There’s a dizzying array of information accessible to telecom operators on the standing, efficiency and well being of their networks; at numerous speeds and ranges of granularity; centered on end-user-experience, peak community functionality, or dozens of different potential indicators.
Operators face appreciable challenges in making the very best use of that knowledge and translating it into actionable intelligence to validate, guarantee and optimize community operations. RCR Wi-fi Information reached out to Ayodele Damola, director of synthetic intelligence/machine studying technique at Ericsson, for his perspective on these challenges and the present panorama and tendencies in gathering, analyzing and leveraging community knowledge in addition to the usage of AI and ML in telecom networks.
This Q&A was carried out by way of electronic mail and has been calmly edited.
RCR: Because the business strikes additional into 5G and 5G Standalone deployments, in addition to disaggregation and cloud-native networks, how would you describe the challenges surrounding navigating network-related knowledge? Is it a matter of recent sources (i.e., microservices resulting in extra knowledge granularity), new quantity or scale of information, or the rate/velocity at which knowledge is offered — or another issue?
The principle problem navigating community knowledge is the rise in complexity. As 5G beneficial properties traction throughout the globe, communication service supplier (CSP) networks have gotten much more advanced. The complexity is because of the new set of providers being supplied, the rise in quantity and varieties of units within the community, availability of recent spectrum frequencies and bands, and the evolution of bodily networks to virtualized networks. This complexity equates to a rise within the quantity of community knowledge generated; community knowledge generated in several time frames along with the presence of extra number of community knowledge. If we take, for instance, the reference parameter rely evolution in 3GPP radio entry networks (RAN) – we noticed that with 2G networks after we had only a few hundred reference parameters that wanted to be configured. In 5G, that quantity has grown to a number of thousand parameters. In 6G, we anticipate a fair bigger variety of reference parameters. Improve in complexity can be mirrored in ‘overabundance of information’ – it turns into tough to seek out related knowledge with out modern filtering and aggregation. One other problem is the dearth of standardization of information throughout proprietary vendor gear resulting in difficulties in leveraging the info for insights. Yet one more problem is knowledge administration – at the moment, it’s an afterthought consisting of brittle system data (SI) pushed pipelines that are inefficient. Principally, knowledge is duplicated and exhausting to manipulate. All these components put new calls for on the extent of effort wanted to handle and management the community, and thereby result in an elevated CSP operation expense (OpEx) and ultimately elevated capital expense (CapEx). It’s turning into clear that it’ll now not be doable to handle networks in a legacy method the place community engineers take a look at dashboards and make adjustments manually to the community. That is the place a expertise like AI enters the image, a expertise that permits automation and thereby reduces complexity.
RCR: Do you assume that telecom operators, by and huge, have an excellent deal with on their community knowledge? What do you assume they do properly, and the place is there room for enchancment?
Entry to community knowledge is a crucial part of most CSP methods at the moment. Community knowledge from multi-vendor networks has been considerably of a problem for CSPs as a result of every vendor has barely completely different community node interfaces – therefore, the format and syntax of the info could also be completely different throughout completely different distributors. Typically, CSPs lack a constant technique for knowledge persistence and publicity. Completely different storage mechanisms (e.g., knowledge lakes) and retention insurance policies complicate the scalable dealing with of information volumes. Sooner or later, open requirements promise to alleviate this problem, particularly by the introduction of the Service Administration and Orchestration (SMO) framework outlined by the ORAN Alliance, which supplies a set of well-defined interfaces enabling CSPs to entry and act on knowledge from each purpose-built and virtualized multi-vendor and multi-technology networks. Community knowledge is uncovered by way of open interfaces (e.g., R1, A1, O1, O2 and many others.) to the completely different SMO features.
Inside the SMO (our implementation is known as the Ericsson Clever Automation Platform) the Information administration and ingest operate permits CSPs to effectively and securely ingest and handle knowledge. The AI/ML and perception technology operate permits the flexibility to course of and analyze the info, deriving insights and facilitating actuations.
RCR: What do community operators need to use their knowledge for? Are you able to give the highest three makes use of for network-related knowledge which might be the first curiosity of MNOs?
Market analysis earlier commissioned by Ericsson reveals that CSPs leverage community knowledge throughout many use circumstances, with the highest three being:
- Buyer Expertise Administration: Resolution helps CSPs predict buyer satisfaction, detect expertise points, perceive root causes, and mechanically takes the subsequent greatest motion to enhance expertise and operational effectivity resulting in churn discount and elevated buyer adoption.
- Safety/Fraud & Income Assurance: Addresses safety administration and income assurance together with billing and charging.
- Cloud & IT Operations: Automation of cloud and IT administration operations together with administrative processes with assist for {hardware} and software program, and the fast isolation of faults.
Moreover, different essential use circumstances embody Community Administration & Operations, Enterprise Operations, Service Assurance for RAN & Core, Community Design and Optimization, RAN Spectrum and Visitors Administration.
RCR: What position is AI/ML taking part in in networks at this second in time? Individuals are very within the potential — what’s the present actuality of sensible AI/ML use in telecom networks, and will you give some real-world examples?
AI/ML guarantees large potential on the subject of optimizing and automating CSP networks. The business continues to be in its early days, however some issues which have been solved with AI/ML present substantial beneficial properties. The complete systemization of CSP networks primarily based on AI/ML, additionally referred to as AI-native networks, continues to be some years away. In the case of the RAN particularly, Ericsson believes that AI/ML will play a key position within the following areas:
- Community evolution: Enhances community planning with extra environment friendly RF planning, web site choice and capability administration. Improves community and repair efficiency and permits new revenues by knowledge pushed and intent-based insights and suggestions.
- Community deployment: Handles provisioning and life cycle administration of advanced networks with optimum prices and velocity to market.
- Community optimization: Clever autonomous features to optimize buyer expertise and return on investments, e.g., RF shaping, visitors and mobility administration, vitality effectivity and many others.
- Community therapeutic: Service continuity and backbone of each primary and complicated incidents, delivering excessive availability whereas holding the operation prices at a minimal.
- AI and automation basis: Permits sooner TTM for – and belief in – excessive efficiency AI and automation use circumstances by way of openness and suppleness.
As well as, Ericsson believes CSPs will profit from an end-to-end managed providers operations answer enabling the clever administration of CSP networks and providers to supply superior connectivity and person expertise. Powered by superior analytics and machine studying algorithms, CSPs will profit from the flexibility to foretell potential community points brought on by {hardware}, software program, or exterior components corresponding to climate disturbances or buyer habits patterns.
Sensible AI/ML use in telecom networks at the moment:
- Capability Planning: Supplies the flexibility to carry out proactive planning primarily based on visitors forecast together with AI/ML prediction of utilization KPIs. The end result would be the optimum capability Expenditure to fulfill a sure QoS stage. Forecast predictions, bottleneck identification and community dimensioning are the primary use circumstances. Advantages embody 20-40% CapEx financial savings much less provider expansions in comparison with conventional strategy, and 83% elevated operational effectivity elevated operational effectivity on dimensioning duties.
- Efficiency Diagnostics: An answer that analyzes CSPs’ RAN to detect and classify cell points. Recognized points are additional investigated all the way down to root trigger stage, enabling quick and correct optimization of end-user efficiency. Advantages embody: 30% improve in capability per optimization full time worker (FTE), and 15% higher downlink velocity in cells with points.
- Improved spectrum effectivity: By accumulating adjoining cell knowledge in actual time, it’s doable to optimize radio hyperlink efficiency utilizing sample recognition. Advantages embody: 15% improved spectrum effectivity and 50% elevated cell edge DL throughput.
- Sustainability: An autonomous mechanism utilizing AI/ML applied sciences and closed-loop automation to scale back each day radio community vitality consumption by as much as 25% with zero impression on person expertise.
In the case of the potential of AI/ML in CSP networks, we envision a journey with a number of steps:
Within the earliest step, there was full human intervention because the community was manually configured. Within the subsequent step, the community configuration was nonetheless carried out by people, however the effort stage was lowered to the configuration of a beneficial set of parameters. Within the rule-based step, the human position was to create a algorithm which then managed the community; this required that the human builders have a deep understanding of how the community features. We’re presently transitioning to the step with autonomous options the place we’ve AI/ML fashions adapting to new conditions and thereby giving CPS extra automated management. Going ahead, the imaginative and prescient is that the community will evolve to totally autonomous with no human intervention aside from setting intents by which the community operates – mainly, offering the ‘what’ necessities to the community with the community performing the wanted ‘how’ actuation and management. It’s a journey just like the evolution of vehicles from manually managed to totally self-driving.
RCR: When it comes to edge vs. centralized cloud, is most knowledge processing nonetheless centralized or do you see issues really turning into extra distributed and taking place on the edge? Is it completely different for telco workloads vs. enterprise workloads?
You will need to distinguish between knowledge assortment, sometimes for the aim of coaching an AI/ML mannequin, and inference, which entails appearing on new knowledge by a skilled mannequin. The method of accumulating knowledge will entail cleansing and sorting the info after which utilizing this knowledge to coach fashions. Whereas the info assortment will occur out within the community in a distributed method, the info processing will sometimes be finished centralized. After the mannequin is skilled, the deployment of the mannequin within the community may very well be both distributed or centralized, relying on the use case. Given the massive quantity of community knowledge and given its real-time nature, inferencing of telco workloads will sometimes be extra distributed in comparison with enterprise workloads.
RCR: There was for some years a giant push towards “knowledge lakes” and storing as a lot knowledge as doable to sift via for enterprise intelligence. Is that this the case for community knowledge as properly, or is there extra choice for real-time intelligence? What does “actual time” really imply proper now, how shut can it get?
The choice for actual time intelligence will depend upon the use case. Community knowledge is generated throughout the community in several time frames, and we will classify use circumstances primarily based on the community knowledge into fast-loop use circumstances (microsecond timeframe) and slow-loop use circumstances (timeframe of days and even weeks) and all of the in-between. The character of the use circumstances throughout completely different timeframes will differ considerably. An actual time or fast-loop use case is, for instance, radio scheduling made in microseconds executed regionally inside a community node and sometimes totally automated i.e., finished and not using a human within the loop. Sluggish-loop use circumstances can be centered extra on long run community tendencies like community benchmarking, requiring community broad coordination with extra time accessible to decide and can possible entail interplay with a human. Information lakes are then suited to the 2 higher loops within the image beneath, whereas the time sensitivity of each the info and the selections made within the two decrease loops converse in opposition to persistence of the decision-bearing knowledge in any sort of information lake.
RCR: What knowledge wants, challenges or adjustments ought to telecom operators be planning for now that you simply see coming within the subsequent 3-5 years?
I see three issues:
- High quality knowledge: CSPs are challenged with how you can outline and develop a high quality knowledge strategy from which all AI options will be delivered. The difficulty is that knowledge continues to be contained in silos throughout most CSPs, from legacy methods to new methods. AI/ML will solely make a distinction if clear knowledge is offered from all sources. Therefore, investing the time in after which coaching of AI algorithms with the appropriate knowledge can be important to lowering false alarms and in bettering AI effectiveness. An adjoining downside is the difficulty of related knowledge – a few of the knowledge generated has little or no to no entropy (i.e., its usefulness is restricted) and figuring out such knowledge is an ongoing problem.
- Information platform: All CSPs require a strong platform to combination, sanitize, and analyze the info. Whereas there are a number of potential options available in the market, there are issues of privateness, safety, and vendor lock-in that make knowledge platform choice tough.
- Information technique: Many CSPs lack an end-to-end knowledge technique which covers knowledge governance, simplification and automatic assortment, and knowledge evaluation. Whereas some CSPs have a Chief Information Officer (CDO), many CSPs haven’t but been capable of implement a companywide knowledge technique throughout numerous organizations resulting in disconnected islands of information.
