PaperlayerView exact revision on arXiv

Source-linked AI summary

Cellular, Wide-Area, and Non-Terrestrial IoT: A Survey on 5G Advances and the Road Towards 6G

Mojtaba Vaezi, Amin Azari, Saeed R. Khosravirad, Mahyar Shirvanimoghaddam, M. Mahdi Azari, Danai Chasaki, Petar Popovski

arXiv:2107.03059v1cs.NIeess.SP

TL;DR

IoT requires large-scale, long-lasting, reliable, and near-real-time connectivity for diverse devices and data. This survey synthesizes cellular, wide-area, and non-terrestrial IoT communication solutions, including 5G enhancements, unlicensed-spectrum technologies, and emerging 6G directions. It reviews mechanisms addressing energy efficiency, reliability, latency, and scalability, while identifying future research directions.

  • Problem

    Providing large-scale, long-lasting, reliable, and near-real-time connectivity for diverse IoT devices and applications remains a major challenge.

  • Method

    The paper comprehensively surveys cellular, wide-area, and non-terrestrial IoT solutions, standards, enabling techniques, and emerging 6G technologies.

  • Results

    The survey reviews solutions for energy efficiency, reliability, low latency, and scalability, including grant-free access, short-packet channel coding, NOMA, and on-device intelligence.

  • Takeaways & Limitations

    The paper provides a supported overview of current IoT connectivity technologies and identifies future research directions toward beyond-5G IoT networks.

Abstract

from arXiv · show

The next wave of wireless technologies is proliferating in connecting things among themselves as well as to humans. In the era of the Internet of things (IoT), billions of sensors, machines, vehicles, drones, and robots will be connected, making the world around us smarter. The IoT will encompass devices that must wirelessly communicate a diverse set of data gathered from the environment for myriad new applications. The ultimate goal is to extract insights from this data and develop solutions that improve quality of life and generate new revenue. Providing large-scale, long-lasting, reliable, and near real-time connectivity is the major challenge in enabling a smart connected world. This paper provides a comprehensive survey on existing and emerging communication solutions for serving IoT applications in the context of cellular, wide-area, as well as non-terrestrial networks. Specifically, wireless technology enhancements for providing IoT access in fifth-generation (5G) and beyond cellular networks, and communication networks over the unlicensed spectrum are presented. Aligned with the main key performance indicators of 5G and beyond 5G networks, we investigate solutions and standards that enable energy efficiency, reliability, low latency, and scalability (connection density) of current and future IoT networks. The solutions include grant-free access and channel coding for short-packet communications, non-orthogonal multiple access, and on-device intelligence. Further, a vision of new paradigm shifts in communication networks in the 2030s is provided, and the integration of the associated new technologies like artificial intelligence, non-terrestrial networks, and new spectra is elaborated. Finally, future research directions toward beyond 5G IoT networks are pointed out.

I. INTRODUCTION

IoT connects diverse devices and data sources across application segments and network ranges, creating requirements for scalable, energy-efficient, reliable, and low-latency wireless connectivity. This survey examines cellular, wide-area, and non-terrestrial IoT networks, their technologies, applications, challenges, and 5G-related enhancements.

  • IoT foundations: IoT combines diverse devices, embedded systems, and heterogeneous streaming data to support smarter or automated decision-making.Devices range from low-cost sensors and wearables to actuators, robots, drones, and autonomous vehicles; generated data varies in volume, velocity, variety, variability, veracity, and value.
  • Network types: IoT networks are commonly classified by coverage range into cellular, wide-area, and short-range categories, with non-terrestrial networks adding extremely wide-area three-dimensional coverage.Non-terrestrial networks generalize cellular and wide-area networks and include UAV- and satellite-based systems.
  • Application types: Massive IoT connects many low-power endpoints transmitting small, often sporadic data volumes, whereas critical IoT uses fewer endpoints for high-bandwidth, low-latency real-time applications.Massive IoT includes sensing and metering; critical IoT includes industrial control, robotics, and autonomous vehicles.
  • Application types: Broadband and industrial IoT extend the application landscape through higher-throughput connectivity and manufacturing-focused interconnection, automation, and real-time data.Broadband IoT supports applications such as advanced wearables, drones, augmented reality, and virtual reality, while industrial IoT supports Industry 4.0.
  • Connectivity challenges: Selecting IoT connectivity requires trade-offs among coverage, power consumption, battery size, deployment cost, data rate, and other network characteristics.No single wireless technology fits every IoT situation; the survey compares cellular and non-cellular technologies across these dimensions.
  • 5G connectivity: 5G is designed as an IoT enabler through eMBB, mMTC, and URLLC use cases corresponding broadly to broadband, massive, and critical or industrial IoT.The survey relates IoT requirements to 5G key performance indicators, including the requirements associated with mMTC and URLLC.

1) Technical requirements:

The survey examines IoT requirements and enabling technologies across 5G, beyond-5G, cellular, wide-area, and non-terrestrial networks. It emphasizes emerging shifts toward 6G involving 3D coverage, AI integration, broader frequency use, IoT-dominant networks, and solutions for diverse KPI demands.

  • Vision for 6G: 6G envisions 3D coverage through non-terrestrial networks, including UAVs and satellite communication networks, across ground, sea, and space.
  • Vision for 6G: AI and machine learning are envisioned across core, access, edge, and device domains to improve network management, automation, efficiency, and smart applications.
  • Vision for 6G: 6G is expected to exploit terahertz and optical bands alongside existing RF and millimeter-wave bands to pursue 1000 Gbps peak data rates.Propagation characteristics, use cases, and hardware requirements differ across these frequency bands.
  • Survey scope: The survey covers connectivity, channel coding, massive access, security, deep and federated learning, non-terrestrial networks, and future energy-efficiency solutions.Its connectivity analysis includes licensed and unlicensed technologies, NOMA, random access, relaying, and mesh-type network architectures.
  • Related works: The paper identifies a gap in surveys covering all 5G IoT KPIs together with beyond-5G architectures, AI, and security solutions.It presents itself as more comprehensive than prior surveys in these areas.
  • Technical requirements: The survey addresses diverse IoT requirements including latency, reliability, survival time, service availability, age of information, battery lifetime, energy efficiency, and connection density.These requirements can be in deep contrast across envisioned IoT use cases.

II. KEY PERFORMANCE INDICATORS

The survey organizes IoT performance requirements around diverse latency, reliability, and service demands, from 5G mission-critical targets to more extreme 6G expectations. It also emphasizes that end-to-end latency includes multiple propagation, processing, and queuing components.

  • Latency and Reliability: 5G targeted 99.999% reliability within 1 ms for 32-byte packets in mission-critical applications.The target was associated with physical-layer and MAC-layer advances in 3GPP Release 15.
  • Latency and Reliability: End-to-end latency includes air and x-haul propagation, processing, and queuing delays.Processing delays and x-haul propagation are identified as critical contributors to experienced round-trip delay.
  • Latency and Reliability: Reliability and latency requirements vary widely across use cases, with presented packet error rates ranging from 10^-4 to 10^-10.Traffic types also motivate different access modes, including point-to-point, device-to-device, multicast, and broadcast.
  • Latency and Reliability: 6G-era cellular, wide-area, and non-terrestrial technologies are expected to support radio latencies near 100 µs at Gbits/s data rates.These requirements are framed as enabling extreme networking solutions and potentially replacing traditional industrial wireline connectivity.
  • Latency and Reliability: The survey compares envisioned use cases and technologies by latency and reliability, while noting that the plotted values are approximate.The comparison draws on reported measurements and envisioned capabilities in the literature.

B. Survival Time and Service Availability

Survival time and service availability extend reliability analysis from individual packet errors to application-level tolerance of consecutive failures. Their importance and acceptable values depend strongly on the application, traffic pattern, and deployment context.

  • Definitions: Survival time measures how long an application may continue without an anticipated message, while service availability measures how often communication QoS requirements are satisfied.Service availability includes end-to-end latency and survival-time requirements.
  • Application Dependence: Motion control systems may tolerate lost packets for roughly one transfer interval, whereas remote harbor-crane control may tolerate up to six consecutive transfer intervals.The difference reflects application-specific operating-accuracy requirements.
  • Design Implications: Introducing survival time shifts the design focus from minimizing packet error rate to minimizing bursts of consecutive errors.Error bursts longer than the survival-time duration must be avoided.
  • Design Implications: Latency and survival time have no direct relation and can vary independently with use case and service area.Short physical distances can support low end-to-end latency for nearby motion-control nodes.
  • Survival Mode: Survival-mode operation becomes necessary when a link misses the last packet within the end-to-end latency budget.A suitable strategy is needed because cyberphysical applications can tolerate failures only while consecutive-failure duration remains bounded.

F. Data Rate and Spectral Efficiency

Data-rate and spectral-efficiency requirements motivate higher-capacity technologies, while IoT access solutions balance coverage, power, cost, and reliability. The survey contrasts cellular, dedicated low-power, and unlicensed approaches for diverse deployments.

  • Data Rate: 5G peak data rates are expected to reach 20 Gbps, while 6G peak data rates are projected near 1,000 Gbps.The survey identifies mmWave, THz, and massive MIMO as facilitating technologies, while not discussing data-rate solutions in detail.
  • Connectivity Landscape: IoT connectivity technologies are classified according to data-rate and transmission-range requirements, spanning Zigbee, WiFi, Bluetooth, BLE, WiFi-HaLow, and RFID.These technologies target different combinations of range, power dissipation, and application needs.
  • Cellular Access: Cellular IoT traffic may use legacy procedures, new low-power protocols alongside legacy traffic, or dedicated radio resources and IoT-specific protocols.The survey investigates the latter two cases rather than legacy 2G-style service.
  • Cellular Access: LTE-M and NB-IoT were introduced to provide low-cost IoT connectivity over dedicated resources, including narrowband physical channels and signals in NB-IoT.NB-IoT also supports deep sleep, while eDRX and PSM provide additional power-saving mechanisms.
  • Unlicensed Spectrum: Unlicensed-spectrum solutions such as SigFox and LoRaWAN simplify access and can reduce device cost and energy use, but reliability is challenged by grant-free access.LoRaWAN centrally allocates spreading factors to balance cell-wide data rate and communication reliability.

3) Comparison of different technologies:

The survey compares wide-area IoT enablers and emerging spectrum options while examining coverage, throughput, latency, reliability, and deployment constraints. It also discusses cooperation and coding techniques for extending coverage, improving efficiency, and supporting short-packet reliability.

  • Comparison of different technologies: Wide-area IoT enablers span unlicensed SigFox and LoRaWAN and licensed LTE-M and NB-IoT, with coverage and throughput as key comparison dimensions.Table V compares these technologies, while Fig. 9 illustrates their throughput/coverage trade-offs.
  • New Spectra: mmWave, THz, and optical spectrum can provide higher data rates, but mmWave is better suited to broadband and critical IoT than massive connectivity.The wide bandwidth of these spectra supports ultra-high rates, whereas mmWave propagation loss limits its suitability for massive connectivity.
  • New Spectra: mmWave propagation loss and blocking limit communication to tens or hundreds of meters, requiring dense access points and multi-link communication for ubiquitous coverage.Small wavelengths enable high-gain beamforming, but dense radio APs and multiple transmission reception points help mitigate blocked links.
  • New Spectra: Large mmWave bandwidth and contained delay spread can support over-the-air latencies of tens of microseconds for industrial use cases.The bands also offer outdoor-to-indoor penetration losses that can reduce external interference and jamming in factories.
  • Adaptive Network Topology: Low-cost, low-power IoT sensors motivate mesh networking, allowing nodes to forward both their own data and other nodes’ data.WSN-based technologies such as Zigbee and LoRa build low-power mesh networks among inexpensive sensors.
  • Adaptive Network Topology: Adaptive cooperation combines D2D relaying and multi-hop transmission to improve reliability, achieving 10^-9 system error probability for 30 devices at 5 dB SNR.Using channel information to activate relaying on demand can improve spectral efficiency and reduce the interference footprint.
  • Short-Packet Communications: Short-packet IoT communications face reduced coding gain and a larger gap to Shannon’s limit as block length decreases.Shorter packets are needed to meet delay-sensitive applications’ transmission-time requirements.

A. State-of-the-Art Channel Coding Techniques for IoT

IoT channel coding requires low-complexity processing, short blocks, and low modulation orders, while candidate codes trade short-block performance, rate flexibility, complexity, and delay. The survey also examines coding for low-capacity or CSI-free channels, where unreliable symbols and impractical feedback complicate adaptation.

  • IoT channel coding requires low-complexity encoders and decoders, short information blocks, and low modulation orders for low-power operation.
  • LDPC-based 5G codes are effective at moderate-to-long block lengths but show poor short-block performance and an error floor that limits short-packet use.
  • TBCC performs best at very short block lengths, while polar codes maintain strong short- and moderate-block BLER without an error floor.Polar decoding nevertheless has relatively high list-decoding complexity and can introduce delay unsuitable for delay-sensitive applications.
  • Turbo codes perform reasonably well at moderate block lengths and offer fast, very-low-complexity decoders, making them favorable candidates for some IoT applications.
  • Low-capacity channels produce very small LLRs, making iterative LDPC or Raptor decoding slow to converge and successive-interference cancellation unreliable.
  • Rateless and repetition-based schemes provide channel adaptation without receiver CSI, but quantization can degrade decoding and short blocks move capacity away from Shannon limits.

D. Decoder Design

Decoder design for IoT must balance error performance, complexity, latency, power consumption, and limited channel-state information. The survey also motivates non-orthogonal access because massive IoT cannot economically assign dedicated resources to every device.

  • D. Decoder Design: Algebraic codes offer strong ML-decoding BLER performance, but near-ML decoding is complex, latency-intensive, and power-inefficient for low-power IoT devices.
  • D. Decoder Design: OSD-based decoding can reduce average complexity, but worst-case searches remain large and iterative processing is difficult to parallelize efficiently.
  • D. Decoder Design: Simplified LDPC decoders such as min-sum and offset min-sum enable parallel hardware implementation, but their approximations cause performance loss, especially at low rates and short-to-moderate block lengths.
  • D. Decoder Design: Polar SC decoding degrades at moderate-to-short block lengths; list decoding improves performance but scales complexity with list size, while serial decoding limits speed.CRC assistance can improve decoding performance but adds further complexity.
  • D. Decoder Design: Polar and LDPC decoders require receiver-side CSI, so decoding fails without it; CRC-based error detection also adds substantial overhead to short packets.
  • Massive IoT combines extremely high device density, sporadic small payloads, limited power, and potentially distant base stations, motivating specialized access designs.
  • NOMA lets multiple users share a resource block, addressing the impracticality of dedicated allocation for exploding numbers of massive IoT devices.Its schemes use signatures to distinguish users and cancel inter-user interference, with distinct uplink and downlink designs.

Transmitter Structure:

NOMA transmitter designs distinguish users through signatures applied at bit or symbol level, while code-domain schemes trade receiver complexity against overloading. Grant-based access requires multi-step coordination, creating latency and collision limitations for massive IoT.

  • Transmitter Structure:: Code-domain NOMA applies user-distinguishing operations at bit or system level, including interleaving, scrambling, and spreading.Bit-level scrambling requires less processing delay and memory than bit-level interleaving; symbol-level spreading is the largest category.
  • Transmitter Structure:: SCMA directly maps incoming bits to spread codewords in layer-specific codebooks, merging bit-to-symbol mapping and spreading.Its sparse codewords support low-complexity reception and additional multidimensional constellation design.
  • Transmitter Structure:: SCMA’s sparser codewords tolerate more overloading for massive uplink IoT connectivity, but reduce coding gain and retain high downlink decoding complexity.The downlink complexity limits SCMA’s ability to support massive IoT connectivity on low-cost devices.
  • Transmitter Structure:: Table VII compares selected code-domain NOMA schemes by complexity and overloading factor, favoring schemes with higher overloading and lower complexity.The table supports evaluating NOMA designs through these two criteria rather than either criterion alone.
  • Transmitter Structure:: Grant-based random access uses preamble selection, a base-station grant, a connection request, and resource-resolution messaging before payload transmission.The procedure can take 10 ms or more; collisions force affected devices to restart, and collision probability increases with active-device count.

2) Grant-free random access:

Grant-free and unsourced access reduce coordination for massive IoT, but require reliable activity detection, efficient coding, and security protections across constrained devices and networks.

  • 2) Grant-free random access:: Grant-free transmission eliminates the scheduling request for uplink data and can reduce access latency compared with grant-based protocols.The base station detects active devices, assigns a common preamble across time slots, and estimates channels from received metadata.
  • 2) Grant-free random access:: Device-activity detection is the main grant-free challenge, with compressed sensing, approximate message passing, massive MIMO, and deep-learning methods proposed to address it.These approaches target detection of active devices and recovery of sparse signals.
  • 2) Grant-free random access:: Unsourced multiple access targets massive uncoordinated users sending small payloads, where the base station recovers message lists without identifying individual transmitters.This setting provides an information-theoretic model for sporadic massive machine-type communication.
  • 2) Grant-free random access:: Practical unsourced massive random access still needs more efficient codebook design and activity-detection algorithms.These are identified as remaining challenges despite advances in low-complexity coding, massive MIMO extensions, and chirp detection.
  • 2) Grant-free random access:: IoT deployments remain vulnerable through physical, software, and network attack paths, while many standards lack consistently applied end-to-end security.Resource and power constraints limit the security capabilities of many IoT devices.

3) Network:

IoT network security spans device, software, and network layers, motivating programmable management and distributed protection while exposing unresolved authentication and key-management constraints.

  • 3) Network:: Wide-area IoT technologies provide authentication and encryption between devices and gateways, yet remain susceptible to impersonation, replay, and availability attacks.Replay attacks reuse captured authentication packets, while routing attacks also affect low-power wide-area networks.
  • 3) Network:: Attackers can target gateway-server links through man-in-the-middle manipulation or overwhelm wide-area servers with service requests.These attacks can delay, disrupt, or modify messages and cause server malfunction.
  • 3) Network:: SDN can optimize resource allocation and move security intelligence to gateways that analyze traffic, identify abnormal behavior, and block or forward packets.This supports resource-constrained IoT devices through centralized gateway-based enforcement.
  • 3) Network:: Blockchain validates transactions distributively and can identify misbehaving IoT devices, while hashing and public-key cryptography support integrity and privacy by design.These properties are presented as relevant to scaling cellular and wide-area IoT deployments.
  • 3) Network:: Traditional key storage and certificate-exchange methods do not scale with IoT volume, leaving key exchange as a future research need.The limitation applies even where public-key cryptography is used.

B. Deep Learning Models

Deep learning extends machine learning by learning complex features from raw data and can support diverse IoT objectives, but deployment remains constrained by computation, data, privacy, and model-design demands.

  • B. Deep Learning Models: Deep neural networks are artificial neural networks with two or more hidden layers that learn hidden features from raw data for regression and classification.Deeper networks can represent more complex features and produce more accurate learning.
  • B. Deep Learning Models: Deep learning improves accuracy and reduces reliance on hand-crafted features by learning complex decision boundaries and automating feature engineering.These benefits distinguish deep learning from conventional machine-learning techniques.
  • B. Deep Learning Models: Mobile and IoT networks can use unsupervised deep-learning models for unlabeled data and train one model for multiple objectives to reduce computational and memory requirements.Restricted Boltzmann machines and generative adversarial networks are examples of models for unlabeled data.
  • B. Deep Learning Models: Deep learning is limited by adversarial vulnerability, costly or infeasible data collection, computational expense on IoT devices, and difficult hyperparameter optimization.These limitations complicate applying deep learning across resource-constrained IoT settings.
  • B. Deep Learning Models: Open research questions concern sensory-data fusion, efficient deployment on constrained devices or in the cloud, reduced dependence on labeled data, distributed learning, and DNN accelerators.The survey frames these questions as requirements for effective IoT learning systems.

C. Autoencoder

The survey presents autoencoders as neural networks that learn compressed representations by reconstructing inputs, supporting anomaly and intrusion detection in IoT. It also reviews reinforcement-learning methods, including deep reinforcement learning, for communication, computing, and resource-management tasks.

  • Autoencoder: Autoencoders compress input data and reconstruct it while minimizing reconstruction differences.Their purpose is to preserve data quality in a smaller representation.
  • Autoencoder: Autoencoders support fault, intrusion, disease, and general anomaly detection by learning to reproduce normal inputs.Anomaly detection identifies rare items or events that differ significantly from expected behavior.
  • Deep Reinforcement Learning: Reinforcement learning trains agents through environmental interaction and rewards to learn action sequences for complex objectives.RL is typically modeled as a Markov decision process.
  • Deep Reinforcement Learning: Deep reinforcement learning combines artificial neural networks with reinforcement learning and is applied across IoT perception, communication, and edge or cloud layers.Applications include robotics, autonomous driving, resource allocation, scheduling, localization, offloading, and caching.
  • Federated Learning: Federated learning distributes model training across devices by sending models to local data and aggregating device updates without accessing raw data.The server repeatedly distributes an initial model, receives locally trained results, and generates a global model.
  • IoT Communication Trade-offs: Reducing signaling can lower IoT energy consumption per transfer, but unlicensed-spectrum systems lose centralized control and performance guarantees.The survey identifies this as a trade-off between device energy use and network control.

F. IoT-Friendly DL

The survey examines techniques for adapting deep learning to resource-constrained IoT devices. It covers model compression, compact architectures, and neural-network accelerators while emphasizing the severe memory gap between microcontrollers and larger computing platforms.

  • Motivation: Deep learning models can contain millions to billions of parameters, making them computationally expensive for resource-constrained IoT and edge devices.Commercial computer-vision CNNs may contain approximately 10^8 parameters.
  • Compression and Acceleration: DNN complexity can be reduced through network compression and acceleration with negligible performance degradation.The survey organizes techniques around low-rank approximation, pruning, quantization, compact models, and specialized hardware.
  • Compression Techniques: Low-rank approximation, pruning, and quantization reduce model computation or representation size by approximating layers, removing weak connections, or reducing weight precision.Quantization can also apply to activations, errors, and gradients.
  • Compression Techniques: Deep compression combines pruning and k-means-based quantization to reduce DNN size and storage requirements by more than an order of magnitude without losing accuracy.The procedure trains, prunes, and quantizes the network sequentially.
  • Compact Models: Compact models are designed as small CNN architectures from scratch, while SqueezeNet reduces filter parameters nine times and overall parameters about 50 times.These designs preserve acceptable accuracy while reducing model size.
  • Hardware Acceleration: Neural-network accelerators use parallel computing and processing in memory to reduce DNN computation and memory requirements.The survey identifies specialized hardware as an active research direction for minimizing energy and storage in IoT devices.
  • Device Constraints: Microcontroller SRAM is about three orders of magnitude smaller than mobile-phone memory and five to six orders smaller than cloud-GPU memory.This gap makes off-the-shelf DNN models impossible to run on many IoT devices.

A. UAVs Integration into IoT Networks

The survey describes UAVs as flexible components of non-terrestrial IoT networks for data collection, relaying, localization, and wireless power transfer. UAV anchors can improve RSS-based localization, while their deployment must account for altitude, spacing, energy, and connectivity constraints.

  • UAV Roles: UAVs collect data from low-power IoT devices and can relay real-time communication between IoT nodes and base stations.They are useful when data is delay-tolerant or permanent ground infrastructure is costly.
  • UAV Optimization: A double deep Q-network maximizes collected data under flying-time and obstacle-avoidance constraints, while deep reinforcement learning jointly optimizes uplink power and channel allocation for energy efficiency.These examples connect UAV trajectory planning and IoT resource allocation.
  • Localization: RSS-based localization avoids the time synchronization required by time-based methods but estimates distance through an RSS-distance function derived from path-loss models.Multilateration requires at least three distance estimates from distinct anchors.
  • Localization Limitations: Ground-anchor shadowing causes large distance-estimation errors and limits RSS-based localization accuracy.The received signal power varies substantially around its mean on ground-to-ground links.
  • UAV Localization: Higher-altitude ground-to-air links exhibit less shadowing, motivating UAVs as aerial anchors for ground IoT localization.UAV anchors combine improved line-of-sight probability with flexible aerial deployment.
  • UAV Localization: Urban localization accuracy depends strongly on UAV altitude, with an optimum altitude reported for the highest accuracy.The cited study also reports optimal inter-UAV distances and sharply lower localization error as the number of UAVs increases.
  • UAV Energy: UAV networks are constrained by mechanical energy consumption, which strongly influences trajectory design and network lifetime.This makes UAV energy modeling important for quality-of-service analysis.
  • Wireless Power Transfer: UAVs can support wireless power transfer and connectivity in difficult environments because their mobility brings them close to IoT nodes.Trajectory, power control, and scheduling are jointly studied for data and energy transfer.

A. Limited Battery Lifetime

The survey identifies battery lifetime, reliability, scalability, and latency as interconnected challenges for IoT access. It reviews grant-free access, wake-up receivers, energy harvesting, mmWave design, and cooperative transmission as routes toward more energy-efficient networks.

  • Battery Lifetime: NB-IoT improves expected battery lifetime over LTE, but synchronization, connection establishment, scheduling, and control-signal waiting leave a gap to 5G expectations.These procedures remain sources of energy consumption in NB-IoT standardization.
  • Grant-Free Access: Grant-free access transmits packets without handshaking, resource reservation, or authentication, reducing signaling for sporadic IoT traffic.It is already used by unlicensed-spectrum technologies such as SigFox and LoRaWAN.
  • Grant-Free Access: Grant-free access requires reliability, coexistence, scalability, and device-operation analysis because shared resources create collisions, interference, and limited control.The survey highlights these as open research problems for licensed and unlicensed IoT deployments.
  • Grant-Free Access: A scalability model is needed to relate provisioned radio resources, access-point density, device energy per packet, and required quality of service.This relation is identified as an active study item in 3GPP standardization.
  • Wake-Up Receivers: Wake-up receivers can awaken battery-driven devices on demand and may harvest ambient energy, but their circuits must balance very low power with false- and missed-alarm reliability.The receiver continuously monitors a channel for its authenticated address.
  • Energy Harvesting: Energy harvesting is promising for remote IoT devices where battery replacement is difficult or costly, including devices embedded inside aircraft.The motivation is to avoid human intervention for inaccessible devices.
  • mmWave Access: Future mmWave IoT access must address waveform and multiple-access design, hardware differences, beam-selection overhead, and the trade-off between narrow-beam gain and rapid selection.Joint beam design and selection should account for network and traffic dynamics.
  • Cooperative Transmission: Distributed cooperative transmission can use relay selection to improve energy efficiency and reduce interference, but multi-hop relaying adds processing latency at relay nodes.The survey notes LDPC as the 5G NR data-channel code in this context.

D. Massive Connectivity

The survey examines connectivity challenges and enabling techniques for massive IoT, spanning access, coding, latency, reliability, security, and learning. It also identifies practical limitations and future directions for cellular IoT networks.

  • NOMA: Downlink NOMA is unlikely for massive IoT because low-cost devices cannot afford global perfect CSI and practical SIC complexity and errors.Uplink NOMA is considered more promising if code-domain encoding complexity is reduced.
  • Random access: Grant-free massive IoT creates high base-station computational complexity and requires device-activity detection, motivating approximate message passing and deep learning.Deep-learning solutions are expected to reduce complexity and jointly improve device detection and channel estimation.
  • Channel coding: Massive IoT requires channel coding that supports high reliability, low latency, extreme energy efficiency, and diverse application requirements.The survey frames channel coding as central to meeting these varied requirements.
  • Channel coding: Short-block-length polar and LDPC codes lack bit-level granularity, while puncturing and extending can degrade performance.These limitations motivate new rate-compatible coding designs for short packets.
  • Channel coding: Very low-rate, low-SNR IoT links require codes that operate efficiently with unreliable or unavailable CSI.The requirement is especially relevant to extreme-low-power devices.
  • Risk-aware learning: Mission-critical IoT motivates risk-aware network design that accounts for rare events with potentially huge losses rather than optimizing average utility alone.Risk-aware scheduling can use nonlinear utility functions such as exponential latency utility.
  • Risk-aware learning: Conditional value-at-risk has been proposed to minimize age-of-information costs while accounting for rare events in real-time IoT reporting.This extends risk-aware learning to freshness-sensitive monitoring applications.
  • Survey scope: The survey synthesizes advances in massive access, energy efficiency, reliability, latency, security, deep learning, and federated learning for cellular IoT.It also identifies steps needed to bridge remaining gaps toward beyond-5G IoT in the 2030s.

XI. APPENDIX A: LIST OF ABBREVIATION

The appendix provides a glossary of abbreviations used throughout the survey, covering wireless generations, network technologies, learning methods, coding, access, security, and performance terms.

  • Generations and networks: The glossary expands cellular generations and network categories, including 5G, 6G, B5G, NTN, LTE, LTE-A, and LPWA.It also includes 3D, 3GPP, and related coverage or access terminology.
  • Coding and decoding: The appendix defines coding and decoding terms such as LDPC, BCH, TBCC, OSD, BP, SC, and HARQ.These cover channel codes, decoding procedures, and retransmission mechanisms.
  • Security and performance: Security, radio, protocol, and performance terms include IDS, PUF, TLS, CSI, CQI, SNR, PER, BLER, KPI, MIMO, and OFDM.The list also covers device, spectrum, synchronization, and physical-layer terminology.
  • Access methods: Wireless access and multiple-access abbreviations include NOMA, SCMA, IDMA, IGMA, PDMA, RDMA, RSMA, RACH, and GFA.The list spans orthogonal, non-orthogonal, coded, pattern-based, and random-access methods.
  • Learning methods: Learning and intelligence abbreviations include AI, ML, DL, DNN, CNN, RNN, GAN, RL, DRL, SVM, RBM, and ReLU.The glossary also includes autoencoders and federated-learning-related terminology through the survey’s broader vocabulary.
Loading 2107.03059v1…