A 6G fingerprint positioning model intelligent reconstruction method for propagating environment inversion cognition

Abstract

This invention discloses an intelligent reconstruction method for 6G fingerprint positioning models based on propagation environment inversion cognition, belonging to the field of wireless communication and positioning technology. Addressing the problems of existing fingerprint positioning methods, such as lack of environmental awareness, inaccurate global change assessment, and high model update costs, this invention acquires multi-base station collaborative multi-source channel observation data in real time to construct a fused fingerprint feature set, inverts and extracts core environmental characteristics, and identifies whether the terminal is located indoors, outdoors, or in a transitional area; constructs a differentiated environmental change assessment model to output change levels; dynamically generates and optimizes hyperparameters through a meta-learning-based intelligent hyperparameter adaptive network, and achieves hierarchical intelligent model reconstruction by combining parameter migration, domain alignment, sample weighting, and structural constraints, ultimately outputting high-precision location estimation results. This invention achieves highly robust and rapidly adaptive intelligent positioning in dynamic and complex 6G scenarios, significantly reducing model update costs.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication and positioning technology, and specifically relates to a method for intelligent reconstruction of a 6G fingerprint positioning model based on propagation environment inversion cognition. Background Technology

[0002] As 6G networks continue to evolve towards ultra-dense networking, integrated sensing, and air-space-ground collaboration, high-precision positioning is becoming a key foundational capability in scenarios such as intelligent transportation, smart campuses, industrial internet, emergency rescue, and human-computer interaction. Fingerprint positioning technology, by establishing a mapping relationship between signal observation characteristics and spatial location, has advantages such as not requiring strict line-of-sight, adapting to complex scenarios, and flexible deployment, thus maintaining broad application prospects in the 6G environment.

[0003] However, most existing fingerprint localization methods construct static fingerprint databases or static models in relatively stable environments, which have the following main shortcomings: First, there is a lack of environmental perception capabilities based on fingerprint features. Existing methods typically use RSSI, CSI, or a small number of statistics directly for location matching, with little further mining of the electromagnetic propagation environment information implicit in these fingerprint data. Walls, glass, metal obstacles, pedestrian density, and openness in the environment can cause significant changes in reflection, diffraction, scattering, multipath number, time delay spread, and angular spread, which can lead to shifts in fingerprint distribution. If the surrounding propagation environment and its changes cannot be inferred from fingerprint features, the localization model will have poor adaptability to environmental disturbances.

[0004] Second, there is a lack of a unified environmental change assessment mechanism covering the entire region. Indoor areas typically exhibit strong reflection, dense multipath propagation, and significant time delay spread; outdoor areas typically exhibit wide propagation, line-of-sight path dominance, and relatively high link stability; transitional areas simultaneously possess characteristics such as abrupt occlusion, mixed propagation mechanisms, and ambiguous regional discrimination. Existing methods mostly employ a single threshold or unified evaluation index, making it difficult to accurately measure the intensity of environmental changes in these three types of areas, and also making it difficult to determine whether the model needs to be updated and to what extent.

[0005] Third, there is a lack of a rapid model migration and reconstruction mechanism based on environmental changes. When the scene changes, traditional fingerprint positioning usually requires the re-collection of a large amount of reference point data and the overall retraining of the model, which is costly and time-consuming, making it difficult to meet the real-time positioning requirements of 6G dynamic environments. Especially in situations where there are frequent indoor-outdoor transitions or complex changes in transition areas, if region-related migration and rapid reconstruction based on existing models cannot be performed, it is difficult to balance positioning accuracy, computational efficiency, and update costs.

[0006] Therefore, there is an urgent need to propose a new 6G intelligent fingerprint positioning model reconstruction method that can perceive environmental propagation attributes from multi-source fingerprint data, establish an environmental change assessment mechanism for indoor areas, outdoor areas and transition areas, and select and reconstruct a fingerprint model that is adapted to the current scenario based on the environmental change results, thereby achieving highly robust, high-precision and rapidly updated intelligent positioning. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a 6G fingerprint positioning model intelligent reconstruction method based on propagation environment inversion cognition, in order to solve the technical problems of difficulty in perceiving environmental changes, inaccurate assessment of changes in indoor and outdoor areas and transition zones, and difficulty in quickly migrating and reconstructing fingerprint models in existing methods.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A method for intelligent reconstruction of a 6G fingerprint positioning model based on propagation environment inversion cognition includes the following steps: Step 1: Real-time acquisition of multi-base station collaborative multi-source channel observation data reported by the terminal, and construction and output of a multi-source fusion sensing fingerprint feature set characterizing the radio wave propagation environment; Step 2: Using the multi-source fusion sensing fingerprint feature set as input, perform inversion cognitive analysis on the radio wave propagation environment around the terminal, extract environmental characterization quantities such as power fading dispersion, time delay spread factor, angle spread factor, reflection enhancement index and diffraction sensitivity index, construct environmental inversion cognitive feature vector, and output the identification results of the indoor area, outdoor area or transition area between indoor and outdoor where the terminal is located. Step 3: Using the environmental inversion cognitive feature vector and the region identification result as input, construct a differentiated environmental change assessment model for the corresponding region, calculate the change intensity index of the current environment relative to the offline pre-stored reference environment of the region, and output the environmental change level; Step 4: Using the region identification results and environmental change level as input, select the source model to be transferred for the corresponding region from the pre-established regional source fingerprint model library. Through the intelligent hyperparameter adaptive network based on meta-learning and dynamically generate the optimal balanced hyperparameter set of the objective function based on the cognitive features of the current environment, combine parameter transfer, domain feature alignment, sample weighting and structural constraints to intelligently reconstruct the fingerprint model and output the target fingerprint localization model adapted to the current environment. Step 5: Using the multi-source fusion sensing fingerprint features observed in real-time online as input, input the target fingerprint localization model to obtain an initial location estimate, perform similarity fine-tuning correction based on the initial location estimate, and output the final location estimate result of the terminal.

[0009] Furthermore, in step 2, the extracted environmental characterization quantities are combined with the average effective propagation path number to construct an environmental inversion cognitive feature vector. :

[0010] in, For power fading dispersion, For delay spread factor, For the angle expansion factor, As an index for enhanced reflection, As a diffraction-sensitive indicator, This represents the average effective propagation path number.

[0011] Furthermore, in step 2, the cognitive feature vector is derived based on the environmental inversion. The region classifier identifies the region where the terminal is located. :

[0012] in, Let be the posterior probability classification function for the region. Indicates indoor area, Indicates outdoor area, Indicates a transitional region.

[0013] Furthermore, in step 3, the dynamic environment change assessment score is obtained by weighting the comprehensive environmental offset, fingerprint distribution drift, and base station coordination consistency index. :

[0014] in, This is the environmental offset. For fingerprint distribution drift, This is a base station coordination consistency indicator. The preset empirical weighting coefficients for the corresponding regions, and satisfy the following conditions: .

[0015] Furthermore, in step 3, a threshold is determined based on offline historical environmental data. The environmental change level is divided into three levels: Slight changes: Satisfied

[0016] Moderate change: Satisfied

[0017] Significant changes: meet the requirements .

[0018] Furthermore, in step 4, the intelligent hyperparameter adaptive network based on meta-learning uses environmental inversion to retrieve cognitive feature vectors. and environmental change assessment score As input, an adaptive and dynamic set of four optimal balanced hyperparameters for optimizing the objective function is generated:

[0019] in, For meta-learning hyperparameter adaptive network mapping function, The optimal balanced hyperparameters are dynamically generated by the meta-learning-based intelligent hyperparameter adaptive network and are used to control the weight coefficients of the localization supervision loss term, parameter migration constraint term, unsupervised domain alignment loss term, and structure preservation constraint term in the overall objective function.

[0020] Furthermore, in step 4, the offline source model parameters are used as initialization parameters, and an overall objective optimization function is constructed based on the generated hyperparameter set. :

[0021] in, The parameters of the target fingerprint localization model to be optimized are: To locate and monitor loss items, For parameter migration constraints, For unsupervised domain alignment loss term, Constraints are used to maintain the structure.

[0022] Furthermore, in step 4, a hierarchical parameter update strategy is implemented based on the level of environmental change: When the environmental change level is slight, only the output layer parameters of the target fingerprint localization model are updated; When the environmental change level is moderate, freeze the parameters of the bottom-level perception feature extraction layer of the target fingerprint localization model and only update the parameters of the high-level mapping layer. When the environmental change level is significant, the target fingerprint localization model is updated with all parameters.

[0023] Furthermore, in step 5, G candidate reference points with the closest spatial feature distances are selected based on the initial position estimate, and the final position estimate of the terminal is obtained by weighting the distances using the inverse of the distance. :

[0024] in, For the first The physical location coordinates of the candidate reference points For online fingerprint and the first The Euclidean distance between the fingerprints of candidate reference points in the reconstructed feature space.

[0025] Furthermore, the regional source fingerprint model library is pre-built in an offline phase, specifically including: Multiple reference points were deployed indoors, outdoors and in transitional areas and their physical coordinates were marked. Multi-source channel data from multiple base stations were collected to establish regional source fingerprint datasets. The corresponding regional source fingerprint models are trained offline based on the source fingerprint datasets of each region. At the same time, the reference environment statistical mean vector, covariance matrix and reference fingerprint correlation matrix of each region are calculated and stored.

[0026] Beneficial effects: First, it achieves a deep fusion of physical layer fingerprint features and environmental inversion cognition. This invention does not simply use fingerprints as a location input, but extracts propagation attributes to construct an environmental inversion cognition feature vector, enabling the positioning system to actively perceive the radio wave propagation environment, overcoming the shortcomings of traditional methods that rely on "passive and blind matching".

[0027] Second, a differentiated environmental change assessment model for the entire region is proposed. This invention establishes a differentiated evaluation framework for indoor, outdoor, and transitional areas, incorporating environmental offset, fingerprint distribution drift, and multi-base station collaborative consistency into a unified model, and assigning weights based on regional physical characteristics to accurately capture the degree of environmental change, avoiding the failure problem caused by a unified evaluation strategy.

[0028] Third, it realizes intelligent adaptive model reconstruction based on meta-learning. This invention innovatively introduces a hyperparameter adaptive network based on meta-learning, upgrading the model reconstruction strategy from "static manual preset" to "dynamic intelligent decision-making"; at the same time, it combines parameter transfer, domain alignment, reliability weighting and structural constraints to carry out hierarchical reconstruction, which greatly improves the accuracy and robustness of model transfer under unlabeled small sample conditions.

[0029] Fourth, it supports multi-base station collaboration and rapid reconfiguration, significantly reducing update costs. By fully utilizing multi-base station collaborative observation, the global structural stability across base stations is ensured. Compared to traditional methods that require full retraining, this invention only needs a small number of online samples to quickly complete model adaptation, meeting the real-time and efficient positioning requirements in dynamic and complex 6G scenarios. Attached Figure Description

[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is the main flowchart of the intelligent reconstruction method for the 6G fingerprint positioning model based on the propagation environment inversion cognition described in this embodiment of the invention. Detailed Implementation

[0031] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0032] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] Example 1 like Figure 1 As shown, the intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition proposed in this invention includes two parts: an offline stage and an online stage. The online stage is executed according to the following steps: Step S1: Construction of Multi-Source Fusion Sensing Fingerprint Features Real-time acquisition of terminal and The uplink probe signals between the cooperating base stations are used to estimate the parameters of the signals to obtain the underlying physical propagation quantity: the taps of the channel impulse response (CIR) are extracted to obtain the first... The first collaborative base station Delay of an effective propagation path With complex amplitude And calculate the corresponding path power. The angle of arrival (AH) is obtained through the channel state information (CSI) matrix estimation algorithm. The first collaborative base station Angle of arrival of an effective propagation path The arrival time (TOA) is obtained by combining timed advance measurements.

[0034] Depend on The collaborative base station collects channel observations from the terminal in real time to construct the first... Local fingerprint vectors of each base station:

[0035] in, Indicates the first Received signal strength (RSSI) observations of each base station. Represents the CSI feature vector. Represents the CIR eigenvector. This represents the TOA observation value of the principal diameter. This represents the AOA observation value of the principal diameter. This indicates the number of effective propagation paths corresponding to the base station.

[0036] For all The local fingerprint vectors of the cooperating base stations are concatenated to obtain the global multi-source fusion sensing fingerprint features:

[0037] in, Represents the transpose operation of a matrix or vector; Indicates all The global multi-source fusion sensing fingerprint feature vector is constructed by concatenating the local fingerprint vectors of the collaborative base stations.

[0038] Step S2: Propagation Environment Inversion Cognition and Region Identification Based on the global multi-source fusion sensing fingerprint features obtained in step S1, the following environmental propagation attribute representations are extracted: Average signal strength of all cooperating base stations:

[0039] Power fading dispersion:

[0040] No. Power-weighted average latency of each base station:

[0041] Delay spread factor:

[0042] No. Power-weighted average angle of arrival of each base station:

[0043] Angular expansion factor:

[0044] Reflection enhancement index:

[0045] in, Indicates the first An index of the propagation path set dominated by specular reflection for each base station.

[0046] Diffraction Sensitive Indicators:

[0047] in, Indicates the first Each base station is indexed by a set of propagation paths dominated by diffraction propagation.

[0048] Average effective propagation path number:

[0049] The power fading dispersion extracted above Delay spread factor Angular expansion factor Reflection enhancement index Diffraction Sensitive Indicators and average effective propagation path Combining and constructing environmental inversion cognitive feature vectors :

[0050] Use a pre-trained region discriminant classifier to identify the specific region where the current terminal is located. :

[0051] in, Let be the posterior probability classification function for the region. Indicates indoor area, Indicates outdoor area, Indicates a transitional region.

[0052] Step S3: Differentiated Environmental Change Assessment Based on the specific area identified in step S2 The reference environment statistical mean vector corresponding to the region is retrieved from the offline regional model library. Covariance Matrix The environmental inversion cognitive feature vector constructed in step 2 at the current sampling time is denoted as... Define the region-related environmental offset for:

[0053] Construct a multi-base station fingerprint correlation matrix using consecutive frames of online fingerprint observation data within the current time window. , its first Line 1 Column elements Indicates the number of times within the current time window. The base station and the first The Pearson correlation coefficient between the fingerprint feature sequences of each base station; Regional reference correlation matrix retrieved from regional model library Calculate the Frobenius norm of the matrix difference and define the fingerprint distribution drift. for:

[0054] Calculate base station coordination consistency index based on correlation coefficient :

[0055] Based on the above three indicators, the dynamic environment change assessment score of the current window is calculated. :

[0056] in, For specific areas Preset empirical weighting coefficients, and satisfying .

[0057] Further based on the preset threshold obtained by calibration through offline historical environmental data The environmental change level is divided into three levels: Slight changes: Satisfied

[0058] Moderate change: Satisfied

[0059] Significant changes: meet the requirements .

[0060] Step S4: Intelligent Model Reconstruction Based on Meta-Learning For the specific area identified in step S2 Select the corresponding offline source model from the offline established regional source fingerprint model library, and denot its pre-trained network parameters as follows: In the current changing environment, online collection of unlabeled small sample sets of the target domain is crucial. ,in This indicates the number of times the image is observed online within the current window. Global multi-source fusion sensing fingerprint features in the target domain This represents the total number of online samples within the current window.

[0061] Retrieve fingerprints from a pre-established regional fingerprint model library for specific regions. Corresponding regional source fingerprint dataset The total number of offline reference points in this sub-regional dataset is denoted as , will the The multi-source fusion sensing fingerprint features are denoted as follows: The corresponding known real physical location coordinates are denoted as .

[0062] Constructing intelligent hyperparameter adaptive networks based on meta-learning The environmental inversion cognitive feature vector constructed in step S2 and the environmental change assessment score calculated in step S3 As network input, the four optimal balance hyperparameter sets for the objective function are adaptively and dynamically generated:

[0063] With offline source model parameters Initialize network parameters and target model parameters. The overall objective function for migration and reconstruction optimization is... Defined as:

[0064] in, The optimal balanced hyperparameters are dynamically generated by the meta-learning-based intelligent hyperparameter adaptive network and are used to control the weight coefficients of the localization supervision loss term, parameter migration constraint term, unsupervised domain alignment loss term, and structure preservation constraint term in the overall objective function.

[0065] The specific calculation methods for each loss item are as follows: Locating and monitoring loss items : In the retrieved offline regional source domain dataset The above calculation is used to determine the accuracy of the anchor foundation positioning solution:

[0066] in, This represents the target localization model mapping function to be optimized.

[0067] Parameter migration constraints Used to limit the abrupt forgetting of knowledge and ensure that the model inherits the basic localization ability of the source model.

[0068] Unsupervised domain alignment loss term A sample reliability weighting mechanism is introduced to minimize the maximum mean difference between the known source domain feature distribution and the online unlabeled target domain feature distribution in the latent space.

[0069] in, This represents the mapping function of the feature extraction layer within the localization model; For the first The reliability weight of each online target domain sample is based on the score of local environmental change of the current sample. Adaptive computation , The preset weight decay adjustment parameters actively suppress the influence of abnormal fingerprints caused by drastic environmental fluctuations on the alignment process.

[0070] Structural retention constraints Used to maintain the consistency of geometric topology among samples within the target domain:

[0071] in, Indicates the first term within the target domain The online sample and the first The adjacency topology weights between online samples are calculated based on the spatial distance of fingerprints.

[0072] Based on the environmental change level output in step S3, perform hierarchical parameter updates: When the environmental change level is slight, only the output layer parameters of the target fingerprint localization model are updated; When the environmental change level is moderate, freeze the parameters of the bottom-level perception feature extraction layer of the target fingerprint localization model and only update the parameters of the high-level mapping layer. When the environmental change level is significant, the target optimization function is updated with all parameters to obtain an adaptive target fingerprint localization model that fits the current dynamic propagation environment.

[0073] Step S5: High-precision online positioning The multi-source fusion sensing fingerprint features observed in real-time online observation are represented as... The initial location estimate is obtained by inputting the target fingerprint localization model reconstructed in step S4. :

[0074] in, This represents the optimal model parameter set of the optimal target fingerprint localization model obtained in step 4.

[0075] Based on the initial position estimate From regional source fingerprint datasets In the process, the spatial features closest to the nearest one are selected. A number of candidate reference points constitute a spatial local fingerprint database; the online real-time observed fingerprint is calculated and compared with the first... The Euclidean distance between the fingerprints of the candidate reference points in the reconstructed feature space:

[0076] in, Indicates the first The offline pre-stored multi-source fusion sensing fingerprint features corresponding to each candidate reference point This represents the mapping function of the intermediate feature extraction layer in the reconstructed model.

[0077] By combining a spatial local fingerprint database to perform fine-grained similarity correction, the final location estimation result of the output terminal is obtained. :

[0078] in, Indicating the first in the spatial local fingerprint database The physical location coordinates of the candidate reference points Indicates online real-time observation of fingerprints and the first The Euclidean distance between the fingerprints of each candidate reference point in the reconstructed feature space, and similarly... Indicates online real-time observation of fingerprints and the first The Euclidean distance between the fingerprints of candidate reference points in the reconstructed feature space.

[0079] Offline Phase: Construction of Regional Source Fingerprint Model Library During the offline phase, multiple reference points were deployed in indoor, outdoor, and transitional areas. The physical coordinates of each reference point were explicitly calibrated using measuring instruments. Real-time data acquisition of RSSI, CSI, CIR, TOA, and AOA from multiple base stations was also performed to establish a regional source fingerprint dataset. :

[0080] in, , indicating different region types; This represents the multi-source fusion sensing fingerprint feature of the i-th reference point. This represents its corresponding real physical location coordinates. This indicates the total number of reference points in the region.

[0081] Based on the regional source fingerprint dataset Train the corresponding region source models offline, and initialize the network parameters of each region source model. Minimize the mean square error of localization:

[0082] The trained indoor source fingerprint model, outdoor source fingerprint model, and transition region source fingerprint model are stored in the region fingerprint model library; simultaneously, for each region type... Based on the corresponding Extracted environmental characteristics, calculate and store their reference environmental statistical mean vectors. Covariance matrix and reference fingerprint correlation matrix It is used for model migration and reconstruction triggered by subsequent changes in the online environment.

[0083] Compared to the approach of rebuilding a complete fingerprint database and training the model as a whole, this embodiment, through the linkage of regional source model database and online environmental change assessment, can quickly complete seamless model migration and reconstruction with fewer unlabeled samples, greatly reducing system operation and maintenance costs and making it more suitable for high-precision real-time positioning application requirements in dynamic and complex 6G scenarios.

[0084] The performance verification experiment of the method of the present invention is as follows: To verify the effectiveness of the intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition proposed in this invention, a simulation experimental scenario was constructed under unified experimental conditions, and the method was compared and verified with existing mainstream fingerprint positioning methods.

[0085] I. Experimental Environment Setup This embodiment uses a MATLAB R2023b and Python 3.10 co-simulation platform to build a 6G multi-base station cooperative positioning environment. The experimental environment is as follows: CPU: Intel Core i7-12700K; GPU: NVIDIA RTX 4070; Memory: 32GB; Operating system: Windows 11.

[0086] The experimental area included three typical propagation scenarios: an indoor area, an outdoor area, and a transitional area between indoor and outdoor. The indoor area measured 20m × 12m, the outdoor area 25m × 20m, and the transitional area 6m × 4m. Three cooperative base stations were selected within the experimental area to participate in positioning observations. Each cooperative base station was able to acquire multi-source channel observation information, including RSSI, CSI, CIR, TOA, and AOA. Reference points were deployed in a uniform grid pattern, covering the main activity paths in the area. A total of 63 reference points were constructed: 30 in the indoor area, 25 in the outdoor area, and 8 in the transitional area. 50 frames of channel observation data were collected at each reference point.

[0087] In the offline phase, a multi-source fusion sensing fingerprint feature set is constructed according to step 1 of this invention, and an indoor source model, an outdoor source model, and a transition area source model are established according to step 6. The training set and the test set are randomly divided in a 7:3 ratio.

[0088] The following environmental changes will be simulated during the online phase: (1) Slightly changing scenarios: pedestrian movement, opening and closing of doors and windows; (2) Moderate change scenario: table and chair movement, change in the position of metal obstacles; (3) Scenarios with significant changes: the appearance of large obstructions, the passing of vehicles, and the switching between indoor and outdoor environments.

[0089] For each variation scenario, 300 sets of online test samples were collected to verify the localization performance and model reconstruction performance.

[0090] II. Comparison Methods To verify the effectiveness of the technical solution of this invention, the following five typical positioning methods were selected as comparison objects: (1) RSSI-KNN fingerprint localization method; (2) WKNN weighted k-nearest neighbor fingerprint localization method; (3) SVR support vector regression localization method; (4) DNN deep neural network localization method; (5) DA-FP domain adaptive migration fingerprint localization method; (6) The method of the present invention.

[0091] All methods used the same training dataset, test dataset, and base station deployment method. For methods involving neural network training, the same training conditions were used for model training and testing to ensure the fairness and comparability of experimental results.

[0092] III. Evaluation Indicators The average positioning error, root mean square error (RMSE), 90% positioning error quantile, online model reconstruction time, and number of online update samples were used as evaluation indicators.

[0093] IV. Positioning Performance Verification Table 1 Comparison of localization performance of different methods

[0094] As shown in Table 1, the method of this invention achieves the lowest mean positioning error and root mean square error. Compared to the RSSI-KNN method, the mean positioning error is reduced by approximately 61.32%. Compared to the DA-FP method, the mean positioning error is reduced by approximately 33.53%. This demonstrates that the propagation environment inversion cognitive mechanism proposed in this invention can effectively perceive changes in the propagation environment and improve positioning accuracy.

[0095] V. Verification of Positioning Performance in Different Areas Table 2 Comparison of average positioning errors in different regions

[0096] As shown in Table 2, the errors of all methods increase in the transition region due to the simultaneous occurrence of phenomena such as abrupt occlusion, enhanced reflection, and mixed propagation mechanisms. This invention identifies the region type through environmental inversion cognitive features and performs migration reconstruction by combining the region-related source model, maintaining high positioning accuracy even in the transition region, with an average positioning error of only 1.31m.

[0097] VI. Verification of Model Reconstruction Efficiency Table 3 Comparison of Model Reconstruction Performance

[0098] As shown in Table 3, this invention requires only 40 online samples to complete model reconstruction. Compared to the overall retraining method of DNN, the model update time is reduced by approximately 85.37%. Compared to the DA-FP transfer update method, the model update time is reduced by approximately 60.38%. This demonstrates that this invention can quickly complete model updates with fewer online samples.

[0099] VII. Ablation Verification of Key Modules of this Invention To further verify the contribution of each innovative module of the present invention to the improvement of localization performance, ablation experiments were conducted by removing key functional modules of the present invention while keeping the experimental environment, training dataset, test dataset and network structure completely consistent.

[0100] The following comparison scheme is specifically constructed: Solution A (Complete Model of the Invention): The invention employs all technical solutions, including an environmental inversion cognitive feature extraction module, an environmental change assessment module, and a meta-learning hyperparameter adaptive model reconstruction module.

[0101] Option B (Removing the environmental inversion cognitive module): Without extracting power fading dispersion, time delay spread factor, angle spread factor, reflection enhancement index, and diffraction sensitivity index, only the original RSSI, CSI, CIR, TOA, and AOA features are used as the localization input.

[0102] Option C (Removal of the Environmental Change Assessment Module): The environmental inversion cognitive feature extraction module is retained, but environmental change level assessment is not performed, and a fixed model update strategy is always used for model migration.

[0103] Option D (Removing the meta-learning intelligent hyperparameter adaptive module): The environmental inversion cognition module and the environmental change assessment module are retained, but fixed hyperparameters are used for model reconstruction instead of dynamic hyperparameter generation.

[0104] Option E (Basic Positioning Model): The basic fingerprint localization model uses only the original multi-source fingerprint features for localization, without employing environmental inversion cognition mechanisms, environmental change assessment mechanisms, or intelligent reconstruction mechanisms.

[0105] Table 4 Ablation Experiment Results of Key Modules of the Invention

[0106] As shown in Table 4, after removing the environmental inversion cognitive module, the average positioning error increased from 1.11m to 1.42m. This indicates that the environmental inversion cognitive feature can effectively characterize changes in the propagation environment and improve the adaptability of the positioning model to environmental disturbances.

[0107] After removing the environmental change assessment module, the average positioning error increased to 1.33m. This demonstrates that the environmental change level assessment mechanism proposed in this invention can accurately determine the degree of environmental change and guide the selection of subsequent model update strategies.

[0108] After removing the meta-learning hyperparameter adaptive module, the average localization error increased to 1.25m, and the model update time increased to 6.4s. This indicates that the meta-learning hyperparameter adaptive network can dynamically generate better transfer reconstruction parameters based on the current propagation environment, thereby improving model reconstruction efficiency and localization accuracy.

[0109] Furthermore, compared with the basic positioning model, the average positioning error of the complete model of this invention is reduced from 1.89m to 1.11m, and the model update time is reduced from 39.8s to 4.2s.

[0110] The above results show that the environmental inversion cognition module, environmental change assessment module, and meta-learning hyperparameter adaptive module proposed in this invention can effectively improve positioning performance and model reconstruction efficiency. The modules have good synergistic gain effects and jointly realize high-precision positioning and rapid model adaptive reconstruction in complex dynamic propagation environments.

[0111] VIII. Experimental Conclusions The experimental results above demonstrate that this invention achieves rapid migration and adaptive updating of the localization model in complex dynamic environments by constructing a propagation environment inversion cognitive feature vector, an environmental change assessment model, and an intelligent reconstruction mechanism based on meta-learning.

[0112] Compared with existing mainstream fingerprint positioning methods, the present invention demonstrates superior performance in terms of positioning accuracy, adaptability to environmental changes, and model reconstruction efficiency, verifying the effectiveness and practical value of the technical solution of the present invention.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent reconstruction of a 6G fingerprint positioning model based on propagation environment inversion cognition, characterized in that, Includes the following steps: Step 1: Real-time acquisition of multi-base station collaborative multi-source channel observation data reported by the terminal, and construction and output of a multi-source fusion sensing fingerprint feature set characterizing the radio wave propagation environment; Step 2: Using the multi-source fusion sensing fingerprint feature set as input, perform inversion cognitive analysis on the radio wave propagation environment around the terminal, extract environmental characterization quantities such as power fading dispersion, time delay spread factor, angle spread factor, reflection enhancement index and diffraction sensitivity index, construct environmental inversion cognitive feature vector, and output the identification results of the indoor area, outdoor area or transition area between indoor and outdoor where the terminal is located. Step 3: Using the environmental inversion cognitive feature vector and the region identification result as input, construct a differentiated environmental change assessment model for the corresponding region, calculate the change intensity index of the current environment relative to the offline pre-stored reference environment of the region, and output the environmental change level; Step 4: Using the region identification results and environmental change level as input, select the source model to be transferred for the corresponding region from the pre-established regional source fingerprint model library. Through the intelligent hyperparameter adaptive network based on meta-learning and dynamically generate the optimal balanced hyperparameter set of the objective function based on the cognitive features of the current environment, combine parameter transfer, domain feature alignment, sample weighting and structural constraints to intelligently reconstruct the fingerprint model and output the target fingerprint localization model adapted to the current environment. Step 5: Using the multi-source fusion sensing fingerprint features observed in real-time online as input, input the target fingerprint localization model to obtain an initial location estimate, perform similarity fine-tuning correction based on the initial location estimate, and output the final location estimate result of the terminal.

2. The intelligent reconstruction method for the 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 1, characterized in that, In step 2, the extracted environmental characterization quantities are combined with the average effective propagation path number to construct an environmental inversion cognitive feature vector. : in, For power fading dispersion, For delay spread factor, For the angle expansion factor, As an index for enhanced reflection, As a diffraction-sensitive indicator, This represents the average effective propagation path number.

3. The intelligent reconstruction method for the 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 2, characterized in that, In step 2, the cognitive feature vector is derived based on the environmental inversion. The region classifier identifies the region where the terminal is located. : in, Let be the posterior probability classification function for the region. Indicates indoor area, Indicates outdoor area, Indicates a transitional region.

4. The intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 1, characterized in that, In step 3, a dynamic environmental change assessment score is obtained by weighting the comprehensive environmental offset, fingerprint distribution drift, and base station coordination consistency index. : in, This is the environmental offset. For fingerprint distribution drift, This is a base station coordination consistency indicator. The preset empirical weighting coefficients for the corresponding regions, and satisfy the following conditions: .

5. The intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 4, characterized in that, In step 3, the threshold value is determined based on the offline historical environmental data. The environmental change level is divided into three levels: Slight changes: Satisfied Moderate change: Satisfied Significant changes: meet the requirements .

6. The intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 1, characterized in that, In step 4, the intelligent hyperparameter adaptive network based on meta-learning retrieves cognitive feature vectors from the environment. and environmental change assessment score As input, an adaptive and dynamic set of four optimal balanced hyperparameters for optimizing the objective function is generated: in, For meta-learning hyperparameter adaptive network mapping function, The optimal balanced hyperparameters are dynamically generated by the meta-learning-based intelligent hyperparameter adaptive network and are used to control the weight coefficients of the localization supervision loss term, parameter migration constraint term, unsupervised domain alignment loss term, and structure preservation constraint term in the overall objective function.

7. The intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 6, characterized in that, In step 4, the offline source model parameters are used as initialization parameters, and the overall objective optimization function is constructed based on the generated hyperparameter set. : in, The parameters of the target fingerprint localization model to be optimized are: To locate and monitor loss items, For parameter migration constraints, For unsupervised domain alignment loss term, Constraints are used to maintain the structure.

8. The intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 7, characterized in that, In step 4, a hierarchical parameter update strategy is implemented based on the level of environmental change: When the environmental change level is slight, only the output layer parameters of the target fingerprint localization model are updated; When the environmental change level is moderate, freeze the parameters of the bottom-level perception feature extraction layer of the target fingerprint localization model and only update the parameters of the high-level mapping layer. When the environmental change level is significant, the target fingerprint localization model is updated with all parameters.

9. The intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 1, characterized in that, In step 5, G candidate reference points with the closest spatial feature distances are selected based on the initial position estimate, and the final terminal position estimate is obtained by weighting the points by the inverse distance. : in, For the first The physical location coordinates of the candidate reference points For online fingerprint and the first The Euclidean distance between the fingerprints of candidate reference points in the reconstructed feature space.

10. The intelligent reconstruction method for 6G fingerprint positioning model based on propagation environment inversion cognition according to claim 1, characterized in that, The regional source fingerprint model library is pre-built in an offline phase, specifically including: Multiple reference points were deployed indoors, outdoors and in transitional areas and their physical coordinates were marked. Multi-source channel data from multiple base stations were collected to establish regional source fingerprint datasets. The corresponding regional source fingerprint models are trained offline based on the source fingerprint datasets of each region. At the same time, the reference environment statistical mean vector, covariance matrix and reference fingerprint correlation matrix of each region are calculated and stored.

Images