Residual graph model training methods, devices, equipment, media and program products

By constructing a set of propagation area masks and calculating various loss functions, the problems of shadow morphology and spatial continuity distortion in severely occluded areas in radio map prediction are solved, and more accurate prediction results are achieved.

CN123115801APending Publication Date: 2026-11-17CHINA UNITED NETWORK COMM GRP CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611505350.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing radio map prediction schemes based on deep neural networks show significant distortions in the shadow morphology and spatial continuity of the prediction results in areas with severe building occlusion, compared to the actual propagation distribution. Numerical transitions near building outlines are blurred into smooth, gradual changes.

Method used

By constructing a set of propagation region masks, the target region is divided into an effective prediction region, a non-line-of-sight region, a building boundary region, and a multi-scale local attenuation region. The loss functions for different regions are calculated, including data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, and multi-scale local attenuation consistency loss. These are combined into a composite training objective to update the model parameters.

Benefits of technology

It maintains the accuracy of numerical transition of spatial structure and building outline in the prediction results in severely shaded areas, solves the problems of shadow morphology and spatial continuity distortion, and achieves more accurate radio map prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115801A_ABST
    Figure CN123115801A_ABST
Patent Text Reader

Abstract

This application provides a method, apparatus, device, medium, and program product for training residual graph models, relating to the field of data processing technology. The specific technical solution is as follows: Map information, transmitter information, and propagation features are input into the model to be trained to obtain a predicted residual graph; the true residual graph is obtained from the true path loss and basic propagation prior; based on the building occupancy map, transmitter location, and visibility information, masks for effective prediction, non-line-of-sight, building boundaries, and multi-scale local attenuation regions are constructed; within each mask, data fitting, non-line-of-sight structural consistency, boundary gradient abrupt change, and multi-scale attenuation consistency losses of the two residual graphs are calculated, and under-prediction unilateral risk loss is calculated within the non-line-of-sight sub-region; the various losses are combined into a composite training objective, and the model parameters are updated based on the composite training objective. This method is applied to residual graph model training scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, device, medium and program product for training residual graph models. Background Technology

[0002] With the dense deployment of 5G and future mobile communication networks, engineering processes such as base station site selection, network coverage assessment, interference analysis, and frequency planning all require grid-level prediction of radio maps of target areas in urban environments, that is, predicting the path loss or received power distribution at each location within the area.

[0003] A radio map prediction scheme based on deep neural networks is proposed. This scheme inputs environmental data such as building occupancy maps, transmitter locations, and transmit / receive distance maps of the target area into a convolutional neural network (CNN). The network directly regresses and outputs a path loss distribution map for the entire target area. During the training phase, the scheme calculates the numerical deviation between the predicted and actual values ​​point-by-point on all valid grids of the map, and averages the squared deviations of all grids to obtain a mean squared error (MSE) index covering the entire map. The scheme uses minimizing this global error index as the sole optimization objective, driving the network parameters to iteratively update until convergence.

[0004] However, in this scheme, because the sole optimization objective treats all valid graticles in the entire map equally, the shadow morphology and spatial continuity in the predicted results are significantly distorted from the actual propagation distribution in areas with severe building occlusion. The numerical transitions that should be steep near the building outlines are blurred into gentle gradations. Summary of the Invention

[0005] This application provides a residual map model training method, apparatus, equipment, medium, and program product to solve the problem that existing methods in areas with severe building occlusion have significant distortions in the shadow morphology and spatial continuity of the predicted results compared to the actual propagation distribution, and the numerical transitions that should be steep near the building outline are blurred into gentle gradations.

[0006] In a first aspect, embodiments of this application provide a residual graph model training method, the method comprising:

[0007] The training samples, consisting of map information of the target area, radio transmitter information, and radio propagation characteristics, are input into the model to be trained to obtain the prediction residual map. The true residual map is obtained based on the actual path loss of radio propagation in the target area and the basic propagation prior.

[0008] Based on the building occupancy map of the target area, the location of the radio transmitter, the coordinates of the radio receiver grid, the building obstruction relationship, and the visibility information from the transmitter to the receiver grid, a set of propagation area masks is constructed. The set of propagation area masks includes an effective prediction area mask, a non-line-of-sight area mask, a building boundary area mask, and a multi-scale local attenuation area mask.

[0009] The data fitting loss between the predicted residual map and the real residual map is calculated within the effective prediction region mask, and the non-line-of-sight structural consistency loss between the predicted residual map and the real residual map is calculated within the non-line-of-sight region mask.

[0010] Within the aforementioned building boundary region mask, calculate the consistency loss of the building boundary gradient abrupt change between the predicted residual map and the actual residual map, and within the aforementioned multi-scale local decay region mask, calculate the consistency loss of the multi-scale local decay between the predicted residual map and the actual residual map.

[0011] In the sub-region of the non-line-of-sight mask, the underpredicted one-sided risk loss of the predicted residual map and the real residual map is calculated. The data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss are combined into a composite training objective. The parameters of the model to be trained are updated based on the composite training objective.

[0012] The technical solution provided in this application brings at least the following beneficial effects: By inputting training samples composed of map information of the target area, radio transmitter information, and radio propagation characteristics into the model to be trained, a predicted residual map is obtained. The true residual map is obtained based on the actual path loss of radio propagation within the target area and the basic propagation prior. This allows the model to focus on learning propagation deviations that the basic propagation prior cannot characterize, providing a learning object for complex attenuation caused by building occupancy. Based on this, a propagation area mask set is constructed according to the building occupancy map of the target area, the radio transmitter location, the radio receiver grid coordinates, building occupancy relationships, and the visibility information from the transmitter to the receiver grid. The target area is divided into four different types of regions: effective prediction area mask, non-line-of-sight area mask, building boundary area mask, and multi-scale local attenuation area mask. This explicitly identifies the key spatial locations related to occupancy, providing a spatial basis for differentiated constraints. Furthermore, the data fitting loss between the predicted residual map and the true residual map is calculated within the effective prediction area mask, ensuring that the overall numerical accuracy is not weakened. Simultaneously, the calculation is advanced to the non-line-of-sight area mask. The non-line-of-sight structural consistency loss between the predicted and true residual maps directly constrains the residual spatial morphology in areas severely occluded by buildings. Further, it calculates the building boundary gradient abrupt change consistency loss between the predicted and true residual maps within a mask in the building boundary region. This forces the predicted residual map to reproduce the steep numerical transitions of the true residual map near the building outline, preventing it from being blurred into a gentle gradient. Furthermore, it calculates the multi-scale local attenuation consistency loss between the predicted and true residual maps within a mask in multi-scale local attenuation regions, stabilizing attenuation details at different spatial scales. The model employs a binding mechanism to maintain the spatial continuity of the residual distribution in occluded areas. Subsequently, within sub-regions of the mask in non-line-of-sight areas, the under-predicted one-sided risk loss between the predicted and actual residual maps is calculated, imposing additional penalties on residual distortions in the most severely occluded areas that are not adequately corrected by other losses. Finally, the data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local attenuation consistency loss, and under-predicted one-sided risk loss are combined into a composite training objective. The parameters of the model are then updated based on this composite training objective, ensuring that constraints across different regions work synergistically on the same model. As a result, model training is no longer dominated by the overall average error in smooth areas. The spatial structure of severely occluded areas, the steepness of numerical transitions at building outlines, and the continuity of multi-scale local attenuation are all directly and progressively reinforced by constraints. This addresses the problems of existing methods where shadow morphology and spatial continuity in severely occluded areas are significantly distorted from the actual propagation distribution, and where steep numerical transitions near building outlines are blurred into gentle gradients.

[0013] One possible implementation is as follows: the map information includes a building occupancy map and a building height map; the radio transmitter information includes the transmitter location; the radio propagation characteristics include a distance map from the transmitter to the receiving grid; the receiving grid coordinates are geometric coordinate data representing each receiving grid location; the distance map reflects the propagation spatial conditions between the transmitter and receiver; the training samples composed of the map information of the target area, the radio transmitter information, and the radio propagation characteristics are input into the model to be trained to obtain a prediction residual map, and the true residual map is obtained based on the true path loss of radio propagation within the target area and the basic propagation prior, including:

[0014] Spatial resolution unification and grid coordinate alignment are performed on the building occupancy map, building height map, transmitter location, receiver grid coordinates, and distance map from transmitter to receiver grid of the training samples to obtain input feature maps with unified spatial resolution.

[0015] The encoder of the model to be trained performs multi-scale downsampling feature extraction on the input feature map to obtain a multi-scale radio propagation feature map. The decoder of the model to be trained performs upsampling and cross-layer feature fusion on the multi-scale radio propagation feature map to obtain the prediction residual map.

[0016] Obtain the above-mentioned real path loss map corresponding to the above training samples. The above-mentioned real path loss map is generated from field measurement data or ray tracing simulation data. The above-mentioned basic propagation prior map is calculated according to the free space propagation loss formula based on the transmitter power, transmitter antenna gain, receiver antenna gain and carrier frequency.

[0017] By performing a grid-by-grid interpolation operation on the above-mentioned true path loss map and the above-mentioned basic propagation prior map in the same geographic grid coordinate system, the above-mentioned true residual map is obtained.

[0018] Another possible implementation involves constructing a set of propagation area masks based on the building occupancy map of the target area, the location of the radio transmitter, the coordinates of the radio receiver grid, building obstruction relationships, and visibility information from the transmitter to the receiver grid. This set includes:

[0019] The above effective prediction region mask is generated based on the evaluable area and outdoor area in the above training sample. The above effective prediction region mask is the area after excluding the building interior area and the unlabeled area.

[0020] The visibility information of each receiving grid is determined based on the intersection of the straight path from the transmitter to each receiving grid with the building occupancy map, the obstruction length, and the superelevation. The receiving grids that are not line-of-sight in the visibility information are marked as the non-line-of-sight area mask. The straight path is determined based on the transmitter position and the receiving grid coordinates.

[0021] Morphological gradient operations are performed on the building occupancy map to obtain the building boundary, and a preset pixel width is extended to both sides of the building boundary to obtain the building boundary region mask;

[0022] The occlusion ratio, building density, visibility degradation and local gradient magnitude of the received grid within each window scale are calculated at multiple window scales to obtain a multi-scale attenuation feature sequence.

[0023] The receiving grid corresponding to any feature in the above multi-scale attenuation feature sequence that exceeds the corresponding threshold is marked as a local attenuation region, and the local attenuation regions under each window scale are merged to obtain the above multi-scale local attenuation region mask.

[0024] Another possible implementation involves calculating the non-line-of-sight structural consistency loss between the predicted residual map and the actual residual map within the non-line-of-sight region mask, including:

[0025] Based on the above non-line-of-sight region mask, a non-line-of-sight predicted residual sub-image is extracted from the above predicted residual image, and a non-line-of-sight real residual sub-image is extracted from the above real residual image based on the above non-line-of-sight region mask. The above non-line-of-sight predicted residual sub-image and the above non-line-of-sight real residual sub-image have the same grid size and spatial coordinate range.

[0026] Local window normalization is performed on the above non-line-of-sight predicted residual sub-map and the above non-line-of-sight true residual sub-map respectively to obtain normalized predicted residual sub-map and normalized true residual sub-map, and the local structural similarity index between the above normalized predicted residual sub-map and the above normalized true residual sub-map is calculated.

[0027] Gradient operator convolution processing is performed on the above non-line-of-sight predicted residual sub-map and the above non-line-of-sight true residual sub-map respectively to obtain the predicted gradient direction map and the true gradient direction map, and the gradient direction consistency metric between the above predicted gradient direction map and the above true gradient direction map is calculated.

[0028] The above local structural similarity index and the above gradient direction consistency metric are weighted and summed to obtain the above non-line-of-sight structural consistency loss.

[0029] Another possible implementation involves calculating the underpredicted one-sided risk loss of the predicted residual map and the actual residual map within a sub-region of the non-line-of-sight mask, including:

[0030] Based on the aforementioned non-line-of-sight region mask and occlusion length threshold, grids with occlusion lengths exceeding the aforementioned occlusion length threshold are filtered from the aforementioned non-line-of-sight region mask to obtain the sub-region of the aforementioned non-line-of-sight region mask. The aforementioned occlusion length threshold is set according to the carrier frequency and the average building height.

[0031] Based on the sub-region of the above non-line-of-sight region mask, a deep occlusion prediction residual sub-image is extracted from the above prediction residual image, and based on the above deep occlusion region mask, a deep occlusion real residual sub-image is extracted from the above real residual image.

[0032] The difference between the above deep occlusion prediction residual submap and the above deep occlusion actual residual submap is calculated to obtain the residual difference map. The grids in the above residual difference map where the predicted residual is less than the actual residual and the difference exceeds the tolerance threshold are marked as under-predicted grids. The tolerance threshold is set according to the path loss measurement accuracy and coverage assessment tolerance.

[0033] A one-sided penalty function is applied to the difference magnitude at the aforementioned under-predicted grids to obtain the penalty value for each under-predicted grid. The penalty values ​​of all the aforementioned under-predicted grids are then summed to obtain the aforementioned one-sided risk loss for under-prediction. The aforementioned one-sided penalty function applies a linear or non-linear incremental penalty to the underestimated magnitude.

[0034] Another possible implementation involves calculating the consistency loss of the building boundary gradient abrupt change between the predicted residual map and the actual residual map within the building boundary region mask, including:

[0035] Based on the above building boundary region mask, a boundary prediction residual sub-map is extracted from the above prediction residual map, and a boundary real residual sub-map is extracted from the above real residual map based on the above building boundary region mask. The above boundary prediction residual sub-map and the above boundary real residual sub-map cover the extended zone areas on both sides of the building boundary.

[0036] High-pass filtering is performed on the above-mentioned boundary prediction residual sub-map and the above-mentioned boundary true residual sub-map respectively to obtain the predicted gradient jump component map and the true gradient jump component map, and the gradient jump component difference between the above-mentioned predicted gradient jump component map and the above-mentioned true gradient jump component map is calculated.

[0037] Convolution processing is performed on the above boundary prediction residual submap and the above boundary true residual submap respectively to obtain the prediction response map and the true response map, and the response difference between the above prediction response map and the above true response map is calculated.

[0038] The weighted summation of the gradient jump component differences and the response differences yields the building boundary gradient abrupt change consistency loss.

[0039] Another possible implementation involves calculating the multi-scale local attenuation consistency loss between the predicted residual map and the actual residual map within the multi-scale local attenuation region mask, including:

[0040] Based on the local attenuation region corresponding to any window scale in the above multi-scale local attenuation region mask, a predicted attenuation sub-map at that scale is extracted from the above predicted residual map, and a real attenuation sub-map at that scale is extracted from the above real residual map based on the local attenuation region corresponding to that window scale.

[0041] Window mean calculation is performed on the predicted attenuation submap and the actual attenuation submap respectively to obtain the predicted local mean sequence and the actual local mean sequence, and the mean difference between the predicted local mean sequence and the actual local mean sequence is calculated.

[0042] Gradient magnitude calculations are performed on the predicted attenuation submap and the actual attenuation submap respectively to obtain the predicted gradient magnitude sequence and the actual gradient magnitude sequence, and the magnitude difference between the predicted gradient magnitude sequence and the actual gradient magnitude sequence is calculated.

[0043] The summation of the mean difference and amplitude difference corresponding to each window scale yields the multi-scale local attenuation consistency loss.

[0044] Secondly, embodiments of this application provide a residual graph model training apparatus, including:

[0045] The input module is used to input training samples consisting of map information of the target area, radio transmitter information and radio propagation characteristics into the model to be trained, to obtain the prediction residual map, and to obtain the real residual map based on the real path loss of radio propagation in the target area and the basic propagation prior.

[0046] The construction module is used to construct a set of propagation area masks based on the building occupancy map of the target area, the location of the radio transmitter, the coordinates of the radio receiver grid, the building occupancy relationship, and the visibility information from the transmitter to the receiver grid. The set of propagation area masks includes an effective prediction area mask, a non-line-of-sight area mask, a building boundary area mask, and a multi-scale local attenuation area mask.

[0047] The first calculation module is used to calculate the data fitting loss between the predicted residual map and the real residual map within the effective prediction region mask, and to calculate the non-line-of-sight structural consistency loss between the predicted residual map and the real residual map within the non-line-of-sight region mask.

[0048] The second calculation module is used to calculate the consistency loss of the building boundary gradient abrupt change between the predicted residual map and the real residual map within the building boundary region mask, and to calculate the consistency loss of the multi-scale local decay between the predicted residual map and the real residual map within the multi-scale local decay region mask.

[0049] The combination module is used to calculate the underpredicted one-sided risk loss of the predicted residual map and the real residual map in the sub-region of the non-line-of-sight mask, and combine the data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss into a composite training objective, and train and update the parameters of the model to be trained based on the composite training objective.

[0050] Another possible implementation is that the map information includes a building occupancy map and a building height map; the radio transmitter information includes the transmitter location; the radio propagation characteristics include a distance map from the transmitter to the receiving grid; the receiving grid coordinates are geometric coordinate data representing the grids at each receiving location; the distance map is used to reflect the propagation spatial conditions between the transmitter and receiver; and the input module is specifically used for:

[0051] Spatial resolution unification and grid coordinate alignment are performed on the building occupancy map, building height map, transmitter location, receiver grid coordinates, and distance map from transmitter to receiver grid of the training samples to obtain input feature maps with unified spatial resolution.

[0052] The encoder of the model to be trained performs multi-scale downsampling feature extraction on the input feature map to obtain a multi-scale radio propagation feature map. The decoder of the model to be trained performs upsampling and cross-layer feature fusion on the multi-scale radio propagation feature map to obtain the prediction residual map.

[0053] Obtain the above-mentioned real path loss map corresponding to the above training samples. The above-mentioned real path loss map is generated from field measurement data or ray tracing simulation data. The above-mentioned basic propagation prior map is calculated according to the free space propagation loss formula based on the transmitter power, transmitter antenna gain, receiver antenna gain and carrier frequency.

[0054] By performing a grid-by-grid interpolation operation on the above-mentioned true path loss map and the above-mentioned basic propagation prior map in the same geographic grid coordinate system, the above-mentioned true residual map is obtained.

[0055] Another possible implementation, the above-mentioned building module is specifically used for:

[0056] The above effective prediction region mask is generated based on the evaluable area and outdoor area in the above training sample. The above effective prediction region mask is the area after excluding the building interior area and the unlabeled area.

[0057] The visibility information of each receiving grid is determined based on the intersection of the straight path from the transmitter to each receiving grid with the building occupancy map, the obstruction length, and the superelevation. The receiving grids that are not line-of-sight in the visibility information are marked as the non-line-of-sight area mask. The straight path is determined based on the transmitter position and the receiving grid coordinates.

[0058] Morphological gradient operations are performed on the building occupancy map to obtain the building boundary, and a preset pixel width is extended to both sides of the building boundary to obtain the building boundary region mask;

[0059] The occlusion ratio, building density, visibility degradation and local gradient magnitude of the received grid within each window scale are calculated at multiple window scales to obtain a multi-scale attenuation feature sequence.

[0060] The receiving grid corresponding to any feature in the above multi-scale attenuation feature sequence that exceeds the corresponding threshold is marked as a local attenuation region, and the local attenuation regions under each window scale are merged to obtain the above multi-scale local attenuation region mask.

[0061] Another possible implementation, the first calculation module mentioned above is specifically used for:

[0062] Based on the above non-line-of-sight region mask, a non-line-of-sight predicted residual sub-image is extracted from the above predicted residual image, and a non-line-of-sight real residual sub-image is extracted from the above real residual image based on the above non-line-of-sight region mask. The above non-line-of-sight predicted residual sub-image and the above non-line-of-sight real residual sub-image have the same grid size and spatial coordinate range.

[0063] Local window normalization is performed on the above non-line-of-sight predicted residual sub-map and the above non-line-of-sight true residual sub-map respectively to obtain normalized predicted residual sub-map and normalized true residual sub-map, and the local structural similarity index between the above normalized predicted residual sub-map and the above normalized true residual sub-map is calculated.

[0064] Gradient operator convolution processing is performed on the above non-line-of-sight predicted residual sub-map and the above non-line-of-sight true residual sub-map respectively to obtain the predicted gradient direction map and the true gradient direction map, and the gradient direction consistency metric between the above predicted gradient direction map and the above true gradient direction map is calculated.

[0065] The above local structural similarity index and the above gradient direction consistency metric are weighted and summed to obtain the above non-line-of-sight structural consistency loss.

[0066] Another possible implementation, where the above-mentioned combined module is specifically used for:

[0067] Based on the aforementioned non-line-of-sight region mask and occlusion length threshold, grids with occlusion lengths exceeding the aforementioned occlusion length threshold are filtered from the aforementioned non-line-of-sight region mask to obtain the sub-region of the aforementioned non-line-of-sight region mask. The aforementioned occlusion length threshold is set according to the carrier frequency and the average building height.

[0068] Based on the sub-region of the above non-line-of-sight region mask, a deep occlusion prediction residual sub-image is extracted from the above prediction residual image, and based on the above deep occlusion region mask, a deep occlusion real residual sub-image is extracted from the above real residual image.

[0069] The difference between the above deep occlusion prediction residual submap and the above deep occlusion actual residual submap is calculated to obtain the residual difference map. The grids in the above residual difference map where the predicted residual is less than the actual residual and the difference exceeds the tolerance threshold are marked as under-predicted grids. The tolerance threshold is set according to the path loss measurement accuracy and coverage assessment tolerance.

[0070] A one-sided penalty function is applied to the difference magnitude at the aforementioned under-predicted grids to obtain the penalty value for each under-predicted grid. The penalty values ​​of all the aforementioned under-predicted grids are then summed to obtain the aforementioned one-sided risk loss for under-prediction. The aforementioned one-sided penalty function applies a linear or non-linear incremental penalty to the underestimated magnitude.

[0071] Another possible implementation is that the second calculation module mentioned above is specifically used for:

[0072] Based on the above building boundary region mask, a boundary prediction residual sub-map is extracted from the above prediction residual map, and a boundary real residual sub-map is extracted from the above real residual map based on the above building boundary region mask. The above boundary prediction residual sub-map and the above boundary real residual sub-map cover the extended zone areas on both sides of the building boundary.

[0073] High-pass filtering is performed on the above-mentioned boundary prediction residual sub-map and the above-mentioned boundary true residual sub-map respectively to obtain the predicted gradient jump component map and the true gradient jump component map, and the gradient jump component difference between the above-mentioned predicted gradient jump component map and the above-mentioned true gradient jump component map is calculated.

[0074] Convolution processing is performed on the above boundary prediction residual submap and the above boundary true residual submap respectively to obtain the prediction response map and the true response map, and the response difference between the above prediction response map and the above true response map is calculated.

[0075] The weighted summation of the gradient jump component differences and the response differences yields the building boundary gradient abrupt change consistency loss.

[0076] Another possible implementation is that the second calculation module mentioned above is specifically used for:

[0077] Based on the local attenuation region corresponding to any window scale in the above multi-scale local attenuation region mask, a predicted attenuation sub-map at that scale is extracted from the above predicted residual map, and a real attenuation sub-map at that scale is extracted from the above real residual map based on the local attenuation region corresponding to that window scale.

[0078] Window mean calculation is performed on the predicted attenuation submap and the actual attenuation submap respectively to obtain the predicted local mean sequence and the actual local mean sequence, and the mean difference between the predicted local mean sequence and the actual local mean sequence is calculated.

[0079] Gradient magnitude calculations are performed on the predicted attenuation submap and the actual attenuation submap respectively to obtain the predicted gradient magnitude sequence and the actual gradient magnitude sequence, and the magnitude difference between the predicted gradient magnitude sequence and the actual gradient magnitude sequence is calculated.

[0080] The summation of the mean difference and amplitude difference corresponding to each window scale yields the multi-scale local attenuation consistency loss.

[0081] Thirdly, this application provides an electronic device comprising: a processor and a memory; the memory stores a program or instructions executable on the processor, wherein the program or instructions, when executed by the processor, implement the method of the first aspect described above.

[0082] Fourthly, this application provides a readable storage medium on which a program or instructions are stored, which, when executed by a computer, implement the method of the first aspect described above.

[0083] Fifthly, this application provides a computer program product stored in a storage medium, which, when executed by a computer, implements the method described in the first aspect.

[0084] In a sixth aspect, embodiments of this application provide a chip including a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the method described in the first aspect.

[0085] The beneficial effects of the second to sixth aspects mentioned above are described in the corresponding description of the first aspect and will not be repeated here. Attached Figure Description

[0086] Figure 1 A schematic diagram of the network architecture for a residual graph model training method provided in this application embodiment;

[0087] Figure 2 A flowchart illustrating a residual graph model training method provided in an embodiment of this application;

[0088] Figure 3 A flowchart illustrating another residual graph model training method provided in this application embodiment;

[0089] Figure 4 A flowchart illustrating another residual graph model training method provided in this application embodiment;

[0090] Figure 5 A flowchart illustrating another residual graph model training method provided in this application embodiment;

[0091] Figure 6 This embodiment provides a flowchart for generating a propagation area mask;

[0092] Figure 7 This embodiment provides a flowchart for generating a propagation area mask;

[0093] Figure 8 This is a schematic diagram of the structure of a residual graph model training device provided in an embodiment of this application;

[0094] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0095] The residual graph model training method, apparatus, equipment, medium, and program products provided in this application will be described in detail below with reference to the accompanying drawings.

[0096] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0097] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0098] The terms "at least one," "at least one of," etc., used in the specification and claims of this application refer to any one, any two, or a combination of two or more of the included items. For example, at least one of a, b, and c can mean: "a," "b," "c," "a and b," "a and c," "b and c," and "a, b, and c," where a, b, and c can be single or multiple. Similarly, "at least two" refers to two or more items, and its meaning is similar to that of "at least one."

[0099] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0100] The embodiments of this application provide a residual graph model training method, apparatus, device, medium, and program product that can be applied to residual graph model training scenarios.

[0101] A radio map prediction scheme based on deep neural networks is proposed. This scheme inputs environmental data such as building occupancy maps, transmitter locations, and transmit / receive distance maps of the target area into a convolutional neural network (CNN). The network directly regresses and outputs a path loss distribution map for the entire target area. During the training phase, the scheme calculates the numerical deviation between the predicted and actual values ​​point-by-point on all valid grids of the map, and averages the squared deviations of all grids to obtain a mean squared error (MSE) index covering the entire map. The scheme uses minimizing this global error index as the sole optimization objective, driving the network parameters to iteratively update until convergence.

[0102] However, in this scheme, because the sole optimization objective treats all valid graticles in the entire map equally, the shadow morphology and spatial continuity in the predicted results are significantly distorted from the actual propagation distribution in areas with severe building occlusion. The numerical transitions that should be steep near the building outlines are blurred into gentle gradations.

[0103] To address the aforementioned technical problems, this application provides a residual map model training method, apparatus, device, medium, and program product. By inputting training samples composed of map information of the target area, radio transmitter information, and radio propagation characteristics into the model to be trained, a predicted residual map is obtained. The true residual map is obtained based on the actual path loss of radio propagation within the target area and the basic propagation prior, allowing the model to focus on learning propagation deviations that the basic propagation prior cannot characterize, providing a learning object for complex attenuation caused by building occupancy. Based on this, a propagation area mask set is constructed according to the building occupancy map of the target area, the radio transmitter location, the radio receiver grid coordinates, building occupancy relationships, and the visibility information from the transmitter to the receiver grid. The target area is divided into four different types of regions: effective prediction area mask, non-line-of-sight area mask, building boundary area mask, and multi-scale local attenuation area mask. This explicitly identifies key spatial locations related to occupancy, providing a spatial basis for differentiated constraints. Furthermore, the data fitting loss between the predicted residual map and the true residual map is calculated within the effective prediction area mask, ensuring that the overall numerical accuracy is not weakened, while simultaneously advancing to… The non-line-of-sight structural consistency loss between the predicted and actual residual maps is calculated within a mask in the non-line-of-sight region. This directly constrains the residual spatial morphology in areas with severe building occlusion. Further, within a mask in the building boundary region, the consistency loss between the predicted and actual residual maps for abrupt gradient changes at the building boundary is calculated. This forces the predicted residual map to reproduce the steep numerical transitions of the actual residual map near the building outline, preventing it from being blurred into a gentle gradient. Furthermore, within a mask in the multi-scale local attenuation region, the consistency loss between the predicted and actual residual maps for multi-scale local attenuation is calculated, ensuring attenuation details at different spatial scales. A stable constraint is obtained to maintain the spatial continuity of the residual distribution in the occluded region. Subsequently, the underpredicted one-sided risk loss between the predicted residual map and the true residual map is calculated in the sub-region of the mask in the non-line-of-sight region. Additional penalties are imposed on the residual distortion in the most severely occluded areas that has not been fully corrected by other losses. Finally, the data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss are combined into a composite training objective. The parameters of the model to be trained are updated based on the composite training objective, so that the constraints of each region work synergistically on the same model. As a result, the model training is no longer dominated by the overall average error in the smooth region. The spatial structure of the severely occluded region, the steepness of the numerical transition at the building outline, and the continuity of the multi-scale local decay are all directly and progressively reinforced by constraints. This solves the problem that existing methods have obvious distortions in the shadow morphology and spatial continuity of the prediction results in the severely occluded region, and the problem that the numerical transition that should be steep near the building outline is blurred into a smooth gradient.

[0104] The following description, in conjunction with the accompanying drawings, details the residual graph model training method, apparatus, equipment, medium, and program products provided in the embodiments of this application.

[0105] Figure 1 This illustration shows a network architecture for a residual graphical model training method provided in an embodiment of this application. For example... Figure 1 As shown, the network architecture includes a residual graph model training device 101 and a terminal device 102. The residual graph model training device 101 and the terminal device 102 are interconnected.

[0106] In some embodiments, the residual graph model training device 101 may be a server, a computer, or a processor or processing unit within a server or computer. The server may be a single server or a server cluster consisting of multiple servers. It should be noted that the embodiments of this application do not limit the specific device form of the residual graph model training device 101. Figure 1 The residual graph model training device 101 is used as an example of a single server.

[0107] In some embodiments, the terminal device 102 may be a mobile phone, tablet computer, laptop computer, handheld computer, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, personal computer (PC), ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc., and the embodiments of this application do not specifically limit it. Figure 1 The example shown is a mobile phone, with terminal device 102 as an example.

[0108] In some embodiments, the residual map model training device 101 is used to input training samples consisting of map information of the target area, radio transmitter information, and radio propagation characteristics into the model to be trained to obtain a predicted residual map, and to obtain a true residual map based on the true path loss of radio propagation within the target area and the basic propagation prior; to construct a propagation area mask set based on the building occupancy map of the target area, the radio transmitter location, the radio receiver grid coordinates, building occupancy relationships, and the visibility information from the transmitter to the receiver grid, the propagation area mask set including an effective prediction area mask, a non-line-of-sight area mask, a building boundary area mask, and a multi-scale local attenuation area mask; and to calculate the data fitting loss between the predicted residual map and the true residual map within the effective prediction area mask. The system calculates the non-line-of-sight structural consistency loss between the predicted residual map and the true residual map within the non-line-of-sight region mask; it calculates the building boundary gradient abrupt change consistency loss between the predicted residual map and the true residual map within the building boundary region mask; it calculates the multi-scale local decay consistency loss between the predicted residual map and the true residual map within the multi-scale local decay region mask; and it calculates the underpredicted one-sided risk loss between the predicted residual map and the true residual map within a sub-region of the non-line-of-sight region mask. The data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss are combined into a composite training objective. The parameters of the model to be trained are updated based on this composite training objective. Terminal device 102 is used to monitor the operating status of the residual map model training device 101.

[0109] It should be noted that the network architecture described in the embodiments of this application is for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and does not constitute a limitation on the technical solutions provided in the embodiments of this application. As network architectures evolve, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0110] See Figure 2 This is a flowchart illustrating a residual graph model training method provided in an embodiment of this application. Figure 2 As shown, the residual graph model training method provided in this application embodiment specifically includes the following steps 201 to 205.

[0111] Step 201: The residual map model training device inputs training samples consisting of map information of the target area, radio transmitter information, and radio propagation characteristics into the model to be trained to obtain the predicted residual map, and obtains the true residual map based on the actual path loss of radio propagation in the target area and the basic propagation prior.

[0112] In some embodiments, the residual map model training device performs spatial resolution unification and grid coordinate alignment processing on the building occupancy map, building height map, transmitter location, receiver grid coordinates, and distance map from the transmitter to the receiver grid of the training samples to obtain an input feature map with unified spatial resolution. This input feature map is then input into the model to be trained. Specifically, spatial resolution unification resamples all input data to the same grid size, and grid coordinate alignment maps each input data grid to a grid within the same geographic grid coordinate system, ensuring that the input data are spatially matched during subsequent grid-by-grid calculations.

[0113] In some embodiments, the residual map model training device performs multi-scale downsampling feature extraction processing on the input feature map through the encoder of the model to be trained, obtaining a multi-scale radio propagation feature map. This multi-scale downsampling feature extraction process extracts propagation features at different receptive field ranges while progressively reducing spatial resolution. Shallow layers retain local details such as building outlines, while deeper layers reflect the large-scale propagation attenuation trend. Then, the decoder of the model to be trained performs upsampling and cross-layer feature fusion processing on the multi-scale radio propagation feature map, fusing the features that progressively restore spatial resolution with the features of the same scale from the encoder, ultimately outputting a predicted residual map. The model to be trained can be any of a Convolutional Neural Network (CNN), U-Net, Transformer, Graph Neural Network, or encoder-decoder network and combinations thereof; this method does not require a specific model structure.

[0114] In some embodiments, the residual map model training device acquires the true path loss map corresponding to the training samples. The true path loss map is generated from field measurement data or ray tracing simulation data. It then calculates the basic propagation prior map according to the free-space propagation loss formula based on the transmitter power, transmitter antenna gain, receiver antenna gain, and carrier frequency. Subsequently, it performs a grid-by-grid difference operation between the true path loss map and the basic propagation prior map in the same geographic grid coordinate system to obtain the true residual map. For example, if the true path loss at a certain receiver grid is 130 dB and the basic propagation prior is 92 dB, then the true residual at that grid is 38 dB. This value reflects the additional loss introduced by environmental factors such as building obstruction in addition to free-space propagation.

[0115] In some embodiments, map information represents the geographic environment data of the target area, including a building occupancy map and a building height map. The building occupancy map represents whether each grid is occupied by a building, and the building height map represents the building height corresponding to the grid occupied by a building. Radio transmitter information represents the configuration data of the transmitter source, including the transmitter location. Radio propagation characteristics represent data related to propagation conditions, including a distance map from the transmitter to the receiving grid. The receiving grid coordinates represent the geometric coordinate data of each receiving grid. The distance map is used to reflect the propagation spatial conditions between the transmitter and receiver. The true residual map represents the grid-by-grid difference distribution between the true path loss and the basic propagation prior. The prediction residual map represents the distribution of the estimated values ​​output by the model to be trained for the portion not covered by the basic propagation prior.

[0116] Step 202: The residual map model training device constructs a propagation area mask set based on the building occupancy map of the target area, the location of the radio transmitter, the coordinates of the radio receiver grid, the building occupancy relationship, and the visibility information from the transmitter to the receiver grid. The propagation area mask set includes an effective prediction area mask, a non-line-of-sight area mask, a building boundary area mask, and a multi-scale local attenuation area mask.

[0117] In some embodiments, the residual graph model training device generates an effective prediction region mask based on the evaluable region and outdoor region in the training samples, that is, excluding the interior region of the building and the unlabeled region, and retaining only the grid that is both outdoors and meets the loss evaluation conditions, as the calculation range of the data fitting loss.

[0118] In some embodiments, the residual map model training device determines the straight-line path from the transmitter to each receiving grid based on the transmitter location and receiving grid coordinates. It then determines the visibility information of each receiving grid based on the intersection of the straight-line path with the building occupancy map, the occlusion length, and the superelevation. Receiving grids that are non-line-of-sight in the visibility information are marked as non-line-of-sight region masks. For example, if the transmitter height is 25 meters, and the straight-line path of a receiving grid passes through a 30-meter-high building, the length of the straight-line path inside the building is the occlusion length, and the portion of the highest intersection point on the straight-line path exceeding the top of the building is the superelevation. Based on this, the receiving grid is determined to be non-line-of-sight.

[0119] In some embodiments, the residual map model training device performs morphological gradient operations on the building occupancy map to obtain the building boundary, that is, extracts the transition position between the building occupancy grid and the building non-occupancy grid by the difference between the dilation result and the erosion result, and then expands the building boundary to both sides by a preset pixel width, for example, by 2 grids on each side, to obtain a building boundary region mask, so that the building boundary region mask covers the numerical transition zone on both sides of the building outline.

[0120] In some embodiments, the residual map model training device calculates the occlusion ratio, building density, visibility degradation degree, and local gradient magnitude of the received grids at multiple window scales to obtain a multi-scale attenuation feature sequence. The occlusion ratio is the ratio of the number of non-line-of-sight grids to the total number of grids within the window; the building density is the proportion of grids occupied by buildings within the window; the visibility degradation degree characterizes the degree of change in visibility information from line-of-sight to non-line-of-sight within the window; and the local gradient magnitude characterizes the gradient magnitude of the building occupancy map or residual correlation quantity within the window. Subsequently, the received grids corresponding to any feature in the multi-scale attenuation feature sequence that exceeds the corresponding threshold are marked as local attenuation regions, and the local attenuation regions at each window scale are merged to obtain a multi-scale local attenuation region mask. For example, multiple window scales can be 3×3, 7×7, and 15×15 grid windows; the larger the window scale, the wider the attenuation range it reflects.

[0121] In some embodiments, building occlusion relationships characterize the spatial geometric relationship in which the propagation path from the transmitter to the receiver grid is truncated by buildings; visibility information characterizes the determination result of whether there is building occlusion between the transmitter and each receiver grid, including line-of-sight and non-line-of-sight categories; the propagation region mask set characterizes the total number of masks used to define the range of each loss calculation, including effective prediction region masks, non-line-of-sight region masks, building boundary region masks, and multi-scale local attenuation region masks.

[0122] Step 203: The residual map model training device calculates the data fitting loss between the predicted residual map and the real residual map within the effective prediction region mask, and calculates the non-line-of-sight structural consistency loss between the predicted residual map and the real residual map within the non-line-of-sight region mask.

[0123] In some embodiments, the residual map model training device calculates the data fitting loss between the predicted residual map and the true residual map within the effective prediction region mask. That is, it calculates the deviation between the predicted residual and the true residual grid by grid by grid on the grid covered by the effective prediction region mask, and sums or averages the deviations of each grid to obtain the data fitting loss. The data fitting loss can be any one of the following: Smooth-L1 loss, Mean Squared Error (MSE) loss, Mean Absolute Error (MAE) loss, or Huber loss, so that the model as a whole closely approximates the true residual distribution.

[0124] In some embodiments, the residual map model training device extracts non-line-of-sight predicted residual sub-images and non-line-of-sight true residual sub-images from the predicted residual map and the true residual map respectively, based on a non-line-of-sight region mask. Both sub-images have the same grid size and spatial coordinate range. Local window normalization is performed on both sub-images, i.e., the residual value is subtracted from the window mean and divided by the window standard deviation within the local window to eliminate the influence of residual amplitude differences at different locations, resulting in normalized predicted residual sub-images and normalized true residual sub-images. A local structural similarity index is calculated between the two sub-images to measure the consistency of their local structures. Gradient operator convolution is then performed on the non-line-of-sight predicted residual sub-images and non-line-of-sight true residual sub-images respectively, for example, by performing convolution with difference operators in the horizontal and vertical directions, resulting in predicted gradient direction maps and true gradient direction maps. A gradient direction consistency metric is calculated between the two sub-images to measure the consistency of the residual change direction. Finally, the local structural similarity index and the gradient direction consistency metric are weighted and summed to obtain the non-line-of-sight structural consistency loss. The loss does not simply increase the weight of the non-line-of-sight grid, but rather constrains the spatial morphology and structural continuity of the residuals in the non-line-of-sight region.

[0125] Step 204: The residual map model training device calculates the consistency loss of the building boundary gradient abrupt change between the predicted residual map and the real residual map within the building boundary region mask, and calculates the multi-scale local decay consistency loss between the predicted residual map and the real residual map within the multi-scale local decay region mask.

[0126] In some embodiments, the residual map model training device extracts boundary predicted residual sub-images and boundary true residual sub-images from the predicted residual map and the true residual map respectively, based on the building boundary region mask. These sub-images cover the extended band regions on both sides of the building boundary. High-pass filtering is applied to both sub-images to extract high-frequency gradient transition components at the building boundary, resulting in predicted gradient transition component maps and true gradient transition component maps. The difference between the two gradient transition components is calculated to constrain whether the predicted residual map exhibits a steep transition at the building boundary comparable to the true residual map. Convolution processing is then applied to the boundary predicted residual sub-images and boundary true residual sub-images, for example, using convolution kernels for detecting edge transitions, to obtain predicted response maps and true response maps. The response difference between the two is calculated. Finally, the gradient transition component difference and the response difference are weighted and summed to obtain the building boundary gradient abrupt change consistency loss, ensuring the model maintains reasonable abrupt changes at building occlusion edges and reducing over-smoothing.

[0127] In some embodiments, the residual map model training device extracts the predicted attenuation sub-map and the true attenuation sub-map at any window scale corresponding to the local attenuation region in the multi-scale local attenuation region mask, respectively, from the predicted residual map and the true attenuation sub-map at that scale; performs window mean calculation on both to obtain the predicted local mean sequence and the true local mean sequence, and calculates the mean difference between the two to constrain the overall level of the local attenuation patch at that scale; then performs gradient magnitude calculation on both to obtain the predicted gradient magnitude sequence and the true gradient magnitude sequence, and calculates the magnitude difference between the two to constrain the strength of the attenuation change at that scale; and sums the mean difference and magnitude difference corresponding to each window scale to obtain the multi-scale local attenuation consistency loss, so that the model maintains the attenuation patch morphology and boundary transition consistent with the true residual distribution at each spatial scale.

[0128] Step 205: The residual map model training device calculates the underpredicted one-sided risk loss of the predicted residual map and the real residual map in the sub-region of the non-line-of-sight mask, and combines the data fitting loss, non-line-of-sight structure consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss into a composite training objective, and trains and updates the parameters of the model to be trained based on the composite training objective.

[0129] In some embodiments, the residual graph model training device filters out grids with occlusion lengths exceeding the occlusion length threshold from the non-line-of-sight region mask based on the non-line-of-sight region mask and the occlusion length threshold, to obtain a sub-region of the non-line-of-sight region mask, wherein the occlusion length threshold is set according to the carrier frequency and the average building height, so that grids occluded by taller and denser buildings are included in the sub-region.

[0130] In some embodiments, the residual map model training device extracts deep occlusion predicted residual sub-images and deep occlusion true residual sub-images from the predicted residual map and the true residual map respectively, based on the sub-regions of the non-line-of-sight region mask. The device performs a difference operation on the two to obtain a residual difference map. Grids in the residual difference map whose predicted residual is less than the true residual and whose difference exceeds the tolerance threshold are marked as underpredicted grids. The tolerance threshold is set according to the path loss measurement accuracy and coverage evaluation tolerance to exclude grids within the normal measurement error range. If the predicted residual is higher than the true residual or the underestimation does not exceed the tolerance threshold, the grid is not marked as an underpredicted grid.

[0131] In some embodiments, the residual map model training device performs a one-sided penalty function operation on the difference magnitude at the underpredicted grid to obtain the penalty value of each underpredicted grid. The one-sided penalty function applies a linear or non-linear increasing penalty to the underestimation magnitude, that is, the larger the underestimation magnitude, the higher the penalty value, while no penalty is applied to the overestimation case where the predicted residual is greater than the true residual. The penalty values ​​of each underpredicted grid are summed to obtain the underpredicted one-sided risk loss, which is used to reduce the risk of underestimating path loss in coverage assessment.

[0132] In some embodiments, the residual graph model training device combines data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted unilateral risk loss into a composite training objective by weighting them according to their respective weights. The weights can be fixed weights or determined by stage scheduling, gradient normalization, uncertainty weighting, or task balancing during the training process. Then, the parameters of the model to be trained are updated through backpropagation to reduce the composite training objective, and the updated parameters are fed back to the model to be trained until the training termination condition is met.

[0133] In some embodiments, the underpredicted one-sided risk loss represents the amount of unidirectional penalty imposed on cases where residuals are underestimated at severely occluded grids; the composite training objective represents the overall optimization objective, composed of the various losses, used to drive model parameter updates.

[0134] In some embodiments, combined with Figure 2 ,like Figure 3 As shown, the above map information includes a building occupancy map and a building height map; the above radio transmitter information includes the transmitter location; the above radio propagation characteristics include a distance map from the transmitter to the receiving grid; the above receiving grid coordinates are geometric coordinate data representing each receiving location grid; the above distance map is used to reflect the propagation spatial conditions between the transmitter and receiver; step 201 can also be implemented through the following steps 201a to 201d.

[0135] Step 201a: The residual map model training device performs spatial resolution unification and grid coordinate alignment processing on the building occupancy map, building height map, transmitter position, receiver grid coordinates, and distance map from transmitter to receiver grid of the training samples to obtain an input feature map with unified spatial resolution.

[0136] In some embodiments, when performing spatial resolution unification processing, the residual map model training device resamples the building occupancy map, building height map, transmitter location, receiver grid coordinates, and distance map from transmitter to receiver grid in the training samples with inconsistent grid sizes to the same grid size, for example, uniformly resampled to a 2m × 2m grid resolution. When performing grid coordinate alignment processing, the grids of each input data after resampling are matched one-to-one with the grids under the same geographic grid coordinate system to eliminate the spatial misalignment between the input data caused by different coordinate origins or grid divisions, thereby obtaining an input feature map with unified spatial resolution, providing spatially matched input for the subsequent grid-by-grid feature extraction by the encoder.

[0137] In some embodiments, the input feature map represents multi-channel rasterized data that is stacked after alignment of a building occupancy map, a building height map, a transmitter location, receiver grid coordinates, and a distance map from the transmitter to the receiver grid. The building occupancy map and the building height map constitute the environment channel, the transmitter location and the receiver grid coordinates constitute the geometric location channel, and the distance map from the transmitter to the receiver grid constitutes the propagation condition channel.

[0138] Step 201b: The residual map model training device performs multi-scale downsampling feature extraction processing on the input feature map through the encoder of the model to be trained to obtain a multi-scale radio propagation feature map, and performs upsampling and cross-layer feature fusion processing on the multi-scale radio propagation feature map through the decoder of the model to be trained to obtain the prediction residual map.

[0139] In some embodiments, the residual map model training device performs multi-scale downsampling feature extraction processing on the input feature map through the encoder of the model to be trained. The encoder consists of successive convolutional layers and downsampling layers. After each downsampling stage, the spatial resolution of the feature map is reduced by one stage and the receptive field is expanded by one stage. Shallow features retain local details such as building outlines, while deep features reflect the propagation attenuation trend over a large area, thereby obtaining a multi-scale radio propagation feature map.

[0140] In some embodiments, the residual map model training device performs upsampling and cross-layer feature fusion processing on the multi-scale radio propagation feature map through the decoder of the model to be trained. The decoder restores the spatial resolution of the feature map step by step, and at each level, it concatenates or adds the upsampled features with the same scale in the encoder along the channel direction, so that the features that restore the spatial resolution carry both deep, large-scale propagation information and shallow, local detail information. Finally, the final stage of the decoder outputs a predicted residual map with the same grid size as the input feature map. The model to be trained can be any of the following: Convolutional Neural Network (CNN), U-Net, Transformer, Graph Neural Network, or encoder-decoder network and combinations thereof. This method does not require a specific model structure.

[0141] Step 201c: The residual graph model training device obtains the real path loss graph corresponding to the training sample. The real path loss graph is generated from field measurement data or ray tracing simulation data, and the basic propagation prior graph is calculated according to the free space propagation loss formula based on the transmitter power, transmitter antenna gain, receiver antenna gain and carrier frequency.

[0142] In some embodiments, when the residual map model training device acquires the true path loss map corresponding to the training sample, the true path loss map is generated from field measurement data or ray tracing simulation data. The field measurement data comes from the actual signal measurement at each receiving location in the target area, and the ray tracing simulation data comes from the path loss simulation results obtained by calculating the propagation mechanism of radio waves emitted from the transmitter on the surface of the building, such as reflection and diffraction. Both sources map the path loss value to each receiving grid in the geographic grid coordinate system.

[0143] In some embodiments, the residual map model training device calculates the basic propagation prior map based on the transmitter power, transmitter antenna gain, receiver antenna gain, and carrier frequency according to the free-space propagation loss formula. The free-space propagation loss formula describes the deterministic relationship between path loss and transmission / reception distance and carrier frequency in an ideal propagation environment without any obstructions. Substituting the transmitter power, transmitter antenna gain, receiver antenna gain, and carrier frequency into this formula yields the path loss distribution of each receiving grid within the target area under free-space propagation conditions, i.e., the basic propagation prior map. The basic propagation prior map characterizes the base map of propagation loss in the target area without considering building obstructions and environmental influences.

[0144] Step 201d: The residual map model training device performs grid-by-grid difference calculations on the above-mentioned real path loss map and the above-mentioned basic propagation prior map in the same geographic grid coordinate system to obtain the above-mentioned real residual map.

[0145] In some embodiments, the residual map model training device performs grid-by-grid interpolation on the true path loss map and the basic propagation prior map in the same geographic grid coordinate system. Specifically, at the same receiving grid, the true path loss value is subtracted from the basic propagation prior value, and the difference is used as the true residual value for that receiving grid. This process is repeated for all receiving grids to obtain the true residual map. For example, if the true path loss at a certain receiving grid is 130 dB and the basic propagation prior is 92 dB, then the true residual at that grid is 38 dB. This value reflects the additional loss introduced by environmental factors such as building obstruction in free space propagation. The true residual map represents the grid-by-grid distribution of the difference between the true path loss and the basic propagation prior, serving as a supervisory target for comparison with the predicted residual map during training.

[0146] In this way, by aligning the environment, geometry, and propagation conditions into an input feature map with a uniform spatial resolution, the encoder and decoder output a predicted residual map, and construct the true residual map using the grid-by-grid difference between the true path loss map and the basic propagation prior map, the model to be trained can focus on learning the propagation bias caused by environmental factors such as occlusion that cannot be characterized by the basic propagation prior, thus providing a numerical comparison basis for the calculation of composite loss within the propagation region mask.

[0147] The residual map model training method provided in this application obtains a predicted residual map by inputting training samples composed of map information of the target area, radio transmitter information, and radio propagation characteristics into the model to be trained. The true residual map is obtained based on the actual path loss of radio propagation within the target area and the basic propagation prior. This allows the model to focus on learning propagation biases that cannot be characterized by the basic propagation prior, providing a learning object for complex attenuation caused by building occupancy. Based on this, a propagation area mask set is constructed according to the building occupancy map of the target area, the radio transmitter location, the radio receiver grid coordinates, building occupancy relationships, and the visibility information from the transmitter to the receiver grid. The target area is divided into four different types of regions: effective prediction area mask, non-line-of-sight area mask, building boundary area mask, and multi-scale local attenuation area mask. This explicitly identifies key spatial locations related to occupancy, providing a spatial basis for differentiated constraints. Furthermore, the data fitting loss between the predicted residual map and the true residual map is calculated within the effective prediction area mask to ensure that the overall numerical accuracy is not weakened. Simultaneously, the calculation is extended to the non-line-of-sight area mask. The non-line-of-sight structural consistency loss between the residual map and the true residual map directly constrains the residual spatial morphology in areas severely occluded by buildings. Further, the consistency loss of the building boundary gradient abrupt change between the predicted and true residual maps is calculated within a mask in the building boundary region. This forces the predicted residual map to reproduce the steep numerical transition of the true residual map near the building outline, preventing it from being blurred into a gentle gradient. Furthermore, the multi-scale local attenuation consistency loss between the predicted and true residual maps is calculated within a mask in multi-scale local attenuation regions, ensuring stable constraints on attenuation details at different spatial scales. To maintain the spatial continuity of the residual distribution in the occluded region, the underpredicted one-sided risk loss is calculated within the sub-regions of the mask in the non-line-of-sight region, applying additional penalties to residual distortions in the most severely occluded areas that are not adequately corrected by other losses. Finally, the data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss are combined into a composite training objective. The parameters of the model under training are updated based on the composite training objective, allowing the constraints of each region to work synergistically on the same model. As a result, model training is no longer dominated by the overall average error in smooth regions. The spatial structure of severely occluded regions, the steepness of numerical transitions at building outlines, and the continuity of multi-scale local decays are all directly and progressively reinforced by constraints. This solves the problem that existing methods often result in significant distortions in the shadow morphology and spatial continuity of predictions in severely occluded areas compared to the actual propagation distribution, and that the numerical transitions near building outlines, which should be steep, are blurred into gentle gradients.

[0148] In some embodiments, combined with Figure 2 ,like Figure 4As shown, step 202 can also be implemented through steps 202a to 202e.

[0149] Step 202a: The residual map model training device generates the effective prediction area mask based on the evaluable area and outdoor area in the training samples. The effective prediction area mask is the area excluding the interior area of ​​the building and the unlabeled area.

[0150] In some embodiments, the residual map model training device generates an effective prediction region mask based on the evaluable and outdoor areas in the training samples. Specifically, it retains grids that belong to both outdoor and evaluable areas as effective grids, while excluding interior building areas and unlabeled areas, thus obtaining the effective prediction region mask. The outdoor area is an open area within the target region not occupied by buildings; the evaluable area is an area with loss assessment conditions, such as areas with true path loss labels or within the model's prediction range; the interior building area is a set of grids occupied by buildings, where there is no need for radio reception; and the unlabeled area is a set of grids lacking true residual labels and unable to participate in loss calculation.

[0151] In some embodiments, the effective prediction region mask is represented in the form of a binary raster map, where the effective raster has a value of 1 and the excluded raster has a value of 0. This mask is used to limit the range of raster cells involved in the calculation when subsequently calculating the data fitting loss.

[0152] Step 202b: The residual map model training device determines the visibility information of each receiving grid based on the intersection relationship between the straight path from the transmitter to each receiving grid and the building occupancy map, the obstruction length, and the superelevation, and marks the receiving grids that are non-line-of-sight in the visibility information as the non-line-of-sight area mask; the straight path is determined based on the transmitter position and the receiving grid coordinates.

[0153] In some embodiments, the residual map model training device determines the straight path from the transmitter to each receiving grid based on the transmitter position and the receiving grid coordinates. This straight path is a spatial connection between the transmitter and the center of the receiving grid. Then, the straight path is compared grid by grid with the building occupancy map to obtain the intersection relationship between the straight path and the building occupancy map. Based on the intersection, the visibility information of each receiving grid is determined by combining the occupancy length and the superelevation. When the straight path intersects with the building and the occupancy length or superelevation meets the determination condition, the visibility information of the receiving grid is determined to be non-line-of-sight; otherwise, it is determined to be line-of-sight. Finally, the receiving grids that are non-line-of-sight in the visibility information are marked one by one, and the non-line-of-sight region mask is obtained by summarizing them.

[0154] In some embodiments, the obstruction length represents the length of a straight path traveling inside a building; the superelevation represents the height difference between the line-of-sight height at the intersection of the straight paths and the height of the building at that point. For example, if the transmitter is 25 meters high and a receiving grid's straight path passes through a 30-meter-high building, the length of the straight path inside the building is the obstruction length, and the portion of the line-of-sight height at the intersection of the straight paths below the top of the building is the superelevation. Based on this, the visibility information of the receiving grid is determined to be non-line-of-sight.

[0155] Step 202c: The residual map model training device performs morphological gradient calculations on the building occupancy map to obtain the building boundary, and expands the building boundary to both sides by a preset pixel width to obtain the building boundary region mask.

[0156] In some embodiments, the residual map model training device performs morphological gradient operations on the building occupancy map to obtain the building boundary. The morphological gradient operation first performs dilation and erosion operations on the building occupancy map, and then takes the difference between the dilation and erosion results. This difference extracts the transition position between the building occupancy grid and the building non-occupancy grid, i.e., the building boundary. Then, a preset pixel width is extended to both sides of the building boundary, that is, with the building boundary as the center, a preset number of grids adjacent to it are included. For example, two grids are extended to both sides of the building boundary to obtain a building boundary region mask. The building boundary region mask covers the numerical transition zone on both sides of the building outline, providing the working area for the subsequent building boundary gradient abrupt change consistency loss.

[0157] In some embodiments, the preset pixel width is a pre-defined number of extended grids, and its value can be determined based on the spatial range of the residual jump at the building boundary.

[0158] Step 202d: The residual map model training device calculates the occlusion ratio, building density, visibility reduction degree and local gradient magnitude of the received grid at each window scale to obtain a multi-scale attenuation feature sequence.

[0159] In some embodiments, the residual map model training device calculates the occlusion ratio, building density, visibility degradation degree, and local gradient magnitude of the received grids at multiple window scales to obtain a multi-scale attenuation feature sequence. The multiple window scales can be, for example, 3×3, 7×7, and 15×15 grid windows; the larger the window scale, the wider the attenuation spatial range it reflects. Specifically, the occlusion ratio is the ratio of the number of grids within the window whose visibility information is non-line-of-sight to the total number of grids within the window; the building density is the ratio of the number of grids within the window occupied by buildings to the total number of grids within the window; the visibility degradation degree characterizes the degree of change in visibility information within the window from line-of-sight to non-line-of-sight; and the local gradient magnitude characterizes the gradient magnitude of the building occupancy map or residual correlation quantity within the window.

[0160] In some embodiments, the multi-scale attenuation feature sequence characterizes the combination of features arranged by scale of the occlusion ratio, building density, visibility reduction degree and local gradient magnitude at each window scale, and is used to reflect the propagation attenuation strength of each receiving grid location in multiple spatial ranges.

[0161] Step 202e: The residual map model training device marks the receiving grid corresponding to any feature in the multi-scale attenuation feature sequence that exceeds the corresponding threshold as a local attenuation region, and merges the local attenuation regions under each window scale to obtain the multi-scale local attenuation region mask.

[0162] In some embodiments, the residual map model training device marks the receiving grid corresponding to any feature in the multi-scale attenuation feature sequence that exceeds the corresponding threshold as a local attenuation region. That is, for a given receiving grid, if any one of its occlusion ratio, building density, visibility degradation degree, or local gradient magnitude at any window scale exceeds the threshold corresponding to that feature, the receiving grid is marked as a local attenuation region. Subsequently, the local attenuation regions marked at each window scale are merged, that is, the union of the local attenuation regions at each window scale is taken to obtain a multi-scale local attenuation region mask. Among them, the threshold corresponding to each feature is a pre-set judgment threshold, which is used to distinguish between regions with obvious attenuation and regions with gentle attenuation.

[0163] Thus, by automatically calculating effective prediction area masks, non-line-of-sight area masks, building boundary area masks, and multi-scale local attenuation area masks from maps, buildings, and transmission and reception relationships, a set of propagation area masks is formed. The target area is divided into key areas with different properties according to the physical laws of wireless propagation, providing a spatial basis for the differentiated calculation of each loss term in different mask areas.

[0164] In some embodiments, combined with Figure 2 ,like Figure 5 As shown, step 203 can also be achieved through steps 203a to 203d.

[0165] Step 203a: The residual map model training device extracts a non-line-of-sight predicted residual sub-map from the predicted residual map based on the non-line-of-sight region mask, and extracts a non-line-of-sight real residual sub-map from the real residual map based on the non-line-of-sight region mask. The non-line-of-sight predicted residual sub-map and the non-line-of-sight real residual sub-map have the same grid size and spatial coordinate range.

[0166] In some embodiments, the residual map model training device uses a non-line-of-sight region mask as the selection criterion, retaining the residual values ​​in the predicted residual map corresponding to the grids within the non-line-of-sight region mask, while excluding the remaining grids, thus obtaining a non-line-of-sight predicted residual sub-map. Similarly, a non-line-of-sight true residual sub-map is extracted from the true residual map using a non-line-of-sight region mask. Since both the non-line-of-sight predicted residual sub-map and the non-line-of-sight true residual sub-map are extracted based on the same non-line-of-sight region mask, they have the same grid size and spatial coordinate range, meaning the two sub-maps correspond one-to-one in spatial location. This allows subsequent grid-by-grid comparisons and structural comparisons between the non-line-of-sight predicted residual sub-map and the non-line-of-sight true residual sub-map to be performed within the same spatial range.

[0167] In some embodiments, the non-line-of-sight predicted residual subgraph represents the residual distribution in the non-line-of-sight region mask of the predicted residual graph; the non-line-of-sight true residual subgraph represents the residual distribution in the non-line-of-sight region mask of the true residual graph.

[0168] Step 203b: The residual map model training device performs local window normalization processing on the above-mentioned non-line-of-sight predicted residual sub-map and the above-mentioned non-line-of-sight real residual sub-map respectively to obtain normalized predicted residual sub-map and normalized real residual sub-map, and calculates the local structural similarity index between the above-mentioned normalized predicted residual sub-map and the above-mentioned normalized real residual sub-map.

[0169] In some embodiments, the residual map model training device performs local window normalization processing on the non-line-of-sight predicted residual submap and the non-line-of-sight true residual submap respectively. That is, the local window slides on the submap, and the residual value within the window is subtracted from the mean residual value within the window and divided by the standard deviation of the residual value within the window. This converts the residual values ​​within each window into a dimensionless relative fluctuation form. After traversing all windows, the normalized predicted residual submap and the normalized true residual submap are obtained. After the local window normalization processing, the differences in residual amplitude at different locations are eliminated, and the submap retains the brightness distribution and fluctuation pattern of the residual in the local space.

[0170] In some embodiments, the residual map model training device calculates a local structural similarity index between the normalized predicted residual submap and the normalized true residual submap. Specifically, it compares the mean similarity, variance similarity, and covariance correlation between the normalized predicted residual submap and the normalized true residual submap within corresponding windows, and summarizes the comparison results of each window into a local structural similarity index. The local structural similarity index measures the degree of consistency between the predicted occlusion shadow and the true occlusion shadow in terms of local structural morphology; a higher index value indicates a closer similarity in the local structures of the two submaps.

[0171] Step 203c: The residual map model training device performs gradient operator convolution processing on the above-mentioned non-line-of-sight predicted residual sub-map and the above-mentioned non-line-of-sight true residual sub-map respectively to obtain the predicted gradient direction map and the true gradient direction map, and calculates the gradient direction consistency metric between the above-mentioned predicted gradient direction map and the above-mentioned true gradient direction map.

[0172] In some embodiments, the residual map model training device performs gradient operator convolution processing on the non-line-of-sight predicted residual submap and the non-line-of-sight true residual submap respectively. Specifically, it convolves the non-line-of-sight predicted residual submap with difference operators in the horizontal and vertical directions, and then convolves it with the non-line-of-sight true residual submap to obtain the gradient components in both directions at each grid cell. Based on this, a predicted gradient direction map and a true gradient direction map are obtained, where the gradient direction map reflects the direction in which the residual value changes most steeply at each grid cell. For example, when an occlusion shadow transitions from dark to light along one side of a building, the residual gradient direction generally points towards the building side; the gradient direction map thus depicts the direction of this shadow transition.

[0173] In some embodiments, the residual map model training device calculates a gradient direction consistency metric between the predicted gradient direction map and the true gradient direction map. This involves comparing the angle between the predicted and true gradient directions at corresponding grid points; the smaller the angle, the more consistent the directions. The consistency results at each grid point are then summarized into a gradient direction consistency metric. This metric measures the degree of agreement between the predicted occlusion shadow transition direction and the true occlusion shadow transition direction.

[0174] Step 203d: The residual graph model training device weights and sums the above local structural similarity index and the above gradient direction consistency metric to obtain the above non-line-of-sight structural consistency loss.

[0175] In some embodiments, the residual graph model training device weights and sums the local structural similarity index and the gradient direction consistency measure, which is to multiply the local structural similarity index and the gradient direction consistency measure by their respective weight coefficients and then sum them to obtain the non-line-of-sight structural consistency loss. The weight coefficients are preset values ​​used to adjust the proportion of local structural morphology constraints and gradient direction constraints in the non-line-of-sight structural consistency loss.

[0176] In some embodiments, the non-line-of-sight structural consistency loss is a comprehensive measure of the differences between the predicted residual map and the actual residual map within the non-line-of-sight mask in terms of both local structural morphology and gradient direction. A lower value indicates that the occlusion shadow morphology output by the model is closer to the reality. This loss is not a simple weighting of the numerical error of the non-line-of-sight grid, but rather a direct constraint on the spatial morphology and structural continuity of the residual within the non-line-of-sight mask.

[0177] Thus, by imposing consistency constraints on the predicted residual map and the real residual map from two levels—local structural morphology and gradient direction—within the mask in the non-line-of-sight region, the occlusion shadows output by the model maintain a spatial morphology, transition direction, and structural continuity that are close to the real propagation distribution, thereby alleviating the problem of the weakening of the non-line-of-sight shadow structure when the global average error dominates the training.

[0178] In some embodiments, when performing step 205 to calculate the underpredicted one-sided risk loss of the predicted residual map and the true residual map within a sub-region of the non-line-of-sight mask, the following implementation methods may also be adopted, but are not limited to: filtering grids with occlusion lengths exceeding the occlusion length threshold from the non-line-of-sight mask based on the non-line-of-sight mask and the occlusion length threshold to obtain a sub-region of the non-line-of-sight mask, wherein the occlusion length threshold is set according to the carrier frequency and the average building height; extracting a deep occlusion predicted residual sub-map from the predicted residual map based on the sub-region of the non-line-of-sight mask, and extracting a deep occlusion predicted residual sub-map from the true residual map based on the deep occlusion mask. A deep occlusion true residual sub-map is extracted from the actual residual map; the difference between the deep occlusion predicted residual sub-map and the deep occlusion true residual sub-map is calculated to obtain a residual difference map; grids in the residual difference map where the predicted residual is less than the true residual and the difference exceeds the tolerance threshold are marked as under-predicted grids, the tolerance threshold being set based on the path loss measurement accuracy and coverage assessment tolerance; a one-sided penalty function is applied to the difference magnitude at the under-predicted grids to obtain the penalty value for each under-predicted grid, and the penalty values ​​of all the under-predicted grids are summed to obtain the under-predicted one-sided risk loss; the one-sided penalty function applies a linear or non-linear incremental penalty to the underestimated magnitude.

[0179] In some embodiments, the residual graph model training device filters out grids in the non-line-of-sight region mask whose occlusion length exceeds the occlusion length threshold based on the non-line-of-sight region mask and the occlusion length threshold. Specifically, for each grid in the non-line-of-sight region mask, the corresponding occlusion length is read. If the occlusion length exceeds the occlusion length threshold, the grid is retained in the sub-region; if the occlusion length does not exceed the occlusion length threshold, the grid is removed from the sub-region. After traversing all grids in the non-line-of-sight region mask, the sub-regions of the non-line-of-sight region mask are obtained. Since the occlusion length has already been obtained when determining visibility information in step 202b, this filtering does not require recalculating the propagation path.

[0180] In some embodiments, the obstruction length threshold is a threshold value used to distinguish between deep obstruction and general obstruction. The obstruction length threshold is set according to the carrier frequency and the average building height. For example, the higher the carrier frequency and the greater the average building height, the deeper the signal attenuation under the same obstruction length. The obstruction length threshold can be adjusted accordingly to ensure that the sub-region covers the grid blocked by thicker buildings. The obstruction length, as previously described, characterizes the length of a straight path traveling inside a building.

[0181] In some embodiments, the residual map model training device extracts the residual values ​​corresponding to the grids located in the sub-region of the predicted residual map based on the sub-region of the non-line-of-sight region mask to obtain a deep occlusion prediction residual sub-map; similarly, a deep occlusion real residual sub-map is extracted from the real residual map based on the deep occlusion region mask, where the deep occlusion region mask is the mask corresponding to the sub-region of the non-line-of-sight region mask, and the two sub-maps have the same grid size and spatial coordinate range, and can be compared grid by grid.

[0182] In some embodiments, the deep occlusion prediction residual subgraph represents the residual distribution within a sub-region of the prediction residual map located in the non-line-of-sight mask; the deep occlusion true residual subgraph represents the residual distribution within a sub-region of the true residual map located in the non-line-of-sight mask.

[0183] In some embodiments, the residual map model training device performs a difference operation between the deep occlusion predicted residual sub-map and the deep occlusion true residual sub-map. That is, at the corresponding grid, the predicted residual value is subtracted from the true residual value, and after traversing all grids in the sub-region, a residual difference map is obtained. Subsequently, the residual map model training device searches for grids in the residual difference map where the predicted residual is less than the true residual and the difference exceeds a tolerance threshold. When the predicted residual at a grid is lower than the true residual and the difference exceeds the tolerance threshold, the grid is marked as an underpredicted grid. When the predicted residual at a grid is higher than the true residual, or the underestimation does not exceed the tolerance threshold, the grid is not marked as an underpredicted grid.

[0184] In some embodiments, the tolerance threshold is a threshold value for determining whether an underestimation exceeds the limit. The tolerance threshold is set based on the path loss measurement accuracy and coverage assessment tolerance, that is, taking into account the error level inherent in the field measurement or ray tracing simulation itself, and the allowable deviation range of the coverage assessment. Underestimations within the normal error range are excluded, and only grates with underestimations exceeding this range are penalized. For example, if the tolerance threshold is 5 dB, and the true residual at a certain grate is 40 dB and the predicted residual is 28 dB, the underestimation is 12 dB and exceeds the tolerance threshold, then the grate is marked as an underpredicted grate; if the underestimation at another grate is 3 dB and does not exceed the tolerance threshold, then the grate is not marked.

[0185] In some embodiments, the residual map model training device performs a one-sided penalty function operation on the difference magnitude at the underpredicted grid, that is, inputs the underestimated magnitude at the underpredicted grid into the one-sided penalty function, and outputs the penalty value of the grid by the one-sided penalty function to obtain the penalty value of each underpredicted grid; then the penalty values ​​of each underpredicted grid are summed, that is, the penalty values ​​of all underpredicted grids in the sub-region are accumulated to obtain the underpredicted one-sided risk loss.

[0186] In some embodiments, the one-sided penalty function applies a linear or non-linear increasing penalty to the underestimation magnitude; that is, the larger the underestimation magnitude, the higher the penalty value, and the two exhibit a linear or non-linear increasing relationship. For example, the penalty value can be linearly proportional to the underestimation magnitude, or non-linearly increase with the square of the underestimation magnitude. The one-sided penalty function does not output a penalty value for overestimation cases where the predicted residual is greater than the actual residual; that is, overestimated grates are not included in the summation, thus ensuring that the underestimation one-sided risk loss only constrains the underestimation direction. The underestimation one-sided risk loss represents the sum of the penalty values ​​for all underestimated grates when the residual is underestimated at severely occluded grates; a higher value indicates a more severe underestimation in the deeply occluded region.

[0187] Thus, by applying a unidirectional incremental penalty to under-predicted graticules whose underestimation exceeds the tolerance threshold within the sub-region of the non-line-of-sight mask, the model tends to avoid residual underestimation at deeply occluded locations, reducing the risk of the trained model underestimating path loss in coverage evaluation.

[0188] In some embodiments, when performing step 204 to calculate the consistency loss of the building boundary gradient abrupt change in the predicted residual map and the true residual map within the building boundary region mask, the following implementation methods may also be adopted, but are not limited to: extracting a boundary predicted residual sub-map from the predicted residual map based on the building boundary region mask, and extracting a boundary true residual sub-map from the true residual map based on the building boundary region mask, wherein the boundary predicted residual sub-map and the boundary true residual sub-map cover the extended zone regions on both sides of the building boundary; and processing the boundary predicted residual sub-map and the true residual sub-map... The aforementioned boundary true residual sub-maps are subjected to high-pass filtering to obtain predicted gradient jump component maps and true gradient jump component maps, and the gradient jump component difference between the predicted gradient jump component maps and the true gradient jump component maps is calculated. Convolution processing is performed on the aforementioned boundary predicted residual sub-maps and the aforementioned boundary true residual sub-maps to obtain predicted response maps and true response maps, and the response difference between the aforementioned predicted response maps and the true response maps is calculated. The gradient jump component difference and the response difference are weighted and summed to obtain the aforementioned building boundary gradient abrupt change consistency loss.

[0189] In some embodiments, the residual map model training device uses a building boundary region mask as a selection criterion to extract the residual values ​​corresponding to the grids located within the building boundary region mask in the predicted residual map, thus obtaining a boundary predicted residual sub-map. Similarly, a boundary true residual sub-map is extracted from the true residual map using the building boundary region mask. Since the building boundary region mask has been extended to both sides of the building boundary by a preset pixel width in step 202c, the boundary predicted residual sub-map and the boundary true residual sub-map cover the extended band area on both sides of the building boundary. That is, the sub-map simultaneously contains the residual values ​​inside and outside the building outline, which can completely present the transition of the residual from one side to the other at the building boundary. Moreover, the two sub-maps have the same grid size and spatial coordinate range, and can be compared grid by grid.

[0190] In some embodiments, the boundary prediction residual subgraph represents the residual distribution within the building boundary region mask in the prediction residual graph; the boundary true residual subgraph represents the residual distribution within the building boundary region mask in the true residual graph.

[0191] In some embodiments, the residual map model training device performs high-pass filtering on the boundary predicted residual submap and the boundary true residual submap, respectively. High-pass filtering is a filtering operation that allows high-frequency components in the residuals to pass through while suppressing low-frequency components. Since the steep jumps in the residuals at the building boundary are high-frequency components and the gradual background undulations are low-frequency components, after high-pass filtering, the gradually changing background portion of the residuals is filtered out, while the jump portion at the boundary is retained, resulting in the predicted gradient jump component map and the true gradient jump component map. For example, the true residual is 15 dB on one side of the building boundary and jumps to 35 dB after crossing the boundary. This jump is represented by a concentrated high response in the true gradient jump component map, while the response in the surrounding gradually changing region is lower.

[0192] In some embodiments, the residual map model training device calculates the gradient jump component difference between the predicted gradient jump component map and the true gradient jump component map. That is, it calculates the deviation between the predicted gradient jump component and the true gradient jump component at the corresponding grid and summarizes it as the gradient jump component difference. This difference reflects the degree of deviation between the jump intensity of the predicted residual map at the building boundary and the true jump intensity.

[0193] In some embodiments, the residual map model training device performs convolution processing on the boundary prediction residual submap and the boundary true residual submap, respectively. Specifically, a convolution kernel used for detecting edge jumps is slidably convolved on the submap. The convolution kernel outputs a higher response value when passing through a residual jump location and a lower response value when passing through a smooth location, thereby obtaining the predicted response map and the true response map. The convolution kernel used for edge jump detection can be, for example, the Sobel operator or the Laplace operator; the type of convolution kernel does not affect the purpose of this processing to detect boundary jumps.

[0194] In some embodiments, the residual map model training device calculates the response difference between the predicted response map and the true response map, that is, it calculates the deviation between the predicted response and the true response at the corresponding grid and summarizes it as the response difference. This difference reflects the degree of deviation between the predicted residual map and the actual situation in terms of the jump position and jump direction at the building boundary from the perspective of edge detection.

[0195] In some embodiments, the residual graph model training device weights and sums the gradient jump component differences and response differences, which is to multiply the gradient jump component differences and response differences by their respective weight coefficients and then sum them to obtain the building boundary gradient mutation consistency loss. The weight coefficients are preset values ​​used to adjust the proportion of jump component constraints and edge response constraints in the building boundary gradient mutation consistency loss.

[0196] In some embodiments, the building boundary gradient abrupt change consistency loss is a comprehensive measure of the difference between the predicted residual map and the true residual map within the mask of the building boundary region in terms of gradient jump components and edge response. The lower the value, the closer the numerical transition of the predicted residual map at the building boundary is to the true residual map.

[0197] Thus, by imposing constraints on the predicted residual map and the true residual map from two aspects, namely gradient jump components and edge response, within the mask of the building boundary region, the residual output by the model maintains a steep numerical transition in the vicinity of the building outline that is comparable to the true propagation distribution, thereby alleviating the problem of the boundary transition being blurred into a gentle gradient.

[0198] In some embodiments, when performing step 204 to calculate the multi-scale local attenuation consistency loss between the predicted residual map and the true residual map within the multi-scale local attenuation region mask, the following implementation methods may also be adopted, but are not limited to: For example, based on the local attenuation region corresponding to any window scale in the multi-scale local attenuation region mask, a predicted attenuation sub-map at that scale is extracted from the predicted residual map, and a true attenuation sub-map at that scale is extracted from the true residual map based on the local attenuation region corresponding to that window scale; window mean calculation is performed on the predicted attenuation sub-map and the true attenuation sub-map respectively to obtain a predicted local mean sequence and a true local mean sequence, and the mean difference between the predicted local mean sequence and the true local mean sequence is calculated; gradient magnitude calculation is performed on the predicted attenuation sub-map and the true attenuation sub-map respectively to obtain a predicted gradient magnitude sequence and a true gradient magnitude sequence, and the magnitude difference between the predicted gradient magnitude sequence and the true gradient magnitude sequence is calculated; the mean difference and the magnitude difference corresponding to each window scale are summed to obtain the multi-scale local attenuation consistency loss.

[0199] In some embodiments, the residual map model training device extracts the residual values ​​corresponding to the grids located within the local attenuation region of any window scale in the multi-scale local attenuation region mask to obtain a predicted attenuation sub-map at that scale. Similarly, a true attenuation sub-map at that scale is extracted from the true residual map based on the local attenuation region corresponding to that window scale. Since the multi-scale local attenuation region mask is obtained by merging the local attenuation regions at each window scale in step 202e, the sub-maps are extracted separately according to the window scale. Each window scale corresponds to a pair of predicted attenuation sub-maps and true attenuation sub-maps. The two sub-maps have the same grid size and spatial coordinate range, and can be compared grid by grid.

[0200] In some embodiments, the predicted attenuation subplot represents the residual distribution within a local attenuation region corresponding to a certain window scale in the predicted residual plot; the true attenuation subplot represents the residual distribution within a local attenuation region corresponding to the same window scale in the true residual plot.

[0201] In some embodiments, the residual map model training device performs window mean calculations on the predicted attenuation submap and the true attenuation submap, respectively. Specifically, it slides a window of the same scale as any given window across the submap, calculates the average residual value within each window, and arranges the average values ​​of each window according to their window positions to obtain a predicted local mean sequence and a true local mean sequence. The predicted local mean sequence and the true local mean sequence reflect the overall attenuation level in each local region at that scale.

[0202] In some embodiments, the residual graph model training device calculates the mean difference between the predicted local mean sequence and the true local mean sequence, that is, it calculates the deviation between the predicted local mean and the true local mean at the corresponding window position and summarizes them into a mean difference. This mean difference reflects the degree of deviation between the overall level of predicted decay and the overall level of true decay at this scale.

[0203] In some embodiments, the residual map model training device performs gradient magnitude calculations on the predicted attenuation submap and the true attenuation submap, respectively. Specifically, it calculates the square root of the sum of the squares of the gradient components of the residuals at each grid point in the submap in both the horizontal and vertical directions to obtain the gradient magnitude at each grid point. These gradient magnitudes are then arranged spatially to obtain the predicted gradient magnitude sequence and the true gradient magnitude sequence. The predicted and true gradient magnitude sequences reflect the strength of the spatial variation in attenuation at that scale.

[0204] In some embodiments, the residual graph model training device calculates the magnitude difference between the predicted gradient magnitude sequence and the true gradient magnitude sequence, that is, it calculates the deviation between the predicted gradient magnitude and the true gradient magnitude at the corresponding position and summarizes them into a magnitude difference. This magnitude difference reflects the degree of deviation between the strength of the predicted decay change and the strength of the true decay change at this scale.

[0205] In some embodiments, the residual map model training device sums the mean difference and amplitude difference corresponding to each window scale. Specifically, it first obtains the mean difference and amplitude difference for each window scale, and then sums the mean difference and amplitude difference for all window scales to obtain the multi-scale local attenuation consistency loss. For example, when using 3×3, 7×7, and 15×15 grid windows, the mean difference and amplitude difference for each of the three scales are obtained separately, and then the six differences are summed to obtain the multi-scale local attenuation consistency loss.

[0206] In some embodiments, the multi-scale local decay consistency loss characterizes the comprehensive measure of the difference between the predicted residual map and the true residual map in terms of both local decay level and decay change strength at multiple window scales. The lower the value, the closer the local decay pattern of the predicted residual map is to the true residual map at each spatial scale.

[0207] In this way, by constraining the overall level and intensity of local attenuation at multiple window scales, the predicted residual map maintains the attenuation patch morphology and transition intensity consistent with the true residual distribution at each spatial scale, thus alleviating the problem that local attenuation details at different ranges are canceled out in the overall average and are difficult to be stably learned by the model.

[0208] The following describes the residual map model training method of this application through specific embodiments. It includes: This invention provides a training method for the physical consistency of propagation area masks for radio map residual prediction. It is applicable to grid prediction models that output any radio map or path loss residual map. The aforementioned model can be CNN, U-Net, Transformer, graph neural network, encoder-decoder network, or a combination thereof; its input can be map information, transmitter information, propagation features, link features, or a combination thereof. This invention is particularly applicable to models that use propagation mechanism grouping input and link-level occlusion profile encoding, but such models are not a necessary prerequisite. The input of this method includes the model to be trained, the model prediction residual map, the true residual map, the building occupancy map, the building height map, the transmitter location, the receiver grid coordinates, the path occupancy statistics, visibility information, local gradient information, the path loss threshold, and the effective prediction area information. The output is a composite loss function used to train the model. This method includes the following steps: S1: Obtain the model to be trained and the residual map. For a training sample, input map information, transmitter information, propagation features, or a combination thereof into the model to be trained to obtain the prediction residual map. S2: Construct an effective prediction region mask. An effective prediction region mask is generated based on the evaluable region, outdoor region, or target prediction region in the samples, excluding invalid regions inside buildings, unlabeled regions, or regions not participating in training. S3: Construct a non-line-of-sight region mask. Based on the visibility from the transmitter to the receiver grid, building occupancy, occupancy length, number of obstructions, or superelevation, determine whether each receiver grid belongs to a non-line-of-sight region. In one embodiment, if the straight path from the transmitter to the receiver grid intersects with the building occupancy map, and the occupancy grid height exceeds the line-of-sight height or the occupancy length exceeds a threshold, then the receiver grid is marked as a non-line-of-sight region. S4: Construct a building boundary region mask. The building boundary is calculated based on the building occupancy map or building height map, and extended to both sides of the boundary by a certain width to form a building boundary band. In one embodiment, the building occupancy map can first be dilated and eroded, then the difference can be taken to obtain the morphological boundary, and then extended by a preset pixel width to form a boundary region mask. S5: Construct a multi-scale local attenuation region mask. Based on maps, occlusion statistics, visibility, local gradients, or path loss thresholds, generate local attenuation region masks at multiple window scales. For example, for each window scale, the occlusion ratio, building density, visibility degradation, local gradient magnitude, or path loss threshold exceeding the limit can be calculated, and areas with significant local attenuation can be selected as masks. Masks at multiple scales can be used separately for loss calculation or merged into a unified mask. The ground truth residual map and model attention map can be used as alternatives but are not necessary sources for the primary solution. S6: Calculate the data fitting loss. Calculate the data fitting loss between the predicted residual map and the ground truth residual map within the effective prediction region.Data fitting loss can be Smooth-L1, mean squared error, mean absolute error, or Huber loss. S7: Calculate non-line-of-sight structural consistency loss. Within the mask in the non-line-of-sight region, compare the local structural similarity, correlation, gradient direction, or local ranking relationship between the predicted residual map and the true residual map to ensure the model maintains a reasonable spatial shape in occluded shadow areas. This loss does not simply increase the weights of non-line-of-sight pixels, but rather constrains the consistency of the residual structure in the non-line-of-sight region. S8: Calculate underpredicted one-sided risk loss. In non-line-of-sight, deeply occluded, or high-risk engineering areas, a one-sided penalty is generated when the predicted path loss is lower than the true path loss and exceeds the tolerance threshold. This loss is used to reduce the risk of underestimating path loss in deeply occluded areas. If the predicted path loss is higher than the true path loss or the underestimation does not exceed the threshold, this penalty can be reduced or not generated. S9: Calculate high-frequency consistency loss at building boundaries. Within the building boundary region mask, compare the local gradients, high-frequency components, boundary direction responses, or Laplace responses of the predicted and actual residual maps to ensure the model maintains reasonable abrupt changes at building occlusion edges and reduces oversmoothing. S10: Calculate multi-scale local attenuation consistency loss. Constrain the residual mean, gradient magnitude, correlation, local ranking, or structural similarity within local attenuation regions at multiple window scales to maintain consistency in attenuation patches and boundary transitions across different spatial scales. S11: Combine training objectives and update the model. Combine the data fitting loss, non-line-of-sight structural consistency loss, underpredicted one-sided risk loss, high-frequency consistency loss at building boundaries, and multi-scale local attenuation consistency loss into a composite training objective, and update the model parameters through backpropagation. Weights can be fixed, or alternative methods such as stage scheduling, gradient normalization, uncertainty weighting, or task balancing can be used.

[0209] In a specific application scenario of the method in this application, taking urban outdoor path loss residual prediction as an example, the model to be trained can be U-Net, Transformer, or other grid prediction networks. The model input can be a building map, transmitter location, distance map, propagation features, or propagation mechanism grouping features and link-level occlusion profile encoding features. The model output is a path loss residual map. During training, an effective prediction region mask is first generated based on the building occupancy map and the evaluable area. Then, a non-line-of-sight region mask is generated based on the intersection relationship between the straight path from the transmitter to each receiving grid and the building occupancy map, the occlusion length, and the superelevation. Next, a building boundary region mask is generated based on the morphological gradient of the building occupancy map, and it is extended to both sides of the boundary to form a boundary band. Furthermore, based on building density, occlusion statistics, visibility degradation degree, local gradient, and path loss threshold, local attenuation region masks are generated at multiple window scales. In the loss calculation phase, the Smooth-L1 data fitting loss is calculated within the effective region; the structural consistency loss is calculated within the non-line-of-sight mask; the underpredicted one-sided risk loss is calculated within the non-line-of-sight or deeply occluded regions; the gradient or high-frequency consistency loss is calculated within the building boundary region; and the local mean, gradient, or correlation consistency loss is calculated within the multi-scale local attenuation region. Finally, the various losses are combined into a composite training objective, and the model parameters are updated through backpropagation. This embodiment demonstrates that the present invention can be used as a training method for arbitrary residual radio map prediction models, and also as a training enhancement module for models employing propagation mechanism grouped input and link-level profile coding.

[0210] Figure 6This embodiment provides a flowchart for generating a propagation area mask, which shows that: First, map and building information, as well as transmitter and receiver grid information, are acquired. The map and building information may include building occupancy maps, building height maps, geographic grids, and outdoor effective areas; the transmitter and receiver grid information may include transmitter location, transmitter height, receiver grid coordinates, receiver height, and related propagation relationship information. Based on the above inputs, the path occlusion and visibility calculation module calculates the occlusion relationship, visibility state, occlusion length, superelevation, or other propagation visibility indicators between the transmitter and each receiver grid, according to the building map, transmitter location, and receiver grid location. This calculation result, along with the map and building information, serves as the basis for generating the propagation area mask. Subsequently, this invention generates an effective prediction area mask, a non-line-of-sight area mask, a building boundary area mask, and a multi-scale local attenuation area mask, respectively. Among them, the effective prediction region mask is used to limit the effective grid range for training; the non-line-of-sight region mask is used to identify non-line-of-sight regions determined by building occlusion or visibility relationships; the building boundary region mask is used to identify areas of building outlines, abrupt changes in building height, or building boundary extension zones; and the multi-scale local attenuation region mask is used to identify local attenuation regions determined by occlusion statistics, visibility degradation, local gradients, or path loss thresholds. These different types of region masks are collectively incorporated into the propagation region mask set for subsequent physical consistency training. Figure 6 This corresponds to steps S2 to S5. Specifically, S2 corresponds to generating the mask for the effective prediction area, S3 to generating the mask for the non-line-of-sight area, S4 to generating the mask for the building boundary area, and S5 to generating the mask for the multi-scale local attenuation area. Figure 6 It can be seen that the propagation area mask in this invention is not the result of manual annotation, but is automatically calculated from the map, buildings, transmitter, receiver grid and propagation visibility relationship.

[0211] Figure 7This embodiment provides a flowchart for generating a propagation region mask, illustrating the following: A residual prediction model to be trained is used as the training object. After receiving inputs such as map information, transmitter information, propagation features, or link features, the residual prediction model outputs a predicted residual map. The corresponding real residual map and predicted residual map in the training data serve as the basis for numerical comparison of various loss terms; the propagation region mask set is used to define the effective area of ​​different loss terms. In the loss calculation stage, this invention sets data fitting loss, non-line-of-sight structural consistency loss, underpredicted unilateral risk loss, high-frequency consistency loss at building boundaries, and multi-scale local attenuation consistency loss. The data fitting loss applies to the effective prediction region, constraining the overall numerical difference between the predicted residual map and the true residual map; the non-line-of-sight structural consistency loss applies to the non-line-of-sight mask region, constraining the non-line-of-sight shadow structure and local spatial morphology; the underprediction unilateral risk loss applies to deeply occluded areas or high-risk engineering areas, reducing the coverage assessment risk caused by underestimated path loss; the building boundary high-frequency consistency loss applies to the building boundary mask region, maintaining high-frequency changes near the building occlusion edge; and the multi-scale local attenuation consistency loss applies to the multi-scale local attenuation mask region, constraining the structural consistency of local attenuation patches at different scales. All these losses are input into the composite training objective module, forming the total loss used for model training. The composite training objective updates the model parameters through backpropagation and feeds the updated parameters back to the residual prediction model to be trained. The dashed arrows in the figure represent the backpropagation update process, illustrating the relationship between the training objective and the model parameters.

[0212] Figure 7 This corresponds to steps S1, S6 to S11 in the main text. Specifically, S1 corresponds to generating the residual prediction model to be trained and the prediction residual map; S6 corresponds to calculating the data fitting loss; S7 corresponds to calculating the non-line-of-sight structural consistency loss; S8 corresponds to calculating the under-prediction unilateral risk loss; S9 corresponds to calculating the high-frequency consistency loss of the building boundary; S10 corresponds to calculating the multi-scale local attenuation consistency loss; and S11 corresponds to constructing the composite training objective and updating the model parameters. Figure 6 and Figure 7 As can be seen, the present invention first generates a set of propagation area masks from maps, buildings, transceiver relationships and propagation visibility relationships, and then applies different propagation area masks to the corresponding physical consistency loss terms, so that the residual radio map prediction model pays attention to the overall numerical error, non-line-of-sight structure, underprediction risk, high-frequency changes of building boundaries and multi-scale local attenuation consistency during the training process.

[0213] In some embodiments, the non-line-of-sight region mask can be generated from straight path occlusion, visibility determination, hyperscale, occlusion length, ray tracing results, link-level profile results, or model estimation visibility.

[0214] In some embodiments, the building boundary region mask may be generated from the morphological gradient of the building occupancy map, the building height gradient, the road boundary, the wall normal, the semantic segmentation result, or the boundary of the 3D map.

[0215] In some embodiments, the multi-scale local attenuation region mask can be generated from a map, occlusion statistics, visibility, local gradient, or path loss threshold; alternatively, it can also be generated with the assistance of a ground truth residual map, a prediction error map, or a model attention heatmap.

[0216] In some embodiments, structural consistency indices may include correlation coefficient, structural similarity, gradient consistency, frequency domain consistency, local ranking consistency, local mean consistency, or edge response consistency.

[0217] In some embodiments, underpredicted one-sided risk loss may be achieved using hinge loss, Smooth-L1, Huber loss, quantile loss, or asymmetric weighted loss.

[0218] In some embodiments, multi-scale local attenuation consistency can be achieved using multi-window mean difference, gradient difference, local correlation, local sorting constraint, pyramid loss, or wavelet / frequency domain loss.

[0219] In some embodiments, the loss weights may be fixed weights, or they may be adjusted by stage scheduling, gradient normalization, uncertainty weighting, task balancing, or relative loss reduction rate.

[0220] In some embodiments, the applicable model is not limited to a certain structure and can be used with CNN, U-Net, Transformer, graph neural network, diffusion model, encoder-decoder network or hybrid model; the input can be map information, transmitter information, propagation features, link features or a combination thereof.

[0221] It should be noted that the above-described method embodiments, or the various possible implementations of the method embodiments, can be executed individually, or, provided there is no conflict, they can be combined with each other. The specific implementation can be determined according to actual usage requirements, and this application embodiment does not impose any restrictions on this.

[0222] As can be seen, the above mainly describes the solutions provided by the embodiments of this application from a methodological perspective. To achieve the above functions, the embodiments of this application provide corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the modules and algorithm steps of the various examples described in the embodiments disclosed herein, the embodiments of this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0223] This application embodiment can divide the residual graph model training method into functional modules based on the above method example. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. Optionally, the module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0224] In some embodiments, this application also provides a residual graph model training apparatus. The residual graph model training apparatus may include one or more functional modules for implementing the residual graph model training method of the above method embodiments.

[0225] For example, Figure 8 This is a schematic diagram of a residual graph model training device provided in an embodiment of this application. Figure 8 As shown, the residual graph model training device 900 includes:

[0226] The input module 901 is used to input training samples consisting of map information of the target area, radio transmitter information and radio propagation characteristics into the model to be trained, to obtain the prediction residual map, and to obtain the real residual map based on the real path loss of radio propagation in the target area and the basic propagation prior.

[0227] The construction module 902 is used to construct a set of propagation area masks based on the building occupancy map of the target area, the location of the radio transmitter, the coordinates of the radio receiver grid, the building occupancy relationship and the visibility information from the transmitter to the receiver grid. The set of propagation area masks includes an effective prediction area mask, a non-line-of-sight area mask, a building boundary area mask and a multi-scale local attenuation area mask.

[0228] The first calculation module 903 is used to calculate the data fitting loss between the predicted residual map and the real residual map within the effective prediction region mask, and to calculate the non-line-of-sight structural consistency loss between the predicted residual map and the real residual map within the non-line-of-sight region mask.

[0229] The second calculation module 904 is used to calculate the consistency loss of the building boundary gradient abrupt change between the predicted residual map and the real residual map within the building boundary region mask, and to calculate the multi-scale local attenuation consistency loss between the predicted residual map and the real residual map within the multi-scale local attenuation region mask.

[0230] The combination module 905 is used to calculate the underpredicted one-sided risk loss of the predicted residual map and the real residual map in the sub-region of the non-line-of-sight mask, and combine the data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss into a composite training objective, and train and update the parameters of the model to be trained based on the composite training objective.

[0231] The residual map model training device provided in this application obtains a predicted residual map by inputting training samples composed of map information of the target area, radio transmitter information, and radio propagation characteristics into the model to be trained. The true residual map is obtained based on the actual path loss of radio propagation within the target area and the basic propagation prior, allowing the model to focus on learning propagation deviations that the basic propagation prior cannot characterize, providing a learning object for complex attenuation caused by building occupancy. Based on this, a propagation area mask set is constructed according to the building occupancy map of the target area, the radio transmitter location, the radio receiver grid coordinates, building occupancy relationships, and the visibility information from the transmitter to the receiver grid. The target area is divided into four different types of regions: effective prediction area mask, non-line-of-sight area mask, building boundary area mask, and multi-scale local attenuation area mask. This explicitly identifies key spatial locations related to occupancy, providing a spatial basis for differentiated constraints. Furthermore, the data fitting loss between the predicted residual map and the true residual map is calculated within the effective prediction area mask to ensure that the overall numerical accuracy is not weakened. Simultaneously, the calculation of the predicted residual map is advanced to the non-line-of-sight area mask. The non-line-of-sight structural consistency loss between the residual map and the true residual map directly constrains the residual spatial morphology in areas severely occluded by buildings. Further, the consistency loss of the building boundary gradient abrupt change between the predicted and true residual maps is calculated within a mask in the building boundary region. This forces the predicted residual map to reproduce the steep numerical transition of the true residual map near the building outline, preventing it from being blurred into a gentle gradient. Furthermore, the multi-scale local attenuation consistency loss between the predicted and true residual maps is calculated within a mask in multi-scale local attenuation regions, ensuring stable constraints on attenuation details at different spatial scales. To maintain the spatial continuity of the residual distribution in the occluded region, the underpredicted one-sided risk loss is calculated within the sub-regions of the mask in the non-line-of-sight region, applying additional penalties to residual distortions in the most severely occluded areas that are not adequately corrected by other losses. Finally, the data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss are combined into a composite training objective. The parameters of the model under training are updated based on the composite training objective, allowing the constraints of each region to work synergistically on the same model. As a result, model training is no longer dominated by the overall average error in smooth regions. The spatial structure of severely occluded regions, the steepness of numerical transitions at building outlines, and the continuity of multi-scale local decays are all directly and progressively reinforced by constraints. This solves the problem that existing methods often result in significant distortions in the shadow morphology and spatial continuity of predictions in severely occluded areas compared to the actual propagation distribution, and that the numerical transitions near building outlines, which should be steep, are blurred into gentle gradients.

[0232] In other embodiments, the map information includes a building occupancy map and a building height map; the radio transmitter information includes the transmitter location; the radio propagation characteristics include a distance map from the transmitter to the receiving grid; the receiving grid coordinates are geometric coordinate data representing each receiving location grid; the distance map is used to reflect the propagation spatial conditions between the transmitter and receiver; the input module 901 is specifically used for:

[0233] Spatial resolution unification and grid coordinate alignment are performed on the building occupancy map, building height map, transmitter location, receiver grid coordinates, and distance map from transmitter to receiver grid of the training samples to obtain input feature maps with unified spatial resolution.

[0234] The encoder of the model to be trained performs multi-scale downsampling feature extraction on the input feature map to obtain a multi-scale radio propagation feature map. The decoder of the model to be trained performs upsampling and cross-layer feature fusion on the multi-scale radio propagation feature map to obtain the prediction residual map.

[0235] Obtain the above-mentioned real path loss map corresponding to the above training samples. The above-mentioned real path loss map is generated from field measurement data or ray tracing simulation data. The above-mentioned basic propagation prior map is calculated according to the free space propagation loss formula based on the transmitter power, transmitter antenna gain, receiver antenna gain and carrier frequency.

[0236] By performing a grid-by-grid interpolation operation on the above-mentioned true path loss map and the above-mentioned basic propagation prior map in the same geographic grid coordinate system, the above-mentioned true residual map is obtained.

[0237] In other embodiments, the above-described building module 902 is specifically used for:

[0238] The above effective prediction region mask is generated based on the evaluable area and outdoor area in the above training sample. The above effective prediction region mask is the area after excluding the building interior area and the unlabeled area.

[0239] The visibility information of each receiving grid is determined based on the intersection of the straight path from the transmitter to each receiving grid with the building occupancy map, the obstruction length, and the superelevation. The receiving grids that are not line-of-sight in the visibility information are marked as the non-line-of-sight area mask. The straight path is determined based on the transmitter position and the receiving grid coordinates.

[0240] Morphological gradient operations are performed on the building occupancy map to obtain the building boundary, and a preset pixel width is extended to both sides of the building boundary to obtain the building boundary region mask;

[0241] The occlusion ratio, building density, visibility degradation and local gradient magnitude of the received grid within each window scale are calculated at multiple window scales to obtain a multi-scale attenuation feature sequence.

[0242] The receiving grid corresponding to any feature in the above multi-scale attenuation feature sequence that exceeds the corresponding threshold is marked as a local attenuation region, and the local attenuation regions under each window scale are merged to obtain the above multi-scale local attenuation region mask.

[0243] In other embodiments, the first computing module 903 described above is specifically used for:

[0244] Based on the above non-line-of-sight region mask, a non-line-of-sight predicted residual sub-image is extracted from the above predicted residual image, and a non-line-of-sight real residual sub-image is extracted from the above real residual image based on the above non-line-of-sight region mask. The above non-line-of-sight predicted residual sub-image and the above non-line-of-sight real residual sub-image have the same grid size and spatial coordinate range.

[0245] Local window normalization is performed on the above non-line-of-sight predicted residual sub-map and the above non-line-of-sight true residual sub-map respectively to obtain normalized predicted residual sub-map and normalized true residual sub-map, and the local structural similarity index between the above normalized predicted residual sub-map and the above normalized true residual sub-map is calculated.

[0246] Gradient operator convolution processing is performed on the above non-line-of-sight predicted residual sub-map and the above non-line-of-sight true residual sub-map respectively to obtain the predicted gradient direction map and the true gradient direction map, and the gradient direction consistency metric between the above predicted gradient direction map and the above true gradient direction map is calculated.

[0247] The above local structural similarity index and the above gradient direction consistency metric are weighted and summed to obtain the above non-line-of-sight structural consistency loss.

[0248] In other embodiments, the above-described combination module 905 is specifically used for:

[0249] Based on the aforementioned non-line-of-sight region mask and occlusion length threshold, grids with occlusion lengths exceeding the aforementioned occlusion length threshold are filtered from the aforementioned non-line-of-sight region mask to obtain the sub-region of the aforementioned non-line-of-sight region mask. The aforementioned occlusion length threshold is set according to the carrier frequency and the average building height.

[0250] Based on the sub-region of the above non-line-of-sight region mask, a deep occlusion prediction residual sub-image is extracted from the above prediction residual image, and based on the above deep occlusion region mask, a deep occlusion real residual sub-image is extracted from the above real residual image.

[0251] The difference between the above deep occlusion prediction residual submap and the above deep occlusion actual residual submap is calculated to obtain the residual difference map. The grids in the above residual difference map where the predicted residual is less than the actual residual and the difference exceeds the tolerance threshold are marked as under-predicted grids. The tolerance threshold is set according to the path loss measurement accuracy and coverage assessment tolerance.

[0252] A one-sided penalty function is applied to the difference magnitude at the aforementioned under-predicted grids to obtain the penalty value for each under-predicted grid. The penalty values ​​of all the aforementioned under-predicted grids are then summed to obtain the aforementioned one-sided risk loss for under-prediction. The aforementioned one-sided penalty function applies a linear or non-linear incremental penalty to the underestimated magnitude.

[0253] In other embodiments, the second computing module 904 described above is specifically used for:

[0254] Based on the above building boundary region mask, a boundary prediction residual sub-map is extracted from the above prediction residual map, and a boundary real residual sub-map is extracted from the above real residual map based on the above building boundary region mask. The above boundary prediction residual sub-map and the above boundary real residual sub-map cover the extended zone areas on both sides of the building boundary.

[0255] High-pass filtering is performed on the above-mentioned boundary prediction residual sub-map and the above-mentioned boundary true residual sub-map respectively to obtain the predicted gradient jump component map and the true gradient jump component map, and the gradient jump component difference between the above-mentioned predicted gradient jump component map and the above-mentioned true gradient jump component map is calculated.

[0256] Convolution processing is performed on the above boundary prediction residual submap and the above boundary true residual submap respectively to obtain the prediction response map and the true response map, and the response difference between the above prediction response map and the above true response map is calculated.

[0257] The weighted summation of the gradient jump component differences and the response differences yields the building boundary gradient abrupt change consistency loss.

[0258] In other embodiments, the second computing module 904 described above is specifically used for:

[0259] Based on the local attenuation region corresponding to any window scale in the above multi-scale local attenuation region mask, a predicted attenuation sub-map at that scale is extracted from the above predicted residual map, and a real attenuation sub-map at that scale is extracted from the above real residual map based on the local attenuation region corresponding to that window scale.

[0260] Window mean calculation is performed on the predicted attenuation submap and the actual attenuation submap respectively to obtain the predicted local mean sequence and the actual local mean sequence, and the mean difference between the predicted local mean sequence and the actual local mean sequence is calculated.

[0261] Gradient magnitude calculations are performed on the predicted attenuation submap and the actual attenuation submap respectively to obtain the predicted gradient magnitude sequence and the actual gradient magnitude sequence, and the magnitude difference between the predicted gradient magnitude sequence and the actual gradient magnitude sequence is calculated.

[0262] The summation of the mean difference and amplitude difference corresponding to each window scale yields the multi-scale local attenuation consistency loss.

[0263] It should be noted that the residual graph model training device can implement all the processes implemented in the above method embodiments and achieve the same beneficial effects. To avoid repetition, it will not be described again here.

[0264] In the case where the functions of the integrated modules described above are implemented in hardware, this application provides a possible structural schematic diagram of the electronic device involved in the above embodiments. For example... Figure 9 As shown, the electronic device 90 includes: a processor 92, a communication interface 93, and a bus 94. Optionally, the electronic device 90 may also include a memory 91.

[0265] Processor 92 may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 92 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 92 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0266] Communication interface 93 is used to connect with other devices via a communication network. This communication network can be Ethernet, wireless access network, wireless local area network (WLAN), etc.

[0267] The memory 91 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0268] As one possible implementation, the memory 91 can exist independently of the processor 92. The memory 91 can be connected to the processor 92 via a bus 94 and is used to store instructions or program code. When the processor 92 calls and executes the instructions or program code stored in the memory 91, it can implement the residual graph model training method provided in the embodiments of this application.

[0269] In another possible implementation, memory 91 can also be integrated with processor 92.

[0270] Bus 94 can be an Extended Industry Standard Architecture (EISA) bus, etc. Bus 94 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0271] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the service calling device can be divided into different functional modules to complete all or part of the functions described above.

[0272] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above residual graph model training method embodiment and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0273] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0274] This application also provides a readable storage medium storing a program or instructions that, when executed by a computer, implement the residual graph model training method provided in the above embodiments. It is understood that all or part of the processes in the above method embodiments can be executed by computer instructions instructing related hardware; the readable storage medium can be any of the foregoing embodiments or memory; the readable storage medium can also be an external storage device of the service invocation device, such as a pluggable hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, flash card, etc., equipped on the service invocation device. Further, the readable storage medium can include both internal storage units of the service invocation device and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the service invocation device. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0275] This application also provides a computer program product, which is stored in a storage medium and, when executed by a computer, implements the residual graph model training method provided in the above embodiments.

[0276] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0277] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0278] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for training a residual graph model, characterized in that, include: The training samples, consisting of map information of the target area, radio transmitter information, and radio propagation characteristics, are input into the model to be trained to obtain the prediction residual map. The true residual map is obtained based on the actual path loss of radio propagation within the target area and the basic propagation prior. Based on the building occupancy map of the target area, the location of the radio transmitter, the coordinates of the radio receiver grid, the building occupancy relationship, and the visibility information from the transmitter to the receiver grid, a set of propagation area masks is constructed. The set of propagation area masks includes an effective prediction area mask, a non-line-of-sight area mask, a building boundary area mask, and a multi-scale local attenuation area mask. The data fitting loss between the predicted residual map and the true residual map is calculated within the effective prediction region mask, and the non-line-of-sight structural consistency loss between the predicted residual map and the true residual map is calculated within the non-line-of-sight region mask; Within the building boundary region mask, calculate the consistency loss of the building boundary gradient abrupt change between the predicted residual map and the real residual map, and within the multi-scale local decay region mask, calculate the consistency loss of the multi-scale local decay between the predicted residual map and the real residual map. In a sub-region of the non-line-of-sight mask, the underpredicted one-sided risk loss of the predicted residual map and the real residual map is calculated. The data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss are combined into a composite training objective. The parameters of the model to be trained are updated based on the composite training objective.

2. The residual graph model training method according to claim 1, characterized in that, The map information includes a building occupancy map and a building height map; the radio transmitter information includes the transmitter location; the radio propagation characteristics include a distance map from the transmitter to the receiving grid; the receiving grid coordinates are geometric coordinate data representing each receiving grid location; the distance map is used to reflect the propagation spatial conditions between the transmitter and receiver; the training samples composed of the map information of the target area, the radio transmitter information, and the radio propagation characteristics are input into the model to be trained to obtain a prediction residual map, and the true residual map is obtained based on the true path loss of radio propagation within the target area and the basic propagation prior, including: Spatial resolution unification and grid coordinate alignment are performed on the building occupancy map, building height map, transmitter location, receiver grid coordinates, and distance map from transmitter to receiver grid of the training samples to obtain input feature maps with unified spatial resolution. The encoder of the model to be trained performs multi-scale downsampling feature extraction processing on the input feature map to obtain a multi-scale radio propagation feature map, and the decoder of the model to be trained performs upsampling and cross-layer feature fusion processing on the multi-scale radio propagation feature map to obtain the prediction residual map. Obtain the true path loss map corresponding to the training sample. The true path loss map is generated from field measurement data or ray tracing simulation data. The basic propagation prior map is calculated according to the free space propagation loss formula based on the transmitter power, transmitter antenna gain, receiver antenna gain and carrier frequency. The true residual map is obtained by performing a grid-by-grid interpolation operation on the true path loss map and the basic propagation prior map in the same geographic grid coordinate system.

3. The residual graph model training method according to claim 2, characterized in that, The process of constructing a propagation area mask set based on the building occupancy map of the target area, the location of the radio transmitter, the coordinates of the radio receiver grid, building obstruction relationships, and visibility information from the transmitter to the receiver grid includes: The effective prediction region mask is generated based on the evaluable area and outdoor area in the training samples. The effective prediction region mask is the area excluding the interior area of ​​buildings and the unlabeled area. The visibility information of each receiving grid is determined based on the intersection of the straight path from the transmitter to each receiving grid with the building occupancy map, the obstruction length, and the superelevation. The receiving grids that are not at line of sight in the visibility information are marked as the non-line-of-sight area mask. The straight path is determined based on the transmitter position and the receiving grid coordinates. Morphological gradient operations are performed on the building occupancy map to obtain the building boundary, and a preset pixel width is extended to both sides of the building boundary to obtain the building boundary region mask; The occlusion ratio, building density, visibility degradation and local gradient magnitude of the received grid within each window scale are calculated at multiple window scales to obtain a multi-scale attenuation feature sequence. The receiving grid corresponding to any feature in the multi-scale attenuation feature sequence that exceeds the corresponding threshold is marked as a local attenuation region, and the local attenuation regions under each window scale are merged to obtain the multi-scale local attenuation region mask.

4. The residual graph model training method according to claim 1, characterized in that, The calculation of the non-line-of-sight structural consistency loss between the predicted residual map and the true residual map within the non-line-of-sight region mask includes: A non-line-of-sight predicted residual sub-image is extracted from the predicted residual image based on the non-line-of-sight region mask, and a non-line-of-sight real residual sub-image is extracted from the real residual image based on the non-line-of-sight region mask. The non-line-of-sight predicted residual sub-image and the non-line-of-sight real residual sub-image have the same grid size and spatial coordinate range. Local window normalization is performed on the non-line-of-sight predicted residual sub-map and the non-line-of-sight true residual sub-map respectively to obtain normalized predicted residual sub-map and normalized true residual sub-map, and the local structural similarity index between the normalized predicted residual sub-map and the normalized true residual sub-map is calculated. Gradient operator convolution processing is performed on the non-line-of-sight predicted residual sub-map and the non-line-of-sight true residual sub-map respectively to obtain the predicted gradient direction map and the true gradient direction map, and the gradient direction consistency metric between the predicted gradient direction map and the true gradient direction map is calculated. The non-line-of-sight structural consistency loss is obtained by weighted summation of the local structural similarity index and the gradient direction consistency metric.

5. The residual graph model training method according to claim 1, characterized in that, The calculation of the underpredicted one-sided risk loss between the predicted residual map and the true residual map within a sub-region of the non-line-of-sight mask includes: Based on the non-line-of-sight region mask and the occlusion length threshold, the grids whose occlusion length exceeds the occlusion length threshold are filtered from the non-line-of-sight region mask to obtain the sub-region of the non-line-of-sight region mask. The occlusion length threshold is set according to the carrier frequency and the average building height. Based on the sub-region of the non-line-of-sight region mask, a deep occlusion prediction residual sub-image is extracted from the prediction residual image, and a deep occlusion real residual sub-image is extracted from the real residual image based on the deep occlusion region mask. The deep occlusion prediction residual submap is compared with the deep occlusion actual residual submap to obtain a residual difference map. Grids in the residual difference map where the predicted residual is less than the actual residual and the difference exceeds the tolerance threshold are marked as under-predicted grids. The tolerance threshold is set according to the path loss measurement accuracy and coverage assessment tolerance. A one-sided penalty function is applied to the difference magnitude at the underpredicted grid to obtain the penalty value for each underpredicted grid. The penalty values ​​of each underpredicted grid are then summed to obtain the underpredicted one-sided risk loss. The one-sided penalty function applies a linear or non-linear incremental penalty to the underestimated magnitude.

6. The residual graph model training method according to claim 1, characterized in that, The step of calculating the consistency loss of the building boundary gradient abrupt change between the predicted residual map and the true residual map within the building boundary region mask includes: Based on the building boundary region mask, a boundary prediction residual sub-image is extracted from the prediction residual image, and a boundary true residual sub-image is extracted from the true residual image based on the building boundary region mask. The boundary prediction residual sub-image and the boundary true residual sub-image cover the extended zone areas on both sides of the building boundary. High-pass filtering is performed on the boundary prediction residual submap and the boundary true residual submap respectively to obtain the predicted gradient jump component map and the true gradient jump component map, and the gradient jump component difference between the predicted gradient jump component map and the true gradient jump component map is calculated. Convolution processing is performed on the boundary prediction residual submap and the boundary true residual submap respectively to obtain the prediction response map and the true response map, and the response difference between the prediction response map and the true response map is calculated. The weighted summation of the gradient jump component difference and the response difference yields the building boundary gradient abrupt change consistency loss.

7. The residual graph model training method according to claim 1, characterized in that, The step of calculating the multi-scale local attenuation consistency loss between the predicted residual map and the true residual map within the multi-scale local attenuation region mask includes: Based on the local attenuation region corresponding to any window scale in the multi-scale local attenuation region mask, a predicted attenuation sub-map at that scale is extracted from the predicted residual map, and a real attenuation sub-map at that scale is extracted from the real residual map based on the local attenuation region corresponding to that window scale. Window mean calculation is performed on the predicted attenuation submap and the true attenuation submap respectively to obtain the predicted local mean sequence and the true local mean sequence, and the mean difference between the predicted local mean sequence and the true local mean sequence is calculated. Gradient magnitude calculations are performed on the predicted attenuation submap and the actual attenuation submap respectively to obtain a predicted gradient magnitude sequence and an actual gradient magnitude sequence, and the magnitude difference between the predicted gradient magnitude sequence and the actual gradient magnitude sequence is calculated. The mean difference and amplitude difference corresponding to each window scale are summed to obtain the multi-scale local attenuation consistency loss.

8. A residual graph model training device, characterized in that, include: The input module is used to input training samples consisting of map information of the target area, radio transmitter information and radio propagation characteristics into the model to be trained, to obtain the prediction residual map, and to obtain the real residual map based on the real path loss of radio propagation in the target area and the basic propagation prior. The construction module is used to construct a propagation area mask set based on the building occupancy map of the target area, the location of the radio transmitter, the coordinates of the radio receiver grid, the building occupancy relationship, and the visibility information from the transmitter to the receiver grid. The propagation area mask set includes an effective prediction area mask, a non-line-of-sight area mask, a building boundary area mask, and a multi-scale local attenuation area mask. The first calculation module is used to calculate the data fitting loss between the predicted residual map and the real residual map within the effective prediction region mask, and to calculate the non-line-of-sight structural consistency loss between the predicted residual map and the real residual map within the non-line-of-sight region mask; The second calculation module is used to calculate the consistency loss of the building boundary gradient abrupt change between the predicted residual map and the real residual map within the building boundary region mask, and to calculate the multi-scale local decay consistency loss between the predicted residual map and the real residual map within the multi-scale local decay region mask. The combination module is used to calculate the underpredicted one-sided risk loss of the predicted residual map and the real residual map in a sub-region of the non-line-of-sight mask, and combine the data fitting loss, non-line-of-sight structural consistency loss, building boundary gradient abrupt change consistency loss, multi-scale local decay consistency loss, and underpredicted one-sided risk loss into a composite training objective, and train and update the parameters of the model to be trained based on the composite training objective.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the residual graph model training method as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a computer, implement the residual graph model training method as described in any one of claims 1-7.

11. A computer program product, characterized in that, The computer program product is stored in a storage medium, and when executed by a computer, the computer program product implements the residual graph model training method as described in any one of claims 1-7.