DYNAMIC TRAFFIC ROUTING AND SWITCHING MANAGEMENT SYSTEM WITH ARTIFICIAL INTELLIGENCE IN 5G NON-TERRELSIAN NETWORK AND SATELLITE INTEGRATION

Abstract

This invention relates to a system (1) that provides handover management and artificial intelligence-assisted network optimization, ensuring coverage extension and continuity of user plane and control plane signaling, especially in remote and rural areas, in relation to hybrid scenarios where 5G (Generation) Non-Terrestrial Network (NTN) architectures and terrestrial radio access network (RAN) and satellite-based access nodes work together, as defined within the scope of 3GPP (3rd Generation Partnership Project).

Description

1 TARIFF 5G NON-TERRELSIAN NETWORK AND SATELLITE INTEGRATION DYNAMIC TRAFFIC DIRECTION AND MANUAL TRANSMISSION WITH ARTIFICIAL INTELLIGENCE CHANGE MANAGEMENT SYSTEM Technical Area This invention is part of the 3GPP (3rd Generation Partnership Project) 5G (Generation) Non-Terrestrial Network defined within this scope Terrestrial Network (NTN) architectures and terrestrial radio access network (Radio Access 10 A hybrid network (RAN) where satellite-based access nodes work together. scenarios related to; expanding coverage, especially in remote and rural areas and handpiece that ensures continuity of signaling between user plane and control plane handover management and AI-powered network optimization It is related to a system that provides. 15 Previous Technique In the known state of the art, NTN handover decisions are based on RSRP / RSRQ and altitude. It is based on the elevation angle thresholds. The rapid orbital motion of LEO satellites is 20 Due to the orbital period (~90–120 min, beam transit time ~2–8 min), static thresholds may be required. early (unnecessary ping-pong handover) or late (RLF – Radio Link Failure) This is triggered by network load, user movement pattern, and Doppler shift. Dynamic adaptation of thresholds according to the instantaneous situation is defined in the current standard. It is not. In LEO satellites, the Doppler shift is ±45 kHz (S-band @ 2 GHz) or 25 Can reach ±200 kHz (Ka-band); Timing Advance (TA) values ​​are 5–40 ms. It varies within this range and needs to be constantly updated due to satellite movement. In current systems, Doppler and TA compensation is performed on the UE side; however, In the handover decision, the estimated values ​​of these parameters (for the next beam) Expected Doppler / TA) is not taken into account – this results in 30 seconds of transition to the target beam. This causes loss of synchronization and high BLER. Terrestrial gNB handover and 2 NTN beam transitions are managed independently. A UE (User Equipment- User Equipment) simultaneously at both the terrestrial cell boundary and the satellite beam When this occurs at the transition point, there is a risk of a double interruption. These two The temporal coordination of the event is absent in current implementations. 5QI, S- Differentiated handover based on NSSAI and service type (eMBB, URLLC, mMTC) 5 And while carrier selection is necessary, existing systems apply the same static load to all services. implements the policy. For URLLC traffic, terrestrial route preference or mMTC Service-aware dynamic decisions such as selecting an energy-efficient satellite path. It is not defined as being integrated with artificial intelligence in current products. NTN In its architecture, transparent payload satellites provide feeder link (satellite-gateway) quality of 10. It directly affects service link (satellite-UE) performance. Rain on the feeder link. attenuation (rain attenuation, ~10 dB in Ka-band) or gateway congestion When this happens, end-to-end QoS can degrade even if the satellite GNB signal quality appears good. Current handover mechanisms only handle service link (UE-satellite) metrics. It takes into account; it does not integrate the feeder link status into the handover decision. 15 Therefore, considering the studies and shortcomings in the current technique... when considered, 3GPP (3rd Generation Partnership Project) 5G Non-Terrestrial Network (NSN) defined within the scope of the Project Network-NTN) architectures and terrestrial radio access network (Radio Access Network- 20 with hybrid scenarios where RAN and satellite-based access nodes work together relevant; expanding coverage and users, especially in remote and rural areas. hand switching that ensures continuity of signaling between the control plane and the control plane a system that provides (handover) management and AI-powered network optimization. It is understood that the system is needed. 25 United States Regulation US2025374135A1, which is included in the known state of the art. The patent document describes 5G core networks and non-terrestrial network cells. It is mentioned that the system in question, in this invention, is a wireless network similar to a 5G (5th generation) network. The network's core network (CN) provides signaling for efficient resource and mobility management. To optimize, provide support for non-terrestrial network (NTN) types. 3 It is adapted to accommodate CN, user equipment (UE), and network capabilities. Supported satellite types (geostationary orbit satellite (GSO) and geostationary incompatible with respect to non-north-orbit satellite (NGSO) types (including signaling with a next-generation radio access network (NG-RAN) to determine if it is not present. It uses and prevents unnecessary calls from being made. CN also uses specific NTN 5 defining a unique tracking area code (TAC) for the cells and an extended list of Tracking Area Identifiers (TAIs) that enables its use It is designed to be used. CN also includes a hand containing an NTN cell. to use a target identifier (ID) used in modifications It has been arranged. 10 Brief Description of the Invention The purpose of this invention is to support the 3GPP (3rd Generation Partnership Project). 5G Non-Terrestrial Network (Non-Terrestrial 15) defined within the scope of the Project Network-NTN) architectures and terrestrial radio access network (Radio Access Network- with hybrid scenarios where RAN and satellite-based access nodes work together relevant; expanding coverage and users, especially in remote and rural areas. hand switching that ensures continuity of signaling between the control plane and the control plane (handover) management and AI-powered network optimization 20 The goal is to implement a system developed for this purpose. Detailed Description of the Invention The “5G Non-Terrestrial Network and 25” project was carried out to achieve the purpose of this invention. Dynamic Traffic Routing and Manual Transmission with Artificial Intelligence in Satellite Integration The "Change Management System" is shown in the attached figure; Figure 1 shows a schematic view of the system that is the subject of the invention. 4 The parts shown in the figure are individually numbered, and these numbers correspond to... The corresponding answers are given below. 1. System 2. Multimodal Feature Vector Generator 5 3. Temporal Fusion Network-Based Hands-Off Window and Carrier Selection Module 4. Dynamic Traffic Routing and Service Quality Constraint Module 5. User Access Traffic Routing, Switching, and Segmentation Support Plane Bridge Module 10 6. Bidirectional Safe Feedback Loop Module 7. Network Data Analysis Function Integration Module 5G Non-Terrestrial Network architectures and terrestrial networks as defined under 3GPP. Hybrid 15 where radio access networks and satellite-based access nodes work together. scenarios related to; expanding coverage, especially in remote and rural areas and handpiece that ensures continuity of signaling between user plane and control plane To provide change management and AI-powered network optimization. The system in question, developed for the purpose of invention (1); - heterogeneous metrics from terrestrial gNB and NTN gNB into a single 20 configured to convert to a normalized input tensor at least one multimodal feature vector generator (2), - when and which carrier to choose from the feature vector array to execute a deep learning model that generates a combined process at least one configured temporal fusion network-based handover 25 window and carrier selection module (3), - Temporal fusion network-based handover window and carrier selection module (3) carrier and timing decision operator policy and Generating the final routing command by verifying it against SLA constraints. at least one dynamic traffic routing and service configured to 30 quality constraint module (4), - when the decision to change hands is made, the user plane on the new carrier at least one configured to prepare the route without interruption User access traffic routing, switching, and splitting supported. plane bridge module (5), - learns from the results of the changeover and with the operator safety filter 5 to operate as a limited model update mechanism at least one bidirectional secure feedback loop configured accordingly module (6), - the possibility of maintaining service quality in a specific cell / segment service quality sustainability analytics output in the form of and 10 User equipment movement direction / speed estimation in the form of user equipment mobility analytical output based on temporal fusion network changeover window and carrier selection module (3) decoder to enable its use as cross-attention input 15. Integration of at least one configured network data analysis function. It includes module (7). The multimodal feature vector generator (2) in the system subject to the invention (1), terrestrial gNB (Next Generation Radio Base Station, Next Generation Node B- Next Generation Radio Base Station, Next Generation Node B) and NTN (Terrestrial 20 Non-Terrestrial Network (GNB) from To convert heterogeneous metrics into a single normalized input tensor. It is structured. The multimodal feature vector generator (2) requires at least one feature vector for each dimension. and perform the maximum (min-max) normalization process, missing measurement. In this case, the exponentially weighted average of the last known value (EWMA, α=0.3) is 25 Interpolation: If the difference is more than 60 seconds, the relevant dimension is set to 0.5 (neutral). encoding, Doppler and TA (Timing Advance) encoding to do, calculate Doppler frequency offset, feeder link Calculating the quality index, power supply link SNR (Signal to noise ratio). Normalize the noise level (total noise ratio) and congestion (congestion) information on a 0-1 range. by adding it to the feature vector, the feed link quality index is less than 0.3. 6 In this case, reducing the score of the relevant NTN carrier reduces the feature vector to 100. Updating with millisecond intervals, Ephemeris derived features with 1 second intervals. It is configured to update. Multimodal feature vector generator (2), RSRP (Reference Signal) from terrestrial gNB, NTN gNB and Ephemeris source. Received Power (Reference Signal Received Power), Doppler, TA error, beam 5 heterogeneous in terms of visibility, supply link quality, and service quality profile. UE energy profile in the form of metrics and battery level, estimated NTN power cost. It is configured to convert to a 26-dimensional normalized tensor. Temporal fusion network-based exchange in the system (1) of the invention 10 (handover) window and carrier selection module (3), what from feature vector array the timing (handover timing) and the decision on which carrier (terrestrial / NTN / hybrid) to use running a deep learning model that generates information in a combined manner; with this model, 2 layers Bi-LSTM (Long Short Term Memory) and final T=30 Processing the feature vector array of the time step, Doppler trend, RTT (Round 15 Trip Time (Round Trip Time) change, in the form of decreased radiation visibility. It is structured to learn temporal dependencies. Temporal fusion. Network-based handover window and carrier selection module (3), multi-head attention (Multi-Head Attention (4 head, d_model=256)) and encoder output Ephemeris prediction sequence (ray transit schedule for the next 60 seconds) and network data analytics 20 Network Data Analytics Function (NWDAF) outputs analytical data (QoS) (Sustainability estimation) combined with cross-attention It is being structured. Temporal fusion network-based exchange window and Carrier selection module (3) output in the form of hand change timing header. to calculate the confidence interval using "Monte Carlo Dropout", 25 in hybrid selection Generating the traffic splitting ratio with an additional regression header, NWDAF's service Quality sustainability analytics output and EU mobility analytics using the output as the cross-attention input for the decoder, the model predicts network-level service quality degradation using its own local observations It is structured to enable its integration. 30 based on temporal fusion networks. The handover window and carrier selection module (3) of the artificial intelligence platform, 7 In O-RAN architecture, there is a clear division of labor between Near-RT RICs and Non-RT RICs. to enable feature vector generation, temporal fusion network inference (inference) and quality of service constraint validation on Near-RT RIC using xApp (to enable it to function as an application); with a cycle time of 10–100 ms. Generating timely changeover decisions, the optimized model from gNB 5 receiving key performance indicator streams and handover commands. to enable the transmission of long-term trajectory prediction, offline model training, A / B testing and operator policy management in a Non-RT RIC environment Executing as an rApp allows you to update model weights and policy parameters. Transferring AI machine learning / A1-policy interface to Near-RT RIC, 10 To report inference results and local performance metrics back to the Non-RT RIC. It is structured accordingly. Dynamic traffic routing and service quality in the system (1) that is the subject of the invention constraint module (4), temporal fusion network based handover window and carrier 15 selection module (3) carrier and timing decision operator policy and SLA (Verified with Service Level Agreement) restrictions It is configured to generate the final routing command. Dynamic traffic. Routing and service quality constraint module (4), TFN (Temporal Fusion Network- Temporal Fusion Network) decision 5QI latency budget, S-NSSAI (Single Network 20 Slice Selection Assistance Information resource constraint, feeder link quality, and Doppler / TA Verify with synchronization constraints, automatic weight assignment based on service type. It is structured to do so. Access traffic routing, transition and logging in the system (1) which is the subject of the invention. (Access Traffic Steering, Switching and Splitting- ATSSS) supported user plane bridge module (5), when the decision to change hands is made, in the new carrier It is configured to prepare the user plane path without interruption. Access traffic routing, transition and partitioning supported user-plane bridge 30 module (5), SMF (Session Management Function- before the change of ownership 8 Session Management Function (UPF) / User Plane Function (function) coordinated with GTP-U (GPRS Tunnelling Protocol - User Plane (GPRS Tunneling Protocol - User Plane) tunnel preparation, access Traffic redirection, passage and division via dual-path, simultaneous traffic, PDCP SN (Packet Data Convergence Protocol 5) Sequence Number (SQN) synchronization for packet sorting and destination Performing controlled traffic diversion operations once the road is stable, perceived It is configured to minimize downtime. The bidirectional safe feedback loop in the system (1) is 10 module (6) learns from the results of handover and operator safety filter It operates as a limited model update mechanism, each hand Automatically record metrics after a changeover event, the last 1000 hand changes. With a small group (mini-batch) consisting of the event, the final full version of the TFN model is produced every 6 hours. Performing fine-tuning operations on connected layers, weekly new and 15 Retraining with the entire training set, including all historical data. The process is structured to validate the model with A / B testing. Dual directional safe feedback loop module (6), NTN hand in specific geographic regions to prevent it from changing the carrier / timing suggested by the model, SLA budget If it exceeds 20%, reject the proposal and revert to the default policy. Applying a dynamic hysteresis margin to the model output, the last N hand changes hysteresis margin according to the average inter-frequency transition time of the event adjust if the handover timing suggested by the model is correct, before the last hand change. If it's shorter than a specified time, filter the suggestion or adjust the timing. shifting to a specified time, using a prohibited frequency band or power level 25 automatic decision blocking, event when machine learning suggestion is rejected saving and adding the model to the update queue, handover results and model confidence, Doppler value, selected carrier, key performance indicator Network data analysis function of local context metrics in the form of impact Inform the integration module (7), network data analysis function integration 30 9 to update the service quality sustainability models of module (7) It is being structured. The network data analysis function (Network Data) in the system (1) is the subject of the invention. Analytics Function- NWDAF) integration module (7), 5 in a specific cell / slice service quality in the form of the possibility of sustainability of service quality sustainability analytics output and user equipment movement direction / speed estimation The temporal fusion network is the analytical output of user equipment mobility in this form. cross-decoder of the based handover window and carrier selection module (3) To enable its use as a cross-attention (decoder) input, hand 10 using the change results to update their own analytical models It is structured accordingly. Industrial application of the invention The system in question (1) has both transparent payload (satellite only relays the signal; gNB is located at the gateway on the ground) and regenerative payload (gNB is on the satellite) (works; satellite on-board processing is performed) Non-Terrestrial Network (Non-Terrestrial Network) It covers network (NTN) architectures. In the transparent payload scenario, the feeder link (Gateway-to-satellite) quality directly affects service link (satellite-to-UE) performance and 20 should be included in the handover decision; in the regenerative payload scenario, the satellite gNB can make independent decisions on it, but inter-satellite handover Handover and on-board processing delay are considered as additional parameters. The invention's feature vector and decision architecture will support both payload types. It is designed in such a way that the feeder_link_quality_index size is 25 for the transparent payload. The active, regenerative payload is coded as 1.0 (neutral). In the system subject to the invention, (2) in the multimodal feature vector generator (1) The feature vector structure (d=26 dimensions) is as follows: - Terrestrial metrics (8 dimensions): RSRP_TN [dBm], RSRQ_TN [dB], SINR_TN [dB], throughput_TN [Mbps], PRB_load_TN [%], RTT_TN [ms], HO_fail_rate_TN [%], CQI_TN [0-15]. - NTN metrics (10 dimensions): RSRP_NTN [dBm], RSRQ_NTN [dB], SINR_NTN [dB], throughput_NTN [Mbps], PRB_load_NTN [%], RTT_NTN 5 [ms], Doppler_shift [Hz], TA_precomp_error [μs], beam_visibility_remaining [s], elevation_angle [degrees]. - Ephemeris derived (3D): time_to_beam_switch [s] (time to beam transition) (duration), next_beam_Doppler_predicted [Hz] (expected Doppler of the target beam), feeder_link_quality_index [0-1] (feeder link reported by gateway 10 (quality). - UE energy profile (2 dimensions): battery_level_normalized [0-1] (reported by UE) Battery level is monitored via UE Power Headroom Report and PowerSavingPreference. is obtained), estimated_NTN_power_cost [0-1] (in case of NTN carrier transition) Estimated additional power consumption rate – 15 times the ratio of NTN path loss to terrestrial path loss. It is derived; in the LEO link budget, UE is typically 6–10 dB higher than terrestrial. TX requires power). - QoS / service profile (3 dimensions): 5QI_normalized [0-1], slice_priority [0-1] (S- NSSAI priority map), service_type_encoding [one-hot: eMBB=0, URLLC=1, mMTC=2]. 20 Doppler frequency offset is calculated using the following formula: Δf_D = (f_c · v_sat · cos(α)) / c Here, f_c: carrier frequency [Hz] (e.g., S-band: 2 GHz, Ka-band: 20 GHz), v_sat: The satellite's LEO orbital velocity is [m / s] (typically ~7500 m / s), c: speed of light (3×10⁸ m / s), α: 25 satellite's velocity vector and UE-satellite line-of-sight It is the angle between them – this angle is different from the elevation angle and It depends on the orbital direction of the satellite. α is directly derived from ephemeris data (satellite It is derived from the position vector (position + velocity vector) and the UE position. When the satellite is directly overhead. (Rare) α ≈ 90° and Δf_D ≈ 0; as approaching the horizon, α → 0° and Δf_D 30 It reaches its maximum (±45 kHz in S-band, ±200 kHz in Ka-band). 11 TA pre-compensation error ε_TA = TA_measured - TA_predicted is coded; |ε_TA| Risk of loss of synchronization if > τ_TA (typical τ_TA = 500 μs) This is marked as high, and the handover urgency score is increased. Temporal fusion network-based exchange in the system (1) of the invention 5 (handover) window and the objective function (Multi- in the carrier selection module (3), Task Loss is as follows: L_total = λ₁·L_timing + λ₂·L_carrier + λ₃·L_qos_penalty + λ₄·L_energy - Handover timing estimation error with L_timing = Huber Loss (δ=1.0) - Carrier classification error with L_carrier = Cross-Entropy Loss is 10 - QoS budget violation with L_qos_penalty = max(0, PDB_predicted - PDB_budget)² Penalty (weight is increased for URLLC services: λ₃_URLLC = 3·λ₃) - UE energy budget with L_energy = max(0, NTN_power_cost - battery_budget)² The penalty is that `battery_budget` is dynamically adjusted according to the battery level: `battery_level` If < 0.2, the budget is tightened (transition to NTN is heavily penalized and terrestrial 15 Network "holding" is preferred); if battery_level > 0.6, the budget is relaxed. mMTC / IoT The weight of λ₄ for the devices is increased by 3× (λ₄_mMTC = 3·λ₄) – this increases battery life. In IoT scenarios where the timeframe is on the order of weeks / months, not hours / days It is critical. Default weights: λ₁=0.35, λ₂=0.25, λ₃=0.25, λ₄=0.15 20 Training details are as follows: - Optimizer: AdamW (lr=1×10⁻³, weight_decay=1×10⁻⁴) - Learning rate scheduler: Cosine Annealing (T_max=50 epoch, η_min=1×10⁻⁵) - Batch size: 256 (each sample = 3 seconds window + label) 25 - Training data: Minimum 500,000 handover events (successful + unsuccessful, labeled); Data source: historical handover logs + simulation data (ray-tracing + trajectory) model) - 5-fold cross-validation; target metrics: time MAE < 2 s, carrier selection accuracy > 90%, QoS violation rate < 5% 30 12 Dynamic traffic routing and service quality in the system (1) that is the subject of the invention The constrained optimization framework in the constraint module (4) is as follows: -Purpose: max U(decision) = w₁·QoS_score + w₂·cost_efficiency + w₃·reliability_score - Limitations: 5 (a) Service-based delay budget: delay(path) ≤ PDB(5QI) [e.g., URLLC] 5QI=82: PDB=10 ms → terrestrial preference] (b) Slice resource constraint: sufficient PRB for the relevant S-NSSAI on the selected carrier must have capacity (c) Feeder link quality constraint: feeder_link_quality > 0.3 (otherwise NTN carrier 10 (disabled) (d) Doppler / TA synchronization constraint: |ε_TA_predicted| < τ_TA in the target beam. - Weights (w_k) can be configured according to operator policy; service type based. Automatic weight assignment: URLLC → w₁=0.6, w₂=0.1, w₃=0.3; eMBB → w₁=0.4, w₂=0.3, w₃=0.3; mMTC → w₁=0.2, w₂=0.5, w₃=0.3 15 Access traffic routing, transition and logging in the system (1) which is the subject of the invention. The working mechanism of the supported user plane bridge module (5) is as follows: It is like this: (a) Pre-preparation: TFN engine handover decision t_HO - Δt_prep (Δt_prep = 2–5 20 s) instantly provides; new UPF path information (target gNB) to SMF via N11 interface. (or NTN gNB) is transmitted; UPF establishes a GTP-U tunnel to the target carrier. (b) Dual-path: Compliant with 3GPP ATSSS (TS 23.256) rules. as source and destination carriers with MPTCP or ATSSS-LL (Lower Layer) stack. Simultaneous data transmission is initiated. Packets arriving from the source path are placed in the buffer. Traffic is gradually shifted once the target route is stable. (c) Controlled traffic shifting: Target path RTT < 1.5× source path RTT and The source path tunnel is closed after it is confirmed that packet loss is less than 1%. The verification time is typically 500 ms–2 s. 13 (d) Fallback: If the target path does not become stable within Δt_prep + 5 s, The source remains en route and handover is postponed; the event enters a feedback loop. It is recorded as "failed transition". Session Continuity: Dual path through PDCP sequence number (SN) synchronization. During this process, packet sorting and duplicate elimination are performed. Audio and 5 Buffer depth for video streams: max 200 ms (URLLC) / 1 s (eMBB). In the system of the invention, (1) the terrestrial and (2) feature vector generator of the multimode feature vector. Heterogeneous metrics (RTT, Doppler, ephemeris, etc.) from NTN access nodes. RSRP, beam visibility, QoS profile) and UE energy profile (battery level, estimated 10 It produces a 26-dimensional input tensor by normalizing the NTN power cost; Doppler It encodes the frequency offset and TA pre-compensation error as explicit features. Temporal fusion network-based handover window and carrier selection. module (3) with LSTM encoder and Multi-Head Attention decoder architecture ephemeris estimation, radio conditions and quality of service (QoS) constraints are all the same 15 By combining them in the objective function, it produces the decision of "when" and "which carrier"; NWDAF uses the QoS Sustainability analytics output as additional input. Access Traffic Steering, Switching and Dividing User Plane Bridge module (5), supported by Splitting-ATSS) and Splitting-ATSS, transition It maintains session continuity during dual-path operation; with SMF / UPF 20 Coordinated tunnel pre-preparation and controlled traffic diversion to detect disruption. minimizes the time. Bidirectional safe feedback loop module (6), It triggers model fine-tuning by automatically tagging failed handover events; Operator safety filter, prohibited zone and machine learning carrying SLA violation risk. He rejects their proposals. 25 The reference standards used in the system in question (1) are: 3GPP TS 38,821 (5G NR – NTN: Solutions for NR to Support Non-Terrestrial Networks), 3GPP TS 38.300 (NR – Overall Description Stage 2: gNB architecture, NTN extensions), 3GPP TS 23.501 (System Architecture for 5G – NTN support, 30 ATSSS), 3GPP TS 23.256 (Support of ATSSS – Access Traffic Steering, 14 Switching and Splitting), 3GPP TS 38.331 (NR RRC – Conditional Handover, NTN configuration parameters), 3GPP TS 38.413 (NG-RAN / NG Application Protocol – NGAP handover procedures), 3GPP TS 23.288 (NWDAF – Network Data Analytics Function), 3GPP TR 38.811 (Study on NR to Support NTN – channel model, propagation), 3GPP TS 38.214 (NR Physical Layer – HARQ NTN 5 (adaptations), ITU-R S.1503 (LEO satellite orbital parameters) are included. Around these fundamental concepts, the subject of the meeting is “5G Non-Terrestrial Network and Satellite”. Dynamic Traffic Routing and Handing Over with Artificial Intelligence in its Integration It is possible to develop a wide variety of applications related to the Management System (1)” 10 and the invention cannot be limited to the examples described here, but mainly to the claims as stated.

Claims

REQUESTS 1. 5G Non-Terrestrial Network architectures defined under 3GPP and terrestrial radio access network and satellite-based access nodes together 5 related to the hybrid scenarios he worked on; especially in remote and rural areas. Coverage expansion and user plane and control plane the exchange management that ensures the continuity of signaling and artificial intelligence Developed to provide intelligent network optimization; - heterogeneous metrics from terrestrial gNB and NTN gNB into a single system. 10 configured to convert to a normalized input tensor at least one multimodal feature vector generator (2), - when and which carrier to choose from the feature vector array to execute a deep learning model that generates a combined process at least one structured temporal fusion network-based exchange window and carrier selection module (3), 15 - Temporal fusion network-based handover window and carrier selection module (3) carrier and timing decision operator policy and Generating the final routing command by verifying it against SLA constraints. at least one dynamic traffic routing and service configured for quality constraint module (4), 20 - when the decision to change hands is made, the user plane on the new carrier at least one configured to prepare the route without interruption User access traffic routing, switching, and splitting supported. plane bridge module (5), - learns from the results of the changeover and with the operator safety filter 25 to operate as a limited model update mechanism at least one bidirectional secure feedback loop configured accordingly module (6), - the possibility of maintaining service quality in a specific cell / segment service quality sustainability analytics output in the form of and 30 User equipment movement direction / speed estimation in the form of user 16 equipment mobility analytical output based on temporal fusion network changeover window and carrier selection module (3) decoder to enable its use as cross-attention input at least one configured network data analysis function integration a system characterized by module (7) (1). 5 2. Normalizing heterogeneous metrics from terrestrial gNB and NTN gNB into a single standardized system. multimode configured to convert to a converted input tensor a feature vector generator (2) as in Claim 1 system (1). 10 3. Perform minimum and maximum normalization operations for each dimension. In case of missing measurements, the exponentially weighted average of the last known value is used. Interpolating; if it is more than 60 seconds short, set the relevant dimension to 0.

5. encoding as, performing Doppler and TA encoding, Doppler frequency 15 calculating offset, calculating feed connection quality index, feed The connection normalizes SNR and congestion status information within a range of 0-1. by adding it to the feature vector, feed link quality index When it is less than 0.3, it lowers the score of the relevant NTN carrier, feature. Updating the vector with a period of 100 milliseconds, Ephemeris derived 20 Multi-mode configured to update features every 1 second. from the above requests characterized by the feature vector generator (2) a system like any other (1).

4. RSRP from terrestrial gNB, NTN gNB and Ephemeris source, Doppler, 25 TA error, beam visibility, power supply connection quality, service quality Heterogeneous metrics in the form of a profile and battery level, estimated NTN power. UE energy profile in the form of cost to 26-dimensional normalized tensor Multimodal feature vector generator configured to convert (2) a 30 as in any of the above claims characterized by system (1). 17 5. When and which carrier to use from a feature vector array? (combined) running a deep learning model that produces results in a 2-layered system. Bi-LSTM and processing the feature vector array of the last T=30 time step, Doppler trend, RTT change, and beam visibility reduction are the five indicators. Temporal fusion structured for learning temporal dependencies network-based handover window and carrier selection module (3) as in any of the above characterized claims system (1).

6. Multi-head attention and encoder output to ephemeris prediction sequence and network data. to combine the analytical function's analytical output with cross-referencing structured temporal fusion network-based exchange window and from the above requests characterized by the carrier selection module (3) a system like any other (1). 15 7. Obtaining the output in the form of a handover timing header, confidence interval Calculating traffic splitting in hybrid mode using "Monte Carlo Dropout" Generating the ratio with an additional regression header, NWDAF's service quality sustainability analytics output and EU mobility analytics output code 20 using the solver's cross-attention input at the network level of the model combining the prediction of service quality deterioration with its own local observations a temporal fusion network-based exchange structured to provide the above is characterized by its window and carrier selection module (3). a system like any of the requests (1). 25 8. The artificial intelligence platform's Near-RT RIC and Non-RT architecture is based on O-RAN. Ensuring a clear division of labor among RICs, feature vectoring production, temporal fusion network extraction and service quality constraints Verification: Ensuring it runs as an xApp on Near-RT RIC; 30 18 Real-time handover decisions with 10–100 ms cycle time. Producing, optimizing the model is a key performance indicator from gNB. enabling it to receive the stream and transmit the handover commands, long future trajectory forecasting, offline model training, A / B testing, and operator Implementing policy management as an rApp in a Non-RT RIC environment, 5 updated model weights and policy parameters using artificial intelligence. Transferring machine learning / A1-policy interface to Near-RT RIC, inference results and local performance metrics are returned to the Non-RT RIC. Temporal fusion network-based handover structured for reporting The above 10 is characterized by its window and carrier selection module (3). a system like any of the requests (1).

9. Temporal fusion network-based handover window and carrier selection. module (3) carrier and timing decision operator policy and SLA 15 Configured dynamic traffic routing and quality of service constraint module (4) like any of the above-mentioned claims characterized by system (1).

10. TFN decision 5QI delay budget, S-NSSAI resource constraint, feeder 20 Verify with link quality and Doppler / TA synchronization constraints, configured to automatically assign weight based on service type. dynamic traffic routing and quality of service constraint module (4) as in any of the above characterized claims system (1). 25 11. When the decision to change ownership is made, the user plane path on the new carrier access traffic configured to be prepared without interruption User-plane bridge with support for guidance, transition, and partitioning. 19 any of the above requests characterized by module (5) a system like one of them (1).

12. Prior to the handover, GTP-U tunnel pre-construction in coordination with SMF / UPF. Preparation, access traffic routing, transit and partitioning with dual carriageway 5 Simultaneous traffic, packet sequencing and destination with PDCP SN synchronization. Once the road is stable, perform controlled traffic diversion operations. access configured to minimize perceived downtime User-level bridge supporting traffic redirection, passage, and partitioning. Any of the above requests characterized by module (5) 10 a system like one of them (1).

13. Learning from the results of the handover and with the operator safety filter. It operates as a limited model update mechanism, each hand Automatically save metrics after a change event, last 1000 hands 15 The small group consisting of the replacement event every 6 hours TFN model perform fine-tuning operations on the last fully connected layers, The entire training set, including all new and historical data, is updated weekly. to retrain with, and to validate the model with A / B testing 20 with configured bidirectional secure feedback loop module (6) as in any of the above characterized claims system (1).

14. Preventing NTN transfers in specific geographic regions, the model If the proposed carrier / timing exceeds the SLA budget by 20%, the proposal will be reduced by 25%. rejecting and reverting to the default policy, in the model output Applying a dynamic hysteresis margin, the last N handover events Adjusting the hysteresis margin based on the average transition time between frequencies. if the handover timing suggested by the model is before the last hand change If it is shorter than the specified time period, filter the suggestion or 30 shifting the timing to a specified period, prohibited frequency band or Automatically blocking decisions that use power level, machine learning When the proposal is rejected, record the event and add the model to the update queue. to add, exchange results and model confidence, Doppler value, The selected carrier's local 5-point impact on key performance indicators. context metrics to network data analysis function integration module (7) informing, network data analysis function integration module (7) service dual structured to update quality sustainability models characterized by the directional safe feedback loop module (6) a system like any of the above requests (1). 10 15. The possibility of maintaining service quality in a specific cell / segment. service quality sustainability analytics output and user User equipment in the form of estimation of equipment movement direction / speed. mobility analytical output temporal fusion network-based handover 15 window and carrier selection module (3) decoder cross attention to ensure it is used as input, to enable the results of the change of ownership a network configured to update analytical models characterized by the data analysis function integration module (7) a system like any of the above requests (1). 20 30

Images