Low-overhead ai-based positioning system for 5g networks

Abstract

The invention introduces a low-overhead, AI-based positioning 5 system for 5G networks, utilizing advanced AI / ML models to optimize positioning accuracy with minimal reporting. It significantly reduces feedback requirements by leveraging truncated, low-resolution channel information while maintaining sub-meter positioning accuracy, even in challenging Non-Line-of-Sight (NLoS) environments. The system employs 10 AI / ML-based super-resolution models to reconstruct complete channel data from compact reports, enabling precise estimation of parameters such as Time of Arrival (ToA) and Angle of Arrival (AoA), or direct location inference. Efficient methods for training data collection are proposed to ensure high- quality datasets, with quality indicators and timestamps. Mechanisms are 15 included to maintain consistency between training and inference phases using context-aware identifiers. Additionally, the invention integrates performance monitoring and update signaling schemes, allowing for continuous model adaptation to dynamic conditions. This results in improved positioning accuracy, reduced computational complexity, and 20 lower power consumption making it suitable for a wide range of 5G applications. Figure 1 (publication).

Description

[0001] Low-Overhead Al-Based Positioning System for 5G Networks

[0002] Field of the Invention

[0003] The present invention relates to wireless communication technology, specifically to methods and systems for enhancing user equipment (UE) positioning accuracy through low-overhead reporting and artificial intelligence or machine learning (AI / ML) based reconstruction in 5G / 5G-NR and beyond communication systems.

[0004] Background of the Invention

[0005] In modern wireless communication systems, accurate positioning of a node such as user equipment (UE) is vital for a wide range of applications, including location-based services, emergency response, network optimization, autonomous mobility, and industrial automation. Several techniques have been standardized for positioning in 5G and beyond, 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 rely on precise time and angle measurements obtained from line-of-sight (LoS) paths between the transmitter and the receiver. However, in real-world scenarios especially urban, indoor, and complex industrial environments LoS paths are frequently blocked, resulting in non-line-of-sight (NLoS) propagation conditions. Under such conditions, conventional positioning methods often produce inaccurate location estimates due to distortions in timing and angular measurements. The accuracy degradation in NLoS settings stems from several factors, including low time and angular resolution, multipath interference, weak signal-to-noise ratios (SNRs), and the lack of direct LoS connections needed to extract reliable positioning metrics.

[0006] Although current communication systems support large bandwidths and utilize advanced antenna arrays to improve temporal and spatial resolution, fundamental limitations remain when SNR is poor, or the propagation environment is highly reflective. Some improvements can be achieved by averaging received signal metrics over time and space, but these enhancements are not sufficient to overcome the core limitations of conventional positioning in NLoS conditions.

[0007] To address these challenges, artificial intelligence and machine learning (AI / ML) techniques have been proposed to enhance positioning accuracy by learning complex propagation patterns from received signal features. AI / ML models can be trained using supervised learning to interpret power delay profiles (PDP), channel impulse responses (CIR), angle-of-arrival characteristics, and Doppler information, enabling them to infer virtual LoS paths even in NLoS-dominant environments. By compensating for NLoS-induced biases, these models can improve the robustness and accuracy of positioning estimates.

[0008] However, the effectiveness of AI / ML-based positioning techniques depends on the quality and richness of the input channel information. In practical systems, the raw CIR, PDP, or delay profile (DP) data needed for high-performance AI / ML inference must be reported from the UE or base station to a centralized server. Reporting such high-resolution channel information imposes substantial uplink feedback overhead, which degrades system performance, particularly in bandwidth-constrained or high-mobility scenarios. Conversely, aggressive reduction of the reported channel data to alleviate feedback pressure can result in significant accuracy loss during positioning.

[0009] Therefore, there is a need for a positioning framework that effectively balances uplink resource efficiency with high positioning accuracy. This can be achieved by introducing a low-overhead positioning method in which reduced channel information generated by intelligently selecting, truncating, and quantizing CIR or PDP samples is transmitted from the receiver node to a reporting entity.

[0010] Objective of the Invention

[0011] The principal objective of this invention is to develop a low-overhead channel information reporting method that enables the receiver to selectively report a compact subset of channel impulse responses (CIR) or power delay profiles (PDP) or delay profile (DP) samples while preserving essential positioning data.

[0012] Another objective of this invention is to design an AI / ML-based super-resolution model capable of reconstructing complete channel information from truncated reports, supporting multiple truncation levels with a single adaptive model.

[0013] Another objective of this invention is to achieve high-accuracy positioning under NLoS conditions by leveraging AI-ML models that infer virtual line-of-sight measurements from sparse and noisy channel observations.

[0014] Another objective of this invention is to enable joint estimation of positioning parameters such as Time of Arrival (ToA), Angle of Arrival (AoA), and Doppler across multiple transmitter nodes using a unified Al-assisted model, thereby improving positioning robustness and efficiency.

[0015] A further objective of this invention is to ensure consistency between the AI-ML training and inference phases by defining and using a context-aware association ID derived from RS configuration and base station metadata.

[0016] Summary of the Invention

[0017] This invention introduces a comprehensive, low-overhead AI / ML-based positioning framework tailored for 5G, 5G-NR, and future 6G networks. It addresses the critical challenge of excessive feedback overhead in transmitting full channel state information (CSI) such as Channel Impulse Response (CIR), Power Delay Profile (PDP), and Doppler Profile (DP) from the user equipment (UE) to the network for positioning purposes, especially in Non-Line-of-Sight (NLoS) environments.

[0018] To mitigate this, the invention proposes a selective reporting mechanism where the UE transmits only a compact subset of channel information. The selection is based on intelligent criteria such as first path delay, estimated Tx-Rx distance, and power thresholds. At the network side, an AI / ML-based super-resolution model reconstructs the complete channel information from these truncated inputs. This reconstructed data is then used to estimate positioning parameters like Time of Arrival (ToA), Angle of Arrival (AoA), and Doppler, or directly infer the UE’s geographic location.

[0019] A key aspect of the invention is its novel framework for training data collection and model performance monitoring, designed to maintain high model accuracy and robustness across dynamic wireless conditions. It ensures alignment between training and inference contexts through the use of context-aware association identifiers (IDs), enabling consistent and reliable model behavior. Additionally, real-time performance monitoring using both label-based and label-free methods supports proactive model updates and lifecycle management.

[0020] It also emphasizes high-quality dataset generation with minimal overhead, using timestamping and quality indicators to ensure the integrity of training inputs. It integrates seamlessly with existing 5G network architectures, offering backward compatibility and future-proof scalability for Rel-19 and beyond.

[0021] By combining efficient data reporting, AI / ML-based signal reconstruction, intelligent model management, and adaptive inference, this invention delivers a high-precision, resource-efficient, and resilient positioning solution. It enhances positioning accuracy, reduces power and computational requirements, and optimizes network resource utilization — making it ideal for a wide range of next-generation wireless deployment scenarios.

[0022] Brief description of the drawings

[0023] 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.

[0024] Figure 1 illustrates various positioning methodologies within a communication network (100) positioning methodologies includes:

[0025] Figure 1(A) illustrates the location server-based Uplink (UL) positioning, according to one embodiment of the present invention. Figure 1(B) illustrates the location server-based Downlink (DL) positioning, according to one embodiment of the present invention.

[0026] Figure 1(C) illustrates the User equipment-based positioning, according to one embodiment of the present invention.

[0027] Figure 2 illustrates the preprocessing of channel information prior to super-resolution (200), in accordance with one embodiment of the present invention.

[0028] Figure 3 illustrates the AI-ML model for channel information super-resolution (300), according to one embodiment of the present invention.

[0029] Figure 4 illustrates a single assisted AI-ML model for a Time of Arrival (ToA) estimator generalized for a validity area (400), according to one embodiment of the present invention.

[0030] Figure 5 illustrates an AI-ML assisted model for the joint estimation of all ToAs (500), in accordance with one embodiment of the present invention.

[0031] Figure 6 illustrates a CNN-super resolution-based direct AI-ML position estimator for truncated Channel Impulse Response (CIR) (600), according to one embodiment of the present invention.

[0032] Figure 7 is a block diagram illustrating an example of a schematic hardware configuration of the network node (700) according to one embodiment of the present invention.

[0033] 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.

[0034] Throughout the drawings, it should be noted that like reference numbers are used to depict the same or similar elements, features, and structures.

[0035] Detailed Description of the invention

[0036] 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 their equivalents. It includes various specific details to assist in that understanding but these are to be regarded as merely exemplary.

[0037] Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the invention. In addition, descriptions of well-known functions and constructions are omitted for clarity and conciseness.

[0038] The terms and words used in the following description and claims are not limited to the bibliographical meanings but are merely used by the inventor to enable a clear and consistent understanding of the invention. Accordingly, it should be apparent to those skilled in the art that the following description of exemplary embodiments of the present invention are provided for illustration purpose only and not for the purpose of limiting the invention as defined by the appended claims and their equivalents.

[0039] It is to be understood that the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to “a component surface” includes reference to one or more of such surfaces.

[0040] By the term “substantially” it is meant that the recited characteristic, parameter, or value need not be achieved exactly, but that deviations or variations, including for example, tolerances, measurement error, measurement accuracy limitations and other factors known to those of skill in the art, may occur in amounts that do not preclude the effect the characteristic is intended to provide.

[0041] Figures discussed below, and the various embodiments used to describe the principles of the present disclosure in this patent document are by way of illustration only and should not be construed in any way that would limit the scope of the disclosure. Those skilled in the art will understand that the principles of the present disclosure may be implemented in any suitably arranged system. The terms used to describe various embodiments are exemplary. It should be understood that these are provided to merely aid the understanding of the description, and that their use and definitions, in no way limit the scope of the invention. Terms first, second, and the like are used to differentiate between objects having the same terminology and are in no way intended to represent a chronological order, unless where explicitly stated otherwise. A set is defined as a non-empty set including at least one element.

[0042] Figure 1 illustrates various positioning methodologies within a communication network (100) configured to enhance positioning accuracy of a user equipment (UE). The communication network (100) comprises a Training Server (105), a User Equipment (UE) (110), a Location Server (115) and a Next Generation Radio Access Network (NG-RAN) (120).

[0043] Each of these network elements cooperates to support positioning functionalities that incorporate artificial intelligence and machine learning (AI-ML) based models operating on low-overhead channel information, wherein the channel information comprises at least one of a channel impulse response (CIR), a power delay profile (PDP), a delay profile (DP), or a channel frequency response (CFR). The network architecture is represented as a dashed enclosure (100), visually depicting the interactions between components. Subfigures FIG. 1(A), FIG. 1(B), and FIG. 1(C) respectively depict uplink (UL)-based positioning, downlink (DL)-based positioning, and user equipment (UE)-based positioning configurations. Each methodology adheres to a common framework comprising truncated channel reporting, AI-ML inference, association identifier validation, and performance monitoring via model performance monitoring metrics (MPMM), as described in Table 1 (given below).

[0044] Figure 1(A) illustrates, according to one embodiment of the present invention, a location server-based uplink (UL) positioning method that enables accurate and efficient positioning in wireless communication networks such as 5G-Advanced and 6G. This method is designed to address the challenges of positioning in complex propagation environments, particularly under non-line-of-sight (NLoS) conditions, while minimizing the uplink feedback overhead. The figure captures the interactions between four primary entities: the Training Server (105), User Equipment (UE) (110), Location Server (115), and the NG-RAN (Next-Generation Radio Access Network) (120), with each entity playing a key role in the positioning lifecycle.

[0045] The process begins with the Location Server (115) receiving a Location Services (LCS) request, which prompts the UE (110) to initiate a capability exchange. The UE requests and then provides its capabilities related to reference signal transmission and channel measurement. Based on this information, the Location Server selects an appropriate positioning method such as uplink time difference of arrival (UL-TDoA) or angle of arrival (AoA) and forwards a positioning information request to the NG-RAN (120). The NG-RAN then determines the necessary uplink reference signal (UL RS) resources based on the selected method, ensuring orthogonality in time, frequency, or spatial domains to facilitate accurate reception and measurement.

[0046] Once the configuration is completed, the UE activates and transmits the UL reference signal. The NG-RAN, acting as the receiver node, performs channel estimation using the received reference signal. This results in channel information such as the channel impulse response (CIR), power delay profile (PDP), or delay profile (DP). From this data, the NG-RAN derives intermediate metrics like time of arrival (ToA), angle of arrival (AoA), Doppler shift, and carrier phase. To reduce feedback overhead, the NG-RAN selects a subset of the channel samples using criteria such as the estimated first path delay, transmit-receive distance, power thresholds, and predefined window sizes (e.g., 64 or 128 samples). These samples are then processed, truncated, and quantized to form a reduced channel information report, which is transmitted to the Location Server.

[0047] The Location Server uses two AI / ML models in sequence to process this reduced report. The first model reconstructs full-resolution channel information from the compressed input. The second model estimates the position of the UE using techniques like ToA and AoA analysis and may also determine line-of-sight (LoS) versus non-line-of-sight (NLoS) conditions. These models are trained offline using data from the Training Server (105), which collects positioning datasets that include CIR / PDP / DP sequences, timestamped location labels, and measurement quality indicators such as mean squared error (MSE).

[0048] To ensure consistent and accurate inference, an association ID is generated during training. This ID encodes information such as RS configuration, TRP / base station identity, and other relevant parameters. During inference, this ID is validated to ensure the training and runtime configurations match. Additionally, the Location Server continuously monitors the performance of the deployed AI models by computing a model performance monitoring metric (MPMM), such as the position or range error. If the MPMM exceeds a preconfigured quality-of-service threshold, the Location Server can request the Training Server to retrain or update the AI model, thus enabling robust lifecycle management of the model.

[0049] Figure 1(B) illustrates a location server-based Downlink (DL) positioning methodology, aligned with one embodiment of the present invention. In this setup, the NG-RAN (120) acts as the transmitting node that sends downlink reference signals (RS) to the User Equipment (UE) (110) over orthogonal time, frequency, or spatial resources. The UE, functioning as the receiver node, performs channel estimation based on the received DL reference signals. This estimation yields critical channel information, such as channel impulse response (CIR), power delay profile (PDP), or delay profile (DP).

[0050] Using this estimated channel data, the UE derives intermediate measurements such as Time of Arrival (ToA), Angle of Arrival (AoA), Doppler shift, and carrier phase. To minimize feedback overhead while maintaining positioning accuracy, the UE selects a subset of high-value samples from the channel estimate. This selection is performed using various criteria such as power thresholding, first-path estimation, or fixed window' size sampling and results in a reduced channel information report. The UE sends this report, along with an association ID, to the Location Server (115). The association ID is computed using metadata related to reference signal configuration (e.g., PRB allocations, subcarrier spacing), transmitter parameters (e.g., TRP ID, spatial beam settings), and receiver-side preprocessing criteria. This ID ensures consistency between training and inference phases by linking each inference to the context in which its training data was generated.

[0051] The Location Server (115) receives the reduced channel report and uses a two-stage AI / ML model cascade to estimate the UE’s position. The first model reconstructs high-resolution channel information from the reduced input, while the second model uses the reconstructed data to estimate the UE's position or intermediate values like ToA or AoA. These AI / ML models are pre-trained and maintained by the Training Server (105), which also manages the Training Data Collection (TDC) process. The TDC dataset includes channel estimates, inferred or known ground-truth location labels, timestamps (in UTC format), and quality indicators such as Mean Squared Error (MSE) or normalized error metrics (between 0 and 1).

[0052] In addition to inference, Figure 1(B) also illustrates how model performance is periodically monitored using a metric called Model Performance Monitoring Metric (MPMM). This metric compares the Al model's inferred location (p) against either a known ground truth (p) or against results from conventional positioning schemes like DL-TDoA or m-RTT. If the MPMM exceeds a predefined threshold, the Location Server may trigger an Al model update request to the Training Server. The updated model is then downloaded and deployed, ensuring that inference performance remains robust overtime, particularly as channel conditions or deployment contexts evolve.

[0053] Figure 1(C) illustrates, in accordance with one embodiment of the present invention, a user equipment (UE)-based positioning methodology. Unlike network- or server-driven positioning techniques, this approach empowers the UE (110) to handle channel processing, inference, and position estimation locally without requiring direct involvement from the Location Server (115) during the positioning procedure itself. In this embodiment, the NG-RAN (120) transmits downlink reference signals (RS) to the UE, which acts as the receiver node. Upon receiving the RS, the UE performs channel estimation, typically deriving Channel Impulse Response (CIR) or other channel metrics such as Power Delay Profile (PDP) or Delay Profile (DP).

[0054] The UE uses a truncated and compact version of the CIR for processing to reduce computational load and storage requirements. This reduced information is preprocessed to generate a vector h~1 \tilde{h}__1 h~1, where missing sample indices are zero-padded while maintaining actual reported values. The window used for this operation is defined by a starting index no, a total window size Nt. and a sample count N't. The resulting vector is then fed into a cascaded AI / ML model framework, beginning with SSRNet (Super-Resolution Sequence Network), which reconstructs a full-resolution version of the channel, denoted as hAt. This reconstructed channel representation is subsequently processed by a second AI / ML model that maps the high-resolution CIR to a position estimate pATo ensure model integrity during inference, the UE references a preloaded association ID, which encodes information about RS configuration, TRP identity, subcarrier spacing, and other metadata defined in Table 1 (given below) of the invention. This association ID acts as a consistency check, confirming that the deployed AI / ML model is valid for the current signal context. If there’s a mismatch, the UE may trigger a model update or refuse to proceed with inference.

[0055] The Training Server (105) plays a critical role in this setup by pretraining the AI / ML models and provisioning them to the UE. It also manages training data collection (TDC), ensuring that each dataset contains CIR / PDP / DP sequences, associated location labels, quality indicators (e.g., mean squared error), and timestamped entries in UTC format (with optional time slot indices).

[0056] For performance monitoring, the UE calculates a Model Performance Monitoring Metric (MPMM) locally. This metric may be computed using one or more of the following: a known ground truth position (if available), feedback from the Location Server (e.g., network-validated location), or residual errors between the Al-inferred position and estimates derived from legacy positioning methods like DL-TDoA or multi-round-trip time (m-RTT). Optionally, the UE may report the inferred position and / or MPMM to the Location Server for purposes such as validation, analytics, or triggering model updates. The system also includes a periodic monitoring mechanism to ensure the long-term reliability and robustness of the positioning models deployed on the UE.

[0057]

[0058]

[0059] Table 1: information that should be collected as a part of TDC and exchanged between the location server transmitter node and receiver node during inference.

[0060] Figure 2 illustrates the preprocessing of channel information prior to super-resolution (200). The block diagram represents the process of preparing truncated channel information, specifically the channel impulse response (Cl R), for input into an AI-ML model designed to reconstruct high- resolution channel information. The diagram outlines a sequence of operations to transform raw CIR measurements into a processed format by inserting zeros at non-selected indices, thereby minimizing reporting overhead while retaining critical channel characteristics for subsequent super-resolution.

[0061] In one embodiment, the sequential flow of operations for preprocessing the CIR at a receiver node, including an input block for raw CIR measurements, a processing block for sample selection and windowing, and an output block for the pre-processed CIR with zeros inserted at non-selected indices. Mathematical notations, such as no, Nt and Nt, are used to define the starting sample index, window size, and number of reported samples, respectively and are integral to the preprocessing methodology.

[0062] The preprocessing operation (200) begins with the input of raw CIR measurements, which represent the channel impulse response derived from reference signals received at the receiver node from one or more transmission reception points (TRPs) over orthogonal resources in at least one of the time, frequency, or spatial domains. The raw CIR is characterized by a sequence of samples obtained at a sampling rate defined as fs= k × Δf × NFFT> where k is the oversampling factor, Af is the subcarrier spacing, and NFFT is the Fast Fourier Transform size. The CIR samples span multiple delay bins and are expressed as a time-domain vector h with amplitude and phase components.

[0063] The first operation in the diagram involves determining the starting sample index, denoted as no, which identifies the initial sample of the CIR to be included in the preprocessing window. The figure explicitly shows that no is computed based on one of three criteria: the first arrival path, the estimated distance between the transmitter and receiver, a power threshold. For the first arrival path, no is calculated as:

[0064]

[0065] where r0represents the delay of the first arrival path, [xj denotes the floor function (rounding to the nearest smaller integer), and [x] denotes the ceiling function (rounding to the nearest larger integer). This computation ensures that no is a non-negative integer constrained by the FFT size minus the window size Nt, aligning the starting index with the earliest significant channel information. Alternatively, when based on the separation distance between the transmitter and receiver, no is derived as:

[0066]

[0067] In this expression, dtr is the estimated distance between the transmitter and receiver and c is the speed of light. The term j

[0068]

[0069] converts the distance into a time delay scaled by the sampling rate, with rounding operations ensuring the index fits within the valid range of the CIR sequence.

[0070] For the power threshold criterion, no is selected as the smallest sample index where the CIR power exceeds a predefined threshold, which is either determined by the receiver or configured by the location server. This threshold-based approach ensures that the starting index captures significant channel energy, as visually exemplified in the diagram with a reference to a processed power delay profile (PDP) where no = 26 for a threshold of 0.1 and window size Nt = 128.

[0071] The next operation selects the window size Nt, which defines the length of the sample window. The figure specifies that Nt is chosen from a set of integer values, such as {32, 64, 96, 128, 192, 256}, either by the receiver based on measured channel information or configured by the location server or transmitter node, ensuring that at least 90% of the channel power is captured within the selected samples. The number of samples to report, denoted as Nt', where Nt' < Nt. The figure shows that Nt' is determined by the expression:

[0072]

[0073] where ex is a non-negative integer less than [logs Nt]. This formula indicates that the number of reported samples is a fraction of the window size, with oc controlling the truncation level (e.g., oc == 1, 2, 3 corresponds to Nt / 2, Nt / 4, Nt / 8, respectively). The selection of Nt samples is based on one of two criteria: the strongest Nt' samples by strength, determined by either the absolute value or power of the samples within the window {no, no + 1,, no + Nt - 1} or the strongest Nt' samples exceeding a configured power threshold. If the number of samples exceeding the threshold, denoted as Kmax, is less than Nt', only Kmax samples are reported, and the remaining Nt' - Kmax samples are not included.

[0074] While the preprocessing method illustrated primarily targets CIR measurements, it is equally applicable to Power Delay Profile (PDP) and Delay Profile (DP) data. For PDP, only the time indices and amplitudes of significant multipath components are retained, whereas DP includes only the delay indices. The same preprocessing framework, including windowing and zero-padding, applies with suitable adjustments in the output format, thereby enabling model-agnostic training and inference across diverse measurement types. The final operation preprocessing the selected Nt' CiR samples by inserting them at their corresponding time indices within a sequence of length equal to the original CIR, with zeros filled at all non-selected indices. The output block depicts the pre-processed CIR sequence, denoted as hi, which is a zero-padded vector of fixed length that retains the temporal structure of the channel. The sequence!^ includes the time indices, amplitudes, and phases of the selected samples, where Fn represents the subset of CIR samples selected for reporting and hi denotes the fullresolution CIR targeted by the downstream super-resolution model.

[0075] To support the creation of high-quality Ai-ML training datasets, the receiver node optionally logs the pre-processed channel information h,, along with relevant configuration metadata, quality indicators, and timestamps. This logging enables the offline or online Training Server to construct ground-truth aligned data collections that improve model generalization. In some embodiments, the logged dataset may include information regarding the spatial context or receiver motion state to enhance the diversity of the training corpus.

[0076] Figure 3 illustrates an AI-ML model for channel information super-resolution (300). The neural network-based model configured to reconstruct a full-resolution version of the channel information from the reduced channel information report using a trained artificial intelligence or machine learning (AI / ML) model. The model transforms a sparse, zero-padded input sequence into a reconstructed full-resolution CIR sequence that approximates the original untruncated CIR.

[0077] In one embodiment, the diagram represents a unidirectional data flow through a single processing block labelled as super-resolution model. The input to the model, shown on the left, is denoted as hj, representing the pre-processed channel impulse response sequence, obtained by inserting the reported CIR samples into their corresponding sample indices and assigning zero values to all unreported sample positions. This fixed-length, sparse sequence hi is provided as the input to the AI-ML model.

[0078] The output of the super-resolution model is labelled as hj, denoting the reconstructed CIR estimate approximating the true channel impulse response hi, using the sparsely populated sequence bias the input. The figure emphasizes the model’s role in mapping the sparse input hito a dense outputhj, enabling high-fidelity reconstruction of channel characteristics from low-overhead feedback.

[0079] The super-resolution operation commences with the input of the pre-processed CIR sequence fu. which is a fixed-length, sparse vector preserving the temporal alignment and structure of the reported samples, including time indices, amplitudes, and phases. The Super-Resolution Model block represents a trained neural network, such as a convolutional neural network (CNN) or recurrent neural network (RNN), configured to operate on structured time-domain CIR sequences. The model infers the full-resolution CIR by learning the statistical correlations between the sparsely reported samples in hi and the corresponding untruncated CIR hi, minimizing the difference between the predicted output hi and hi using a suitable loss metric during training. The output sequence Fn. maintains the same dimensionality as hj but contains dense values across all sample positions, as inferred by the AI-ML model, without modifying the original reported samples.

[0080] The model is further configured to operate as a single adaptive model capable of delivering high-resolution outputs for various input truncation levels, such as factors of 8, 4, 2, and 1, corresponding to different values of Nt'. This adaptive configuration reduces memory requirements at the model entity by a factor of four compared to using dedicated models for each truncation level, as a single model can handle multiple truncation scenarios efficiently.

[0081] Figure 4 illustrates a single assisted AI-ML model for a Time of Arrival (ToA) estimator generalized for a validity area (400). A machine learningbased architecture wherein a single artificial intelligence and machine learning (AI-ML) model is employed to infer the Time of Arrival (ToA) from high-resolution channel impulse response (CIR) data, with generalization over a defined validity area to reduce model complexity and memory requirements by avoiding the need for separate models per transmitter.

[0082] In one embodiment, the diagram represents a unidirectional data flow through a single processing block labelled as single assisted AI-ML model. The input to the model, shown on the left, is denoted as hirepresenting the high-resolution CIR sequence reconstructed by a super-resolution model from preprocessed truncated CIR data. The output, shown on the right, is labelled as

[0083]

[0084] denoting the estimated Time of Arrival of the primary propagation path, typically the first significant arrival or strongest component in the CIR. The figure emphasizes the model’s role in providing a lightweight and scalable solution for ToA estimation across a validity area.

[0085] The ToA estimation operation commences with the input of the high-resolution CIR sequence hi, which contains time-domain CIR values, including time indices, amplitudes, and phases, as reconstructed by the super-resolution model. The Single Assisted Ai-ML Model block represents a trained neural network, such as a convolutional neural network (CNN) or recurrent neural network (RNN), designed to map the input sequence hito a scalar ToA value T,. The model is trained on labeled channel data collected from within a spatial region, or validity area, where propagation characteristics, such as multipath delay profiles, angular spreads, and Doppler shifts, exhibit statistically bounded variation.

[0086] This generalization enables the model to predict ToA for various deployment scenarios within the validity area without requiring retraining or per-link adaptation, supporting legacy positioning schemes such as Downlink Time Difference of Arrival (DL-TDoA) and muiti-Round Trip Time (m-RTT). The model infers q by learning to detect meaningful ToA signatures from the high-resolution CIR, optimizing the error between predicted and ground-truth ToA values during training. The estimated Time of Arrival (ToA) values produced by the single assisted AI-ML model can be directly consumed by conventional positioning systems such as DL-TDoA, multi-Round Trip Time (m-RTT), or other delay-based estimators. This compatibility ensures backward integration with existing infrastructure while enhancing robustness in multipath-prone or NLoS environments through learned inference.

[0087] Figure 5 illustrates an Ai-ML assisted model for the joint estimation of all ToAs (500). A Neural network-based architecture configured to jointly estimate the Times of Arrival (ToAs) from multiple received signals, each associated with a different transmission reception point (TRP), using a unified AI-ML inference model to enhance estimation accuracy and efficiency in multi-TRP environments.

[0088] In one embodiment, the diagram represents a parallel, multi-input, multi-output processing pipeline. The input, shown on the left, consists of a set of high-resolution CIR sequences, denoted as { h., fi2,..., hN}, where each hi represents the preprocessed CIR sequence for the ithTRP, obtained from a super-resolution model as described in the invention. The central block, labelled as AI-ML Assisted Model for Joint Estimation of all the ToAs, processes these inputs simultaneously. The output shown on the right, is a set of ToA estimates {x1,x2,...,xN}, where each xsdenotes the estimated Time of Arrival for the Ithtransmitter node. Each scalar output x;corresponds to the delay at which the earliest significant propagation path from the ithTRP arrives at the receiver. The figure emphasizes the model’s ability to capture both individual CIR characteristics and inter-link statistical dependencies across transmitters.

[0089] The mapping from inputs

[0090]

[0091] hN}each reflecting the propagation characteristics from a respective TRP to the receiver node, including time indices, amplitudes, and phases, is processed by the AI-ML model block, which represents a trained neural network, potentially comprising convolutional, recurrent, or attention-based layers, configured to learn per-link features and inter-link correlations among the CIRs. The model outputs a vector of ToA estimates

[0092]

[0093] ---TN} where each -corresponds to the delay of the earliest significant propagation path from the ithTRR

[0094] This joint estimation approach enhances accuracy and efficiency in multi-TRP environments, supporting positioning schemes such as Downlink Time Difference of Arrival (DL-TDoA) and multi-Round Trip Time (m-RTT), by leveraging a single inference operation across all TRPs, reducing memory overhead and computation time compared to per-link models. The model is trained to minimize a loss function that penalizes deviations between predicted ToA values and ground-truth labels across all transmitters. The jointly estimated ToAs produced by the unified model can be directly fed into DL-TDoA, m-RTT, or hybrid positioning schemes, thereby eliminating the need for redundant per-link inference chains. This approach allows for scalable deployment across multi-TRP environments while maintaining full compatibility with standardized location computation mechanisms.

[0095] Figure 6 illustrates a CNN-super resolution-based direct AI-ML position estimator for truncated Channel Impulse Response (CIR) (600). The end-to-end machine learning architecture configured to estimate the position of a user equipment (UE) directly from truncated CIR data using a two-stage deep learning pipeline, eliminating the need for intermediate parameter estimation.

[0096] In one embodiment, the sequential data flow through two interconnected blocks. The input, shown on the left, is denoted as hi, representing the preprocessed CIR sequence with reported samples at their corresponding indices and zeros at unreported positions, as prepared by the preprocessing methodology of the invention. The first block, labeled as the “CNN-based Super-Resolution Model,” transforms fn into a high-resolution CIR hj. The second block, labeled as the “Direct Position Estimator,” processes hj to output the estimated position p. The figure emphasizes the direct inference of position without intermediate measurements, suitable for challenging channel conditions.

[0097] The position estimation operation commences with the input of the pre-processed CIR sequence hj, which is processed by the CNN-based super-resolution model to reconstruct the high-resolution CIR hrThe model applies learned convolutional filters to infer missing sample values, producing

[0098]

[0099] as an estimate of the full-resolution CIR, including time indices, amplitudes, and phases. For multi-TRP scenarios, the model may process multiple CIR sequences^, h2, hN], producing corresponding high-resolution CIRs {h h2, hN}.

[0100] The high-resolution CIRs are then input to the direct position estimator, which infers the position p, representing a multi-dimensional location estimate (e.g., 2D or 3D coordinates), without computing intermediate measurements such as Time of Arrival (ToA) or Angle of Arrival (AoA). The model is trained on labeled datasets comprising CIR sequences and ground-truth position coordinates, optimizing the direct mapping from CIR to position. This direct approach, while lacking explainability due to the absence of intermediate parameters, offers superior performance in non-line-of-sight (NLoS)-dominated environments and supports positioning applications in 5G and beyond networks, providing sub-meter accuracy with minimal feedback overhead.

[0101] While the direct positioning model offers exceptional performance, particularly in NLoS conditions, it does so at the expense of interpretability. Unlike AI-ML-assisted pipelines that expose intermediate metrics such as ToA or AoA, the direct model functions as an opaque function from input to output. In applications where explainabiiity or traceability is essential such as industrial safety or regulatory compliance the system may revert to assisted methods, trading off a small reduction in accuracy for improved interpretability. When the AI-ML model is deployed locally at the UE, model performance monitoring is carried out on-device using internal heuristics, known ground-truth location, or assistance data. The UE computes the MPMM by comparing its inference with known location data or optimization error from fallback techniques. If the MPMM crosses a configured limit, the UE may initiate a model replacement process or notify the Location Server, allowing remote model lifecycle management and reliability assurance.

[0102] Figure 7 is a block diagram illustrating an example of a schematic hardware configuration of the network node (700). A schematic representation of a network node (700) designed to enhance positioning accuracy of a user equipment (UE) in a communication network through Al-ML-assisted positioning functionalities. The network node (700) comprises interconnected hardware components, including a Network Interface (710), a Processor (720), a Memory (730), and a Storage (740), enclosed within a dashed rectangular boundary labelled with reference numeral 700. These components collectively enable data communication, processing, storage, and execution of AI-ML-based positioning tasks, ensuring efficient handling of channel information and position estimation.

[0103] The Network Interface (710) receives reference signals from one or more Transmission Reception Points (TRPs) over orthogonal resources in at least one of the time, frequency, or spatial domains. It facilitates bidirectional communication with external networks or devices, transmitting and receiving data such as channel impulse responses (CIRs), positioning reports, and AI-ML inference results, A bidirectional arrow connects the Network Interface (710) to the Processor (720), indicating seamless data flow for processing incoming signals and sending processed outputs.

[0104] The Processor (720) serves as the core computational unit of the network node (700). It estimates channel information, comprising at least one of a channel impulse response (CIR), power delay profile (PDP), delay profile (DP), or channel frequency response (CFR), based on the received reference signals; selects a subset of samples from the estimated channel information using criteria such as a first path delay estimate, estimated transmit-receive distance, total power threshold, or a predefined window size of 32, 64, 128, or 256 samples; generates a reduced channel information report by truncating samples within the window, selecting based on power or amplitude thresholds, and quantizing amplitude and phase; reconstructs full-resolution channel information using two trained AI / ML models, where a first model reconstructs full-resolution data from reduced input and a second model estimates the UE position using time of arrival (ToA), angle of arrival (AoA), or doppler-based techniques, also determining line-of-sight (LoS) or non-line-of-sight (NLoS) probabilities; collects positioning datasets from transmitter-receiver node pairs, including doppler shift, positioning measurements, and timestamps; generates and validates an association identifier (ID) from transmission and reception parameters to ensure training-inference consistency; and monitors AI-ML model performance using metrics like position error, range error, or angular error, updating models when thresholds are exceeded. Bidirectional arrows connect the Processor (720) to the Network Interface (710), Memory (730), and Storage (740), enabling coordinated data exchange and task execution.

[0105] The Memory (730) provides volatile storage for temporary data during computation. It stores channel information, selected sample indices, and model inference results, supporting rapid access to runtime model parameters, CIR vectors, and inference outputs. A bidirectional arrow to the Processor (720) ensures efficient read and write operations for processing tasks.

[0106] The Storage (740), positioned adjacent to the Memory (730), offers non-volatile storage for persistent data. It holds AI / ML model parameters, training datasets, and historical channel and positioning data, including pretrained models, CIR logs, and configuration metadata. A bidirectional arrow to the Processor (720) allows retrieval of stored models and logging of inference outcomes for future retraining.

[0107] The logical links between components highlight their coordinated operation under the Processor's (720) control, enabling the network node (700) to process CIR-based data for accurate positioning. This deploymentagnostic architecture supports instantiation as a UE, gNB, edge cloud node, or Location Server, ensuring scalability across diverse network environments. The Processor (720) evaluates model accuracy via a Model Performance Monitoring Metric (MPMM), comparing inferences against ground-truth or conventional estimates, and retrieves updated models from Storage (740) if performance degrades, maintaining service reliability.

[0108] 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 improving positioning accuracy of a user equipment (UE) in a communication network, the method comprising:receiving reference signals at one or more receiver nodes from one or more Transmission Reception Points (TRPs) over orthogonal resources in at least one of the time, frequency, or spatial domains;estimating channel information at the receiver node based on the received reference signals;selecting a subset of samples from the estimated channel information based on one or more of:a first path delay estimate;an estimated transmit-receive distance;a total power threshold; anda predefined window size;transmitting the selected subset of samples as a reduced channel information report to a reporting node;reconstructing, at the reporting node, a full-resolution version of the channel information from the reduced channel information report using two trained artificial intelligence or machine learning (AI / ML) models; and estimating, at the reporting node, the location of the user equipment based on the reconstructed full-resolution channel information using at least one of time of arrival (ToA), angle of arrival (AoA), or doppler-based positioning techniques.

2. The method as claimed in claim 1, further comprising:training a first AI / ML model to reconstruct full-resolution channel information from reduced input; andtraining a second AI / ML model to estimate the position of the receiver node from the reconstructed channel information.

3. The method as claimed in claim 1, wherein the channel information comprises at least one of a channel impulse response (CIR), a power delay profile (PDP), a delay profile (DP), or a channel frequency response (CFR).

4. The method as claimed in claim 1, wherein the reduced channel information is computed by:truncating CIR samples within a specified window size; selecting samples based on power or amplitude thresholds or by identifying the strongest samples; andquantizing the amplitude and phase of the selected samples.

5. The method as claimed in claim 1, wherein collecting positioning datasets from one or more transmitter-receiver node pairs, the datasets including at least one of doppler shift, positioning measurements, and timestamps of measurement acquisition.

6. The method as claimed in claim 1, further comprising:generating an associate identifier (ID) from transmission and reception parameters during training; andvalidating the associate identifier (ID) during inference to ensure consistency between training and inference phases.

7. The method as claimed in claim 1, wherein monitoring the performance of the AI-ML models periodically using performance metrics, including position error, range error, or angular error, and updating the AI-ML models when performance metrics exceed a predefined threshold.

8. The method as claimed in claim 1, wherein selecting the subset of samples comprises:computing a starting index using a first path delay or an estimated transmit-receive distance;determining the starting index using a sample rate defined by a product of an FFT size and a subcarrier spacing;selecting a sample window of predefined size starting from the starting index, wherein the predefined size is selected from a set comprising 32, 64, 128, or 256 samples; andretaining only samples whose power exceeds a predefined threshold.

9. The method as claimed in claim 1, wherein the estimation of the user equipment location further includes determining at least one of line-of-sight (LoS) or non-line-of-sight (NLoS) probabilities based on the reconstructed channel information.

10. A network node for enhancing User Equipment (UE) positioning accuracy in a communication network, the network node comprising:a network interface configured to receive reference signals from one or more Transmission Reception Points (TRPs) over orthogonal resources in at least one of the time, frequency, or spatial domains;a processor configured to:estimate channel information based on the received reference signals;select a subset of samples from the estimated channel information based on at least one of:a first path delay estimate,an estimated transmit-receive distance,a total power threshold, anda predefined window size;generate a reduced channel information report comprising the selected subset of samples;reconstruct a full-resolution version of the channel information from the reduced channel information report using two trained artificial intelligence or machine learning (AI / ML) models;estimate a position of the receiver node or UE using the reconstructed full-resolution channel information based on at least one of time of arrival (ToA), angle of arrival (AoA), or doppler-based positioning techniques;a memory operably connected to the processor and configured to store channel information, selected sample indices, and model inference results; anda storage unit configured to store AI / ML model parameters, training datasets, and historical channel and positioning data.

11. The network node as claimed in claim 10, wherein the processor is further configured to:train a first AI / ML model to reconstruct full-resolution channel information from reduced input; andtrain a second AI / ML model to estimate the position of the receiver node from the reconstructed channel information.

12. The network node as claimed in claim 10, wherein the reduced channel information is computed by:truncating CIR samples within a specified window size; selecting samples based on power or amplitude thresholds or by identifying the strongest samples; andquantizing the amplitude and phase of the selected samples.

13. The network node as claimed in claim 10, wherein the processor is configured to collect positioning datasets from one or more transmitter-receiver node pairs, the datasets including at least one of doppler shift, positioning measurements, and timestamps of measurement acquisition.

14. The network node as claimed in claim 10, wherein the processor is configured to:generate an associate identifier (ID) from transmission and reception parameters during training; andvalidate the associate identifier (ID) during inference to ensure consistency between training and inference phases.

15. The network node as claimed in claim 10, wherein the processor is further configured to periodically monitor the performance of the AI / ML models using performance metrics, including position error, range error, or angular error, and update the AI / ML models when the performance metrics exceed a predefined threshold.

16. The network node as claimed in claim 10, wherein the processor is configured to:compute a starting index using a first path delay or an estimated transmit- receive distance;determine the starting index using a sample rate defined by a product of an FFT size and a subcarrier spacing;select a sample window of a predefined size starting from the starting index, wherein the predefined size is selected from a set comprising 32, 64, 128, or 256 samples; andretain only samples whose power exceeds a predefined threshold.

17. The network node as claimed in claim 10, wherein the processor is further configured to determine at least one of line-of-sight (LoS) or non-line-of-sight (NLoS) probabilities based on the reconstructed channel information.

Images