A 5G non-line-of-sight compensation method and device based on a fusion learning mode

Abstract

The application provides a 5G non-line-of-sight compensation method and device based on a fusion learning mode, steps of the method include model pre-training and 5G non-line-of-sight compensation processing: in the step of model pre-training, a preset waveform generation network and an arrival time prediction network are trained based on a training data set, a first loss function is calculated based on the waveform generation network, a second loss function is calculated based on the arrival time prediction network, and fusion learning is performed based on the first loss function and the second loss function; the steps of the 5G non-line-of-sight compensation processing include: performing fast Fourier transform on an initial 5G signal to construct input features; inputting the input features into the waveform generation network, wherein the waveform generation network outputs compensation features; inputting the corrected positioning reference signal into the arrival time prediction network, wherein the arrival time prediction network outputs a predicted arrival time deviation based on a multi-layer perception.

Description

Technical Field

[0001] This invention relates to the field of wireless positioning technology, and in particular to a 5G non-line-of-sight compensation method and apparatus based on fusion learning. Background Technology

[0002] Wireless positioning systems (especially those based on 5G PRS / TOA) are often strongly affected by non-line-of-sight (NLOS) and multipath propagation in obstructed environments such as indoor spaces and urban canyons. NLOS introduces significant time delays / biases, causing a positive skew in Time of Arrival (TOA) ranging (often called ranging bias), which in turn amplifies the positioning error exponentially.

[0003] Traditional NLOS processing methods can be summarized into two categories: 1. Weighting / Robust Estimation: After identification, the weights are reduced or robust estimation (WLS, Huber, UKF / SRUKF, etc.) is used. Disadvantage: Limited compensation for severe biases.

[0004] 2. Geometry / Map-based / Virtual Station, Ray-tracing: This method uses environmental geometry or virtual base stations to convert reflection paths into equivalent line of sight (LOS). Disadvantages: It requires high-precision maps and high computational cost, making it difficult to deploy universally. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a 5G non-line-of-sight compensation method based on fusion learning, in order to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides a 5G non-line-of-sight compensation method based on fusion learning, the method comprising the steps of model pre-training and 5G non-line-of-sight compensation processing: In the model pre-training step, a preset waveform generation network and arrival time prediction network are trained based on the training dataset. The waveform generation network is configured with multiple waveform generation sub-networks connected in sequence. A first loss function is calculated based on the waveform generation network, and a second loss function is calculated based on the arrival time prediction network. The waveform generation network and the arrival time prediction network are then fused and learned based on the first loss function and the second loss function. The steps of the 5G non-line-of-sight compensation processing include: The initial 5G signal is subjected to a fast Fourier transform to construct input features. The input features are three-dimensional features, with the three dimensions being the time domain dimension, the frequency domain dimension, and the feature dimension, respectively. The input features are input into the waveform generation network, and the waveform generation network outputs compensation features. The compensation features correspond to the dimensions of the input features, wherein the corresponding feature in the frequency domain dimension is the corrected positioning reference signal. The corrected positioning reference signal is input into the arrival time prediction network, which is based on the predicted arrival time deviation output by the multilayer perceptron.

[0007] Using the above scheme, the input stage receives the current frame PRS waveform and channel environment parameters, extracts the time-domain waveform, frequency-domain features, and environment vectors to form a multi-domain input tensor; in the generation and denoising stage, the waveform generation network removes NLOS and multipath interference through an encoder-decoder structure, outputting a corrected frequency-domain PRS grid; in the TOA compensation stage, the input arrival time prediction network predicts the TOA deviation; in the delay correction and fusion stage, the corrected arrival time is calculated, and multiple ranging results are fused according to confidence levels to obtain a high-confidence TOA input for positioning filtering; in summary, this scheme can efficiently compensate for severe deviations and also meets the requirements of general deployment.

[0008] In some embodiments of the present invention, in the step of calculating the first loss function based on the waveform generation network, the first loss function is calculated using the following formula: in, This represents the corrected frequency domain positioning reference signal output by the waveform generation network. This represents the tag's frequency domain positioning reference signal. This represents the inverse fast Fourier transform. Represents the short-time Fourier transform. This represents the feature corresponding to the frequency domain dimension in the input features. This represents the corrected time-domain positioning reference signal output by the waveform generation network. and All of these are preset calculation parameters. This represents the value of the first loss function.

[0009] In some embodiments of the present invention, in the step of calculating the second loss function based on the arrival time prediction network, the second loss function is calculated using the following formula: in, This represents the value of the second loss function. This represents the predicted time of arrival deviation of the time-of-arrival prediction network output. This indicates the predicted time-of-arrival deviation of the label. This represents the prediction uncertainty of the time-of-arrival prediction network output. These are the preset calculation parameters.

[0010] In some embodiments of the present invention, in the step of performing fusion learning on the waveform generation network and the time-of-arrival prediction network based on the first loss function and the second loss function, the fusion learning loss function is calculated based on the first loss function and the second loss function, using the following formula: ; in, This represents the value of the fusion learning loss function. This indicates that the loss value is fine-tuned. , and All of these are preset weight parameters.

[0011] In some embodiments of the present invention, the waveform generation subnetwork includes an encoder network and a decoder network. In the step of inputting the input features into the waveform generation network and the waveform generation network outputting compensation features, the encoder network is provided with a first processing channel, a second processing channel, a third processing channel, and a fusion processing network. The first processing channel obtains the first feature based on a one-dimensional convolutional network; The second processing channel obtains the second feature based on complex-valued convolutional network processing; The third processing channel obtains the third feature based on multilayer perceptron processing; The fusion processing network receives the first feature, the second feature, and the third feature, and processes them based on an attention mechanism to obtain the fused feature.

[0012] In some embodiments of the present invention, the first processing channel is provided with a plurality of first processing blocks arranged in sequence and a fully connected layer, each first processing block being provided with a one-dimensional convolutional layer and a pooling layer arranged in sequence; the second processing channel is provided with a plurality of second processing blocks arranged in sequence and a fully connected layer, each second processing block being provided with a two-dimensional convolutional layer and a pooling layer arranged in sequence; the third processing channel is provided with a plurality of fully connected layers arranged in sequence, and the third processing channel constructs a multilayer perceptron through the plurality of fully connected layers.

[0013] In some embodiments of the present invention, in the step of receiving a first feature, a second feature, and a third feature in the fusion processing network and obtaining a fused feature based on an attention mechanism, the second feature and the third feature are processed using an attention mechanism, and the output of the attention mechanism is concatenated with the first feature to obtain the fused feature.

[0014] In some embodiments of the present invention, in the step of inputting the input features into the waveform generation network and the waveform generation network outputting compensation features, the decoder network is provided with a plurality of decoding units arranged in sequence, each decoding unit consisting of an upsampling layer and a convolutional layer arranged in sequence, and the decoder network outputs compensation features.

[0015] In some embodiments of the present invention, in the step of inputting the corrected positioning reference signal into the arrival time prediction network, and the arrival time prediction network outputting the predicted arrival time deviation based on the multilayer perceptron, the corrected positioning reference signal and the fusion feature of the last waveform generation subnetwork in the waveform generation network are input into the arrival time prediction network, and the multilayer perceptron of the arrival time prediction network outputs the predicted arrival time deviation.

[0016] A second aspect of the present invention also provides a 5G non-line-of-sight compensation device based on a fusion learning approach. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method described above.

[0017] A third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned 5G non-line-of-sight compensation method based on fusion learning.

[0018] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0019] Those skilled in the art will understand that the objectives and advantages achievable with this invention are not limited to those specifically described above, and that the above and other objectives achievable with this invention will become clearer from the following detailed description. Attached Figure Description

[0020] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0021] Figure 1 This is a schematic diagram of one embodiment of the 5G non-line-of-sight compensation method based on fusion learning according to the present invention; Figure 2 This is a schematic diagram of the waveform generation network in this scheme; Figure 3 This is a schematic diagram of the overall structure of the waveform generation network and the arrival time prediction network in this scheme. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0023] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0024] With the development of deep learning and generative models, existing technologies have emerged that directly use the original signal (such as CIR or PRS waveforms) as input and "denoise" or regress bias through supervised / generative models; others use filters / Kalman variants (such as SRUKF) to dynamically correct the residuals. Existing technologies mainly improve in the above directions, but there are few systematic engineering implementations of end-to-end generation / denoising methods that directly use the Positioning Reference Signal (PRS) waveform + Time of Arrival (TOA) as input and output "corrected PRS or TOA".

[0025] In the field of non-line-of-sight (NLOS) error compensation, traditional compensation methods are mainly based on two main approaches: geometric modeling and statistical filtering. Geometric modeling methods typically assume that the reflection path can be equivalently represented by a mirror source or a virtual base station (VBS). They calculate the additional propagation distance by analyzing the signal propagation path, reflector geometry, or ray tracing models, and then correct the ranging results. Theoretically, these methods can eliminate the additional time delay introduced by the reflection path with relatively high accuracy and have clear physical interpretability. However, their implementation heavily relies on prior knowledge of the environmental structure and accurate modeling, making them poorly adaptable to complex, variable, or dynamic indoor scenes. Furthermore, ray tracing and geometric optimization involve large computational costs, making it difficult to meet real-time requirements.

[0026] Another mainstream approach is error correction based on statistical estimation and filtering, such as Weighted Least Squares (WLS), Robust Estimator (M-estimator), Extended Kalman Filter (EKF), and Square Root Unscented Kalman Filter (SRUKF). These methods achieve adaptive correction for anomalous ranging by modeling or dynamically weighting the ranging residuals. Their advantage is that they do not require environmental geometry information and can be directly integrated into existing positioning frameworks. However, they also have significant limitations: firstly, the filtering performance depends on the accuracy of the noise statistical assumptions, and they are prone to failure when the NLOS bias exhibits a non-Gaussian distribution or abrupt changes; secondly, these methods typically only perform numerical corrections on the ranging results at the back end, and cannot directly recover the physical signal morphology distorted by multipath effects, thus the information loss at the signal level cannot be compensated for.

[0027] In recent years, some studies have attempted to introduce machine learning or data-driven compensation methods to predict and compensate for NLOS bias by learning the mapping relationship between signal features and ranging errors. For example, decision trees, XGBoost, or feedforward neural networks are used to regress multidimensional channel features to estimate and correct ranging bias. These methods do not rely on precise environmental models and are highly adaptable, but their effectiveness is limited by the number and diversity of training samples. When the environment changes or channel statistical characteristics shift, the model often needs to be retrained or fine-tuned, making it difficult to guarantee long-term stability and cross-scenario generalization ability.

[0028] In addition, some studies have proposed end-to-end regression or generative compensation methods based on deep neural networks. These methods directly use the original waveform (such as CIR or PRS) as input and output corrected ranging or reconstructed clean waveforms, thereby recovering information perturbed by NLOS at the signal level. While these methods have shown good compensation accuracy in experiments, they suffer from large model parameters, computational complexity, long inference latency, and reliance on a large amount of labeled data with ground truth for supervised training, which limits their practical deployment.

[0029] In summary, existing NLOS compensation methods each have their advantages in geometric modeling, statistical filtering, and data-driven learning, but they also share common problems: they either rely on prior environmental information and have poor real-time performance, or they depend on large-scale labeled data and have insufficient generalization, or they can only correct ranging at the numerical level without recovering the signal itself. Therefore, there is still a lack of a unified solution that can directly correct PRS waveforms at the signal level, output accurate and reliable TOA compensation results, and balance real-time performance and deployability.

[0030] like Figure 1 As shown, this invention proposes a 5G non-line-of-sight compensation method based on fusion learning. The method includes model pre-training and 5G non-line-of-sight compensation processing. The model pre-training includes step S100, which involves training a preset waveform generation network and arrival time prediction network based on a training dataset. The waveform generation network is configured with multiple waveform generation sub-networks connected in sequence. A first loss function is calculated based on the waveform generation network, and a second loss function is calculated based on the arrival time prediction network. The waveform generation network and the arrival time prediction network are then fused and learned based on the first loss function and the second loss function. The 5G non-line-of-sight compensation processing includes: Step S210: Perform a fast Fourier transform on the initial 5G signal to construct input features. The input features are three-dimensional features, with the three dimensions being the time domain dimension, the frequency domain dimension, and the feature dimension. In the specific implementation process, in order to fully express the distortion characteristics of PRS waveforms under non-line-of-sight (NLOS) conditions, this invention introduces a multi-domain feature construction mechanism in the input layer, which comprehensively utilizes information from the time domain, frequency domain, and environmental domain to obtain a multi-scale characterization of signal propagation distortion.

[0031] Time-domain characteristics: including the original received PRS waveform (complex form) and its real envelope, instantaneous power, group delay curve, etc.; these characteristics are used to reflect the degree of signal distortion and energy accumulation characteristics in the time dimension. Frequency domain characteristics: The frequency domain power spectral density, phase spectrum, and main-lobe ratio are obtained by performing a Fast Fourier Transform (FFT) on the original received PRS waveform. Feature dimension features, i.e., environmental features: defined as , , respectively represent the channel RMS delay spread, signal-to-noise ratio, and phase variance, used to describe the macroscopic statistical characteristics of the propagation environment; The above features collectively form the three-dimensional tensor input in the network: ,in Features in the time domain dimension For frequency domain dimension features, These are the feature dimensions corresponding to environmental characteristics.

[0032] Step S220: Input the input features into the waveform generation network, and the waveform generation network outputs compensation features. The compensation features correspond to the dimensions of the input features, wherein the corresponding feature in the frequency domain dimension is the corrected positioning reference signal. In the specific implementation process, in order to recover the PRS signal distorted by reflection, multipath and NLOS propagation, this invention proposes a waveform generation sub-network structure based on Conditional Denoising Autoencoder (CDAE). Its input is a multi-domain feature tensor and its output is an ideal line-of-sight frequency domain PRS grid after denoising and reconstruction, that is, the frequency domain dimension feature in the final output.

[0033] Step S230: The corrected positioning reference signal is input into the arrival time prediction network, which is based on the arrival time deviation predicted by the multilayer perceptron.

[0034] Using the above scheme, the input stage receives the current frame PRS waveform and channel environment parameters, extracts the time-domain waveform, frequency-domain features, and environment vectors to form a multi-domain input tensor; in the generation and denoising stage, the waveform generation network removes NLOS and multipath interference through an encoder-decoder structure, outputting a corrected frequency-domain PRS grid; in the TOA compensation stage, the input arrival time prediction network predicts the TOA deviation; in the delay correction and fusion stage, the corrected arrival time is calculated, and multiple ranging results are fused according to confidence levels to obtain a high-confidence TOA input for positioning filtering; in summary, this scheme can efficiently compensate for severe deviations and also meets the requirements of general deployment.

[0035] In some embodiments of the present invention, in the step of calculating the first loss function based on the waveform generation network, the first loss function is calculated using the following formula: in, This represents the corrected frequency domain positioning reference signal output by the waveform generation network. This represents the tag's frequency domain positioning reference signal. This represents the inverse fast Fourier transform. Represents the short-time Fourier transform. This represents the feature corresponding to the frequency domain dimension in the input features. This represents the corrected time-domain positioning reference signal output by the waveform generation network. and All of these are preset calculation parameters. This represents the value of the first loss function.

[0036] Using the above scheme, the training phase of this scheme uses an ideal LOS frequency domain grid. With the objective of minimizing the time-frequency joint loss, Ensure frequency domain amplitude and phase consistency; Maintain consistency in the time-domain waveform structure; Constraint spectral smoothing and energy consistency; Represents the L2 norm, The design, which uses a first norm, enables multi-domain coupled reconstruction and environmentally adaptive denoising of PRS signals.

[0037] In some embodiments of the present invention, in the step of calculating the second loss function based on the arrival time prediction network, the second loss function is calculated using the following formula: in, This represents the value of the second loss function. This represents the predicted time of arrival deviation of the time-of-arrival prediction network output. This indicates the predicted time-of-arrival deviation of the label. This represents the prediction uncertainty of the time-of-arrival prediction network output. These are the preset calculation parameters.

[0038] Using the above scheme, the first item of this scheme Used to introduce adaptive weights into the regression error term: when adverse propagation conditions (such as low SNR, heavy occlusion) lead to unreliable predictions. It will automatically increase, thus reducing the impact of this sample on the overall optimization; the second item The normalization factor in Gaussian likelihood is introduced to prevent the network from growing indefinitely. The visitor deliberately downplayed the first loss; the third... By directly incorporating the frequency domain consistency constraint into the TOA loss, this scheme achieves a joint optimization mechanism for "waveform reconstruction quality and time delay compensation performance".

[0039] Furthermore, to reduce reliance on labeled data, this invention introduces a reinforcement learning framework based on policy gradient. The agent uses the minimum localization error after TOA correction as the reward signal and achieves adaptive optimization of unlabeled samples through interaction with the localization module.

[0040] The reward function is defined as: This represents the estimated position vector obtained by using the corrected TOA as input and a positioning algorithm. This is the actual position vector at that moment; Represents the L2 norm; Joint parameter updates are based on policy gradients: Joint parameter updates are based on policy gradients. Let the expected reward objective function be used. This represents the logarithmic probability of the output action under the current parameters; This represents the gradient operator with respect to all network parameters; This represents the mathematical expectation of the sampling trajectory; This represents all learnable parameters of the generating subnetwork (CDAE); This represents all learnable parameters of the TOA compensation subnetwork. Through this mechanism, the model can continuously optimize latency compensation performance in actual deployments.

[0041] In some embodiments of the present invention, in the step of performing fusion learning on the waveform generation network and the time-of-arrival prediction network based on the first loss function and the second loss function, the fusion learning loss function is calculated based on the first loss function and the second loss function, using the following formula: ; in, This represents the value of the fusion learning loss function. This indicates that the loss value is fine-tuned. , and All of these are preset weight parameters.

[0042] This invention employs a phased training strategy: Supervised pre-training phase: Using calibrated samples, the parameters of the CDAE and TOA sub-networks are optimized to enable them to have preliminary waveform correction and bias regression capabilities; Enhanced fine-tuning phase: On unlabeled samples, online optimization is performed through the reinforcement learning reward function to ensure that the model maintains stable adaptive performance in dynamic channels.

[0043] The overall optimization objective function is: in , , The relative importance of signal reconstruction accuracy, TOA regression accuracy, and reinforcement learning policy optimization to the overall training objective is measured separately, where: This is a parameterization strategy composed of a waveform generation subnetwork and a TOA compensation subnetwork. An instant reward defined based on positioning error. Represents the action variable Distribution by strategy After sampling, the expected value is obtained.

[0044] In the dataset preparation phase for model training, to achieve algorithm verification and model training of this invention, multi-scenario measured data were collected in a 5G experimental site. The experimental scenarios included typical channel environments such as open line-of-sight (LOS), single reflection, double reflection, and diffuse reflection, covering the relative positions of different base stations (BS) and user equipment (UE). The collected data included: 1. Standardized PRS reference sequence transmitted by the base station; 2. Distorted PRS waveform received by the user equipment (including time domain and frequency domain forms). 3. The initial TOA estimate for each frame of signal; 4. Channel statistics, such as RMS delay spread (r_RMS), signal-to-noise ratio (SNR), phase variance (φ_var), etc. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio.

[0045] During the data calibration phase, high-precision laser ranging equipment and geometric truth calculation methods are used to obtain the true TOA value for each link. Based on this, the TOA deviation is calculated. .

[0046] Using precise ranging equipment (such as laser rangefinders) and geometric truth calculations, the true TOA value for each link is obtained, and the TOA deviation is calculated accordingly. .

[0047] For waveform level calibration, this invention uses cross-correlation matching and phase alignment to calibrate the ideal LOS PRS waveform. Compared with the measured waveform Registration yields the target corrected waveform label. This calibration process only needs to be performed on a small number of samples and is used in the initialization phase of semi-supervised training.

[0048] like Figure 2 As shown, in some embodiments of the present invention, the waveform generation sub-network includes an encoder network and a decoder network. In the step of inputting the input features into the waveform generation network and the waveform generation network outputting compensation features, the encoder network is provided with a first processing channel, a second processing channel, a third processing channel, and a fusion processing network. The first processing channel obtains the first feature based on a one-dimensional convolutional network; The second processing channel obtains the second feature based on complex-valued convolutional network processing; The third processing channel obtains the third feature based on multilayer perceptron processing; The fusion processing network receives the first feature, the second feature, and the third feature, and processes them based on an attention mechanism to obtain the fused feature.

[0049] In the specific implementation process, the input of the first processing channel is the time domain dimension of the input features, the input of the second processing channel is the frequency domain dimension, and the input of the third processing channel is the feature dimension.

[0050] Using the above scheme, the encoder part of this scheme consists of three feature extraction branches that are structurally independent but fused at a high level, respectively performing multi-granularity modeling for time domain, frequency domain and environmental information; the three features are spliced ​​and fused with attention to obtain the latent representation, wherein the attention mechanism uses the environmental features of the feature dimension as the weight generation function to guide the network to adaptively strengthen or suppress features of different frequencies / time periods under dynamic channel conditions, so that the model has stable expressive ability in different scenarios.

[0051] In some embodiments of the present invention, the first processing channel is provided with a plurality of first processing blocks arranged in sequence and a fully connected layer, each first processing block being provided with a one-dimensional convolutional layer and a pooling layer arranged in sequence; the second processing channel is provided with a plurality of second processing blocks arranged in sequence and a fully connected layer, each second processing block being provided with a two-dimensional convolutional layer and a pooling layer arranged in sequence; the third processing channel is provided with a plurality of fully connected layers arranged in sequence, and the third processing channel constructs a multilayer perceptron through the plurality of fully connected layers.

[0052] In some embodiments of the present invention, in the step of receiving a first feature, a second feature, and a third feature in the fusion processing network and obtaining a fused feature based on an attention mechanism, the second feature and the third feature are processed using an attention mechanism, and the output of the attention mechanism is concatenated with the first feature to obtain the fused feature.

[0053] In the specific implementation process, the decoder inputs the fusion features and the environmental features corresponding to the input features of the waveform generation network. The environmental modulation is introduced through the Conditional BatchNorm to achieve adaptive waveform correction. The decoder adopts a transposed convolution and skip connection structure to restore high-frequency details and phase continuity.

[0054] In some embodiments of the present invention, in the step of inputting the input features into the waveform generation network and the waveform generation network outputting compensation features, the decoder network is provided with a plurality of decoding units arranged in sequence, each decoding unit consisting of an upsampling layer and a convolutional layer arranged in sequence, and the decoder network outputs compensation features.

[0055] In the specific implementation process, the waveform generation network is configured with three waveform generation sub-networks, which are arranged sequentially. The input and output of each waveform generation sub-network are three-dimensional features. Specifically, if it is the first waveform generation sub-network, the input feature is the initial input feature; if it is not the first waveform generation sub-network, the input feature is the three-dimensional compensation feature output by the previous waveform generation sub-network.

[0056] like Figure 3 As shown, Figure 3 middle In some embodiments of the present invention, in the step of inputting the corrected positioning reference signal into the arrival time prediction network, and the arrival time prediction network outputting the predicted arrival time deviation based on the multilayer perceptron, the fusion feature of the corrected positioning reference signal and the last waveform generation subnetwork in the waveform generation network is input into the arrival time prediction network, and the multilayer perceptron of the arrival time prediction network outputs the predicted arrival time deviation.

[0057] In practical implementation, the output of the arrival time prediction network receives the waveform generation network. The fused features of the last waveform generation subnetwork in the waveform generation network are input into the arrival time prediction network, and the arrival time deviation is predicted by the multilayer perceptron (MLP).

[0058] To further improve the accuracy of time of arrival (TOA) ranging using the above scheme, this scheme proposes an TOA prediction network that combines regression prediction mechanism and reinforcement learning optimization method. This network can not only output the delay correction amount, but also provide the corresponding prediction confidence, thereby providing quantifiable measurement reliability for subsequent positioning systems.

[0059] In summary, this solution addresses the shortcomings of existing technologies by proposing an end-to-end NLOS compensation system based on generation / denoising + regression + uncertainty estimation. This solution can directly recover PRS waveforms perturbed by multipath / NLOS, thereby repairing information at the waveform level and improving the accuracy of subsequent TOA / AoA / multipath extraction. It can directly correct TOA (ranging bias) and output a correction value with confidence, enabling better fusion of positioning filters and improving terminal positioning accuracy. It reduces dependence on high-precision maps / environmental geometry and generalizes across multiple scenarios through a data-driven approach. This solution balances real-time performance and deployability, supporting edge device deployment through model compression / distillation and online fine-tuning mechanisms. This solution provides interpretability and a safe fallback mechanism, allowing fallback to traditional filters or conservative weighting strategies in unknown domains or when model uncertainty is high.

[0060] This solution provides a 5G NLOS compensation method based on generative reinforcement learning for waveform-level correction and precise time-of-arrival (TOA) compensation of the Positioning Reference Signal (PRS) in complex indoor and outdoor multipath environments. The method takes the original PRS waveform (time-domain or frequency-domain representation) as input and constructs a dual-path network structure that fuses generative denoising and reinforcement regression. It directly removes distortions caused by reflections and multipath propagation at the signal layer and outputs the physically corrected PRS along with the corresponding TOA correction. This solution does not rely on environmental geometric models or large-scale labeled data and can adapt to non-line-of-sight errors under different propagation conditions, achieving real-time, low-latency compensation and improved positioning accuracy.

[0061] The beneficial effects of this plan include: 1. This solution can directly recover the PRS waveform: Compared with methods that only regress the deviation, this invention can correct the waveform at the signal level, thereby improving the accuracy of subsequent TOA / AoA / multipath structure extraction.

[0062] 2. This solution combines TOA correction and localization optimization: By introducing localization loss or output confidence into the training process and combining it with a filter, link-level improvements can be directly transformed into improved localization accuracy.

[0063] 3. This solution reduces dependence on environmental geometry: it is primarily data-driven, combined with simulation enhancement and a small amount of experimental fine-tuning, to adapt to unknown / dynamic environments.

[0064] 4. This solution ensures scalability and engineering feasibility: the model can be distilled and quantized, and is compatible with edge deployment; the strategy generator (RL) can provide online adaptation and safe rollback.

[0065] 5. This solution proposes an end-to-end compensation network structure that takes the PRS waveform (time domain or frequency domain) and the original TOA measurement as input, and outputs the corrected PRS waveform and / or TOA value. This structure combines an encoder-decoder denoising generator with a regression prediction head, and achieves coordinated optimization of the signal layer and the positioning layer by jointly optimizing the waveform reconstruction loss, time delay deviation regression loss, and positioning error auxiliary loss, thereby fundamentally reducing the ranging error caused by NLOS propagation.

[0066] This invention also provides a 5G non-line-of-sight compensation device based on a fusion learning approach. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method described above.

[0067] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned 5G non-line-of-sight compensation method based on fusion learning. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0068] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0069] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0070] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A 5G non-line-of-sight compensation method based on fusion learning, characterized in that, The method includes model pre-training and 5G non-line-of-sight compensation processing: In the model pre-training step, a preset waveform generation network and arrival time prediction network are trained based on the training dataset. The waveform generation network is configured with multiple waveform generation sub-networks connected in sequence. A first loss function is calculated based on the waveform generation network, and a second loss function is calculated based on the arrival time prediction network. The waveform generation network and the arrival time prediction network are then fused and learned based on the first loss function and the second loss function. The steps of the 5G non-line-of-sight compensation processing include: The initial 5G signal is subjected to a fast Fourier transform to construct input features. The input features are three-dimensional features, with the three dimensions being the time domain dimension, the frequency domain dimension, and the feature dimension, respectively. The input features are fed into the waveform generation network, which outputs compensated features. These compensated features correspond to the dimensions of the input features, with the corresponding feature in the frequency domain being the corrected positioning reference signal. The waveform generation sub-network includes an encoder network and a decoder network. The encoder network has a first processing channel, a second processing channel, a third processing channel, and a fusion processing network. The first processing channel uses a one-dimensional convolutional network to obtain a first feature; the second processing channel uses a complex-valued convolutional network to obtain a second feature; the third processing channel uses a multilayer perceptron to obtain a third feature; and the fusion processing network receives the first, second, and third features and processes them using an attention mechanism to obtain a fused feature. The corrected positioning reference signal is input into the arrival time prediction network, which is based on the predicted arrival time deviation output by the multilayer perceptron.

2. The 5G non-line-of-sight compensation method based on fusion learning as described in claim 1, characterized in that, In the step of calculating the first loss function based on the waveform generation network, the first loss function is calculated using the following formula: in, This represents the corrected frequency domain positioning reference signal output by the waveform generation network. This represents the tag's frequency domain positioning reference signal. This represents the inverse fast Fourier transform. Represents the short-time Fourier transform. This represents the feature corresponding to the frequency domain dimension in the input features. This represents the corrected time-domain positioning reference signal output by the waveform generation network. and All of these are preset calculation parameters. This represents the value of the first loss function.

3. The 5G non-line-of-sight compensation method based on fusion learning as described in claim 2, characterized in that, In the step of calculating the second loss function based on the arrival time prediction network, the second loss function is calculated using the following formula: in, This represents the value of the second loss function. This represents the predicted time of arrival deviation of the time-of-arrival prediction network output. This indicates the predicted time-of-arrival deviation of the label. This represents the prediction uncertainty of the time-of-arrival prediction network output. These are the preset calculation parameters.

4. The 5G non-line-of-sight compensation method based on fusion learning as described in claim 3, characterized in that, In the step of fusing the waveform generation network and the time-of-arrival prediction network based on the first loss function and the second loss function, the fusing learning loss function is calculated based on the first loss function and the second loss function, using the following formula: ； in, This represents the value of the fusion learning loss function. This indicates that the loss value is fine-tuned. , and All of these are preset weight parameters.

5. The 5G non-line-of-sight compensation method based on fusion learning as described in claim 1, characterized in that, The first processing channel is provided with a plurality of first processing blocks arranged in sequence and a fully connected layer. Each first processing block is provided with a one-dimensional convolutional layer and a pooling layer arranged in sequence. The second processing channel is provided with a plurality of second processing blocks arranged in sequence and a fully connected layer. Each second processing block is provided with a two-dimensional convolutional layer and a pooling layer arranged in sequence. The third processing channel is provided with a plurality of fully connected layers arranged in sequence, and the third processing channel constructs a multilayer perceptron through the plurality of fully connected layers.

6. The 5G non-line-of-sight compensation method based on fusion learning according to claim 1, characterized in that, In the step of receiving the first feature, the second feature, and the third feature in the fusion processing network and obtaining the fused feature based on the attention mechanism, the second feature and the third feature are processed using the attention mechanism, and the output of the attention mechanism is concatenated with the first feature to obtain the fused feature.

7. The 5G non-line-of-sight compensation method based on fusion learning as described in claim 1, characterized in that, In the step of inputting the input features into the waveform generation network and the waveform generation network outputting compensation features, the decoder network is provided with multiple decoding units arranged in sequence. Each decoding unit consists of an upsampling layer and a convolutional layer arranged in sequence, and the decoder network outputs compensation features.

8. The 5G non-line-of-sight compensation method based on fusion learning according to claim 1, characterized in that, In the step of inputting the corrected positioning reference signal into the arrival time prediction network, and the arrival time prediction network outputting the predicted arrival time deviation based on the multilayer perceptron, the fusion feature of the corrected positioning reference signal and the last waveform generation subnetwork in the waveform generation network is input into the arrival time prediction network, and the multilayer perceptron of the arrival time prediction network outputs the predicted arrival time deviation.

9. A 5G non-line-of-sight compensation device based on fusion learning, characterized in that, The device includes a computer device, which includes a processor and a memory, wherein computer instructions are stored in the memory, and the processor is configured to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method as described in any one of claims 1 to 8.

Images