Utilization of AI in 5G Edge Networks
Marton Aron Horvath BME Department of Telecommunications and Media
Informatics, Ericsson Budapest, Hungary marton.aron.horvath@ericsson.com
Abstract
As society advances toward the information era, the telecommunications network, a major aspect of this transformation, is undergoing rapid and profound change.
Innovative methods, such as Artificial Intelligence (AI), are necessary to accomplish these transitions. In my proposal, the artificial intelligence is not designed to help the system that hosts it, but rather the terminal node that benefits from it by processing its incoming data stream before it arrives. I constructed an experimental Edge Computing system with a tightly integrated AI solution in order to attain this. The System comprises a traffic generator, an Ericsson UPF application, an artificial intelligence (AI) component, and software serving as the Edge Application. The produced traffic was routed through the UPF before being processed by the AI component and forwarded to the application.
Keywords— 5G, Edge Computing, Mobile Networks, User Plane Function, Artificial Intelligence
I. INTRODUCTION
The fifth generation of the mobile networks brings a large number of new use-case scenarios, each generation of mobile networks improves both in bandwidth and in latency, although previous generation provides sufficient network speed for casual users, the emphasis in 5G is placed on the reduction of latency. The provided latency meets the requirements for time-sensitive applications, such as manufacturing and transportation demands. This new technology is suitable for use in the factorial sector as a wireless communication technology, as it provides reliable, low-latency communication over an established infrastructure. These facts indicate an increasing demand for private 5G networks.
In the case of private networks, computation moves from the central, or from cloud to the network's edge; this is referred to as edge computing. Edge computing is a paradigm of distributed computing that relocates computation and data storage closer to the data sources. This should improve response times and reduce bandwidth usage. Edge computing is an architecture, not a specific technology, and a topology- and location-aware form of distributed computing. Another important characteristic of private 5G networks is that data do not leave the confines of a factory, making communication secure.
Control Plane and User Plane are separated by 5G; the control plane carries signaling traffic, the user plane carries user data (the content of communications), and the control plane carries administrative traffic. User Plane is what we would like to move on-premises, but since it cannot function without Control Plane, it can be placed on-premises or left in the cloud;
there are solutions for both.
To summarize the data flows through User Plane and its alignment always on-site in a scenario of a private 5G network. My intention was to assist in the processing of the data using Artificial Intelligence (AI). In other words, as data flows through the network, it is processed, and the network acts as a "smart wire" that not only transfers but also modifies the data (Figure 1 AI extended flow of the data .
Figure 1 AI extended flow of the data
This solution is most beneficial at the edge, but it is also applicable in a centralized environment. What I would like to provide is a framework for a multitude of new use cases. The solution not only decreases the processing time, but also provides a solution, so the client should not worry about its implementation. In section 2, I outline the possible use cases.
My research was intended to build a solution within an existing system by modifying its existing components;
however, because the 5G network already has its own architecture, I was unable to build a solution from scratch. The goal of my research was to build a solution within an existing system by modifying its existing components. Because of this, It was required to conduct an in-depth analysis of the architecture, look into the features it possesses, and decide where to deploy my deep learning solution while explaining the benefits and drawbacks associated with each potential alternative. In section 3, I will discuss the architecture of the 5G network.
After determining the location of the solution, I was required to design an architecture that was capable of hosting the solution and conduct a Proof of Concept test utilizing that architecture. I have tried the solution in a variety of settings, ranging from a demonstration setting all the way up to a real 5G network. These topics are going to be covered in Section 4.
In the last two Sections, I will summarize what I have accomplished and draw some judgments about what the recently established framework is capable of hosting and what it is not capable of hosting.
II. USE CASE SCENARIOS
Since edge AI in 5G networks allows low-latency applications, it could open up many new use cases. My study's goal was to provide a foundation for several AI applications.
I wanted to develop an architecture that could provide all of these solutions. 5G networks serve manufacturing use cases, therefore they include features 4G doesn't. This paradigm allows me to tackle industrial concerns like long life cycles, brownfield protocols, and others. Some solutions are listed above:
1. Add external sensors for operating safety. Cameras are employed instead of light barriers to recognize people or other obstacles in the wrong place (Figure 2). This artificial intelligence uses machine vision with a single output. Add this bit to Profinet data flow. Thus, a proxy that can combine OT dataflow with external data solves this problem. No OT system change is needed here. Self-driving automobiles and factory AGVs could benefit from that use-case. A simple video stream can detect human workers and avoidable objects with an Object Detection algorithm.
Figure 2 Enchance Functional Safety
2. Using sensor virtualization, we can use cheaper sensors to educate artificial intelligence (AI) to make robots and automated guided vehicles (AGVs) smarter. Instead of a costly LIDAR (the SICK S30A- 4111CL costs 3,400 euros), a few stereo cameras may be used. Reduce AGV costs in this more complicated case. If we don't replace the controllers, we can convert them to camera-to-lidar.
3. Increasing indoor position accuracy. Indoor localization requires 2 cm or less precision, however the typical accuracy is +-5 meters, depending on topography. However, we may meet the 2- centimeter criterion by optically measuring the AGV's position with the CCTV camera. Mobile robots employ other ways, hence that use case doesn't exist (magnetic stripe, QR-code, different SLAM methods, etc.). However, improving precision would make mobile positioning effective.
Companies employ differential GPS to locate autos outdoors within 2 cm. There is no indoor solution.
Machine vision algorithms and video cameras can locate an object in three dimensions. Mathematical algorithms can make this change.
Stereotactic localization using video cameras was tested using a typical Brown-Roberts-Wells (BRW) phantom simulator and BRW angiographic localizer. 1.5 mm localization precision. Machine vision can be used to freehand stereo tactically localize surgical instruments. If the computer is quick enough, these systems can continuously verify instrument locations inside the cranial vault. [1]
Thus, fewer video cameras could yield more exact results than current methods. Thus, combining a few camera feeds with an extension of deep learning can give the application the precision we need.
4. Cellular robot hints (based on NWDAF - network data analytics function). 5G cellular networks have several new features, such as the Network Data Analytics Function (NWDAF), which lets network operators use their own Machine Learning (ML)- based data analytics methods or third-party solutions. [2]
When an Industry 4.0 robot travels and welds, we may warn it of cell handover or network issues. This chained service involves the NWDAF. [3] Our AI module might change industrial protocol-only Operational Technology (OT) into IT with a NWDAF-compliant interface using the REST-only NWDAF. This planned NWDAF use case involves co-located external services.
AI-NWDAF cooperation may benefit both parties.
Both could share data, improving both functions.
The following example shows synergy.
5. Mass IoT data collection.
We could collect data, pre-process it, and transfer it to the industrial system using an AI in the core, even though a Lightweight M2M (LwM2M) protocol does not allow low-power devices to cooperate. [4]
In this scenario, LwM2M is encrypted, making the chained technique impracticable.
Use an independent service.
These were some examples, but several other scenarios are possible also. The architecture that can host various use-cases and provide a framework is most crucial. The background knowledge needed to understand my approach is provided in the following section.
III. THE 5GARCHITECTURE
The 5G architecture is adequate to host an AI solution, Figure 3 displays the essential components of the 5G core network, which is the architecture since 3GPP REL-15. [5] A gateway application is a User Plane Function (UPF). User Equipment (UE), such as a mobile phone, connects to Radio Access Network (RAN), data flows from RAN to UPF, which acts as a router, and finally to Data Network (DN). User Plane connectivity is RAN-DN data flow. 5G Protocol Data Unit (PDU) Sessions are device specific. QoS is achieved in the PDU Session by generating QoS Flows with unique IDs. [6]
Figure 3 UPF in 5G network
The data flows though that path, the procession of the data should be happened on the fly during the data transfer. The components which are not part of the PDU session are part of the controller, they are not able to host a solution. Between the UE and the RAN AI cannot be placed, since its catching and retransmitting radio waves should be avoidable. Data Network is the destination, I would like to transfer data to them in a processed from. Placing new component before or after UPF would come with a huge overhead. Hence, the only available option is placing the AI inside the UPF. The data travels inside it with DPDK, hence it faster than normal transmitting between software components. [7]
In real-time systems, Edge UPF AI implementation may be advantageous. If so, we can help other services quickly evaluate our data and directly affect the mobile network.
Figure 4 Traditional implementation of AI
Consider the next scenario, a 5G private network connected AGV works at a manufacturing plant. The manufacturing site edge cloud is where the robot control program gets and gives commands. In the manufacturing facility, the UPF receives radio signals and sends them to the robot controller. Local clouds host the UPF and robot controller. AGVs have cameras. This mobile network transmitted video stream is also analyzed by artificial intelligence at the edge. The camera must be processed properly for the AGV to operate safely. Traditional robot controller applications require this AI (Figure 4).
Figure 5 Architecture of planned Edge AI solution Figure 5 shows a new design for efficient AI at the network edge. The UPF might route the AGV's control and camera signals to the robot controller. AI, a standalone service, gathers camera signals and provides functionality. The robot controller can find, connect to, and use its API to get the appropriate AGV operation result.
Several AI services may broadcast the same signal stream, each of which wants to process it, or an application may terminate it. For instance, integrating data from numerous cameras to generate a free-viewpoint video initially would allow the AI to identify both the issue and its specific location. The third A.I. translates the English result into Spanish before it can be used on a SCADA system.
Processing entities must chain.
UPF can detect each data stream and route it to the right service. After processing, the UPF might decide which service to chain till the data reaches its destination.
IV. THE SYSTEM ARCHITECTURE
In my proposed architecture, the traffic generated by User Equipments, since solution targets the industrial sector, where these are robots, sensors, and other machinery assist factory processes rather than mobile phones. Let's focus on the robots—one AGV, a fleet of AGVs, or a huge group of robot arms, AGVs, and cobots. Despite the fact that factories have varying levels of automation and the goal is to replace human workers with robots, the solution could be useful in a more automated setting as well. [8]
In Figure 6, Automated Guided Vehicles communicate directly with Radio Access Network (RAN) and subsequently to the User Plane Function (UPF) on a server.
The UPF's control plane, the Controller, runs in containers on another machine. 5G emphasizes Control and User Plane Separation (CUPS). 5G technologies divide the user plane and control plane, allowing a UPF on-prem and the CP in the operator data center. AMF maintains connection while SMF manages session, controlling the UPF. UDM tracks system sim cards, policies, etc., and robots use them. UDR is a vast database. The demo deployment omits the Controller, but that scenario requires it.
RAN sends data to UPF over the N3 interface. Information flow has a session with many parameters. The packet arrives in a ramification, can take either slow path or fast path, slow path investigate the packet fast do not. To claffify traffic, need to study the first packets only. Deep Packet Inspection applies session rules if needed to packets. For example, it may slow down the bandwidth due to permitted but unsupported activity
or discontinue charging for services provided by the mobile network provider without debiting the monthly limit. After the investigation, the packet can take the fast lane without additional examination. If the ruleset defines Service Chaining, it proceeds to that stage.
The Deep Neural Network output is sent to the application since the AI is implemented in the Service Chaining, not the application. When the Deep Learning Solution (DLS) is an object recognition algorithm that uses an AGV's camera output as input, we can deliver the bounding boxes of detected items from the video stream. Maybe the output is chained further, and different Artificial Intelligence services use it as input. The chained service could detect anomalies and avert mishaps, so the traffic flow that arrives in the application carries warning messages and the program can react to them.
Figure 6 Architecure of the real-life scenario
In a real 5G scenario there are more components, Figure 7 shows the design of the 5G core and its settings. The blue components are the ones which were designed or modified, while the others were still deployed and configured. In this case, a live video stream was broadcasted from a laptop's built-in camera. UTP connected the laptop to a Wistron NeWeb 5G modem. The modem, which includes a 5G SIM card, can handle 4G traffic, but I set it to use exclusively 5G.
The modem, DOT4489, is connected to the DOT through coaxial wires. Normally, air would transmit data, but it was used to minimize frequency conflict issues. Instead, 5G setups have four antennae.
Low-power radio transceiver Radio Dot. Radio waves from the Radio Dot cover broadband networks indoors. The Radio Dot Interface receives signal and power from the Indoor Radio Unit (IRU, IRU 88640). (RDI). This RDI can be digital using Ethernet (which I use in my solution) or analog using Intermediate Frequencies (IF), depending on the Radio Dot.
One of an IRU's eight or sixteen Radio Dots was used. Radio Dots come in a range of frequency bands. Tri-Band Dots have three sides that can be handled separately, while Dual-Band Dots have two.
BBUs process baseband signals in telecommunications networks. Baseband is the transmission frequency before modulation. The classic Radio Access Network (RAN) has a baseband unit connected to one or more RRUs near the antenna (s).
RAN comprises the baseband and RF processing units.
The baseband unit—the base station's "hub"—processes uplink and downlink data flow and controls RRU operation.
A BBU's DSP converts analog signals to digital or vice versa.
The central control plane is a cloud-native Packet Core Controller. Access, session, mobility, and gateway controls support 5G use scenarios. The control plane generally runs on the Service Provider's side, Edge UPF, but here was implemented a Local Packet Gateway, which can run on the premises or on the Edge, which reduces latency and protects data since it stays in the factorial area. The solution required UPF port and NIC management adjustments.
5G enables several new edge computing use cases.
Connectivity is key to addressing the edge opportunity, but edge deployments strain infrastructure and applications. The edge user plane is crucial for keeping company data on- premises and reducing latency. The user plane should be small, easy to deploy and manage, and short-lived to be effective. The Local Packet Gateway is ideal for this.
Figure 7 The architecture and parameters of the real 5G core In the Flightt Rack (FR-19) (Figure 8), run the servers which needed to my solution, one of them runs AI container, which functions as an AI as a Service. In that scenario a webcam generates the traffic, an on another laptop displays the live-processed video stream. That laptop can conduct REST queries to set up the AI module and get data.
Figure 8 Flight rack and modules of the real 5G core Existing 5G network components have been slightly modified and incorporated into my solution. My proposed solution intelligently augments the network, allowing me to provide new functionality to devices without modifying the devices themselves.
V. POC SOLUTION AND RESULTS WITH IT
A valid solution was considered to producing for all solutions while creating a Proof-of-Concept (PoC). A solution was provided that legitimized all other solutions, even though the network may set up many scenarios and apply alternative deep learning solutions or chains.
Object Detection is sufficient for Proof of Concept since it poses several challenges that can be solved, such as receiving a video stream, processing it, encoding the processed frames to return them to a video stream, and transmitting the stream forward. To address the aforesaid issues, data processing must be extensive. Despite having a sophisticated model, object detection solutions should be executed rapidly to decrease latency. Catching the video feed promptly is essential.
Video encoding demands the most resources, so the optimal solution is essential. Software has many uses. AGVs have Line Sensors, LIDAR, Microphones, and Cameras. AGVs need cameras, which are affordable. Object detection may help avert accidents. Thus, the robot's driving should be safe in populated areas. The method can also be used with any camera that can send video to a network. The method is versatile.
The first and most important stage is choosing a solution model. First, the definition of object detection, then description the different models, and finally the explanation of the solution technique.
Table 1 lists the fastest and most efficient model parameters.
SSD detects items instantly. Faster R-CNN generates boundary boxes for object classification using a region proposal network. The operation is performed at seven frames per second, the most accurate available. Real-time processing demands less. SSD speeds things up by eliminating the region proposal network. SSD adds multi- scale features and default boxes to compensate for accuracy loss. These advancements allow SSD to match the accuracy of the Faster R-CNN despite using lower-resolution images, enhancing its speed. It outperforms the Faster R-CNN in accuracy and processing speed (Table 1). (Mean average precision (mAP) can be used to assess forecast accuracy.) The SSD300 algorithm was selected after considering these results.
Table 1 Object Detection algorithm comparison
System VOC2007 test (mAP)
FPS Number of Boxed
Input resolution
Faster R-CNN 73.2 7 ~6000 ~1000x600
YOLO 63.4 45 98 448x448
SSD300 77.2 46 8732 300x300
SSD512 79.8 19 24564 512
Training a new model from scratch needs a lot of data, time, and processing power. A pre-trained model was used because the best models are trained on the cloud, which has practically unlimited GPU performance. SSD is the fastest and most accurate algorithm, thus it gets used.
My initial objective was to develop a comprehensive solution including video streaming protocols, buffering, and every possible feature. However, buffering resulted in transmission delays.
Then I opted to simplify the transmission method to eliminate buffering delays since it appeared that normal video streaming solutions were introducing unnecessary delays to the system.
I sent the frames one by one, but a frame is too large to convey in a single packet, any file can be broken into chunks and transferred over the system and rearranged them in another side (Figure 9).
Figure 9 Send frames in chunks and rearrange them
UDP fragments larger packets, which is a concern for us because DPDK does not support fragmentation. Before sending, the frame might be jpeg-encoded. Because UDP does not send packets in order, the system has built to be error tolerant.
That approach eliminated the buffer latency, allowing real- time video stream processing. The terminating node receives the stream instantly. On Ericsson's Innovation Day, the 5G Core, represented by the Flight Rack in Figure 10, processed real-time video. Event photo. The monitor on the right displays the video stream supplied by the monitor on the right to the laptop on the left, while the left monitor displays the slides. Modems, DOTs, and other devices are on the right side of the table.
Figure 10 2022 Ericsson’s Innoday, Cognitive Edge over Local Packet Gateway
With an API the AI module can be set up, as it is necessary to configure a running system on the fly. If we consider further, the API should additionally configure which chains apply to which streams.
VI. CONCLUSIONS
A Proof-of-Concept architecture was constructed that selects packets and runs Deep Learning in the edge 5G system.
After gaining knowledge edge 5G, I'd like to forecast which use cases can be implemented and which cannot (Table 2).
Table 2 Feasibility of the Use-case Scenarios
Enhance
functional safety YES
Though my video streaming solution had high latency, we could just glance at frames and communicate the critical info in plain text.
Virtual sensors YES I developed a similar scenario, and it results as fesible.
Increasing the precision of indoor positioning
MODERATELY
Video processing is more precise than Ultra Sonic systems, but the design must be functional. However, if the AGV moves at 1 m/s, 200 ms lag represents 20cm inaccuracy.
Support cellular hints for mobile
robots YES
A Couple of research described its feasibility.
[7]
Massive gathering of IoT data
YES
DPI can transmit specific streams to additional processing and let the remainder pass through the system for large dataflow.
As a result, I developed a framework that is capable of hosting a variety of applications, it could be a plethora of use- cases. The framework should be able to manage the conversion between old protocols and new cloud applications, notwithstanding the fact that most industries employ brown-field protocols. Infinite use cases are possible.
REFERENCES
[1] e. a. M. P. Heilbrun, Stereotactic Localization and Guidance Using a Machine Vision Technique, Karger AG, 1992.
[2] e. a. S. Sevgican, Intelligent Network Data Analytics Function in 5G Cellular Networks using Machine Learning, IEEE, 2020.
[3] e. a. M. Beshley, A Self-Optimizing Technique Based on Vertical Handover for Load Balancing in
Heterogeneous Wireless Networks Using Big Data Analytics, MDPI, 2021.
[4] C. A. L. P. a. F. J. Lin, Incorporating OMA Lightweight M2M protocol in IoT/M2M standard architecture, Milan, Italy: IEEE, 2016.
[5] "Release 15 Description; Summary of Rel-15 Work Items, TR 21.915," 3GPP, 2018.
[6] e. a. P. Marsch, 5G System Design: Architectural and Functional Considerations and Long Term Research, Wiley, 2018.
[7] e. a. Michail-Alexandros Kourtis, Enhancing VNF performance by exploiting SR-IOV and DPDK packet processing acceleration, IEEE, 2015.
[8] A. A. &. S.-A. Andréasson, Balanced automatization levels in manufacturing systems, Göteborg, Sweden:
Springer, 1995.
[9] e. a. P. Popovski, Wireless Access for Ultra-Reliable Low-Latency Communication: Principles and Building Blocks, IEEE, 2018.
