AI-Powered, Autonomous User Experience Optimization and Investment Planning System for Multi-Layer Telecommunications Networks

Abstract

The invention relates to a planning system that integrates the mobile access (RAN), fixed fiber access (FTTH / GPON / XGS-PON), and IP backbone (core) layers of telecommunications operators into a single artificial intelligence (AI) graphene model, compatible with 5G network slicing, satellite-D2D (device-to-device) integration, and DePIN (Decentralized Physical Infrastructure Network) based data incentive mechanisms. This system enables anomaly detection, root cause analysis, capacity planning, capital (CAPEX) and operational expenditure (OPEX) optimization, and subscriber churn prediction, all focused on real user experience (QoE).

Description

1 TARIFF AI-Powered, Autonomous Users in Multilayer Telecommunications Networks Experience Optimization and Investment Planning System Technical Area The invention relates to 5G network slicing, satellite-D2D (device-to-device) integration, and DePIN 5. (Decentralized Physical Infrastructure Network) based data incentive mechanisms operators working in harmony and operating in the telecommunications sector mobile access (RAN), fixed fiber access (FTTH / GPON / XGS-PON) and IP backbone by combining the (backbone / core) layers under a single artificial intelligence (AI) graphene model; anomaly detection, root cause analysis, and capacity 10 focused on real user experience (QoE). planning, capital (CAPEX) and operational expenditure (OPEX) optimization It relates to a planning system that predicts subscriber churn. State of the Art Today, the networks of operators operating in the telecommunications sector Monitoring, analysis, and management activities are carried out independently by different network layers. This is carried out through silo-based approaches, which are considered as such. In this context, performance indicators related to the mobile access network (RAN), fixed fiber telemetry data relating to the access network and traffic relating to the IP backbone / core network and routing data are mostly collected separately by different monitoring and management systems. are being collected and evaluated. 20 Evaluation of cellular and radio access performance in mobile access networks. For this purpose, for example, the use of physical resource blocks (PRBs), handover success rate and Similar KPIs are tracked. In fixed fiber access networks, OLT / ONT Indicators such as telemetry, GPON occupancy rates, and optical power are used. IP In the backbone and core networks, NetFlow / IPFIX traffic data, BGP information and 25 Monitoring is carried out through similar network indicators. In these approaches, data obtained from different network layers are mostly Tools and instrument panels specific to the respective network area, without being related to each other. They are evaluated separately based on these factors. Therefore, mobile access, fixed access and IP are considered. Holistic analysis of the relationships occurring between the vertebral layers 30 2 This is not possible or this analysis requires manual evaluation by expert users. It is needed. The silo approach in the current technique can be described as cross-domain blindness. This causes technical problems. For example, the issues users experience on mobile networks. Is the increase in latency due to resource utilization or cell load in the relevant cell? 5 or a bottleneck or capacity issue occurring on the IP backbone through which the traffic is carried. It is possible to automatically determine whether the problem arose due to a specific reason. This is not the case. Data from different network layers cannot be evaluated independently. Therefore, data from different domains is needed to determine the source of the problem. These 10 need to be manually collected and linked by the teams. This situation prolongs the troubleshooting and repair processes and increases the average repair time. This leads to an increase in (MTTR). The current technology also includes network performance indicators and real-world user experience. A direct and integrated correlation cannot be made between (QoE) and a network. The acceptable level of a specific KPI value for an element or cell is 15 This means that the actual experience of the relevant users is also acceptable. This may not be the case. For example, a cell may have PRB usage at a level of 70%. Even in this situation, users experience high buffering in video content. It is possible to experience this problem. However, current KPI-focused approaches, These kinds of relationships between network KPIs and users' real experiences are holistic. It is unable to model it in this way. In addition, capacity increase and investment in existing telecommunication networks. Planning processes are mostly carried out using reactive or periodic approaches. Capacity increase decisions are made based on user complaints, current performance levels, or Taking into account the capacity that may arise in the future after periodic evaluations 25 This limits the ability to determine the need in advance. For example, which mobile cell to add to Whether a 5G carrier needs to be added or in which OLT region capacity needs to be increased or not. Future changes in user experience regarding the need to perform division Taking this into account, it is not possible to proactively foresee the situation. This situation, one on the one hand, unnecessary capital investments, and on the other hand, when the need arises, 30 insufficient capacity can lead to a deterioration in user experience and It reduces the efficiency of capital utilization. 3 In the current technique, root cause analysis is also mainly carried out by specialist personnel. separate examination of information obtained from manual or different monitoring systems It relies on. When a malfunction or performance degradation occurs, mobile access, combining data from different domains such as fixed access and IP backbone They need to be brought together and the relationship between them established by experts. Different 5 Manually associating data sources reveals the true cause of the malfunction. This increases the time required for determination and therefore leads to an increase in the MTTR value. This is the reason. On the other hand, subscriber churn prediction is mostly based on customer surveys and usage data. through customer-centric indicators such as behavior or billing periods and 10 It is usually implemented with a delay. Network-related user experience. Although the impact of disruptions on subscriber churn can become apparent beforehand, current approaches compare network QoE changes with future churn behavior. the relationship in an integrated way using data from different layers of the network It is not possible to model it. Therefore, the negative 15 in user experience Detecting changes before subscriber loss occurs and acting accordingly Planning of preventive actions remains limited. Application number US20220329524A1, forecasting network capacity, customer to evaluate experience and prioritize investment situations It is related to the variable machine learning method. However, the application includes mobile access and fixed 20. Fiber access and IP backbone / core layer data are integrated into a single unified graphene topology. Integration is not carried out through this method. In conclusion, due to the negative aspects described above and the current solutions being the subject of discussion... Due to its shortcomings, an improvement is needed in the relevant technical field. It has been made. 25 Purpose of the Invention The invention was created by drawing inspiration from existing situations and overcoming the aforementioned drawbacks. It aims to solve the problem. The main purpose of the invention is to provide mobile access (RAN), fixed fiber access (FTTH / GPON / XGS-PON) and IP backbone / core layer data is stored on a single unified graphene 30 4 An AI-powered and autonomous network intelligence system that integrates across the topology, The goal is to provide user experience optimization and investment planning systems. The purpose of the invention is to transmit data in real-time and / or periodically from different network domains. The collected telemetry, performance, traffic, and user experience data are combined into a common dataset. The goal is to ensure that data is collected and processed within the processing infrastructure. This includes cells, OLTs, and 5-bit systems. Nodes are physical and logical entities such as routers and user clusters; they manage traffic flows, a composite graphene where handover paths and dependencies are represented as edges Creating the model and analyzing this model using GNN-based methods that is intended. Another aim of the invention is to detect anomalies in different network layers temporally and 10 Automatic detection at the cross-domain level, taking topological relationships into account. The goal is to investigate and determine the source of the anomaly. For this purpose, Isolation Forest, Temporal Transformer and Spatial Graph Convolution methods are used together. Additionally... Possible root causes of problems identified using Bayesian causal graphs and SHAP-based methods The aim is to determine the causes and their level of contribution. 15 Another aim of the invention is to determine future capacity from past and current network data. Needs forecasting and reinforcement learning methods with LSTM-based prediction. using proactive methods that aim for maximum QoE improvement with minimum CAPEX. The goal is to develop capacity and investment proposals. Another aim of the invention is to combine network QoE data with CRM data to create Survival 20. Churn risk can be predicted in advance on a subscriber basis using Analysis and Gradient Boosting methods. It involves predicting and generating early warnings. The invention enables network monitoring, anomaly detection, root cause analysis, QoE optimization, and capacity utilization. Planning, CAPEX / OPEX optimization, and churn prediction are all handled by a single AI-native and autonomous system. It aims to integrate on the platform. As a result of the implementation, MTTR's 25 30–45% reduction in OPEX, 8–15% reduction in CAPEX efficiency, and 12–18% reduction in churn. The goal is to improve performance by 5–8%. The invention also applies to 5G networks. compatible with slicing, satellite-D2D and DePIN based architectures It can be configured. The invention offers federated learning support. It supports 30 user devices. The collected QoE data is not sent raw to a central server; it is processed on the device itself. Partial model updates are made, and only weight updates are shared with the central office. This ensures compliance with GDPR and KVKK (Turkish Personal Data Protection Law). The platform is designed to scale horizontally to accommodate more than 10 million subscribers. It has been designed. Data processing cluster (Kafka, Spark Streaming), distributed graphene database. (JanusGraph or Neo4j cluster) and GPU cluster (TensorFlow Serving) 5 Orchestration is done with Kubernetes. At the request of some business partners, the system is a Decentralized Physical Infrastructure Network. It can support the (DePIN) mechanism. QoE shared by users from their devices. Data collection is incentivized with token rewards on a blockchain smart contract. This is how data collection works. It increases the sustainability and reliability of the network. 10 The target end-to-end delay for anomaly detection is < 5 seconds. Root cause analysis and CAPEX. Optimization is done at the minute level (offline training models once a day). (updated). Therefore, the system uses a lambda architecture (velocity layer + batch processing layer). It has been built. When access to actual operator data is restricted, synthetic data is used for training the invention. 15 A network simulator is used. This simulator is based on OMNeT++ or ns-3, and It generates realistic mobile / stationary / backbone traffic models. The structural and characteristic features and all the advantages of the invention are given in the figures below. This becomes clearer thanks to the detailed explanation written with references to these figures. This will be understood, and therefore the evaluation should also take these forms and detailed explanations into account. 20 It needs to be done by taking precautions. Figures that will help understand the invention. Figure 1 shows a representative block diagram of the system described in the invention. Explanation of Part References 1. Multi-domain data collection module 25 2. Unified graphene model generator 3. Processor 4. Cross-domain anomaly engine 6 5. Root cause inference engine 6. Optimization Unit 7. Risk analysis unit 8. Autonomous decision layer Detailed Description of the Invention 5 This detailed explanation describes artificial intelligence in multi-layered telecommunication networks, which is the subject of the invention. intelligence-powered, autonomous user experience optimization and investment planning system. preferred structures are solely for the purpose of better understanding the subject. It is explained. The system described in the invention consists of 10 mobile RAN, fixed fiber and IP backbone infrastructures. By combining the obtained data under a single unified graphene model, anomaly detection and root canal analysis can be performed. root cause analysis, capacity and investment optimization, and subscriber churn risk analysis. It is an AI-powered and autonomous system that performs this task. The system in question consists of a multi-domain data acquisition module (1), and a combined graphene model generator. (2), processor (3), cross-domain anomaly engine (4), root cause inference engine (5), 15 optimization unit (6), risk analysis unit (7) and autonomous decision layer (8) It consists of. Multidomain data collection module (1) collects data from different telecommunication network domains. It is a data collection unit that works 24 / 7 to collect data. Multi-field data collection module (1), Data such as cell-based PRB usage and handover matrix from the mobile RAN layer; 20 OLT / ONT telemetry from the fixed fiber layer, GPON occupancy, and optical power deviation, etc. data from the IP backbone layer includes Netflow / IPFIX, BGP route instability, and MPLS collects data such as LSP usage. Multidomain data collection module (1), TCP re-enabled from user SDK on demand. It can also collect data related to user experience, such as transmission rate. 25 The data in question is uploaded to the system in real-time (streaming) and / or periodically. It can be transferred. Message queue structures like Kafka can be used for data transfer. Multi-field on the data in different formats and protocols received by the data collection module (1) By performing protocol conversion and timestamp normalization, different Data from various sources are processed within a common time and data structure. 30 7 The combined graphene model generator (2) is obtained by the multi-domain data acquisition module (1). It represents the different domain data obtained within a common network topology. Unified The physical and logical entities of the network are nodes by the graphene model generator (2). Nodes are modeled as nodes, and the relationships between these entities are modeled as edges. For example, a Cell Node_A can be defined as an OLT Node_B and a router Node_C. Cell 5 Entities belonging to different layers, such as OLTs and routers, exist within the same multi-layered topology. Relationships such as traffic flow, handover relationship, and backlink dependency are associated with edges. It is shown via. In this way, mobile RAN, fixed fiber, and IP backbone layers provide independent data. not as separate sets, but as a single unified 10 while preserving the physical and logical relationships between them. It is represented within a graphene structure. The processor (3) is a multilayer created by the combined graphene model builder (2). It enables the processing of graphene topology. The processor (3) includes graphene convolution (graph By performing (convolution) operations, a failure or malfunction occurring at a node is detected. The potential for congestion to spread to related neighboring nodes and different network layers is 15. is calculated. This process allows us to determine the performance of only a single network element. not the change itself, but other factors such as the cell, OLT, and router to which the change is linked. The potential impact on assets is also assessed. The graphene structure processed by the processor (3) is processed by the cross-domain anomaly engine (4) and a combined 20 that can be used by other AI-based analysis components in the system It creates a network representation. Cross-domain anomaly engine (4) uses time series from different network domains and It performs anomaly detection on graphene-based data. Cross-domain anomaly detection. By applying Isolation Forest to the time series data coming from the engine (4), the outliers Points are determined. Then, using the Temporal Transformer, the past tense 25 is defined. Dependencies are modeled and the nature of the identified anomaly is classified. Within this scope... Types of anomalies such as sudden fluctuations or trend breaks can be identified. In addition, The anomaly detected using Spatial Graph Convolution is a composite graphene. The potential for propagation to neighboring nodes within the system is calculated. As a result of these operations, the cross-domain anomaly engine (4) generates an anomaly score 30 and tags the network domain from which the anomaly originated or is associated. Thus, a performance or user experience problem is only identified when it occurs. 8 Not based on the point, but considering associated domains such as mobile RAN, fixed fiber, and IP backbone. It is collected and evaluated. Root cause inference engine (5) determined by cross-domain anomaly engine (4) It works to determine the possible causes of anomalies. Root cause inference engine (5), There are 5 possible causes for a problem identified using Bayesian causal graphs and SHAP values. It creates a probabilistic distribution of causes for the possible causes. For example, a high video buffering problem was detected in the Ankara region. In this case, the root cause inference engine (5) identifies the problem with GPON OLT with 62% probability. It is caused by blockage and, with a 21% probability, peering saturation. It can generate a probabilistic output. In this way, 10 obtained from different network layers. By using the relationships between the data, the possible root causes of the problem and their contributions can be identified. levels are determined. The optimization unit (6) estimates the future network capacity needs and this Based on the forecast, it enables the creation of investment proposals with capacity expansion. The optimization unit (6) uses LSTM-based time series forecasting to predict the next four 15 It estimates the weekly traffic load. The resulting estimates are based on a combined graphene topology and User experience (QoE) metrics are used in the evaluation. The situation in the reinforcement learning (RL) agent used within the optimization unit (6) The information relates to graphene topology and QoE metrics, actions influence investment decisions, and the reward is QoE. This represents the value obtained based on the CAPEX cost of the increase. Thanks to this structure, 20 Optimization unit (6), cell-based 5G carrier return on investment (ROI) Investment recommendations such as evaluation, OLT split recommendation, and backbone link upgrade priority. It can create. For example, the optimization unit (6) stated that “5G carrier should be added to Cell_ID 1234, OLT_ID An action plan could be suggested such as "5678 should be split, backbone link X should be increased". 25 Risk analysis unit (7), network user experience data and customer relationship management By evaluating customer service (CRM) data together, it performs churn risk analysis on a subscriber basis. Risk analysis unit (7), video obtained from user SDK or other sources Subscriber-based QoE metrics such as initialization latency and TCP retransmission rate; CRM 30 with data such as billing history and customer service call count obtained from their systems It combines. 9 Each of the combined data is analyzed using a Gradient Boosting-based XGBoost model. The probability of churning for a subscriber is calculated. Also, the Kaplan-Meier survival model is used. Using this method, the risk of losing a subscriber over a specific time period is predicted and weekly. A churn risk score is generated. This allows for the identification of network-related changes in user experience. By jointly assessing the risk of subscriber churn, an early warning system is implemented. 5 Autonomous decision layer (8), cross-domain anomaly engine (4), root cause inference engine (5), the outputs produced by the optimization unit (6) and the risk analysis unit (7) are given to the operator transfers to their existing systems. Autonomous decision layer (8), anomaly alert, root cause The report includes capacity and CAPEX recommendations, churn risk lists, and RESTful API and / or Connect to existing NOC boards, OSS / BSS systems or automation 10 via WebSocket. They can transmit it to their vehicles. For example, “Add 5G carrier to Cell ID 1234” created by the optimization unit (6). The proposal is to connect to the relevant NOC or automation system via the Autonomous decision layer (8). can be transferred. Similarly, root cause analysis results and churn risk lists are also relevant. can be transmitted to systems. 15 Integration of system outputs into existing operational infrastructures with autonomous decision layer (8) This is ensured by the system being in the form of containerized microservices on Kubernetes. It is executable. Thanks to its microservice architecture, the system can be scaled horizontally. can be provided. As a result, in the system in question, mobile 20 via Multi-domain data collection module (1) Data is collected from RAN, fixed fiber, and IP backbone layers; combined graphene model With the generator (2), these layers are combined into a single graphene topology; processor Graphene-based dispersion analyses are performed with (3); cross-domain anomaly Network anomalies are detected with the engine (4); root cause extraction is done with the engine (5). The possible causes of anomalies are determined; the optimization unit (6) and capacity and investment 25 recommendations are being formulated; subscriber churn risks are being calculated by the risk analysis unit (7) and Results obtained with the autonomous decision layer (8) are applied to existing operational systems is being transferred. This structure allows for manual adjustments, which traditionally had to be done separately in different network domains. Data analysis, anomaly detection, root cause analysis, capacity planning, investment 30 optimization and churn risk analysis processes on a single combined graphene model. This is ensured by artificial intelligence support and in an autonomous manner.

Claims

REQUESTS 1. AI-powered, autonomous user experience in multi-layered telecommunications networks. It is an optimization and investment planning system, and its feature is; 5  receiving data from different telecommunication network domains, the data it receives protocol conversion and timestamp normalization Multi-field data collection module (1),  Different domains obtained by the multi-domain data collection module (1) composite graphene 10 that represents data within a common network topology model builder (2),  Multilayer generated by the combined graphene model generator (2) Processor that handles graphene topology (3),  Time series and graphene-based data from different network domains Cross-15, which creates an anomaly score by performing anomaly detection. domain anomaly engine (4),  possible anomalies identified by the cross-domain anomaly engine (4) to determine the causes, for the possible causes of an identified problem a root cause inference engine that generates a probabilistic cause distribution (5),  Predicting future network capacity needs using time series forecasting 20 estimating optimization unit (6),  Combining network user experience data with customer relationship management data risk analysis unit that performs risk analysis on a subscriber basis by evaluating (7),  Cross-domain anomaly engine (4), root cause inference engine (5), 25 outputs produced by the optimization unit (6) and the risk analysis unit (7) autonomous decision layer that transfers to the operator's existing systems (8) It includes.

2. Optimization and investment planning system according to Claim 1, its feature is; mobile RAN 30 from the layer at the cell level, from the fixed fiber layer and from the IP backbone layer It includes a multi-field data collection module (1) that receives data.

3. It is an optimization and investment planning system according to Claim 1, and its characteristic feature is that the network The physical and logical entities are the nodes, and the relationships between these entities are the edges. It includes a combined graphene model generator (2) that models. 35 11 4. Optimization and investment planning system according to Claim 1, its feature is; graphene. By performing convolution operations, a malfunction occurring at a node or congestion spreading to related neighboring nodes and different network layers It contains a processor (3) that calculates its potential.

5. It is an optimization and investment planning system according to Claim 1, and its feature is; incoming time 5 Identifying outliers in the series data, past dependencies cross-domain modeling and classifying the nature of the detected anomaly The anomaly is that it includes the motor (4).

Images