doi.org/10.1109/GLOCOM.2018.8647469
  19. [19] Abdelfadeel, K. Q., Cionca, V., & Pesch, D. (2018). Fair adaptive data rate allocation and power control in LoRaWAN. In 2018 IEEE 19th International Symposium on" A World of Wireless, Mobile and Multimedia Networks"(WoWMoM) (pp. 14-15). IEEE. https://doi.org/10.1109/WoWMoM.2018.8449737
  20. [20] Mikhaylov, K., Petäjäjärvi, J., & Janhunen, J. (2017). On LoRaWAN scalability: Empirical evaluation of susceptibility to inter-network interference. In 2017 European Conference on Networks and Communications (EuCNC) (pp. 1-6). IEEE. https://doi.org/10.1109/EuCNC.2017.7980757
  21. [21] Haxhibeqiri, J., den Abeele, F., Moerman, I., & Hoebeke, J. (2017). LoRa scalability: A simulation model based on interference measurements. Sensors, 17(6), 1193. https://doi.org/10.3390/s17061193
  22. [22] den Abeele, F., Haxhibeqiri, J., Moerman, I., & Hoebeke, J. (2017). Scalability analysis of large-scale LoRaWAN networks in ns-3. IEEE Internet of Things Journal, 4(6), 2186–2198. https://doi.org/10.1109/JIOT.2017.2768498
  23. [23] Elshabrawy, T., & Robert, J. (2019). Capacity planning of LoRa networks with joint noise-limited and interference-limited coverage considerations. IEEE Sensors Journal, 19(11), 4340–4348. https://doi.org/10.1109/JSEN.2019.2897156
  24. [24] Bor, M., & Roedig, U. (2017). LoRa transmission parameter selection. In 2017 13th International conference on distributed computing in sensor systems (DCOSS) (pp. 27-34). IEEE. https://doi.org/10.1109/DCOSS.2017.10
  25. [25] Lodhi, M. A., Obaidat, M. S., Wang, L., Mahmood, K., Qureshi, K. I., Chen, J., & Hsiao, K.F. (2024). Tiny machine learning for efficient channel selection in LoRaWAN. IEEE Internet of Things Journal, 11(19), 30714–30724. https://doi.org/10.1109/JIOT.2023.3342277
  26. [26] Margelis, G., Piechocki, R., Kaleshi, D., & Thomas, P. (2015). Low throughput networks for the IoT: Lessons learned from industrial implementations. 2015 IEEE 2nd World Forum on Internet of Things (WF-IoT) (pp. 181- 186). IEEE. https://doi.org/10.1109/WF-IoT.2015.7389049
  27. [27] Farhad, A., & Pyun, J.Y. (2023). LoRaWAN meets ML: A survey on enhancing performance with machine learning. Sensors, 23(15), 6851. https://doi.org/10.3390/s23156851
  28. [28] Ayoub Kamal, M., Alam, M. M., Sajak, A. A. B., & Mohd Su’ud, M. (2023). Requirements, deployments, and challenges of LoRa technology: A survey. Computational Intelligence and Neuroscience, 2023(1), 5183062. https://doi.org/10.1155/2023/5183062
  29. [29] Guerrero, M., Cano, C., Vilajosana, X., & Thubert, P. (2021). Towards dependable IoT via interface selection: predicting packet delivery at the end node in LoRaWAN networks. Sensors, 21(8), 2707. https://doi.org/10.3390/s21082707
  30. [30] Landge, P., Singh, Y., Mohapatra, H., & Edalatpanah, S. (2025). Efficient resource management in IoT network through ACOGA Algorithm. Computer modeling in engineering & sciences, 143(2), 1661. http://dx.doi.org/10.32604/cmes.2025.065599
  31. [31] Fhima, H., Mahfoudhi, S., Almutiq, M., & Bouallegue, R. (2026). Accurate LoRaWAN localization in harbors: A comparative study of VMD-BiLSTM and VMD-GAT-BiLSTM for Spatiotemporal RSSI Forecasting. IEEE Access. https://doi.org/10.1109/ACCESS.2026.3727851
  32. [32] Moradbeikie, A., Zare, M., Keshavarz, A., & Lopes, S. I. (2024). RSSI-based LoRaWAN dataset collected in a dynamic and harsh industrial environment with high humidity. Data in Brief, 53, 110120. https://doi.org/10.1016/j.dib.2024.110120
  33. [33] Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks. In International Conference on Machine Learning (pp. 1321-1330). PMLR. https://proceedings.mlr.press/v70/guo17a/guo17a.pdf
  34. [34] Bouguera, T., Diouris, J.F., Chaillout, J.J., Jaouadi, R., & Andrieux, G. (2018). Energy consumption model for sensor nodes based on LoRa and LoRaWAN. Sensors, 18(7), 2104. https://doi.org/10.3390/s18072104
  35. [35] Demšar, J. (2006). Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research, 7(Jan), 1–30. https://www.jmlr.org/papers/volume7/demsar06a/demsar06a.pdf
  36. [36] Holm, S. (1979). A simple sequentially rejective multiple test procedure. Scandinavian journal of statistics, 65–70. https://www.jstor.org/stable/4615733