Efficient data collection and model architecture for ai-ML based positioning
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2026-02-04
- Publication Date
- 2026-08-13
Smart Images
Figure IB2026051044_13082026_PF_FP_ABST
Abstract
Description
[0001] EFFICIENT DATA COLLECTION AND MODEL ARCHITECTURE FOR Al- ML BASED POSITIONING
[0002] Field of the Invention
[0003] The present invention relates to wireless communication systems, and more particularly to efficient data collection and model architecture for artificial intelligence and machine learning (AI-ML) based positioning in non-line-of-sight (NLoS) scenarios.
[0004] Background of the Invention
[0005] Modern communication systems employ various methods to determine the position of a node, including downlink time difference of arrival (DL-TDoA), uplink time difference of arrival (UL-TDoA), downlink angle of departure (DL-AoD), uplink angle of arrival (UL-AoA), multiple round-trip time (m-RTT), and enhanced cell ID (ECID). These techniques generally rely on the availability of reliable line-of-sight measurements for achieving acceptable positioning accuracy. However, in deployment environments dominated by non-line-of-sight (NLoS) conditions, such techniques frequently produce inaccurate position estimates due to distorted timing and angular measurements reported by receiver nodes.
[0006] The inaccuracies in NLoS scenarios arise from multiple technical factors, including limited timing and angular resolution, severe multipath propagation, poor signal-to-noise ratios (SNRs), and the absence of a direct propagation path between transmitting and receiving nodes. Although modern wireless systems support large bandwidths and multiple antenna configurations, these enhancements alone do not sufficiently mitigate the fundamental degradation of positioning measurements under NLoS conditions. While time-domain and spatial averaging techniques may partially reduce noise effects, they do not adequately compensate for structural measurement distortions caused by NLoS propagation.To address these limitations, artificial intelligence and machine learning (AI-ML) techniques have been proposed for positioning applications, wherein models are trained to learn propagation characteristics such as power profiles, delay profiles, angle profiles, and Doppler profiles. However, the effectiveness of such AI-ML-based positioning approaches is highly dependent on the richness and completeness of the reported channel information, which must capture sufficient site-specific and environment-specific characteristics.
[0007] In practice, existing wireless communication systems often provide only limited or noisy measurement reports that lack the detailed channel characteristics required for AI-ML models to reliably infer accurate position estimates, particularly in complex NLoS environments. As a result, AI-ML models trained on incomplete or sparse measurement data frequently fail to generalize across deployment scenarios, leading to degraded positioning performance.
[0008] Furthermore, direct AI-ML-based positioning methods typically require user equipment or network nodes to report detailed channel information, such as full channel impulse response data or high-resolution channel state information, to a centralized positioning or training server. While such detailed reporting can improve positioning accuracy, it imposes a substantial uplink signaling burden, increases processing complexity, and may negatively impact overall network performance. Conversely, reducing the amount of reported channel information to alleviate uplink overhead often results in insufficient data for accurate positioning, thereby creating a fundamental trade-off between positioning accuracy and reporting efficiency.
[0009] Accordingly, there exists a technical challenge in achieving accurate AI-ML-based positioning under NLoS conditions while simultaneously reducing channel information reporting overhead. Existing solutions do not provide an effective mechanism to reconcile this trade-off, highlighting the need for improved methods and systems that can enable accuratepositioning using limited measurement reports without imposing excessive communication and computational burden on wireless networks.
[0010] Objective of the Invention
[0011] The principal objective of the present invention is to address the technical challenge of achieving accurate positioning in wireless communication networks under non-line-of-sight (NLoS) conditions while reducing the amount of channel information required to be reported by user equipment or network nodes, thereby minimizing signaling overhead without compromising positioning accuracy.
[0012] Another objective of the present invention is to enable reconstruction of channel characteristics relevant for positioning from concise input data, such that accurate location estimation can be performed without requiring transmission of raw or complete channel measurements.
[0013] Another objective of the present invention is to improve the robustness and adaptability of AI-ML-based positioning models by facilitating systematic association of input data and output data used during training and inference, thereby enhancing model performance across diverse deployment environments.
[0014] Another objective of the present invention is to ensure consistency between training and inference phases of AI-ML-based positioning models by accounting for essential network and propagation parameters, thereby reducing performance degradation caused by model drift in real-world deployments.
[0015] Another objective of the present invention is to support monitoring and maintenance of AI-ML-based positioning performance by enabling detection of performance degradation and facilitating model update or retraining based on observed positioning errors or changes in data distribution.
[0016] A further objective of the present invention is to provide a scalable and flexible AI-ML-based positioning framework that is applicable to bothtwo-dimensional and three-dimensional positioning scenarios and is compatible with advanced and next-generation wireless communication systems, including 5G-Advanced, sixth-generation networks, and emerging loT applications.
[0017] Summary of the Invention
[0018] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.
[0019] According to an aspect of the present disclosure, a method for artificial intelligence and machine learning (AI-ML) based positioning in wireless communication networks is provided. The method includes receiving concise channel information from at least one receiver node, the concise channel information comprising timing information, angle information, and a line-of-sight or non-line-of-sight (LoS / NLoS) confidence indication. The method further includes processing the concise channel information using a first AI-ML model to reconstruct channel characteristics relevant for positioning and processing the reconstructed channel characteristics using a second AI-ML model to estimate a position of a target node.
[0020] In some aspects, the concise channel information may further include a timestamp associated with a channel measurement and a quality indicator indicative of measurement reliability. The timing information may include one or more of time of arrival, relative time of arrival, or reference signal time difference, and the angle information may include angle of arrival or angle of departure, depending on a receiver configuration.
[0021] According to another aspect of the present disclosure, a system for AI-ML based positioning in wireless communication networks is provided. The system includes a receiver configured to obtain concise channel information from one or more receiver nodes and a processor configured toexecute a first AI-ML model for reconstructing channel characteristics from the concise channel information and a second AI-ML model for estimating a position of a target node based on the reconstructed channel characteristics.
[0022] According to another aspect of the present disclosure, a non-transitory computer-readable medium storing instructions is provided, which, when executed by a processor, cause the processor to perform AI-ML based positioning by receiving concise channel information, reconstructing channel characteristics using a first AI-ML model, and estimating a position of a target node using a second AI-ML model.
[0023] The foregoing general description of the illustrative embodiments and the following detailed description thereof are merely exemplary aspects of the teachings of this disclosure and are not restrictive.
[0024] Brief description of the drawings
[0025] The figures described below depict various aspects of the system and methods disclosed herein. It should be understood that each figure depicts an embodiment of a particular aspect of the disclosed system and methods, and that each of the figures is intended to accord with a possible embodiment thereof. Further, wherever possible, the following description refers to the reference numerals included in the following figures, in which features depicted in multiple figures are designated with consistent reference numerals.
[0026] FIG. 1 illustrates sequence diagrams for different AI-ML-based positioning processes (100) in wireless communication networks, in accordance with one embodiment of the present invention.
[0027] FIG. 1(a) illustrates a sequence diagram for a location server-based uplink (UL) positioning process, where the user equipment (UE) transmits signals to the network, according to one embodiment of the present invention.FIG. 1(b) depicts a sequence diagram for a location server-based downlink (DL) positioning process, in accordance with one embodiment of the present invention.
[0028] FIG. 1(c) shows a sequence diagram for a user equipment-based positioning process, according to one embodiment of the present invention.
[0029] FIG. 2 illustrates a positioning model architecture for 2D positioning using AI-ML techniques (200), in accordance with one embodiment of the present invention.
[0030] FIG. 3 depicts a model architecture for AI-ML based 3D positioning (300), according to one embodiment of the present invention.
[0031] FIG. 4 shows a positioning framework for AI-ML assisted positioning (400), in accordance with one embodiment of the present invention.
[0032] FIG. 5 illustrates a positioning framework with three configurations for AI-ML based positioning systems (500), in accordance with one embodiment of the present invention.
[0033] FIG. 6 is a block diagram presents an example of a schematic hardware configuration of the network node (600), according to one embodiment of the present invention.
[0034] Persons skilled in the art will appreciate that elements in the figures are illustrated for simplicity and clarity and may have not been drawn to scale. For example, the dimensions of some of the elements in the figure may be exaggerated relative to other elements to help to improve understanding of various exemplary embodiments of the present disclosure.
[0035] Throughout the drawings, it should be noted that like reference numbers are used to depict the same or similar elements, features, and structures.
[0036] Detailed Description of the Invention
[0037] The following description, with reference to the accompanying drawings, is provided to assist in a comprehensive understanding of exemplary embodiments of the invention as defined by the claims and theirequivalents. It includes various specific details to facilitate such understanding; however, these details are to be regarded as merely exemplary. Accordingly, persons skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the invention. In addition, descriptions of well-known functions, procedures, and constructions are omitted for clarity and conciseness.
[0038] The terms and expressions used in the following description and claims are not intended to be limiting, but are used to enable a clear and consistent understanding of the invention. Accordingly, the following description of exemplary embodiments is provided for illustration purposes only and is not intended to limit the invention as defined by the appended claims and their equivalents. It is to be understood that the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. The term “substantially” is used to indicate that the stated characteristic or value need not be achieved exactly, and that reasonable variations, tolerances, and measurement inaccuracies known to those skilled in the art may be present without materially affecting the intended technical effect. A set is defined as a non-empty set including at least one element.
[0039] As used throughout this disclosure, the term “concise channel information” refers to a reduced-dimensional or compressed representation of channel measurements transmitted with lower reporting overhead than raw or complete channel information. The concise channel information comprises at least timing information, angular information, a line-of-sight or non-line-of-sight (LoS / NLoS) confidence indication, and an associated quality indicator, and is configured to preserve positioning-relevant characteristics while reducing uplink signaling and processing burden.
[0040] As further used herein, the term “virtual complete channel representation” refers to reconstructed channel information generated by an artificial intelligence or machine learning (AI-ML) model from the concisechannel information. The virtual complete channel representation approximates characteristics of a complete channel measurement by compensating for missing, noisy, or degraded channel components resulting from limited reporting, multipath propagation, and non-line-of-sight conditions, and is also referred to herein as reconstructed complete channel information.
[0041] Throughout this disclosure, a “first AI-ML model” refers to an AI-ML model configured to reconstruct a virtual complete channel representation from concise channel information, while a “second AI-ML model” refers to an AI-ML model configured to estimate a position of a target node based on the reconstructed virtual complete channel representation. As used herein, a “receiver node” may comprise a user equipment, a base station, a transmission reception point, or a reference node, depending on the positioning configuration and deployment scenario.
[0042] Figure 1 illustrates sequence diagrams (100) for different AI-ML-based positioning processes in wireless communication networks, according to aspects of the present disclosure. These processes include location server-based uplink (UL) positioning, location server-based downlink (DL) positioning, and user equipment-based positioning, each designed to optimize positioning accuracy using AI-ML models while minimizing network overhead. As used herein, a “receiver node” may comprise a user equipment, a base station, a transmission reception point, or a reference node, depending on the positioning configuration and deployment scenario.
[0043] Figure 1(a) illustrates a sequence diagram for a location serverbased uplink (UL) positioning process (100a), in which the user equipment (UE) transmits uplink reference signals and the network performs positioning measurements to estimate the UE’s location.
[0044] In one embodiment, the procedure begins when the Location Server receives a Location Services (LCS) request, which initiates the UL positioning process. Following receipt of the LCS request, the UE and theLocation Server exchange capability information through UE Request Capabilities and UE Provide Capabilities messages. This exchange allows the Location Server to determine the UE’s supported positioning features and measurement capabilities. The NG-RAN then provides TRP (Transmission Reception Point) information to the Location Server, enabling awareness of the network infrastructure relevant for positioning.
[0045] Based on the received capabilities and infrastructure information, the Location Server selects one or more positioning methods and sends a Positioning Information Request to the NG-RAN. In response, the NG-RAN determines appropriate uplink reference signal (UL RS) resources and returns a Positioning Information Response. The Location Server subsequently provides the UL RS configuration to the UE and triggers positioning activation, after which the UE activates UL RS transmission and transmits the configured reference signals.
[0046] Upon receiving the uplink reference signals, the NG-RAN computes positioning measurements, such as timing information, angle information, received power, and line-of-sight or non-line-of-sight indicators. The computed measurements are processed to generate concise channel information comprising selected timing information, angular information, a line-of-sight or non-line-of-sight confidence indication, and associated quality indicators, optionally including timestamps. These measurements may be associated with timestamps and quality indicators to improve robustness.
[0047] The NG-RAN provides the concise channel information to the Location Server, wherein a first AI-ML model reconstructs a virtual complete channel representation from the concise channel information. The Location Server then performs position estimation by processing the reconstructed virtual complete channel representation using a second AI-ML model, and finally responds to the original LCS request with the estimated UE position.
[0048] In parallel with the positioning procedure, the Training Server provides context assistance data and AI-ML model responses, supportingAI-ML model identification, training, selection, or inference as part of the positioning process. The lower portion of Figure 1(a) illustrates a periodic AI-ML model performance monitoring mechanism, including AI-ML model identification, training or selection, model performance monitoring requests and responses, performance metric calculation, and model update information. This monitoring framework enables continuous evaluation and adaptation of AI-ML models, ensuring sustained positioning accuracy under dynamic network and propagation conditions.
[0049] In some embodiments, the Location Server, Training Server, or another network node monitors a statistical distribution of positioning errors generated during inference over a plurality of positioning instances. The monitored error distribution may be compared against a stored or reference error distribution corresponding to a trained or validated model state. When a deviation between the monitored statistical distribution and the reference distribution exceeds a predefined threshold, the system triggers retraining, fine-tuning, or model update of at least one of the first AI-ML model or the second AI-ML model. This distribution-based monitoring enables detection of model drift, changes in propagation environment, or data distribution shifts, and ensures sustained positioning accuracy over time.
[0050] Figure 1(b) illustrates a sequence diagram for a location serverbased downlink (DL) positioning process (100b), in which the network transmits downlink reference signals and the user equipment (UE) performs positioning measurements based on the received signals. The procedure begins when the Location Server receives a Location Services (LCS) request, which initiates the DL positioning process.
[0051] Following receipt of the LCS request, the UE and the Location Server exchange capability information through UE Request Capabilities and UE Provide Capabilities messages, enabling the Location Server to determine the UE’s supported positioning features. The NG-RAN provides TRP information to the Location Server, ensuring that the Location Server hasup-to-date information about the network infrastructure relevant for positioning.
[0052] Based on the received capability and infrastructure information, the Location Server selects one or more positioning methods and performs Method Information Transfer toward the NG-RAN. The Location Server further provides assistance data to both the UE and the NG-RAN, including configuration information required for downlink positioning measurements. The NG-RAN then provides reference signal (RS) information and performs positioning activation, after which the NG-RAN transmits downlink reference signals for positioning purposes.
[0053] Upon receiving the downlink reference signals, the UE computes positioning measurements, such as timing information, angle information, received power, and line-of-sight or non-line-of-sight indicators, and generates concise channel information therefrom. The UE then provides the concise channel information to the Location Server. The Location Server applies a first AI-ML model to reconstruct a virtual complete channel representation and applies a second AI-ML model to estimate the UE’s position. The Location Server responds to the original LCS request with the estimated location.
[0054] AI-ML techniques are integrated into the DL positioning process through interaction with the Training Server, which supports AI-ML model identification, training, and selection. The Training Server provides context assistance data and AI-ML model responses that may assist the UE or Location Server in processing measurements or selecting appropriate models for positioning.
[0055] The lower portion of Figure 1(b) depicts a periodic AI-ML model performance monitoring mechanism, including model performance monitoring requests, content assistance data exchange, AI-ML model responses, performance metric calculation, model update information, and monitoring responses. This monitoring framework enables continuousevaluation and adaptation of AI-ML models, ensuring sustained positioning accuracy under varying radio and environmental conditions.
[0056] Figure 1(c) illustrates a user equipment-based positioning process (100c), in which the UE performs position estimation locally, thereby reducing reliance on centralized computation at the Location Server.
[0057] Following receipt of the LCS request, the UE and the Location Server exchange capability information through UE Request Capabilities and UE Provide Capabilities messages, allowing the Location Server to determine the UE’s positioning and processing capabilities. The NG-RAN performs TRP Information Transfer, providing the Location Server with relevant network infrastructure information required for positioning configuration.
[0058] Based on the received information, the Location Server selects an appropriate positioning method and provides assistance data to the UE and the NG-RAN. The assistance data includes configuration parameters necessary for positioning measurements and AI-ML inference. The NG-RAN transmits downlink reference signals for positioning, and the UE receives these signals for measurement processing.
[0059] The UE computes positioning measurements from the received signals and generates concise channel information, which is processed locally using a first AI-ML model to reconstruct a virtual complete channel representation. The UE then applies a second AI-ML model to estimate its location locally and provides the estimated location information to the Location Server, which responds to the original LCS request without performing the core position computation.
[0060] The lower portion of Figure 1(c) depicts a periodic AI-ML model performance monitoring mechanism, including content assistance data exchange, AI-ML model identification or selection, AI-ML model responses, performance metric calculation, model update information, and performance monitoring responses. This framework enables continuous evaluation and lifecycle management of AI-ML models used for UE-basedpositioning, ensuring robust performance under varying radio and environmental conditions.
[0061] The UE-based positioning approach enables reduced network load, lower latency, and increased autonomy in position estimation, while still allowing network-assisted configuration and centralized model lifecycle management.
[0062] FIG. 2 illustrates a positioning model architecture (200) designed for 2D positioning using AI-ML techniques, according to an embodiment. The positioning model architecture 200 processes a reconstructed virtual complete channel representation, generated by a first AI-ML model from concise channel information, as its input, and operates through multiple stages to extract relevant spatial features and generate precise position estimates for a target node. For clarity, the reconstructed virtual complete channel representation may be denoted as an input signal h(t) in the illustrated embodiment.
[0063] The model leverages deep learning techniques, including convolutional layers, transformer layers, and adaptive filtering, to enhance the accuracy and robustness of the positioning process, particularly in non-line-of-sight (NLoS) environments. In some embodiments, recurrent neural networks (RNNs) may additionally be employed to model temporal dependencies in sequential reconstructed channel representations.
[0064] The positioning model architecture 200 begins with a sequence of two-dimensional convolutional layers (Conv-2D) that extract spatial features from the reconstructed virtual complete channel representation. These convolutional layers are particularly effective at identifying local patterns in channel characteristics, such as power profiles, delay profiles, and angle profiles, which are relevant for positioning tasks. After each Conv-2D layer, a batch normalization layer is applied to normalize activations, reducing internal covariate shifts and improving training stability.
[0065] To introduce non-linearity and ensure effective gradient propagation, each batch normalization layer is followed by a leaky ReLU activationfunction. Unlike standard ReLU, leaky ReLU allows for a small non-zero gradient when a unit is inactive, thereby mitigating dead neuron effects and improving convergence. The combination of Conv-2D layers, batch normalization, and leaky ReLU activation is repeated NCNN times, enabling progressive extraction of increasingly complex spatial features.
[0066] Following the convolutional layers, the positioning model architecture 200 integrates a transformer encoder layer configured to capture long-range dependencies and contextual relationships within the reconstructed channel representation. Through self-attention mechanisms, the transformer layer models spatial correlations across different signal components, enabling improved generalization across varying network topologies and propagation environments.
[0067] To further refine the extracted features, an adaptive filter is applied after the transformer layer. The adaptive filter dynamically adjusts its parameters based on characteristics of the reconstructed channel representation, including inferred reliability derived from line-of-sightornon-line-of-sight confidence indications and associated quality indicators, thereby enabling robust handling of multipath fading, interference, and non-line-of-sight propagation effects.
[0068] The combination of the transformer layer and adaptive filtering is repeated NTBB times, providing multiple levels of feature abstraction and refinement. This iterative processing improves the model’s ability to disentangle complex spatial dependencies present in reconstructed channel representations.
[0069] In the final stage, the positioning model architecture 200 employs one or more fully connected (FC) layers to integrate the extracted features and generate precise position estimates. These layers act as high-level feature aggregators, synthesizing learned spatial relationships into a meaningful output representation. The fully connected layers may be repeated NFC times to capture higher-order dependencies and improve final prediction accuracy.The output of the positioning model architecture 200 comprises two coordinate values, x and y, representing a two-dimensional position estimate of the target node.
[0070] The positioning model architecture 200 is adaptive and configurable, allowing its structure and parameters to be modified based on dataset size, deployment environment, and characteristics of input measurement reports. Adjustable parameters may include, but are not limited to: (1 ) the number of Conv-2D layers, transformer layers, and fully connected layers; (2) the learning rate used during training; and (3) feature extraction strategies conditioned on propagation characteristics. This flexibility enables effective deployment across diverse wireless communication environments, including 5G-Advanced and sixth-generation (6G) networks.
[0071] The positioning model architecture 200 is typically employed as the second AI-ML model in a two-stage positioning framework. In a first stage, concise channel information received from one or more receiver nodes is processed using a first AI-ML model to reconstruct a virtual complete channel representation. In a second stage, the reconstructed virtual complete channel representation is processed by the positioning model architecture 200 to estimate the precise two-dimensional position of the target node.
[0072] This two-stage approach significantly reduces uplink reporting overhead while maintaining high-precision positioning performance comparable to systems that rely on raw channel impulse response measurements.
[0073] FIG. 3 depicts a model architecture (300) for AI-ML-based three-dimensional positioning, according to aspects of the present disclosure. The model architecture 300 processes a reconstructed virtual complete channel representation, generated by a first AI-ML model from concise channel information, as its input, and operates through multiple stages to generate position estimates in three-dimensional space (x, y, z). For clarity, thereconstructed virtual complete channel representation may be denoted as an input signal h(t) in the illustrated embodiment.
[0074] While the model architecture 300 shares structural similarities with the two-dimensional positioning model shown in FIG. 2, key modifications enable accommodation of an additional vertical dimension corresponding to a z-coordinate, which is required for three-dimensional positioning tasks. By incorporating deep learning techniques such as convolutional layers, transformer layers, adaptive filtering, and skip connections, the model architecture 300 effectively estimates precise three-dimensional locations under diverse network conditions, including non-line-of-sight (NLoS) environments.
[0075] Like the two-dimensional positioning model, the model architecture 300 begins with a sequence of two-dimensional convolutional layers (Conv-2D) configured to extract spatial features from the reconstructed virtual complete channel representation. These layers identify patterns related to channel characteristics including power profiles, delay profiles, and angular variations that are relevant for accurate positioning. Each Conv-2D layer is followed by a batch normalization layer to stabilize activations and improve convergence behavior during training.
[0076] To introduce non-linearity, a leaky ReLU activation function is applied after batch normalization. Unlike standard ReLU functions, leaky ReLU allows a small non-zero gradient for inactive units, thereby reducing dead neuron effects and improving robustness. The combination of Conv-2D layers, batch normalization, and leaky ReLU activation is repeated NCNN times, enabling progressively richer feature extraction. Skip connections are incorporated to preserve spatial information from earlier layers, ensuring that both low-level and high-level features contribute to positioning accuracy.
[0077] Beyond convolutional layers, the model architecture 300 integrates a transformer encoder layer configured to capture long-range dependencies within the reconstructed channel representation. The self-attentionmechanism employed by the transformer layer enables modeling of spatial relationships across multiple signal components, which is particularly beneficial for three-dimensional positioning scenarios where propagation effects vary across multiple spatial dimensions.
[0078] Following the transformer layer, an adaptive filter is applied to dynamically refine extracted features. The adaptive filter adjusts its parameters based on characteristics of the reconstructed virtual complete channel representation, including inferred reliability derived from line-of-sight or non-line-of-sight confidence indications and associated quality indicators, thereby compensating for signal distortions, multipath effects, and elevation-dependent propagation variations encountered in three-dimensional space.
[0079] The transformer-adaptive filter combination is repeated NTBB times, providing multiple levels of feature abstraction and refinement. Skip connections ensure retention of essential spatial information across deeper network layers while refining complex positioning characteristics.
[0080] The final stage of the model architecture 300 comprises one or more fully connected layers that integrate extracted features and generate a final position estimate. These layers combine spatial dependencies learned by the transformer layers and refine the positioning outputs to generate precise three-dimensional coordinates (x, y, z). The fully connected layers may be repeated NFC times to enable high-dimensional feature integration. Skip connections in this stage further help preserve information from earlier processing stages, reducing estimation errors in the final positioning output.
[0081] The key distinction between the three-dimensional positioning model of FIG. 3 and the two-dimensional positioning model of FIG. 2 lies in the output layer. While the two-dimensional model outputs only x and y coordinates, the model architecture 300 outputs three coordinate values x, y, and z, thereby enabling estimation of elevation or vertical positioning. This capability is particularly advantageous for applications requiring full spatialawareness, including indoor localization, aerial positioning, and multi-floor navigation.
[0082] Similar to the two-dimensional positioning model, the model architecture 300 is employed as the second AI-ML model in a two-stage positioning framework. In a first stage, concise channel information received from one or more receiver nodes is processed using a first AI-ML model to reconstruct a virtual complete channel representation. In a second stage, the reconstructed virtual complete channel representation is processed by the model architecture 300 to estimate the precise three-dimensional position (x, y, z) of the target node.
[0083] This two-stage approach reduces uplink feedback overhead while achieving positioning accuracy comparable to systems relying on raw channel impulse response measurements.
[0084] A significant advantage of the model architecture 300 is its flexibility. By modifying the output layer, the architecture may be adapted to support both two-dimensional and three-dimensional positioning scenarios. This enables deployment of a common AI-ML positioning architecture across multiple applications, reducing implementation complexity and computational overhead.
[0085] FIG. 4 illustrates a positioning framework (400) for AI-ML-assisted positioning, according to an embodiment of the present disclosure. The positioning framework 400 operates within a two-stage AI-ML positioning architecture and leverages AI-ML models to process channel-related information and generate positioning-related outputs, thereby improving localization accuracy in wireless communication networks. By analyzing timing information, angle information, and line-of-sight or non-line-of-sight (LoS / NLoS) indications derived from reconstructed channel characteristics, the framework enhances positioning precision, particularly in challenging non-line-of-sight (NLoS) environments.
[0086] At the core of the positioning framework 400 is an AI-ML-based processing pipeline that receives reconstructed virtual complete channelrepresentation, which is generated by a first AI-ML model from concise channel information reported by one or more receiver nodes. The concise channel information comprises reduced-overhead measurements such as timing information, angle information, LoS / NLoS confidence indications, and associated quality indicators. The reconstructed virtual complete channel representation approximates characteristics of a full channel measurement while compensating for missing, noisy, or degraded measurements.
[0087] The AI-ML processing within the positioning framework 400 produces multiple positioning-related outputs that are used by a second AI-ML model or positioning function to estimate the position of a target node. These outputs include timing-related measurements, angular measurements, and link-state confidence indicators that collectively improve localization accuracy.
[0088] One category of outputs generated within the positioning framework 400 is timing information, which may include: (1) Time of Arrival (ToA), representing the absolute time at which a signal reaches a receiver, (2) Relative Time of Arrival (RTOA), representing time differences between signals received from multiple transmitters, and (3) Reference Signal Time Difference (RSTD), representing relative timing differences with respect to a reference signal. These timing estimates enable accurate distance calculations and form a fundamental basis for position estimation.
[0089] The positioning framework 400 further extracts angle information from the reconstructed channel representation. The type of angular measurement depends on the receiver configuration: (1 ) For a base station or transmission reception point, the framework provides Angle of Arrival (AoA), indicating the direction from which a signal arrives, and (2) For user equipment or a reference node, the framework provides Angle of Departure (AoD), indicating the direction of signal transmission. The combined use of timing and angular information enables enhanced triangulation and geometric positioning accuracy.To address signal degradation caused by obstructed propagation paths, the positioning framework 400 generates a line-of-sight or non-line-of-sight (LoS / NLoS) indication represented as a continuous confidence value between 0 and 1. A value closer to 1 indicates a higher likelihood of line-of-sight propagation, whereas a value closer to 0 indicates a higher likelihood of non-line-of-sight propagation due to reflections, diffractions, or scattering effects. Intermediate values represent varying degrees of confidence regarding propagation conditions.
[0090] Each positioning-related output generated within the framework, including timing information, angle information, and LoS / NLoS confidence, is associated with a quality indicator ranging from 0 to 1. A lower value indicates reduced reliability of the measurement, while a higher value indicates increased confidence and accuracy. These quality indicators enable selective weighting of measurements during position estimation.
[0091] The positioning framework 400 dynamically adapts position estimation based on the LoS / NLoS confidence and associated quality indicators. When a high LoS confidence is detected, greater weight is assigned to corresponding reconstructed timing and angular measurements. Conversely, when NLoS conditions are inferred, the framework compensates for potential errors by reducing reliance on unreliable measurements, applying error compensation techniques, prioritizing multi-anchor information, or incorporating additional assistance data. This adaptive behavior mitigates multipath effects and improves robustness of the final position estimate.
[0092] For clarity, input information provided to AI-ML models within the framework is referred to herein as Part-A information, while output information generated by AI-ML models or positioning functions is referred to as Part-B information. The separation of Part-A and Part-B information facilitates structured data handling, model training consistency, and scalable deployment of AI-ML-based positioning systems.FIG. 5 illustrates a positioning framework (500) for AI-ML-based positioning systems, according to aspects of the present disclosure. The positioning framework 500 operates within the disclosed two-stage AI-ML positioning architecture and incorporates three different configurations for processing positioning-related information. Each configuration leverages one or more AI-ML models to process input information (referred to herein as Part-A information) and generate output information (referred to herein as Part-B information) for positioning purposes. The framework provides flexibility to support multiple deployment scenarios while maintaining low reporting overhead, high positioning accuracy, and secure handling of information.
[0093] In a first configuration, the AI-ML processing pipeline receives Part-A information comprising concise channel information reported by one or more receiver nodes, along with associated timestamps and quality indicators. The concise channel information is first processed by a first AI-ML model to reconstruct a virtual complete channel representation. A second AI-ML model then processes the reconstructed virtual complete channel representation to generate Part-B outputs comprising a location estimate, an associated timestamp, and a quality indicator. This configuration enables efficient end-to-end positioning while avoiding transmission of raw channel measurements, and is suitable for applications requiring low latency and reduced signaling overhead.
[0094] In a second configuration, the AI-ML processing pipeline similarly receives Part-A information derived from concise channel information and associated metadata. The first AI-ML model reconstructs a virtual complete channel representation, after which an AI-ML model generates intermediate positioning-related measurements as Part-B outputs, together with timestamps and quality indicators. These intermediate measurements may represent reconstructed signal characteristics or abstracted positioning features that are subsequently consumed by additional positioning functions or AI-ML models. This configuration is particularly advantageous in multi-stage AI-ML positioning architectures and hybrid positioning systems where intermediate representations improve final accuracy.
[0095] In a third configuration, the AI-ML processing pipeline receives Part-A information comprising intermediate measurements, timestamps, and quality indicators generated by earlier processing stages. One or more AI-ML models process this information to generate Part-B outputs comprising a final location estimate. This configuration supports distributed positioning architectures in which measurement reconstruction, feature extraction, and final position estimation are performed at different network entities, thereby enabling refined processing and scalability in large-scale deployments.
[0096] To ensure data security and privacy, the positioning framework 500 categorizes assistance and contextual information into two classes. Sensitive or proprietary information associated with transmitter or receiver identities, network topology, or configuration parameters is protected using a single unique identifier referred to herein as an Associated ID. Nonsensitive information required for positioning computation is explicitly shared, allowing AI-ML models to operate effectively without exposing proprietary network details.
[0097] Another key feature of the positioning framework 500 is its ability to adapt position estimation based on line-of-sight or non-line-of-sight (LoS / NLoS) confidence indications and associated quality indicators. When a high likelihood of line-of-sight propagation is inferred, greater weight is assigned to corresponding reconstructed timing and angular measurements. Conversely, under non-line-of-sight conditions, the framework reduces reliance on unreliable measurements, applies error compensation techniques, or prioritizes alternative positioning information, thereby improving robustness in multipath environments.
[0098] By integrating multiple processing configurations, secure information handling, and confidence-aware adaptive positioning, the positioning framework 500 provides a scalable and flexible solution for a wide range of positioning applications. Whether generating final position estimates directlyfrom reconstructed channel representations, producing intermediate positioning measurements, or supporting distributed multi-stage processing, the framework maintains consistency with the disclosed two-stage AI-ML architecture and is well-suited for deployment in fifth-generation advanced and sixth-generation wireless communication networks.
[0099] Figure 6 illustrates an exemplary block diagram of a network node (600) configured to support AI-ML-based positioning, training data collection, inference execution, and lifecycle management in accordance with the present invention. The network node may correspond to a user equipment, a base station, a transmission reception point, a reference node, a location server, or a training server, depending on the deployment scenario and functional role within the wireless communication network.
[0100] The network node (600) comprises a network interface (605), a processor (610), a memory (615), and a storage unit (620). The network interface (605) is configured to enable communication with other network entities over wired or wireless links. In accordance with the present invention, the network interface facilitates transmission and reception of positioning reference signals, channel measurements, concise channel information, reconstructed positioning information, intermediate positioning measurements, assistance data, Part-A information, Part-B information, AI-ML model identifiers, model outputs, and model performance monitoring messages.
[0101] The processor (610) is operatively coupled to the network interface and is configured to execute one or more software modules, algorithms, or machine-readable instructions stored in the memory (615) or the storage unit (620). The processor performs operations including extraction of channel measurements, generation of concise channel information, reconstruction of a virtual complete channel representation using a first AI-ML model, generation of intermediate positioning measurements where applicable, and estimation of a position of a target node using a second AI-ML model. In some embodiments, the processor executes AI-ML modelscomprising neural network architectures including convolutional layers, transformer layers, adaptive filtering, and fully connected layers, as described herein.
[0102] In some embodiments, the processor is further configured to extract spatial, temporal, and angular features from the reconstructed virtual complete channel representation, including features indicative of multipath propagation, signal dispersion, and obstruction effects. Such feature extraction enhances robustness of positioning under non-line-of-sight conditions by enabling the second AI-ML model to learn and compensate for NLoS-induced distortions, thereby improving positioning accuracy in challenging propagation environments.
[0103] The memory (615) is coupled to the processor and is configured to temporarily store operational data during execution. Such data may include Part-A information comprising concise channel information, timestamps, line-of-sight or non-line-of-sight confidence indications, and quality indicators, as well as Part-B information comprising reconstructed channel representations, intermediate measurements, or location estimates. The memory may further store assistance information used to ensure consistency between training and inference phases, including explicit or implicit parameters and associated identifiers representing proprietary configuration information.
[0104] The storage unit (620) is configured to provide persistent storage for longer-term data and program instructions. In accordance with the invention, the storage unit may store AI-ML model parameters, training datasets, logged associations between Part-A and Part-B information, model performance metrics, historical inference results, and model update or retraining information. The storage unit thereby enables efficient lifecycle management of AI-ML positioning models, including training, validation, deployment, monitoring, and update operations.
[0105] The storage unit may further store historical positioning error metrics and reference statistical distributions associated with trained AI-ML models.Such stored distributions may be used by the processor or training server to detect deviations in inference-time error behavior and to determine when retraining, fine-tuning, or model updates are required.
[0106] In operation, the network node (600) uses the network interface to receive positioning signals, channel measurements, or assistance data, and to transmit measurement reports, reconstructed channel representations, intermediate positioning information, or location estimates. The processor processes received signals to generate concise channel information and executes A I -ML models using data stored in the memory and storage. The resulting outputs may include a virtual complete channel representation, intermediate positioning measurements, or final position estimates, which may be locally consumed by the network node or transmitted to other network entities for further processing.
[0107] In some embodiments, the receiver or network interface is configured to perform filtering, normalization, noise suppression, outlier rejection, or formatting of concise channel information prior to providing the concise channel information to the first AI-ML model. Such pre-processing reduces the impact of measurement noise, improves input data consistency, and enhances reconstruction accuracy of the virtual complete channel representation.
[0108] Through coordinated interaction of the network interface, processor, memory, and storage unit, the network node enables low-overhead data collection, accurate AI-ML-based positioning, and consistent performance across training and inference stages, including under non-line-of-sight dominated propagation conditions.
[0109] A number of implementations have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the disclosure. Accordingly, other implementations are within the scope of the following claims.
Claims
We Claim:
1. A method for AI-ML-based positioning in a wireless communication networks, comprising:receiving, from at least one receiver node, concise channel information having reduced reporting overhead, the concise channel information comprising at least one of timing information, angle information, a line-of-sight or non-line-of-sight (LoS / NLoS) confidence indication, and an associated quality indicator;processing the concise channel information using a first AI-ML model to reconstruct a virtual complete channel representation, wherein the reconstruction compensates for missing, noisy, or degraded channel characteristics resulting from limited measurement reporting and non-line-of-sight propagation; andprocessing the reconstructed virtual complete channel representation using a second AI-ML model to estimate a position of a target node,wherein the LoS / NLoS confidence indication and the quality indicator are used to adapt or weight the position estimation.
2. The method as claimed in claim 1, wherein the concise channel information further comprises a timestamp associated with a channel measurement.
3. The method as claimed in claim 1, wherein the quality indicator represents a reliability level of the channel measurement and ranges between a minimum value indicating lowest reliability and a maximum value indicating highest reliability.
4. The method as claimed in claim 1, wherein the timing information comprises at least one of:Time of Arrival (ToA),Relative Time of Arrival (RTOA), andReference Signal Time Difference (RSTD).
5. The method of claim 1 , wherein the angle information comprises:Angle of Arrival (AoA) measured at a base station or transmission reception point, orAngle of Departure (AoD) measured at user equipment or a reference node.
6. The method as claimed in claim 1, wherein the LoS / NLoS indication represents a continuous confidence value indicative of a likelihood of a line-of-sight propagation path.
7. The method as claimed in claim 6, further comprising: adapting or weighting the position estimation comprises increasing reliance on reconstructed timing and angle information when the LoS / NLoS confidence indicates a higher likelihood of line-of-sight propagation..
8. The method as claimed in claim 1, wherein the first A I -ML model comprises at least one of:A deep neural network (DNN),A convolutional neural network (CNN),A transformer-based model, orA recurrent neural network (RNN).
9. The method as claimed in claim 1 , further comprising:logging associations between the concise channel information and corresponding reconstructed virtual complete channel representations during a training phase, and using the logged associations to maintain consistency between training and inference of at least one of the first AI-ML model or the second AI-ML model.
10. The method as claimed in claim 1 , further comprising:monitoring a statistical distribution of positioning errors during inference, and retraining or fine-tuning at least one of the first AI-ML model or the second AI-ML model when a deviation from a reference distribution exceeds a predefined threshold.
11. The method as claimed in claim 1, wherein the concise channel information comprises a reduced-dimensional representation relative to full channel measurements, thereby reducing uplink signaling overhead.
12. The method as claimed in claim 1, wherein the concise channel information is aggregated from a plurality of receiver nodes prior to reconstruction.
13. The method as claimed in claim 1, wherein the first AI-ML model is configured to condition reconstruction of the virtual complete channel representation on the LoS / NLoS confidence indication and the associated quality indicator, wherein the reconstruction comprises adaptively weighting timing components or angular components based on inferred measurement reliability.
14. A system for AI-ML-based positioning in wireless communication networks, comprising:a receiver configured to receive concise channel information having reduced reporting overhead from at least one receiver node;a processor configured to:process the concise channel information using a first AI-ML model to reconstruct a virtual complete channel representation, and process the reconstructed virtual complete channel representation using a second AI-ML model to estimate a position of a target node,wherein the processor is further configured to adapt the position estimation based on a LoS / NLoS confidence indication and an associated quality indicator.
15. The system as claimed in claim 14, wherein the processor is configured to extract features from the reconstructed virtual complete channel representation to enhance positioning accuracy in non-line-of-sight environments..
16. The system as claimed in claim 14, wherein the receiver is configured to aggregate concise channel information over multiple time instances prior to reconstruction.
17. The system as claimed in claim 14, wherein the second AI-ML model is trained using a dataset comprising at least one of real-world positioning data or synthetic positioning data.
18. The system as claimed in claim 14, wherein the receiver is configured to receive positioning assistance data comprising at least one of:Reference signal configurations,Beam information, andInterference indicators.
19. The system as claimed in claim 14, wherein the processor is further configured to periodically update at least one of the first or second AI-ML models based on newly collected positioning data.
20. The system as claimed in claim 14, wherein the receiver is configured to filter and pre-process the concise channel information prior to input to the first AI-ML model.
21. The system as claimed in claim 14, wherein the system is implemented in a 5G-Advanced or 6G wireless communication network environment.
22. A network node for use in a wireless communication network, the network node being configured to support artificial-intelligence or machinelearning (AI-ML) based positioning of a target node, the network node comprising:a network interface configured to exchange positioning reference signals, concise channel information, assistance data, and positioning-related information with one or more network entities;a processor operatively coupled to the network interface;a memory coupled to the processor; anda storage unit coupled to the processor,wherein the processor is configured to execute machine-readable instructions to:reconstruct a virtual complete channel representation from concise channel information using a first AI-ML model, and estimate a position of the target node using the reconstructed virtual complete channel representation and a second AI-ML model.
23. The network node as claimed in claim 22, wherein the concise channel information further comprises a timestamp and a quality indicator representing measurement reliability.