AI/ML BASED SS BURST SET & CSI-RS TRS CONFIGURATION OPTIMIZATION AND IMPROVING NR NETWORK POWER AND SPECTRAL EFFICIENCY

Abstract

A system for optimizing one of a 5G New Radio (NR) network or a 4G Long Term Evolution (LTE) network operation, includes: an artificial intelligence (AI) engine configured to: a) implement at least one of the following: i) collection of raw training data set; ii) training of an AI agent; iii) inference generation triggered by UE; iv) using at least one of network key performance indicator (KPI) report and network operator inputs; and v) network KPI collection after application of inferences; and b) at least one of infer and apply at least one of i) network-wide optimal per-gNB or per transmission/reception point (TRP) Synchronization Signal (SS) burst set, and ii) associated channel state information reference signal (CSI-RS) configurations for downlink (DL) reference signal transmissions, thereby enhancing at least one of network transmission power efficiency and spectral efficiency.

Claims

1. A system for optimizing one of a 5G New Radio (NR) network or a 4G Long Term Evolution (LTE) network operation, comprising: an artificial intelligence (AI) engine configured to: a) implement at least one of the following: i) collection of raw training data set; ii) training of an AI agent; iii) inference generation triggered by UE; iv) using at least one of network key performance indicator (KPI) report and network operator inputs; and v) network KPI collection after application of inferences; and b) at least one of infer and apply at least one of i) network-wide optimal per-gNB or per-transmission/reception point (TRP) Synchronization Signal (SS) burst set, and ii) associated channel state information reference signal (CSI-RS) configurations for downlink (DL) reference signal transmissions, thereby enhancing at least one of network transmission power efficiency and spectral efficiency.

2. The system according to claim 1, further comprising: a core network entity including at least one of Service Management and Orchestration (SMO) element, Non-Real Time (Non-RT) Radio Access Network Intelligent Controller (RIC), and Near Real Time (Near-RT) RIC configured for processing the collected raw training data set over one of 3GPP or Open Radio Access Network (O-RAN) defined interface to produce a processed training data set.

3. The system according to claim 2, wherein the raw training data set is generated by a network access node including at least one of TRP, gNB, distributed unit (DU), and centralized unit (CU).

4. The system according to claim 2, wherein the raw training data set includes at least one of: physical random-access channel (PRACH) receive beam index, number of PRACH instances exceeding energy threshold for the beam index, number of connected user equipments (UEs) in one of the TRP or the gNB over a time window of observation, and UE-reported reference signal received power (RSRP).

5. The system according to claim 4, wherein the core network entity is configured to process the collected raw training data set by applying a deep neural network technique to define implicit relationship of the collected raw training data set with at least one of: UE mobility; deployment-dependent angular spread; observed interferences; time-dependent network load or network usage pattern; geographical location-dependent network usage; and RSRP.

6. The system according to claim 2, wherein the training of the AI agent includes using the processed training data set to derive optimal interferences on reference signal (RS) configuration including at least one of optimal SS burst set configuration, optimal CSI-RS configuration, and optimal joint SS burst set and CSI-RS configuration for each one of TRP or gNB in the network based on minimizing {SS beam span, number of SS blocks in SS burst set, Number of CSI-RS followed} dependent cost functions.

7. The system according to claim 2, wherein at least one of i) applying and communicating the generated inference by the AI engine to at least one receptor node, and ii) collecting the network KPIs is implemented via a 3GPP-defined interface between the core network entity and the access network node.

8. The system according to claim 6, wherein the optimal SS burst set configuration is for maximizing initial access latency (IAL) up to a selected target IAL, and wherein the optimal SS burst set configuration is derived based on at least one of the following: a) based on the gNB physical random access channel (PRACH) detection energy signal-to-noise-ratio (SNR) threshold, compute {reception (Rx) beam ID, number of UE’s PRACH correlation magnitude exceeding SNR threshold} for an observation time window; b) collect information on number of connected UEs with the gNB for the observation time windows; and c) infer UE mobility based on multiple windows of observations.

9. The system according to claim 4, wherein the training of the AI agent includes using the processed training data set to derive optimal interferences on reference signal (RS) configuration including at least one of optimal SS burst set configuration, optimal CSI-RS configuration, and optimal joint SS burst set and CSI-RS configuration for each one of TRP or gNB in the network based on minimizing {SS beam span, number of SS blocks in SS burst set, Number of CSI-RS followed} dependent cost functions.

10. The system according to claim 9, wherein the optimal SS burst set configuration is for maximizing initial access latency (IAL) up to a selected target IAL, and wherein the optimal SS burst set configuration is derived based on at least one of the following: a) based on the gNB physical random access channel (PRACH) detection energy signal-to-noise-ratio (SNR) threshold, compute {reception (Rx) beam ID, number of UE’s PRACH correlation magnitude exceeding SNR threshold} for an observation time window; b) collect information on number of connected UEs with the gNB for the observation time windows; and c) infer UE mobility based on multiple windows of observations.

11. A method of optimizing one of a 5G New Radio (NR) network or a 4G Long Term Evolution (LTE) network operation, comprising: implementing, using an artificial intelligence (AI) engine, at least one of the following: i) collection of training data; ii) training of an AI agent; iii) inference generation triggered by UE; iv) using at least one of network key performance indicator (KPI) report and network operator inputs; and v) KPI collection after application of inferences; and at least one of inferring and applying, using the AI engine, at least one of i) network-wide optimal per-gNB Synchronization Signal (SS) burst set ,and ii) associated channel state information reference signal (CSI-RS) configurations for downlink (DL) reference signal transmissions, thereby enhancing at least one of network transmission power efficiency and spectral efficiency.

12. The method according to claim 11, further comprising: processing, by a core network entity including at least one of Service Management and Orchestration (SMO) element, Non-Real Time (Non-RT) Radio Access Network Intelligent Controller (RIC), and Near Real Time (Near-RT) RIC, the collected raw training data set over one of 3GPP or Open Radio Access Network (O-RAN) defined interface.

13. The method according to claim 12, wherein the raw training data are generated by an access node including at least one of transmission/reception point (TRP), gNB, distributed unit (DU), and centralized unit (CU).

14. The method according to claim 12, wherein the raw training data set includes at least one of: physical random-access channel (PRACH) receive beam index, number of PRACH instances exceeding energy threshold for the beam index, number of connected user equipments (UEs) in one of the TRP or the gNB over a time window of observation, and UE-reported reference signal received power (RSRP).

15. The method according to claim 14, wherein the processing of the collected raw training data set includes applying a deep neural network technique to define implicit relationship of the collected raw training data set with at least one of: UE mobility, deployment-dependent angular spread, observed interferences, time-dependent network load or network usage pattern, geographical location-dependent network usage, and RSRP.

16. The method according to claim 12, wherein the training of the AI agent includes using the processed training data set to derive optimal interferences on reference signal (RS) configuration including at least one of optimal SS burst set configuration, optimal CSI-RS configuration, and optimal joint SS burst set and CSI-RS configuration for each one of TRP or gNB in the network based on minimizing {SS beam span, number of SS blocks in SS burst set, Number of CSI-RS followed} dependent cost functions.

17. The method according to claim 12, wherein at least one of i) applying and communicating the generated inference by the AI engine to at least one receptor node, and ii) collecting the network KPIs is implemented via a 3GPP-defined interface between the core network entity and the access network node.

18. The method according to claim 16, wherein the optimal SS burst set configuration is for maximizing initial access latency (IAL) up to a selected target IAL, and wherein the optimal SS burst set configuration is derived based on at least one of the following: a) based on the gNB physical random access channel (PRACH) detection energy signal-to-noise-ratio (SNR) threshold, compute {reception (Rx) beam ID, number of UE’s PRACH correlation magnitude exceeding SNR threshold} for an observation time window; b) collect information on number of connected UEs with the gNB for the observation time windows; and c) infer UE mobility based on multiple windows of observations.

19. The method according to claim 14, wherein the training of the AI agent includes using the processed training data set to derive optimal interferences on reference signal (RS) configuration including at least one of optimal SS burst set configuration, optimal CSI-RS configuration, and optimal joint SS burst set and CSI-RS configuration for each one of TRP or gNB in the network based on minimizing {SS beam span, number of SS blocks in SS burst set, Number of CSI-RS followed} dependent cost functions.

20. The method according to claim 19, wherein the optimal SS burst set configuration is for maximizing initial access latency (IAL) up to a selected target IAL, and wherein the optimal SS burst set configuration is derived based on at least one of the following: a) based on the gNB physical random access channel (PRACH) detection energy signal-to-noise-ratio (SNR) threshold, compute {reception (Rx) beam ID, number of UE’s PRACH correlation magnitude exceeding SNR threshold} for an observation time window; b) collect information on number of connected UEs with the gNB for the observation time windows; and c) infer UE mobility based on multiple windows of observations.

Description

BRIEF DESCRIPTION OF THE FIGURES

[0015] FIG. 1 illustrates the details of an example embodiment of the SS Burst set.

[0016] FIG. 2 is a flowchart illustrating an example methodology for training for a set of gNBs.

[0017] FIG. 3 is a flowchart illustrating an example process for learning the optimal SS Burst Configuration for maximizing the initial access latency.

DETAILED DESCRIPTION

[0018] According to an example embodiment of the present disclosure, an optimization method is provided to save time and frequency resources while maximizing the target key performance indicator (KPI), e.g., initial access time, limited to T.sub.IA .sub._.sub.max. According to an example embodiment of the present disclosure, instead of deploying the networks with theoretically obtained parameter settings, the disclosed method uses network-wide optimizations of the parameters based on AI/ML algorithms to drive ({S.sub.D, N.sub.ssh:t,.sub.z,....sub.ltsector )9.sub.NB/or each gNB.sub.i configurations by exploiting UE mobility, use case, time of the day for the configuration parameters. In the context of this proposed optimization framework, the following relationships are defined by the Al/ML, agent:

[00005]$S_{D_k}:=f\left(deploymentdependentangularspread;Interference;\mathrm{..}\right)$

[00006]$N_{S{S}_k}:=f\left(UEMobilityStatistics;NumberofConnectedUsers;\mathrm{..}\right)$

[0019] In the above expressions (collectively referenced as (3.0)), f(x) is a linear function of variable x (see, e.g., Equations (1.0) and (2.0) above). Thus, the goal of the AI/ML agent is to identify functional relationships for S.sub.D and N.sub.ss with deploymentdependent parameters using the observations available at gNB with an optimization target for initial access time KPI. The more independent observations are fed to the AI/ML, agent, the better the result of functional relationship establishment. An example list of observations available at gNB (which list is not intended to be limiting for the example method according to the present disclosure) is provided below: [0020] a) Receive beam number (beam index) for PRACH and the statistics for the detected PRACH transmissions sets; [0021] b) Cell-specific and beam-specific RSRP of the received PRACH signals at gNB and statistics of SNR threshold T, measurement of PRACH correction power; [0022] c) Number of received PRACH signals for a receive beam ID at sample time instant; [0023] d) UE mobility or geolocation information (direct feedback or Xapp based); [0024] e) Number of detected PRACH transmission in every UL Receive Beams at gNB; [0025] f) Other relevant performance measurements, e.g., those which deal with PHY-layer beam measurements (such as those mentioned in 3GPP TS 28.552); and [0026] g) RSRP reported by the UEs for any other DL reference signal transmissions.

[0027] The following portions of the present disclosure are directed to the optimization framework (problem) to be solved by the AI/ML agent to derive optimal values for each TRP. This optimization framework (problem) can be extended further to optimize CSI-RS resource allocation following each SS Burst Set transmission in the network. Furthermore, the AI/ML model can include PRACH collision and detection failure to achieve more realistic network optimization.

[0028] To optimize the network level transmission power and improve spectral efficiency, the optimization problem can be formally defined over a set of observation windows {T.sub.w} at the gNB as follows:

[00007]$\begin{array}{l}\underset{{\eta }_1,{\eta }_2.Math.{\eta }_{SSBeamidx}}{\min }\left\{S_D\forall SSBeamswith\#ofRACHdetection\ge N_{th}\right\}\\ Overall\left\{T_w\right\}\&\end{array}$

[00008]$\underset{\left\{{}^{\upsilon }\text{}{\eta }_1{\upsilon }_{{\eta }_2},.Math.{\upsilon }_{{\eta }_{SSBeamIdx}}{,}^{.N\text{}}\text{}U{E}_{Connected\_Tw}\right\}}{\min }\left\{\text{}{N}_{SS}\text{}\le \text{}{N}_{S{S}_{3GPPmax}}\right\}\text{}overall\text{}\left\{T_W\right\}$

[00009]$\begin{array}{l}\text{such}\text{that}\underset{S_{D,}{N}_{SS}}{max}{T}_{IA}\\ T_{IA}\le T_{IA\_max}\end{array}$

In the above expressions (collectively referenced as (4.0)),, T.sub.IA_.sub.max is the maximum IA latency stipulated by the network operator for satisfactory network operation without degrading the KPI of the network; ƞ.sub.1, ..., ƞ.sub.j, . . ƞ.sub.SSBeamIdx are the SS beam index at the gNB (receive beam index for PRACH); V is the inferred UE mobility indicator in the direction of beam ID ƞ.sub.j; NI.sub./Econnected .sub.rw is the number of connected UEs at the observation time window T.sub.W; N.sub.th is the minimum number of PRACH instance detected in a particular Rx beam direction at the gNB to be considered for above optimization problem.

[0029] In this section, PRACH receive power and detected PRACH instances at the gNB will be discussed Assuming p.sub.UE is the transmitted power of the UE for each PRACH transmissions, then under the Poisson Point Process (PPP) distribution of the UE and the gNB, the measured SNR at the gNB from UE.sub.i can be expressed as follows:

[00010]$SN{R}_i=\frac{PU{E}^{G}UE,i|h_i{|}^{2}P{L}_i{S}_i}{{\sigma }^2}$

In the above expression, σ.sup.2 is the normalized noise power and S.sub.i is the LOS condition indicator for the UE in consideration, h is the channel fading function, G.sub.UE,i is the directional gain of the main lobe of the beam, and PL.sub.i is the pathloss factor for UE.sub.i to the target gNB. The gNB successfully detects the UE.sub.i when the received SNR (or correlation power) of the received PRACH achieves a signal-to-noise ratio (SNR) threshold T and the gNB detects the PRACH signal. For simplicity of analysis, such an event can be expressed as the following without considering the PRACH collision:

[00011]$\text{I}\left\{ma{x}_{xi\in \Phi }SN{R}_i\ge T\right\}=1:PRACHDetetcionSuccess$

[00012]$\text{I}\left\{ma{x}_{xi\in \Phi }SN{R}_i<T\right\}=0:PRACHDetectionFailure,$

where I{.square-solid.} is the indicator function. PRACH detection failure can happen either because BS.sub.i is not transmitting SS block in the UE receive beam direction, or the received signal power at the UE does not meet the detection threshold T. Based on the above indicator function output, the gNB can count the number of PRACH receptions from the UEs in the receive beam direction ƞ.sub.j.

[0030] According to an example embodiment of the present disclosure, the first task (“Action 1”) of the AI/ML agent in the optimization process is to establish (through learning) the functional relationship between angular spans {Δ.sub.θ, A.sub.Ø} (hence S.sub.D) and select the set of receive beam directions where detected PRACH transmissions exceed N.sub.th which is formulated as hypothesis in (4.0)

[0031] Observations at each gNB/TRP is divided over a set of observation time window {T.sub.w; w ≥ 1} in a typical day (or any predefined time span - weekdays/weekend, etc.) of operation. In each T.sub.w, the gNB can log, e.g., the following example information (but the method according to the present disclosure is not limited to these example information items): [0032] a) Number of PRACH detections from respective receive beam directions/IDs; [0033] b) Average number of PRACH detections over all receive beam IDs of interest (based on Action 1); [0034] c) Number of connected UEs (in RRC Connected state) served. Rate of change and number of PRACH detections over the set of observation time window {T.sub.w; w ≥ 1} can be used as an indirect measure of the UE mobility in the cell. The AI/ML agent can learn the mobility pattern based on the above observations at the gNB.

[0035] According to an example embodiment of the present disclosure, the second task (“Action 2”) of the AI/ML agent in the optimization process is to find the functional relationship among the derived UE mobility information, number of connected UEs in the gNB with N.sub.SS, and the number of SS blocks in the SS Burst Set (upper bounded by NSS3.sub.CPPmax) as per (4.0).

[0036] By implementing the above-described first and second tasks, the AI/ML agent (e.g., reinforcement learning (RL) agent) learns from the gNB observations and derives optimal set of values {S.sub.D, N.sub.ss} to maximize the value of T.sub.IA which is upper bounded by the operator-specified maximum value. In each step, AI/ML agent action will be rewarded for the decisions which jointly minimizes {S.sub.D, N.sub.SS} with restriction

[00013]$T_{IA}\le T_{IA\_ma\text{}x.Math.}$

[0037] Joint optimization of {S.sub.D, N.sub.ss} with restriction T.sub.IA ≤ T.sub.IA_max will result in following advantages: [0038] a) usage of minimum number of SS blocks needed to support UE angular distributions, mobility, and number of connected UEs through satisfying target KPI for initial access; [0039] b) optimal transmissions of SS beams in spatial directions where UE activity is mainly observed and varying over large time windows; and [0040] c) flexible configuration of T.sub.ss to allow optimal CSI-RS configurations for each gNBs (refer to equation (2.0) above).

[0041] These factors directly impact the network transmit power and spectral efficiency.

[0042] According to an example methodology for training for a set of gNBs (illustrated in FIG. 2, beginning with the start block 201), initialization (see, e.g., block 202 of FIG. 2) for training for a set of gNBs involves specifying the following: [0043] a. Maximum initial access latency (T.sub.IA.sub._.sub.max) KPI operator inputs; [0044] b. Network operating Frequency (FR1 or FR2) [0045] c. Set of SS Beam Weight vectors for network operating frequency (e.g., FR1 or FR2) to choose from for each gNB.

[0046] According to the example methodology for training for a set of gNBs, for each gNB, at least one of the following inputs can be utilized (see, e.g., FIG. 2, block 203): [0047] 1. UE indirect best beam ID reports per gNB (Rx Beam IDs over which PRACH is received); [0048] 2. best beam SNR measured over received PRACH correlation magnitude; [0049] 3. derived information at the gNB (e.g., UE mobility information, number of connected UEs); and [0050] 4. UE location History (Xapp based/R1-16 Location service based).

[0051] Based on the above-described input, a trained AI/ML agent (e.g., RL agent) for selecting the optimal SS Burst configuration is produced, which selection of the optimal SS Burst configuration can involve: [0052] a. Phase 1 (FIG. 2, block 204): SS Burst Set Configuration (L, SS Block frequency diversity, T.sub.ss, CSI-RS TRS); [0053] b. Phase 2 (FIG. 2, block 205): SS Beam (common beam) Weight vector Selection; and [0054] c. Phase 3 (FIG. 2, block 206): CSI-RS configuration set for beam measurement.

[0055] As shown in FIG. 3, an example process for learning the optimal SS Burst Configuration for maximizing the initial access latency (IAL) up to the target IAL can involve the following (beginning with the start block 301): [0056] a. Based on the gNB PRACH detection energy SNR threshold, compute (Rx Beam ID, number of UE’s PRACH correlation magnitude > Threshold} for an observation time window (FIG. 3, block 302), [0057] b. Gather information on number of connected UEs with the gNB for the observation time windows (FIG. 3, block 303); [0058] c. Infer UE mobility based on multiple windows of observations (FIG. 3, block 304); [0059] d. AI/ML Agent (e.g., RL Agent) evaluates relationships (e.g., as mentioned above in Action 1 and Action 2) to minimize {S.sub.D, N.sub.ss} based on observations at gNB using deep reinforcement learning (DRL) (FIG. 3, block 305); [0060] e. Calculate T.sub.IA upper limit bounded by the 3GPP-specified max value T.sub.IA_max for every gNB (FIG. 3, block 306); and [0061] f. AI/ML engine updates parameter set, minimizes cost function or maximizes reward (FIG. 3, block 307).

[0062] According to an example methodology in accordance with the present disclosure, the AI/ML agent can learn to define an optimization framework for a set of multiple gNBs using both SS and CSI-RS optimizations, e.g., based on multi-TRP/gNB observations. In this case, the training data set used for the AI/ML agent training will also incorporate the set of beamforming weight vectors to apply for different SS beam directions and shape, along with the CSI-RS configurations parameters available at the gNB. This will allow choosing the best SS beam directions in a given gNB to minimize the interference from neighboring gNBs, as well as enable optimal configurations for CSI-RS followed by SS Burst set, and hence will further optimize the network performance in terms of lowering the transmit power and improving spectral efficiency. Further variations and extensions of the above learning methodology will be readily apparent to those skilled in the art.

Definitions

[0063] 3GPP: 3rd Generation Partnership Project [0064] AGC: Automatic gain control [0065] AI/ML: Artificial Intelligence/Machine Learning [0066] BS: Base Station [0067] CCH: Control channel [0068] CPRI: Common Public Radio Interface [0069] CUS-plane: Control, user, and synchronization plane [0070] CSI-RS: channel state information reference signal [0071] CSI-RS TRS: CSI-RS Tracking Reference Signal [0072] DL: Downlink [0073] eCPRI: Enhanced Common Public Radio Interface [0074] eNB: eNodeB (4G LTE base station) [0075] FEC: Forward error correction [0076] FH: Fronthaul [0077] FS: Functional split [0078] FR1: Frequency Range 1 [0079] FR2: Frequency Range 2 [0080] gNB: gNodeB (5G NR base station) [0081] GPS: Global positioning system [0082] HW: Hardware [0083] IA: Initial Access [0084] IAL: Initial Access Latency [0085] M-plane: Management plane [0086] MIMO: Multiple Input Multiple Output [0087] Massive MIMO: Massive Multiple Input Multiple Output [0088] Near-RT RIC: Near Real Time RIC [0089] Non-RT RIC: Non Real Time RIC [0090] NR: new radio interface and radio access technology for cellular networks [0091] KPIs: Key performance indicators [0092] O-CU: O-RAN compliant Centralized Unit [0093] O-DU: O-RAN compliant Distributed Unit [0094] O-RU: O-RAN compliant Radio Unit [0095] OPEX: Operating expenses [0096] PRACH: Physical random-access channel [0097] PRB: Physical resource block [0098] PTP: Precision time protocol [0099] PPP: Poisson point process [0100] RRC: Radio Resource Controller [0101] RIC: RAN Intelligent Controller [0102] TRP: Transmission/Reception Point [0103] RACH: Random access channel [0104] RAT: Radio access technology [0105] RE: Resource element [0106] RoE: Radio over Ethernet [0107] RSRP: Reference Signal Received Power [0108] SMO: Service Management and Orchestration [0109] SW: Software [0110] SS: Synchronization Signal [0111] SSB: Synchronization Signal Block [0112] SyncE: Synchronous Ethernet [0113] SNR: Signal to Noise Ratio [0114] TCH: Traffic channel [0115] UL: Uplink [0116] UE: User Equipment [0117] xAPP: applications realizing chosen telecom functions [0118] PSS: Primary Synchronization Signal [0119] SSS: Secondary Synchronization Signal [0120] PBCH: Physical Broadcast Channel [0121] DMRS: Demodulation Reference Signal [0122] OFDM: Orthogonal Frequency Division Multiplexing [0123] dB: Decibel