Dynamic optimization of CSI-RS transmission for user-beam configurations in o-ran ecosystem

Abstract

A system and method that allows dynamic optimization of Channel State Information–Reference Signal (CSI-RS) transmission, feedback reporting, and user-beam configurations in Open Radio Access Network (O-RAN) ecosystems, while simultaneously adapting system parameters based on real-time network conditions to achieve improved spectral efficiency, reduced interference, and minimized power consumption, thereby addressing limitations of static provisioning in fifth-generation (5G) mobile networks.

Description

0017507WGU / 4688Dynamic Optimization of CSI-RS Transmission for User-Beam Configurations in O-RAN EcosystemBACKGROUND1. Field of the Disclosure

[0001] The present disclosure pertains to telecommunications. The system enhances configurations of Channel State Information-Reference Signal (CSI-RS) transmission, CSI-feedback reporting, and user-beam strategies in Open Radio Access Network (O-RAN) ecosystems for 5G-based mobile networks.2. Description of Related Art

[0002] In the following sections, Next Generation Radio Access Network (NG-RAN) architecture disaggregation and the associated constituent functions are discussed. NG-RAN has evolved from a monolithic gNodeB into separate Central Units (CUs), Distributed Units (DUs), and Radio Units (RUs). Within this disaggregated architecture, Central Units perform higher-layer control-plane and user-plane functions, Distributed Units execute lower-layer real-time processing, and Radio Units undertake radio-frequency operations.

[0003] For the Open RAN ecosystem depicted in FIG. 1, a Service Management and Orchestration (SMO) domain 101 hosts non-real-time controllers, while near-real-time Radio Intelligent Controllers (RICs) 102 execute per-slot optimizations via standardized Al, E2, and 01 interfaces. These interfaces facilitate the collection of configuration parameters, acquisition of performance measurements, and deployment of control policies across multiple cells.

[0004] In existing 5G deployments, a gNodeB periodically transmits Channel State Information-Reference Signals (CSI-RS), as shown in FIG. 2, to enable User Equipment (UE) to perform channel measurements and report metrics such as Precoding Matrix Indices (PMIs), Rank Indices (RIs), Layer Indices (Lis), and Reference Signal Received Power (RSRP). These CSI reports play a role in selecting beamforming weights, scheduling UEs on physical resource blocks, and managing beam steering.

[0005] However, multiple configuration parameters govern CSI-RS transmission. These include the number of CSI-RS ports, the orthogonality domain (time, frequency, or code), signal density per resource block, transmission periodicity, port-to-antenna mapping, and selection among various codebook types for feedback reporting. In addition, sub-band sizing for feedback and enabling or disabling RI-PMI reporting further diversify the configuration space.

[0006] Alternatively, non-codebook-based beamforming relies on periodic Sounding Reference Signal (SRS) transmissions from UEs to DUs, exploiting channel reciprocity to derive beamweight coefficients without explicit CSI-RS feedback. This approach can reduce uplink feedback overhead but depends heavily on uplink channel quality and reciprocity calibration.

[0007] Although codebook-based CSI schemes incur significant overhead due to frequent uplink feedback, leading to increased time-frequency resource occupancy and reduced spectral efficiency, periodic transmissions of high-resolution CSI-RS and CSI reports can negatively impact resource utilization, especially when conservative assumptions — such as maximum user mobility — guide static design choices.

[0008] Moreover, when CSI configurations are statically or semi-statically provisioned at deployment, the network is designed to accommodate infrequent extreme scenarios, yet persistent static provisioning results in resource waste. Static provisioning can result in increased inter-cell interference and elevated power consumption in both network nodes and UEs, leading to higher capital expenditure and reduced UE battery life.

[0009] In some deployments, UE mobility varies both across a site and over time. High-speed users require more frequent CSI-RS transmissions to maintain channel accuracy, whereas low-speed users benefit from less frequent updates. Similarly, traffic load fluctuation demands adaptability: a high-load scenario may favor multi-user MIMO configurations with many CSI-RS ports and advanced codebooks, whereas lighter traffic can be served by fewer ports and simpler codebook settings.

[0010] Furthermore, the spatial distribution of UEs across Synchronization Signal Block (SSB) beams changes dynamically, so a fixed number of CSI-RS beams per SSB beam introduces unnecessary overhead when user density is low. Variations in coherence bandwidth and delayspread — driven by the multipath environment — also render static frequency-domain CSI-RS and feedback configurations suboptimal.

[0011] Angular spread and per-UE channel quality differences impact the choice between codebook- and non-codebook-based beamforming modes; however, a universal static selection cannot accommodate these dynamic channel conditions. Because CSI-RS transmission, feedback reporting, and beamforming strategies are often applied uniformly across cells, static logic fails to capture cell-specific and time-varying network dynamics.

[0012] Accordingly, there exists a significant demand for mechanisms that dynamically adjust CSI-RS transmission parameters, feedback reporting configurations, and user-beam strategies in response to real-time network conditions. Such mechanisms are designed to streamline resource utilization, improve spectral efficiency, mitigate interference, and reduce power consumption without relying on static provisioning that does not align with the temporal and spatial variability of modem mobile networks.SUMMARY

[0013] Accordingly, what is needed is a method for optimizing configurations in an Open Radio Access Network (0-RAN) for 5G mobile networks by coordinating a Near-Real-Time RAN Intelligent Controller (Near-RT RIC) together with a Service Management and Orchestration function (SMO) and / or Non-Real-Time RIC, and by modifying channel state informationreference signal (CSI-RS) transmission and CSI-feedback reporting at an E2 node of a Distributed Unit (DU).

[0014] The method is further desired to involve the SMO / Non-RT RIC adjusting CSI-RS transmission parameters, including the number of ports, port allocation over resource elements in a resource block, transmission periodicity, transmission density, CSI-RS port-to-physical-antenna-port mappings or beamforming weights, and the number of CSI-RS beams per synchronization signal beam, while simultaneously modifying CSI-feedback reporting parameters such as codebook configuration and sub-band size.0017507WGU / 4688

[0015] A further objective is to provide a method where the Near-RT RIC updates user-beam configurations by selecting codebook-based or non-codebook beamforming modes, specifying per-slot precoding matrix indices when PMI-based beam management is enabled, specifying perslot P2 beamforming weights or beam indices when Pl-P2-based beam management is enabled, and adjusting channel quality indicator (CQI) and modulation and coding scheme (MCS) indexes for user equipment.

[0016] A method is also desired for training an artificial intelligence / machine learning model using Deep-Q Learning-based reinforcement learning and CNN-LSTM-based prediction learning to optimize per-cell CSLRS transmissions, per-cell CSLfeedback reporting, policies for selecting beamforming modes based on CSLRS reference signal received power, and per-UE slot-specific codebook-based beam indexes and CQI / MCS indexes.

[0017] It is further desired to provide a method where, upon completion of training, the SMO / Non-RT RIC deploys the beamforming-mode policy, the trained AI / ML model, and enrichment information — such as histories of PMIs, P2 beam indexes, CQIs, and MCS indexes — to the Near-RT RIC. In this method, the Near-RT RIC configures the beamforming mode for each UE by associating CSI-RS reference signal received power with the received policy and inferring slot-specific beam indexes and CQI / MCS indexes for a sequence of slots using the trained model.

[0018] Modern 5G 0-RAN deployments are characterized by distributed intelligence and the need for seamless, dynamic adaptation of radio parameters. In one configuration, a gNodeB implements a smooth transition between an old CSI configuration and a new CSI configuration by operating concurrently on both during a transition interval. The gNodeB periodically transmits RRC configuration and reconfiguration messages carrying the new CSI configuration to user equipment in groups, applies zero padding to precoding matrices to match dimensions across the old and new configurations, and maintains a transition timer (csiConfgGraceTimer) with a value significantly higher than the RRC signaling periodicity. User equipment unable to update continues operating on the old configuration until the timer expires, after which the gNodeB ceases operation on the old configuration and non-updated equipment may announce radio link failure.0017507WGU / 4688

[0019] In another confi uration, the Near-RT RIC and SMO / Non-RT RTC are coupled via standardized E2 and Al interfaces, respectively, allowing the Near-RT RIC to receive enrichment information and policies generated by the SMO / Non-RT RIC. The connection module within the Near-RT RIC receives real-time CSI measurements from the DU, while a normalization and indexing module correlates those measurements with the deployed policies and model outputs. The dashboard module then provides an analyst interface for monitoring per-UE beamforming decisions, CQI / MCS adaptations, and transition status of CSI configurations.

[0020] The above-described and other features and advantages of the present disclosure will be appreciated and understood by those skilled in the art from the following detailed description, drawings, and appended claims.DESCRIPTION OF THE DRAWINGS

[0021] FIG. 1 A is a flow diagram at the SMO / Non-RT RIC showing data collection, AI / ML training, deployment of policies / models, and CSI-RS / feedback configuration updates; it initiates the outer-loop process that provides inputs and artifacts used in FIG. IB.

[0022] FIG. IB is a near-real-time flow between the Near-RT RIC, E2 nodes, and O-RUs showing receipt of DL LI measurements, per-slot inference, and writing per-UE beamforming mode, beam indexes, and CQI / MCS; it consumes the policies, models, and configurations produced in FIG. 1A.

[0023] FIG. 2 is a block diagram of offline AI / ML training in a simulation with a DRL agent choosing CSI / feedback configurations and mode policy, and a CNN-LSTM predicting per-UE PMIs and CQI / MCS over n TTIs; it generates the trained artifacts deployed in FIG. 1A and used operationally in FIG. IB.

[0024] FIG. 3 is a call flow of gNodeB transitioning from an old to a new CSI configuration with concurrent operation, periodic RRC updates, optional UE batching, and zero-padding of precoders; it is triggered by configuration changes written in FIG. 1A and ensures continuity for the near-real-time actions of FIG. IB.DETAILED DESCRIPTION

[0025] The present disclosure relates to systems and methods for dynamic optimization of Channel State Information-Reference Signal (CSI-RS) transmission, CSLfeedback reporting, and user-beam configurations in an Open Radio Access Network (0-RAN) ecosystem for fifthgeneration (5G) mobile networks. The embodiments herein enable coordinated control between a Service Management and Orchestration function and / or a Non-Real-Time RAN Intelligent Controller operating in non-real time and a Near-Real-Time RAN Intelligent Controller operating in near real time, with actions executed at E2 nodes of Distributed Units. By jointly adapting non-real-time and near-real-time parameters, the system addresses inefficiencies inherent in static or semi-static provisioning, including unnecessary overhead, degraded spectral efficiency, increased inter-cell interference, and elevated power consumption.

[0026] In an embodiment, the 0-RAN architecture includes a set of E2 nodes instantiated as Distributed Units that perform lower-layer real-time processing and serve one or more cells. Each DU is coupled to one or more Radio Units that handle radio-frequency operations and to control-plane and user-plane Central Units, and is further integrated to an SMO / Non-RT RIC over the 01 interface and to a Near-RT RIC over the E2 interface. The SMO / Non-RT RIC hosts rApp functions that execute data collection, preprocessing, optimization, and model training. The Near-RT RIC hosts xApp functions responsible for per-slot inference and actuation, and receives policies, trained models, and enrichment information from the SMO / Non-RT RIC via standardized interfaces.

[0027] The system operates in an outer-loop and inner-loop control arrangement. In the outer loop, at non-real-time periodicities such as on the order of minutes to hours, the SMO / Non-RT RIC collects configuration and performance data across multiple cells and trains an AI / ML model suite to select CSI-RS transmission configurations and CSLfeedback reporting configurations that jointly optimize target KPIs while minimizing control and measurement overhead. Once trained, the SMO / Non-RT RIC writes the selected configurations to the DU via O1 / O2 mediated processes and deploys the beamforming-mode policy and the trained per-UE prediction models together with enrichment information to the Near-RT RIC. In the inner loop, at near-real-time periodicities such as every slot or several Transmission Time Intervals, theNear-RT RIC subscribes to and receives downlink Layer 1 measurements from the DU over E2, associates CSLRSRP with the deployed beamforming-mode policy to select codebook-based or non-codebook-based modes per UE, and uses the trained model and enrichment information to infer per-UE slot sequences of codebook-based beam indexes and CQI / MCS indexes. The inferred configurations are then written to the DU over E2 for execution.

[0028] In one embodiment, the non-real-time optimization space for CSLRS transmission includes the number of CSLRS ports, port orthogonality domain selection, port allocation mapping over resource elements within a resource block, transmission periodicity, transmission density including the number of CSLRS signals per resource block, port-to-physicaLantenna mapping and beamforming weights, and the number of CSLRS beams per Synchronization Signal beam. The non-real-time optimization space for CSLfeedback reporting includes enabling or disabling RI-PMI reporting, selecting among supported codebook families such as Type I, Type II, enhanced-Type II and related variations, configuring port selection, and configuring sub-band size for feedback. These non-real-time decisions configure the available user-beam codebook and the corresponding overhead footprint that is reflected in time-frequency resource occupancy.

[0029] In another embodiment, the near-real-time decision space for user-beam configurations includes selecting codebook-based or non-codebook-based beamforming mode per UE according to a policy indexed by CSLRSRP, setting per-slot Precoding Matrix Indices where PMLbased beam management is enabled, setting per-slot P2 beam weights or beam indexes where P1-P2-based beam management is enabled, and setting per-slot CQI and MCS indexes inferred over a sequence of slots. The Near-RT RIC maintains a connection subsystem that receives real-time measurements from E2 nodes, a normalization and indexing subsystem that correlates these measurements against the deployed policies and model states, and a decision subsystem implementing trained inference pipelines. In some embodiments, a dashboard subsystem exposes observability over per-UE decisions, KPI trends, and the transition status of CSI configuration updates.

[0030] In a further embodiment, the system supports the coexistence and transition between an old CSI configuration and a new CSI configuration at the gNodeB to ensure service continuitywhen the SMO / Non-RT RTC updates parameters at the DU. Upon receiving a new CSI configuration, the gNodeB initiates a transition timer with a duration significantly greater than the periodicity of RRC configuration signaling. During this grace interval, the gNodeB concurrently operates with both old and new CSI configurations, periodically transmits RRC configuration and reconfiguration messages that carry the new CSI configuration, and allows UEs that have not yet updated to continue to be served under the old configuration. When the old and new configurations lead to different precoder dimensionalities, the gNodeB applies zero padding at muted-port locations to harmonize precoder dimensions for compatibility. After the transition timer expires, the gNodeB ceases operation under the old configuration and UEs that remain non-updated may declare radio link failure.

[0031] The AI / ML subsystem comprises two complementary techniques. A Deep-Q Learningbased reinforcement learning agent performs joint, two-stage optimization across cells by selecting non-real-time CSLRS and CSLfeedback configurations and a beamforming-mode policy indexed by CSI-RSRP. A CNN-LSTM predictor performs sequence inference to forecast per-UE slot-specific codebook-based beam indexes and CQI / MCS indexes using historical observations. The DRL agent is trained offline in a simulation environment such as a digital twin to avoid performance degradation associated with online exploration. The agent observes a state that may include indicators for whether UEs achieve KPI targets, chooses actions from a discrete action space that enumerates supported configuration tuples per cell and beamforming-mode policy tuples, and receives a reward proportional to the fraction or count of UEs meeting KPI targets. The training employs an r-greedy policy, experience replay, and a deep neural network to estimate Q-values. The CNN-LSTM consumes sequences of historical PMIs, RIs, Lis, CQIs and associated MCS indexes and outputs predictions for the next sequence of slots at a near-realtime horizon.

[0032] In an embodiment where the optimization is expressed as a joint two-stage stochastic program, the first-stage binary decision variables select CSI-RS and CSI-feedback configurations per cell from supported sets, and the second-stage decisions choose per-UE beam indexes and CQI / MCS indexes conditioned on the first-stage selections. The objective minimizes a weighted sum of CSI overhead and the expected negative of the number of UEs achieving KPI targets, subject to feasibility constraints including support of chosen configurations at each cell,availability of beam indexes determined by the first-stage configuration, and bounds derived from the CQI / MCS lookup table supported by the cell. In one formulation, the indicator function of per-UE KPI attainment contributes to the objective’s expectation. A coefficient balances the tradeoff between overhead minimization and KPI maximization. Alternative training strategies that avoid explicit stochastic programming rely on the described DRL agent to implicitly solve this tradeoff by shaping the reward.

[0033] In a practical deployment, the SMO / Non-RT RIC collects configuration management parameters such as supported CSI-RS and CSI-feedback choices per cell, performance measurements including CSI-RS and SSB beam-specific UE measurement reports and per-cell PRB utilization, and KPI targets supplied by the operator. The Near-RT RIC subscribes to downlink Layer 1 measurements from the E2 node, including PMI, RI, LI, CSI-RSRP and CQI / MCS indexes or partial channel feedback in open-loop MIMO configurations. The Near-RT RIC aggregates and reports UE-level data over 01 to the SMO / Non-RT RIC to support rApp training inputs. Following model training, the SMO / Non-RT RIC deploys to the Near-RT RIC a beamforming-mode policy mapping CSI-RSRP to codebook-based or non-codebook-based selection, a trained CNN-LSTM predictor that infers per-UE slot-specific codebook-based beam indexes and CQI / MCS indexes, and enrichment information that includes histories of PMIs, P2 beam indexes, CQIs, and MCS indexes. The Near-RT RIC applies the policy to select beamforming mode per UE, invokes the predictor to infer near-term sequences, and writes per-UE per-slot configurations to the DU for execution over the E2 interface.

[0034] The embodiments herein support a range of parameterizations. For CSI-RS ports, the set of supported values can include 2, 4, 8, 12, 16, 24, and 32. Transmission periodicity can be selected from 5 ms, 10 ms, 20 ms up to 80 ms depending on UE mobility. CSI density can be set to 0.5, 1, or 3 signals per resource block, where a density of 3 may be used for tracking reference signals. Sub-band sizes for feedback can be 4, 8, 16, or 32 physical resource blocks. Codebook selection can favor Type I for near-cell UEs in lower traffic conditions and enhanced-Type II for high-load scenarios or cell-edge UEs to improve resolution. For user-beam strategies, noncodebook-based beamforming that relies on SRS and channel reciprocity may be preferred for UEs with high CSI-RSRP, while codebook-based beamforming may be selected when SRS quality is degraded or for cell-edge scenarios with poor channel conditions. In high-speedmobility conditions, the method can reduce periodicity intervals to track fast channel variation, while stationary or low-speed UEs can benefit from relaxed periodicity to reduce overhead and power consumption.

[0035] The system can be instantiated as software modules deployed on dedicated hardware or virtualized in a cloud-native environment. The SMO / Non-RT RIC can be hosted in a centralized data center with scalable compute to train models, whereas the Near-RT RIC operates closer to the edge to meet near-real-time constraints. The DU and RU can be implemented as hardware appliances or virtualized network functions. The solution supports a broad range of user equipment types and traffic distributions and can ingest multi-cell enrichment information to coordinate cross-cell interference mitigation by adjusting beam configurations with awareness of spatial and temporal traffic patterns.

[0036] The gNodeB transition technique ensures robustness to RRC message losses, intermittent connectivity, or UE-specific reconfiguration failures. During the grace period, the gNodeB maintains compatibility by applying zero padding on precoding matrices to ensure consistent dimensions across old and new configurations. The gNodeB may transmit RRC reconfiguration messages in batches, for example dividing a set of RRC-connected UEs into subgroups and sending reconfiguration messages in staggered transmissions to improve spectrum utilization and mitigate signaling congestion. In variants, the gNodeB dynamically scales the grace timer based on current traffic load or the number of UEs attached and adjusts reconfiguration retries for UEs with poor connectivity. Fallback mechanisms can prioritize non-updated UEs for subsequent reconfiguration attempts and may temporarily boost signaling robustness for those UEs. At timer expiry, the gNodeB completes the transition by ceasing operation under the old configuration; residual non-updated UEs may announce radio link failure and re-establish with the new configuration.

[0037] The described framework provides continuous monitoring and retraining. rApps observe operational KPIs across cells after deployment and, if performance degrades or falls short of targets, initiate a fallback to default configurations and request retraining or updates to the AI / ML models. Upon retraining, updated policies and models are redeployed to the Near-RT0017507WGU / 4688 RIC, and the cycle of non-real-time parameter updates and near-real-time per-UE actuation continues.

[0038] FIG. 1 A illustrates a high-level call flow executed at the SMO / Non-RT RIC and the rApp layer for data collection, preprocessing, model training, configuration deployment, and performance monitoring. Observation and measurement data are collected, including configuration management inputs such as the supported CSI-RS transmission and CSI-feedback reporting configurations per cell, performance measurements that include CSI-RS and SSB beam-specific UE measurements obtained by tracing RRC messages and per-cell PRB usage, and operator-provided KPI targets. The collected data are preprocessed to generate AI / ML training datasets and are supplied to the rApps for model training. Upon completion of training, the SMO / Non-RT RIC issues a request to deploy both the beamforming-mode selection policy and the trained AI / ML model that predicts codebook-based beam indexes and CQI / MCS indexes. Enrichment information, including histories of codebook-based beam indexes and CQI / MCS indexes, is packaged for deployment. The SMO / Non-RT RIC writes or modifies CSI-RS transmission and CSI-feedback reporting configurations at the E2 nodes. The figure further depicts performance monitoring loops in which KPIs, measurement reports, and observations are fed back for evaluation. If performance is unsatisfactory, default or fallback decisions can be enacted, trained models and policies can be deleted, and retraining triggers can be generated to update the AI / ML components. The call flow shows repeated cycles of performance evaluation, fallback, deletion, retraining triggers, and re-deployment to maintain alignment with network dynamics.

[0039] FIG. IB depicts the complementary near-real-time inference and actuation flow executed between the Near-RT RIC, the E2 nodes, and the O-RUs. The E2 nodes provide downlink Layer 1 measurements per UE to the Near-RT RIC, including PMI, RI, LI, CSLRSRP, and CQI / MCS indexes, either periodically or upon events. The Near-RT RIC processes the data and performs AI / ML inference with a near-real-time periodicity, which in some embodiments is less than 5 ms per decision epoch. Based on the deployed beamforming-mode policy and the trained model, the Near-RT RIC writes and modifies, at the E2 nodes, per-UE beamforming mode selections between codebook-based and non-codebook-based, per-UE slot-specific codebook-based beam indexes such as PMIs and P2 beam indexes, and per-UE slot-specific CQI / MCS indexes0017507WGU / 4688 aggregated over a near-term sequence of slots. The E2 nodes propagate the updated configurations to the O-RUs for execution, while operational data continue to be collected and fed back to both the Near-RT RIC and the SMO / Non-RT RIC to close the control loops.

[0040] FIG. 2 shows the AI / ML training arrangement employed at the SMO / Non-RT RIC in conjunction with a simulation environment that can be instantiated as a digital twin. The DRL agent observes a state that indicates, for each UE, whether KPI targets are achieved, and it receives a reward proportional to the number or fraction of UEs achieving the KPI targets. The action space of the DRL agent includes, for each cell, selecting CSLRS transmission and CSL feedback reporting configurations from the supported pool and selecting a tuple comprising a CSLRSRP threshold and a beamforming mode to form a policy mapping for mode selection. The simulation evaluates actions at periodic intervals of n TTIs, where the beamforming mode per UE is selected by associating CSLRSRP with the DRL policy and a CNN-LSTM predictor infers PMIs, RIs, Lis, and CQI / MCS indexes over the sequence of n TTIs. Over a longer horizon of N TTIs with N much greater than n, the simulation environment returns the next state and reward, allowing the DRL agent to update Q-value estimates using sampled mini-batches from an experience replay memory. The agent follows an c-greedy exploration policy, and the deep neural network maps states to Q-values for all possible actions. This offline training yields policies and models that can be deployed to the Near-RT RIC.

[0041] FIG. 3 presents the CSI configuration update procedure at the gNodeB that ensures a smooth transition between an old CSI configuration and a new CSI configuration. The gNodeB starts while operating under a first CSI configuration and transmits RRC messages carrying that configuration to attached UEs. Upon receipt of a new CSI configuration from the SMO / Non-RT RIC framework, the gNodeB begins operating concurrently on both the old and new configurations and initiates a grace timer whose duration is significantly longer than the RRC signaling periodicity. During the grace interval, the gNodeB periodically transmits RRC configuration and reconfiguration messages carrying the new CSI configuration, optionally grouping UEs for staggered reconfiguration to improve spectrum utilization and reduce signaling congestion. The figure includes an alternative branch for the case in which the number of attached UEs is greater than zero and a loop that continues until the grace timer expires.Throughout the loop, UEs that successfully receive and apply the new configuration operateunder the new configuration, while UEs that fail to update continue to be served under the old configuration. Where necessary, the gNodeB applies zero padding to the precoding matrix to harmonize matrix dimensions between the old and new configurations. When the grace timer expires, the gNodeB stops operating under the old configuration. UEs that remain on the old configuration at that time can announce radio link failure and reattach under the new configuration. Variations include dynamically adjusting the grace timer according to network conditions and applying targeted reconfiguration retries for UEs with poor connectivity to maximize successful updates.

[0042] The examples and embodiments described herein are provided for illustrative purposes only and are not intended to limit the scope of the described subject matter. Certain details, such as widely recognized principles of telecommunications and 0-RAN architecture, can be omitted for brevity and clarity. Furthermore, the described subject matter includes various modifications, rearrangements, and adaptations of the configurations, methods, and systems discussed, provided they fall within the scope of the appended claims. Those skilled in the art will recognize that the described subject matter can be applied to a wide range of applications and implementations without deviating from its underlying principles.

[0043] In one embodiment, the Near-Real-Time RAN Intelligent Controller (Near-RT RIC) and the Service Management Orchestration (SMO) and / or Non-Real -Time RAN Intelligent Controller (Non-RT RIC) are implemented as software modules running on dedicated hardware within the 0-RAN ecosystem, where the SMO / Non-RT RIC operates at a centralized data center and the Near-RT RIC operates closer to the edge of the network. In another embodiment, the SMO / Non-RT RIC and Near-RT RIC are virtualized and deployed on cloud-native platforms, allowing for dynamic scaling and resource allocation based on network demands. The configurations of Channel State Information-Reference Signal (CSI-RS) transmission and CSI-feedback reporting may vary, such as adjusting the number of CSI-RS ports to values like 4, 8, or 16, or modifying the transmission periodicity to intervals of 10ms, 20ms, or 40ms, depending on the traffic load and user equipment (UE) speed. In yet another embodiment, the user-beam configurations are dynamically adjusted by the Near-RT RIC, where codebook-based beamforming is used for cell-edge UEs with poor channel quality, while non-codebook-based beamforming is employed for near-cell UEs with strong channel conditions. Additionally, thesystem may utilize different AI / ML models for optimization, such as a Deep-Q Learning-based model for reinforcement learning or a CNN-LSTM model for predicting per-UE slot-specific parameters, depending on the complexity and real-time requirements of the network. The E2 node of the Distributed Unit (DU) can also support various hardware configurations, such as FPGA-based accelerators or general-purpose processors, to handle the computational demands of modifying CSLRS and user-beam configurations. These embodiments ensure adaptability and efficiency in optimizing the O-RAN for diverse 5G network scenarios.

[0044] By providing a Near-Real-Time RAN Intelligent Controller (Near-RT RIC) and a Service Management Orchestration (SMO) and / or Non-Real-Time RAN Intelligent Controller (Non-RT RIC), the system enables a dual-layered control mechanism that separates non-real-time and near-real-time operations. This arrangement ensures that configurations of Channel State Information-Reference Signal (CSI-RS) transmission and CSLfeedback reporting are optimized at non-real-time intervals, while user-beam configurations and Channel Quality Indicator (CQI) and / or Modulation Coding Scheme (MCS)-indexes are adjusted in near-real-time. This separation allows for efficient resource allocation and adaptation to dynamic network conditions.

[0045] Modifying CSI-RS transmission and CSI-feedback reporting configurations at the E2 node of a Distributed Unit (DU) ensures that the system can adapt to variations in traffic load, user mobility, and channel conditions. For example, the number of CSI-RS ports, transmission periodicity, and codebook configurations can be dynamically adjusted to match the real-time requirements of the network, reducing resource wastage caused by static provisioning and improving spectral efficiency.

[0046] Adjusting user-beam configurations and CQI / MCS-indexes for individual User Equipment (UE) in near-real-time ensures that beamforming strategies and modulation schemes are tailored to the specific channel conditions of each UE. This improves data throughput, reduces retransmissions, and minimizes power consumption for both the network and the UEs.

[0047] The integration of the Near-RT RIC and SMO / Non-RT RIC within the O-RAN architecture provides a scalable and flexible solution that aligns with the temporal and spatial variability of modem 5G networks. This approach addresses inefficiencies associated with staticconfigurations, enhances spectral efficiency, mitigates inter-cell interference, and reduces power consumption, thereby improving overall network performance and user experience.

[0048] In one embodiment, the modification of CSI-RS transmission configurations can involve adjusting the number of ports to optimize beam resolution based on traffic load, such as using 4 ports for low traffic scenarios and 32 ports for high traffic scenarios. The port allocation over Resource Elements (REs) in a Resource Block (RB) can be dynamically mapped to minimize interference, with orthogonality achieved in the time, frequency, or code domain. Transmission periodicity can be tailored to UE speed, such as 5ms for high-speed UEs and 20ms for low-speed UEs, ensuring efficient resource utilization. Transmission density can be varied, such as using a density of 0.5 for standard CSI-RS signals and 3 for Tracking Reference Signal transmission. The mapping of CSI-RS ports to physical antenna ports can be optimized using AI / ML models to dynamically adjust beamforming weights based on real-time channel conditions. The number of CSI-RS beams per Synchronization Signal (SS) beam can be configured to match UE distribution, such as increasing the number of beams in areas with high UE density to improve coverage and reduce intercell interference. In another embodiment, the modification of CSI-feedback reporting configurations can include enabling or disabling codebook configuration parameters based on channel quality, such as using Type I codebooks for near-cell UEs and enhanced-Type II codebooks for cell-edge UEs. Sub-band size can be adjusted dynamically, such as using smaller sub-bands for environments with high coherence bandwidth to improve feedback granularity. These embodiments ensure adaptability to network dynamics, enhancing spectral efficiency, reducing power consumption, and improving overall network performance.

[0049] By dynamically modifying the number of CSI-RS ports, the system can optimize beam resolution based on traffic load and user distribution, ensuring efficient utilization of radio resources. For instance, fewer ports may be allocated during low traffic conditions, while a higher number of ports can be used during high traffic scenarios to support multi-user MIMO configurations.

[0050] Adjusting the port allocation over Resource Elements (REs) in a Resource Block (RB) minimizes inter-cell interference and improves spectral efficiency. This dynamic mappingensures orthogonality in the time, frequency, or code domain, reducing signal overlap and enhancing channel measurement accuracy.

[0051] Modifying transmission periodicity based on UE mobility allows the system to adapt to varying coherence times. High-speed UEs benefit from shorter periodicity intervals to maintain accurate channel state information, while low-speed UEs can operate with longer intervals, reducing overhead and power consumption.

[0052] Dynamically varying the transmission density or the number of CSI-RS signals per RB ensures that the system can allocate resources effectively for different scenarios, such as using higher density for tracking reference signals in environments with rapid channel variations.

[0053] Optimizing CSI-RS port-to-physical antenna port mapping or beamforming weights enables the system to adapt beam configurations to real-time channel conditions, improving signal quality and reducing power consumption.

[0054] Adjusting the number of CSI-RS beams per Synchronization Signal (SS) beam based on UE distribution ensures that resources are allocated efficiently, reducing unnecessary overhead in areas with low user density and enhancing coverage in high-density regions.

[0055] Modifying codebook configuration parameters dynamically allows the system to select appropriate codebooks based on channel quality. For example, Type I codebooks can be used for near-cell UEs with strong channel conditions, while enhanced-Type II codebooks may be preferred for cell-edge UEs to improve resolution and performance.

[0056] Adjusting sub-band size for feedback reporting based on coherence bandwidth ensures that feedback granularity matches the channel environment, improving the accuracy of CSI reports and reducing resource wastage.

[0057] These modifications collectively enhance spectral efficiency, reduce inter-cell interference, and minimize power consumption, addressing the limitations of static provisioning in 5G networks.

[0058] In one embodiment, the Artificial Intelligence / Machine Learning (AI / ML) model is trained using Deep-Q Learning-based Reinforcement Learning (DRL) to optimize configurationsof CSI-RS transmission, CST-feedback reporting, and user-beam strategies. The DRL agent evaluates actions in a simulated environment, such as a Digital Twin, where the agent selects configurations for CSI-RS transmission and feedback reporting, as well as beamforming modes (Codebook-based or non-Codebook-based) based on CSI-RSRP values. In another embodiment, the AI / ML model incorporates CNN-LSTM-based prediction learning to infer per-UE slotspecific Codebook-based beam indexes, such as Precoding Matrix Indices (PMIs) and P2 beam indexes, along with CQIs and MCS-indexes for a sequence of slots. The CNN-LSTM model processes historical data, including PMIs, RI, LI, and CQIs / MCS-indexes, to predict future configurations dynamically. In yet another embodiment, the AI / ML model is trained offline using a simulation environment where the DRL agent employs an cc-greedy policy to balance exploration and exploitation, ensuring effective configuration decisions. The training process involves storing experience tuples in a replay memory and using deep neural networks to predict Q-values for various actions. Additionally, the AI / ML model can be deployed in real-time to adapt configurations based on network dynamics, such as UE speed, traffic load, and angular spread, ensuring enhanced Performance Indicators (Pls). In a further embodiment, the AI / ML model is designed to optimize the tradeoff between minimizing overhead and improving Pls by incorporating design coefficients and constraints, such as supported configurations per cell and CQI / MCS lookup tables. These embodiments demonstrate the adaptability of the AI / ML model to various network conditions and operational requirements, ensuring efficient and dynamic optimization of CSI-RS transmission, CSI-feedback reporting, and user-beam configurations.

[0059] The integration of an Artificial Intelligence / Machine Learning (AI / ML) model trained using Deep-Q Learning-based Reinforcement Learning and CNN-LSTM-based prediction learning enables dynamic optimization of configurations for CSI-RS transmission, CSI-feedback reporting, and user-beam strategies. By leveraging Deep-Q Learning, the system can evaluate and select optimal configuration tuples for CSI-RS and CSI-feedback reporting across multiple cells, while simultaneously determining beamforming mode policies indexed by CSI-RSRP. This ensures that the configurations are tailored to real-time network conditions, such as traffic load, UE mobility, and channel quality, thereby reducing resource wastage and improving spectral efficiency.0017507WGU / 4688

[0060] The CNN-LSTM model processes historical CST-feedback data, including PMIs, RIs, Lis, CQIs, and MCS-indexes, to predict per-UE slot-specific configurations dynamically. This predictive capability allows the system to anticipate future channel conditions and adapt beamforming strategies accordingly, ensuring that UEs receive optimal signal quality and throughput. The combination of reinforcement learning and sequence prediction minimizes the time complexity of configuration updates, aligning the optimization process with the fastchanging dynamics of modern 5G networks.

[0061] By training the AI / ML model offline in a simulation environment, such as a digital twin, the system avoids performance degradation associated with online exploration. The offline training process ensures that the model is robust and capable of handling diverse network scenarios, including high-speed mobility, varying traffic loads, and angular spread differences. This approach enhances the reliability and scalability of the optimization framework, enabling efficient deployment in large-scale 0-RAN ecosystems.

[0062] Overall, the use of AI / ML models for optimizing CS RS transmission, CSLfeedback reporting, and user-beam configurations improves spectral efficiency, reduces inter-cell interference, and minimizes power consumption, while ensuring that the network adapts seamlessly to temporal and spatial variations in user and channel conditions.

[0063] In one embodiment, the gNodeB (gNB) operates on both the old and new CSI configurations during the transition interval, ensuring compatibility by applying zero padding to the precoding matrix to align the dimensions of the configurations. In another embodiment, the gNB transmits RRC configuration and reconfiguration messages carrying the new CSI configuration to UEs in staggered groups, such as dividing UEs into smaller batches to optimize spectrum utilization and reduce signaling congestion. In yet another embodiment, the transition timer (csiConfgGraceTimer) is configured to a value significantly higher than the RRC signaling periodicity, allowing UEs multiple opportunities to receive and process the reconfiguration messages. Additionally, in one variation, the gNB dynamically adjusts the timer value based on network conditions, such as the number of UEs connected or the traffic load, to further enhance the transition process. In a further embodiment, UEs that fail to update their configurations during the transition interval continue to operate under the old CSI configuration, ensuringuninterrupted service, and only after the timer expires does the gNB cease support for the old configuration, at which point such UEs may announce radio link failure (RLF). This approach ensures a smooth and efficient transition while minimizing service disruptions and maintaining network performance.

[0064] Operating concurrently on both the old and new CSI configurations during the transition interval ensures service continuity for UEs that have not yet updated their configurations. This arrangement prevents abrupt service disruptions and allows UEs to continue functioning under the old configuration while the new configuration is gradually introduced.

[0065] Periodic transmission of RRC configuration and reconfiguration messages carrying the new CSI configuration provides multiple opportunities for UEs to receive and process the updated configuration. Grouping UEs for staggered reconfiguration reduces signaling congestion and optimizes spectrum utilization, ensuring efficient use of network resources during the transition.

[0066] Applying zero padding to the precoding matrix harmonizes the dimensions of the old and new CSI configurations, enabling compatibility between the two configurations. This prevents errors in beamforming operations and ensures consistent signal quality during the transition.

[0067] Maintaining a transition timer with a value significantly higher than the RRC signaling periodicity allows UEs sufficient time to update their configurations. UEs that fail to update during the transition interval are served under the old configuration, ensuring uninterrupted service. Once the timer expires, the gNodeB ceases operation under the old configuration, and UEs that remain non-updated may announce radio link failure, enabling the network to identify and address connectivity issues effectively.

[0068] In one embodiment, the gNodeB (gNB) operates concurrently on both the old and new CSI configurations during the transition interval, ensuring seamless service continuity for user equipments (UEs) that have not yet updated to the new configuration. In another embodiment, the gNB transmits radio resource control (RRC) configuration and reconfiguration messages carrying the new CSI configuration in batches, where UEs are grouped based on criteria such as signal strength, location, or priority, to optimize spectrum utilization and reduce signalingoverhead. In yet another embodiment, the gNB applies zero padding to the precoding matrix to align the dimensions of the old and new CSI configurations, ensuring compatibility and preventing performance degradation during the transition. Additionally, the transition timer (csiConfgGraceTimer) can be configured with varying durations depending on network conditions, such as traffic load or UE density, to provide sufficient time for UEs to update their configurations. In a further embodiment, the gNB can implement adaptive signaling strategies, such as increasing the frequency of RRC reconfiguration messages for UEs with poor connectivity, to maximize the likelihood of successful updates. For UEs unable to update their configurations within the transition interval, the gNB can employ fallback mechanisms, such as temporarily boosting signal strength or prioritizing these UEs in subsequent reconfiguration attempts, before ceasing operation on the old CSI configuration and allowing these UEs to announce radio link failure (RLF).

[0069] By serving UEs under the old CSI configuration during the transition interval, uninterrupted service is ensured for UEs that fail to update their configurations, preventing abrupt service disruptions and maintaining network stability.

[0070] The use of a transition timer with a value significantly higher than the RRC signaling periodicity provides UEs multiple opportunities to receive and process reconfiguration messages, increasing the likelihood of successful updates and reducing the risk of service interruptions.

[0071] After the transition timer expires, the gNB ceases operation on the old CSI configuration, allowing the network to identify UEs that remain non-updated. These UEs may announce radio link failure (RLF), enabling the network to effectively manage connectivity issues and ensure that only updated UEs operate under the new configuration.

[0072] This approach minimizes the risk of network performance degradation during configuration transitions by ensuring compatibility between old and new configurations, thereby maintaining service continuity and improving the overall reliability of the system.

[0073] The foregoing detailed description illustrates exemplary embodiments of the disclosed system and methods. It will be understood that certain functions can be performed in an order different from that shown or can be combined or separated without departing from the scope ofthe claimed subject matter. The described modules can be implemented in software, hardware, or any combination thereof, and functional partitions can be altered to suit deployment requirements. The scope of the disclosure encompasses all alternatives, modifications, and equivalents as can be included within the spirit of the claims.

Claims

CLAIMSWhat is claimed is:

1. A method for optimizing configurations in an Open Radio Access Network (O-RAN) for 5G-based mobile networks, the method comprising the steps of:providing a Near-Real-Time RAN Intelligent Controller (Near-RT RIC);providing a Service Management Orchestration (SMO) and / or Non-Real-Time RAN Intelligent Controller (Non-RT RIC);modifying, with the SMO and / or Non-RT RIC at an E2 node of a Distributed Unit (DU), configurations of Channel State Information-Reference Signal (CSLRS) transmission and CSL feedback reporting for at least one cell; andmodifying, with the Near-RT RIC at the E2 node of the DU, user-beam configurations and Channel Quality Indicator (CQI) and / or Modulation Coding Scheme (MCS)-indexes for at least one User Equipment (UE).

2. The method of claim 1,wherein the step of modifying of the CSl-RS transmission further includes modifying:a number of ports, a port allocation over Resource Elements (REs) in a Resource Block (RB), a transmission periodicity, a transmission density or number of CSLRS signals per RB, a CSLRS port to physical antenna port mapping or beamforming weights, a number of CSLRS beams per Synchronization Signal (SS) beam, and combinations thereof; andwherein the step of modifying of the CSLfeedback reporting further includes modifying:codebook configuration parameters, a sub-band size, and combinations thereof.

3. The method of claim 2, wherein the step of modifying of the user-beam configurations further includes modifying:codebook-based or non-Codebook-based beamforming mode, per-slot Precoding Matrix Indices (PMIs) when PMI-based beam management is enabled, per-slot P2 beamformingweights or beam-indexes when Pl-P2-based beam management is enabled, CQIs or MCS-indexes, and combinations thereof.

4. The method of claim 3, further comprising an Artificial Intelligence / Machine Learning (AI / ML) model, the method further comprising the step of:training the AI / ML model to optimize configurations of CS RS transmission per cell, configurations of CS feedback reporting per cell, a policy to select user-beamforming mode based on CS RSRP, and per-UE slot-specific Codebook-based beam indexes and CQIs or MCS-indexes, wherein the training includes Deep-Q Learning-based Reinforcement Learning and CNN-LSTM-based prediction learning.

5. The method of claim 4, wherein the step of the SMO and / or Non-RT RIC modifying configurations of CSLRS transmissions and CSLfeedback reporting occurs after the AI / ML model training is completed.

6. The method of claim 5, wherein the SMO and / or Non-RT RIC deploys the policy for beamforming mode selection to the Near-RT RIC and deploys the trained AI / ML model to predict the per-UE slot-specific Codebook-based beam indexes and CQIs or MCS-indexes to the Near-RT RIC.

7. The method of claim 6, wherein the SMO and / or Non-RT RIC deploys enrichment information to the Near-RT RIC, the enrichment information selected from the group consisting of: a history of PMIs, P2 beam indexes, CQIs, MCS-indexes, and combinations thereof.

8. The method of claim 6, further comprising the steps of:configuring the beamforming mode with the Near-RT RIC per UE by associating CSI-RSRP with a policy received from the SMO and / or Non-RT RIC; andby associating the CSI-feedbacks with the trained AI / ML model received from the SMO and / or Non-RT RIC, the Near-RT RIC inferring per-UE slot-specific Codebook-based beam indexes and CQIs or MCS-indexes for a sequence of slots.

9. A method implemented at a gNodeB (gNB) to ensure a smooth transition between an old CSI configuration and a new CSI configuration, the method comprising the steps of:operating, during a transition interval, concurrently on the old and the new CSI configurations;periodically transmitting radio resource control (RRC) configuration and reconfiguration messages carrying the new CSI configuration, including sending reconfiguration messages to user equipments (UEs) in groups;applying zero padding to a precoding matrix to match dimensions across the old and the new CSI configurations; andmaintaining a transition timer (csiConfgGraceTimer) having a value significantly higher than the RRC signaling periodicity, so that UEs unable to update operate on the old CSI configuration until the transition timer expires.

10. The method of claim 9, whereinUEs that are unable to update their CSI configurations during the transition interval are served under the old CSI configuration until the transition timer expires; andafter which the gNB stops operating on the old CSI configuration and said UEs may announce radio link failure (RLF).

Images