Wireless positioning model training, wireless positioning methods, apparatuses, and computing devices

Abstract

This application relates to a wireless positioning model training, wireless positioning method, apparatus, and computing device. The method includes: acquiring wireless measurement data of a target network area and environmental semantic data characterizing the spatial or functional attributes of the target network area, as well as location annotation information corresponding to the wireless measurement data; inputting the wireless measurement data into a wireless positioning model to be trained based on the location annotation information, environmental semantic data, and a set of semantic tags corresponding to the wireless measurement data; obtaining location prediction results and semantic prediction results corresponding to the wireless measurement data through the wireless positioning model; and training the wireless positioning model based on the location annotation information, the set of semantic tags, and the location and semantic prediction results to obtain a trained wireless positioning model. By introducing semantic information for joint model training, the positioning accuracy of the trained model is improved, thereby enhancing the positioning precision of wireless positioning.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a wireless positioning model training, wireless positioning method, apparatus, computing device, computer-readable storage medium, and computer program product. Background Technology

[0002] With the continuous evolution of wireless communication systems, positioning technology based on wireless measurement data has gradually become one of the important basic capabilities of network intelligence. Existing wireless positioning schemes typically take reference signal received power, reference signal received quality, channel quality indication, timing advance, neighbor cell information, uplink reference signal measurement results or channel characteristics as inputs, and output the target terminal's location coordinates, grid position or relative position results through machine learning models or deep learning models.

[0003] However, current wireless positioning methods suffer from insufficient positioning accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide a wireless positioning model training, wireless positioning method, device, computing equipment, computer-readable storage medium, and computer program product that can improve positioning accuracy in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a method for training a wireless positioning model, including:

[0006] Acquire wireless measurement data and environmental semantic data of the target network area, as well as the location annotation information corresponding to the wireless measurement data; the environmental semantic data is used to characterize the spatial or functional attributes of the target network area.

[0007] Based on location annotation information and environmental semantic data, construct a set of semantic tags corresponding to wireless measurement data;

[0008] The wireless measurement data is input into the wireless positioning model to be trained, and the location prediction result and semantic prediction result corresponding to the wireless measurement data are obtained through the wireless positioning model.

[0009] Based on location annotation information and semantic label set, as well as location prediction results and semantic prediction results, the wireless positioning model is trained to obtain the trained wireless positioning model.

[0010] In conjunction with the first aspect, in one embodiment, the environmental semantic data includes spatial element information, a plan view, and spatial management rules. Based on the location annotation information and the environmental semantic data, a semantic tag set corresponding to the wireless measurement data is constructed, including:

[0011] Spatial geometric calculations are performed based on location labeling information and spatial element information to generate spatial semantic labels corresponding to wireless measurement data.

[0012] The location annotation information is matched with the planar schematic diagram to generate functional semantic tags corresponding to the wireless measurement data;

[0013] Generate custom semantic tags corresponding to wireless measurement data based on space management rules;

[0014] By combining spatial semantic labels, functional semantic labels, and custom semantic labels, a set of semantic labels corresponding to wireless measurement data is obtained.

[0015] In conjunction with the first aspect, in one embodiment, wireless measurement data is input into a wireless positioning model to be trained, and the location prediction result and semantic prediction result corresponding to the wireless measurement data are obtained through the wireless positioning model, including:

[0016] Perform data preprocessing on the wireless measurement data to obtain the processed wireless measurement data;

[0017] The processed wireless measurement data is input into the wireless positioning model, and the processed wireless measurement data is encoded by the input encoding layer in the wireless positioning model to obtain the unified feature representation corresponding to the wireless measurement data.

[0018] The unified feature representation is input into the shared representation layer in the wireless positioning model, and the location semantic fusion feature is extracted based on the unified feature representation to obtain the high-level fusion feature corresponding to the unified feature representation.

[0019] The high-level fusion features are input into the multi-task output layer of the wireless positioning model. The location prediction result corresponding to the wireless measurement data is obtained through the location output head of the multi-task output layer. The main semantic prediction result and the fine-grained semantic prediction result corresponding to the wireless measurement data are obtained through the main semantic output head and the fine-grained semantic output head of the multi-task output layer, respectively.

[0020] In conjunction with the first aspect, in an exemplary embodiment, a wireless positioning model is trained based on location annotation information and a set of semantic tags, as well as location prediction results and semantic prediction results, to obtain a trained wireless positioning model, including:

[0021] Based on the location annotation information and location prediction results, construct the location loss of the wireless positioning model;

[0022] Based on the main semantic prediction results, fine-grained semantic prediction results, and semantic tag set, the semantic loss of the wireless positioning model is constructed.

[0023] The wireless positioning model is trained based on location loss and semantic loss to obtain the trained wireless positioning model.

[0024] In conjunction with the first aspect, in one embodiment, based on the semantic prediction results and the semantic label set, a semantic loss for the wireless positioning model is constructed, including:

[0025] Based on the main semantic prediction results and the semantic label set, construct the main semantic loss of the wireless positioning model;

[0026] When a coarse-grained label in the semantic label set is the same as any preset coarse-grained label, a fine-grained semantic loss for the wireless positioning model is constructed based on the fine-grained semantic prediction result and the semantic label set. The main semantic loss and the fine-grained semantic loss are combined to obtain the semantic loss of the wireless positioning model.

[0027] When the coarse-grained labels in the semantic label set are different from any preset coarse-grained label, the main semantic loss is used as the semantic loss of the wireless localization model.

[0028] In conjunction with the first aspect, in one embodiment, acquiring wireless measurement data of a target network area includes:

[0029] Acquire terminal-side measurement data and network-side measurement data within the target network area;

[0030] The terminal-side measurement data and the network-side measurement data are fused to obtain fused measurement data.

[0031] Based on terminal-side measurement data, network-side measurement data, and fused measurement data, the raw wireless measurement data of the target network area is obtained.

[0032] Secondly, this application also provides a wireless positioning method, including:

[0033] Acquire wireless measurement data of the network area to be located;

[0034] The wireless measurement data to be located is input into the trained wireless positioning model, and the wireless positioning model outputs the location prediction result and semantic prediction result of the network area to be located; the wireless positioning model is trained by any of the wireless positioning model training methods in the first aspect.

[0035] In conjunction with the first aspect, in one embodiment, it further includes:

[0036] Perform confidence evaluation on the semantic prediction results, and generate additional semantic features based on the obtained semantic confidence and semantic prediction results;

[0037] The semantic augmented features and the wireless measurement data to be located are input into a pre-built semantic augmented feature enhancement model, and the optimized location prediction result of the network area to be located is obtained through the semantic augmented feature enhancement model.

[0038] In conjunction with the first aspect, in one embodiment, it further includes:

[0039] Obtain the predicted semantic distribution of the semantic prediction results of the network region to be located, and the training semantic distribution of the semantic prediction results output by the wireless positioning model during the training phase.

[0040] If the offset between the predicted semantic distribution and the training semantic distribution is greater than a preset offset threshold, an early warning signal is generated indicating that the model performance has degraded or the scene has changed.

[0041] Based on the warning signal, the wireless positioning model is triggered to undergo model retraining, incremental learning, or fine-tuning of model parameters.

[0042] Thirdly, this application also provides a wireless positioning model training device, comprising:

[0043] The first acquisition module is used to acquire wireless measurement data and environmental semantic data of the target network area, as well as the location annotation information corresponding to the wireless measurement data; the environmental semantic data is used to characterize the spatial attributes or functional attributes of the target network area.

[0044] The building module is used to construct a set of semantic tags corresponding to wireless measurement data based on location annotation information and environmental semantic data;

[0045] The input module is used to input wireless measurement data into the wireless positioning model to be trained, and to obtain the location prediction result and semantic prediction result corresponding to the wireless measurement data through the wireless positioning model.

[0046] The training module is used to train the wireless positioning model based on location annotation information, semantic label set, location prediction results and semantic prediction results, so as to obtain the trained wireless positioning model.

[0047] Fourthly, this application also provides a wireless positioning device, comprising:

[0048] The second acquisition module is used to acquire wireless measurement data of the network area to be located.

[0049] The positioning module is used to input the wireless measurement data to be positioned into the trained wireless positioning model, and output the location prediction result and semantic prediction result of the network area to be positioned through the wireless positioning model; the wireless positioning model is trained by any of the wireless positioning model training methods in the first aspect.

[0050] Fifthly, this application also provides a computing device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of either the first or second aspect.

[0051] Sixthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method of either the first or second aspect.

[0052] In a seventh aspect, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method of either the first or second aspect.

[0053] The aforementioned wireless positioning model training method, wireless positioning method, device, computing device, computer-readable storage medium, and computer program product acquire wireless measurement data of the target network area and environmental semantic data characterizing the spatial or functional attributes of the target network area, as well as location annotation information corresponding to the wireless measurement data. Based on the location annotation information, environmental semantic data, and the semantic tag set corresponding to the wireless measurement data, the wireless measurement data is input into the wireless positioning model to be trained. The wireless positioning model obtains the location prediction result and semantic prediction result corresponding to the wireless measurement data. Based on the location annotation information, semantic tag set, and location and semantic prediction results, the wireless positioning model is trained to obtain a trained wireless positioning model. The wireless measurement data of the network area to be located is then input into the trained wireless positioning model to obtain the location prediction result and semantic prediction result for the network area to be located. By introducing environmental semantic data of the target network area to generate a corresponding semantic tag set and integrating semantic information into the training process of the wireless positioning model, the adaptability of the trained wireless positioning model to the actual environment is ensured, thereby improving the positioning accuracy of the trained wireless positioning model and achieving an improvement in wireless positioning precision. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a diagram illustrating the application environment of a wireless positioning model training method in one embodiment.

[0056] Figure 2 This is a flowchart illustrating a wireless positioning model training method in one embodiment;

[0057] Figure 3 This is a model architecture diagram of a wireless positioning model in one embodiment;

[0058] Figure 4 This is a flowchart illustrating a wireless positioning method in one embodiment;

[0059] Figure 5 This is a structural block diagram of a wireless positioning model training device in one embodiment;

[0060] Figure 6 This is a structural block diagram of a wireless positioning device in one embodiment;

[0061] Figure 7 This is a diagram of the internal structure of a computing device in one embodiment. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] The wireless positioning model training method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, various sample terminals within the target network area communicate with the computing device 102 via a network. A data storage system can store the data that the computing device 102 needs to process. The data storage system can be integrated into the computing device 102 or placed in the cloud or on other network servers. The computing device 102 acquires wireless measurement data of the target network area and environmental semantic data characterizing the spatial or functional attributes of the target network area, as well as location labeling information corresponding to the wireless measurement data. Based on the location labeling information and environmental semantic data, it constructs a semantic label set corresponding to the wireless measurement data. The wireless measurement data is input into the wireless positioning model to be trained. The wireless positioning model obtains the location prediction result and semantic prediction result corresponding to the wireless measurement data. Based on the location labeling information, the semantic label set, and the location prediction result and semantic prediction result, the wireless positioning model is trained to obtain the trained wireless positioning model. The sample terminals can be, but are not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. The head-mounted device can be a virtual reality (VR) device, an augmented reality (AR) device, smart glasses, etc. The computing device 102 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. In a preferred deployment, the computing device 102 can also be a smart computing board deployed on the base station side (such as an AI acceleration board mounted on the baseband processing unit (BBU), an embedded AI processing unit within the base station, or an edge computing node integrated within the base station. Deploying this method on a base station smart computing board allows for localized inference using real-time wireless measurement data acquired from the base station side, thereby reducing location service latency and improving network intelligence efficiency.

[0064] A semantic tag set can be understood as a combination of tags with semantic meaning. It can include spatial semantic tags, functional semantic tags, and custom semantic tags. Spatial semantic tags are used to characterize the physical space type where the sample is located. Physical space types include at least one of roads, buildings, squares, parks, parking lots, and underground spaces. Functional semantic tags are used to characterize the business function attributes of the area where the sample is located. Business function attributes include at least one of toll areas, waiting areas, security check areas, boarding areas, shop areas, passageways, and restricted areas. Custom semantic tags are used to characterize area attributes defined according to manual rules, operational rules, or business rules. It can also include coarse-grained semantic tags and fine-grained semantic tags. Coarse-grained semantic tags are used to characterize the major category of the scenario to which the sample belongs. Fine-grained semantic tags are used to characterize the sub-area category under the corresponding major category of the scenario. Fine-grained semantic tags include at least one of road segment numbers, floor numbers, functional area numbers, or other local area identifiers.

[0065] In one exemplary embodiment, such as Figure 2 As shown, a wireless positioning model training method is provided, which can be applied to... Figure 1 Taking computing device 102 as an example, the explanation includes the following steps S201 to S204. Wherein:

[0066] Step S201: Obtain wireless measurement data and environmental semantic data of the target network area, as well as location labeling information corresponding to the wireless measurement data; the environmental semantic data is used to characterize the spatial or functional attributes of the target network area.

[0067] Wireless measurement data can be understood as measurement data collected and transmitted through wireless communication technologies (such as Wi-Fi, Bluetooth, LoRa, NB-IoT, ZigBee, 5G, etc.). Environmental semantic data can be understood as data used to characterize the spatial or functional attributes of the location area, which may include GIS (Geographic Information System) vector data, local map data, regional division data, and management rule data. Location annotation information can be understood as data used to characterize the known location of the sample, which may include two-dimensional coordinates, three-dimensional coordinates, relative base station offset, raster number, and latitude and longitude. Environmental semantic data can be obtained from map data, network planning data, building floor plans, three-dimensional city models, point of interest data, location management systems, or manually annotated data. The environmental semantic data includes at least one of road centerlines, building outlines, floor boundaries, indoor and outdoor area boundaries, functional area divisions, service area boundaries, and point of interest categories. When constructing training samples, the environmental semantic data is converted to a unified coordinate reference system based on the coordinate reference system used for location annotation information, and the corresponding semantic label set is determined according to the inclusion relationship, distance relationship, or proximity relationship between the sample location coordinates and spatial elements.

[0068] For example, the computing device 102 acquires terminal-side measurement data of sample terminals in the target network area and network-side measurement data acquired based on uplink reference signals, fuses the terminal-side measurement data and network-side measurement data to obtain fused measurement data, and finally obtains wireless measurement data of the target network area based on the terminal-side measurement data, network-side measurement data and fused measurement data, as well as obtains environmental semantic data of the target network area and location labeling information corresponding to the wireless measurement data.

[0069] Step S202: Construct a set of semantic tags corresponding to the wireless measurement data based on the location annotation information and environmental semantic data.

[0070] Optionally, the computing device 102 performs spatial geometric calculations based on spatial element information in the location annotation information and environmental semantic information to generate spatial semantic tags corresponding to the wireless measurement data. It also performs matching processing on the planar schematic diagrams in the location annotation information and environmental semantic information to generate functional semantic tags corresponding to the wireless measurement data, and generates custom semantic tags corresponding to the wireless measurement data based on the spatial management rules in the environmental semantic information. Finally, it combines the spatial semantic tags, functional semantic tags, and custom semantic tags to obtain a set of semantic tags corresponding to the wireless measurement data.

[0071] Step S203: Input the wireless measurement data into the wireless positioning model to be trained, and obtain the location prediction result and semantic prediction result corresponding to the wireless measurement data through the wireless positioning model.

[0072] The wireless positioning model can be understood as a joint learning model with semantic awareness. The location prediction result can be understood as the positioning information of each sample terminal output by the model. Similarly, the semantic prediction result can be understood as the semantic label information of each sample terminal output by the model.

[0073] In one embodiment, the wireless positioning model includes an input encoding layer, a shared representation layer, and a multi-task output layer. The input encoding layer is used to encode the features of wireless measurement data to obtain a wireless feature representation with a unified dimension; the shared representation layer is used to perform nonlinear transformation and fusion on the wireless feature representation to obtain a shared representation for location prediction and semantic prediction; the multi-task output layer includes a location output header and a semantic output header. The location output header is used to output location coordinates, grid numbers, or relative position results, and the semantic output header is used to output coarse-grained semantic prediction results and / or fine-grained semantic prediction results.

[0074] Optionally, the input encoding layer is implemented using a multilayer perceptron (MLP), convolutional network, recurrent network, Transformer encoder, or a combination thereof; the shared representation layer is implemented using a residual network, self-attention network, or multilayer perceptron; and the position output head and semantic output head are implemented using fully connected layers, respectively.

[0075] For example, the computing device 102 performs data preprocessing on the wireless measurement data to obtain processed wireless measurement data. The processed wireless measurement data is then input into the wireless positioning model to be trained. The processed wireless measurement data is encoded through the input encoding layer in the wireless positioning model to obtain a unified feature representation corresponding to the wireless measurement data. The unified feature representation is then input into the shared representation layer in the wireless positioning model. Based on the unified feature representation, location semantic fusion features are extracted to obtain the high-level fusion features corresponding to the unified feature representation. The high-level fusion features are then input into the multi-task output layer of the wireless positioning model. The location prediction result corresponding to the wireless measurement data is obtained through the location output head of the multi-task output layer. The main semantic prediction result and the fine-grained semantic prediction result corresponding to the wireless measurement data are obtained through the main semantic output head and the fine-grained semantic output head of the multi-task output layer, respectively.

[0076] Step S204: Based on the location annotation information and semantic label set, as well as the location prediction results and semantic prediction results, the wireless positioning model is trained to obtain the trained wireless positioning model.

[0077] During training, a location loss L_pos is constructed based on the location prediction results and location annotation information, and a semantic loss L_sem is constructed based on the semantic prediction results and semantic label set. A total loss function is then constructed using a weighted approach.

[0078] L_total = αL_pos+ βL_sem_main + γL_sem_fine

[0079] Where L_sem_main represents the coarse-grained semantic loss, L_sem_fine represents the fine-grained semantic loss, and α, β, and γ are weight coefficients. The positional loss can use mean squared error (MSE), mean absolute error (MAE), smooth L1 loss, or grid classification cross-entropy loss; the semantic loss can use cross-entropy loss, focal loss, or a combination thereof. The aforementioned weight coefficients can be dynamically adjusted based on the training stage, task convergence speed, or validation set error.

[0080] Optionally, the computing device 102 constructs the location loss of the wireless positioning model based on the location annotation information and the location prediction result, constructs the main semantic loss of the wireless positioning model based on the main semantic prediction result and the semantic label set, constructs the fine-grained semantic loss of the wireless positioning model based on the fine-grained semantic prediction result and the semantic label set when the coarse-grained label in the semantic label set is the same as any preset coarse-grained label, and combines the main semantic loss and the fine-grained semantic loss to obtain the semantic loss of the wireless positioning model; or, when the coarse-grained label in the semantic label set is not the same as any preset coarse-grained label, the main semantic loss is used as the semantic loss of the wireless positioning model, and finally, the wireless positioning model is trained based on the location loss and the semantic loss to obtain the trained wireless positioning model.

[0081] In the aforementioned wireless positioning model training method, wireless measurement data of the target network area and environmental semantic data representing the spatial or functional attributes of the target network area, as well as location annotation information corresponding to the wireless measurement data, are acquired. Based on the location annotation information, environmental semantic data, and the semantic label set corresponding to the wireless measurement data, the wireless measurement data is input into the wireless positioning model to be trained. The wireless positioning model obtains the location prediction result and semantic prediction result corresponding to the wireless measurement data. Based on the location annotation information, semantic label set, location prediction result, and semantic prediction result, the wireless positioning model is trained to obtain the trained wireless positioning model. By introducing environmental semantic data of the target network area to generate the corresponding semantic label set and integrating semantic information into the training process of the wireless positioning model, the adaptability of the trained wireless positioning model to the actual environment is ensured, thereby improving the positioning accuracy of the trained wireless positioning model and achieving an improvement in wireless positioning precision.

[0082] In one embodiment, environmental semantic data includes spatial element information, a plan view, and spatial management rules. Based on location annotation information and environmental semantic data, a set of semantic tags corresponding to the wireless measurement data is constructed, including: performing spatial geometric calculations based on location annotation information and spatial element information to generate spatial semantic tags corresponding to the wireless measurement data; matching location annotation information with the plan view to generate functional semantic tags corresponding to the wireless measurement data; generating custom semantic tags corresponding to the wireless measurement data according to spatial management rules; and combining spatial semantic tags, functional semantic tags, and custom semantic tags to obtain a set of semantic tags corresponding to the wireless measurement data.

[0083] Spatial element information can be understood as a real geographic object depicted in the digital world through geometric shapes (points / lines / areas) and attribute information, which may include road centerlines, building outlines, polygonal areas, or points of interest areas; a plan view can be understood as a simplified graphic representation that is not based on precise scale, used to show spatial layout, relative positions, and functional zoning, which may include floor plans, station facility maps, airport maps, or shopping mall maps; spatial management rules can be understood as a series of regulations, standards, and norms that need to be followed when planning, approving, using, and maintaining various elements (buildings, roads, plots, pipelines, etc.) in geographic space, which may include manually defined area boundaries, control rules, or service rules.

[0084] For example, the computing device 102 performs spatial set calculations based on the location annotation information and the road centerline, building outline, polygon area or point of interest area to obtain spatial semantic labels corresponding to the wireless measurement data. Based on the location annotation information and the floor plan, station facility map, airport map or shopping mall map, it performs matching processing to generate functional semantic labels corresponding to the wireless measurement data. It also generates custom semantic labels corresponding to the wireless measurement data based on manually divided area boundaries, control rules or service rules. The spatial semantic labels, functional semantic labels and custom semantic labels are combined to obtain a set of semantic labels corresponding to the wireless measurement data.

[0085] According to the above implementation method, by performing multi-dimensional semantic tag annotation, the semantic tags in the semantic tag set are guaranteed to provide a complete description of the sample's location, providing rich data resources for subsequent semantic information association.

[0086] In one embodiment, wireless measurement data is input into a wireless positioning model to be trained, and the location prediction result and semantic prediction result corresponding to the wireless measurement data are obtained through the wireless positioning model. This includes: performing data preprocessing on the wireless measurement data to obtain processed wireless measurement data; inputting the processed wireless measurement data into the wireless positioning model, encoding the processed wireless measurement data through the input encoding layer in the wireless positioning model to obtain a unified feature representation corresponding to the wireless measurement data; inputting the unified feature representation into the shared representation layer in the wireless positioning model, extracting location semantic fusion features based on the unified feature representation to obtain a high-level fusion feature corresponding to the unified feature representation; inputting the high-level fusion feature into the multi-task output layer of the wireless positioning model, obtaining the location prediction result corresponding to the wireless measurement data through the location output head of the multi-task output layer, and obtaining the main semantic prediction result and fine-grained semantic prediction result corresponding to the wireless measurement data through the main semantic output head and fine-grained semantic output head of the multi-task output layer, respectively.

[0087] The main semantic prediction result can be understood as the major category label corresponding to the location of the sample, which may include at least one of spatial semantic category, functional semantic category and multi-label semantic result; the fine-grained semantic prediction result can be understood as the minor category label corresponding to the location of the sample, which may include at least one of road segment number, floor number, functional area number or local sub-scene category.

[0088] Optionally, the model architecture of the wireless fixed model is as follows: Figure 3 As shown, the computing device 102 performs at least one of the following data preprocessing operations on the wireless measurement data: outlier removal, missing value processing, feature normalization, discrete feature mapping, neighborhood feature alignment, and temporal segment organization, to obtain processed wireless measurement data. The processed wireless measurement data is then input into a wireless positioning model. The processed wireless measurement data is encoded by the input encoding layer in the wireless positioning model to obtain a unified feature representation corresponding to the wireless measurement data. The unified feature representation is then input into the shared representation layer in the wireless positioning model. Based on the unified feature representation, location semantic fusion feature extraction is performed to obtain the high-level fusion feature corresponding to the unified feature representation. The high-level fusion feature is then input into the multi-task input layer of the wireless positioning model. Through the main semantic output head and fine-grained semantic output head of the multi-task output layer, the main semantic prediction result and fine-grained semantic prediction result corresponding to the wireless measurement data are obtained, respectively.

[0089] Based on the aforementioned implementation method, by designing a multi-layered structure within the model, data is processed and feature constructed progressively through the multi-layered architecture, ensuring the traceability of data processing and improving the traceability of model failures.

[0090] In an exemplary embodiment, a wireless positioning model is trained based on location annotation information and a set of semantic tags, as well as location prediction results and semantic prediction results, to obtain a trained wireless positioning model. This includes: constructing a location loss of the wireless positioning model based on location annotation information and location prediction results; constructing a semantic loss of the wireless positioning model based on the main semantic prediction results, fine-grained semantic prediction results, and a set of semantic tags; and training the wireless positioning model based on the location loss and semantic loss to obtain a trained wireless positioning model.

[0091] The location loss is used to constrain the difference between the location prediction result and the location labeling information, and may include at least one of coordinate regression loss, raster classification loss or other location error loss; the semantic loss is used to constrain the difference between the semantic prediction result and the semantic labels in the semantic label set.

[0092] For example, the computing device 102 constructs the location loss of the wireless positioning model based on the difference between the location standard information and the location prediction result, constructs the semantic loss of the wireless positioning model based on the main semantic prediction result, the fine-grained semantic prediction result and the semantic label set, obtains the first loss weight pre-constructed for the location loss and the second loss weight pre-constructed for the semantic loss, and performs a weighted summation of the location loss and the semantic loss according to the first loss weight and the second loss weight to obtain the joint loss of the wireless positioning model, updates the model parameters of the wireless positioning model based on the joint loss, and returns to execute the step of inputting wireless measurement data into the wireless positioning model until the joint loss is less than or equal to a preset joint loss threshold, or the deviation between the joint losses of consecutive target counts is less than or equal to a preset deviation threshold, and the wireless positioning model at this time is determined as the wireless positioning model that has been trained.

[0093] According to the above implementation method, by training the wireless positioning model by combining semantic loss and location loss, an explicit collaborative relationship is formed between the location task and the semantic task, which can make fuller use of the correlation between location and scene to improve the positioning performance of the trained wireless positioning model.

[0094] In one embodiment, constructing the semantic loss of the wireless positioning model based on the semantic prediction result and the semantic label set includes: constructing the main semantic loss of the wireless positioning model based on the main semantic prediction result and the semantic label set; when the coarse-grained label in the semantic label set is the same as any preset coarse-grained label, constructing the fine-grained semantic loss of the wireless positioning model based on the fine-grained semantic prediction result and the semantic label set; combining the main semantic loss and the fine-grained semantic loss to obtain the semantic loss of the wireless positioning model; when the coarse-grained label in the semantic label set is not the same as any preset coarse-grained label, using the main semantic loss as the semantic loss of the wireless positioning model.

[0095] The main semantic loss is used to constrain the difference between the main semantic prediction result and the coarse-grained labels in the semantic label set. It may include at least one of the main semantic loss, such as cross-entropy loss, Focal Loss, or other classification loss. The fine-grained semantic loss is used to constrain the difference between the fine-grained semantic prediction result and the fine-grained labels in the semantic label set.

[0096] Optionally, the computing device 102 constructs the main semantic loss of the wireless positioning model based on the difference between the main semantic prediction result and the coarse-grained labels in the semantic label set; when the coarse-grained labels in the semantic label set are the same as any preset coarse-grained labels, the fine-grained semantic loss of the wireless positioning model is constructed based on the difference between the fine-grained semantic prediction result and the fine-grained labels in the semantic label set; the main semantic loss and the fine-grained semantic loss are combined to obtain the semantic loss of the wireless positioning model; when the coarse-grained labels in the semantic label set are not the same as any preset coarse-grained labels, the main semantic loss is used as the semantic loss of the wireless positioning model.

[0097] Based on the aforementioned implementation method, a conditional training mechanism is introduced to calculate the fine-grained semantic loss. The fine-grained semantic loss is calculated only when the semantic label set meets the preset conditions. If the semantic label set does not meet the preset conditions, the fine-grained semantic loss is not calculated. This reduces invalid supervision for different output heads and lowers the training cost of model training.

[0098] In one embodiment, acquiring wireless measurement data of a target network area includes: acquiring terminal-side measurement data and network-side measurement data within the target network area; fusing the terminal-side measurement data and network-side measurement data to obtain fused measurement data; and obtaining wireless measurement data of the target network area based on the terminal-side measurement data, network-side measurement data, and fused measurement data.

[0099] Among them, the terminal-side measurement data can be understood as the physical information of the sample terminal itself and the detected wireless signal information, which may include MDT (Minimization of Drive-Tests) data or MR (Measurement Report) data. The MDT data or MR data includes at least one of the following: serving cell measurement results, neighboring cell measurement results, and auxiliary features. The serving cell measurement results and neighboring cell measurement results include at least one of the following: reference signal received power, reference signal received quality, channel quality indication, timing advance, carrier frequency identifier, physical cell identifier, number of neighboring cells, and signal-to-noise ratio. The network-side measurement data can be understood as the channel measurement results obtained based on the uplink reference signal. The channel measurement results include at least one of the following: channel impulse response, power delay distribution, angle of arrival, delay spread, angle spread, and beam measurement results.

[0100] For example, the computing device 102 receives raw terminal-side measurement data uploaded by sample terminals in the target network area, and acquires raw network-side measurement data collected by the computing device 102. The raw terminal-side measurement data and the raw network-side measurement data are cleaned and aligned in the time dimension to obtain their respective corresponding terminal-side measurement data and network-side measurement data. Then, the terminal-side measurement data and the network-side measurement data are fused to obtain fused data. Finally, wireless measurement data is constructed based on the terminal-side measurement data, the network-side measurement data, and the fused data.

[0101] Based on the above implementation method, by acquiring measurement data from different data sources and constructing wireless measurement data based on data from different dimensions, a rich data foundation is provided for subsequent model training, thereby improving the training accuracy of the model.

[0102] In one exemplary embodiment, such as Figure 4 As shown, a wireless positioning method is provided, which is applied to... Figure 1 The computing device 102 includes steps S401 and S402. Wherein:

[0103] Step S401: Obtain the wireless measurement data of the network area to be located.

[0104] Step S402: Input the wireless measurement data to be located into the trained wireless positioning model, and output the location prediction result and semantic prediction result of the network area to be located through the wireless positioning model; the wireless positioning model is trained by any of the wireless positioning model training methods in the above embodiments.

[0105] Optionally, the computing device 102 acquires terminal-side measurement data uploaded by each sample terminal in the network area to be located, as well as network-side measurement data collected by the computing device 102 based on the uplink reference signal. It then fuses the terminal-side and network-side measurement data to obtain fused data. Based on the terminal-side, network-side, and fused data, it generates wireless measurement data to be located. Subsequently, it inputs the wireless measurement data to be located into a trained wireless positioning model. The wireless positioning model outputs the location prediction result and semantic prediction result for the network area to be located. The wireless positioning model is trained through the following steps:

[0106] The computing device 102 acquires terminal-side measurement data from sample terminals in the target network area and network-side measurement data acquired based on uplink reference signals. It fuses the terminal-side and network-side measurement data to obtain fused measurement data. Finally, based on the terminal-side, network-side, and fused measurement data, it obtains the wireless measurement data for the target network area, as well as the environmental semantic data and location annotation information corresponding to the wireless measurement data. Subsequently, it performs spatial geometric calculations based on spatial element information in the location annotation information and environmental semantic information to generate spatial semantic labels corresponding to the wireless measurement data. It then matches the planar schematic diagrams in the location annotation information and environmental semantic information to generate functional semantic labels corresponding to the wireless measurement data, and generates custom semantic labels corresponding to the wireless measurement data based on spatial management rules in the environmental semantic information. Finally, it combines the spatial semantic labels, functional semantic labels, and custom semantic labels to obtain a set of semantic labels corresponding to the wireless measurement data. The computing device 102 performs data preprocessing on the wireless measurement data to obtain processed wireless measurement data. It inputs the processed wireless measurement data into the wireless positioning model to be trained, and encodes the processed wireless measurement data through the input encoding layer in the wireless positioning model to obtain a unified feature table corresponding to the wireless measurement data. The unified feature representation is input into the shared representation layer of the wireless positioning model. Location semantic fusion feature extraction is performed based on the unified feature representation to obtain the high-level fusion feature corresponding to the unified feature representation. The high-level fusion feature is then input into the multi-task output layer of the wireless positioning model. The location prediction result corresponding to the wireless measurement data is obtained through the location output head of the multi-task output layer. The main semantic prediction result and fine-grained semantic prediction result corresponding to the wireless measurement data are obtained through the main semantic output head and fine-grained semantic output head of the multi-task output layer, respectively. The computing device 102 constructs the location loss of the wireless positioning model based on the location annotation information and the location prediction result, and then calculates the location loss based on the main semantic prediction result. Based on the result and semantic label set, the main semantic loss of the wireless positioning model is constructed. If the coarse-grained label in the semantic label set is the same as any preset coarse-grained label, the fine-grained semantic loss of the wireless positioning model is constructed based on the fine-grained semantic prediction result and the semantic label set. The main semantic loss and the fine-grained semantic loss are combined to obtain the semantic loss of the wireless positioning model. Alternatively, if the coarse-grained label in the semantic label set is not the same as any preset coarse-grained label, the main semantic loss is used as the semantic loss of the wireless positioning model. Finally, the wireless positioning model is trained based on the position loss and the semantic loss to obtain the trained wireless positioning model.

[0107] In one embodiment, the method further includes: performing a confidence assessment on the semantic prediction result, and generating semantic additional features based on the semantic confidence and semantic prediction result obtained from the assessment; inputting the semantic additional features and the wireless measurement data to be located into a pre-built semantic additional feature enhancement model, and obtaining the optimized location prediction result of the network area to be located through the semantic additional feature enhancement model.

[0108] The semantic augmentation model can be understood as another wireless positioning model trained with semantic augmentation features on the input.

[0109] For example, computing device 102 performs at least one of the following inference processes on the semantic prediction results output by the wireless positioning model, including:

[0110] First, semantic constraint correction is applied to the location prediction results: when the semantic prediction results indicate that the target is located in a road scene, the location prediction results are projected or corrected to the geometric constraint direction of the nearest road; when the semantic prediction results indicate that the target is located in a building scene, unreasonable location results falling outside the building are suppressed; when the semantic prediction results indicate that the target is located in an indoor scene, the 3D location results are corrected by combining the floor prediction results or floor priors.

[0111] Second: Select the corresponding positioning model, positioning parameters or post-processing strategy based on the main semantic prediction results: select different dedicated positioning sub-models based on indoor or outdoor semantics; switch different post-processing rules based on the semantics of roads, buildings or open areas; decide whether to start the fine positioning process based on semantic confidence.

[0112] Third, the reliability of the location prediction results is evaluated based on semantic confidence. When the semantic confidence is lower than the preset threshold, the corresponding location prediction results can be marked as low confidence results for use by the upper-layer business system for alarm, filtering or downgrading.

[0113] Subsequently, the computing device 102 constructs semantic additional features based on the semantic confidence obtained in the inference phase and the semantic prediction results output by the previous model. The semantic additional features and the wireless measurement data to be located are input into a pre-constructed wireless positioning model that includes semantic additional features, and the optimized location prediction results and optimized semantic prediction results are obtained.

[0114] Based on the aforementioned implementation method, by designing a two-stage positioning optimization, the semantic prediction result obtained from the first positioning is subjected to reasoning processing, and new additional input features are constructed based on the reasoning result and the semantic prediction structure. The additional input features and the wireless measurement data to be located are input into the pre-constructed semantic additional feature enhancement model for two-stage positioning, thereby obtaining the location prediction result of the network area to be located, thus improving the positioning accuracy of the network area to be located.

[0115] In one embodiment, the method further includes: obtaining the predicted semantic distribution of the semantic prediction results of the network region to be located, and the training semantic distribution of the semantic prediction results output by the wireless positioning model during the training phase; generating an early warning signal of model performance degradation or scene change when the offset between the predicted semantic distribution and the training semantic distribution is greater than a preset offset threshold; and triggering model retraining, incremental learning, or model parameter fine-tuning of the wireless positioning model according to the early warning signal.

[0116] For example, the computing device 102 obtains the predicted semantic distribution information of the semantic prediction results of the network region to be located, and obtains the training semantic distribution information of the semantic prediction results output by the wireless positioning model during the training phase. If the offset between the predicted semantic distribution information and the training semantic distribution information is greater than a preset offset threshold, an early warning signal of model performance degradation or scene change is generated. If the early warning signal is detected, model retraining, incremental learning or model parameter fine-tuning for the wireless positioning model is triggered.

[0117] According to the above implementation method, by comparing the semantic distribution of the semantic prediction results in the application stage and the training stage, if there is a significant shift in the semantic distribution in the application stage compared with the semantic distribution in the training stage, it indicates that the semantic information fusion of the trained wireless positioning model is not tight enough. Therefore, the model adjustment action for the wireless positioning model is triggered to ensure the tight coupling between the wireless positioning model and the semantic information in the corresponding environment, thereby improving the positioning accuracy of wireless positioning.

[0118] In one exemplary embodiment, a wireless positioning method incorporating semantics is provided, comprising the following steps, wherein:

[0119] S1, acquire wireless measurement data and the location labeling information corresponding to the wireless measurement data;

[0120] In step S1, the wireless measurement data may include at least one of the following:

[0121] MDT or MR data reported by the terminal side;

[0122] Measurement data acquired on the network side based on the uplink reference signal;

[0123] Fusion measurement data consisting of two or more of the above types of data.

[0124] MDT data or MR data includes at least one of the following: serving cell measurement results, neighboring cell measurement results, and auxiliary characteristics. Serving cell measurement results and neighboring cell measurement results include at least one of the following: reference signal received power, reference signal received quality, channel quality indication, timing advance, carrier frequency identifier, physical cell identifier, number of neighboring cells, and signal-to-noise ratio.

[0125] The network-side measurement data includes channel measurement results obtained based on the uplink reference signal. The channel measurement results include at least one of the following: channel impulse response, power delay distribution, angle of arrival, delay spread, angle spread, and beam measurement results.

[0126] Location annotation information is used to characterize the known location of the sample. Location annotation information includes at least one of the following: two-dimensional coordinates, three-dimensional coordinates, relative base station offset, grid number, and latitude and longitude.

[0127] S2, construct a semantic label set for the training samples based on the location annotation information and environmental semantic data;

[0128] Environmental semantic data is used to characterize the spatial or functional attributes of a location area. Environmental semantic data includes at least one of the following: GIS vector data, local map data, regional division data, and management rule data.

[0129] The semantic tag set includes at least one of the following semantic tags:

[0130] Spatial semantic tags are used to characterize the physical space type in which the sample is located. The physical space type includes at least one of roads, buildings, squares, parks, parking lots, and underground spaces.

[0131] Functional semantic tags are used to characterize the business function attributes of the area where the sample is located. The business function attributes include at least one of the following: payment area, waiting area, security check area, boarding area, shop area, passage, and restricted area.

[0132] Custom semantic tags are used to characterize regional attributes defined according to manual rules, operational rules, or business rules.

[0133] The semantic tag set can include coarse-grained semantic tags and fine-grained semantic tags. Coarse-grained semantic tags are used to represent the major scene category to which the sample belongs, while fine-grained semantic tags are used to represent the sub-region category within the corresponding major scene category. Fine-grained semantic tags include at least one of road segment numbers, floor numbers, functional area numbers, or other local area identifiers.

[0134] The generation of semantic tag sets can be achieved through at least one of the following methods:

[0135] Spatial geometric calculations are performed based on the sample location and the road centerline, building outline, polygonal region or point of interest region to generate spatial semantic labels;

[0136] Based on the sample location, perform matching processing with floor plans, station facilities maps, airport maps, or shopping mall maps to generate functional semantic tags;

[0137] Custom semantic tags are generated based on manually defined area boundaries, control rules, or service rules.

[0138] Therefore, each training sample includes at least three parts: wireless measurement features, location labels, and semantic label sets, thus forming a semantically enhanced wireless positioning sample.

[0139] S3, preprocesses and encodes the wireless measurement data to obtain a unified feature representation;

[0140] In step S3, preprocessing operations are performed on the wireless measurement data. The preprocessing operations include at least one of outlier removal, missing value handling, feature normalization, discrete feature mapping, neighbor feature alignment, and time-series segment processing.

[0141] To adapt to input data from different sources, formats, and dimensions, the method employs an input coding layer to encode the preprocessed wireless measurement features, resulting in a unified feature representation.

[0142] The input coding layer receives at least one type of wireless measurement data and converts it into a unified tensor representation. Preferably, the input coding layer also generates feature validity markers, which indicate whether the corresponding wireless measurement item is missing, abnormal, or only available under a specific communication standard.

[0143] The input encoding layer has a pluggable structure. Different encoding strategies can be used for MDT data, MR data, SRS measurement data, or fused measurement data; however, the outputs of different encoding strategies are all converted into shared feature inputs that can be processed by the subsequent unified model.

[0144] S4 takes the unified feature representation as input to the wireless positioning model to be trained, outputs the location prediction result and semantic prediction result, and trains it based on the multi-task joint loss function.

[0145] In step S4, a unified feature representation is used to train the wireless localization model. The wireless localization model includes an input coding layer, a shared representation layer, and a multi-task output layer.

[0146] The shared representation layer is connected to the input encoding layer and is used to extract high-level features that are commonly related to both the location and semantic tasks from the unified feature representation. The shared representation layer can employ a multilayer perceptron, residual multilayer perceptron, attention network, convolutional network, or a combination thereof.

[0147] The multi-task output layer is connected to the shared representation layer and includes at least a location output header and a main semantic output header; preferably, the multi-task output layer also includes at least one fine-grained semantic output header.

[0148] The location output header is used to output the location prediction results, which include at least one of the following: two-dimensional coordinates, three-dimensional coordinates, relative base station offset, grid index, latitude and longitude, road segment location, and floor number.

[0149] The main semantic output header is used to output the main semantic prediction result, which includes at least one of spatial semantic category, functional semantic category, and multi-label semantic result.

[0150] The fine-grained semantic output header is used to output fine-grained semantic prediction results, which include at least one of the following: road segment number, floor number, functional area number, or local sub-scene category.

[0151] Preferably, a semantic-aware interaction module, a gating decoupling module, a feature modulation module, or an attention modulation module can be set between the shared representation layer and the multi-task output layer to enhance the correlation modeling capability between location tasks and semantic tasks.

[0152] The wireless positioning model is trained using a multi-task joint loss function. This function includes a location loss, a main semantic loss, and at least one fine-grained semantic loss.

[0153] Location loss is used to constrain the difference between the location prediction result and the location label. Location loss includes at least one of coordinate regression loss, raster classification loss, or other location error loss.

[0154] The main semantic loss is used to constrain the difference between the main semantic prediction result and the main semantic label. The main semantic loss includes at least one of cross-entropy loss, Focal Loss or other classification loss.

[0155] Fine-grained semantic loss is used to constrain the difference between fine-grained semantic prediction results and fine-grained semantic labels. To reduce invalid supervision, the fine-grained semantic loss is calculated using a conditional training mechanism.

[0156] The conditional training mechanism includes: when the coarse-grained semantic label corresponding to the sample meets the preset conditions, the fine-grained semantic output head corresponding to the preset conditions is enabled to perform loss calculation; when the coarse-grained semantic label corresponding to the sample does not meet the preset conditions, loss calculation is not performed on the corresponding fine-grained semantic output head.

[0157] The preset conditions include at least one of the following:

[0158] When the coarse-grained semantic label is "road", the road segment classification task is enabled.

[0159] When the coarse-grained semantic label is "building", enable the floor classification task;

[0160] When the coarse-grained semantic label is a specific location area, the functional area classification task is enabled.

[0161] Preferably, the conditional training mechanism includes at least one of the following implementation methods:

[0162] Training methods based on semantic condition masks;

[0163] Training method based on hierarchical semantic propagation;

[0164] Training method based on independent judgment of multidimensional semantic labels.

[0165] The weights of the loss for each task can be set as fixed hyperparameters, or they can be dynamically adjusted according to the training stage, sample distribution, class imbalance, or model convergence status.

[0166] S5, during the inference phase, perform at least one of the following on the location prediction results: semantic constraints, model or parameter selection, post-processing switching, and result confidence evaluation, based on the semantic prediction results;

[0167] In step S5, the trained wireless positioning model simultaneously outputs location prediction results and semantic prediction results during the inference phase. The semantic prediction results are used to perform at least one of the following inference processes:

[0168] Semantic constraint corrections are applied to the location prediction results;

[0169] Select the corresponding localization model, localization parameters, or post-processing strategy based on the semantic category;

[0170] The reliability of the localization results is assessed based on semantic confidence.

[0171] Execute scenario consistency judgment based on the joint results of multiple semantic dimensions.

[0172] Semantic constraint modification includes at least one of the following:

[0173] When the semantic prediction result indicates that the target is located in a road scene, the location prediction result is projected or corrected to the direction of the nearest road geometric constraint.

[0174] When the semantic prediction result represents the target as being located in a building scene, suppress results that fall in unreasonable locations outside the building.

[0175] When the semantic prediction result indicates that the target is located in an indoor scene, the 3D position result is corrected by combining the floor prediction result or the floor prior.

[0176] The selection of model, parameters, or post-processing strategies includes at least one of the following:

[0177] Different dedicated positioning sub-models are selected based on indoor or outdoor semantics;

[0178] Different post-processing rules are switched based on the semantics of roads, buildings, or open areas;

[0179] Whether to initiate the fine-grained localization process is determined based on semantic confidence.

[0180] Reliability assessment includes outputting a semantic confidence score. When the semantic confidence score is lower than a preset threshold, the corresponding location result can be marked as a low-confidence result for use by upper-layer business systems for alarms, filtering, or degradation.

[0181] S6, Optionally, the semantic prediction result or semantic confidence vector is fed back as an additional input to the localization model or refinement model to obtain the optimized location prediction result.

[0182] In step S6, the method may further include a semantic feedback optimization step. The semantic feedback optimization step includes: fusing the semantic prediction results, semantic category codes, or semantic confidence vectors obtained in the inference stage as additional input features with the original wireless measurement features, and inputting the fused features into the localization model or refinement model to obtain optimized location prediction results.

[0183] Preferably, the semantic feedback optimization step forms a two-stage localization process, which includes:

[0184] Initial location prediction results and semantic prediction results are obtained based on the primary model;

[0185] Construct additional semantic features based on semantic prediction results;

[0186] The semantically added features and the original wireless measurement features are input together into the refinement model;

[0187] Output the final location prediction result.

[0188] In addition, the semantic output distribution during the inference phase can be continuously monitored. When the semantic output distribution shifts significantly relative to the training phase, this shift can be used as an early warning signal of model performance degradation or scene changes, thereby triggering model retraining, incremental learning, or parameter fine-tuning.

[0189] Among them, wireless measurement data, location annotation information, environmental semantic data, wireless positioning model, multi-task joint loss function and semantic prediction results constitute a complete technical closed loop of "sample construction - joint learning - semantic reasoning - result optimization".

[0190] The method can be deployed on network-side positioning servers, edge nodes, or base station-side computing units, or on terminal-side processing modules. The training process can be executed in the cloud or on the network side, while the inference process can be executed on the network side, on the terminal side, or in a network-terminal collaborative manner.

[0191] Different deployment methods do not change the core technical ideas of this application regarding semantic label construction, joint learning, conditional training, semantic-driven reasoning, and semantic feedback optimization.

[0192] Example of specific implementation:

[0193] Example 1: A semantic-aware localization method based on MDT data.

[0194] The system acquires MDT data reported by the terminal, which includes reference signal received power, reference signal received quality, channel quality indication, and timing advance of the serving cell and neighboring cells. It then performs spatial matching based on the sample location and GIS road centerlines and building outlines to generate spatial semantic labels for the samples. Subsequently, the samples are input into a semantically aware joint learning model, which outputs location prediction results, main semantic prediction results for roads / buildings / other categories, and fine-grained prediction results for road segments. The fine-grained prediction results for road segments are only used in loss calculations for road-type samples. During the inference phase, road projection correction is performed on the location prediction results based on the road-type semantic results.

[0195] Example 2: Indoor and Outdoor Semantic Aware Localization Method Based on SRS Data. The method acquires channel measurement data from the network side based on uplink SRS and extracts channel impulse response features; it performs matching between sample location and scene semantic region to generate indoor, outdoor, or transition zone semantic labels for the sample; it synchronously outputs location results and spatial semantic results through a joint learning model; and during the inference phase, it switches the corresponding localization post-processing strategy based on the indoor and outdoor semantic results.

[0196] Example 3: A two-stage optimization localization method based on semantic feedback.

[0197] The primary model first outputs the initial location result and semantic prediction result; then the confidence vector corresponding to the semantic prediction result is concatenated with the original wireless measurement features and input into the refinement model for secondary estimation; the refinement model uses semantic priors to correct the initial location result, thereby reducing the positioning error in road intersections, building boundaries and indoor-outdoor transition areas.

[0198] Example 4: A location-level localization method that integrates functional semantics.

[0199] In locations such as shopping malls, airports, train stations, or hospitals, functional semantic labels are generated for samples based on local facility maps. These functional semantic labels include at least one of the following: waiting area, shop area, passageway, waiting area for outpatient services, and security check area. The model outputs functional semantic results along with location results, enabling the location results to be directly used for business triggering, regional service push, and venue management.

[0200] Compared with existing public disclosures, this application has the following technical advantages:

[0201] 1. By using semantically aware interaction design, an explicit collaborative relationship is formed between location tasks and semantic tasks, which can make fuller use of the correlation between location and scene to improve positioning performance.

[0202] 2. Semantic information is used in the inference stage for position constraints, model selection, parameter switching, confidence assessment and joint decision-making. Furthermore, a two-stage optimization process is formed through semantic feedback, so that semantic information plays a substantial role in both the training and inference stages, thereby forming a complete technical closed loop from sample construction, model learning to inference optimization.

[0203] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0204] Based on the same inventive concept, this application also provides a wireless positioning model training device for implementing the wireless positioning model training method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the wireless positioning model training device provided below can be found in the limitations of the wireless positioning model training method described above, and will not be repeated here.

[0205] In one exemplary embodiment, such as Figure 5 As shown, a wireless positioning model training device is provided, comprising: a first acquisition module 501, a construction module 502, an input module 503, and a training module 504, wherein:

[0206] The first acquisition module 501 is used to acquire wireless measurement data and environmental semantic data of the target network area, as well as the location labeling information corresponding to the wireless measurement data; the environmental semantic data is used to characterize the spatial attributes or functional attributes of the target network area.

[0207] Module 502 is used to construct a set of semantic tags corresponding to wireless measurement data based on location annotation information and environmental semantic data.

[0208] The input module 503 is used to input wireless measurement data into the wireless positioning model to be trained, and obtain the location prediction result and semantic prediction result corresponding to the wireless measurement data through the wireless positioning model.

[0209] Training module 504 is used to train the wireless positioning model based on location annotation information and semantic label set, as well as location prediction results and semantic prediction results, to obtain the trained wireless positioning model.

[0210] In one embodiment, the environmental semantic data includes spatial element information, a plan view, and spatial management rules. The construction module 502 is further used to perform spatial geometric calculations based on location annotation information and spatial element information to generate spatial semantic tags corresponding to the wireless measurement data; to perform matching processing on the location annotation information and the plan view to generate functional semantic tags corresponding to the wireless measurement data; to generate custom semantic tags corresponding to the wireless measurement data according to the spatial management rules; and to combine the spatial semantic tags, functional semantic tags, and custom semantic tags to obtain a set of semantic tags corresponding to the wireless measurement data.

[0211] In one embodiment, the input module 503 is further configured to perform data preprocessing on the wireless measurement data to obtain processed wireless measurement data; input the processed wireless measurement data into the wireless positioning model, and encode the processed wireless measurement data through the input encoding layer in the wireless positioning model to obtain a unified feature representation corresponding to the wireless measurement data; input the unified feature representation into the shared representation layer in the wireless positioning model, and extract location semantic fusion features based on the unified feature representation to obtain high-level fusion features corresponding to the unified feature representation; input the high-level fusion features into the multi-task output layer of the wireless positioning model, obtain the location prediction result corresponding to the wireless measurement data through the location output head of the multi-task output layer, and obtain the main semantic prediction result and fine-grained semantic prediction result corresponding to the wireless measurement data through the main semantic output head and fine-grained semantic output head of the multi-task output layer, respectively.

[0212] In an exemplary embodiment, the training module 504 is further configured to construct the location loss of the wireless positioning model based on the location annotation information and the location prediction result; construct the semantic loss of the wireless positioning model based on the main semantic prediction result, the fine-grained semantic prediction result and the semantic label set; and train the wireless positioning model based on the location loss and the semantic loss to obtain the trained wireless positioning model.

[0213] In one embodiment, the training module 504 is further configured to construct the main semantic loss of the wireless positioning model based on the main semantic prediction result and the semantic label set; when the coarse-grained label in the semantic label set is the same as any preset coarse-grained label, construct the fine-grained semantic loss of the wireless positioning model based on the fine-grained semantic prediction result and the semantic label set, and combine the main semantic loss and the fine-grained semantic loss to obtain the semantic loss of the wireless positioning model; when the coarse-grained label in the semantic label set is not the same as any preset coarse-grained label, use the main semantic loss as the semantic loss of the wireless positioning model.

[0214] In one embodiment, the first acquisition module is further configured to acquire terminal-side measurement data and network-side measurement data within the target network area; fuse the terminal-side measurement data and network-side measurement data to obtain fused measurement data; and obtain wireless measurement data of the target network area based on the terminal-side measurement data, network-side measurement data, and fused measurement data.

[0215] In one exemplary embodiment, such as Figure 6 As shown, a wireless positioning device is provided, including: a second acquisition module 601 and a positioning module 602, wherein:

[0216] The second acquisition module 601 is used to acquire wireless measurement data of the network area to be located.

[0217] The positioning module 602 is used to input the wireless measurement data to be located into the trained wireless positioning model, and output the location prediction result and semantic prediction result of the network area to be located through the wireless positioning model; the wireless positioning model is trained by the wireless positioning model training method of each embodiment.

[0218] In one embodiment, the wireless positioning device is further configured to perform a confidence assessment on the semantic prediction result, and generate semantic additional features based on the semantic confidence and semantic prediction result obtained from the assessment; input the semantic additional features and the wireless measurement data to be located into a pre-built semantic additional feature enhancement model, and obtain the optimized location prediction result of the network area to be located through the semantic additional feature enhancement model.

[0219] In one embodiment, the wireless positioning device is further configured to acquire the predicted semantic distribution of the semantic prediction results of the network region to be located, and the training semantic distribution of the semantic prediction results output by the wireless positioning model during the training phase; generate an early warning signal of model performance degradation or scene change when the offset between the predicted semantic distribution and the training semantic distribution is greater than a preset offset threshold; and trigger model retraining, incremental learning or model parameter fine-tuning of the wireless positioning model according to the early warning signal.

[0220] The modules in the aforementioned wireless positioning model training device and wireless positioning device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computing device in hardware form, or stored in the memory of the computing device in software form, so that the processor can call and execute the corresponding operations of each module.

[0221] In one exemplary embodiment, a computing device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computing device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores wireless measurement data of the target network area, environmental semantic data, location labeling information corresponding to the wireless measurement data, a set of semantic tags corresponding to the wireless measurement data, location prediction results corresponding to the wireless measurement data, and semantic prediction results. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a wireless positioning model training method.

[0222] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computing device on which the present application is applied. The specific computing device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0223] In one exemplary embodiment, a computing device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the various method embodiments.

[0224] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the various method embodiments.

[0225] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the various method embodiments.

[0226] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0227] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0228] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0229] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for training a wireless positioning model, characterized in that, The method includes: Acquire wireless measurement data and environmental semantic data of the target network area, as well as location annotation information corresponding to the wireless measurement data; the environmental semantic data is used to characterize the spatial or functional attributes of the target network area. Based on the location annotation information and the environmental semantic data, construct a set of semantic tags corresponding to the wireless measurement data; The wireless measurement data is input into the wireless positioning model to be trained, and the location prediction result and semantic prediction result corresponding to the wireless measurement data are obtained through the wireless positioning model. Based on the location annotation information and semantic tag set, as well as the location prediction result and semantic prediction result, the wireless positioning model is trained to obtain the trained wireless positioning model.

2. The method according to claim 1, characterized in that, The environmental semantic data includes spatial element information, a plan view, and spatial management rules. The step of constructing a semantic tag set corresponding to the wireless measurement data based on the location annotation information and the environmental semantic data includes: Spatial geometric calculations are performed based on the location annotation information and the spatial element information to generate spatial semantic labels corresponding to the wireless measurement data; The location annotation information is matched with the planar schematic diagram to generate functional semantic tags corresponding to the wireless measurement data; Generate custom semantic tags corresponding to the wireless measurement data according to the space management rules; By combining the spatial semantic tags, the functional semantic tags, and the custom semantic tags, a set of semantic tags corresponding to the wireless measurement data is obtained.

3. The method according to claim 2, characterized in that, The step of inputting the wireless measurement data into the wireless positioning model to be trained, and obtaining the location prediction result and semantic prediction result corresponding to the wireless measurement data through the wireless positioning model, includes: The wireless measurement data is preprocessed to obtain the processed wireless measurement data. The processed wireless measurement data is input into the wireless positioning model, and the processed wireless measurement data is encoded by the input encoding layer in the wireless positioning model to obtain a unified feature representation corresponding to the wireless measurement data. The unified feature representation is input into the shared representation layer in the wireless positioning model, and location semantic fusion feature extraction is performed based on the unified feature representation to obtain the high-level fusion feature corresponding to the unified feature representation; The high-level fusion features are input into the multi-task output layer of the wireless positioning model. The location prediction result corresponding to the wireless measurement data is obtained through the location output head of the multi-task output layer. The main semantic prediction result and the fine-grained semantic prediction result corresponding to the wireless measurement data are obtained through the main semantic output head and the fine-grained semantic output head of the multi-task output layer, respectively.

4. The method according to claim 3, characterized in that, The step of training the wireless positioning model based on the location annotation information and semantic tag set, as well as the location prediction result and semantic prediction result, to obtain the trained wireless positioning model includes: Based on the location annotation information and the location prediction result, the location loss of the wireless positioning model is constructed; Based on the main semantic prediction result, the fine-grained semantic prediction result, and the semantic tag set, the semantic loss of the wireless positioning model is constructed; The wireless positioning model is trained based on the location loss and the semantic loss to obtain the trained wireless positioning model.

5. The method according to claim 4, characterized in that, The construction of the semantic loss of the wireless positioning model based on the semantic prediction result and the semantic label set includes: Based on the main semantic prediction results and the semantic label set, the main semantic loss of the wireless positioning model is constructed; If the coarse-grained label in the semantic label set is the same as any preset coarse-grained label, the fine-grained semantic loss of the wireless positioning model is constructed based on the fine-grained semantic prediction result and the semantic label set. The semantic loss of the wireless positioning model is obtained by combining the main semantic loss and the fine-grained semantic loss. When the coarse-grained labels in the semantic label set are not the same as any preset coarse-grained label, the main semantic loss is used as the semantic loss of the wireless positioning model.

6. The method according to any one of claims 1-4, characterized in that, The acquisition of wireless measurement data for the target network area includes: Acquire terminal-side measurement data and network-side measurement data within the target network area; The terminal-side measurement data and the network-side measurement data are fused to obtain fused measurement data. The wireless measurement data of the target network area is obtained based on the terminal-side measurement data, the network-side measurement data, and the fused measurement data.

7. A wireless positioning method, characterized in that, The method includes: Acquire wireless measurement data of the network area to be located; The wireless measurement data to be located is input into the trained wireless positioning model, and the wireless positioning model outputs the location prediction result and semantic prediction result of the network area to be located; the wireless positioning model is trained by the wireless positioning model training method as described in any one of claims 1 to 6.

8. The method according to claim 7, characterized in that, The method further includes: A confidence assessment is performed on the semantic prediction results, and additional semantic features are generated based on the assessed semantic confidence and the semantic prediction results. The semantic augmented features and the wireless measurement data to be located are input into a pre-built semantic augmented feature enhancement model, and the optimized location prediction result of the network area to be located is obtained through the semantic augmented feature enhancement model.

9. The method according to claim 7, characterized in that, The method further includes: Obtain the predicted semantic distribution of the semantic prediction results of the network region to be located, and the training semantic distribution of the semantic prediction results output by the wireless positioning model during the training phase. If the offset between the predicted semantic distribution and the training semantic distribution is greater than a preset offset threshold, an early warning signal for model performance degradation or scene change is generated. Based on the warning signal, the wireless positioning model is triggered to undergo model retraining, incremental learning, or model parameter fine-tuning.

10. A wireless positioning model training device, characterized in that, The device includes: The first acquisition module is used to acquire wireless measurement data and environmental semantic data of the target network area, as well as location annotation information corresponding to the wireless measurement data; the environmental semantic data is used to characterize the spatial attributes or functional attributes of the target network area. The construction module is used to construct a set of semantic tags corresponding to the wireless measurement data based on the location annotation information and the environmental semantic data; The input module is used to input the wireless measurement data into the wireless positioning model to be trained, and obtain the location prediction result and semantic prediction result corresponding to the wireless measurement data through the wireless positioning model; The training module is used to train the wireless positioning model based on the location annotation information and semantic label set, as well as the location prediction result and semantic prediction result, to obtain the trained wireless positioning model.

11. A wireless positioning device, characterized in that, The device includes: The second acquisition module is used to acquire wireless measurement data of the network area to be located. The positioning module is used to input the wireless measurement data to be positioned into a trained wireless positioning model, and output the location prediction result and semantic prediction result of the network area to be positioned through the wireless positioning model; the wireless positioning model is trained by the wireless positioning model training method as described in any one of claims 1 to 6.

12. A computing device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images