Improvements in and relating to a telecommunication network

The AI-DDF dynamically selects the optimal plane for AI data transmission in 3GPP networks based on size, sensitivity, and conditions, addressing inefficiencies in current systems and ensuring timely and efficient data transfer, thus improving AI operations within 3GPP networks.

GB2630440APending Publication Date: 2024-11-27SAMSUNG ELECTRONICS CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
GB2024005119
Authority / Receiving Office
GB · GB
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-11
Filing Date
2024-04-10
Publication Date
2024-11-27

AI Technical Summary

Technical Problem

Current 3GPP networks lack mechanisms for dynamic selection between Control Plane and User Plane for transmitting Artificial Intelligence/Machine Learning data, which can lead to inefficiencies due to size, time sensitivity, and resource constraints, particularly in Federated Learning scenarios where data must be transmitted efficiently to meet strict deadlines.

Method used

The introduction of an Al Data Delivery Function (AI-DDF) that determines the optimal plane (Control Plane or User Plane) for routing AI-related data based on parameters such as data size, time sensitivity, latency, and network conditions, allowing for dynamic and intelligent selection to ensure timely and efficient data transfer.

Benefits of technology

Enables efficient and timely delivery of AI-related data, optimizing resource usage and meeting deadlines by dynamically selecting the appropriate plane for data transmission, thereby enhancing the performance and reliability of AI operations within 3GPP networks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

A method of routing Artificial Intelligence, AI, related data in a telecommunication network, comprising determining if the AI-related data is to be routed via a User Plane, UP, or a Control Plane, CP. The AI data may comprise at least one of model training data, inference data and datasets. Determining if the AI-related data is to be routed via the UP or the CP may take into account at least one AI related parameter. The AI related parameter may be selected from: data size, time sensitivity, latency, data format, AI model functionality, model termination and network condition. The AI-related data may be split into more than one part and each one or more part may be sent on a different plane selected from UP and CP. The method may be performed by a telecommunication system. An AI Data Delivery Function, AI-DDF, may be provided in one or more of a User Equipment, UE, base station, gNB, or User Plane Function, UPF of the telecommunication system. The AI-DDF may be operable to select CP or UP for transfer of AI-related data. The AI-DDF may be operable to adjust a priority of AI-related data.
Need to check novelty before this filing date? Find Prior Art

Description

The present invention relates to a telecommunication network which operates using a Control Plane and a User Plane, whereby different data traffic types are typically sent over one or other planes. The use of Artificial Intelligence / Machine Learning, AIML, introduces new challenges, one of which is the choice of the appropriate plane for communication. A key innovation in Fifth Generation, 5G, system architecture compared to Evolved Packet Core, EPC, is the introduction of Service-Based Architecture (SBA). With SBA, every network function in the 5G Core, 5GC, such as Access and Mobility management Function, AMF, Session Management Function, SMF, Unified Data Management, UDM, Policy Control Function, PCF, etc., can communicate with another network function, NF, via RESTful APIs over an HTTP protocol. A transport protocol (e.g., TCP, UDP, SCTP, etc.) should be utilized to carry HTTP payloads. The SBA simplifies interactions between NFs given that interfaces can be defined at the application layer and they can be easily expandable. For example, if AMF wishes to interact with a newly proposed NF in 5GC, then this can be performed by a simple software update. With SBA, each network function expresses its functionalities through Service-Based Interfaces (SBI), which include a set of services. Each service includes a set of service operations (i.e., a set of RESTful APIs). This way, an NF can be a service producer, and a set of other NFs can be a service consumer. Figure 1 shows a high-level architecture of 5GC where several NFs are connected to one another via a bus. Typically, a Service Communication Proxy (SCP) governs communications between NFs in 5GC. SCP provides several benefits for 5GC networks such as load balancing, routing, message periodization, overload control, etc. Figure 2 illustrates a non-roaming 5G system (5GS) architecture with some key NFs interacting with one another. As is clear, AMF is well connected to other NFs, given that it is an anchor point for relaying messages from Next Generation Radio Access Network, NG-RAN, and User Equipment, UE, over N2 and N1 reference points, respectively. Similarly, SMF is also connected to several other NFs because it manages Protocol Data Unit, PDU, sessions. Federated Learning (FL) is an important machine learning technique that allows a set of participants (UEs) to engage in distributed model training without exposing their own parameters (data) other than a set of weights to outside entities. It is predicted that FL will be significantly used in several Fifth and Sixth Generation, 5G / 6G use cases and thus, it can generate a significant amount of data traffic on 5G / 6G networks. Therefore, it is important that it is handled gracefully within 3GPP networks. Figure 3 shows key interactions between the main entities involved in a distributed FL model training session. Participants can be a car, robot, smartphone, and drone and the central FL server may be located in a 5G / 6G cloud within or outside the 3GPP network. It is worth highlighting that a central FL server may be located in other locations such as RAN, 5GC and / or UE, especially in a hierarchical FL model. A general working principle of FL model training is as follows: • Participants advertise their willingness to be part of a training session. • The FL server selects a set of participants to be part of the next training cycle based on a set of criteria that may be dynamically changed between training cycles. • The FL server then distributed the latest trained global model to selected participants. • Each participant then starts its local model training when it receives the global model and its related configurations. The local training may use a particular or locally available dataset or real environment parameters. • Each participant then sends the trained model to the FL server once its local training is completed (i.e., when the local model is converged and is stable). • The FL server then aggregates all locally trained models from participants, creating a new global model which can be further distributed between participants in the next round of training sessions. Synchronous Federated Learning (Sync-FL) is another form of FL where participants have a strict deadline for their local model training completion and also uploading the results to the central FL server. If a participant can’t meet the deadline, its results (i.e., trained models) may be unused by the central FL server, wasting 3GPP resources across UE, RAN, and Core Network. On that basis, typically, the central FL server indicates these time constraints to participants so that their compute and network resources can be adjusted accordingly. That said, completing local model training at UE is also computationally intensive and thus requires a significant amount of power which is a crucial resource, especially for mobile devices with limited battery capacity. Table 1 in Figure 4 shows Latency and user experienced Uplink / Downlink, UL / DL, data rates for uncompressed FL. Sync-FL is a crucial example to highlight the importance of minimizing partial or total disturbance of data collection or data transfer for AI / ML operations. Another important scenario could be related to multi-agent, multi-device ML operations where typically a big data processing task (e.g., distributed training) is split amount a set of devices (UEs) by ML agents. Figure 5 shows the achievable learning performance (i.e., Experience of Learning) towards a given task when data collection / transfer is distributed (Dashed line) and not (Solid line). When AI / ML data (training results) is disturbed for any reason (and thus not arriving on time to its destination), it will directly influence the learning performance (the vertical and horizontal arrows between the dotted and solid lines shows this performance gap). There may be several reasons that the AI / ML data (training results, etc.) may be disturbed and thus will not be arrived at its destination on time, for example: • Lack of network resources (e.g., radio resources due to temporal degradation, higher noise / interference level, highly crowded situations, partial / total break-down, and so on) preventing input data from being delivered in time. • Lack of computational resources at UE (e.g., GPU / CPU) at the time of training. • Lack of reach training dataset at UE. • Lack of power at UE during the course of the training Figure 6 shows a scenario where a UE attempts to prevent data disturbance by intelligently scheduling the UL data transmission for a particular set of bits so that the required UL data transmission deadline of one second is met. In Figure 6(a), the UE delivered the concerning data with a size of 3 bits in two seconds, which caused the UL transmission deadline of one second to be missed. Instead, in Figure 6(b), the UE used a higher data rate of 3bps, delivering the data in one second (meeting the UL transmission deadline of one second). In this way, not only can the UE meet its application requirements but also, between time to+1 and to+2, the base station has more resources to grant to other UEs for their UL transmission. In the prior art, the 3GPP network has different planes to carry different data types. For example, for carrying control plane (CP) data which can be transported in a separate path than the user plane (UP) data. The CP data includes all control signalling exchanges between 3GPP entities. The UP data is the user / application data that needs to be exchanged between user device and servers in DN (data network). The CP packets typically have a very small payload (size), while the UP packets may have a large one. As a result, the 3GPP systems have been designed to handle CP and UP packets based on these basic assumptions. In other words, the CP path is currently designed to handle small messages / packets between entities such as UE, RAN and CN. These assumptions have been violated by the AIML data, which needs to be used by the 3GPP entities such as base stations (gNB) or NFs in 5GC to improve their functionality. The AIML data are typically large in size, but at the same time, they may need to be carried via the CP if entities of the 3GPP system use AIML to improve their functionality. That said, the AIML data is not always large in size, and it depends on several factors (use cases, scenarios, etc.). For example, if an AIML model needs to be regularly trained in 5GC and then transported from 5GC to UE, then this type of data transportation is required to carry a large volume of data. Similarly, moving around large training datasets across multiple devices / entities is frequently required, especially for distributed model training (which is a common case in Federated Learning). In other scenarios, the AIML data may be small in size. For example, in the case of model inference or monitoring, where an AIML model has already been deployed, and regular measurement data is needed to be exchanged between entities. Another important aspect to consider when deciding on a plane to use for transporting the AIML data is how time-sensitive the data is. For example, in the case of online training, it may be suitable to use a CP path as it may provide a higher priority data transmission compared to a UP path. This is crucial when AIML-related data (e.g., measurements) is needed for model training in short intervals. Apart from AIML data size and time sensitivity, the location of an AIML model to be deployed and / or to be trained is important to consider. For example, when an AIML model needs to be deployed in a gNB, then the CP path between UE and gNB may be a better transport option compared to the UP path, which may need to pass via the CN (i.e., User Plane Function, UPF). Within prior art 3GPP standards, no mechanisms exist that permit dynamic plane / path selection, e.g., between CP and UP, for transmitting the AIML data. It is an aim of embodiments of the invention to enable entities in the 3GPP system to decide which path (e.g., whether CP or UP or both) to use to transport AIML data (including model training, inference and datasets). It is a further aim to determine that If AIML models can be transferred to UE in multiple ways (e.g., from CP or UP or both) simultaneously how is it decided which plane (CP or UP) to choose. Other problems in the prior art, whether mentioned herein or not, are also addressed by embodiments of the invention. According to the present invention there is provided an apparatus and method as set forth in the appended claims. Other features of the invention will be apparent from the dependent claims, and the description which follows. According to a first aspect of the present invention, there is provided a method of routing Artificial Intelligence, Al, related data in a telecommunication network, comprising determining if the Al-related data is to be routed via a User Plane, UP, or a Control Plane, CP. In an embodiment, the Al data comprises at least one of model training data, inference data and datasets. In an embodiment, the step of determining if the Al-related data is to be routed via the UP or the CP takes into account at least one Al related parameter. In an embodiment, the at least one Al related parameter is selected from: data size, time sensitivity, latency, data format, Al model functionality, model termination and network condition. In an embodiment, the Al-related data is split into more than one part and each one or more part may be sent on a different plane selected from UP and CP. According to a second aspect of the present invention, there is provided a telecommunication system arranged to perform the method of the first aspect. In an embodiment, an Al Data Delivery Function, AI-DDF, is provided in one or more of a User Equipment, UE, base station, gNB, or User Plane Function, UPF. In an embodiment, the AI-DDF is operable to select CP or UP for transfer of Al-related data. In an embodiment, the AI-DDF is operable to adjust a priority of Al-related data. Although a few preferred embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes and modifications might be made without departing from the scope of the invention, as defined in the appended claims. For a better understanding of the invention, and to show how embodiments of the same may be carried into effect, reference will now be made, by way of example only, to the accompanying diagrammatic drawings in which: Figure 1 shows a 5G network with various interfaces, as known in the art; Figure 2 shows Non-Roaming 5G System Architecture in reference point representation as known in the art; Figure 3 shows a representation of Federated Learning as known in the art; Figure 4 shows Table 1, relating to Latency and user experienced UL / DL data rates for uncompressed FL, known in the art; Figure 5 shows Performance gap vs. experience of learning for a given task, as known in the art; Figure 6 shows example of disturbance of data transfer within a preferred deadline of 1 sec (t = t0+1) with the data size of 3 bits, as known in the art; Figure 7 shows AI-DDF components at UE, RAN and gNB, according to an embodiment of the invention; Figure 8 shows how the eAS layer may operate in parallel to the NAS and NGAP layers, according to an embodiment of the invention; Figure 9 shows all possible combinations where the UE, RAN, CN and AF may interact with one another either via UP, CP or a combination according to an embodiment of the invention; Figure 10 shows Table 2, an example of rules that the network (more specifically, PCF) may configure at UE according to an embodiment of the invention; Figure 11 shows the 3GPP system which includes the AI-DDF components at UE, RAN and UPF according to an embodiment of the invention; Figure 12 shows a new layer to converge both UP and CP stacks according to an embodiment of the invention; Figure 13 shows an integration of the AI-LCMF into 5GC according to an embodiment of the invention; and Figure 14 shows a dedicated AF which takes responsibility for AIML LCM for different use cases, or an AF can be used per each AIML use case according to an embodiment of the invention. The AI-DDF (Al Data Delivery Function) is designed to deliver AIML models and / or related data over UP from CN to UE (and vice-versa), but it may also be configured to take responsibility for data collection for AIML purposes. The AI-DDF can be instantiated at UE, gNB and UPF, allowing the AIML data is exchanged between these entities over UP while controlling how AIML data is prioritised according to the AIML data requirements. Figure 7 shows how the AI-DDF may be integrated into the prior art architecture of the 3GPP system. With AI-DDF, any AIML data with potentially any size can be delivered between UE, RAN and CN with no issue. The AI-DDF may also adjust the priority (by changing Quality of Service Flow Indicator, QFI) of AIML data if needed. Evolved Access Stratum (eAS) is new layer within the Control Plane, CP, of the radio protocol stack. The eAS terminates at gNB and supports service-based interface (SBI) that operates over HTTP. This way, applications running at the UE side can use the eAS layer to interact with components at NG-RAN (e.g., RAN Intelligent Control and / or gNB) and also NFs of 5GC / 6GC. This layer may utilize JavaScript Object Notation, JSON, for its interactions which is platform independent data formatting. As highlighted in Figure 8, the eAS layer may operate on top of the RRC layer. This means that the RRC layer may encapsulate the HTTP traffic arriving from the eAS layer at UE and de-capsulate it at the eAS layer of gNB. The eAS layer may operate in parallel to the Non-Access Stratum, NAS, and Next Generation Application Protocol, NGAP, layers. This is important for introducing new technologies within the 3GPP system, allowing incremental deployment and backward compatibility for already existing products. Throughout this application, the following should be noted: • AIML data or data related to a given AI / ML model (or models), for example, data related to AIML model construction, AIML model topology, neural network weights, datasets for training (or datasets for monitoring, inference, and / or other AI / ML procedures) and / or other measurements and statistics for model training, and / or any other data which can be used by AIML components and / or other components handling AIML aspects (or traffic). • AIML data may also be related to the AIML model metadata, which includes information on model performance (e.g., prediction accuracy, entropy, loss function, etc.), model description (e.g. architecture (e.g. number of layers, activation functions, connections between neurons), version history, software, hardware), model training (e.g., data used for the model, training completion time window, etc.). • When it is said that a particular component and / or entity is handling the AIML data, this may include all steps of the AIML Life Cycle Management (LCM). For example, the AIML LCM may include data collection, model monitoring, model training, AIML inference, model update, AIML data transport, AIML model format, etc.). • The embodiments described herein may also be used for non-AIML use cases. • Embodiments herein are applicable to 5G and beyond systems, including 6G, so when it is referred to as 5GC (5G Core), it also applicable to 6GC (6G Core). Similarly, when it is referred to as NG-RAN, it is also referred to as 6G-RAN. Similarly, when it is referred to as UE, it is also referred to as 6G-UE. • The term “packets” may be used interchangeably with “messages”. • The term “plane” may be used interchangeably with “path”. To transmit AIML data between 3GPP entities (including UE, RAN, CN) and to meet their data transmission requirements and constraints, it is reasonable to determine a plane (either UP or CP or both) dynamically (and perhaps intelligently). For example, when UE wishes to send some AIML-related data to gNB, it should dynamically determine whether to use CP (via a Signaling Radio Bearer, SRB) or UP (either via a UPF or directly to a gNB). This determination may be made by considering several AIML-related parameters (e.g., AIML data size, timesensitivity, format, and AIML model functionality, model termination, etc.), plus underlying network and / or radio conditions at UE, gNB and CN. There are several ways in which the UE and gNB can interact with CN (as shown in Figure 9): 1. UE <— CP (SRB) —> gNB <— CP (N2+SBI) —> CN 2. UE <— CP (SRB) —> gNB <— UP (N3)--------> UPF <- (SBI / N4) —> CN 3. UE <— UP (DRB) —> gNB <— CP (N2+SBI) —> CN 4. UE <— UP (DRB) —> gNB <— UP (N3)--------> UPF <- (SBI / N4) —> CN There are several ways in which the UE and gNB can interact with AF (which may be untrusted and thus hosted outside the 3GPP network) and the other way around (as shown in Figure 9): 1. UE <— CP —> gNB <--- CP —> UPF <--- UP —> AF 2. UE <— CP —> gNB <--- UP —> UPF <--- UP —> AF 3. UE <— UP —> gNB <--- UP —> UPF <--- UP —> AF 4. UE <— UP —> gNB <--- UP —> UPF <--- CP —> NEF <--- CP —> AF 5. UE <— CP —> gNB <--- CP —> NEF <--- CP —> AF 6. UE <— UP —> UPF <— CP —> NEF <— CP —> AF 7. UE <— UP —> AF There could be several entities able to dynamically determine how to send AIML data to the next entity (hop): .  UE (*) <- CP / UP -> gNB (*) <- CP / UP -> CN (*) .  UE (*) <- CP / UP -> gNB (*) <- CP / UP -> UPF / NEF (*) <- CP / UP -> AF (*) In one embodiment the UE or network considers the AIML data size, time-sensitivity (or latency requirements), data priority, other data requirements. Here, the network configures the UE (or group of UEs) to select a UP or CP or both for transferring / delivering a model (or models) and / or data related to the model (or models), taking into consideration different data requirements. In one example, these requirements may include data latency, data size, data importance / priority, and / or other requirements, and / or other network defined requirements. In another example, the network determines the model (or models) and / or model (or models) related data transfer / delivery solution, i.e. transfer / deliver over UP or CP or both. In another example, the network configures the UE (or group of UEs) to select a UP or CP or both for transferring / delivering a model (or models) and / or data related to the model (or models), taking into consideration the model (and / or functionality) use case, scenario, configurations, sites, and / or other information related to the model (or functionality). In another example, the network may configure the UE (or group of UEs) to select a UP or CP or both for transferring / delivering a model (or models) and / or data related to the model (or models), based on policy, rules, and / or other information available in the network. For example, information related to congestion levels on the UP or CP, and / or importance of current and predicted (future) non-AI / ML traffic that is using (or will be using) the UP or CP or both. In another example, the UE (or group of UEs) may select a UP or CP or both for transferring / delivering a model (or models) and / or data related to the model (or models), based on UE conditions, for example, UE resources and energy, or other reasons determined by the UE. In another example, the UE may send assistance information, for example, a request (or recommendation, etc.) for the network to use a UP or CP or both in transferring / delivering the model (models) and / or data related to the model (or models) to the UE. In a related example, the network may acknowledge (confirm) or reject the UE request (or recommendation). Optionally, the network may provide a suitable failure (or rejection) cause, for example, ‘UP or CP basedSolutionNotSuppoted’, or any other suitable naming. In another example, the network may send assistance information, the UE may send assistance information, for example, a request (or recommendation, etc.) to the UE to use a UP or CP or both in transferring / delivering the model (models), or part of a model, or a trained model, and / or data related to a given model (or models) to the network. In a related example, the UE may acknowledge (or confirm) or reject the network request (or recommendation, etc.). Optionally, the UE may provide a suitable failure (or rejection) cause value, for example, ‘UP or CP basedSolutionNotSuppoted’, or any other suitable naming. In a related example, the AMF / SMF (or any other network entity or function) can determine whether the AI / ML model and / or data related to the model, may be transferred / delivered over UP or CP or both. Optionally, the AMF / SMF (or any other network entity or function) may inform the RAN (e.g. gNB, NG-RAN) of this decision, for example, (as part of or included in) NG (or S1) messages / signalling between the AMF (or MME) and NG-RAN (or gNB, or eNB). In one embodiment, the data collected at the UE may be provided for at least one network entity (and / or function) performing (or involved in) at least one AI / ML LCM procedure. In another embodiment, the UE can support the capability of (dynamically) selecting a suitable data transfer / delivery solutions, i.e. via CP or UP or both. In one example, this new capability may be part of the UE capability reporting / indication that is exchanged with the network. In another embodiment, the network can support the capability of (dynamically) selecting a suitable data transfer / delivery solutions, i.e. via CP or UP or both. In one example, this new capability may be part of the network capability reporting / indication that is exchanged with the network. In one embodiment, the model (in full or partially) may be delivered / transferred to the UE (or group of UEs) and / or any other network entity (and / or function) using a UP-based solution, while the meta data related (or associated) to this model (or models) is transferred / delivered to the UE (or group of UEs) and / or any other network entity (and / or function) using a CP-based solution. There are several cases where the UE needs to send AIML data to the network (either gNB / NG-RAN, CN, AF and / or DN) given that, typically, all of the data that may be needed for training AIML models may come from UE. In one scenario, the AIML data may be time-sensitive, and it should be delivered to gNB within a certain time window provided that the data should be input into an online AIML model (e.g., in the case of beam management and / or CSI feedback improvement scenarios and / or in the case of some Federated Learning (FL) use cases). In these scenarios, using CP may be a good option for achieving high data transmission priority with low latency, provided that CP between UE and gNB has only one hop. The UE may also determine which SRBs to use depending on the priority (or time sensitivity) of AIML data. That said, this scenario may be reasonable when the AIML data in question is not very large in size. If it is so, then the UE may prefer to use UP instead because using CP is not currently suited for large data transmissions due to some fundamental shortcomings rooted in SRB design, e.g., lack of proper segmentation ability and small RRC buffer size (<64KB). Even if using SRBs does not impose the above limitations, the gNB may still provide high scheduling priority to packets carried over SRBs in both UL and DL directions, and that may cause delaying other important CP packets (e.g., those related to Non-Access Stratum (NAS) signalling). In another scenario, the AIML data may be time-sensitive and has a large size. In such cases, it may be sensible to use a UP path (i.e., a DRB) with high priority. Table 2, shown in Figure 10 highlights a few of these scenarios. In another embodiment the UE considers the AIML data termination points. If the AIML data will be used within gNB, then sending it to gNB directly via UP (or CP) is more efficient than sending it to UPF, which then needs to be bounced back to gNB (from UPF either via N3 or N4). In other words, the termination points of AIML data (more precisely the location of the AIML data consumers) are equally important as other parameters. For example, the decision of using UP or CP may be completely different, if the AIML data is destined for an NF in 5GC and / or an Over the Top, OTT, server in the Data Network, DN. There may be several scenarios where multiple entities within the 3GPP system may need to consume the same AIML data for different purposes. This may also influence the decision of which plane / path to use for transporting the AIML data. If the AIML data is not going to be used by gNB or 5GC (e.g., Operations and Management OAM), then it may be reasonable to transport it via UP directly between UE and its destination (which could be an OTT server or an AF residing outside the 3GPP network). Having said that, there may be several scenarios where AIML data may be collected from UE by an AF residing outside the 3GPP network, and then the collected data is inputted back to the entities of the 3GPP networks via Network Exposure Function, NEF. An example of such a scenario is DCAF (Data Collection Application Function) which is used by Network Data Analytics Function, NWDAF, instances of 5GC. In another embodiment, the [UE considers the underlying network condition to decide a path for AIML data transfer. In one scenario, the UE may collect some data and / or analytics from the network (e.g., from NWDAF) to decide whether to use CP or not for delivering certain AIML data. For example, suppose the CP within CN or between NG-RAN and CN (over N2) is almost past a certain congestion level (according to a predefined policy provided by the network). In that case, the UE may prefer to transfer its AIML data to CN via UP either through AF (e.g., DCAF) or UPF. This scenario may be reasonable when a NF in 5GC is a consumer of the AIML data. For example, this holds for AIML models used at Location Management Function, LMF, for improving UE’s positioning. In the LMF use case, the UE may have several paths to select for transporting the AIML data to LMF (see Figure 9): 1. Via LLP signalling, which piggybacks on NAS messages which need to be initially terminated at AMF of 5GC. In other words, the UE should send AIML data over N1 (NAS signalling) to AMF and then it is relayed form AMF to LMF via the CP of 5GC. 2. Via AF, such as DCAF, where UE sends the AIML data over application signalling initially to AF, and then from AF, it is delivered to LMF via NEF. 3. Via UPF, where UE is sent the AIML data over an existing PDU Session to UPF first, and then from UPF, it is relayed to LMF either directly if UPF supports Service Based Interface (SBI) or via SMF (N4). 4. Via gNB and UPF, where the UE initially send data to gNB over RRC signalling (i.e., over an SRB) but from the gNB to the CN the transmission would not be direct due to congestion on N2. Thus, the gNB is sent AIML data to UPF over N3 first and then UPF relays it to LMF either directly if UPF supports Service Based Interface (SBI) or SMF (via N4). In another embodiment, the UE uses the AI-DDF component to dynamically select a plane for AIML data. In one solution, the network may push some rules to the AI-DDF (Al-Data Delivery Function) component at UE. The AI-DDF components were originally proposed to handle AIML data via UP in the 3GPP system. However, it may also support extra functionality to determine whether to use CP or UP for delivering certain AIML data according to the provided rules by the network. If CP is determined by AI-DDF, then, the AI-DDF may use the eAS layer, which operates on top of the RRC layer, as described previously, or existing Access Stratum which is controlled by the RRC layer. Whether to use eAS or AS, an SRB should be configured, with a certain priority, and used for data transportation between UE and gNB. The network may deliver AIML rules to AI-DDF in different ways. For example, the network can deliver it from PCF to SMF and from SMF to UE via NAS signalling, or it may deliver it from PCF to UPF (if UPF supports SBI) and from UPF, it is relayed directly to the AI-DDF component at UE over UP. Alternatively, the PCF may initially push the AIML data to SMF, and then from SMF, the related PCC rules are created and pushed to other AI-DDFs at UE and UPF accordingly. A note should be taken that the SMF may also push AIML rules to a third-party AF via NEF (Network Exposure Function). This way, the AF could also subscribe to some analytics from 5GC (more specifically from NWDAF instances) and determine whether to use application signalling over UP or 5GC control plane for transporting certain AIML data. Having such flexibility may permit a dedicated AIML AF to be introduced to manage AIML data from a third-party network (outside the 3GPP network). This way, the dedicated AIML AF may collect data across the 3GPP system and store it in a database while permitting 3GPP entities to access the collected data. Figure 11 shows the 3GPP system which includes the AI-DDF components at UE, RAN and UPF. The AI-DDF was originally designed to handle AIML data over UP, embodiments herein provide a new functionality that permits AI-DDF to dynamically select a path / plane for transporting AIML data. The AI-DDF may obtain some analytics from the network for this purpose. In another embodiment, the UE uses a 6G radio protocol stack with a convergence layer which sits above RRC and SDAP. In one solution, the UE may follow a new radio protocol stack which has a new common layer above both CP and UP stacks. This new layer (known as “Packet Convergence Layer”- PCL) can permit data transportation between UE and gNB in a highly flexible manner where both CP and UP can be used at the same time (i.e., AIML data packets can be sent via both CP and UP stacks simultaneously if needed). This layer is common across both UP and CP protocol stacks, and thus, it may reside above RRC and PDCP layers. It requires mechanisms for packet ordering similar to the Packet Data Convergence Protocol, PDCP, layer. This way, packets can be split across both SRBs and DRBs and then reassembled at the PCL. The PCL layer inherits most of PDCP layer's functionality (e.g., packet reordering and in-order delivery, sequence numbering, packet duplication, duplicate discarding, etc.). The PCL may permit the pooling SRBs and DRBs together and prioritises packets (both data and control packets) with a finer control. With current standards, SRB1 has higher priority than SRB2. SRB1 is mostly used for the RRC signalling (e.g., signalling related to handover) while SRB2 is mostly used for NAS signalling (e.g., signalling related to PDU Session establishment). Other SRBs with different priorities are also exist such as SRB4 which is currently used for delivering application related feedback from UE to NG-RAN. The priority of flows within a DRB is dependent on QoS flows (or more specifically by 5QI) of each flow within a PDU Session. There may be several scenarios where using SRBs can assist UP data to be delivered to the network with high priority (e.g., in the case of AIML workflows with deadline in their task completion, such as those commonly used in distributed Federated Learning). Similarly, there may be several scenarios where using DRBs is more appropriate for carrying CP data, such as those related to AIML (e.g., transferring models between 3GPP entities and / or AIML datasets). Figure 12 illustrates the radio protocol stacks at UE and gNB with PCL. The RRC signalling may be used to control the PCL layer. A note should be taken that when NG-RAN uses a Cloud-RAN, C-RAN, deployment, the PCL of gNB-CU-CP and gNB-CU-UP need to connect to each other via the E1 interface. Such separation may introduce some latency if packets are split across both stacks on a per packet basis (or a similar granularity) so extra considerations may be needed to as what types of packet should be split in this way. In other deployments where gNB-CU is not separated for CP and UP, then such considerations do not apply. The PCL layer may also optionally support SBL This way, the PCL layer can bring significant flexibility to the 3GPP system, provided that the PCL at UE can directly interact with components at UE, and similarly, the PCL at gNB can interact with components at NG-RAN. For example, if there is a RAN Intelligent Controller (RIC) layer, then PCL may interact with xApps and rApps of the RIC layer or if NWDAF of 5GC is expanded to NG-RAN for 6G and beyond networks, then the PCL layer may directly interact with NFs of 5GC. Finally, the PCL may support new interactions between UE and CN which are currently either required to be handled via NAS. In another embodiment, thegNB determines how to transport the AIML data to CN and UE. As highlighted in Figure 9, the gNB may also decide to select CP or UP fortransporting AIML data to CN and UE. A similar principle that also applied to the UE case can also be used for the gNB case. For example, the gNB may follow a policy provided by the network (e.g., 5GC) on how to handle AIML data under different network conditions (similar to Table 2, Figure 10). If the AIML data is not destined for an NF in 5GC or if a CP link between gNB and CN (e.g., N2) is congested, then it may be possible for the gNB to send the AIML data via N3 (UP) over an existing PDU Session to UPF. Similarly, if the AIML data is destined for UE, then whether to use an SRB or an DRB requires considering several parameters together (e.g., AIML data size, priority, time-sensitivity, type, format, underlying network and / or radio conditions). Overall, SRBs are more suitable for small high-priority packets while DRBs can be used for large low-priority packets. It is worth highlighting that in the case of UP, the gNB may need to have an established UP connection to UE (this is not possible with current standards) and also be able to use an N3 tunnel (i.e., over an existing PDU Session) for transporting AIML data to a UPF. These functionalities need to be configured during the UE registration procedure and / or PDU Session establishment procedure or later on via a PDU Session Modification procedure (especially when a solution such as AI-DDF is considered). The UE capability information may also need to be indicated to the network during the above procedures, and then appropriate NFs (e.g., AMF, SMF, and UPF) should be selected which support these functionalities. The gNB may also consider other aspects, such as the AIML data consumers. For example, when UPF and LMF are both the consumers of the AIML data, then the gNB may prefer to transport the AIML data to UPF first and ask UPF to further relay it to LMF. Alternatively, the AIML data could also be stored in a centralized location (database) in NG-RAN and / or 5GC and / or a third-party server in a cloud with possibilities of exchanging AIML data between these databases. In another embodiment, 5GC manages AIML activities in 3GPP system. In one solution, a new NF in the 5GC may be introduced to handle AIML Life Cycle Management (LCM) in the 3GPP system (e.g., across UE, NG-RAN, and 5GC). This new NF, AI-LCMF (Al-Life Cycle Management Function), may take responsibility for how AIML models should be trained, monitored, used and updated in different use cases. In another solution, an instance of AI-LCMF may be used per AIML use case. Figure 13 shows an integration of the AI-LCMF into 5GC. In another related solution, the AI-LCMF network function may also optimize AIML-related data collection for various use cases when the data should be consumed by different AIML entities / components simultaneously. This may also include how AIML data should be stored, e.g., in which databases at which location (e.g., UE, RAN, CN and / or 3rd party databases). This is somewhat similar to some of the functionalities of the DCCF (Data Collection and Coordination Function) of 5GC. In another related solution, the AI-LCMF may also take responsibility for how the AIML data should be transported between different entities of the 3GPP system under different network and / or radio conditions. In this solution, the AI-LCMF may receive AIML policies from PCF and then dynamically create required rules to be pushed to UE, RAN, UPF, AF, etc. The AI-LCMF may dynamically update / create new rules and in turn update related entities. In order to update AIML rules dynamically, the AI-LCMF may require to subscribe to some data and network analytics from related NFs of 5GC (e.g., NWDAF, DCCF, SMF, UDM, etc.) or other entities within the 3GPP system. The AI-LCMF may also utilize AIML frameworks for some of these functionalities. In one solution where the AI-DDF component is deployed at UE, gNB and UPF, the AI-LCMF may closely interact with AI-DDF instances and provide them with the required information for how the AIML data should be collected and stored in different use cases and / or how the AIML data should be transported between entities of the 3GPP system. In another embodiment, an AIML AF manages AIML activities / LCM. In one solution, an AF may push a set of AIML-related rules to UPF, gNB and UE (for a UE or group of UEs) via NEF and / or PCF / SMF. In fact, the SMF can be the main component to distribute these new / updated rules to these entities because the SMF is already supporting to distribute the PCC rules to UPF and UE. A note should be taken that the actual AIML-related rules may not be sent from AF to PCF / SMF, but instead, a set of indications might be sent from AF about how the AIML data should be handled at a given time or in near future. It is the PCF / SMF to decide what AIML rules should be added, removed and / or updated at UE, RAN and UPF. In other words, the PCF / SMF may or may not follow the indications received from AF. The AF could directly interact with the UE’s applications, so it may access some information that the network is not aware of. Relaying such information to the network may assist the network to update AIML-related rules accordingly. A note should be taken that the AF may also subscribe to the set of analytics from 5GC / 6GC and / or even UE (if it supports SBA) in order to handle the AIML data for a UE or a group of UE. The AF could be trusted by the 3GPP network or not. In the case of untrusted AF, the NEF can only be used to interact with 5GC / 6GC. Whether the AF is trusted or untrusted may also impact interactions with UE. In another solution, if the AI-LCMF component is used, then the AF can push AIML rules to UE, gNB and UPF via the AI-LCMF component. As discussed earlier, a set of indications may be sent from AF to AI-LCMF about how the AIML data should be handled at a given time instead of the actual AIML rules. Alternatively, AF can interact with PCF / SMF initially, and then the PCF / SMF may interact with AI-LCMF if required. In this solution, it could be the responsibility of AI-LCMF to update AIML rules at UE, gNB and UPF. This way, the AI-LCMF component may also require to subscribe to a set of data and analytics related to a UE or a group UEs from 5GC, RAN and / or UE. In another related solution, if the AI-DDF component is used at UE, RAN (i.e., gNB), and CN (i.e., UPF), then the AF may interact with PCF / SMF first in order to push (add / remove rules or update existing rules) in AI-DDF instances at UE, gNB and UPF. This way, the AI-DDF component becomes responsible for handling AIML data under different network conditions. In another related solution, the AIML AF may also decide whether to use CP of 5GC for delivering and / or collecting the AIML data or directly from UE and / or UPF. Figure 14 shows the AIML AF and components in 5GC that can assist AF to handle the AIML data at UE, RAN, and CN. The figure includes multiple solutions such as AI-DDF, and AI-LCMF. At least some of the example embodiments described herein may be constructed, partially or wholly, using dedicated special-purpose hardware. Terms such as ‘component’, ‘module’ or ‘unit’ used herein may include, but are not limited to, a hardware device, such as circuitry in the form of discrete or integrated components, a Field Programmable Gate Array (FPGA) or Application Specific Integrated Circuit (ASIC), which performs certain tasks or provides the associated functionality. In some embodiments, the described elements may be configured to reside on a tangible, persistent, addressable storage medium and may be configured to execute on one or more processors. These functional elements may in some embodiments include, by way of example, components, such as software components, object-oriented software components, class components and task components, processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables. Although the example embodiments have been described with reference to the components, modules and units discussed herein, such functional elements may be combined into fewer elements or separated into additional elements. Various combinations of optional features have been described herein, and it will be appreciated that described features may be combined in any suitable combination. In particular, the features of any one example embodiment may be combined with features of any other embodiment, as appropriate, except where such combinations are mutually exclusive. Throughout this specification, the term “comprising” or “comprises” means including the component(s) specified but not to the exclusion of the presence of others. Attention is directed to all papers and documents which are filed concurrently with or previous to this specification in connection with this application and which are open to public inspection with this specification, and the contents of all such papers and documents are incorporated herein by reference. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and / or all of the steps of any method or process so disclosed, may be combined in any combination, except combinations where at least some of such features and / or steps are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise. Thus, unless expressly stated otherwise, each feature disclosed is one example only of a generic series of equivalent or similar features. The invention is not restricted to the details of the foregoing embodiment(s). The invention extends to any novel one, or any novel combination, of the features disclosed in this specification (including any accompanying claims, abstract and drawings), or to any novel one, or any novel combination, of the steps of any method or process so disclosed.

Claims

1. A method of routing Artificial Intelligence, Al, related data in a telecommunication network, comprising determining if the Al-related data is to be routed via a User Plane, UP, or a Control Plane, CP.

2. The method of claim 1 wherein the Al data comprises at least one of model training data, inference data and datasets.

3. The method of any preceding claim wherein the step of determining if the Al-related data is to be routed via the UP or the CP takes into account at least one Al related parameter.

4. The method of claim 3 wherein the at least one Al related parameter is selected from: data size, time sensitivity, latency, data format, Al model functionality, model termination and network condition.

5. The method of any preceding claim wherein the Al-related data is split into more than one part and each one or more part may be sent on a different plane selected from UP and CP.

6. A telecommunication system arranged to perform the method of any preceding claim.

7. The telecommunication system of claim 6 wherein an Al Data Delivery Function, Al-DDF, is provided in one or more of a User Equipment, UE, base station, gNB, or User Plane Function, UPF.

8. The telecommunication system of claim 7 wherein the AI-DDF is operable to select CP or UP for transfer of Al-related data.

9. The telecommunication system of claim 7 or 8 wherein the AI-DDF is operable to adjust a priority of Al-related data.

Citation Information

Patent Citations

  • Dynamic radio bearer selection associated with ai / ML operations

    WO2024097324A1

  • Communication method and apparatus

    WO2024139985A1

  • Devices and methods of communication

    WO2024152200A1