A three-dimensional positioning model training method and device, and related equipment

Abstract

The present disclosure provides a three-dimensional positioning model training method and device and related equipment. It obtains a source domain model of a source low-altitude network and a measurement report data set of a target low-altitude network; performs domain adaptation transfer learning on the source domain model according to each target domain three-dimensional wireless fingerprint vector to obtain a transfer adaptation model, and predicts the three-dimensional coordinates of the low-altitude terminal based on the target domain three-dimensional wireless fingerprint vector from the transfer adaptation model; calculates the three-dimensional geometric pseudo coordinates of the low-altitude terminal corresponding to the measurement report data according to the time advance parameter and the angle of arrival ranging parameter in the measurement report data, and performs parameter fine-tuning on the transfer adaptation model according to the difference between the three-dimensional coordinates predicted by the transfer adaptation model and the three-dimensional geometric pseudo coordinates to obtain a three-dimensional positioning model. The present disclosure effectively improves the positioning accuracy of the three-dimensional positioning model through cross-airspace model transfer learning combined with distribution alignment.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a training method, apparatus and related equipment for a three-dimensional positioning model. Background Technology

[0002] Currently, constructing a digital twin model of the wireless network is a key technical means in the analysis and optimization of the coverage characteristics and operational status of low-altitude wireless networks. This model needs to combine the wireless signal characteristics and spatial location information in the measurement data. Due to the significant three-dimensional spatial characteristics of low-altitude communication scenarios, namely large variations in terminal altitude and complex operating trajectories, obtaining minimal drive-test data containing accurate latitude, longitude, and altitude information is costly. Therefore, in practical applications, the amount of measurement report data is usually much larger than the minimal drive-test data.

[0003] To address the lack of location information in measurement report data, a fingerprint-based positioning method has been proposed. This method learns wireless signal propagation characteristics using minimized drive-test data from known locations within the same network domain. The latitude and longitude are then inferred by matching the signal characteristics of the measurement report data with the minimized drive-test signal fingerprints stored in the fingerprint database. However, this method heavily relies on the large amount of minimized drive-test data collected by the target low-altitude network itself. When minimized drive-test data is lacking in the target low-altitude network, it is difficult to construct a signal fingerprint database that covers the entire three-dimensional space and is sufficiently representative, leading to a significant decrease in the positioning accuracy of the low-altitude terminal or even its failure.

[0004] Therefore, there is an urgent need for a training method that can transfer the existing 3D fingerprint model of the source low-altitude network to the 3D positioning model of the target low-altitude network, so as to effectively improve the positioning accuracy of low-altitude terminals when the target low-altitude network lacks accurate location labeling data.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] This disclosure provides a training method, apparatus, and related equipment for a three-dimensional positioning model, which at least to some extent overcomes the problem that the positioning accuracy of low-altitude terminals is significantly reduced due to the inability of related technologies to effectively solve the problem of insufficient precise location labeling data in the target low-altitude network, i.e., the difficulty in constructing a signal fingerprint database that covers the complete three-dimensional space and has sufficient representativeness under the condition of minimizing the scarcity of road test data.

[0007] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0008] According to one aspect of this disclosure, a method for training a three-dimensional positioning model is provided. The method includes: acquiring a source domain model of a source low-altitude network and a measurement report dataset of a target low-altitude network, the measurement report dataset including multiple measurement report data; performing domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain to obtain a transfer adaptation model, and using the transfer adaptation model to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vectors of the target domain; wherein the three-dimensional wireless fingerprint vectors of the target domain are determined based on the measurement report dataset of the target low-altitude network, and the domain adaptation transfer learning is used to suppress the spatial propagation feature domain offset between the source low-altitude network and the target low-altitude network through multi-level distributed alignment constraints; calculating the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data based on the time lead parameter and the angle of arrival ranging parameter in the measurement report data, and fine-tuning the parameters of the transfer adaptation model based on the difference between the three-dimensional coordinates predicted by the transfer adaptation model and the three-dimensional geometric pseudo-coordinates to obtain a three-dimensional positioning model; the three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vectors of the target domain corresponding to the target measurement report data.

[0009] In some embodiments, obtaining the source domain model of the source low-altitude network includes: obtaining the source domain three-dimensional wireless fingerprint vector of the source low-altitude network; constructing a dual-channel joint network including a fingerprint matching channel and a regression prediction channel, and jointly training the dual-channel joint network based on each source domain three-dimensional wireless fingerprint vector and the actual three-dimensional coordinates corresponding to the source domain three-dimensional wireless fingerprint vector to obtain the source domain model, wherein the source domain three-dimensional wireless fingerprint vector is determined based on the minimized drive test dataset of the source low-altitude network.

[0010] In some embodiments, obtaining the source domain 3D wireless fingerprint vector of the source low-altitude network includes: obtaining a minimized drive test dataset of the source low-altitude network; the minimized drive test dataset includes multiple minimized drive test data and the actual 3D coordinates of the low-altitude terminal corresponding to each minimized drive test data; performing low-altitude propagation feature enhancement processing on each minimized drive test data based on the base station's 3D engineering parameters to construct the corresponding source domain 3D wireless fingerprint vector; determining the target domain 3D wireless fingerprint vector based on the target low-altitude network's measurement report dataset includes: performing the low-altitude propagation feature enhancement processing on each measurement report data based on the base station's 3D engineering parameters to construct the corresponding target domain 3D wireless fingerprint vector; the low-altitude propagation feature enhancement processing includes: altitude compensation processing, angle mapping correction, and vertical lobe gain correction; the altitude compensation processing is used to correct signal attenuation at different altitudes based on free space loss, the angle mapping correction is used to quantify the 3D propagation attenuation of low-altitude signals and correct antenna measurement angle deviation, and the vertical lobe gain correction is used to suppress vertical signal deviation through weighting and beamforming.

[0011] In some embodiments, the step of jointly training the dual-channel joint network based on each source domain 3D wireless fingerprint vector and the corresponding actual 3D coordinates of the source domain 3D wireless fingerprint vector to obtain a source domain model includes: constructing a first sample dataset based on each source domain 3D wireless fingerprint vector and the corresponding actual 3D coordinates of the source domain 3D wireless fingerprint vector; inputting the source domain 3D wireless fingerprint vectors in the first sample dataset into the dual-channel joint network, and using the fingerprint matching channel in the dual-channel joint network to extract features from the source domain 3D wireless fingerprint vectors to obtain source domain fingerprint embedding features; using the regression prediction channel in the dual-channel joint network to predict the 3D coordinates corresponding to the source domain fingerprint embedding features based on the source domain fingerprint embedding features; calculating a joint loss value based on the difference between the 3D coordinates corresponding to the source domain fingerprint embedding features and the actual 3D coordinates using a joint loss function; adjusting the parameters of the dual-channel joint network based on the joint loss value, and iteratively training the adjusted model using unused sample data from the first sample dataset until the training cutoff condition is met to obtain the source domain model.

[0012] In some embodiments, performing domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain to obtain a transfer adaptation model includes: constructing a target domain sample dataset based on multiple target domain three-dimensional wireless fingerprint vectors, and constructing a source domain sample dataset based on multiple source domain three-dimensional wireless fingerprint vectors; sequentially inputting sample data from the target domain sample dataset and the source domain sample dataset into the transfer adaptation model; using the fingerprint matching channel of the transfer adaptation model to extract features from the target domain three-dimensional wireless fingerprint vectors and the source domain three-dimensional wireless fingerprint vectors respectively, obtaining corresponding target domain fingerprint embedding features and source domain fingerprint embedding features, and extracting target domain samples from the target domain sample dataset. The height distribution, and the height distribution of source domain samples in the source domain sample dataset; based on the difference between the target domain fingerprint embedding features and the source domain fingerprint embedding features, determine the single-sample embedding feature distance loss, and use the maximum mean difference to measure the overall distribution difference between all target domain fingerprint embedding features and all source domain fingerprint embedding features to obtain the maximum mean difference global distribution loss; based on the height distribution of the target domain samples and the height distribution of the source domain samples, determine the relative entropy height distribution alignment loss, and perform the domain adaptation transfer learning on the source domain model based on the single-sample embedding feature distance loss, the maximum mean difference global distribution loss, and the relative entropy height distribution alignment loss to obtain the transfer adaptation model.

[0013] In some embodiments, the step of calculating the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data based on the time advance parameter and the angle of arrival ranging parameter in the measurement report data, and fine-tuning the parameters of the migration adaptation model based on the difference between the three-dimensional coordinates predicted by the migration adaptation model and the three-dimensional geometric pseudo-coordinates to obtain a three-dimensional positioning model, includes: acquiring the three-dimensional coordinates of the base station, and extracting the time advance parameter and the angle of arrival ranging parameter from each of the measurement report data respectively; for each measurement report data, calculating the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data through spatial triangulation based on the time advance parameter and the angle of arrival ranging parameter of the measurement report data, and the three-dimensional coordinates of the base station; calculating the corresponding three-dimensional calibration loss value based on the difference between each three-dimensional coordinate predicted by the migration adaptation model and the corresponding three-dimensional geometric pseudo-coordinate using a three-dimensional calibration loss function; and fine-tuning the parameters of the migration adaptation model based on the three-dimensional calibration loss value to obtain the three-dimensional positioning model.

[0014] In some embodiments, if minimal road test data exists in the target low-altitude network, the method further includes: acquiring the minimal road test data existing in the target low-altitude network and the actual three-dimensional coordinates of the low-altitude terminal corresponding to the minimal road test data; for each minimal road test data in the target low-altitude network, performing low-altitude propagation feature enhancement processing on the minimal road test data to construct a corresponding three-dimensional wireless fingerprint vector; constructing a second sample dataset based on the three-dimensional wireless fingerprint vector corresponding to each minimal road test data and the actual three-dimensional coordinates corresponding to the minimal road test data; sequentially inputting the data in the second sample dataset into the three-dimensional positioning model, and using the three-dimensional positioning model to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vector; calculating a strongly supervised regression loss value based on the difference between the three-dimensional coordinates of the low-altitude terminal predicted by the three-dimensional positioning model and the corresponding actual three-dimensional coordinates using a strongly supervised regression loss function; adjusting the parameters of the three-dimensional positioning model based on the strongly supervised regression loss value, and iteratively training the adjusted three-dimensional positioning model using unused sample data in the second sample dataset until the training cutoff condition is met to obtain the target three-dimensional positioning model.

[0015] According to another aspect of this disclosure, a training apparatus for a three-dimensional positioning model is also provided. The apparatus includes: an acquisition module for acquiring a source domain model of a source low-altitude network and a measurement report dataset of a target low-altitude network, the measurement report dataset including multiple measurement report data; and a transfer learning module for performing domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain to obtain a transfer adaptation model, and using the transfer adaptation model to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vectors of the target domain; wherein the three-dimensional wireless fingerprint vectors of the target domain are determined based on the measurement report dataset of the target low-altitude network, and the domain... The adaptive transfer learning is used to suppress the spatial propagation feature domain offset between the source low-altitude network and the target low-altitude network through multi-level distributed alignment constraints. The parameter fine-tuning module is used to calculate the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data based on the time advance parameter and the angle of arrival ranging parameter in the measurement report data, and to fine-tune the parameters of the transfer adaptation model based on the difference between the three-dimensional coordinates predicted by the transfer adaptation model and the three-dimensional geometric pseudo-coordinates to obtain a three-dimensional positioning model. The three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the target domain three-dimensional wireless fingerprint vector corresponding to the target measurement report data.

[0016] According to another aspect of this disclosure, an electronic device is also provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform a training method for a three-dimensional localization model as described above by executing the executable instructions.

[0017] According to another aspect of this disclosure, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the training method for the three-dimensional positioning model described in any of the preceding claims.

[0018] According to another aspect of this disclosure, a computer program product is also provided, comprising: a computer program or instructions, wherein the computer program or instructions, when executed by a processor, implement a training method for a three-dimensional positioning model that achieves any one of the above.

[0019] The training method for the 3D positioning model provided in the embodiments of this disclosure first obtains the source domain model of the source low-altitude network and the measurement report dataset of the target low-altitude network, the measurement report dataset including multiple measurement report data; then, domain adaptation transfer learning is performed on the source domain model based on the 3D wireless fingerprint vectors of each target domain, wherein the target domain 3D wireless fingerprint vectors are determined based on the measurement report dataset of the target low-altitude network, in order to suppress the spatial propagation feature domain offset between the source low-altitude network and the target low-altitude network through multi-level distributed alignment constraints, thereby obtaining a transfer adaptation model, and the transfer adaptation model predicts the 3D coordinates of the low-altitude terminal based on the target domain 3D wireless fingerprint vectors, so that the model can still maintain good performance even in the absence of target domain labels. It exhibits good generalization ability and positioning accuracy. Finally, based on the time lead parameters and arrival angle ranging parameters in the measurement report data, the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data are calculated. Based on the difference between the three-dimensional coordinates predicted by the transfer adaptation model and the three-dimensional geometric pseudo-coordinates, the parameters of the transfer adaptation model are fine-tuned to obtain the three-dimensional positioning model. This further corrects the prediction bias of the model in the target domain. Without the need for additional high-precision minimized road test data collection, the actual positioning accuracy and reliability of the final three-dimensional positioning model in the target low-altitude network are significantly improved. Thus, the three-dimensional positioning model can accurately predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vector of the target domain corresponding to the target measurement report data.

[0020] Furthermore, it significantly reduces the dependence of the target low-altitude network on high-precision minimized road test data and expensive annotation resources, effectively overcoming the problems of high data acquisition costs, weak cross-domain model generalization ability, and limited 3D positioning accuracy in existing low-altitude positioning technologies.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0023] Figure 1 A schematic diagram of the system architecture of a training method for a three-dimensional localization model according to an embodiment of the present disclosure is shown. Figure 2 A flowchart illustrating a training method for a three-dimensional localization model according to an embodiment of this disclosure is shown. Figure 3 A flowchart illustrating the training process of a source domain model according to an embodiment of this disclosure is provided. Figure 4 A flowchart illustrating the generation process of a migration adaptation model in an embodiment of this disclosure is shown. Figure 5 A flowchart illustrating the generation process of a three-dimensional positioning model in an embodiment of this disclosure is provided. Figure 6 A flowchart illustrating the generation process of another three-dimensional positioning model in an embodiment of this disclosure is provided. Figure 7 A schematic diagram of a training device for a three-dimensional positioning model is shown in an embodiment of the present disclosure; Figure 8 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0024] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0025] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0026] As mentioned earlier, to address the lack of location information in measurement report data, related technologies have proposed a fingerprint-based positioning method. This method involves learning wireless signal propagation characteristics using minimized drive-test data from known locations within the same network domain. The latitude and longitude are then inferred by matching the signal characteristics of the measurement report data with the minimized drive-test signal fingerprints stored in the fingerprint database. However, this method heavily relies on the large amount of minimized drive-test data collected by the target low-altitude network itself. When the target low-altitude network lacks or has only a small amount of minimized drive-test data, it is difficult to construct a signal fingerprint database that covers the complete three-dimensional space and is sufficiently representative. This leads to a significant decrease in the positioning accuracy of the low-altitude terminal, or even its failure.

[0027] Research has revealed that measurement report data in the target low-altitude network includes Time Advance (TA) and Angle of Arrival (AoA) measurements. Based on the signal propagation delay information provided by TA and the signal arrival direction information provided by AoA, a triangulation positioning method can be used to calculate the three-dimensional geometric pseudo-coordinates of the low-altitude terminal by combining the TA and AoA measurements from at least two base stations. In this disclosure, these three-dimensional geometric pseudo-coordinates serve as weakly supervised annotation information to fine-tune and calibrate the signal fingerprint map model migrated from the source low-altitude network. This enables effective adaptation and migration of the source domain-trained three-dimensional fingerprint positioning model to the target domain, even when the target low-altitude network lacks precise location annotation data (i.e., minimizing the scarcity of drive test data). This reduces the reliance on manually annotated data collection in the target domain and enhances the model's practicality and versatility in actual deployment scenarios.

[0028] In view of this, this disclosure provides a method for training a three-dimensional positioning model, comprising: acquiring a minimized road test dataset of a source low-altitude network and a measurement report dataset of a target low-altitude network; the minimized road test dataset includes multiple minimized road test data and the actual three-dimensional coordinates of the low-altitude terminal corresponding to each minimized road test data, and the measurement report dataset includes multiple measurement report data; constructing a dual-channel joint network including a fingerprint matching channel and a regression prediction channel, and jointly training the dual-channel joint network based on the three-dimensional wireless fingerprint vectors of each source domain and the actual three-dimensional coordinates corresponding to the three-dimensional wireless fingerprint vectors of the source domain to obtain a source domain model, wherein the source domain three-dimensional wireless fingerprint vectors are determined based on the minimized road test dataset of the source low-altitude network; and performing domain adaptation transfer learning on the source domain model according to the three-dimensional wireless fingerprint vectors of each target domain to obtain a model. A transfer adaptation model is developed, which predicts the 3D coordinates of the low-altitude terminal based on the target domain 3D wireless fingerprint vector. The target domain 3D wireless fingerprint vector is determined based on the measurement report dataset of the target low-altitude network. Domain adaptation transfer learning is used to suppress the spatial propagation feature domain offset between the source and target low-altitude networks through multi-level distributed alignment constraints. Based on the time lead parameter and arrival angle ranging parameter in the measurement report data, the 3D geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data are calculated. The parameters of the transfer adaptation model are fine-tuned based on the difference between the 3D coordinates predicted by the transfer adaptation model and the 3D geometric pseudo-coordinates, resulting in a 3D positioning model. This 3D positioning model is used to predict the 3D coordinates of the low-altitude terminal based on the target domain 3D wireless fingerprint vector corresponding to the target measurement report data.

[0029] It should be noted that the above embodiments are applicable to low-altitude line-of-sight (LOS) propagation environments. In open low-altitude scenarios, signal propagation between low-altitude terminals and base stations mainly occurs via line-of-sight paths. Therefore, the mapping relationship between signal propagation characteristics, such as path loss and angle of arrival, and the terminal's spatial location is highly deterministic and reproducible. When the network configurations of the source and target low-altitude networks are basically the same—for example, when parameters such as operating frequency band, base station antenna vertical lobe gain, and antenna elevation angle are the same or similar—the signal fingerprint distribution patterns in the two domains have an inherent similarity. This provides a theoretical basis for transferring the 3D fingerprint localization model trained in the source domain to the target domain. However, due to subtle differences in actual deployment environments, this 3D fingerprint localization model cannot be directly applied. It requires adaptive adjustments to the model parameters through fine-tuning and calibration to match the signal characteristic distribution of the target domain. Therefore, in the case of a lack of sufficiently minimized drive test data for the target low-altitude network in the low-line-of-sight propagation environment, the existing 3D fingerprint model in the source domain can be used to achieve effective migration with a few fine-tunings. This significantly reduces the dependence on data collection in the target domain, reduces data acquisition costs, and improves the practicality and generalization ability of the model in actual deployment.

[0030] To facilitate understanding, before introducing the embodiments of this disclosure, the following explanations are provided for several terms involved in the embodiments of this disclosure: Time Advance (TA) is a key parameter in mobile communication systems used to compensate for signal transmission delay. It synchronizes the base station's signal reception by adjusting the uplink transmission time of the terminal. Its value reflects the propagation delay between the terminal and the base station.

[0031] Angle of Arrival (AoA) is a positioning technology that uses the direction of arrival of signals to determine the location of a target. It utilizes the relative azimuth or angle information between the receiving node and the anchor node, combined with triangulation, to determine the target's location.

[0032] Measurement Report (MR): A measurement report is a set of measurement data actively reported by a terminal device (such as a mobile phone) to the base station. It reflects the quality of wireless signals, coverage, and interference. Measurement report data does not include location information such as latitude and longitude.

[0033] Minimization of Drive Tests (MDT) is an automated data acquisition technique defined in 3GPP standards, designed to reduce the costs of traditional manual drive tests. According to 3GPP standards (such as TS 37.320), MDT data supports the acquisition of three-dimensional positioning information, meaning that MDT data contains precise location information.

[0034] The specific implementation methods of the embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0035] Figure 1 A schematic diagram of an exemplary application system architecture is shown, illustrating a training method for a 3D localization model that can be applied according to embodiments of this disclosure. For example... Figure 1 As shown, the system architecture includes a terminal device 101, a network 102, and a network-side device 103; wherein, the network 102 is used as a medium to provide a communication link between the terminal device 101 and the network-side device 103, and can be a wired network or a wireless network.

[0036] Optionally, the aforementioned wireless or wired networks use standard communication technologies and / or protocols. The network is typically the Internet, but can also be any network, including but not limited to Local Area Networks (LANs), Metropolitan Area Networks (MANs), Wide Area Networks (WANs), mobile, wired or wireless networks, private networks, or any combination of virtual private networks. In some embodiments, technologies and / or formats, including Hyper Text Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network. Furthermore, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Networks (VPNs), and Internet Protocol Security (IPSec) can be used to encrypt all or some links. In other embodiments, custom and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies.

[0037] Optionally, the terminal device in this embodiment may also be referred to as UE (User Equipment). In specific implementation, the terminal device may be a mobile phone, a tablet personal computer, a laptop computer, a personal digital assistant (PDA), a mobile internet device (MID), a wearable device, or an in-vehicle device, etc. It should be noted that the specific type of terminal device is not limited in the embodiments of this invention.

[0038] Network-side equipment can be base stations, relays, or access points, etc. Base stations can be 5G and later versions of base stations (e.g., 5G NR NB), or base stations in other communication systems (e.g., eNB base stations). It should be noted that the specific type of network-side equipment is not limited in the embodiments disclosed herein.

[0039] Those skilled in the art will know that Figure 1The number of terminals, networks, and network-side devices shown is merely illustrative; any number of terminals, networks, and network-side devices can be included as needed. This disclosure does not limit the scope of the embodiments.

[0040] Under the above system architecture, this disclosure provides a method for training a three-dimensional positioning model, which can be executed by any electronic device with computing power.

[0041] In some embodiments, the training method for the 3D positioning model provided in this disclosure can be executed by a terminal device in the system architecture described above; in other embodiments, the training method for the 3D positioning model provided in this disclosure can be executed by a server in the system architecture described above; in still other embodiments, the training method for the 3D positioning model provided in this disclosure can be implemented by the terminal device and the server in the system architecture through interaction. For ease of description, the server will be referred to as the execution entity of the training method for the 3D positioning model provided in this disclosure.

[0042] Figure 2 A flowchart illustrating a training method for a three-dimensional localization model according to an embodiment of this disclosure is shown, such as... Figure 2 As shown, the training method for the three-dimensional localization model provided in this embodiment includes the following steps: S202, Obtain the source domain model of the source low-altitude network and the measurement report dataset of the target low-altitude network.

[0043] In S202, the source domain model of the source low-altitude network refers to the converged model obtained by jointly training a dual-channel joint network consisting of fingerprint matching channels and regression prediction channels, based on the source domain 3D wireless fingerprint vectors and their corresponding actual 3D coordinates contained in the minimized drive test dataset of the source low-altitude network. It possesses the ability to map from the source domain signal fingerprint space to the 3D geographic location space, serving as the foundation model for subsequent cross-domain transfer learning. The measurement report dataset includes multiple measurement report data. These data include, but are not limited to, one or more of the following wireless fingerprint features: Serving Cell RSRP (Reference Signal Received Power) / RSRQ (Reference Signal Received Quality), Co-frequency / Different Frequency Neighbor Cell PCI (Physical Cell Identification) / RSRP / RSRQ, and TA / AOA. These features are characterized by wide coverage, strong continuity, large data scale, statistical stability, and low positioning accuracy. It should be noted that the latitude and longitude of the terminal can be calculated based on the TA and AoA parameters, but the positioning accuracy of this method is low, and the generated tags belong to weakly supervised information.

[0044] In some embodiments, determining a target domain three-dimensional wireless fingerprint vector based on a target low-altitude network measurement report dataset includes: performing the low-altitude propagation feature enhancement processing on each measurement report data based on base station three-dimensional engineering parameters to construct a corresponding target domain three-dimensional wireless fingerprint vector. In this embodiment, the low-altitude propagation feature enhancement processing includes: altitude compensation processing, angle mapping correction, and vertical lobe gain correction; altitude compensation processing is used to correct signal attenuation at different altitudes based on free space loss; angle mapping correction is used to quantify the three-dimensional propagation attenuation of low-altitude signals and correct antenna measurement angle deviations; and vertical lobe gain correction is used to suppress vertical signal deviations through weighting and beamforming. The server can first perform data preprocessing, standardizing and normalizing the measured values ​​such as RSRP, RSRQ, SINR (Signal-to-Interference-plus-Noise Ratio), AoA, and TA in each measurement report data. Then, the server can perform low-altitude three-dimensional propagation compensation and enhancement processing on the normalized data. Through the aforementioned altitude compensation processing, angle mapping correction, vertical lobe gain correction, and pitch angle feature modeling, it can more accurately characterize the three-dimensional signal propagation characteristics in complex low-altitude environments, and construct a unified format three-dimensional wireless fingerprint vector with stronger spatial representation capabilities and robustness. This improves the accuracy of subsequent parameter fine-tuning, thereby enhancing the positioning precision of the 3D positioning model.

[0045] For example, the target domain three-dimensional wireless fingerprint vector enhanced with low-altitude propagation characteristics can be represented as:

[0046] It should be noted that the target domain 3D wireless fingerprint vector mentioned above is used as input to the transfer adaptation model, and its parameter types and numbers are only examples. In actual applications, the target domain 3D wireless fingerprint vector may include, but is not limited to, one or more of RSRP, RSRQ, AoA, SINR, TA, and altitude h. Furthermore, in the context of low-altitude network 3D positioning, pitch angle feature modeling involves incorporating the vertical angle between the base station antenna and the terminal (such as a drone) as an independent, computable wireless fingerprint feature into the input vector of the positioning model, and establishing its mathematical relationship with signal propagation attenuation and altitude changes.

[0047] In some embodiments, obtaining the source domain model of the source low-altitude network includes: obtaining the source domain three-dimensional wireless fingerprint vector of the source low-altitude network; constructing a dual-channel joint network including a fingerprint matching channel and a regression prediction channel, and jointly training the dual-channel joint network based on each source domain three-dimensional wireless fingerprint vector and the actual three-dimensional coordinates corresponding to the source domain three-dimensional wireless fingerprint vector to obtain the source domain model, wherein the source domain three-dimensional wireless fingerprint vector is determined based on the minimized drive test dataset of the source low-altitude network.

[0048] In this embodiment, the source domain 3D wireless fingerprint vector refers to a 3D feature vector that characterizes the wireless propagation environment at a specific point in low-altitude space, collected and constructed based on a minimized drive test dataset within the source low-altitude network coverage area. Specifically, the server can first construct a dual-channel joint network including a fingerprint matching channel and a regression prediction channel; then, the server can jointly train the dual-channel joint network based on each source domain 3D wireless fingerprint vector and the corresponding actual 3D coordinates. This allows the server to simultaneously utilize the discrete features of fingerprint matching and the continuous features of regression prediction, fully mining the spatial location information in the source domain data to obtain a high-precision source domain model, significantly improving the feature extraction capability and positioning accuracy of the source domain model on known data.

[0049] Specifically, see Figure 3 The figure illustrates a flowchart of a training process for a source domain model according to an embodiment of the present disclosure, which includes the following steps: S302, construct the first sample dataset based on the three-dimensional wireless fingerprint vectors of each source domain and the actual three-dimensional coordinates corresponding to the three-dimensional wireless fingerprint vectors of that source domain.

[0050] In this embodiment of the disclosure, the server can be based on the three-dimensional wireless fingerprint vectors of each source domain. The actual three-dimensional coordinates corresponding to the three-dimensional wireless fingerprint vector of the source domain Establish corresponding mapping relationships Then, the first sample dataset is constructed based on this mapping relationship.

[0051]

[0052] in, It is the x-axis coordinate of sample number s. It is the y-axis coordinate of sample number s. The z-axis coordinate.

[0053] S304, input the source domain 3D wireless fingerprint vector from the first sample dataset into the dual-channel joint network, and use the fingerprint matching channel in the dual-channel joint network to extract features from the source domain 3D wireless fingerprint vector to obtain the source domain fingerprint embedding features.

[0054] In this embodiment of the disclosure, the server can sequentially input the three-dimensional wireless fingerprint vectors of each source domain into the dual-channel joint network, and then use the fingerprint matching channel in the dual-channel joint network to achieve geometric clustering of similar samples in the wireless fingerprint space through metric learning. Thus, the fingerprint matching channel can accurately capture the features in the three-dimensional wireless fingerprint vectors of the source domain and obtain the source domain fingerprint embedding features.

[0055] S306 utilizes the regression prediction channel in the dual-channel joint network to predict the three-dimensional coordinates corresponding to the source domain fingerprint embedding feature based on the source domain fingerprint embedding feature.

[0056] In this embodiment of the disclosure, the server can directly predict the three-dimensional coordinates of the low-altitude terminal by using the regression prediction channel in the dual-channel joint network based on the source domain fingerprint embedding features of the neural network.

[0057] S308 calculates the joint loss value based on the difference between the three-dimensional coordinates corresponding to the source domain fingerprint embedding features and the actual three-dimensional coordinates using the joint loss function.

[0058] In this embodiment of the disclosure, the server can calculate the joint loss value L based on the difference between the three-dimensional coordinates corresponding to the source domain fingerprint embedding features and the actual three-dimensional coordinates using the joint loss function shown in formula (1).

[0059]

[0060] in, Corresponding to the fingerprint matching channel, The metric loss (i.e. triplet loss) is used to ensure that fingerprints with similar propagation features are clustered more tightly, thereby improving localization stability and making the output coordinates approximate the real label to ensure 3D localization accuracy. Corresponding to the regression prediction channel, The loss refers to the three-dimensional coordinate regression loss, which is the Euclidean distance between the true value (i.e., the actual three-dimensional coordinates) and the predicted value (i.e., the predicted three-dimensional coordinates). , This is a weighted hyperparameter used to balance fingerprint clustering accuracy and 3D localization accuracy.

[0061] S310: Adjust the parameters of the dual-channel joint network based on the joint loss value, and iteratively train the adjusted model using unused sample data from the first sample dataset until the training cutoff condition is met, thus obtaining the source domain model.

[0062] In this embodiment of the disclosure, the server can adjust the parameters of the dual-channel joint network based on the joint loss value, and iteratively train the adjusted model using unused sample data from the first sample dataset until the training cutoff condition is met, thereby obtaining the source domain model. .

[0063]

[0064] in, Source domain fingerprint embedding features, corresponding to the output of the fingerprint matching channel, are used to map the input fingerprint vector to the embedding space to measure the similarity between fingerprints; This is a 3D coordinate regression model, corresponding to the regression prediction channel, used to learn the general mapping law from low-altitude wireless fingerprint embedding features to 3D spatial coordinates. In the joint loss function... The item is reflected in the Optimization constraints The item is reflected in the The optimization constraints are such that the two achieve joint optimization of feature discriminativity and localization accuracy through shared embedding representations. It should be noted that... and This is only used to illustrate the optimization objectives of each channel in the joint loss function, to reflect the overall framework of joint training. It is not used as an adjustable parameter in subsequent model forward inference and transfer fine-tuning, and there is no issue of arbitrary weight configuration. After training, the source domain model... The source low-altitude network, whose coordinates are known to correspond to the fingerprint vectors, has converged to a stable state and serves as the initial model for subsequent cross-domain transfer learning.

[0065] It should be noted that by utilizing the discrete features of fingerprint matching and the continuous features of regression prediction, the spatial location information in the source domain data can be fully extracted to obtain a high-precision source domain model, thereby significantly improving the feature extraction capability and positioning accuracy of the source domain model on known data.

[0066] In some embodiments, obtaining the source domain three-dimensional wireless fingerprint vector of the source low-altitude network includes: obtaining a minimized drive test dataset of the source low-altitude network; the minimized drive test dataset includes multiple minimized drive test data and the actual three-dimensional coordinates of the low-altitude terminal corresponding to each minimized drive test data; and performing low-altitude propagation feature enhancement processing on each minimized drive test data based on the base station's three-dimensional engineering parameters to construct the corresponding source domain three-dimensional wireless fingerprint vector.

[0067] The minimized drive test dataset includes multiple minimized drive test data sets and the actual three-dimensional coordinates of the low-altitude terminals corresponding to each minimized drive test data set. Specifically, the minimized drive test data may include, but is not limited to, one or more of the following wireless fingerprint features: serving cell RSRP / RSRQ, co-frequency / different-frequency neighbor cell PCI / RSRP / RSRQ, and latitude and longitude location information. It has the characteristics of high data positioning accuracy, strong ability to reflect real user experience, limited data scale, uneven coverage, and high collection cost and complexity.

[0068] It is understandable that the process of enhancing the low-altitude propagation characteristics of each minimized road test data can be similar to the process of enhancing the low-altitude propagation characteristics of each measurement report data described above, and will not be repeated here. Specifically, the server can first perform unified data preprocessing, standardizing and normalizing the RSRP, RSRQ, SINR (Signal-to-Interference-plus-Noise Ratio), horizontal AoA, pitch AoA, and TA measurement values ​​in each measurement report data and each minimized road test data; then the server can perform low-altitude three-dimensional propagation compensation enhancement processing on the normalized data, so as to more accurately characterize the three-dimensional signal propagation characteristics in the complex low-altitude environment through the above-mentioned height compensation processing, angle mapping correction, vertical lobe gain correction, and pitch angle feature modeling, and construct a unified format three-dimensional wireless fingerprint vector x with stronger spatial representation capabilities and robustness, laying the data foundation for subsequent high-precision positioning. Represents the target's three-dimensional wireless fingerprint vector; This represents the source domain three-dimensional wireless fingerprint vector.

[0069] S204, perform domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain to obtain the transfer adaptation model, and use the transfer adaptation model to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vectors of the target domain.

[0070] In S204, domain adaptation transfer learning is used to suppress the spatial propagation feature domain shift between the source low-altitude network and the target low-altitude network through multi-level distribution alignment constraints.

[0071] In this embodiment of the disclosure, the server can perform domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain. This is done by suppressing the spatial propagation feature domain offset between the source low-altitude network and the target low-altitude network through multi-level distributed alignment constraints, thereby obtaining a transfer adaptation model. The transfer adaptation model then predicts the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vectors of the target domain, enabling the model to maintain good generalization ability and positioning accuracy even in the absence of target domain labels.

[0072] It should be noted that after the source domain model is trained, the feature encoder corresponding to the fingerprint matching channel in the source domain model... A stable source domain fingerprint embedding mapping relationship has been learned, which can be regarded as the source domain model of the source low-altitude network. Based on this, the server will transmit the measurement report data of the target low-altitude network. Inputting this feature encoder yields the target domain fingerprint embedding features. To suppress the spatial propagation feature shift between the source and target domains, the server minimizes a multi-level distribution alignment objective function, selects the optimal sample subset for the source domain model from the existing target domain measurement report dataset, and performs transfer adaptation on the source domain model based on this subset.

[0073] Specifically, see Figure 4 The figure illustrates a flowchart of the generation process of a migration adaptation model according to an embodiment of the present disclosure, which includes the following steps: S402, construct a target domain sample dataset based on multiple target domain 3D wireless fingerprint vectors, and construct a source domain sample dataset based on multiple source domain 3D wireless fingerprint vectors.

[0074] S404, input the sample data from the target domain sample dataset and the source domain sample dataset into the transfer adaptation model in sequence, and use the fingerprint matching channel of the transfer adaptation model to extract features from the target domain 3D wireless fingerprint vector and the source domain 3D wireless fingerprint vector respectively, to obtain the corresponding target domain fingerprint embedding features and source domain fingerprint embedding features, and extract the height distribution of the target domain samples in the target domain sample dataset and the height distribution of the source domain samples in the source domain sample dataset.

[0075] S406. Based on the difference between the fingerprint embedding features of the target domain and the fingerprint embedding features of the source domain, the single-sample embedding feature distance loss is determined, and the maximum mean difference (MMD) is used to measure the overall distribution difference between all fingerprint embedding features of the target domain and all fingerprint embedding features of the source domain, thus obtaining the MMD global distribution loss.

[0076] In this embodiment of the disclosure, the server can use formula (2) based on the fingerprint embedding features of the target domain. Source domain fingerprint embedding features The difference between them determines the single-sample embedding feature distance loss. By minimizing the embedding feature distance, the embedding features of the target task are brought closer to the feature space of the source task, thereby achieving wireless feature domain alignment. Then, the server can use the maximum mean discrepancy (MMD) (full name: Maximum Mean Discrepancy) to measure the overall distribution difference between all target domain fingerprint embedding features and all source domain fingerprint embedding features using formula (3) to obtain the MMD global distribution loss. By minimizing the MMD global distribution loss, the feature distribution of the target domain is aligned with the feature distribution of the source domain at the mean level, that is, the wireless propagation feature distributions of the two domains are as similar as possible, thereby eliminating the domain offset.

[0077]

[0078] in, The source domain three-dimensional wireless fingerprint vector, The target domain is a three-dimensional wireless fingerprint vector.

[0079] S408. Based on the height distribution of the target domain samples and the height distribution of the source domain samples, the relative entropy height distribution alignment loss is determined. Then, based on the single sample embedding feature distance loss, the maximum mean difference global distribution loss, and the relative entropy height distribution alignment loss, domain adaptation transfer learning is performed on the source domain model to obtain the transfer adaptation model.

[0080] In this embodiment of the disclosure, the server can use formula (4) to calculate the height distribution h of the target domain samples. t and the height distribution h of the source domain samples s Determine the KL (Kullback-Leibler Divergence, referring to relative entropy) height distribution alignment loss for the source task hidden layer h. s and the target task hidden layer h t The probability distribution is aligned using KL algorithm to maintain the consistency of the feature space structure and constrain the uncertainty distribution, thus avoiding the problem of distortion of the localization pattern caused by the distribution alignment. Then, the server can perform domain adaptation transfer learning on the source domain model based on single sample embedding feature distance loss, MMD global distribution loss and KL height distribution alignment loss using formula (5). This allows the fingerprint distributions of different heights / different regions to be aligned, so that the model can be transferred to the target low-altitude domain and a transfer-adapted model can be obtained.

[0081]

[0082] It should be noted that, after the above multi-level distribution alignment optimization, the optimal sample subset selected from the target domain measurement report data corresponds to the target domain fingerprint embedding features. It can be mapped to the low-altitude propagation feature space learned by the source task and substituted into the source domain model. Forward reasoning is performed to initially predict the 3D coordinates of the target domain samples using the existing 3D localization capabilities of the source task, achieving effective transfer of localization knowledge from the source domain to the target domain. At this point, the measurement report data from the target low-altitude network has been integrated into the source network model through embedded spatial mapping, providing an initial state for subsequent fine-tuning of model parameters for the target domain. This allows the fine-tuning process to quickly adapt to the signal distribution characteristics of the target domain, based on existing knowledge in the source domain and utilizing a small amount of weakly supervised information from the target domain, further improving the model's localization accuracy in the target low-altitude network. This embodiment, through cross-scene transfer learning with multi-source statistical distribution alignment, effectively achieves efficient cross-domain knowledge transfer during the transfer from the source domain to the target domain by implementing three alignment objectives: distribution consistency, feature structure consistency, and the introduction of high-dimensional geographical regularity consistency.

[0083] S206. Based on the time lead parameter and arrival angle ranging parameter in the measurement report data, calculate the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data, and fine-tune the parameters of the transfer adaptation model according to the difference between the three-dimensional coordinates predicted by the transfer adaptation model and the three-dimensional geometric pseudo-coordinates to obtain the three-dimensional positioning model.

[0084] In S206, the three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the target domain three-dimensional wireless fingerprint vector corresponding to the target measurement report data.

[0085] In this embodiment, the server can calculate the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data based on the time advance parameter TA and the angle of arrival ranging parameter AOA in the measurement report data. Then, based on the difference between the three-dimensional coordinates predicted by the transfer adaptation model and the three-dimensional geometric pseudo-coordinates, the parameters of the transfer adaptation model are fine-tuned to obtain a three-dimensional positioning model. This further corrects the prediction bias of the model in the target domain. Without the need for additional high-precision minimized road test data collection, the value of the existing measurement report data is maximized, significantly improving the actual positioning accuracy and reliability of the final three-dimensional positioning model in the target low-altitude network. Thus, the three-dimensional positioning model can be used to determine the target domain three-dimensional wireless fingerprint vector corresponding to the target measurement report data. Accurately predict the three-dimensional coordinates of low-altitude terminals .

[0086]

[0087] in, These are the final coordinates of the MR data output from the fine-tuned, transferred model.

[0088] Specifically, see Figure 5The figure illustrates a flowchart of a three-dimensional positioning model generation process according to an embodiment of the present disclosure, which includes the following steps: S502 acquires the three-dimensional coordinates of the base station and extracts the time advance parameter and the arrival angle ranging parameter from each measurement report data.

[0089] S504, for each measurement report data, based on the time advance parameter and arrival angle ranging parameter of the measurement report data, as well as the three-dimensional coordinates of the base station, calculates the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data through spatial triangulation.

[0090] In this embodiment of the disclosure, for each measurement report data, the server can calculate the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data through spatial triangulation based on the time advance parameter and angle of arrival ranging parameter of the measurement report data, as well as the three-dimensional coordinates of the base station. This allows for model fine-tuning through the three-dimensional geometric pseudo-coordinates.

[0091]

[0092] S506 calculates the corresponding 3D calibration loss value based on the difference between each 3D coordinate predicted by the transfer adaptation model and the corresponding 3D geometric pseudo-coordinates using the 3D calibration loss function.

[0093] In this embodiment of the disclosure, the server can calculate the corresponding three-dimensional calibration loss value based on the difference between each three-dimensional coordinate predicted by the transfer adaptation model and the corresponding three-dimensional geometric pseudo-coordinates using the three-dimensional calibration loss function shown in formula (6). This is used to measure the consistency between the 3D coordinates predicted by the transfer adaptation model and the geometric constraints (i.e., 3D geometric pseudo-coordinates).

[0094]

[0095] in, The target domain three-dimensional wireless fingerprint vector; The predicted 3D coordinates for the transfer adaptation model; These are three-dimensional geometric pseudo-coordinates.

[0096] In some embodiments, the target domain three-dimensional wireless fingerprint vector may be target domain measurement report data that has been optimized and filtered through the aforementioned domain adaptation transfer learning. The corresponding fingerprint embedding features have been mapped to the low-altitude propagation feature space learned by the source task. The server will then filter the... Input transfer adaptation model Perform forward inference to obtain the model's predicted 3D coordinates; simultaneously, using the same... The associated TA and AoA measurements are used to calculate the three-dimensional geometric pseudo-coordinates. As a weakly supervised geometric constraint, the three-dimensional calibration loss function defined in Equation (6) measures the Euclidean distance between the model's predicted coordinates and the geometric pseudo-coordinates, and is used to measure the prediction bias of the transfer adaptation model in the target domain. This loss function serves as the optimization objective in the subsequent parameter fine-tuning stage, by minimizing... This allows the transfer adaptation model to calibrate its parameters using weakly supervised geometric information, even when there are no precise location labels in the target domain. This aligns the general mapping rules learned from the source domain with the actual propagation characteristics of the target domain, thereby effectively improving the model's 3D positioning accuracy in the target low-altitude network.

[0097] S508, based on the 3D calibration loss value, fine-tunes the parameters of the transfer adaptation model to obtain the 3D positioning model.

[0098] In this embodiment of the disclosure, the server can fine-tune the parameters of the transfer adaptation model based on the three-dimensional calibration loss value, so as to perform small-step, lightweight online fine-tuning and updating of the parameters of the transfer adaptation model, and obtain the calibrated model (i.e., the three-dimensional positioning model) as shown below.

[0099]

[0100] in, As mentioned above η is the online calibration learning rate. A partial parameter freezing strategy is adopted to update only the parameters related to the three-dimensional coordinate regression, so as to reduce computational complexity and ensure system stability. This represents the gradient direction of the calibration loss with respect to the model parameters. By differentiating the parameters, we can identify the parameters that cause the prediction to deviate from geometric consistency, and update them along the gradient direction to make the predicted position satisfy the geometric consistency constraint.

[0101] It should be noted that by using the TA and AoA in the target network measurement report data to construct an approximate and physically consistent 3D geometric position estimate (i.e., 3D geometric pseudo-coordinates), and using it as a weakly supervised reference signal to calibrate and fine-tune the transferred model, 3D scene adaptation across low-altitude networks can be achieved, thereby effectively eliminating the systematic offset of the model and improving the positioning accuracy in the target low-altitude network.

[0102] It should be noted that if the target low-altitude network lacks minimal road test data and location labels, regression training cannot be performed, meaning it does not involve... At this point, relying solely on distribution alignment Φ( This enables model transfer, and the aligned model is directly used as the final 3D positioning model.

[0103] In one feasible implementation, if there is a small amount of minimized drive test data containing high-confidence 3D locations in the target low-altitude network, the server uses this data to perform regression enhancement training on the transferred model to further improve the positioning accuracy and obtain the final positioning model (i.e., the target 3D positioning model).

[0104] See Figure 6 The figure illustrates a flowchart of another three-dimensional positioning model generation process in an embodiment of this disclosure, which includes the following steps: S602, acquire the minimum road test data existing in the target low-altitude network and the actual three-dimensional coordinates of the low-altitude terminal corresponding to the minimum road test data.

[0105] S604: For each minimized drive test data in the target low-altitude network, perform low-altitude propagation feature enhancement processing on the minimized drive test data to construct the corresponding three-dimensional wireless fingerprint vector.

[0106] S606. Based on the three-dimensional wireless fingerprint vector corresponding to each minimized road test data and the actual three-dimensional coordinates corresponding to the minimized road test data, a second sample dataset is constructed.

[0107] In this embodiment of the disclosure, the server can take the three-dimensional wireless fingerprint vector corresponding to each minimized road test data and the actual three-dimensional coordinates corresponding to the minimized road test data as a strongly supervised sample pair, and then construct a second sample dataset based on all strongly supervised sample pairs.

[0108] S608, the data in the second sample dataset are sequentially input into the three-dimensional positioning model, and the three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vector.

[0109] S610 calculates the strongly supervised regression loss value based on the difference between the three-dimensional coordinates of the low-altitude terminal predicted by the three-dimensional positioning model and the corresponding actual three-dimensional coordinates using the strongly supervised regression loss function.

[0110] In this embodiment of the disclosure, the server can calculate the strongly supervised regression loss value L based on the difference between the three-dimensional coordinates of the low-altitude terminal predicted by the three-dimensional positioning model and the corresponding actual three-dimensional coordinates using the strongly supervised regression loss function shown in formula (7). MDT .

[0111]

[0112] in, The three-dimensional coordinates of the low-altitude terminal predicted by the three-dimensional positioning model; These are the actual three-dimensional coordinates.

[0113] S612, adjust the parameters of the 3D localization model based on the strongly supervised regression loss value, and iteratively train the adjusted 3D localization model using unused sample data in the second sample dataset until the training cutoff condition is met, thus obtaining the target 3D localization model.

[0114] In this embodiment of the disclosure, the server can base the loss value L on the strongly supervised regression value. MDT The parameters of the 3D localization model are adjusted, and the adjusted 3D localization model is iteratively trained using unused sample data from the second sample dataset (this iterative training is updated in a limited number of rounds) until the training cutoff condition is met, thus obtaining the enhanced model. f s calib,0 (i.e., the target's three-dimensional positioning model).

[0115]

[0116] Furthermore, the server can input the target domain 3D wireless fingerprint vector corresponding to any measurement report data into the target 3D positioning model, and the target 3D positioning model can directly output the 3D coordinates of the low-altitude terminal corresponding to the measurement report data. .

[0117]

[0118] It should be noted that by introducing a small amount of road test data minimizing the target domain for regression enhancement training, the model can be finely calibrated using high-confidence real location labels, effectively eliminating the systematic bias and nonlinear distortion remaining from cross-domain migration. This process transforms the model's generalization ability into absolute positioning accuracy in the target domain, achieving a performance leap from distribution alignment to precise positioning with extremely low data cost, significantly enhancing the model's robustness and practicality in target low-altitude networks.

[0119] It should be noted that the acquisition, storage, use, and processing of data in this disclosed technical solution comply with the relevant provisions of national laws and regulations. The various types of data, such as personal identity data, operational data, and behavioral data related to individuals, customers, and groups, obtained in the embodiments of this disclosure have all been authorized.

[0120] Based on the same inventive concept, this disclosure also provides a training device for a three-dimensional positioning model, as shown in the following embodiment. Since the principle by which this device embodiment solves the problem is similar to that of the above-described method embodiment, the implementation of this device embodiment can refer to the implementation of the above-described method embodiment, and repeated details will not be elaborated further.

[0121] Figure 7 This diagram illustrates a training device for a three-dimensional positioning model according to an embodiment of the present disclosure, such as... Figure 7As shown, the device includes: an acquisition module 701, a transfer learning module 702, and a parameter fine-tuning module 703; The acquisition module 701 is used to acquire the source domain model of the source low-altitude network and the measurement report dataset of the target low-altitude network, wherein the measurement report dataset includes multiple measurement report data. The transfer learning module 702 is used to perform domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain to obtain a transfer adaptation model, and the transfer adaptation model predicts the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vectors of the target domain; wherein, the three-dimensional wireless fingerprint vectors of the target domain are determined based on the measurement report dataset of the target low-altitude network, and the domain adaptation transfer learning is used to suppress the spatial propagation feature domain offset between the source low-altitude network and the target low-altitude network through multi-level distribution alignment constraints; The parameter fine-tuning module 703 is used to calculate the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data based on the time advance parameter and the arrival angle ranging parameter in the measurement report data, and to fine-tune the parameters of the transfer adaptation model based on the difference between the three-dimensional coordinates predicted by the transfer adaptation model and the three-dimensional geometric pseudo-coordinates to obtain the three-dimensional positioning model; the three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the target domain three-dimensional wireless fingerprint vector corresponding to the target measurement report data.

[0122] Optionally, the acquisition module includes: an acquisition subunit, used to acquire the minimized road test dataset of the source low-altitude network and the measurement report dataset of the target low-altitude network; the minimized road test dataset includes multiple minimized road test data and the actual three-dimensional coordinates of the low-altitude terminal corresponding to each minimized road test data, and the measurement report dataset includes multiple measurement report data; a joint training module, used to construct a dual-channel joint network including a fingerprint matching channel and a regression prediction channel, and to jointly train the dual-channel joint network based on the three-dimensional wireless fingerprint vectors of each source domain and the actual three-dimensional coordinates corresponding to the three-dimensional wireless fingerprint vectors of the source domain to obtain the source domain model, wherein the three-dimensional wireless fingerprint vectors of the source domain are determined based on the minimized road test dataset of the source low-altitude network.

[0123] A sub-unit is used to acquire a minimized drive test dataset of the source low-altitude network. The minimized drive test dataset includes multiple minimized drive test data and the actual three-dimensional coordinates of the low-altitude terminal corresponding to each minimized drive test data. Based on the three-dimensional engineering parameters of the base station, low-altitude propagation feature enhancement processing is performed on each minimized drive test data to construct the corresponding source domain three-dimensional wireless fingerprint vector. Based on the three-dimensional engineering parameters of the base station, the low-altitude propagation feature enhancement processing is performed on each measurement report data to construct the corresponding target domain three-dimensional wireless fingerprint vector. The low-altitude propagation feature enhancement processing includes: altitude compensation processing, angle mapping correction, and vertical beam gain correction. Altitude compensation processing is used to correct signal attenuation at different altitudes based on free space loss. Angle mapping correction is used to quantify the three-dimensional propagation attenuation of low-altitude signals and correct antenna measurement angle deviation. Vertical beam gain correction is used to suppress vertical signal deviation through weighting and beamforming.

[0124] Optionally, the joint training module includes: The first construction unit is used to construct the first sample dataset based on the three-dimensional wireless fingerprint vectors of each source domain and the actual three-dimensional coordinates corresponding to the three-dimensional wireless fingerprint vectors of that source domain. The first feature extraction unit is used to input the source domain three-dimensional wireless fingerprint vector in the first sample dataset into the dual-channel joint network, and use the fingerprint matching channel in the dual-channel joint network to extract features from the source domain three-dimensional wireless fingerprint vector to obtain the source domain fingerprint embedding features. The first prediction unit is used to predict the three-dimensional coordinates corresponding to the source domain fingerprint embedding feature based on the source domain fingerprint embedding feature using the regression prediction channel in the dual-channel joint network. The first calculation unit is used to calculate the joint loss value based on the difference between the three-dimensional coordinates corresponding to the source domain fingerprint embedding features and the actual three-dimensional coordinates through the joint loss function; The iterative training unit is used to adjust the parameters of the dual-channel joint network based on the joint loss value, and iteratively train the adjusted model using unused sample data in the first sample dataset until the training cutoff condition is met to obtain the source domain model.

[0125] Optionally, the transfer learning module includes: The second building unit is used to build a target domain sample dataset based on multiple target domain three-dimensional wireless fingerprint vectors, and to build a source domain sample dataset based on multiple source domain three-dimensional wireless fingerprint vectors. The second feature extraction unit is used to input the sample data from the target domain sample dataset and the source domain sample dataset into the transfer adaptation model in sequence, and use the fingerprint matching channel of the transfer adaptation model to extract features from the target domain 3D wireless fingerprint vector and the source domain 3D wireless fingerprint vector respectively, to obtain the corresponding target domain fingerprint embedding features and source domain fingerprint embedding features, and extract the height distribution of the target domain samples in the target domain sample dataset and the height distribution of the source domain samples in the source domain sample dataset. The second calculation unit is used to determine the single-sample embedding feature distance loss based on the difference between the target domain fingerprint embedding features and the source domain fingerprint embedding features, and to use the maximum mean difference to measure the overall distribution difference between all target domain fingerprint embedding features and all source domain fingerprint embedding features, so as to obtain the maximum mean difference global distribution loss. The domain adaptation transfer learning unit is used to determine the relative entropy height distribution alignment loss based on the height distribution of the target domain samples and the height distribution of the source domain samples, and to perform domain adaptation transfer learning on the source domain model based on the single sample embedding feature distance loss, the maximum mean difference global distribution loss, and the relative entropy height distribution alignment loss to obtain the transfer adaptation model.

[0126] Optionally, the parameter fine-tuning module includes: The data acquisition unit is used to acquire the three-dimensional coordinates of the base station and extract the time advance parameter and the angle of arrival ranging parameter from each measurement report data. The third calculation unit is used to calculate the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to each measurement report data by spatial triangulation, based on the time advance parameter and arrival angle ranging parameter of the measurement report data, as well as the three-dimensional coordinates of the base station. The fourth calculation unit is used to calculate the corresponding three-dimensional calibration loss value based on the difference between each three-dimensional coordinate predicted by the transfer adaptation model and the corresponding three-dimensional geometric pseudo-coordinates using the three-dimensional calibration loss function. The parameter fine-tuning unit is used to fine-tune the parameters of the transfer adaptation model based on the 3D calibration loss value to obtain the 3D localization model.

[0127] Optionally, the training device for the 3D localization model also includes: The data acquisition module is used to acquire the minimum road test data existing in the target low-altitude network and the actual three-dimensional coordinates of the low-altitude terminal corresponding to the minimum road test data when there is a small amount of minimum road test data in the target low-altitude network. The processing module is used to perform low-altitude propagation feature enhancement processing on each minimized drive test data in the target low-altitude network and construct the corresponding three-dimensional wireless fingerprint vector. The sample construction module is used to construct a second sample dataset based on the three-dimensional wireless fingerprint vector corresponding to each minimized road test data and the actual three-dimensional coordinates corresponding to that minimized road test data. The prediction module is used to input the data from the second sample dataset into the three-dimensional positioning model in sequence, and use the three-dimensional positioning model to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vector. The calculation module is used to calculate the strongly supervised regression loss value based on the difference between the three-dimensional coordinates of the low-altitude terminal predicted by the three-dimensional positioning model and the corresponding actual three-dimensional coordinates, using the strongly supervised regression loss function. The iterative training module is used to adjust the parameters of the 3D localization model based on the strongly supervised regression loss value, and iteratively train the adjusted 3D localization model using unused sample data in the second sample dataset until the training cutoff condition is met, thereby obtaining the target 3D localization model.

[0128] It should be noted that the examples and application scenarios implemented by the modules in the above device embodiments and the corresponding steps in the method embodiments are the same, but are not limited to the content disclosed in the above method embodiments. It should also be noted that the above modules, as part of the device, can be executed in a computer system such as a set of computer-executable instructions.

[0129] Those skilled in the art will understand that various aspects of this disclosure can be implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, which can be collectively referred to herein as a "circuit", "module" or "system".

[0130] Based on the same inventive concept, this disclosure also provides an electronic device, which includes: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the training method of the three-dimensional localization model described above by executing the executable instructions. Since the principle by which this electronic device solves the problem is similar to that of the above method embodiments, the implementation of this electronic device embodiment can refer to the implementation of the above method embodiments, and repeated details will not be described again.

[0131] The following reference Figure 8 To describe an electronic device 800 according to such an embodiment of the present disclosure. Figure 8 The electronic device 800 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0132] like Figure 8As shown, the electronic device 800 is manifested in the form of a general-purpose computing device. The components of the electronic device 800 may include, but are not limited to: at least one processing unit 810, at least one storage unit 820, and a bus 830 connecting different system components (including storage unit 820 and processing unit 810).

[0133] The storage unit stores program code that can be executed by the processing unit 810, causing the processing unit 810 to perform the steps described in the "Exemplary Methods" section above according to various exemplary embodiments of this disclosure. For example, the processing unit 810 can execute the following steps of the above method embodiment: acquiring a source domain model of the source low-altitude network and a measurement report dataset of the target low-altitude network, the measurement report dataset including multiple measurement report data; performing domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vector of each target domain to obtain a transfer adaptation model, and using the transfer adaptation model to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vector of the target domain; wherein, the three-dimensional wireless fingerprint vector of the target domain is determined based on the measurement report dataset of the target low-altitude network, and the domain adaptation transfer learning is used to suppress the spatial propagation feature domain offset between the source low-altitude network and the target low-altitude network through multi-level distribution alignment constraints; calculating the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data based on the time lead parameter and the angle of arrival ranging parameter in the measurement report data, and fine-tuning the parameters of the transfer adaptation model based on the difference between the three-dimensional coordinates predicted by the transfer adaptation model and the three-dimensional geometric pseudo-coordinates to obtain a three-dimensional positioning model; the three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vector of the target domain corresponding to the target measurement report data.

[0134] Storage unit 820 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 8201 and / or cache memory 8202, and may further include a read-only memory (ROM) 8203.

[0135] The storage unit 820 may also include a program / utility 8204 having a set (at least one) of program modules 8205, such program modules 8205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0136] Bus 830 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0137] Electronic device 800 can also communicate with one or more external devices 840 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 800, and / or with any device that enables electronic device 800 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 850. Furthermore, electronic device 800 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 860. As shown, network adapter 860 communicates with other modules of electronic device 800 via bus 830. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 800, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0138] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0139] Based on the same inventive concept, this disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a training method for a 3D positioning model as described above. Since the principle by which this computer-readable storage medium solves the problem is similar to that of the above-described method embodiments, the implementation of this computer-readable storage medium embodiment can refer to the implementation of the above-described method embodiments, and repeated details will not be elaborated further.

[0140] More specific examples of computer-readable storage media in this disclosure may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0141] In this disclosure, a computer-readable storage medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting a program for use by or in connection with an instruction execution system, apparatus, or device.

[0142] Optionally, the program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0143] In practical implementation, program code for performing the operations of this disclosure can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0144] Based on the same inventive concept, this disclosure also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements a training method for a three-dimensional positioning model according to any one of the above method embodiments. Since the principle by which this computer program product embodiment solves the problem is similar to that of the above method embodiments, the implementation of this computer program product embodiment can refer to the implementation of the above method embodiments, and repeated details will not be elaborated further.

[0145] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0146] Furthermore, although the steps of the method in this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0147] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0148] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A training method for a three-dimensional positioning model, characterized in that, include: Obtain the source domain model of the source low-altitude network and the measurement report dataset of the target low-altitude network, wherein the measurement report dataset includes multiple measurement report data; Domain adaptation transfer learning is performed on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain to obtain a transfer adaptation model. The three-dimensional coordinates of the low-altitude terminal are then predicted by the transfer adaptation model based on the three-dimensional wireless fingerprint vectors of the target domains. The three-dimensional wireless fingerprint vectors of the target domains are determined based on the measurement report dataset of the target low-altitude network. The domain adaptation transfer learning is used to suppress the spatial propagation feature domain offset between the source low-altitude network and the target low-altitude network through multi-level distribution alignment constraints. Based on the time advance parameter and arrival angle ranging parameter in the measurement report data, the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data are calculated. Based on the difference between the three-dimensional coordinates predicted by the migration adaptation model and the three-dimensional geometric pseudo-coordinates, the parameters of the migration adaptation model are fine-tuned to obtain a three-dimensional positioning model. The three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the target domain three-dimensional wireless fingerprint vector corresponding to the target measurement report data.

2. The training method according to claim 1, characterized in that, The acquisition of the source domain model of the source low-altitude network includes: Obtain the source domain three-dimensional wireless fingerprint vector of the source low-altitude network; A dual-channel joint network containing a fingerprint matching channel and a regression prediction channel is constructed. Based on the three-dimensional wireless fingerprint vectors of each source domain and the actual three-dimensional coordinates corresponding to the three-dimensional wireless fingerprint vectors of the source domain, the dual-channel joint network is jointly trained to obtain the source domain model. The source domain three-dimensional wireless fingerprint vectors are determined based on the minimized drive test dataset of the source low-altitude network.

3. The training method according to claim 2, characterized in that, The acquisition of the source domain three-dimensional wireless fingerprint vector of the source low-altitude network includes: Obtain the minimum road test dataset of the source low-altitude network; the minimum road test dataset includes multiple minimum road test data and the actual three-dimensional coordinates of the low-altitude terminal corresponding to each minimum road test data; Based on the three-dimensional engineering parameters of the base station, the low-altitude propagation feature enhancement processing of each minimized drive test data is performed to construct the corresponding source domain three-dimensional wireless fingerprint vector. The determination of the target domain three-dimensional wireless fingerprint vector based on the measurement report dataset of the target low-altitude network includes: Based on the three-dimensional engineering parameters of the base station, the low-altitude propagation feature enhancement processing is performed on each measurement report data to construct the corresponding target domain three-dimensional wireless fingerprint vector. The low-altitude propagation feature enhancement processing includes: altitude compensation processing, angle mapping correction, and vertical lobe gain correction; the altitude compensation processing is used to correct signal attenuation at different altitudes based on free space loss, the angle mapping correction is used to quantify the three-dimensional propagation attenuation of low-altitude signals and correct antenna measurement angle deviation, and the vertical lobe gain correction is used to suppress vertical signal deviation through weighting and beamforming.

4. The training method according to claim 2, characterized in that, The dual-channel joint network is jointly trained based on the three-dimensional wireless fingerprint vectors of each source domain and the actual three-dimensional coordinates corresponding to those source domain three-dimensional wireless fingerprint vectors to obtain the source domain model, including: Based on each of the source domain three-dimensional wireless fingerprint vectors and the actual three-dimensional coordinates corresponding to the source domain three-dimensional wireless fingerprint vectors, a first sample dataset is constructed. The source domain 3D wireless fingerprint vector in the first sample dataset is input into the dual-channel joint network, and the fingerprint matching channel in the dual-channel joint network is used to extract features from the source domain 3D wireless fingerprint vector to obtain source domain fingerprint embedding features. Using the regression prediction channel in the dual-channel joint network, the three-dimensional coordinates corresponding to the source domain fingerprint embedding feature are predicted based on the source domain fingerprint embedding feature. The joint loss value is calculated based on the difference between the three-dimensional coordinates corresponding to the source domain fingerprint embedding features and the actual three-dimensional coordinates using the joint loss function. The parameters of the dual-channel joint network are adjusted based on the joint loss value, and the adjusted model is iteratively trained using unused sample data from the first sample dataset until the training cutoff condition is met, thereby obtaining the source domain model.

5. The training method according to claim 2, characterized in that, The step of performing domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain to obtain a transfer adaptation model includes: A target domain sample dataset is constructed based on multiple target domain three-dimensional wireless fingerprint vectors, and a source domain sample dataset is constructed based on multiple source domain three-dimensional wireless fingerprint vectors. The sample data from the target domain sample dataset and the source domain sample dataset are sequentially input into the transfer adaptation model. The fingerprint matching channel of the transfer adaptation model is used to extract features from the target domain 3D wireless fingerprint vector and the source domain 3D wireless fingerprint vector respectively, to obtain the corresponding target domain fingerprint embedding features and source domain fingerprint embedding features. The height distribution of the target domain samples in the target domain sample dataset and the height distribution of the source domain samples in the source domain sample dataset are also extracted. Based on the difference between the target domain fingerprint embedding features and the source domain fingerprint embedding features, the single-sample embedding feature distance loss is determined, and the maximum mean difference is used to measure the overall distribution difference between all the target domain fingerprint embedding features and all the source domain fingerprint embedding features, thus obtaining the maximum mean difference global distribution loss. Based on the height distribution of the target domain samples and the height distribution of the source domain samples, the relative entropy height distribution alignment loss is determined, and the domain adaptation transfer learning is performed on the source domain model based on the single sample embedding feature distance loss, the maximum mean difference global distribution loss, and the relative entropy height distribution alignment loss to obtain the transfer adaptation model.

6. The training method according to claim 1, characterized in that, Based on the time lead parameter and arrival angle ranging parameter in the measurement report data, the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data are calculated. Then, based on the difference between the three-dimensional coordinates predicted by the migration adaptation model and the three-dimensional geometric pseudo-coordinates, the parameters of the migration adaptation model are fine-tuned to obtain a three-dimensional positioning model, including: Obtain the three-dimensional coordinates of the base station, and extract the time advance parameter and the angle of arrival ranging parameter from each of the measurement report data; For each measurement report data, based on the time advance parameter of the measurement report data, the arrival angle ranging parameter, and the three-dimensional coordinates of the base station, the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data are calculated by spatial triangulation. The corresponding three-dimensional calibration loss value is calculated based on the difference between each three-dimensional coordinate predicted by the transfer adaptation model and the corresponding three-dimensional geometric pseudo-coordinate using the three-dimensional calibration loss function. The parameters of the migration adaptation model are fine-tuned based on the three-dimensional calibration loss value to obtain the three-dimensional localization model.

7. The training method according to claim 1, characterized in that, If minimal drive test data exists in the target low-altitude network, the method further includes: Obtain the minimized road test data existing in the target low-altitude network and the actual three-dimensional coordinates of the low-altitude terminal corresponding to the minimized road test data; For each minimized drive test data in the target low-altitude network, low-altitude propagation feature enhancement processing is performed on the minimized drive test data to construct a corresponding three-dimensional wireless fingerprint vector; A second sample dataset is constructed based on the three-dimensional wireless fingerprint vector corresponding to each of the minimized road test data and the actual three-dimensional coordinates corresponding to the minimized road test data. The data in the second sample dataset are sequentially input into the three-dimensional positioning model, and the three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vector; The strongly supervised regression loss value is calculated based on the difference between the three-dimensional coordinates of the low-altitude terminal predicted by the three-dimensional positioning model and the corresponding actual three-dimensional coordinates using the strongly supervised regression loss function. The parameters of the 3D localization model are adjusted based on the strongly supervised regression loss value, and the adjusted 3D localization model is iteratively trained using unused sample data from the second sample dataset until the training cutoff condition is met, thereby obtaining the target 3D localization model.

8. A training device for a three-dimensional positioning model, characterized in that, The device includes: The acquisition module is used to acquire the source domain model of the source low-altitude network and the measurement report dataset of the target low-altitude network, wherein the measurement report dataset includes multiple measurement report data. The transfer learning module is used to perform domain adaptation transfer learning on the source domain model based on the three-dimensional wireless fingerprint vectors of each target domain to obtain a transfer adaptation model, and the transfer adaptation model predicts the three-dimensional coordinates of the low-altitude terminal based on the three-dimensional wireless fingerprint vectors of the target domain; wherein, the three-dimensional wireless fingerprint vectors of the target domain are determined based on the measurement report dataset of the target low-altitude network, and the domain adaptation transfer learning is used to suppress the spatial propagation feature domain offset between the source low-altitude network and the target low-altitude network through multi-level distribution alignment constraints; The parameter fine-tuning module is used to calculate the three-dimensional geometric pseudo-coordinates of the low-altitude terminal corresponding to the measurement report data based on the time advance parameter and the arrival angle ranging parameter in the measurement report data, and to fine-tune the parameters of the migration adaptation model based on the difference between the three-dimensional coordinates predicted by the migration adaptation model and the three-dimensional geometric pseudo-coordinates to obtain a three-dimensional positioning model; the three-dimensional positioning model is used to predict the three-dimensional coordinates of the low-altitude terminal based on the target domain three-dimensional wireless fingerprint vector corresponding to the target measurement report data.

9. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the training method for the three-dimensional localization model according to any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method for the three-dimensional positioning model according to any one of claims 1 to 7.

Images