Method and system for multi-dimensional deep FAKE imputation of telecom network observability data for network intelligence-as-a-service

Abstract

A system and a method for generating a realistic data deep fake of missing observability data, integrity-compromised data, or statistical outliers using Generative AI-based imputation for Operations, Administration and Maintenance (OAM) observability data for a Radio Access Network.

Description

METHOD AND SYSTEM FOR MULTI-DIMENSIONAL DEEP FAKE IMPUTATION OF TELECOM NETWORK OBSERVABILITY DATA FOR NETWORK INTELLIGENCE- AS-A-SERVICEDESCRIPTION OF THE RELATED TECHNOLOGY

[0001] The present disclosure relates to systems and methods for radio access networks. More particularly, the present disclosure relates to the design of operation, administration and management of various network elements and Artificial Intelligence and Analytics for mobile telecommunication systems of 4G, 5G, and further generations.OVERVIEW OF IMPLEMENTATIONS

[0002] Described is a system and method for generating a realistic data deep fake of missing observability data, integrity-compromised data, or statistical outliers using Generative Al-based imputation for Operations, Administration and Maintenance (0AM) observability data for a Radio Access Network.BRIEF DESCRIPTION OF THE DRAWINGS

[0003] Figure 1 is a block diagram of a system architecture.

[0004] Figure 2 shows an example of a User Plane Stack.

[0005] Figure 3 shows an example of a Control Plane Stack.

[0006] Figure 4A shows an example of a Separation of CU-CP (CU-Control Plane) and CU-UP (CU-User Plane).

[0007] Figure 4B shows an example of a Separation of CU-CP (CU-Control Plane) and CU-UP (CU-User Plane).

[0008] Figure 5 shows a DL (Downlink) Layer 2 Structure.

[0009] Figure 6 shows an exemplary logical flow for implementing an RB allocation policy.0017389INP / 4688

[0010] Figure 7 shows an L2 Data Flow example.

[0011] Figure 8A shows an example of an O-RAN architecture.

[0012] Figure 8B shows an example of an O-RAN architecture.

[0013] Figure 8C shows an example of an O-RAN architecture.

[0014] Figure 9 describes an E-UTRAN architecture.

[0015] Figure 10 describes an EN-DC architecture.

[0016] Figure 11 illustrates a discriminator loss backpropagation.

[0017] Figure 12 illustrates a generator loss backpropagation.

[0018] Figure 13 illustrates a recurrent GAN with LSTM variant for generator and discriminator.

[0019] Figure 14 illustrates a GAN with LSTM Training Architecture.

[0020] Figure 15 illustrates shows a GAN with LSTM - Inference-only architecture.

[0021] Figure 16 illustrates an architecture of an R-GAIN implementation in an 0-RAN deployment.

[0022] Figure 17 illustrates an evaluation of a data deep-fake of Wideband Channel Quality Indication (CQI).

[0023] Error! Reference source not found, illustrates evaluation of a data deep fake of Data Radio Bearer throughput volume and time.

[0024] Figure 19 illustrates an evaluation of a data deep fake of Physical Radio Block (PRB) utilization and PRB congestion.0017389INP / 4688

[0025] Figure 20 illustrates an evaluation of a data deep fake of actual number of initial Evolved Radio Access Bearer (E-RAB) establishment attempts and actual number of E-RAB release attempts.

[0026] Figure 21 shows an evaluation of a data deep fake of an actual number of samples with high Reference Signal Received Power (RSRP), medium RSRP and poor RSRP.

[0027] Figure 22 shows an evaluation of data deep fake of a number of Radio Resource connection (RRC) establishment attempts, number of Random Access Channel (RACH) attempts and number of contention RACH reports.

[0028] Figure 23 shows an evaluation of data deep fake of number of a number of handover attempts and the number of too-early HO attempts.

[0029] Figure 24 shows an evaluation of a data deep fake of a number of samples with high, medium and low neighbor cell interference.DETAILED DESCRIPTION OF THE IMPLEMENTATIONS

[0030] Reference is made to Third Generation Partnership Project (3GPP) and the Internet Engineering Task Force [IETF] and related standards bodies in accordance with embodiments of the present disclosure. The present disclosure employs abbreviations, terms and technology defined in accord with Third Generation Partnership Project (3GPP) and / or Internet Engineering Task Force (IETF) technology standards and papers, including the following standards and definitions. 3GPP and IETF technical specifications (TS), standards (including proposed standards), technical reports (TR) and other papers are incorporated by reference in their entirety hereby, define the related terms and architecture reference models that follow.

[0031] 0-RAN. WG4. MP.0-R003-vl3.00

[0032] 3GPP TS 23.501 V 18.1.02023-04-05

[0033] 3GPP TS 38.300 V 17.4.003-28-20230017389INP / 4688

[0034] 3GPP TS 38.401 V 17.4.02023-04-03

[0035] 3GPP TS 38.501 V 18.1.02023-04-05

[0036] 3GPP TS 38.425 17.3.0, 2023-04-03

[0001] Acronyms3GPP: Third generation partnership projectADC: Analog-to-Digital ConverterACER: Gain and Adjacent Channel Leakage RatioASIC: Application Specific Integrated CircuitAWGN: Additive White Gaussian NoiseBFW: Beamforming weightBS: Base StationC-RAN: cloud radio access networkCU: Central unitCQI: Channel Quality IndicatorDFE: Digital Front EndDL: DownlinkDCI: Downlink Control InformationDU: Distributed unitEPC: Evolved Packet CoreeNB: evolved Node BgNB: g NodeBEN-DCDAC: Digital-to-Analog ConverterDMRS: Demodulation Reference Signal0017389INP / 4688FCAPS: Fault Management data, Configuration Management data.Performance Management data, Trace data, and so onGAN: Generative Adversarial NetworksloT: Internet of ThingsLI: Layer 1L2: Layer 2L3: Layer 3LSTM: Long Short Term MemoryRLC: Radio Link ControlRRC: Radio Resource ControlRU: Radio UnitU-plane: User planeUPF: User Plane FunctionUE: user equipmentUL: uplinkMIMO: multiple-in multiple-outMME: Mobility Management EntityMR-DC: Multi-Radio Dual ConnectivityM-plane: Management plane interface between SMO and O-RUNB: NarrowbandNIaaS: Network Intelligence-as-a-ServiceNR: New RadioNR-U: New Radio - User PlaneOAM: Operations, Administration and Maintenance0017389INP / 4688 OFDM: orthogonal frequency-division multiplexingOPEX: Operating ExpenseO-RAN: Open Radio Access NetworkPA: Power AmplifierPDCP: Packet Data Convergence ProtocolPDCCH: Physical Downlink Control ChannelPDSCH: Physical Downlink Shared ChannelPUCCH: Physical Uplink Control ChannelPUSCH: Physical Uplink Shared ChannelPDCP: Packet Data Convergence ProtocolQCI: QoS Class IdentifierQFI: QoS Flow IdQoS: Quality of ServiceRLC: Ratio Link ControlMAC: Medium Access ControlPHY: Physical LayerPRG: Physical Resource block GroupRAT: Radio Access TechnologyRB: Resource BlockRLC: Radio Link ControlRU: Radio UnitRMM: Radio resource managementSINR: Signal-to-Interference and Noise RatioSN: Signal Node0017389INP / 4688SR: Scheduling RequestSRS: Sounding Reference SignalSMO: Service Management and Orchestration systemS-GW: Serving Gateway

[0037] Described are implementations of technology for a cloud-based Radio Access Networks (RAN), where a significant portion of the RAN layer processing is performed at a central unit (CU) and a distributed unit (DU). Both CUs and DUs are also known as the baseband units (BBUs). CUs are usually located in the cloud on commercial off the shelf servers, while DUs can be distributed, while the RF and real-time critical functions can be processed in the remote radio unit (RU).

[0038] RAN Architectures

[0039] Figure 1 is a block diagram of a system 100 for implementations as described herein. System 100 includes a NR UE 101, a NR gNB 106. The NR UE and NR gNB are communicatively coupled via a Uu interface 120.

[0040] NR UE 101 includes electronic circuitry, namely circuitry 102, that performs operations on behalf of NR UE 101 to execute methods described herein. Circuity 102 may be implemented with any or all of (a) discrete electronic components, (b) firmware, and (c) a programmable circuit 102A.

[0041] NR gNB 106 includes electronic circuitry, namely circuitry 107, that performs operations on behalf of NR gNB 106 to execute methods described herein. Circuity 107 may be implemented with any or all of (a) discrete electronic components, (b) firmware, and (c) a programmable circuit 107A.

[0042] Programmable circuit 107A, which is an implementation of circuitry 107, includes a processor 108 and a memory 109. Processor 108 is an electronic device configured of logic circuitry that responds to and executes instructions. Memory 109 is a tangible, non-transitory, computer-readable storage device encoded with a computer program. In this regard, memory 109 stores data and0017389INP / 4688instructions, i.e., program code, that are readable and executable by processor 108 for controlling operations of processor 108. Memory 109 can be implemented in a random-access memory (RAM), a hard drive, a read only memory (ROM), or a combination thereof. One of the components of memory 109 is a program module, namely module 110. Module 110 has instructions for controlling processor 108 to execute operations described herein on behalf of NR gNB 106.

[0043] The term "module" is used herein to denote a functional operation that may be embodied either as a stand-alone component or as an integrated configuration of a plurality of subordinate components. Thus, each of module 105 and 110 can be implemented as a single module or as a plurality of modules that operate in cooperation with one another.

[0044] While modules 110 are indicated as being already loaded into memories 109, and module 110 can be configured on a storage device 130 for subsequent loading into their memories 109. Storage device 130 is a tangible, non-transitory, computer-readable storage device that stores module 110 thereon. Examples of storage device 130 include (a) a compact disk, (b) a magnetic tape, (c) a read only memory, (d) an optical storage medium, (e) a hard drive, (f) a memory unit comprising multiple parallel hard drives, (g) a universal serial bus (USB) flash drive, (h) a random-access memory, and (i) an electronic storage device coupled to NR gNB 106 via a data communications network.

[0045] Uu Interface (120) is the radio link between the NR UE and NR gNB, which is compliant to the 5G NR specification.

[0046] UEs 101 can be dispersed throughout a wireless communication network, and each UE can be stationary or mobile. A UE includes: an access terminal, a terminal, a mobile station, a subscriber unit, a station, and so on A UE can also include be a cellular phone (e.g., a smart phone), a personal digital assistant (PDA), a wireless modem, a wireless communication device, a handheld device, a laptop computer, a cordless phone, a wireless local loop (WLL) station, a tablet, a0017389INP / 4688camera, a gaming device, a drone, a robot / robotic device, a netbook, a smartbook, an ultrabook, a medical device, medical equipment, a healthcare device, a biometric sensor / device, a wearable device such as a smartwatch, smart clothing, smart glasses, a smart wristband, and / or smart jewelry (e.g., a smart ring, a smart bracelet, and so on], an entertainment device (e.g., a music device, a video device, a satellite radio, and so on], industrial manufacturing equipment, a global positioning system [GPS] device, or any other suitable device configured to communicate via a wireless or wired medium. UEs can include UEs considered as machine-type communication (MTC] UEs or enhanced / evolved MTC (eMTC] UEs. MTC / eMTC UEs that can be implemented as loT UEs. loT UEs include, for example, robots / robotic devices, drones, remote devices, sensors, meters, monitors, cameras, location tags, and so on, that can communicate with a BS, another device (e.g., remote device], or some other entity. A wireless node can provide, for example, connectivity for or to a network (e.g., a wide area network such as Internet or a cellular network] via a wired or wireless communication link.

[0047] One or more UEs 101 in the wireless communication network can be a narrowband bandwidth UE. As used herein, devices with limited communication resources, e.g. smaller bandwidth, are considered as narrowband UEs. Similarly, legacy devices, such as legacy and / or advanced UEs can be considered as wideband UEs. Wideband UEs are generally understood as devices that use greater amounts of bandwidth than narrowband UEs.

[0048] The UEs 101 are configured to connect, for example, communicatively couple, with an or RAN. In embodiments, the RAN is an NG RAN or a 5G RAN, an E-UTRAN, an MF RAN, or a legacy RAN, such as a UTRAN or GERAN. The term " NG RAN” or the like refers to a RAN 110 that operates in an NR or 5G system, the term " E-UTRAN” or the like refers to a RAN that operates in an LTE or 4G system, and the term " MF RAN” or the like refers to a RAN that operates in an MF system 100. The UEs 101 utilize connections (or channels], respectively, each of which comprises a physical communications interface or layer. The connections can comprise several0017389INP / 4688different physical DL channels and several different physical UL channels. As examples, the physical DL channels include the PDSCH, PMCH, PDCCH, EPDCCH, MPDCCH, R-PDCCH, SPDCCH, PBCH, PCFICH, PHICH, NPBCH, NPDCCH, NPDSCH, and / or any other physical DL channels mentioned herein. As examples, the physical UL channels include the PRACH, PUSCH, PUCCH, SPUCCH, NPRACH, NPUSCH, and / or any other physical UL channels mentioned herein.

[0049] The RAN can include one or more AN nodes or RAN nodes. These access nodes can be referred to as BS, gNBs, RAN nodes, eNBs, NodeBs, RSUs, MF-APs, TRxPs or TRPs, and so forth, and comprise ground stations (e.g., terrestrial access points) or satellite stations providing coverage within a geographic area (e.g., a cell). The term " NG RAN node” or the like refers to a RAN node that operates in an NR or 5G system (e.g., a gNB), and the term " E-UTRAN node” or the like refers to a RAN node that operates in an LTE or 4G system (e.g., an eNB). According to various embodiments, the RAN nodes can be implemented as one or more of a dedicated physical device such as a macrocell base station, and / or a low power base station for providing femtocells, picocells or other like cells having smaller coverage areas, smaller user capacity, or higher bandwidth compared to macrocells.

[0050] In some embodiments, all or parts of the RAN nodes can be implemented as one or more software entities running on server computers as part of a virtual network, which may be referred to as a CRAN and / or a vBBU. In these embodiments, the CRAN or vBBU can implement a RAN function split, such as a PDCP split in which RRC and PDCP layers are operated by the CRAN / vBBU and other L2 protocol entities are operated by individual RAN nodes; a MAC / PHY split wherein RRC, PDCP, RLC, and MAC layers are operated by the CRAN / vBBU and the PHY layer is operated by individual RAN nodes; or a "lower PHY” split in which RRC, PDCP, RLC, MAC layers and upper portions of the PHY layer are operated by the CRAN / vBBU and lower portions of the PHY layer are operated by individual RAN nodes. This virtualized framework allows the freed-up processor cores of the RAN nodes to perform other virtualized applications. In some implementations, an0017389INP / 4688individual RAN node can represent individual gNB-DUs that are connected to a gNB-CU via individual Fl interfaces. In these implementations, the gNB-DUs can include one or more remote radio heads (RRH), and the gNB-CU each be operated by a server that is located in the RAN or by a server pool in a similar manner as the CRAN / vBBU. One or more of the RAN nodes can be next generation eNBs (ng-eNBs), which are RAN nodes that provide E-UTRA user plane and control plane protocol terminations toward the UEs 101, and are connected to a 5GC via an NG interface. In MF implementations, the MF-APs are entities that provide MultiFire radio services, and maybe similar to eNBs in an 3GPP architecture.

[0051] In some implementations, access to a wireless interface can be scheduled. A scheduling entity (e.g.: BS, gNB, and so on) allocates bandwidth resources for devices and equipment in its service area or cell. As scheduling entity can be configured to schedule, assign, reconfigure, and release resources for one or more subordinate entities. In some examples, a UE 101 (or other device) may function as master node scheduling entity, scheduling resources for one or more secondary node subordinate entities (e.g., one or more other UEs 101). Thus, in a wireless communication network with a scheduled access to time — frequency resources and having a cellular configuration, a P2P configuration, and a mesh configuration, a scheduling entity and one or more subordinate entities may communicate utilizing the scheduled resources.

[0052] BS or gNB can be equipped with T antennas and UE 101 can be equipped with R antennas, where in general T>1 and R>1. At BS, a transmit processor is configured to receive data from a data source for one or more UEs 101 and select one or more modulation and coding schemes (MCS) for each UE based on channel quality indicators (CQIs) received from the UE 101. The BS is configured to process (e.g., encode and modulate) the data for each UE 101 based on the MCS(s) selected for the UE 101, and provide data symbols for all UEs. A transmit processor is also configured to process system information (e.g., for static resource partitioning information (SRPI), and so on) and control information (e.g., CQI0017389INP / 4688requests, grants, upper layer signaling, and so on) and can provide overhead symbols and control symbols. Processor 108 can also generate reference symbols for reference signals (e.g., the cell-specific reference signal (CRS)) and synchronization signals (e.g., the primary synchronization signal (PSS) and the secondary synchronization signal (SSS)). A transmit (TX) multiple-input multiple-output (MIMO) processor can be configured perform spatial processing (e.g., preceding) on the data symbols, the control symbols, the overhead symbols, and / or the reference symbols, if applicable, and can be configured to provide T output symbol streams to T modulators (MODs). Each modulator can be configured to process a respective output symbol stream (e.g., for OFDM, and so on) to obtain an output sample stream. Each modulator can further be configured to process (e.g., convert to analog, amplify, filter, and upconvert) the output sample stream to obtain a downlink signal. T downlink signals from modulators can be transmitted via T antennas.

[0053] An overview of 5G NR Stacks is as follows. 5G NR (New Radio) user and control plane functions with monolithic gNB (gNodeB) are shown in the figures below. For the user plane, PHY (physical), MAC (Medium Access Control), RLC (Radio Link Control), PDCP (Packet Data Convergence Protocol) and SDAP (Service Data Adaptation Protocol) sublayers are terminated in the gNB on the network side. For the control plane, RRC (Radio Resource Control), PDCP, RLC, MAC and PHY sublayers are terminated in the gNB on the network side and NAS (Non-Access Stratum) is terminated in the AMF (Access Mobility Function) on the network side. Figure 2 shows an example of a User Plane Stack as described in 3GPP TS 38.300. Figure 3 shows an example of a Control Plane Stack as described in 3GPP TS 38.300.

[0054] An NG-RAN (NG-Radio Access Network) architecture from 3GPP TS 38.401 is described below. Fl is the interface between gNB-CU (gNB - Centralized Unit) and gNB-DU (gNB - Distributed Unit), NG is the interface between gNB-CU (or gNB) and 5GC (5G Core), El is the interface between CU-CP (CU-Control Plane) and CU-UP (CU-User Plane), and Xn is interface between gNBs.0017389INP / 4688

[0055] A gNB can comprise a gNB-CU-CP, multiple gNB-CU-UPs and multiple gNB-DUs. The gNB-CU-CP is connected to the gNB-DU through the Fl-C interface and to the gNB-CU-UP through the El interface. The gNB-CU-UP is connected to the gNB-DU through the Fl-U interface and to the gNB-CU-CP through the El interface. One gNB-DU is connected to only one gNB-CU-CP and one gNB-CU-UP is connected to only one gNB-CU-CP. Figure 4A shows an example of an NG-RAN Architecture as described in 3GPP TS 38.501. Figure 4B shows an example of a Separation of CU-CP (CU-Control Plane) and CU-UP (CU-User Plane) as described in 3GPP TS 38.401.

[0056] A Layer 2 (L2) of 5G NR is split into the following sublayers as described in 3GPP TS 38.300):• Medium Access Control (MAC): The MAC sublayer offers Logical Channels (LCs) to the RLC sublayer. This layer runs a MAC scheduler to schedule radio resources across different LCs (and their associated radio bearers).• Radio Link Control (RLC): The RLC sublayer offers RLC channels to the PDCP sublayer. The RLC sublayer supports three transmission modes: RLC-Transparent Mode (RLC-TM), RLC-Unacknowledged Mode (RLC-UM) and RLC-Acknowledgement Mode (RLC-AM). RLC configuration is per logical channel. It hosts ARQ (Automatic Repeat Request) protocol for RLC-AM mode.• Packet Data Convergence Protocol (PDCP): The PDCP sublayer offers Radio Bearers (RBs) to the SDAP sublayer. There are two types of Radio Bearers: Data Radio Bearers (DRBs) for data and Signaling Radio Bearers (SRBs) for control plane.• Service Data Adaptation Protocol (SDAP): The SDAP offers QoS Flows to the 5GC (5G Core). This sublayer provides mapping between a QoS flow and a DRB. It marks QoS Flow Id in DL (downlink) as well as UL (uplink packets).0017389INP / 4688

[0057] Figure 5 shows a DL (Downlink) Layer 2 Structure as described in 3GPP TS 38.300. Figure 6 shows an UL (uplink) Layer 2 Structure in accord with 3GPP TS38.300. Figure 7 shows an L2 Data Flow example in accord with 3GPP TS 38.300 ([H] denotes headers or subheaders in Figure 7.)

[0058] O-RAN, which is based on disaggregated components and connected through open and standardized interfaces, is based on 3GPP NG-RAN. An overview of O-RAN with disaggregated RAN (CU, DU, and RU), near-real-time RIC and non-real-time RIC is shown in the figure below. Here, DU (Distributed Unit) and CU (Centralized Unit) are typically implemented using COTS (Commercial off-the-shelf) hardware.

[0059] Figures 8A-8C show an example of an O-RAN architecture. In Figure 8A, the CU and the DU are connected using the Fl interface (with Fl-C for control plane and Fl-U for user plane traffic) over the midhaul (MH) path. One DU can host multiple cells (for example, one DU can host 24 cells) and each cell can support many users. For example, one cell may support 600 RRC Connected users and out of these 600, there may be 200 Active users (i.e.; users which have data to send at a given point of time).

[0060] A cell site can comprise of multiple sectors and each sector can support multiple cells. For example, one site can comprise three sectors and each sector can support 8 cells (with 8 cells in each sector on different frequency bands). One CU-CP can support multiple DUs and thus multiple cells. For example, a CU-CP can support 1000 cells and around 100,000 UEs. Each UE can support multiple DRBs and there can be multiple instances of CU-UP to serve these DRBs. For example, each UE can support 4 DRBs, and 400,000 DRBs (corresponding to 100,000 UEs) can be served by five CU-UP instances (and one CU-CP instance).

[0061] DU can be located in a private data center or it can be located at a cellsite too. CU can also be located in a private data center or even hosted on a public cloud system. DU and CU can be tens of kilometers away. CU can communicate with0017389INP / 46885G core system which can also be hosted in the same public cloud system (or can be hosted by a different cloud provider). RU (Radio Unit) is located at cell-site and communicated with DU via a fronthaul (FH) interface.

[0062] The E2 nodes (CU and DU) are connected to the near-real-time RIC using the E2 interface. The E2 interface is used to send data (e.g., user, cell, slice KPMs) from the RAN, and deploy control actions and policies to the RAN at near-real-time RIC. The application or service at the near-real-time RIC that deploys the control actions and policies to the RAN are called xApps. The near-real-time RIC is connected to the non-real-time RIC using the Al interface.

[0063] SMO manages multiple regional networks, and O-RAN NFs (O-CUs, Near-RT RIC, O-DUs) can be deployed in a regional data center which is connected to multiple cell sites or in cell site which is close to localized O-RU according to network requirements. Since SMO Functions and O-RAN NFs are micro services and deployment-independent logical functions, SMO Functions and O-RAN NFs can be composed of multiple deployment instances deployed in the same O-Cloud or in a different O-Cloud in regional data center, or in cell site according to network requirements (ex. capacity, latency, security, and so on) if the secure connection among SMO Functions and O-RAN NFs are available.

[0064] As shown in Figures 8B and 8C, an O-RAN compliant SMO defines TE& IV, RAN NF 0AM, Non-RT RIC, and NFO, FOCOM services. SMO interacts with 0-RAN NFs with 01 interface. SMO interacts with O-RU with Open FH M-Plane interface and interacts O-Cloud via the 02 interface. O-RAN NF 0AM manages 0-RAN NF CM, FM, PM and creates O-RAN NF inventory and topology in TE& IV.FOCOM / NFO manages O-Cloud resources and creates O-Cloud resources inventory and topology in TE& IV. Analytics / rApp in Non-RT RIC can subscribe O-RAN NFs PM / FM, O-Cloud PM / FM data based on O-RAN NF 0AM and FOCOM / NFO.Analytics / rApp in Non-RT RIC can retrieve the O-RAN NF and O-Cloud resource inventory and topology.0017389INP / 4688

[0065] NIaaS, or Network Intelligence-as-a-Service, is a holistic Al framework for telecommunication systems that offers a suite of Al-native micro-services, that include data engineering framework services for handling observability and configuration data pipelines, data science framework services for generating statistics relevant to data and towards detecting statistical outliers, Al ML model training framework services for training causation, classification, forecasting, recommendation models and so on, Al ML life cycle management framework services for managing the life cycle of models such as version upgrade, rollback and so on, Al ML inference framework services for live data feeds, Al test framework services for digital twins, and Al ML visualization framework services, and so on. A NIaaS framework in the RIC can offer a rich set of sophisticated data analytics and AI / ML services for xApps. A TS xApp can use, for example, Reinforcement Learning (RL) to generate control actions, learn throughput KPI as a function of the input features from the state information and action spaces, and update its state-action-reward mapping. AI / ML can be initially trained offline in the Non-RT RIC and then downloaded in the Near-RT RIC, where it continues to learn online against live E2 data. A TS xApp can use evolving 0-RAN standard API procedures to interact with the Near-RT RIC platform services, which further use E2AP procedures to interface with the E2 nodes. The APIs can be based on cloud-native gRPC and RIC Message Routing (or RMR - a socket interface SI-95 based low-latency transport protocol] in accord with 0-RAN. These APIs can facilitate seamless integration and interoperability of operator, vendor, and 3rd party xApps with the platform.

[0066] An E-UTRAN architecture is illustrated in Figure 9. The E-UTRAN comprises eNBs, providing the E-UTRA U-plane (PDCP / RLC / MAC / PHY) and control plane (RRC) protocol terminations towards the UE. The eNBs are interconnected with each other by the X2 interface. The eNBs are also connected by the SI interface to the EPC (Evolved Packet Core), more specifically to the MME (Mobility Management Entity) by the Sl-MME interface and to the Serving Gateway (S-GW) by the Sl-U interface. The SI interface supports a many-to-many relation between MMEs / Serving Gateways and eNBs.0017389INP / 4688

[0067] E-UTRAN also supports MR-DC via E-UTRA-NR Dual Connectivity (EN-DC), in which a UE is connected to one eNB that acts as a MN and one en-gNB that acts as a SN. An EN-DC architecture is illustrated in Figure 10. The eNB is connected to the EPC via the SI interface and to the en-gNB via the X2 interface. The en-gNB might also be connected to the EPC via the Sl-U interface and other en-gNBs via the X2-U interface. In EN-DC, and en-gNB comprises gNB-CU and gNB-DU(s).

[0068] E-UTRAN also supports and NG-RAN architecture. An NG-RAN node is either:a gNB, providing NR user plane and control plane protocol terminations towards the UE; oran ng-eNB, providing E-UTRA user plane and control plane protocol terminations towards the UE. (3GPP TS 38.30017.3.0.)

[0069] As shown in Figures 9-10, the gNBs and ng-eNBs are interconnected with each other by the Xn interface. The gNBs and ng-eNBs are also connected by the NG interfaces to the 5GC, more specifically to the AMF (Access and Mobility Management Function) by the NG-C interface and to the UPF (User Plane Function) by the NG-U interface. The gNB and ng-eNB host functions for Radio Resource Management such as: Radio Bearer Control, Radio Admission Control, Connection Mobility Control, Dynamic allocation of resources to UEs in both uplink and downlink (scheduling), connection setup and release; session Management; QoS Flow management and mapping to data radio bearers; and Dual Connectivity. Tight interworking between NR and E-UTRA. NB-IoT UE is supported by ng-eNB.

[0070] The gNB and ng-eNB host functions such as functions for Radio Resource Management: Radio Bearer Control, Radio Admission Control, Connection Mobility Control, Dynamic allocation of resources to UEs in both uplink and downlink (scheduling), connection setup and release; session Management; QoS Flow management and mapping to data radio bearers; Dual Connectivity; and Tight interworking between NR and E-UTRA. NB-IoT UE is supported by ng-eNB.0017389INP / 4688

[0071] In an example, control information (e.g., scheduling information) may be provided for broadcast and / or multicast operation. The UE can monitor different bundle sizes for the control channel depending on the maximum number of repetitions.

[0072] Description of Implementations

[0073] Network and User Experience Key Performance Indicators [KPIs] are central for telco operators. Operators are rated in markets based on their KPIs, which directly impact the growth of the mobile subscriber base for these operators in their respective markets. Telco operations thus require continuous monitoring of the KPIs, aided by insightful analytics and intelligence, towards meeting KPI compliance and subsequently towards improving them consistently. The telco network KPI families include:o Connectivity and Registration: This KPI family measures the success of connectivity of mobile User Equipment (UEs) to RAN and packet core of the telco mobile network.o Mobility and load balancing: This KPI family measures the success and stability of mobility of User Equipment (UE) in the network between network entities, network functions, and so ono Coverage quality and Availability: This KPI family measures the quality of network coverage.o Context and Traffic Management: This KPI family measures the success of control and management plane aspects of the UE context and UE-subscribed traffic bearers.0017389INP / 4688o User-plane and Integrity: This KPI family measures the user-plane quality of UE-subscribed data traffic.o Quality-of-Experience: This KPI family measures the Quality of Experience of mobile subscriber UEs.o Resource Management and Efficiency: This KPI family measures the efficiency and management of the network resources. Examples include (i) RAN KPIs, such as spectral efficiency, RAN energy efficiency, radio resource utilization, active user load, RAN OPEX, and so on, (ii) Packet Core KPIs, such as network slice energy efficiency, packet core OPEX, (iii) Infrastructure KPIs, such as transactions per second per virtual or physical core, and so on.

[0074] Intelligence about the KPIs of a given cell (or a network function or a network entity or a UE) is derived by the Operations, Administration and Maintenance (0AM) observability data pertaining to that cell-of-interest (or network function or network entity or UE) but also pertaining to significantly-impacting neighbor cells (or NFs or network entities or other UEs), which include Fault Management (FM) data. Configuration Management (CM) data, Performance Management (PM) data, Trace data, and so on for observability and configuration (collectively referred to as 0AM FCAPS data). Deriving intelligence about the KPIs of the cells (or network function or network entity or UE) is impacted by the availability of the 0AM FCAPS data.

[0075] The telco data volume, generated by the network functions and / or their associated network entities, is enormous. As a ballpark, from the NF domain side, there are cell-sites or NFs, ranging from 100’s or few 1,000’s in a sub-market (city / county / district), 10,000’s in a regional market (federal states / regional zones) to 100,000’s in a nation. In a sub-market / city, there are 1,000’s to a few 10,000’s cells in a sub-market / city. There are 10,000’s to 100,000’s UEs in a city / edge sub-0017389INP / 4688market. Sometimes, this can scale till IM’s UEs in a metro city (such as NYC, London, Paris, Shanghai, Mumbai, and so on). The instances of NFs and their associated entities (such as cells) are logically grouped into slice instances or slice subnet instances, and there can be 10’s to a few 100’s of slice or slice subnet instances. From the infrastructure side, there are units to 10’s of Kubernetes (k8s) clusters in a city / sub-market, 100’s in a region / market and 1,000’s nation-wide. And in each k8s cluster, there are units to a few 10's of worker nodes (VMs / servers). Also, NFs have several tens of containerized applications, executing the protocol stack functionality, hosted as micro-services with each micro-service running across multiple pods (units to a few 10’s) across worker nodes. Now, each NF, cell, UE or slice / slice subnet instance can generate several 100’s to a few 1,000’s of observability data and event notifications, in the order of units of minutes. Some of the observability data such as trace messages, time-aggregated PM reports, and so on, are bulkier (where each such file / stream from each UE or a network entity, such as cell, or a worker node's pod can run into 10’s - 100's of KBs every minute or unit of mins, and thus, the net observability data volume reported from across the entire network is significantly voluminous).

[0076] However, it is to be noted that one of the other major problems with teleco observability data reporting lies with their availability. In particular, when there is missingness and integrity compromises in the observability data being reported, then deriving intelligence about the KPI, such as KPI predictions, root cause analysis and explainability, forecasting and alerts, recommendations and optimizations, and so on is adversely impacted. With volumes of observability telco data as discussed above, the missingness in data for a given window of time (in the order of weeks) is:o around 5% due to unplanned operational or software issues and transport bottlenecks,o around 10 - 15% due to planned operational upgrades, and0017389INP / 4688o 20% and above due to transport link or infrastructure outages and lack of hardware resources support for dimensioning volume.

[0077] This outage, concerning observability data, may last continuously and persistently, for a few hours to the order of number of days, which adversely impacts intelligence from being generated during the period of time that the data goes missing.

[0078] In the process, disclosed is a multi-dimensional data imputation technique by generating realistic deep fakes of observability data using Generative Al techniques, when the data goes missing due to the above factors. In particular, the disclosure describes a Generative Al-based model, based on Generative Adversarial Neural Networks, by which, a single Al deep fake model can be used for multi-dimensional imputation, i.e., to impute any number of missing observability data inputs (such as one or more PMs, one or more CMs, one or more trace data, one or more alarms, and so on across one or more cells and / or network entities) during any time window.

[0079] As explained above, NIaaS is a holistic Al framework for telecommunication (teleco) systems that offers a suite of Al-native micro-services, that include data engineering framework services for handling observability and configuration data pipelines, data science framework services for generating statistics relevant to data and towards detecting statistical outliers, Al ML model training framework services for training causation, classification, forecasting, recommendation models and so on, Al ML life cycle management framework services for managing the life cycle of models such as version upgrade, rollback and so on, Al ML inference framework services for live data feeds, Al test framework services for digital twins, and Al ML visualization framework services, and so on.

[0080] This multi-dimensional imputation technique for generating observability data using Generative Al-based "deep fakes” is offered as part of NIaaS data science framework services. That is, the NIaaS framework producing data science services registers its capability of generating data deepfakes for 0AM0017389INP / 4688observability data and the corresponding data types, which can then be discovered and leveraged by authorized service consumers for Al-driven intelligence and optimization.

[0081] Techniques Employed as of the Present Disclosure:• Linear interpolation techniques, such as linear averaging or moving average, were used for interpolation of missing values. However, such techniques are useful only if the values change linearly with time. But the telecon observability data do not necessarily change linearly with time. Moreover, there are hundreds to thousands of observability data for predictive modelling of KPIs, and linear interpolations often require different models to be developed for different types of observability data.• Time-series based imputation techniques, such as ARIMA time series, are used to unearth non-linear patterns in data across time. While these models unearth non-linear patterns in data across time, they do not account for spatial dependencies across the data types towards accurate predictions. Moreover, time-series based imputation techniques also suffer from scalability, i.e., different time-series models will have to be developed for different types of data, and the number of such models keeps growing with the number of 0AM observability data types generated by the NFs.• Deep learning-based imputation techniques, such as Recurrent Neural Networks - Long Short Term Memory [LSTM], account for temporal patterns as well as spatial dependencies among 0AM observability data types towards predicting and imputing the missing observability data. While LSTMs perform relatively well on accounting for temporal dependencies and spatial dependencies in time-series data towards predicting and imputing the target KPI, they do not perform as well in imputing the different types of missing 0AM observability data.0017389INP / 4688

[0082] Hence, in this disclosure, described is a multi-dimensional data deepfake technique for imputing teleco 0AM observability data using Generative Adversarial Networks [GAN] that trains a single model for imputing any number of missing or outlier data types by unearthing the intricate spatial and temporal relationship across the different 0AM observability data types, of one or more interdependent network functions and / or network entities (such as, cells] and / or network resources. The multi-dimensional data deep-fake model imputes over 100 missing and / or outlier data types for each cell across different time intervals, and yields an accuracy of over 72% when compared against actual values.

[0083] Aspects of the Disclosure

[0084] 1.1 GAN Definition

[0085] GANs have two neural networks, a generator G and a discriminator D, which are trained simultaneously through adversarial training. The generator G creates new data instances that are similar to some existing data, while the discriminator D tries to distinguish between real and generated data. The training process involves a minimax game, where the generator G and discriminator D are in competition.

[0086] 1.1.1 Loss Function for discriminator [LD:

[0087] The discriminator D aims to maximize the probability of correctly classifying real samples and minimizing the probability of incorrectly classifying generated samples as real. The discriminator's loss function can be defined as the negative log likelihood of these classifications:LD= -£z~pW[log(D(x)] - Ez p(z)[log (1 - (G(z))] Where,p(x is real data distribution,p(z) is noise distribution,G(z) represents generated fake data samples, i.e., the output of the discriminator, given noise z,0017389INP / 4688D x) represents probability that x is real,D(G(z)) represents probability that G(z), the fake sample, is classified as real.Here,

[0088] — Ez„.p(z)[log (1 — £)(G(z))] refers to the negative log likelihood that G(z) is classified as fake, i.e., if the probability that G(z~), the fake sample, is classified as real, the discriminator’s loss function is higher.

[0089] “X-pm [log (X ( )] refers to the negative log likelihood that x is classified as real, i.e., if the probability that x, the real sample, is classified as real, the discriminator's loss function is lower.

[0090] That is, the discriminator's loss function is lower if the discriminator correctly classifies G’(z) as fake, and x as real. The goal of the GAN algorithm is to have the generator fool the discriminator (when the discriminator is classifying "fake” data as real) increasingly, and increase its loss.

[0091] Figure 11 shows a discriminator loss backpropagation (Ref: https: / / developers.google.com / machine-learning / gan / discriminator).

[0092] 1.1.2 Loss Function for generator (LG):

[0093] The generator's objective is to generate data samples G(z) that the discriminator cannot distinguish from real data. The generator's loss function can be defined as the negative log likelihood of the discriminator accepting the generated samples as real:= -£Z~P(Z) [log (£> (6 (z))]

[0094] The objective for the generator G is to generate samples G(z) in such a way thatD(G(z) approaches 1, making it difficult for the discriminator D to distinguish between real and fake.0017389INP / 4688

[0095] That is, the generator’s loss function is lower if the discriminator D falsely classifies G(z) as real. The goal of the GAN algorithm is to have the generator G minimize its loss.

[0096] Figure 12 shows a generator loss backpropagation (Ref:https: / / developers.google.com / machine-learning / gan / generator].

[0097] The generator loss is fed back to the generator G and discriminator D neural networks, using which the neural networks subsequently update their weights and biases.

[0098] 1.1.3 Minimax loss and Cross-entropy variant:

[0099] The discriminator and generator loss are fed back to the neural networks, as seen in Figures 11 and 12. The objective of GANs is to train the generator network (G) to produce data samples that are indistinguishable from real data according to the discriminator network (D). This process can be defined as a minimax optimization problem. The objective function, also known as the loss function, can be expressed as:min maxFYD, G)G D

[0100] V D, G) represents the value function that is trying to be minimized and maximized simultaneously. The training process of GANs involves finding the Nash Equilibrium, where the generator and discriminator reach a stable state. This is achieved by solving the minimax problem:minmaxEtD, Gj = minmax(Ln+ 1GG D C D

[0101] In this minimax game, the discriminator D is trying to minimize LD, and the generator G is trying to minimize LGwhile maximizing LD. The optimal0017389INP / 4688solution occurs when the generator G produces data that is indistinguishable from real data, and the discriminator D is 50% confident for both real and generated data.

[0102] Solving this minimax game involves iterative optimization techniques like stochastic gradient descent (SGD) or variants like Adam based on backpropagation techniques, where both the generator G and discriminator D are updated in alternating steps to reach a stable equilibrium.

[0103] 1.1.4 Wasserstein Loss:

[0104] Unlike the minimax loss where the discriminator D determines if the generated data is real or fake, the Wasserstein loss function is based on a modification of the GAN in which the discriminator does not actually classify if the generated data instance is real or fake, but it instead outputs a number based on the Wasserstein Distance (also known as Earth Mover’s distance) between the real and generated (fake) data distributions. The Wasserstein distance can be informally described as the minimum "cost" required to transform the distribution of generated data into the distribution of real data.

[0105] The idea is to measure how far apart the real and generated distributions are, rather than classifying whether the data samples are real or fake. In Wasserstein GAN, the discriminator D is also called critic, and it does NOT output a probabilistic value that is strictly within 0-1 range.

[0106] In Wasserstein GAN, the loss function for the discriminator D is given by:LD= Ex~p(x)[D(x)] - £z~p(z)[D(G(z))]

[0107] The generator’s objective is to minimize the negative of the discriminator's output, which is given by:0017389INP / 4688LG - -EZ~P(Z)[D(G Z)'>)]

[0108] 1.1.5 With Recurrent Neural Network variant:

[0109] When an RNN layer is added on top of the GAN, the GAN framework remains similar to the standard one, but the generator and discriminator networks now operate on sequences or time-series data. RNNs are designed to capture sequential information and temporal dependencies in data. The hidden state of the RNN at time step t(h(t)) is calculated based on the input at time stepand the previous hidden step ht_x):h(t) = RNN xt,ht-1)

[0110] This recurrence relation allows the RNN to maintain information about past inputs and capture sequential patterns in the data. The discriminator D with an RNN layer can be represented as follows:

[0111] Discriminator with RNN (DRNN)

[0112] The discriminator D with an RNN layer can be represented as follows:DRNN^X1, X2,..., XT) — a(WRNN.hT+ bRNN)Where,xltx2,...,xTare the input data points over time,hTis the final hidden state of the RNN after processing the entire sequence, W / 'RJ JV and bRNNare weights and biases for the RNN output,<7 is the activation function (like sigmoid) transforming the RNN function into a probability.

[0113] Generator G with RNN [GRNN

[0114] The generator G with an RNN layer can be represented similarly, generating a sequence of data points:0017389INP / 4688GRNN (Zl<Z2> ■■■ >Zt) ~ (X1-X2’ ■■■ ’XT)

[0115] The training objective for the GAN with an RNN layer involves optimizing the parameters of both the generator and discriminator to achieve a stable equilibrium. This involves minimizing the following loss function:min max[Ex^w[log(RWW(%))] + £'z^(z)[log(l -RWW(GRWW(z))] GRNN DRNN

[0116] The RNNs in both the generator and discriminator are unfolded through time steps to compute gradients using techniques like Backpropagation Through Time.

[0117] 1.2 Problem statement

[0118] In a telecommunications network, a network function or an infrastructure platform (in which the NFs are deployed) and its associated entity or resource can generate 0AM observability data (PM, CM, FM, events, Trace, state information, and so on) periodically or during network events, based on consumer subscriptions. The amount of data can be in the order of several hundreds to 1000s of data types, and the periodicity can typically range in the order of minutes with some low-latency consumers (such as 0-RAN Near-RT RIC) requiring the data to be made available in the order of hundreds of milli-seconds. However, data consumption is affected by missingness in data, compromises in data integrity, and statistical outliers, which subsequently impacts deriving and generating intelligence. The data and the related missingness are quantified in Section A. The magnitude of adverse impact is higher, if the subscribed reporting periodicity indicates a higher frequency of reporting. The problem statement is thus to overcome the adverse impact of data missingness, integrity compromises, data drifts and outliers and so on from generating intelligence about network and infrastructure KPIs in large-scale telco deployments.

[0119] Exemplary advantages of the disclosure are:0017389INP / 46881. Generating realistic data deep fakes of missing observability data or integrity- compromised data or statistical outliers using Generative Al-based imputation techniques.2. The imputation is multi-dimensional in nature, that is, a single Generative Al model is used to impute any number of observability data, such that the Wasserstein loss between the generated data and the real data are minimized.

[0120] In more formal terms, lettG RNbe the N-dimensional 0AM observability data pertaining to network entity i at time t. Let Xi tc Xi tbe the subset of 0AM observability data pertaining to network entity i at time t which is impacted by data missingness or statistical outliers or data drifts due to integrity compromises. Let G(Z)i t£ RNbe the generator’s output of the fake data pertaining to network entity i at time t, corresponding to Xi t. The goal is to generate G(Z)i t, for any i and t, such that the Wasserstein loss between Xi tand G(Zft, for the corresponding i and t, is minimized.

[0121] 1.3 Methodology

[0122] This section details the training and inference methodologies.

[0123] 1.3.1 Training methodology

[0124] 1.3.1.1 Principles

[0125] Training methodology is based on the following principles:

[0126] 1. Time-series data and identifications-.

[0127] The state data comprises of 0AM FCAPS data (comprising PM, CM, FM, Trace, Logs, and so on pertaining to the network entity that are generated from the corresponding NF or infrastructure component. The state data is standardized, scaled and normalized.0017389INP / 4688

[0128] The observability data is obtained from the state data and is prepared as time-series data to be fed to the recurrent GAN training engine, in order to leverage temporal recurrence of the observability state across the network entities, in addition to spatial dependencies across the observability data, towards enabling the generator for better learnings of inter-dependencies across observability data towards realistic generations of data deep-fakes.

[0129] Temporal recurrence is leveraged only for observability data pertaining to individual network entities, and so, time-series data should be entityspecific. Temporal recurrence is based on time-of-day, day-of-week (i.e., weekday, weekend, mid-week), seasonality, and so on. Hence, the observability dataset should also have network entity identifier (so as to explore temporal recurrence of observability data for individual entities, and NOT to explore temporal recurrence across entities), time of day, day of week and seasonality.

[0130] In order to enable the model to accurately learn about both temporal and spatial dependencies, the LSTM variant of GAN is used.

[0131] 2. Genuinely-missing data in the original dataset and deliberately-imputed data:

[0132] The original observability data can have missingness (reported as NULL). However, in order to train the model to generate "realistic deep fakes”, deliberately mark a subset of non-missing (i.e., NON-NULL) observability data as "missing”, so as to enable the generator to be trained with a loss function towards learning to accurately impersonate such non-missing observability data. This enables the trained model to generate and impute realistic deep fake of genuinely-missing data from the original dataset.

[0133] It is noted that any subset of data can go missing in the observability data reported from the network function or infrastructure component. Thus, deliberately mark varying subsets of observability data across the training0017389INP / 4688records / streams as ‘missing’. That is, between any two rows in the training records, the subset of observability data, deliberately marked as ‘missing’ is different.

[0134] 3. Mask Ma trix:

[0135] The mask matrix is a 0-1 matrix to enable the generator to distinguish deliberately-masked data, that needs to be imputed, from the original data and the genuinely-missing data in the dataset.

[0136] 4. Loss Functions;

[0137] Generator-specific Loss Function-. Towards enabling the generator neural network to impute data deep fakes realistically such that they are close enough to the original data, there is a discriminator-specific loss function used by the generator network during the model training phase that uses supervised learning to minimize the error.

[0138] Wasserstein Loss Function-. The Wasserstein loss for the discriminator and generator, described in Section 1.1.4, are fed back to the discriminator and generator neural networks via backpropagation towards enabling the generator to outsmart the discriminator, while at the same time, ensuring that the discriminator is not fooled easily.

[0139] 5. Hint Generator and Hint score: A hint matrix is introduced to guide the discriminator in learning specific features or properties in the data. The hint generator matrix can supply auxiliary data or hints about the true labels, specific features, or attributes in the training data. By doing this, the discriminator learns to distinguish between real and generated data not only based on the visual features but also based on the contextual information or characteristics associated with the data.

[0140] In GANs, mode collapse is a common problem where the generator produces limited varieties of outputs. By providing hints through a matrix, the discriminator has a more nuanced basis to evaluate real versus generated data,0017389INP / 4688potentially encouraging the generator to produce a wider range of outputs. The additional hints can act as a regularization mechanism, guiding the discriminator's gradients in ways that stabilize the adversarial training process. This is particularly useful in applications where the GAN is prone to divergence. In this setup, the hint matrix is often a conditional input that is meant to help the discriminator learn class-specific features or handle labelled data towards influencing the training dynamics.

[0141] 6. Multi-layer Generator and discriminator with LSTM feedforward: The generator and discriminator GAN network with LSTM variant, known as Recurrent -Generative Adversarial Network, is shown in Fig 13. Each network entity-specific time-series record is fed to the generator network which is an LSTM variant of GAN, comprising:• RNN with individual LSTM time blocks corresponding to the individual timestep-specific sub-records of a time-series record. The output (LSTM cell state and hidden state) generated from each LSTM time block is fed forward to the next LSTM time block, in order to capture temporal recurrence.• Within each LSTM time block of the generator, use a deep learning neural network for the three gates (i.e., the forget gate, the input gate and the output gate) and the sigmoid layer (to shape the values of the activation function to [-1,1]), and the weights and biases of the neural networks are Back-propagated through time (BPTT) using appropriate optimizer functions (such as Stochastic Gradient Descent, Adam Optimizer, and so on) and loss functions (such as MSE, Wasserstein loss, and so on) towards generating realistic data deep fakes for the generator, and towards intelligently classifying fake data from real data at the discriminator.

[0142] Figure 13 shows a recurrent GAN with LSTM variant for generator and discriminator.

[0143] 1.3.1.2 Training dataset preparation0017389INP / 4688

[0144] Let Ai TG JiNxN(2-dimensional real numbers] be defined as an N-dimensional time-series record of N-dimensional 0AM observability data for entity i and time window T, given by:

[0145] where Xi t^G is the observation vector for entity i with respect to the individual time-step t(k) of the time window T, comprising Key Performance Indicators [KPI], Performance Measurement [PM] counters, Configuration Management (CM) data. Trace data and Alarms Fault Management (FM) data pertaining to the network entity i (and other related entities with respect to i ) at time-step t(k). LetG Xi t^ be any feature j corresponding to entity i and time-step t( / c) of the observability vector.

[0146] The observability data is based on the state vector data reported by the Network Function or infrastructure component, of which i is an associated entity. The state vector pertaining to the time window T, corresponding to entity i, is given by Di TSi,tW

[0147] Let Si t G Si t^ be an 0AM feature j corresponding to entity i and time-step t( / c) of the state vector, reported by the network function or infrastructure component. NOTE that, based on the description in Section C.3.1.1, in addition to the 0AM observability and state data reported in the state vector, the state vector includes:l(i) indicating the local ID of the entity i0017389INP / 4688T indicating the time of the day,D indicating the day of the week,- S indicating the seasonality (such as month, season, and so on].

[0148] The observation feature, corresponding to the state data feature i,j,t(k isxi,j,t(k)(si,j,t(ky> if 0xi,j,t(k):= j = 0\ if $i,j,t(k) Si,j,t(k)Here, Stj,^c is the randomly-chosen subset of the state vector Siit kcomprising randomly-selected features for imputation, against which "deep fake" data can be generated by training our GAN -based models.

[0149] Figure 14 shows a GAN with LST Training Architecture

[0150] LetBi TE JlNxN(2-dimensional floating-point numbers] be defined as an N-dimensional time-series record of N-dimensional 0AM observability data, given by:Zi.tWZi.t(2)-Zi,t(K)

[0151] Let zi)7-Ezi,t k) be the random Gaussian-distributed noise corresponding to ytjitk) i-e., any feature j of entity i and time-step t( / c) of the vectorzi,t(k)> where:f 0; if = 1[JV'(O, <72); i / yiJ>tm= o0017389INP / 4688

[0152] Let Ci TG ZNxN(2-dimensional integers) be defined as an N- dimensional time-series record of N-dimensional 0AM observability data, given by:

[0153] Letbe the 0 — 1 mask value corresponding to, i.e., any feature j of entity i and time-step t( / c) of the observation vector Xi>t kandsi,j,t(k) > i-e., any feature j of entity i and time-step t(fc) of the state vectorSi,t(k where:( 1; if -i:=j ifXi,j,t<ik)= —1andsi,j,t(Jc) = 0\0, if 1 and Sijt^k^ G

[0154] The description for Figurel4 is as follows:• From Matrix D comprising of the time-series state vector that includes network entity and time identifiers, matrices A, B, C are generated, as discussed above.• These matrices are fed to the generator network towards imputing data deep fakes. In order to impute them realistically, the generator uses a loss function (such as Mean Squared Error, i.e., MSE) for training against the original time-series state matrix D.• The matrix C is fed as input to the hint generator that generates a hint score matrix with l’s for original non-null and genuinely-missing null data, and a hint score for data deliberately marked as missing in the matrix " A”.• The generated matrix with data deep fake imputations and the hint matrix are fed as inputs to the discriminator neural network for generating the Wasserstein distance outputs, from which the Wasserstein loss is computed.0017389INP / 4688• The Wasserstein loss is backpropagated to the discriminator as well as to the generator networks. To the generator networks, there is convolution involving the Wasserstein loss function and the MSE loss for back- propagating the weights and biases to the generator network.

[0155] 1.3.2 Inference methodology

[0156] The recurrent GAN model with the LSTM variant, trained as above, is deployed in the inference engine which receives observability data feed from the network entity or infrastructure component. The engine can simultaneously do online update of the model or can act in pure inference-only mode.

[0157] In inference-only mode, there is no deliberate missingness that is introduced in the time-series observability vector. The data, originally missing in the state vector corresponding to the network entity, generated from the network function, is inputted by the generator using realistic data deep fakes.• The network entity-specific time-series state vectors, represented by Matrix " D", is broken down into the following matrices:■ Entity-specific time-series observability vector, given by matrix " A” (where the missing data is replaced by -1). The network entity identifier and the time-stamp identifier fields are retained,■ Matrix " B" including Gaussian-distributed noise, corresponding to the missing data, for which deep fake data can be generated by the generator network, and■ Mask Matrix " C", that uses 0 to indicate missing data in the corresponding rows / columns in Matrix “A” and 1 to indicate other non-null data in the corresponding rows / columns in Matrix " A”.■ These matrices are fed to the GAN generator network model, trained as described in Section C.3.1, that is deployed in the inference engine,0017389INP / 4688towards generating realistic data deep fakes. This is shown in Figure 5.■ In the online update mode jointly along with inference, the ratio of model update instances to inference-only instances is controlled by the learning rate e G [0,1], During online update, the engine engages in re-training, as discussed in Section C.3.1, towards updating the model (in terms of the policy function or value function] with further parameter state space exploration by introducing additional missingness in any random subset of the time-series state vector.

[0158] Figure 15 shows a GAN with LSTM - Inference-only architecture.

[0159] Section 2: Evaluations

[0160] 2.1: Topology and implementation

[0161] Described are implementations for imputing data deep fakes is evaluated in an operational 5G NSA EN-DC Open RAN production deployment:Number of sites = 25Number of eNB CU-CPs = 3Number ofgNB CU-CPs = 3Number of eNB CU-UPs = 3Number of gNB CU-UPs = 3Number of eNB DUs = 25Number of gNB DUs = 25Number of sectors per DU = 3Number of E-UTRA cells per DU = 4Number of NR cells per DU = 4Total number of LTE and NR cells ~= 100Maximum number of active UEs per 4G cell = 2000017389INP / 4688Number of OAM states generated per cell (from both CU and DU), including = 125OAM reporting periodicity = Every 5 minsNumber of LSTM blocks in the recurrent neural network = 24 (2 hours, with 5 mins reporting each).

[0162] Figure 16 shows an architecture of an R-GAIN implementation, that was deployed in Open RAN. Based on discussions in Section 1, the matrix D has the dimensions 100 x 24 x 125, described as follows:Number of network entities (cells) = 100, since each LTE / NR cell as a network entity is considered.Number of timesteps = 24, since each time-series records of 2 hours of length, where each constituent sub-record corresponds to a 5-min reporting periodicity and is timestamped. Hence, 12 units of 5-min timestep-specific sub-records every hour, and 24 timesteps for 2 hours.

[0163] Each network entity (LTE / NR cell)-specific time-series data is organized as 3 matrices (A, B, C as described in Section C), and hence there are 375 columns, spanning across all 3 matrices, and with two additional neurons, the number of input features in the input layer to the generator becomes 377, as shown in Fig 6. The hidden LSTM layer configurations with the neurons are also shown in Fig 6, and the output layer corresponds to 125 neurons, since any missing subset of the observability data can be imputation by the generator towards generating realistic data deep fakes. Likewise, the discriminator also determines the output, towards distinguishing as real or fake, corresponding to any subset of the 125 features "generated” by the generator. This architecture implementation is outlined in Fig 6.

[0164] 2.2: Results0017389INP / 4688

[0165] This section presents the data deep fake imputation accuracy results of a few representative input features out of the 125 observability features, considered for a given cell.

[0166] Figure 17 shows an evaluation of a data deep-fake of Wideband Channel Quality Indication [CQI], Fig 17 evaluates the accuracy and correlation of the imputed number of CQI samples and the corresponding actual number of CQI samples, corresponding to a set of CQI values {1, 4, 7, 10, 13, 15}, across all cells. The accuracy values indicate a highly realistic deep faking of the number of CQI samples for different values of the CQI, and a correlation of over 85%.

[0167] Figure 18 shows an evaluation of a data deep fake of Data Radio Bearer throughput volume and time. Fig. 18 evaluates the accuracy and correlation of the inputted number of DRB throughput volume and the DRB throughput time with the corresponding actual DRB throughput volume and DRB throughput time, across all cells. The accuracy values indicate a highly realistic deep faking of the DRB throughput volume and throughput time, and a correlation of over 90%.

[0168] Figure 19 shows an evaluation of a data deep fake of Physical Radio Block (PRB) utilization and PRB congestion. Fig 19 evaluates the accuracy and correlation of the imputed PRB utilization and the percentage of time the PRBs are congested, with the corresponding actual PRB utilization and PRB congestion, across all cells. The accuracy values indicate a highly realistic deep faking of the PRB utilization and congestion, and a correlation of over 99%.

[0169] Figure 20 shows an evaluation of a data deep fake of actual number of initial Evolved Radio Access Bearer (E-RAB) establishment attempts and actual number of E-RAB release attempts. Fig.20 evaluates the accuracy and correlation of the imputed number of initial E-RAB attempts and the number of E-RAB release attempts, with the corresponding actual number of initial E-RAB attempts and number of E-RAB release attempts, across all cells. The accuracy values indicate a0017389INP / 4688highly realistic deep faking of the E-RAB establishment and release attempts, and a high correlation of over 98%.

[0170] Figure 21 shows an evaluation of data deep fake of an actual number of samples with high Reference Signal Received Power (RSRP), medium RSRP and poor RSRP. Fig.21 evaluates the accuracy and correlation of the imputed number of samples corresponding to the distribution of very high RSRP values, medium RSRP values and poor RSRP values, with the corresponding actual number of samples corresponding to very high, medium and poor RSRP values, across all cells. The accuracy values indicate a highly realistic deep faking of the RSRP distributions, and high correlations for sample distributions of these measurements.

[0171] Figure 22 shows an evaluation of data deep fake of a number of Radio Resource connection (RRC) establishment attempts, number of Random Access Channel (RACH) attempts and number of contention RACH reports. Figure 22 evaluates the accuracy and correlation of the imputed number of RRC connection establishment attempts, RACH attempts and number of contentious RACH reports with the corresponding actual number of these measurement values, across all cells. The accuracy values indicate a highly realistic deep faking of these measurements.

[0172] Figure 23 shows an evaluation of a data deep fake of number of a number of handover attempts and the number of too-early HO attempts. Figure 23 evaluates the accuracy and correlation of the imputed number of connected mode mobility handover (HO) and number of too early HO attempts, against the actual values of these measurements, across all cells. The accuracy values indicate a highly realistic deep faking of these measurements, with a correlation score of over 92%.

[0173] Figure 24 shows an evaluation of a data deep fake of a number of samples with high, medium and low neighbor cell interference. Figure 24 evaluates the accuracy and correlation of the imputed number of samples with high, medium and low neighbor cell interference against the actual number of samples pertaining to high, medium and low neighbor cell interference measurements, across all cells.0017389INP / 4688The accuracy values indicate a highly realistic deep faking of these measurements, with a correlation score of over 93%.

[0174] Other Applications

[0175] The data deep fake multi-dimensional imputation capability is offered as a service by the data science services producer of the NIaaS framework for consumption by authorized consumers. Consumers can subscribe to the capability of the NIaaS framework to automate generate of realistic multi-dimensional data deep fake imputations, when the observability data goes missing or when the data reporting is adversely impacted by statistical outliers.

[0176] It will be understood that implementations and embodiments can be implemented by computer program instructions. These program instructions can be provided to a processor to produce a machine, such that the instructions, which execute on the processor, create means for implementing the actions specified herein. The computer program instructions can be executed by a processor to cause a series of operational steps to be performed by the processor to produce a computer-implemented process such that the instructions, which execute on the processor to provide steps for implementing the actions specified. Moreover, some of the steps can also be performed across more than one processor, such as might arise in a multi-processor computer system or even a group of multiple computer systems. In addition, one or more blocks or combinations of blocks in the flowchart illustration can also be performed concurrently with other blocks or combinations of blocks, or even in a different sequence than illustrated without departing from the scope or spirit of the disclosure.

Claims

0017389INP / 4688What is claimed:

1. A method comprisinggenerating a realistic data deep fake of missing observability data, integrity-compromised data, or statistical outliers using Generative Al-based imputation for Operations, Administration and Maintenance (OAM) observability data for a Radio Access Network.

2. The method of claim 1, further comprising:using a single Generative Al model used to impute a number of the observability data, so thata Wasserstein loss between the generated data and the real data are minimized, whereinXi tE RNis an N-dimensional 0AM observability data pertaining to network entity i at time t;Xi t£ xi tis a subset of 0AM observability data pertaining to network entity i at time t which is impacted by data missingness or statistical outliers or data drifts due to integrity compromises; andG(Z); t£ RNis a generator’s output of the fake data pertaining to network entity i at time t, corresponding to Xi tso that G(Z)jtis generated for any i and t, and so that the Wasserstein loss between Xi:tand G(Zt, for the corresponding i and t, is minimized.

3. The method of claim 1, further comprising:obtaining the observability data from a state data comprising 0AM Fault Management (FM) data, Configuration Management (CM) data, Performance Management (PM) data (FCAPS) pertaining to a network entity that are generated from a corresponding Network Function (NF) or infrastructure component.0017389INP / 4688preparing the observability data as time-series data for a training dataset; andtraining the Generative Al with a recurrent Generative Adversarial Network (GAN) training engine to leverage temporal recurrence of an observability state across the network entities, in addition to spatial dependencies across the observability data, to enabling the generator for better learnings of inter-dependencies across observability data towards realistic generations of data deep-fakes.

4. The method of claim 3, wherein the recurrent GAN comprises a Long Short Term Memory (LSTM) variant.

5. One or more non-transitory computer readable media (NTCRM) comprising instructions that, when executed by one or more processors, cause a computer to:generate a realistic data deep fake of missing observability data, integrity-compromised data, or statistical outliers using Generative Al-based imputation.

6. The one or more NTCRM of claim 5, wherein the execution of the instructions causes the computer to:use a single Generative Al model used to impute a number of the observability data, so that a Wasserstein loss between the generated data and the real data are minimized, whereinX_(i,t)GRAN is an N-dimensional OAM observability data pertaining to network entity i at time t;X_(i,t] cX _(i,t) is a subset of OAM observability data pertaining to network entity i at time t which is impacted by data missingness or statistical outliers or data drifts due to integrity compromises; andG(Z] _(i,t)cR N is a generator’s output of the fake data pertaining to network entity i at time t, corresponding to XA_(i,t) so that [G(Z)]_(i,t] is generated0017389INP / 4688for any i and t, and so that the Wasserstein loss between XA_(i,t) and [G(Z) _(i,t), for the corresponding i and t, is minimized.

7. The one or more NTCRM of claim 5, wherein the execution of the instructions causes the computer to:obtain the observability data is obtained from a state data comprising OAM Fault Management (FM) data, Configuration Management (CM) data, Performance Management (PM) data (FCAPS) pertaining to a network entity that are generated from a corresponding Network Function (NF) or infrastructure component;prepare the observability data as time-series data for a training dataset; andtrain the Generative Al with a recurrent Generative Adversarial Network (GAN) training engine to leverage temporal recurrence of an observability state across the network entities, in addition to spatial dependencies across the observability data, to enabling the generator for better learnings of interdependencies across observability data towards realistic generations of data deep -fakes.

8. The one or more NTCRM of claim 7, wherein the recurrent GAN comprises a Long Short Term Memory (LSTM) variant.

Images