System Intelligence for UAV-Based Mission Critical with Challenging 5G/B5G Connectivity
Cristiano Bonato Both, João Borges, Luan Gonçalves, Cleverson Nahum, Ciro Macedo, Aldebaro Klautau, Kleber Cardoso
SSystem Intelligence for UAV-Based Mission Critical withChallenging 5G/B5G Connectivity
Cristiano Bonato Both a , João Borges b , Luan Gonçalves b , Cleverson Nahum b ,Ciro Macedo c , Aldebaro Klautau b , Kleber Cardoso c a University of Vale do Rio dos Sinos - UNISINOS - São Leopoldo, Brazil b Federal University of Pará - UFPA - Belém, Brazil c Federal University of Goiás - UFG - Goiânia, Brazil
Abstract
Unmanned aerial vehicles (UAVs) and communication systems are fundamentalelements in Mission Critical services, such as search and rescue. In this article, weintroduce an architecture for managing and orchestrating 5G and beyond networksthat operate over a heterogeneous infrastructure with UAVs’ aid. UAVs are usedfor collecting and processing data, as well as improving communications. Theproposed System Intelligence (SI) architecture was designed to comply with recentstandardization works, especially the ETSI Experiential Networked Intelligencespecifications. Another contribution of this article is an evaluation using a testbedbased on a virtualized non-standalone 5G core and a 4G Radio Access Network(RAN) implemented with open-source software. The experimental results indicate,for instance, that SI can substantially improve the latency of UAV-based servicesby splitting deep neural networks between UAV and edge or cloud equipment.Other experiments explore the slicing of RAN resources and efficient placementof virtual network functions to assess the benefits of incorporating intelligence inUAV-based mission critical services.
Keywords:
Artificial intelligence, mission critical services, UAVs, beyond 5G
1. Introduction
Annually, the economic losses from global disasters sum hundreds of billionsof dollars and reap thousands of human lives. The impact on wildlife has more ∗ Corresponding author
Email address: [email protected] (Cristiano Bonato Both) https://ourworldindata.org/grapher/economic-damage-from-natural-disasters Preprint submitted to Future Generation Computer Systems February 5, 2021 a r X i v : . [ c s . N I] F e b iffuse numbers [1], but all of them expose a dramatic situation. Communicationsystems are fundamental tools in these scenarios, since they provide the meansfor proper coordination among the several teams involved in diverse tasks, fromsaving lives to repairing infrastructures. More recently, Unmanned Aerial Vehi-cles (UAVs) have been added as another valuable tools in the context of MissionCritical (MC) services, offering help on localizing people and animals, transport-ing medicines, improving communications, among other relevant tasks [2]. Forexample, Search and Rescue (SAR) missions are services usually performed in re-mote areas where the telecommunication infrastructure is inoperative or severelydamaged due to disasters, and UAVs can be extremely helpful. However, the de-ployment and operation of UAVs in the Mission Critical (MC) context still dependsstrongly in human intervention, what limits its applicability and efficiency. Similarproblem is observed with communication systems, mainly involving data transport,which is becoming the most relevant even in disasters scenarios. Considering theseproblems, standardization and adoption of Artificial Intelligence/Machine Learn-ing (AI/ML) solutions are promising approach to tackle these issues [3].In general, the investments in MC scenarios’ resources aim at preventing losses,not in promoting profit. Therefore, it is essential to increase scale and reduce costsin any MC-related solution. The adoption of worldwide standards is a promisingapproach to achieve this goal. Additionally, international cooperation in MC sce-narios is quite common and standards become again a natural choice. Nowadays,the 5th Generation (5G) and its evolution, i.e., Beyond 5G (B5G), are a key set ofstandards in the general context and also in the MC scenarios. MC communicationshave been a concern in the standardization bodies since long time, however, the fo-cus has been in a specific set of MC services considered critical to teams involvedin MC scenario. Additionally, the support for UAV is very recent, only consid-ered effectively in the last 3rd Generation Partnership Project (3GPP) Releases, 15[4] and 16 [5]. Nevertheless, these standards assume UAVs acting mainly as UserEquipment (UE) and commanded by human beings. While fleets of autonomousUAVs, able to deploy and operate fully functional networks and services, have beeninvestigated in the literature [6, 7, 8], this context is still considered futuristic andout of scope for the standardization bodies in the near future. In this article, weargue that AI/ML can turn this vision into reality sooner than is being expected.AI/ML is already being widely adopted in UAV for autonomous operation[9]. Moreover, the standardization bodies of communication systems are in an ad-vanced stage of defining how to adopt and deploy AI/ML [10]. However, the mainefforts are focused on traditional infrastructure, i.e., without considering UAV aspart of the whole system and acting not only as UE but also as part of the Ra-dio Access Network (RAN) and even the 5G Core (5GC). Furthermore, UAVscan offer edge computing resources, which enables several new applications and2ervices. Finally, UAVs can provide temporary communication systems that arefully functional end-to-end. Software systems such as 5GC and the virtualized anddisaggregated Radio Access Network (RAN) are receiving AI/ML updates to de-crease the need for human intervention. While these updates may be enough fortraditional infrastructure, they cannot deal with the challenging conditions faced byUAVs in MC scenarios. Additionally, each AI/ML solution is limited to its relatedsystem, disregarding the broader context that involves multiple interconnected sys-tems, e.g., 5GC, RAN, and Multi-access Edge Computing (MEC).The contributions of this article are two-fold: (i) we introduce the System In-telligence (SI) as an architecture for managing and orchestrating 5G/B5G commu-nication systems that operate over a heterogeneous infrastructure, which includesUAVs. Besides describing SI and contextualizing it concerning existing standards,(ii) this article also contributes with results obtained with a testbed that supportsexperiments with UAV-based communications and AI in future networks and com-puter systems.The proposed SI must support advanced MC services in several scenarios, in-cluding those in which only UAVs are available. SI focus on keeping all the es-sential systems interacting properly and operating with the best performance possi-ble given the adverse conditions. Moreover, SI was designed following standards,mainly the European Telecommunications Standards Institute (ETSI) ExperientialNetworked Intelligence (ENI) specifications [10]. Regarding the results, one of theexperiments indicates that SI can improve MC services by splitting neural networksbetween UAV and edge equipment, such that the latency decreases by 29.87%. Inanother experiment, it is shown that adequately slicing the resources leads to a de-crease in average latency from 66 ms to 34 ms, approximately, when consideringthe downlink communication between UAV and edge.The remaining of this article is organized as follows. In Section 2, we presentthe standardized approach’s background, considering the main initiatives towardsAI/ML solutions. In Section 3, we introduce our System Intelligence solution, andin Section 4, we show the experiments with AI for UAV-based SAR. We discussthe related work in Section 5 and present final remarks and suggestions for futurework in Section 6.
2. Background
In the last years, telecommunications standardization bodies, such as Inter-national Telecommunication Union (ITU-T), ETSI, 3GPP, as well as alliancesbetween operators and manufacturers as Open Radio Access Network (O-RAN),published specifications about the design, development, and deploy of Artificial3ntelligence (AI)/Machine Learning (ML) for the 5G ecosystem and also its evolu-tion, i.e., B5G. Together, these specifications encompass a wide scope, includingthe core, MEC, and RAN. Systems using AI in 5G/B5G networks will surely bebased on these specifications. However, most AI systems used in 5G are not com-pliant to standards at the current maturity stage yet. Moreover, before AI systemsbecome widespread, grasping all related specifications may be a daunting task.The specifications overlap and have gaps concerning some issues. Initially, thissection briefly introduces the architectures defined by these standards, which cover5GC, RAN, and MEC, and also the integration between them. After, the sectionpresents a summary of the main initiatives considering the AI/ML’s perspective forthe 5G/B5G ecosystem. Besides providing a short review of these specificationsfor the reader’s convenience, it is also vital to contextualize the proposed SI andindicate what can be used from the standards and what was missing.The network system is mostly standardized by the 3GPP 5G, composed of the5GC built as a Service-Based Architecture (SBA) [4] and the Next-Generation Ra-dio Access Network (NG-RAN). An SBA allows flexible and stateless positioningof virtual environments in the network segments that make up a 5G system. More-over, this architecture refers to how virtual networks’ functions are created anddeployed flexibly, widely using the concept of cloud computing to develop, de-ploy, and manage services. In this context, Releases 15 [4] and 16 [5] introduceseveral features to 5GC that are useful in the context of MC systems. For example,services can be implemented to expand the capabilities of a MC system, decom-posing functions with low granularity, making the service light, and having a highcapacity for sharing.When contrasted with 3GPP specifications, the O-RAN Alliance introduceda complementary set of NG-RAN standards, gaining relevant support from thetelecommunications industry. O-RAN addresses the split of the 5G base-station(gNB) in three parts: (i) O-RAN Radio Unit (O-RU), (ii) O-RAN DistributedUnit (O-DU), and (iii) Central Unit (CU). These splits for radio access tech-nologies can be designed, developed, and deployed for the sake of saving costsor dealing with restrictions in energy consumption, such as in UAVs networks. Inthis context, 3GPP and O-RAN define a disaggregated RAN, composed of mul-tiple Virtualized Network Functions (VNF). Therefore, we use the term virtualRAN (vRAN) throughout the article to emphasize that both NG-RAN and 5GCcan be implemented as a collection of VNF that must be appropriately placed andchained to accomplish their tasks [11].Given the importance of edge computing in MC scenarios, ETSI efforts in thisarea are briefly described. ETSI specified a MEC system as software-only enti-ties to operate on top of a network edge’s virtualization infrastructure [12]. Thissystem consists of a platform to run applications on a particular virtualization in-4 ore Network (CN)MECEI
MC Services
Cloud
MECMC Services
Access Network (AN) MC U/Na vRANNSSF NWDAFUDM AUSF UDRNEF AMFNRF SMF UPFPCF
MC ServicesMEC Orchestrator
MECSystem
Naf V i r t ua li z a t i on I n f r a s t r u c t u r e Service MECPlatform
MEC PlatformManager
N6N3 N2 N1 gNB
MC Services
Near-RT RIC
O-RAN Network Functions
O-Cloud
O-RU
Service Management and Orchestration Framework
Non-RT RIC vRAN N3 N2
Figure 1: A standardized approach for supporting MC services based on the integration of 3GPP 5Gsystem (mainly core), O-RAN, and ETSI MEC system. The AI/ML components already defined inthe standards are highlighted in blue. frastructure, a platform manager to handle the specific functionality’s management,and an orchestrator, which controls the whole system and services. Moreover, theMEC system is aligned with the SBA principles, and it was designed to be tightlyintegrated with 5GC [13]. Some of the essential MEC services depend on this inte-gration, with Application Programming Interfaces (APIs) such as Radio NetworkInformation API, Location API, UE Identity API, and Bandwidth ManagementAPI.Figure 1 illustrates how MC services can be supported by a standardized setof systems composed of a 5GC, vRAN according to the O-RAN architecture, andan ETSI MEC system. All these systems can be seen as a large and complexcollection of VNFs that must adequately be managed and orchestrated to supportusers’ application and services. The MC services add the challenge of managingand orchestrating these VNFs over an infrastructure composed mainly of UAVs andthat can constantly change. Additionally, the figure illustrates how interconnectedare multiple systems. In fact, Service-Based Interface (SBI) Network applicationfunction (Naf), and the reference points N1, N2, N3, and N6 are only some ofthe interconnections among the 5GC, vRAN, and MEC system. Finally, the figurehighlights (in blue) the components related to AI/ML already introduced by thestandards.UAVs have been used to aid military and rescue operations under challenging5reas [14, 15], performing several tasks, such as surveillance, inspection, and map-ping [16]. The presence of UAVs in these scenarios tends to increase, with thembeing used for even more applications, including disaster management [17, 18],performing MC tasks, leveraging the capacity of the new generation of mobile net-works. However, it is good to note that MC services provided in disaster scenariosusually need to deal with reduced or even lack of radio coverage [19]. For instance,recent hurricanes occurrences, such as Hurricane Maria, resulted in the degradationof the functionality of about 95.6% and 76.6% of cellular sites in the affected areas[20]. In this context, the rescue team needs to ensure that its UAVs will receive theradio resource allocation dynamically during the mission through a network slice,guaranteeing the necessary throughput and latency, for example, video streamingand remote control.In MC scenarios, 5GC, vRAN, and MEC must be improved to be operationalunder the eventual harsh conditions imposed by the environment. In other words,MC is challenging enough to require functionalities that are not still incorporated incurrent specifications. When seeking a solution that supports MC, part of this solu-tion can be obtained from 3GPP specifications that describe the 5GC architecture.For example, when considering the AI required by MC systems, the componentcalled Network Data Analytics Function (NWDAF) is the approach of 3GPP formeeting the AI perspective in the 5G system written in Release 15 [4] and 16 [5].This component is responsible for collecting several types of information from thenetwork and its users. Any core component and external access can consume theservices provided by NWDAF. The analytical data produced by NWDAF can beused by an AI agent that specifies actions in the network context, for example,for UAV-based critical missions. In this case, the 5GC plays proactively and takesreal-time decisions to provide the necessary services to 5G users, and NWDAFbecomes a central point for analytics in the 5G network.Complementing 5GC by 3GPP, the O-RAN architecture defines an AI perspec-tive, including the AI-enabled RAN Intelligent Controller (RIC) for both non-Real-Time (non-RT) and near-Real-Time (near-RT) [9]. The non-RT functions includeservice and policy management, higher layer procedure optimization, and modeltraining for the near-RT RAN functionality. The near-RT RIC is fit with radio re-source management and improves operational functions such as seamless handovercontrol, Quality of Service (QoS) management, and connectivity management.The specifications published by 3GPP and O-RAN are fundamentals and ap-plied for using of AI/ML solution on 5GC and vRAN, respectively. In a broaderscope, ITU-T and ETSI published a set of manuscripts to define terminology, re-quirements, functionalities, and other essential concepts for characterizing AI in-tegration into communication networks. We organize the manuscripts from thesetwo standardization bodies considering the AI perspective in Figure 2 and sum-6 rtificial Intelligence
ArchitectureAssurance Level Evaluate QoS Data HandlingOrchestrationITU-3172ITU-3175ITU-3173 ITU-3170 ITU-3174ITU-3111ENI ETSI 007ENI ETSI 005ENI ETSI 004ENI ETSI 003ENI ETSI 002ENI ETSI 006
Figure 2: Artificial intelligence standards for communication networks by ITU and ETSI. marize these specifications in Table 1. ITU-T introduces six main specificationsthat cover AI in a next-generation network. These recommendations focus on sev-eral aspects of the design of an AI framework. For example, ITU-3111 presentsthe network management and orchestration framework, and ITU-3170 shows therequirements for machine learning-based QoS assurance for the network. More-over, ITU-3172 introduces the architectural framework for machine learning, andITU-3173 discusses the framework for evaluating the intelligence levels of net-works. Furthermore, ITU-3174 considers the framework for data handling to en-able machine learning, and ITU-3175 covers the functional architecture of machinelearning-based quality of service assurance in the networks.Figure 2 also identifies some key ETSI documents. ENI ETSI 006 and 007 ad-dress concepts and definitions of categories for AI applications to networks, suchas planning and optimization, service provisioning and assurance, data manage-ment, operator experience, etc. Moreover, ENI ETSI 0055 presents a high-levelarchitecture for experiential networked intelligence [10], as well as additional con-tents showing experimental network requirements (ENI ETSI 002), context-awarepolicy management (ENI ETSI 003), and terminology for experimental networks7 able 1: AI/ML specifications by ITU-T and ENI ETSI
Specification Title
ITU-3111 Network management and orchestration frameworkITU-3170 Requirements for ML-based quality of service assurance forthe IMT-2020 networkITU-3172 Architectural framework for ML in future networks includingIMT-2020ITU-3173 Framework for evaluating intelligence levels of future networksincluding IMT-2020ITU-3174 Framework for data handling to enable ML in future networksincluding IMT-2020ITU-3175 Functional architecture of ML-based QoS assurance forthe IMT-2020 networkENI ETSI 002 ENI requirementsENI ETSI 003 Context-Aware Policy Management Gap AnalysisENI ETSI 004 Terminology for Main Concepts in ENIENI ETSI 005 Experiential Networked Intelligence - System ArchitectureENI ETSI 006 Proof of Concepts FrameworkENI ETSI 007 Definition of Categories for AI Application to Networks(ENI ETSI 004). This architecture is referred to as an Assisted System (AS) com-posed of three classes that represent: (i) no AI-based capabilities, (ii) AI is not inthe control loop, and (iii) AI capabilities in its control loop. Additionally, the ar-chitecture designs an API Broker to serve as a gateway between different systems.Even with such a relatively short review, it is clear that the body of recent workaiming at defining how AI/ML should be used in future networks is considerable.Before some of these recommendations become widely adopted in the industry, itis essential to identify how they can be put together and what remains to be defined.The next section uses some key aspects of these recommendations and describesmissing elements to compose an intelligent system suitable for critical missions.
3. System Intelligence
In this section, we describe our proposal in detail, i.e., SI, and how it can dealwith the challenging scenarios previously introduced. Since adherence to stan-8ards is one of our primary concerns, we built SI mainly in compliance with ETSIENI [21]. However, while ENI is a general-purpose abstract specification, SI is aninstantiation of the ENI ideas, being focused on MC services scenarios and takinginto account a complex infrastructure that includes computing devices, gNBs, and,mainly, UAVs.In Figure 3, we illustrate the introduction of SI and its integration with the stan-dardized systems previously presented. Similar to the ENI specification [21], thesesystems that are managed by or receive recommendations from SI are named
As-sisted Systems . Moreover, as in [21], we recognize that each assisted system maypresent a distinct level in terms of capabilities related to AI/ML-based decision-making. However, we adopt a different terminology to categorize the systems,which is more appropriate to the MC services context, as described in the follow-ing.
MECMC Services SI vRAN SI API Translator
MEC5GC vRANunawarebasic-awareadvanced-awareAI/MLMC Services
Figure 3: System Intelligence integration with the main Assisted Systems of the Mission Criticalscenarios.
We consider the MEC System as
AI/ML unaware , which means that it has noAI/ML-based decision-making capabilities. We categorize 5GC as
AI/ML basic- ware , since it has some AI/ML-based decision-making capabilities, mainly char-acterized by the NWDAF component. Finally, we consider vRAN, as defined byO-RAN, an AI/ML advanced-aware system, which means the existence of sophis-ticated AI/ML-based decision-making capabilities, including an internal AI in thecontrol loop. MC services are provided by multiple AI/ML applications, whichcan be either basic-aware or advanced-aware. From any category, SI only needsa set of APIs that allows it to obtain data from the assisted system (related to itsstate) and to provide commands or recommendations for how to act, in general, toachieve a specific goal. It is important to highlight that no change is necessary forany assisted system to interact with SI appropriately. Actually, not even APIs needany adjustment. As illustrated in Figure 3, a component named
API Translator isresponsible for translating between APIs of SI and APIs of the Assisted Systems,if necessary.
Knowledge Management Context Awareness Cognition Management Situational Awareness Model-Driven Engineering Policy Management
Input ProcessorOutput GeneratorAPI TranslatorAPI Translator
Semantic Bus (Inputs)Semantic Bus (Outputs) SI gNB MEC5GC vRAN gNB
Figure 4: System Intelligence architecture based on ETSI ENI architecture.
The main elements of SI are presented in Figure 4. The
Input Processor ele-ment handles several types of data from different sources using connection pointsbetween the
API Translator and the external systems, e.g., 5GC, MEC, and vRAN.These data are normalized and forwarded to the
Semantic Bus , which connectswith the six internal SI elements [9]. These internal elements generate results from10rocessing and making decisions based on these standardized data. The resultscan be new facts or new hypotheses, which can later be converted into actions (bythe
Output Generator ) to be applied to the systems. In the following, we describeeach of the internal SI elements and introduce their deployment considering MCservices and the infrastructure components, especially UAVs.
Knowledge Management
This component defines formalism for representing information and knowl-edge, enabling the SI to analyze, apply, and validate decision-making processes.Knowledge Management works with data and information, using a knowledgerepresentation that defines mechanisms for the set of entities’ characteristics andbehavior. Moreover, this component enables SI to plan actions and determine con-sequences by AI/ML and reasoning to direct action on the set of entities. In thiscontext, this component handles which context and situation information is appliedto the raw data, transforming it to information and then knowledge.Assuming a UAV-based MC use case, UAVs can provide essential data aboutthe context and situation. For example, object detection with AI/ML algorithmsapplied to videos from searching and rescuing areas can help locating victims.The AI/ML algorithms needed during the search stage may change from thoserequired for diagnostic and rescue. For instance, assume that Deep Neural Net-works (DNNs) are used for object detection [22], but the set of objects needs to bedetected change over time. Therefore, SI needs to have Knowledge Managementto fit different AI/ML algorithms loaded at UAVs for each context and situation.Moreover, UAVs can be positioned for providing connectivity in remote areas, be-ing configured on-the-fly as repeaters, just amplifying the signals, or as a full BaseStation (BS) [23]. In this case, Knowledge Management using context and situ-ation information represented by estimated interference signals and coverage ar-eas, can assist in the decision-making that regards choosing the better connectivitystrategy for each MC service.
Context Awareness
The Context Awareness component enables various data and information to beeasily correlated and integrated with the other SI components. In this case, thiscomponent allows SI to provide customized services and resources correspond-ing to that context. This component’s fundamental characteristic is to enable SIto adapt its behavior according to changes in the context. For example, the con-textual history may be useful for driving policy decisions for current and futureinteractions. Moreover, context knowledge offers a greater level of reliability andusefulness over the whole systems. 11I must store data regarding the network, services, and users, such as the num-ber of users and the bandwidth requested along with an MC. For example, it ispossible to ensure that applications that benefit from low latency are prioritizedover others with less importance. The data considered by Context Awareness ismandatory for instance, in SAR scenario that deals with the transmission of Con-trol Signal (CS) to inform UAVs about the flight plan for an MC [24, 7] task, suchas person identification or supply delivery.
Cognition Management
This management allows SI to understand data and information input in the sys-tem, i.e., defining how these were produced. In this case, the component providesfour functions: (i) perform inferences to generate new knowledge, (ii) change ex-isting knowledge, (iii) use raw data and historical data to learn what is happening ina distinct context and situation, e.g., why the data were generated, and which com-ponents could be affected, and (iv) determine new actions to guarantee the goalsof SI. Finally, the Cognition Management component uses these four functions tovalidate and generate new knowledge.An MC service usually operates in dynamic scenarios regarding computation-communication trade-offs between the UAVs and network infrastructure. In thiscontext, SI could provide Cognition Management to choose, change, and deployplacement strategies for the computational vision and VNFs, evaluating their ef-fects on CPU and RAM usage, Moreover, all the changes over the configurationof systems and flight plans must be handled through their history to generate newknowledge.
Situational Awareness
This component enables SI to understand what happened and how it influencesthe SI goals. This component works to know how and why the current situationpresents such results. SI observes various situations evolving, examining them forpatterns within each condition and between different cases. This observation pro-cess includes five actions: (i) collecting data, (ii) understanding the significance ofthis data, (iii) determining what to perform in response to a given event, (iv) makinga decision, and (v) evaluating these actions. In this way, it enables the applicationof context and policies to a distinct situation using inference and historical data tolearn what is happening in one specific context, why, and what should be done inresponse to it. It is essential to highlight the difference between Context Aware-ness and Situational Awareness. The first one describes the state and environmentin which an entity exists. The second one incorporates contextual information andother inputs to understand the meaning of data and behavior of the entire assistedsystem and its operational environment.12ne of UAV’s essential applications in MC service is to provide a quick evalu-ation of the situation in SAR areas. UAVs promote versatility, fast response time,and capacity to support several services. Therefore, all this requires ability tolearn about various evolving situations and examine patterns within each condi-tion among different MC services. In this case, these services can allow Situa-tional Awareness to keep predicting the situation’s progression and how it affectsthe goals to achieve the MC task.
Model-Driven Engineering
Model-Driven Engineering is an approach to software development where mod-els are used to understand, design, implement, deploy, operate, maintain, and mod-ify software systems. This component focuses on business logic, using an abstractmethodology. Therefore, this model supports three essential purposes: (i) to ensurethat several data models used in SI maintain a consistent definition and understand-ing of concepts, (ii) to enable different policies at various levels of abstraction tocommunicate with each other using a common vocabulary and data dictionary, and(iii) to develop from a specification of policy to its implementation.The integration of several standards for supporting MC services, such as 3GPP5GC, O-RAN, and ETSI MEC, demands software development models to provideabstraction levels and a shared vocabulary between external systems and SI. It isessential to define data models and APIs for supporting this integration among thestandards. Moreover, another crucial feature of model-driven is to design differentAccess Point Name (APN) to direct users to various service networks, for example,to guarantee the connectivity of UAVs, vRAN, MEC, 5GC, and SI, to offer thebetter MC service.
Policy Management
Policy Management provides uniform and intuitive mechanisms for providingconsistent recommendations and commands to ensure the scalable decisions direct-ing SI behavior. Three types of policies can be used in SI. The Imperative Policyuses statements to change the state of a set of targeted objects explicitly. Declara-tive Policy works based on ideas to describe what needs to be done without defininghow to execute this task. The Intent Policy applies statements from a restricted nat-ural language to express the policy’s goals, but not how to accomplish these goals.SI can handle any combination of these policies to define recommendations andcommands to support and manage the system.An MC service is fundamental for controlling the network’s behavior, apply-ing security and control rules related to UAV’s session management, mainly forfunctionalities associated with vRAN. This component should provide a mobil-ity policy to add the control of access restrictions to MC services in a given area.13oreover, the policy should include the management of topics associated with pri-ority access to the channel of given UAV to others detriment. Furthermore, thismanagement should provide metrics related to QoS and information regarding thedata flow, which is obtained by regularly monitoring events, for example, consid-ering the transmission bit rate on the partitioning strategy for the computationalvision among UAVs, vRAN, MEC, and 5GC.The next section presents experiments that aim at assessing specific aspects ofusing SI in MC scenarios. The description of a SI system working in a closed-loop,retrieving data, and imposing actions are out of this article’s scope. However, thefollowing section presents evidence using an actual testbed on how MC servicesperformance can be improved via SI.
4. Experiments with AI for UAV-Based SAR
This section describes SAR experiments as a use case to make more concretethe discussion about the benefits of having SI in MC services. The goal of theexperiments is twofold: (i) provide concrete examples of how SI orchestratingcommunications and AI/ML applications can positively impact MC services and(ii) demonstrate how open-source software and low-cost off-the-shelf equipmentcan be put together to assess essential issues related to AI/ML in 5G/B5G MCscenarios.The 5G/B5G network is where SI can orchestrate tasks such as the placementof VNFs in different network elements. UAVs, for instance, can behave as UEsor BSs, under the control of SI. Regarding AI/ML, UAVs fly around a disasterarea with equipment for performing computational vision (more specifically, objectdetection via DNNs). SI has the flexibility of distributing AI/ML processing amongequipment in the cloud, edge, and UAVs themselves. This flexibility allows SIto trade-off energy consumption and latency, for instance. Motivated by a SARsituation in which the objects to be detected change along with the mission, it isassumed that SI can partition (split) the layers of a DNN into subsets and allocatedistinct equipment to execute each subset of layers. For instance, the first layers ofa convolutional DNN, which detect simpler features, can be kept fixed and executedby UAVs (acting as a UE). The last layers, which are specialized to the objects ofinterest, can be implemented by MEC (in vRAN or in the core) and continuouslyadapted according to the time evolution of the SAR mission.The communication network used in the experiments is implemented in a testbedmade available to promote reproducible results. This testbed is based on the Ope- https://github.com/lasseufpa/connected-ai-testbed Table 2: Testbed functionalities with associated hardware and software
Function CPU RAM Software
UPF, AMF, HSS, SMF, PCRF (5GC) i5-7500 8 GB Free5GCRCC (vRAN) i5-7500 8 GB OpenAirInterfaceRRH (vRAN) i5-7500 8 GB OpenAirInterfaceBackhaul i5-7500 8 GB MininetUAV (Jetson Nano) A57 4 GB Pytorch, OpenCVWe devised two sets of experiments. The first concerns evaluating the impactof having SI splitting the execution of a DNN between a UAV (as UE) and MEC(in vRAN, i.e., in edge). The second set of experiments regards vRAN slicingand VNF placement, which we assume to be also orchestrated by SI. In bothcases, SI is not fully implemented in a closed-loop. Instead of automatically actingaccording to its inputs, some configurations regarding AI/ML and VNF placementare manually defined and interpreted as SI actions. This configuration simplifiesthe experiments and their description. A fully working SI is out of this article’s15cope and will be described in future work. We describe the two sets of experimentsand their results in the following.
Factors such as computational resources, energy consumption, and latencymust be considered for real-time image understanding through AI/ML techniqueswhen UAVs capture visual information in MCs. These aspects should be addressedby analyzing the computational cost usage through splitting the processing betweenthe UAV and other processing units. For this set of experiments, an SSD-VGG16DNN was trained to detect four classes: person, car, bicycle, and motorcycle. Theadopted dataset was UFPark, described by Nascimento et al. [25], which consistsof videos recorded in a campus parking lot. When datasets with videos recordedfrom UAVs are available, it will be interesting to compare the performance withmodels trained with videos obtained from fixed cameras, as adopted in this work.The original video resolution was subsampled to 300 ×
300 pixels with threecolor channels, and the DNN input is a frame of this video. The trained DNN had40 layers and . × parameters (weights), each represented with 32 bits.DNN was split into two subsets of layers according to three configurations. Inthe first configuration, UAV executed from the first to the third layer. Therefore,the quantized scores (activation values) of this third layer were sent through thenetwork, and a cloud or edge equipment executed the remaining 37 layers. Theother two configurations adopted splits after the sixth and tenth layers, respectively.Figure 5 depicts the results for these three configurations concerning the averagenumber of Frames per Second (FPS) (bottom) and the latency (top), consideringthe time to execute both subsets of DNN layers and the time to transmit the data inthe uplink from UAV to the edge or cloud equipment.The results in Figure 5 explore three different split configurations in both edgeand cloud scenarios, and identify that, in this case, it is beneficial to have 10 DNNlayers processed by UAV. In contrast, the remaining 30 layers are processed atthe edge or cloud. These scenarios need to assess QoS measures and adapt theapplications running through the transport network indicates that SI can improveMC services by properly orchestrating communication aspects and also the AI/MLapplications. SAR and other MC scenarios require communication networks that maximizethe probability of successful mission completion. As indicated by the previousexperiments, orchestrating AI/ML applications can bring significant advantages inrealistic situations. Therefore, we now illustrate how SI can optimize the networkfor improved robustness to the MC services. In this set of experiments, the SI16 La t en cy ( m s ) EdgeCloud3 6 10
Num ber of layers processed by t he UAV F PS EdgeCloud
Figure 5: Performance of AI/ML application using different splits of a DNN between UAV and aterrestrial equipment at edge or cloud. system is assumed to orchestrate VNFs and manage radio resources to attend twodistinct sets of requirements, imposed by applications running on UAV (as UE)and three traditional UEs (e.g., smartphones), respectively. These four devices areconnected to a BS using Long Term Evolution (LTE) with a bandwidth of 5 MHzfor downlink, which corresponds to a total rate of approximately 18 Mbps to bedistributed among the connected devices.UAV executes an MC service for the SAR mission that requires a constantbit rate of 13 Mbps. The other UEs are requesting 5 Mbps each, for applicationsthat are not critical. We contrast two situations: (a) SI imposes a strategy based onvRAN slicing to protect the UAV communication versus (b) fair scheduling withoutsuch vRAN slicing. Figure 6 shows these two situations via the throughput valuesfor UAV and the three traditional UEs’ accumulated rates. Moreover, the UAVapplication does not reach the required 13 Mbps, considering the scenario withoutvRAN slicing (circle markers). When considering the SI-driven scenario usingvRAN slicing (x markers), the slice exclusively created for UAV can guarantee17
20 40 60 80 100 120
Tim e (seconds) T h r oughpu t ( M b i t s / s e c ) Rate of the UAV with slicingSum of the rate of three UEs without slicingRate of the UAV without slicingSum of the rate of three UEs with slicing
Figure 6: Throughput in RAN slicing experiment. the 13 Mbps target. In this case, the three traditional UEs only have access to theremaining radio resources, which are equally distributed among them.In Figure 7, the curves contrast the results with and without slicing but con-cerning latency between UAV and equipment at the edge or cloud that is servingUAV data. This scenario considers that the same devices of the previous vRANslicing experiment are connected to the network and request the same throughput.The application server and 5GC (with VNFs, such as UPF and AMF) are eitherlocated at the edge or cloud. For emulating the transport network in the cloudscenario, Mininet was used to impose a backhaul topology with a total latency of100 ms. Moreover, the curves in Figure 7 show the effect of both actions: vRANslicing and VNF placement. VNF allocations on edge provide better latency valuesthan allocating at the cloud. Furthermore, vRAN slicing reduces the latency valueexperienced by UAV in both edge and cloud scenarios. Overall, the results indicatethe impact of SI actions in guaranteeing SAR demands related to throughput andlatency. 18
20 40 60 80 100 120
Tim e (seconds) La t en cy ( m s ) Latency from UAV to cloud without slicingLatency from UAV to cloud with slicingLatency from UAV to edge without slicingLatency from UAV to edge with slicing
Figure 7: Latency when using RAN slicing and VNF placement.
5. Related Work Regarding UAV-Based MC
Previously, mainly in Section 2, we reviewed existing academic articles andstandards to contextualize the proposed SI. In this section, we complement theliterature overview by focusing on a non-exhaustive list of recent manuscripts thatdiscuss integrating the main elements of AI/ML-based MC services, eventuallyusing UAVs. As discussed, MC applications (e.g., SAR) must be supported bytelecommunication infrastructures, considering many elements of vRAN, edge,core, and even cloud [26]. In this context, one can observe the great attentionof academia and industry for the utilization of UAVs in SAR [27]. The follow-ing paragraphs provide information on how these elements can be put together andhelp understand the proposed SI’ relations with previous works.Besides the mentioned communication infrastructure elements, many MC ap-plications rely on computational vision [8]. The performance of MC applicationssuch as SAR can be largely improved by considering AI/ML, edge, and multi-UAVtechnologies together [26]. Aligned with this perspective, two recent surveys werepublished [28, 22]. Xu et al. [28] discussed concepts that allow the distinctionbetween Edge Intelligence and Intelligent Edge. For the authors, Edge Intelligence19ocuses on intelligent applications in the edge environment with edge computingassistance and protection of users’ privacy. On the other hand, Intelligent Edgeaims to solve edge computing problems using AI solutions, e.g., resource alloca-tion optimization. Moreover, Wang et al. [22] emphasize that Edge Intelligenceseeks to facilitate Deep Learning (DL) services via edge computing. Furthermore,DL can be integrated into edge computing frameworks to build an Intelligent Edgefor dynamic, adaptive edge maintenance and management. Finally, the authorsdiscuss the challenges regarding Edge intelligence and Intelligent Edge. The mainrequirement is to design a complete system framework covering data acquisition,service deployment, and the placement of AI/ML models considering processingand network resources.Pham et al. [7] introduced an article presenting the research on the integrationof MEC with 5G and beyond. The authors discuss MEC for UAV communicationand the integration between 5GC and vRAN. The authors explain how UAVs canimprove wireless communications, providing cost-effective, fast, flexible, and ef-ficient deployments. Moreover, UAVs can provide on-the-fly communications andestablish Line-of-Sight (LoS) communication links to users in a complementarynetwork for SAR emergencies and disaster reliefs. In this context, Pham et al. [7]describe two typical scenarios integrating UAV and MEC. An application consid-ers UAVs operating as aerial users of the cellular-connected network. In this case,the MEC server-based BS can provide seamless and reliable wireless communica-tions for UAVs to improve computation performance. Another application refersto UAVs working as aerial BSs and equipped with a MEC server. Therefore, MEC-enabled UAV servers give opportunities for mobile users to offload heavy compu-tation tasks. Finally, the authors highlight several challenges integrating UAV withthe MEC system, such as mobility control and trajectory optimization, communi-cation and computation resource optimization, energy-aware resource allocation,and user grouping and UAV association.Specifically, about SAR as an MC task, Queralta et al. [2] show the researchefforts within the AutoSOS project. This project designs an autonomous multi-robot SAR assistance platform using AI models for object detection. The platformoperates in the reconnaissance missions over the sea by executing adaptive DL al-gorithms in UAVs and boats. Moreover, UAVs can autonomously reconfigure theirspatial arrangement to allow multi-hop communication, for example, when a directconnection between a UAV transmitting information and the vessel is unavailable.This reconfiguration uses algorithms for autonomous decision-making at the edgedevices, i.e., UAVs, considering independent task migration and communicationpriority decisions within the multi-UAV system. In this context, the authors dis-cuss the need to integrate a single multi-agent control loop using DL techniques foradvanced vision, communication constraints, spatial awareness, and computation20istribution. Moreover, Queralta et al. highlight that this integration cannot requirehigh computational resources since a UAV’s low energy consumption is needed toincrease its autonomy.
Table 3: Related work involving UAVs assisted by AI/ML and focused on MC services over 5G/B5Ginfrastructures
Multi MCArticle UAV Edge AI 5GC MEC vRAN Cloud Services [28] (cid:88) (cid:88) (cid:88) [22] (cid:88) (cid:88) (cid:88) (cid:88) [7] (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) [2] (cid:88) (cid:88) (cid:88) (cid:88) [29] (cid:88) (cid:88) (cid:88) (cid:88)
Glimpsing a 6th Generation (6G) network, Zhang et al. [29] propose a UAV-to-Everything (U2X) communication framework. The authors argue that UAVs areunsuited for achieving a high data rate by directly connecting the terrestrial cellularnetworks. Therefore, the authors apply three techniques for U2X communications:cooperative sensing and transmission protocol, UAV trajectory design, and radioresource management considering vRANs. Together, they provide a feasible ar-chitecture for UAV sensing utilization in the 6G network. Moreover, the authorsdiscuss the open problems of U2X communications, such as UAV Cooperationwith U2X Communications to reduce the cost of power and spectrum resources, aswell as MEC with U2X Communications to minimize the computation workloadof the base station and to improve QoS.We summarize this short sample of the state-of-the-art in Table 3, which indi-cates the fundamental elements for supporting UAV-based critical missions con-sidered in each article, such as Multi-UAV, edge, AI, MEC, vRAN, cloud, and MCservices. According to our literature review, the manuscripts do not discuss theintegration of all elements that concern the proposed SI. However, the literatureclearly indicates the trend towards an intelligent global system with a cognitivecontrol loop using AI/ML techniques.
6. Conclusion and Future Work
Standardization and wide adoption of AI/ML are key strategies to reach scale,reduce cost, and achieve MC services efficiency that heavily depends on commu-nication and computing systems. UAVs significantly contribute to improving MC21ervices but demand standards and efficient AI/ML for optimized performance. Inthis article, we presented SI, an architecture for providing full AI/ML capabilitiesfor standardized systems supporting MC services. In addition to describing thewhole architecture and its interaction with the assisted system, we also presentedexperiments that illustrate SI benefits. The initial results show the challenges im-posed in some search and rescue scenarios while also offering the potential gainsobtained with the introduction of SI.The described experiments are timely for researching 6G and future generationcomputer systems because the current state-of-the-art in UAV-based has not yetfound the requirements for MC tasks, such as the timeliness of data processing.The present research seeks solutions to how UAV can improve their sensing rangeand how UAVs can become more intelligent through cooperation. Aspects such asdelays need to be taken into account, and simulators are rather limited concerningmimicking all variability found in critical missions.While the experiments of this article exercised some of the essential compo-nents of SI architecture, implementing this software that covers all its componentsis still lacking. Therefore, as future work, we intend to develop and publicly avail-able a functional implementation of SI that minimally illustrates all components.Given the size and complexity of this task, we plan to perform it in phases. Eachone is complemented by additional use cases that illustrate the components in-volved and the benefits obtained. Additionally, we are interested in investigat-ing how to evolve the AI/ML awareness level of assisted systems. For example,the necessary changes to turn MEC at least AI/ML basic-aware and turn 5G coreAI/ML advanced-aware. In this context, the concept of ML pipeline introduced bythe specification ITU-T Y.3172 seems a promising approach for adding awarenessto assisted systems.
Acknowledgment
This article was conducted with partial financial support from the NationalCouncil for Scientific and Technological Development (CNPq) under grant number130555/2019-3.
References [1] A. Iwasaki, T. Noda, A framework for quantifying the relationship betweenintensity and severity of impact of disturbance across types of events andspecies, Scientific Reports 8 (2018) 1–7.222] J. P. Queralta, et al., AutoSOS: Towards Multi-UAV Systems Support-ing Maritime Search and Rescue with Lightweight AI and Edge Comput-ing, CoRR abs/2005.03409 (2020). URL: https://arxiv.org/abs/2005.03409 .[3] L. Bonati, et al., Open, Programmable, and Virtualized 5G Networks: State-of-the-Art and the Road Ahead, Computer Networks 182 (2020) 107516.[4] 3GPP-TR21.915, Technical Specification Group Services and System As-pects; Release 15 Description, Technical Report, 3rd Generation PartnershipProject (3GPP), 2018-12. Version 0.5.0.[5] 3GPP-TR21.916, Technical Specification Group Services and System As-pects; Release 16 Description, Technical Report, 3rd Generation PartnershipProject (3GPP), 2020-09. Version 0.6.0.[6] M. Mozaffari, et al., A Tutorial on UAVs for Wireless Networks: Applica-tions, Challenges, and Open Problems, IEEE Communications Surveys &Tutorials 21 (2019).[7] Q. Pham, et al., A Survey of Multi-Access Edge Computing in 5G and Be-yond: Fundamentals, Technology Integration, and State-of-the-Art, IEEEAccess 8 (2020) 116974–117017.[8] C. Dario, et al., A Survey of Computer Vision Methods for 2D Object Detec-tion from Unmanned Aerial Vehicles, Journal of Imaging 6 (2020) 78–116.[9] R. Shafin, et al., Artificial Intelligence-Enabled Cellular Networks: A CriticalPath to Beyond-5G and 6G, IEEE Wireless Communications 27 (2020) 212–217.[10] Y. Wang, et al., From Design to Practice: ETSI ENI Reference Architectureand Instantiation for Network Management and Orchestration Using Artifi-cial Intelligence, IEEE Communications Standards Magazine 4 (2020) 38–45.[11] A. Ghosh, et al., 5G Evolution: A View on 5G Cellular Technology Beyond3GPP Release 15, IEEE Access 7 (2019) 127639–127651.[12] D. Sabella, et al., ETSI White Paper No. 20 - Developing Software for Multi-Access Edge Computing, Technical Report, European TelecommunicationsStandard Institute (ETSI), 2019. 2313] S. Kekki, et al., ETSI White Paper No. 28 – MEC in 5G networks, TechnicalReport, European Telecommunications Standards Institute (ETSI), 2018.[14] M. Azmat, S. Kummer, Potential applications of unmanned ground and aerialvehicles to mitigate challenges of transport and logistics-related critical suc-cess factors in the humanitarian supply chain, Asian Journal of Sustainabilityand Social Responsibility 5 (2020) 1–22.[15] M. Hossain, et al., Integration of smart watch and geographic informationsystem (GIS) to identify post-earthquake critical rescue area part. i. develop-ment of the system, Progress in Disaster Science 7 (2020) 100116.[16] D. Orfanus, E. P. de Freitas, F. Eliassen, Self-organization as a supportingparadigm for military UAV relay networks, IEEE Communications Letters20 (2016) 804–807.[17] B. Li, Z. Fei, Y. Zhang, UAV communications for 5G and beyond: Recentadvances and future trends, IEEE Internet of Things Journal 6 (2018) 2241–2263.[18] Z. Qadir, et al., Addressing disasters in smart cities through UAVs path plan-ning and 5G communications: A systematic review, Computer Communica-tions (2021).[19] A. A. Alsaeedy, E. K. Chong, 5G and UAVs for Mission-Critical Communi-cations: Swift Network Recovery for Search-and-Rescue Operations, MobileNetworks and Applications 25 (2020) 2063–2081.[20] P. Safety, Homeland Security,“2017 Atlantic hurricane season impact oncommunications,”, Technical Report, Tech. Rep., 2018. https://docs. fcc.gov/public/attachments/DOC-353805A1. pdf, 2018.[21] J. Strassner, et al., ETSI GS ENI 005 V1.1.1 (2019-09) – Experiential Net-worked Intelligence (ENI); System Architecture, Technical Report, EuropeanTelecommunications Standard Institute (ETSI), 2019.[22] X. Wang, et al., Convergence of Edge Computing and Deep Learning: AComprehensive Survey, IEEE Communications Surveys Tutorials 22 (2020)869–904.[23] M. Moradi, et al., Skycore: Moving Core to the Edge For Untethered andReliable UAV-Based LTE Networks, GetMobile: Mobile Computing andCommunications 23 (2019) 24–29.2424] 3GPP-TR22.829, Technical Specification Group Services and System As-pects; Enhancement for Unmanned Aerial Vehicles; Technical Report Stage1 (Release 17), Technical Report, 3rd Generation Partnership Project (3GPP),2019-09.[25] I. Nascimento, et al., Public dataset of parking lot videos for computationalvision applied to surveillance, in: 19th IEEE International Conference OnMachine Learning And Applications (ICMLA) (Accepted for publication),2020.[26] T. Klamt, et al., Flexible Disaster Response of Tomorrow: Final Presenta-tion and Evaluation of the CENTAURO System, IEEE Robotics AutomationMagazine 26 (2019) 59–72.[27] W. Shule, et al., UWB-Based Localization for Multi-UAV Systems and Col-laborative Heterogeneous Multi-Robot Systems, Procedia Computer Science175 (2020) 357–364.[28] D. Xu, et al., Edge Intelligence: Architectures, Challenges, and Applications,2020. https://arxiv.org/abs/2003.12172 .[29] S. Zhang, H. Zhang, L. Song, Beyond D2D: Full Dimension UAV-to-Everything Communications in 6G, 2020. https://arxiv.org/abs/2004.01920https://arxiv.org/abs/2004.01920