Radio map prediction method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a radio map prediction method and device, electronic equipment and a storage medium, and belongs to the technical field of wireless communication. The method comprises the following steps: acquiring a spatial form map of a target area, and selecting a plurality of target observation positions in the target area, the plurality of target observation positions being sparsely distributed; inputting the spatial form map, the observation position coordinate determined by each target observation position and the signal intensity value into a trained prediction model to carry out local feature coding processing to obtain local spatial features, and carrying out global feature coding processing on the spatial form map under a coarse granularity level to obtain global spatial features; and fusing the local spatial features and the global spatial features under different granularity levels to obtain multi-granularity coding features, and performing prediction based on the multi-granularity coding features to obtain a target radio map. According to the method and the device, the accuracy of predicting the radio map can be improved under the condition that the acquired observation information is sparse observation information.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a radio map prediction method, apparatus, electronic device, and storage medium. Background Technology

[0002] A radio map is a spatial information used to describe the electromagnetic spectrum coverage and intensity within a specific area. Radio maps can help relevant personnel understand the signal quality, interference situation, and network coverage effectiveness within a specific area.

[0003] In related technologies, radio maps are predicted using observation information at multiple pixel levels within a specific area, where dense observation information almost covers the entire area. However, in practical applications, only observation information from a limited number of locations can often be collected. Therefore, the accuracy of radio maps predicted using sparse observation information is relatively poor. Summary of the Invention

[0004] The main objective of this application is to propose a radio map prediction method, apparatus, electronic device, and storage medium that can improve the accuracy of the predicted radio map when the collected observation information is sparse.

[0005] To achieve the above objectives, a first aspect of this application proposes a radio map prediction method, the method comprising:

[0006] A spatial morphological map representing the signal occlusion situation in the target area is obtained, and multiple target observation locations are selected in the target area, wherein the multiple target observation locations are sparsely distributed;

[0007] For each target observation location, determine the corresponding observation location coordinates on the spatial morphology map, and measure the signal strength value at the target observation location;

[0008] The spatial morphology map, observation location coordinates, and corresponding signal intensity values ​​are input into the trained prediction model to perform local feature encoding on the observation location coordinates and signal intensity values ​​at a fine-grained level to obtain local spatial features, and to perform global feature encoding on the spatial morphology map at a coarse-grained level to obtain global spatial features.

[0009] Multi-granularity encoded features are obtained by fusing local and global spatial features at different granularity levels, and multi-granularity decoded features are obtained by decoding the multi-granularity encoded features. Based on the multi-granularity decoded features, a target radio map representing the signal strength at various locations in the target area is predicted.

[0010] In some embodiments, selecting multiple target observation locations within a target area includes:

[0011] Spatial analysis is performed on the spatial morphology map to determine the areas covered by ground features and open ground areas.

[0012] The spatial morphology map is divided into multiple initial regions, and at least one candidate region is selected from the multiple initial regions, wherein the area of ​​the open ground area in the candidate region is within a preset area threshold range.

[0013] For each candidate region, at least one boundary location between the open ground area and the land cover area in the candidate region is determined as the target observation location, wherein the boundary location is close to the corner of the land cover area.

[0014] In some embodiments, a first self-attention component is provided in the prediction model;

[0015] At a fine-grained level, local feature encoding is performed on the observation location coordinates and signal intensity values ​​to obtain local spatial features, including:

[0016] Local feature encoding is performed on the observation location coordinates and signal strength values ​​respectively to obtain the initial observation location features and the initial observation signal features;

[0017] The initial local fusion features are obtained by superimposing the initial observation location features and the initial observation signal features. The initial local fusion features are then processed by local feature extraction based on the preset first self-attention component to obtain the updated local fusion features.

[0018] The updated local fusion feature is used as the new initial local fusion feature. The new initial local fusion feature is then subjected to at least one more local feature extraction process until the preset first condition is met. The last updated local fusion feature is then used as the local spatial feature at the fine-grained level.

[0019] In some embodiments, a second self-attention component is provided in the prediction model;

[0020] Global feature encoding is performed on the spatial morphology map at a coarse-grained level to obtain global spatial features, including:

[0021] Global feature encoding is performed on the spatial morphology map to obtain initial global spatial features;

[0022] The initial global spatial features are processed by global feature extraction based on the second self-attention component to obtain updated global spatial features. The updated global spatial features are used as new initial global spatial features. The new initial global spatial features are processed by global feature extraction at least once more until the preset second condition is met. The last updated global spatial features are then used as global spatial features at the coarse-grained level.

[0023] In some embodiments, the prediction model includes a cross-attention component and a multilayer perceptron.

[0024] Multi-granularity coding features are obtained by fusing local and global spatial features at different granularity levels, including:

[0025] The query vector is determined based on local spatial features, and the key vector and value vector are determined based on global spatial features.

[0026] Based on the cross-attention component, cross-attention calculation is performed on the query vector, key vector, and value vector to obtain cross-attention features;

[0027] Based on a multilayer perceptron, nonlinear transformation is applied to the cross-attention features to obtain multi-granularity encoded features.

[0028] In some embodiments, the prediction model is trained through the following steps, including:

[0029] The initial prediction model, the sample spatial morphology map characterizing the signal occlusion in the sample area, and the real radio map corresponding to the sample spatial morphology map are obtained. Among them, the observation locations of multiple samples are sparsely distributed.

[0030] Multiple sample observation locations are selected in the sample area. For each sample observation location, the coordinates of the sample observation location on the sample spatial morphology map are determined, and the sample signal intensity value of the sample observation location is measured.

[0031] The sample spatial morphology map, sample observation location coordinates, and sample signal intensity values ​​are input into the initial prediction model. Local feature encoding is performed on the sample observation location coordinates and sample signal intensity values ​​to obtain the sample local spatial features at the fine-grained level. Global feature encoding is performed on the sample spatial morphology map to obtain the sample global spatial features at the coarse-grained level.

[0032] The sample multi-granularity coding features are obtained by fusing the local spatial features and global spatial features of the samples at different granularity levels. The sample multi-granularity coding features are then decoded to obtain sample multi-granularity decoding features. Based on the sample multi-granularity decoding features, a sample radio map representing the signal intensity at each location in the sample area is predicted.

[0033] The sample loss value is determined based on the sample radio map and the real radio map. The model parameters of the initial prediction model are then adjusted based on the sample loss value to obtain the trained prediction model.

[0034] In some embodiments, determining the sample loss value based on the sample radio map and the real radio map includes:

[0035] For any sample observation location, determine the predicted signal strength value of the sample observation location on the sample radio map, and determine the actual signal strength value of the sample observation location on the real radio map;

[0036] Calculate the squared difference between the predicted signal strength value and the actual signal strength value, and determine the sample loss value based on the squared difference.

[0037] To achieve the above objectives, a second aspect of this application provides a radio map prediction apparatus, the apparatus comprising:

[0038] The acquisition module is used to acquire a spatial morphological map representing the signal occlusion situation in the target area, and select multiple target observation locations in the target area, wherein the multiple target observation locations are sparsely distributed;

[0039] The observation location information determination module is used to determine the observation location coordinates of each target observation location on the spatial morphology map and to measure the signal strength value of the target observation location.

[0040] The encoding processing module is used to input the spatial morphology map, observation location coordinates and corresponding signal intensity values ​​into the trained prediction model, so as to perform local feature encoding processing on the observation location coordinates and signal intensity values ​​at a fine-grained level to obtain local spatial features, and perform global feature encoding processing on the spatial morphology map at a coarse-grained level to obtain global spatial features.

[0041] The target prediction module is used to fuse local and global spatial features at different granularity levels to obtain multi-granularity encoded features, and to decode the multi-granularity encoded features to obtain multi-granularity decoded features. Based on the multi-granularity decoded features, a target radio map representing the signal strength at various locations in the target area is predicted.

[0042] To achieve the above objectives, a third aspect of the present application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the radio map prediction method of the first aspect described above.

[0043] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the radio map prediction method of the first aspect described above.

[0044] The radio map prediction method, apparatus, electronic device, and storage medium proposed in this application acquire a spatial morphological map representing the signal obstruction situation in a target area, and select multiple target observation locations in the target area, wherein the multiple target observation locations are sparsely distributed. For each target observation location, the coordinates of the observation location on the spatial morphological map are determined, and the signal strength value of the target observation location is measured. The spatial morphological map, observation location coordinates, and corresponding signal strength values ​​are input into a trained prediction model to perform local feature encoding processing on the observation location coordinates and signal strength values ​​at a fine-grained level to obtain local spatial features, and global feature encoding processing on the spatial morphological map at a coarse-grained level to obtain global spatial features. The local spatial features and global spatial features at different granularities are fused to obtain multi-granularity encoded features, and the multi-granularity encoded features are decoded to obtain multi-granularity decoded features. Based on the multi-granularity decoded features, a target radio map representing the signal strength of each location in the target area is predicted. This application can improve the accuracy of the predicted radio map when the collected observation information is sparse. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of an optional implementation environment for the radio map prediction device provided in this application embodiment;

[0046] Figure 2 This is an optional flowchart of the radio map prediction method provided in the embodiments of this application;

[0047] Figure 3 yes Figure 2 Step 101 in the flowchart is an optional implementation.

[0048] Figure 4 This is an example of an optional spatial morphology map illustrating the radio map prediction method provided in this application embodiment;

[0049] Figure 5 This is a schematic diagram of an optional boundary location for the radio map prediction method provided in this application embodiment;

[0050] Figure 6 This is an optional prediction processing flow and training diagram of the radio map prediction method provided in this application embodiment;

[0051] Figure 7 yes Figure 2 Step 103 is an optional implementation flowchart;

[0052] Figure 8 yes Figure 2 Another optional implementation flowchart for step 103 in the diagram;

[0053] Figure 9 yes Figure 2 Step 104 in the flowchart is an optional implementation.

[0054] Figure 10 This is a schematic diagram of an optional prediction model training process for the radio map prediction method provided in this application embodiment;

[0055] Figure 11 yes Figure 10 Step A.5 is an optional implementation flowchart;

[0056] Figure 12 This is a schematic diagram illustrating an optional radio map generation method provided in this application embodiment for radio map prediction.

[0057] Figure 13 This is a schematic diagram of an optional radio map generation result using the radio map prediction method provided in this application embodiment;

[0058] Figure 14 This is a schematic diagram of an optional device module of the radio map prediction device provided in the embodiments of this application;

[0059] Figure 15 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0060] 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.

[0061] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0063] First, let's analyze some of the terms used in this application:

[0064] The Self-Attention Transformer module is a key component in deep learning architectures, primarily used to capture complex dependencies between elements in input data. It dynamically emphasizes or diminishes the importance of different elements to the currently processed element by calculating the relevance weights between each element and all other elements (i.e., the self-attention mechanism), thereby achieving effective information integration and feature extraction. This mechanism makes the Self-Attention Transformer module perform exceptionally well when processing sequential data. It is widely used in fields such as natural language processing and image recognition, and is particularly adept at handling long-distance dependency problems. It also supports parallel computation to improve training efficiency.

[0065] Next, the technical background related to the embodiments of this application will be introduced:

[0066] A radio map is a spatial information used to describe the electromagnetic spectrum coverage and intensity within a specific area. Radio maps can help relevant personnel understand the signal quality, interference situation, and network coverage effectiveness within a specific area.

[0067] In related technologies, radio maps are predicted using observation information at multiple pixel levels within a specific area, where dense observation information almost covers the entire area. However, in practical applications, only observation information from a limited number of locations can often be collected. Therefore, the accuracy of radio maps predicted using sparse observation information is relatively poor.

[0068] Based on this, embodiments of this application provide a radio map prediction method, apparatus, electronic device, and storage medium, which can improve the accuracy of the predicted radio map when the collected observation information is sparse.

[0069] For example, such as Figure 1 As shown, Figure 1This is a schematic diagram of an optional implementation environment for the radio map prediction device provided in this application embodiment. The implementation environment includes a client 11 and a server 12. The radio map prediction method proposed in this application embodiment (which can also be simply referred to as the "prediction device" for ease of description) is deployed on the server 12. The client 11 sends the selected target observation location and a spatial morphology map representing the signal obstruction situation in the target area to the server 12. Then, the server 12 acquires the data sent by the client 11, wherein multiple target observation locations are sparsely distributed. For each target observation location, the server determines the location of the target observation location on the corresponding observation location in the spatial morphology map. The method involves marking and measuring the signal strength at the target observation location; inputting the spatial morphology map, observation location coordinates, and corresponding signal strength values ​​into the trained prediction model; performing local feature encoding on the observation location coordinates and signal strength values ​​at a fine-grained level to obtain local spatial features; and performing global feature encoding on the spatial morphology map at a coarse-grained level to obtain global spatial features; fusing local and global spatial features at different granularities to obtain multi-granular coded features; and decoding these multi-granular coded features to obtain multi-granular decoded features; and predicting a target radio map representing the signal strength at various locations within the target area based on these multi-granular decoded features. Thus, this embodiment improves the sensitivity to local signal changes and prediction accuracy by encoding local spatial features at a fine-grained level; enhances the ability to grasp the overall signal distribution trend by extracting global spatial features at a coarse-grained level; and significantly improves the accuracy and spatial resolution of radio map prediction by fusing and decoding multi-granular features, thereby providing high-quality data support for applications such as wireless network optimization and signal coverage analysis.

[0070] The server 12 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Additionally, the server 12 can also be a node server in a blockchain network. The client 11 can be a mobile phone, computer, smart voice interaction device, smart wearable device, smart home appliance, in-vehicle terminal, etc., but is not limited to these. The client 11 and the server 12 can be connected directly or indirectly through wired or wireless communication, which is not limited in this embodiment.

[0071] It should be noted that in this application embodiment, when information related to user characteristics, such as basic user information or user identity, is required, the user's permission or consent will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. In addition, when this application embodiment needs to obtain sensitive personal information of a user, the user's individual permission or consent will be obtained first. Only after obtaining the user's individual permission or consent will the necessary data for the normal operation of this application embodiment be obtained. For example, before obtaining a spatial morphology map, the authorization or consent of relevant personnel will be obtained; otherwise, a spatial morphology map that cannot be used in this application embodiment will be obtained. Furthermore, other relevant data obtained by the optimization device in this application are all authorized data, and will not be elaborated upon here.

[0072] In this application embodiment, the description will focus on the prediction device, which can be integrated into a computer device, such as a server. Figure 2 As shown, Figure 2 This is an optional flowchart of the radio map prediction method provided in the embodiments of this application. Figure 2 The method may include, but is not limited to, the following steps 101 to 104. When the prediction device executes the radio map prediction method, the specific process is as follows. It should be noted that this embodiment... Figure 2 The order of steps 101 to 104 is not specifically limited. The order of steps can be adjusted or some steps can be reduced or added according to actual needs.

[0073] Step 101: Obtain a spatial morphological map representing the signal occlusion situation in the target area, and select multiple target observation locations in the target area, wherein the multiple target observation locations are sparsely distributed.

[0074] Step 101 will be described in detail below.

[0075] In this context, the target area refers to the specific geographic area that requires signal strength analysis and modeling in radio map prediction tasks. In practical applications, the target area is usually not flat but includes environmental factors that affect radio signal propagation, such as man-made structures and natural vegetation.

[0076] A spatial morphology map is a map used to describe signal obstruction within a target area. It is typically generated based on information such as terrain and building distribution, and reflects the signal propagation characteristics within the target area. Signal refers to the received signal strength or related radio parameters in the electromagnetic spectrum, including but not limited to key information such as received signal power, interference power, power spectral density, and channel gain. The electromagnetic spectrum, as the core carrier of low-altitude target activity, carries critical information such as communication, navigation, and control; its global perception and situational modeling capabilities directly determine the effectiveness of low-altitude surveillance. It should be noted that a spatial morphology map only indicates which objects within the target area might obstruct signals; it does not mean that the spatial morphology map already indicates the signal strength at every location within the target area.

[0077] Furthermore, to facilitate data processing by predictive devices, spatial morphology maps are typically binary maps, where '0' represents unobstructed areas and '1' represents obstructed areas. Unobstructed areas refer to regions that do not significantly obstruct or attenuate the propagation of radio signals, including but not limited to open fields, open roads, water surfaces, and spaces without tall buildings or terrain obstacles blocking signal transmission. Signals can propagate relatively freely within unobstructed areas. Obstructed areas refer to geographical or physical features that can obstruct or attenuate signal propagation, including but not limited to buildings (such as high-rise buildings, residential areas, commercial centers, and other complex building complexes), terrain obstacles (such as mountains, hills, and other undulating terrain), and vegetation cover (such as dense forests, trees in large parks, and other vegetation).

[0078] In this embodiment, the target observation location refers to the specific geographical location selected for measuring key parameters such as the received detailed signal strength during the radio map prediction task. In practical applications, limitations in deployment costs, insufficient measurement equipment, and complex environments make it difficult to deploy dense target observation locations. For example, in wide-area monitoring or low-altitude surveillance scenarios, the scale and energy consumption requirements of the sensor network often limit the number of observation nodes within the target area. Furthermore, in special situations such as disaster emergency communication reconnaissance, it may be impossible to pre-deploy sufficient measurement equipment. This extremely sparse spatial sampling condition results in very few available observation sampling points. Therefore, the multiple target observation locations selected in this embodiment are sparsely distributed within the target area, but the prediction device can use the data acquired from them to determine the local spatial characteristics of the target area, and the spatial morphology. Figure 1 This will provide a basis for the subsequent construction of accurate radio environment maps.

[0079] In some embodiments, such as Figure 3 As shown, Figure 3 yes Figure 2Step 101, an optional implementation flowchart, selects multiple target observation locations in the target area, including the following steps:

[0080] 101.1.1 Perform spatial analysis on the spatial morphology map to determine the land cover area and open ground area of ​​the spatial morphology map;

[0081] 101.1.2 Divide the spatial morphology map into multiple initial regions, and select at least one candidate region from the multiple initial regions, wherein the area of ​​the open ground area in the candidate region is within a preset area threshold range;

[0082] 101.1.3 For each candidate region, at least one boundary location between the open ground area and the land cover area in the candidate region shall be determined as the target observation location, wherein the boundary location is close to the corner of the land cover area.

[0083] Steps 101.1.1 to 101.1.3 are described in detail below.

[0084] In some embodiments, the selection of a limited number of target observation locations is crucial because the data subsequently measured at these locations directly affects the accuracy of the generated radio map. Reasonable target observation locations can effectively cover the key features of the target area and provide a comprehensive overview, allowing for the inference of global signal propagation information based on the signal propagation information at that location. Furthermore, compared to randomly selected observation points in traditional methods, the target observation locations selected in this embodiment are more representative of the target area. With limited computational resources, this embodiment reduces the computation of invalid data by selecting more critical target observation locations, thus avoiding wasted computational resources and helping to improve the accuracy of the subsequently predicted radio map.

[0085] Furthermore, in the embodiments of this application, the spatial morphology map M is obtained. b Next, the spatial morphology map M will be analyzed first. b Spatial analysis is conducted to clarify the spatial morphology of map M. b The map defines the areas covered by features and the open ground areas. Covered areas refer to the regions on the spatial map that are obscured by features such as buildings, trees, and hills; while open ground areas refer to the regions on the spatial map that are almost entirely unobstructed. Open ground areas typically offer better signal propagation conditions. Figure 4 As shown, Figure 4 This is an optional spatial morphology map illustration of the radio map prediction method provided in this application embodiment. It is a spatial morphology map corresponding to a certain target area, wherein the white part is the area covered by ground features and the black part is the open ground area.

[0086] Furthermore, after completing the spatial morphology map M b Following spatial analysis, the spatial morphology map M will be analyzed. b The process involves segmenting the data into multiple initial regions. Then, in this embodiment, candidate regions are selected by comparing the area of ​​an open ground region with a preset area threshold. Specifically, the initial regions obtained through segmentation include both land cover areas and open ground areas. If the area of ​​the open ground area in the initial region is too large, the lack of physical conditions conducive to signal propagation (reflection and scattering) will lead to incomplete radio signal coverage or degraded quality. Conversely, if the area of ​​the open ground area in the initial region is too small, there may be significant obstruction and reflection between obstructions (such as buildings), resulting in complex signal propagation paths and unstable signal strength and quality. Therefore, initial regions with excessively large or small open ground areas are not considered in this embodiment; only initial regions with open ground areas within the preset area threshold range are selected as candidate regions.

[0087] Furthermore, based on the number of selectable target observation locations in the actual situation, target observation locations are selected from the candidate regions. Specifically, if the number of target observation locations equals the number of candidate regions, one target observation location can be selected from each candidate region; if the number of target observation locations is greater than the number of candidate regions, candidate regions can be randomly selected, and one target observation location can be selected from each of the selected candidate regions; if the number of target observation locations is less than the number of candidate regions, one target observation location can first be selected from each candidate region, then candidate regions can be randomly selected, and the remaining number of target observation locations can be selected from the selected candidate regions. Of course, the method for selecting target observation locations from the candidate regions can be set according to the actual situation. This is only an example and does not represent a limitation of the embodiments of this application.

[0088] Furthermore, when determining the target observation location in any candidate region, the intersection between the open ground area and the area covered by ground features is first identified as the boundary location, and the target observation location is selected from these boundary locations. When the number of selectable target observation locations is limited, this method maximizes the signal propagation advantage of the open ground area while effectively capturing the influence characteristics of the covered area. Consequently, the data subsequently measured based on this target observation location will be more representative. In extreme observation environments, the radio map generated using this location-related data has higher accuracy compared to traditional methods.

[0089] Furthermore, in the process of selecting target observation locations from the boundary locations, the boundary locations near the corners of the area covered by ground features can also be used as target observation locations. For example... Figure 5 As shown, Figure 5 This is a schematic diagram of an optional boundary location for the radio map prediction method provided in this application embodiment. Figure 5 The dashed rectangle represents the spatial topography map, the solid rectangle represents the area covered by features, and the area between the dashed and solid rectangles represents the open ground area. The four solid lines of the solid rectangle represent the boundary between the open ground area and the area covered by features, while the four corners of the solid rectangle represent the corners of the area covered by features (for simplicity). Figure 5 (Only one corner is shown in the example; the other three are not marked.) Corners are the intersection of multiple boundary locations. While boundary locations can provide complex signal characteristics and environmental information, corners are more effective at capturing signal interactions between open ground areas and areas covered by ground features. Therefore, choosing a location near a corner as a target observation location is more representative and can provide more crucial information for subsequent radio map generation.

[0090] Step 102: For each target observation location, determine the observation location coordinates on the spatial morphology map corresponding to the target observation location, and measure the signal strength value of the target observation location.

[0091] Step 102 is described in detail below.

[0092] In some embodiments, after selecting multiple target observation locations, for each target observation location, the observation location coordinates and signal strength values ​​corresponding to that target observation location are collected, so that the information at the fine-grained level obtained from the target observation location can be combined and analyzed with the spatial morphology map determined at the coarse-grained level to generate a high-precision radio map under limited information and computing resources.

[0093] Furthermore, to determine the corresponding observation location coordinates of the target observation location on the spatial morphology map, a three-dimensional spatial coordinate system can be established first, and the spatial morphology map can be gridded, defining the center point of each grid as the observation location. Next, the geographic coordinates (such as latitude and longitude) of the target observation location are converted to the coordinate system of the spatial morphology map, and an interpolation algorithm (such as nearest neighbor interpolation or bilinear interpolation) is used to determine the specific coordinates of the target observation location on the map. Then, the signal strength is measured at the observation location using a signal receiving device, and the corresponding signal strength value is recorded for subsequent analysis and processing.

[0094] Furthermore, signal strength measurement can be achieved using a signal device: at the target observation location, the signal device is placed at a predetermined height and orientation to ensure accurate signal reception. The signal device can be a signal receiver or a radio spectrum analyzer, or adjusted according to actual conditions; this application embodiment does not impose any limitations on this. Moreover, multiple measurements can be taken, recording multiple initial signal strength values ​​at different time periods, and averaging these multiple initial signal strength values ​​to obtain the signal strength value at the target observation location.

[0095] Step 103: Input the spatial morphology map, observation location coordinates, and corresponding signal intensity values ​​into the trained prediction model to perform local feature encoding on the observation location coordinates and signal intensity values ​​at a fine-grained level to obtain local spatial features, and to perform global feature encoding on the spatial morphology map at a coarse-grained level to obtain global spatial features.

[0096] Step 103 will be described in detail below.

[0097] In some embodiments, such as Figure 6 As shown, Figure 6 This application provides an optional prediction processing flow and training diagram for the radio map prediction method, which combines a spatial morphology map and the observation position coordinates S determined at the target observation position. C (i.e. Figure 6 (x) k ,y k )), and the corresponding signal strength value S V (i.e. Figure 6 values ​​in k The information is input into the trained prediction model, which will perform feature encoding on the information at different levels and obtain the corresponding spatial features at different levels. This allows for the subsequent fusion and supplementation of the spatial features obtained from multiple levels to obtain multi-granularity encoded features, resulting in a radio map with high stability and strong anti-interference capability.

[0098] Furthermore, as shown in the figure, the multi-granularity feature extraction module in this embodiment adopts a dual-stream structure, processing pixel-level information (observation location coordinates and signal strength values) and block-level information (spatial morphology map) respectively: the pixel-level branch is used to capture fine-grained correlations between observation points, while the block-level branch focuses on the geometric shape of the ground feature distribution to reduce redundant computation under sparse observation conditions. Furthermore, through the dual-stream mechanism, the multi-granularity feature extraction module deployed in the prediction device can simultaneously learn local spatial features F. O With global spatial features F B This enhances the ability to model electromagnetic environments.

[0099] Next, let's introduce the pixel-level branch:

[0100] In some embodiments, such as Figure 7 As shown, Figure 7 yes Figure 2 Step 103, an optional implementation flowchart, involves performing local feature encoding on the observation location coordinates and signal intensity values ​​at a fine-grained level to obtain local spatial features, including the following steps:

[0101] 103.1.1 Local feature encoding is performed on the observation location coordinates and signal strength values ​​respectively to obtain the initial observation location features and the initial observation signal features;

[0102] 103.1.2 Initial local fusion features are obtained by superimposing initial observation location features and initial observation signal features. Local feature extraction processing is performed on the initial local fusion features based on the preset first self-attention component to obtain updated local fusion features.

[0103] 103.1.3 The updated local fusion feature is used as the new initial local fusion feature. The new initial local fusion feature is subjected to at least one more local feature extraction process until the preset first condition is met. The last updated local fusion feature is then used as the local spatial feature at the fine-grained level.

[0104] Steps 103.1.1 to 103.1.3 are described in detail below.

[0105] In some embodiments, two linear layers θ are used c and θ v For the observation position coordinates S respectively C and signal strength value S V Local feature encoding is performed to obtain initial observation location features and initial observation signal features; then, the two features are fused by superposition to obtain initial local fused features. Next, based on the first self-attention component, the initial local fusion features are... Local feature extraction is performed to obtain the current updated local fusion feature. Then, the updated local fusion feature is used as the new initial local fusion feature, and at least one more local feature extraction is performed on the new initial local fusion feature until a preset first condition is met. Finally, the last obtained updated local fusion feature is used as the local spatial feature F at the fine-grained level. O This process can be described using a formula. <1> Represented as:

[0106]

[0107] Where, φ OThis is the first self-attention component; for the nth iteration, the current initial local fusion feature is... Based on the first self-attention component, the current initial local fusion features are... Perform local feature extraction to obtain the updated local fusion features in the current iteration n. The number of iterations n is within the preset threshold range [1, N]. o ] within, N o This refers to the preset number of local iteration rounds.

[0108] The first self-attention component is essentially a self-attention Transformer module. It dynamically generates weights by calculating the similarity between elements in the input sequence, thus allowing the prediction model to focus on more important information. The first self-attention component typically consists of multiple self-attention layers and a feedforward neural network, enabling parallel processing of sequence data and generating more accurate local spatial features by capturing long-range dependencies.

[0109] In this embodiment of the application, the first condition is that the number of local feature extraction processes reaches a preset number of local iteration processing rounds N. o Of course, the first condition can also be that the amount of feature change during the local feature extraction process is lower than a certain threshold, the stability of updating the local fusion features reaches a certain level, or the performance indicators (such as accuracy, loss value, etc.) of the first self-attention model on the validation set meet predetermined requirements. The first condition can be adjusted according to the actual situation, and the embodiments of this application do not limit it.

[0110] Furthermore, compared to traditional methods that require collecting information from observation points within each grid of a spatial morphology map to form dense observation information, this application embodiment significantly reduces the need for dense observation information by focusing on the most representative observation information for local feature encoding processing. This allows limited computing resources to be used for processing key observation information, reducing processing complexity and resource requirements while improving the ability to capture important local spatial features.

[0111] Next, we will introduce block-level branches:

[0112] In some embodiments, such as Figure 8 As shown, Figure 8 yes Figure 2 Another optional implementation flowchart for step 103 involves performing global feature encoding on the spatial morphology map at a coarse-grained level to obtain global spatial features, including the following steps:

[0113] 103.2.1 Perform global feature encoding on the spatial morphology map to obtain initial global spatial features;

[0114] 103.2.2 Based on the second self-attention component, global feature extraction processing is performed on the initial global spatial features to obtain updated global spatial features. The updated global spatial features are used as new initial global spatial features. Global feature extraction processing is performed on the new initial global spatial features at least once more until the preset second condition is met. The last updated global spatial features are then used as global spatial features at the coarse-grained level.

[0115] Steps 103.2.1 to 103.2.2 are described in detail below.

[0116] In some embodiments, a linear layer θ is first utilized. b For multiple initial region sets C of the spatial morphology map b Perform linear processing, and simultaneously perform position embedding E. p To obtain the initial global spatial features Next, based on the second self-attention component, the initial global spatial features are... Global feature extraction is performed to obtain updated global spatial features. These updated global spatial features are then used as new initial global spatial features. Global feature extraction is then performed on these new initial global spatial features at least once more, until a preset second condition is met. The last updated global spatial feature is then used as the global spatial feature F at the coarse-grained level. g This process can be described using a formula. <2> Represented as:

[0117]

[0118] Where, φ b This is the second self-attention component; for the nth iteration, the current initial global fusion feature is... Based on the second self-attention component, the current initial global fusion features are... Perform global feature extraction to obtain the updated global fusion features for the current iteration n. The number of iterations n is within the preset threshold range [1, N]. b ] within, N b This is the preset number of global iteration rounds.

[0119] The prediction model also includes a second self-attention component. This second self-attention component functions similarly to the first one; the terms "first" and "second" are merely distinguishing terms and do not imply any special meaning. The second self-attention component is specifically used for global feature encoding of the spatial morphology map. Essentially, it is also a type of self-attention Transformer module. It dynamically generates weights by calculating the similarity between elements in the input sequence, allowing the prediction model to focus on more important information. The second self-attention component typically consists of multiple self-attention layers and a feedforward neural network, enabling parallel processing of sequence data and generating more accurate local spatial features by capturing long-range dependencies.

[0120] In this embodiment of the application, the second condition is that the number of global feature extraction processes reaches a preset number of global iteration processing rounds N. o Of course, the second condition can also be that the amount of feature change during the global feature extraction process is lower than a certain threshold, the stability of updating the global fusion features reaches a certain level, or the performance indicators (such as accuracy, loss value, etc.) of the second self-attention model on the validation set meet predetermined requirements. The second condition can be adjusted according to the actual situation, and the embodiments of this application do not limit it.

[0121] In one optional implementation, a standard visual transformer (ViT) encoder using the Sinusoidal position encoding method can be employed to process the spatial morphology map and extract global building spatial features F. B Among them, Sinusoidal positional encoding uses sine and cosine functions to generate a set of vectors that provide a unique representation for each input position, helping the prediction device capture the relative positional relationships between elements in the input data; ViT is a deep learning model based on the Transformer architecture. ViT segments the input image into fixed-size patches, flattens these patches, and uses them as the input sequence. It uses a self-attention mechanism to capture the relationships between different regions in the image, thereby effectively learning the global features and contextual information of the image.

[0122] Step 104: Fuse local spatial features and global spatial features at different granularity levels to obtain multi-granularity encoded features, and decode the multi-granularity encoded features to obtain multi-granularity decoded features. Based on the multi-granularity decoded features, predict the target radio map representing the signal strength at each location in the target area.

[0123] Step 104 is described in detail below.

[0124] In some embodiments, such as Figure 9 As shown, Figure 9 yes Figure 2 Step 104 is an optional implementation flowchart. By fusing local and global spatial features at different granular levels, multi-granularity coded features are generated, thereby effectively capturing the details and overall information of the target area. Multi-granularity decoding features can more accurately characterize the signal strength at various locations in the target area, thus helping to generate high-quality target radio maps in the future.

[0125] In some embodiments, fusing local spatial features and global spatial features at different granularity levels to obtain multi-granularity encoded features includes the following steps:

[0126] 104.1.1 Determine the query vector based on local spatial features, and determine the key vector and value vector based on global spatial features;

[0127] 104.1.2 Based on the cross-attention component, cross-attention calculation is performed on the query vector, key vector, and value vector to obtain cross-attention features;

[0128] 104.1.3 Based on the multilayer perceptron, nonlinear transformation processing is performed on the cross-attention features to obtain multi-granularity coding features.

[0129] Steps 104.1.1 to 104.1.3 are described in detail below.

[0130] In some embodiments, the prediction model includes a multimodal feature fusion module, which is used to align and fuse features of different granularities. Through cross-scale information interaction, the model's understanding of spectral features at different scales is enhanced, enabling it to effectively infer the electromagnetic distribution of unobserved regions even under extremely sparse sampling conditions, thereby improving the robustness of spectral estimation.

[0131] Furthermore, the multimodal feature fusion module is equipped with a cross-attention component and a multilayer perceptron. The multimodal feature fusion module is based on the formula... <3> Multi-granularity encoded features are obtained by fusing local and global spatial features at different granularity levels.

[0132]

[0133] Where Q is the query vector, K is the key vector, and V is the value vector; This represents the cross-attention mechanism, where X and Y represent two different sequences; the global spatial features F... B As a formula <3> Y in the local spatial feature F O As X; The first feature is the cross-attention feature; then, the cross-attention feature is processed by nonlinear transformation based on the multilayer perceptron (MLP) to obtain the multi-granularity coding feature F.

[0134] Among them, the Multilayer Perceptron (MLP) is a basic feedforward neural network composed of multiple layers of neurons, including an input layer, one or more hidden layers, and an output layer. Each neuron is connected to the neurons in the previous layer through an activation function, enabling it to learn the nonlinear characteristics of the input data. MLPs are typically trained using the backpropagation algorithm, which adjusts the weights to minimize the error between the predicted output and the actual target.

[0135] Finally, the decoder module is used to decode the integrated multi-granularity features F into a radio map M′. Unlike traditional point-by-point inference methods, this embodiment constructs a radio spectrum distribution based on global features, which can achieve efficient and accurate radio map estimation under extremely sparse spatial sampling conditions while maintaining low computational cost, and has excellent generalization ability. Compared with traditional methods, the main advantages of this embodiment are: (1) It can adapt to extremely sparse data sampling, and under conditions where only... (1) Even with sampled pixels, it can still accurately infer the global spectrum distribution; (2) It performs efficient multi-scale information fusion through multi-granularity Transformer feature extraction and multi-modal feature fusion module, making full use of building geometric information and improving the model's adaptability to complex environments; (3) It avoids the high computational overhead of traditional point-by-point reasoning and achieves high-precision estimation with low computational cost.

[0136] In some embodiments, such as Figure 10 As shown, Figure 10 This is a schematic diagram of an optional prediction model training process for the radio map prediction method provided in this application embodiment. The prediction model is trained through the following steps:

[0137] A.1 Obtain the initial prediction model, the sample spatial morphology map characterizing the signal occlusion in the sample area, and the real radio map corresponding to the sample spatial morphology map, wherein multiple sample observation locations are sparsely distributed.

[0138] A.2 Select multiple sample observation locations in the sample area. For each sample observation location, determine the coordinates of the sample observation location on the sample spatial morphology map and measure the sample signal intensity value at the sample observation location.

[0139] A.3 Input the sample spatial morphology map, sample observation location coordinates, and the sample signal intensity value into the initial prediction model. Perform local feature encoding on the sample observation location coordinates and sample signal intensity value to obtain the sample local spatial features at the fine-grained level. Perform global feature encoding on the sample spatial morphology map to obtain the sample global spatial features at the coarse-grained level.

[0140] A.4 The sample local spatial features and sample global spatial features at different granularity levels are fused to obtain sample multi-granularity coding features, and the sample multi-granularity coding features are decoded to obtain sample multi-granularity decoding features. Based on the sample multi-granularity decoding features, a sample radio map representing the signal strength at each location in the sample area is predicted.

[0141] A.5 Determine the sample loss value based on the sample radio map and the real radio map, and adjust the model parameters of the initial prediction model according to the sample loss value to obtain the trained prediction model.

[0142] Steps A.1 to A.5 are described in detail below.

[0143] In some embodiments, this application employs self-supervised training of the initial prediction model to obtain a trained prediction model that is better suited for practical application. To achieve this, this application acquires a real radio map. A real radio map is a map reflecting the actual distribution of radio signal strength within a specific target area, based on data actually measured and collected. It includes the influence of factors such as terrain, buildings, and vegetation on the signal, and accurately represents the radio signal strength received at different locations. The real radio map provides an important reference for training the initial prediction model.

[0144] Furthermore, the specific implementation methods of steps 104.2.1 to 104.2.4 are similar to those of steps 101 to 104. However, the data used in the training process is labeled with the word "sample" to distinguish it from the application process. Therefore, they will not be described in detail here.

[0145] Furthermore, during training, after obtaining the sample radio map corresponding to the sample area, the sample loss value is calculated based on the sample radio map and the real radio map. This loss value is then used to adjust the model parameters of the initial prediction model, resulting in the trained prediction model. The trained prediction model can process observation data obtained from the spatial morphology map of the target area and sparse target observation locations, thus enabling it to output a high-precision radio map even with limited observation information.

[0146] The adjusted model parameters typically include various parameters related to the model structure and learning process, such as weights and biases, which determine how the model maps input data to output results. Additionally, they may include hyperparameters such as learning rate, regularization coefficient, and activation function selection, which affect the model's training speed and generalization ability. Of course, the specific model parameters adjusted can be set according to actual circumstances, and this application embodiment does not impose any limitations on this.

[0147] In some embodiments, such as Figure 11 As shown, Figure 11 yes Figure 10 Step A.5, an optional implementation flowchart, determines the sample loss value based on the sample radio map and the real radio map, including the following steps:

[0148] B.1 For any sample observation location, determine the predicted signal strength value of the sample observation location on the sample radio map, and determine the actual signal strength value of the sample observation location on the real radio map;

[0149] B.2 Calculate the squared difference between the predicted signal strength value and the actual signal strength value, and determine the sample loss value based on the squared difference.

[0150] Steps B.1 to B.2 are described in detail below.

[0151] In some embodiments, according to the following formula <4> Calculate the sample loss value:

[0152]

[0153] Among them, L MSE M′ is the Mean Squared Error (MSE) function; i M represents the predicted signal strength value of the i-th sample observation location on the sample radio map; i is the actual signal strength value of the i-th sample observation location on the real radio map; N is the total number of points.

[0154] To help readers better understand the beneficial effects of the embodiments of this application, the following experiments compare the embodiments of this application with traditional methods:

[0155] (1) As Figure 12 As shown, Figure 12 This is a schematic diagram illustrating an optional radio map generation method provided in this application's embodiment for radio map prediction. Figure 12 This paper presents a comparison of the performance of various deep learning models in traditional methods for generating radio maps at different spatial sampling rates. The inputs are the input samples, with a global size of 256*256 and approximately 5 sampling points. The presentation sequentially showcases radio maps generated by traditional methods including the Convolutional Block Attention Module (CBAM), the Swin Transformer UNet (Swin-UNet), the U-shaped network (UNet), the Radio-UNet for radio signal processing, and the Pixel-Transformer (PiT) (displayed as heatmaps). RadioFormer generates a radio map for this embodiment (also displayed as a heatmap). The colors of the radio maps, from blue to yellow, represent changes in generation probability or intensity. It can be seen that as the number of sampling points increases (from 5 to 1000), the RadioFormer model performs particularly well across all sampling rates, with its generated results most closely resembling the ground truth (GT).

[0156] (2) Figure 13 As shown, Figure 13 This is a schematic diagram of an optional radio map generation result from the radio map prediction method provided in this application embodiment. Figure 13 The paper presents a comparison of the generation performance between traditional methods and the radio map prediction method (RadioFormer) proposed in this application under different sampling rates (Observation Numbers). Here, RMSE represents the root mean square error, and the smaller the value, the better.

[0157] (3) As shown in Table 1, Table 1 is a comparative schematic diagram of the experimental results of an optional quantitative analysis of the radio map prediction method provided in the embodiments of this application. In this table, RMSE represents the root mean square error (the smaller the value, the better), SSIM represents the structural similarity index (the larger the value, the better), and PSNR represents the peak signal-to-noise ratio (the larger the value, the better). It can be found that the RadioFormer model used in the embodiments of this application performs better than the traditional method.

[0158] Table 1

[0159]

[0160] (4) As shown in Table 2, Table 2 is a schematic diagram of optional computational data of the radio map prediction method provided in the embodiments of this application. In this table, Flops, Parameters, and Inference Time are the computational amount, parameter amount, and inference speed of the model, respectively. It can be seen that the RadioFormer model used in the embodiments of this application performs well in terms of computational complexity, parameter amount, and inference time. It is suitable for application in real-world scenarios with high requirements for computational resources and inference speed.

[0161] Table 2

[0162]

[0163] like Figure 14 As shown, Figure 14 This is a schematic diagram of an optional device module of the radio map prediction device provided in this application embodiment. The radio map prediction device may include the following modules 201 to 204:

[0164] The acquisition module 201 is used to acquire a spatial morphological map representing the signal occlusion situation in the target area, and select multiple target observation locations in the target area, wherein the multiple target observation locations are sparsely distributed;

[0165] The observation location information determination module 202 is used to determine the observation location coordinates of each target observation location on the spatial morphology map and to measure the signal strength value of the target observation location.

[0166] The encoding processing module 203 is used to input the spatial morphology map, the observation location coordinates and the corresponding signal intensity values ​​into the trained prediction model, so as to perform local feature encoding processing on the observation location coordinates and signal intensity values ​​at a fine-grained level to obtain local spatial features, and to perform global feature encoding processing on the spatial morphology map at a coarse-grained level to obtain global spatial features.

[0167] The target prediction module 204 is used to fuse local spatial features and global spatial features at different granularity levels to obtain multi-granularity encoded features, and to decode the multi-granularity encoded features to obtain multi-granularity decoded features. Based on the multi-granularity decoded features, a target radio map representing the signal strength at each location in the target area is predicted.

[0168] The specific implementation of this radio map prediction device is basically the same as the specific implementation of the radio map prediction method described above, and will not be repeated here.

[0169] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described radio map prediction method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0170] like Figure 15 As shown, Figure 15 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device includes:

[0171] The processor 301 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0172] The memory 302 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 302 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 302 and called by the processor 301 to execute the radio map prediction method of the embodiments of this application.

[0173] Input / output interface 303 is used to implement information input and output;

[0174] The communication interface 304 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0175] Bus 305 transmits information between various components of the device (e.g., processor 301, memory 302, input / output interface 303, and communication interface 304);

[0176] The processor 301, memory 302, input / output interface 303, and communication interface 304 are connected to each other within the device via bus 305.

[0177] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described radio map prediction method.

[0178] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0179] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0180] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0181] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0182] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0183] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0184] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0185] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0186] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0187] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0188] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0189] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A radio map prediction method, characterized in that, include: A spatial morphological map representing the signal obstruction situation in the target area is obtained, and multiple target observation locations are selected in the target area, wherein the multiple target observation locations are sparsely distributed; For each of the target observation locations, the observation location coordinates corresponding to the target observation location on the spatial morphology map are determined, and the signal strength value of the target observation location is measured. The spatial morphology map, the observation location coordinates, and the corresponding signal intensity value are input into the trained prediction model to perform local feature encoding on the observation location coordinates and the signal intensity value at a fine-grained level to obtain local spatial features, and to perform global feature encoding on the spatial morphology map at a coarse-grained level to obtain global spatial features. Multi-granularity encoded features are obtained by fusing local spatial features and global spatial features at different granularity levels, and multi-granularity decoded features are obtained by decoding the multi-granularity encoded features. Based on the multi-granularity decoded features, a target radio map representing the signal strength at each location in the target area is predicted.

2. The radio map prediction method according to claim 1, characterized in that, The selection of multiple target observation locations within the target area includes: Spatial analysis is performed on the spatial morphology map to determine the land cover area and open ground area of ​​the spatial morphology map; The spatial shape map is divided into multiple initial regions, and at least one candidate region is selected from the multiple initial regions, wherein the area of ​​the open ground area in the candidate region is within a preset area threshold range; For each of the candidate regions, at least one boundary location between the open ground area and the feature-covered area in the candidate region is determined as the target observation location, wherein the boundary location is close to the corner of the feature-covered area.

3. The radio map prediction method according to claim 1, characterized in that, The prediction model includes a first self-attention component. The step of performing local feature encoding on the observation location coordinates and the signal intensity value at a fine-grained level to obtain local spatial features includes: Local feature encoding is performed on the observation location coordinates and the signal intensity value respectively to obtain the initial observation location features and the initial observation signal features; The initial local fusion features are obtained by superimposing the initial observation location features and the initial observation signal features. The initial local fusion features are then processed by local feature extraction based on the preset first self-attention component to obtain updated local fusion features. The updated local fusion feature is used as the new initial local fusion feature. The new initial local fusion feature is subjected to the local feature extraction process at least once more until a preset first condition is met. The last updated local fusion feature is then used as the local spatial feature at the fine-grained level.

4. The radio map prediction method according to claim 1, characterized in that, The prediction model includes a second self-attention component. The process of performing global feature encoding on the spatial morphology map at a coarse-grained level to obtain global spatial features includes: The spatial morphology map is subjected to global feature encoding to obtain initial global spatial features; Based on the second self-attention component, global feature extraction processing is performed on the initial global spatial features to obtain updated global spatial features. The updated global spatial features are used as the new initial global spatial features. The new initial global spatial features are subjected to global feature extraction processing at least once more until a preset second condition is met. The last obtained updated global spatial features are then used as the global spatial features at the coarse-grained level.

5. The radio map prediction method according to claim 1, characterized in that, The prediction model includes a cross-attention component and a multilayer perceptron. The fusion of local spatial features and global spatial features at different granularity levels yields multi-granularity encoded features, including: The query vector is determined based on the local spatial features, and the key vector and value vector are determined based on the global spatial features; Based on the cross-attention component, cross-attention calculation is performed on the query vector, the key vector, and the value vector to obtain cross-attention features; Based on the multilayer perceptron, the cross-attention features are processed by nonlinear transformation to obtain the multi-granularity coding features.

6. The radio map prediction method according to claim 1, characterized in that, The prediction model is trained through the following steps: An initial prediction model, a sample spatial morphology map characterizing the signal occlusion in the sample area, and a real radio map corresponding to the sample spatial morphology map are obtained, wherein multiple sample observation locations are sparsely distributed. Multiple sample observation locations are selected in the sample area. For each sample observation location, the coordinates of the sample observation location corresponding to the sample spatial morphology map are determined, and the sample signal intensity value of the sample observation location is measured. The sample spatial morphology map, the sample observation location coordinates, and the sample signal intensity value are input into the initial prediction model. Local feature encoding is performed on the sample observation location coordinates and the sample signal intensity value to obtain the sample local spatial features at a fine-grained level. Global feature encoding is performed on the sample spatial morphology map to obtain the sample global spatial features at a coarse-grained level. The sample local spatial features and the sample global spatial features at different granularity levels are fused to obtain sample multi-granularity coding features, and the sample multi-granularity coding features are decoded to obtain sample multi-granularity decoding features. Based on the sample multi-granularity decoding features, a sample radio map representing the signal strength at each location in the sample area is predicted. The sample loss value is determined based on the sample radio map and the real radio map. The model parameters of the initial prediction model are adjusted based on the sample loss value to obtain the trained prediction model.

7. The radio map prediction method according to claim 6, characterized in that, The step of determining the sample loss value based on the sample radio map and the real radio map includes: For any of the sample observation locations, determine the predicted signal strength value of the sample observation location on the sample radio map, and determine the actual signal strength value of the sample observation location on the real radio map; Calculate the squared difference of signal strength between the predicted signal strength value and the actual signal strength value, and determine the sample loss value based on the squared difference of signal strength.

8. A radio map prediction device, characterized in that, include: The acquisition module is used to acquire a spatial morphological map representing the signal occlusion situation in the target area, and select multiple target observation locations in the target area, wherein the multiple target observation locations are sparsely distributed; The observation location information determination module is used to determine the observation location coordinates of each target observation location on the spatial morphology map and to measure the signal strength value of the target observation location. The encoding processing module is used to input the spatial morphology map, the observation location coordinates, and the corresponding signal intensity value into the trained prediction model, so as to perform local feature encoding processing on the observation location coordinates and the signal intensity value at a fine-grained level to obtain local spatial features, and to perform global feature encoding processing on the spatial morphology map at a coarse-grained level to obtain global spatial features. The target prediction module is used to fuse the local spatial features and the global spatial features at different granularity levels to obtain multi-granularity encoded features, and to decode the multi-granularity encoded features to obtain multi-granularity decoded features. Based on the multi-granularity decoded features, a target radio map representing the signal strength at each location in the target area is predicted.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the radio map prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the radio map prediction method according to any one of claims 1 to 7.

Images