Radio map estimation method and model training method and device

Abstract

The present disclosure relates to the technical field of wireless communication and intelligent sensing, and discloses a radio map estimation method and a model training method and device. The technical means based on the probability triggering mode are used to determine the sampling position from each trajectory point of the actual motion trajectory and to construct a sparse observation set, so as to realize the accurate screening of the effective observation position in the real trajectory sampling scene, eliminate the redundancy introduced by the non-discriminative processing of the trajectory point signal measurement value, and guarantee the construction accuracy of the sparse observation set under the actual deployment condition. The technical means of mapping the sparse observation set and the passable mask to the discrete space unit and constructing a multi-channel model input tensor containing at least the sparse observation value, the sampling mask and the passable mask make the model input carry three types of information of the observation value distribution, the sampling coverage range and the spatial passability, effectively solve the problem of confusing the missing value of the unsampled position with the real low signal strength, and improve the sensing ability of the model to the spatial signal distribution.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and intelligent sensing technology, specifically to a radio map estimation method, model training method, and apparatus. Background Technology

[0002] With the widespread application of drones and mobile sensing devices in the field of wireless communication, radio map estimation has become a key technology in fields such as spectrum sensing and network coverage prediction. Its core task is to reconstruct the complete spatial distribution of wireless signals within a target area using limited measurement data.

[0003] In related technologies, radio map estimation methods generally assume that sampling points are independent and randomly distributed, and use this as a premise to build training data and estimation models. However, in real-world applications, mobile devices such as drones or vehicles can only collect measurement data along continuous trajectories, and the sampling points exhibit significant spatial continuity and local clustering characteristics, which is fundamentally different from the aforementioned assumption of independent random distribution. This difference leads to a discrepancy between the data distribution learned during model training and the actual data distribution during deployment. Under trajectory sampling conditions, the reconstruction error increases significantly, and the system's practicality is severely reduced.

[0004] Therefore, how to eliminate the impact of inconsistent sampling distribution during the training and deployment stages on reconstruction accuracy in real-world scenarios where mobile devices sample along continuous trajectories, thereby achieving high-precision radio map estimation, has become an urgent technical problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a radio map estimation method, a model training method, and an apparatus to solve the problem of how to eliminate the impact of inconsistent sampling distribution during the training and deployment stages on reconstruction accuracy in real-world scenarios where mobile devices sample along continuous trajectories, thereby achieving high-precision radio map estimation.

[0006] In a first aspect, this disclosure provides a radio map estimation method, including: The environmental information of the target area is obtained, the target area is divided into multiple discrete spatial units, and a passability mask is constructed based on the environmental information; the passability mask is used to identify the spatial areas within the target area that mobile devices are allowed to pass through. The signal measurement values ​​collected at each trajectory point when the mobile device moves along the actual motion trajectory in the target area are obtained. The sampling position is determined from each trajectory point based on the probability triggering method, and a sparse observation set is constructed. Map the sparse set of observations and the passable mask to discrete spatial cells to construct a model input tensor that contains at least sparse observations, a sampling mask, and a passable mask; The model input tensor is a radio map estimation model trained with sparse observation data that includes trajectory constraint characteristics, and the output is the radio map estimation result of the target area.

[0007] Secondly, this disclosure provides a method for training a radio map estimation model, including: Based on the environmental information of the target area, the target area is divided into multiple discrete spatial units, and a passable mask is constructed. Based on the passable mask, a simulated continuous trajectory that satisfies motion constraints is generated. Randomly triggered sampling is performed at each trajectory point of the simulated continuous trajectory to construct a simulated sparse observation set. The simulated sparse observation set and the passable mask are mapped to discrete spatial cells to construct a model input tensor that contains at least the simulated sparse observations, the sampling mask, and the passable mask; The model is trained by using the model input tensor and the corresponding complete radio map as training sample pairs, and the reconstruction error on the effective spatial area marked by the passable mask is used as the loss function.

[0008] Thirdly, this disclosure provides a radio map estimation apparatus for performing the radio map estimation method of the first aspect described above, including: The mask construction module is used to obtain environmental information of the target area, divide the target area into multiple discrete spatial units, and construct a passable mask based on the environmental information; the passable mask is used to identify the spatial area within the target area that mobile devices are allowed to pass through. The random sampling module is used to acquire signal measurement values ​​collected at each trajectory point when the mobile device moves along the actual motion trajectory in the target area. Based on the probability triggering method, the sampling position is determined from each trajectory point to construct a sparse observation set. The input building module is used to map sparse observation sets and passable masks to discrete spatial units, and construct a model input tensor that contains at least sparse observations, sampling masks, and passable masks; The map estimation module is used to input the model input tensor into a radio map estimation model trained with sparse observation data containing trajectory constraints, and output the radio map estimation results for the target area.

[0009] Fourthly, this disclosure provides a training apparatus for a radio map estimation model, used to execute the training method for the radio map estimation model described in the second aspect above, comprising: The mask construction module is used to divide the target area into multiple discrete spatial units based on the environmental information of the target area and construct a passable mask; The trajectory sampling module is used to generate a simulated continuous trajectory that satisfies motion constraints based on a passable mask, and to perform random-triggered sampling at each trajectory point of the simulated continuous trajectory to construct a simulated sparse observation set. The input building module is used to map the simulated sparse observation set and the passable mask to discrete spatial cells, and to build a model input tensor that contains at least the simulated sparse observations, the sampling mask, and the passable mask; The model training module is used to train the radio map estimation model by using the model input tensor and the corresponding complete radio map as training sample pairs, and using the reconstruction error on the effective spatial area marked by the passable mask as the loss function.

[0010] Fifthly, this disclosure also provides an electronic device, including: a memory and a processor, with a communication connection between the memory and the processor, the memory storing computer instructions, and the processor executing the computer instructions to perform the radio map estimation method of the first aspect described above, and / or to perform the training method of the radio map estimation model of the second aspect described above.

[0011] In a sixth aspect, this disclosure also provides a computer-readable storage medium storing computer instructions for causing a computer to implement the radio map estimation method of the first aspect described above, and / or to execute the training method of the radio map estimation model of the second aspect described above.

[0012] In a seventh aspect, this disclosure also provides a computer program product, including computer instructions for causing a computer to execute the radio map estimation method of the first aspect described above, and / or to execute the training method of the radio map estimation model of the second aspect described above.

[0013] The radio map estimation method, model training method, and apparatus of the above embodiments of this disclosure firstly employ a technique based on probability triggering to determine sampling locations from each trajectory point of the actual motion trajectory and construct a sparse observation set. This achieves accurate screening of effective observation locations in real trajectory sampling scenarios, eliminates redundancy introduced by indiscriminate processing of trajectory point signal measurements, and ensures the accuracy of sparse observation set construction under actual deployment conditions. Secondly, the technique involves mapping the sparse observation set and a traversability mask together to discrete spatial units and constructing a multi-channel model input tensor that includes at least sparse observation values, a sampling mask, and a traversability mask. This allows the model input to simultaneously carry three types of information: observation distribution, sampling coverage, and spatial traversability. This effectively solves the problem of confusion between missing values ​​at unsampled locations and real low signal strength, and improves the model's ability to perceive spatial signal distribution.

[0014] Finally, by inputting the above model into a tensor and training it with sparse observation data containing trajectory constraint characteristics, the inconsistency between the sampling distribution assumptions during the training phase and the trajectory sampling characteristics during the actual deployment phase is eliminated, thereby achieving high-precision estimation of the radio map of the target area and effectively reducing the reconstruction error of mobile devices under real trajectory sampling conditions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of this disclosure, the accompanying drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic flowchart of a radio map estimation method provided in an embodiment of this disclosure is shown.

[0017] Figure 2 This diagram illustrates a qualitative comparison of the radio map reconstruction performance under different combinations of sampling strategies and training modes of a radio map estimation method provided in this embodiment of the present disclosure.

[0018] Figure 3 A comparative schematic diagram of three mechanisms—random sampling, fixed-interval sampling, and random-triggered sampling—is shown in an embodiment of the radio map estimation method provided in this disclosure.

[0019] Figure 4 A flowchart illustrating a radio map estimation model training method provided in an embodiment of this disclosure is shown.

[0020] Figure 5 This diagram illustrates a quantitative comparison of the radio map reconstruction performance of a radio map estimation method provided in this disclosure under different training-test sampling distributions.

[0021] Figure 6 This illustration shows a quantitative comparison of the root mean square error of various radio map estimation models of a radio map estimation method provided in this disclosure under different training-test sampling distribution configurations.

[0022] Figure 7 An exemplary schematic diagram of the architecture of a radio map estimation apparatus according to an embodiment of the present disclosure is shown.

[0023] Figure 8 An exemplary schematic diagram of the architecture of a training apparatus for a radio map estimation model according to an embodiment of the present disclosure is shown.

[0024] Figure 9 A schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0025] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this disclosure.

[0026] With the widespread application of drones and mobile sensing devices in wireless communication, constructing high-precision radio maps using a small number of sampling points has become a critical issue. Most radio map estimation methods in related technologies assume that the sampling points are independent and randomly distributed. While this assumption offers some convenience in theoretical analysis and models, it differs significantly from real-world application scenarios.

[0027] In real-world environments, sampling data is typically collected by drones or vehicles along continuous trajectories. The sampling points exhibit clear spatial continuity and correlation, rather than being independently distributed. This discrepancy leads to a mismatch between the data distribution learned by the model during training and the actual data distribution during deployment, resulting in significant performance degradation. Specifically, models trained under random sampling conditions exhibit significantly increased reconstruction errors when tested under trajectory sampling conditions, severely impacting the system's usability.

[0028] Further analysis revealed two key issues with trajectory sampling. First, because sampling points are continuously distributed along the path, the information between adjacent sampling points is highly similar, leading to a decrease in effective information. Second, because sampling is concentrated in local areas, the overall spatial coverage is insufficient, making it difficult for the model to recover the signal distribution in unobserved areas.

[0029] Therefore, radio map estimation methods in related technologies struggle to maintain stable performance under real sampling conditions.

[0030] To address the aforementioned problems, various embodiments of this disclosure provide a radio map estimation method, comprising: acquiring environmental information of a target area; dividing the target area into multiple discrete spatial units; and constructing a passability mask based on the environmental information; wherein the passability mask is used to identify spatial areas within the target area that mobile devices are allowed to pass through; acquiring signal measurement values ​​collected at each trajectory point when the mobile device travels along its actual motion trajectory within the target area; determining sampling positions from each trajectory point based on a probability-triggered method; constructing a sparse observation set; mapping the sparse observation set and the passability mask to discrete spatial units; constructing a model input tensor that at least includes sparse observation values, a sampling mask, and a passability mask; inputting the model input tensor into a radio map estimation model trained with sparse observation data containing trajectory constraint characteristics; and outputting the radio map estimation result of the target area.

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0032] Please refer to Figure 1 , Figure 1 A flowchart illustrating a radio map estimation method provided in an embodiment of this disclosure is shown, as follows: Figure 1 As shown, the process of this method may include the following steps: Step S101: Obtain environmental information of the target area, divide the target area into multiple discrete spatial units, and construct a passable mask based on the environmental information.

[0033] In this embodiment, the target area can refer to the spatial range in which the wireless signal distribution is to be estimated, and its spatial boundaries and estimation accuracy requirements can be determined according to the actual application scenario.

[0034] Environmental information can refer to information that reflects the spatial structure and obstacle distribution within a target area, and is used to identify the boundaries between areas within the target area where mobile devices are allowed to pass and areas where they are not.

[0035] For example, environmental information may include, but is not limited to, building floor plans, 3D city models, electronic maps, Geographic Information System (GIS) data, Building Information Modeling (BIM) models, LiDAR point clouds, or raster maps generated by sensors mounted on drones / vehicles.

[0036] Furthermore, after acquiring the environmental information of the target area, the target area can be uniformly divided according to a preset spatial resolution to obtain multiple discrete spatial units. Each discrete spatial unit together constitutes a spatial reference grid covering the target area.

[0037] For example, in a two-dimensional scene, the target area It can be divided into A discrete spatial unit, here, the target region Can be set according to preset spatial resolution The target area can be further divided into subdivisions in a 3D scene. Discrete spatial units.

[0038] in, and This can correspond to the number of grids in two horizontal spatial directions within the target area; This corresponds to the number of grid cells in the vertical direction of the target area.

[0039] Furthermore, the center spatial coordinates of each discrete spatial unit can be denoted as follows in a two-dimensional scene: In a 3D scene, it is denoted as ,in, It can represent the spatial index of discrete spatial units.

[0040] Furthermore, a passability mask is used to identify the spatial regions within a target area that mobile devices are permitted to traverse. Here, a passability mask can be a binary matrix that identifies the spatial regions within a target area that mobile devices are permitted to traverse, using discrete spatial units as the granularity, and can be denoted as... .

[0041] Among them, for discrete spatial units that allow mobile devices to pass through Its passability mask value is For discrete spatial units located within buildings, restricted areas, or other obstacle zones, the mask value is... .

[0042] As an example, when constructing a passability mask based on environmental information, the following methods can be used depending on the specific format of the environmental information: When the environmental information is a vector map or building outline, the polygon outline can be projected onto the target coordinate system first, and then the rasterization judgment can be performed according to whether the center point of each discrete spatial unit is located inside the obstacle polygon to determine the passable mask value of each discrete spatial unit. When the environmental information is a lidar point cloud or depth map, obstacle distribution can be generated based on the point cloud height, voxel occupancy status or occupancy probability, and the discrete spatial units corresponding to the occupied areas can be marked as impassable. When environmental information comes from historical flight logs, inaccessible locations, collision risk locations, or restricted areas in historical flight records can be directly written into the passability mask, so that the passability mask reflects the passability constraint experience accumulated by the mobile device in historical missions.

[0043] In a 3D scene, when constructing a passability mask, the following constraints can be further superimposed on the 2D passability: When a discrete spatial unit is located at a height lower than the building height or exceeds the permitted flight altitude range of a mobile device, the discrete spatial unit is marked as impassable. In addition, constraints such as terrain height, maximum climb or descent angle, safety interval requirements, and no-fly zones can be incorporated according to specific application requirements to make it suitable for scenarios such as UAV layered height sampling, indoor multi-story radio map estimation, and urban three-dimensional spectrum perception.

[0044] By determining whether the position of each discrete spatial unit falls into an impassable region, a passability mask is constructed, so that the spatial passability distribution within the target region is explicitly expressed in matrix form.

[0045] Step S102: Obtain signal measurement values ​​collected at each trajectory point when the mobile device moves along the actual motion trajectory in the target area, determine the sampling position from each trajectory point based on the probability triggering method, and construct a sparse observation set.

[0046] In this embodiment, the actual motion trajectory can refer to the continuous spatial path traversed by the mobile device in the target area when performing the actual signal acquisition task, and the trajectory point can refer to each spatial position where the mobile device can perform signal acquisition during the process of moving along the actual motion trajectory.

[0047] Each trajectory point can be indexed by time sequence. Mark, denoted as Its spatial coordinates are aligned with the center coordinates of the discrete spatial unit to which it belongs, corresponding to the following in a two-dimensional scene: In a three-dimensional scene, the corresponding Signal measurements can refer to the measurement data obtained by the wireless signal acquisition hardware on the mobile device at various trajectory points, reflecting the wireless signal strength at that location, denoted as... .

[0048] Wireless signal acquisition hardware on mobile devices may include, but is not limited to, software-defined radio receivers, spectrum analyzers, wireless signal scanning modules, cellular network or LoRa or Wi-Fi measurement terminals, and received power detection modules. Signal measurement values. This may include, but is not limited to, Received Signal Strength Indicator (RSSI), Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), or Signal to Interference plus Noise Ratio (SINR).

[0049] When performing signal acquisition tasks, mobile devices simultaneously record the precise spatial coordinates of each sampling location through the Global Navigation Satellite System (GNSS), Real-Time Kinematic (RTK), or Inertial Measurement Unit (IMU) positioning module, and then align the position coordinates with the signal measurement values ​​to form a sparse observation set.

[0050] Furthermore, the probability-triggered method can refer to the process where, during the movement of a mobile device along its actual trajectory, a preset trigger probability is applied. A mechanism for making random sampling decisions for each trajectory point.

[0051] For each trajectory point on the actual motion trajectory, based on the preset trigger probability Random sampling decisions are made, and trajectory points that trigger sampling are determined as sampling locations, and these locations and their corresponding signal measurements are included in the sparse observation set; trajectory points that do not trigger sampling are only recorded as passing locations, and their signal measurements are not recorded.

[0052] A sparse observation set can refer to the set of observation data consisting of each sampling location and its corresponding signal measurement value, denoted as . ,in It is the set of all sampling locations, reflecting the sparse and non-uniformly distributed measured signal information of the mobile device within the target area.

[0053] It is understandable that, since mobile devices travel along continuous trajectories, there is a high degree of spatial correlation between signal measurements between adjacent trajectory points. By selecting some trajectory points as sampling locations through probability triggering, the spatial distribution of sampling locations can be made somewhat random while maintaining the continuity of the trajectory, effectively reducing the redundancy between adjacent observations.

[0054] Step S103: Map the sparse observation set and the passability mask to discrete spatial cells to construct a model input tensor that contains at least sparse observations, a sampling mask, and a passability mask.

[0055] In this embodiment, each discrete spatial cell is used as the basic unit. The signal measurement values ​​and passable masks at each sampling position in the sparse observation set are aligned to the corresponding spatial cells to form a multi-channel rasterized representation with a unified spatial reference.

[0056] sparse observation matrix It can refer to the signal intensity matrix obtained by mapping a sparse set of observations to discrete spatial cells. The sampled discrete spatial cells are filled with the corresponding signal measurement values, and the unsampled discrete spatial cells are filled with default values.

[0057] Sampling mask It can refer to a binary matrix that identifies whether each discrete spatial unit has an actual sampling record. Sampled discrete spatial units are marked as 1, and unsampled discrete spatial units are marked as 0.

[0058] The above sparse observation matrix Sampling mask The passable mask constructed in step S101 above Concatenate along the channel dimension to form the model input tensor. .

[0059] In a two-dimensional scene, the size of the model input tensor is... In a 3D scene, the size of the model input tensor is .

[0060] By including three types of information in the model input simultaneously—sparse observations, sampling masks, and passable masks—the model can clearly distinguish between sampled and unsampled areas, and between passable and impassable areas, effectively avoiding misjudging the default values ​​of unsampled locations as real low signal strength areas.

[0061] Step S104: Input the model input tensor into the radio map estimation model trained with sparse observation data containing trajectory constraint characteristics, and output the radio map estimation result of the target area.

[0062] In this embodiment, the radio map estimation model can refer to a machine learning model that takes sparse observation data as input and the complete spatial distribution of radio signals in the target area as the output target, denoted as... ,in These are model parameters, and the model structure is not limited to a specific network architecture.

[0063] Furthermore, sparse observation data containing trajectory constraint characteristics can refer to training data obtained by sampling simulated continuous trajectories generated using simulated actual motion characteristics of mobile devices during the training phase. The spatial distribution of sampling points of this type of data has continuity and local clustering characteristics consistent with the actual trajectory sampling of mobile devices.

[0064] Radio map estimation models in related technologies typically use independently and randomly distributed sampling points as training data. However, when real mobile devices collect data, they can only travel along continuous trajectories, resulting in a significant difference in data distribution between the training phase and the actual deployment phase, which in turn leads to a substantial decrease in the model reconstruction accuracy.

[0065] Please refer to Figure 2 , Figure 2 This diagram illustrates a qualitative comparison of the radio map reconstruction performance under different combinations of sampling strategies and training modes of a radio map estimation method provided in this disclosure. Figure 2 As shown, Figure 2 The top half of the leftmost column corresponds to random sampling input conditions, and the bottom half corresponds to trajectory-based sampling (TBS) input conditions. Under each input condition, the reconstruction results of four typical radio map estimation models, namely UNet, PMNet, RadioUNet, and RadioFormer, are given in two training modes: training with random sampled data (Random-Trained, blue box) and training with trajectory-constrained sampled data (TBS-Trained, green box). The rightmost column is the ground truth (GT).

[0066] Depend on Figure 2 It can be seen that when the input is random sampling and the training and deployment distributions are consistent (Random-Trained input), each model can output reconstruction results that are relatively close to the real map. When the input is switched to trajectory-constrained sampling, but the model is still trained with random sampling data, the reconstruction results of each model show significant distortion, and the spatial distribution characteristics of the wireless signal deviate severely from the real map. This intuitively proves that the inconsistency between the independent random sampling assumption in the training phase and the trajectory-constrained sampling characteristics in the actual deployment phase leads to a significant decrease in reconstruction performance.

[0067] When the model was trained with trajectory-constrained sampling data (TBS-Trained), the reconstruction accuracy of all four models was significantly restored under the same trajectory-constrained sampling input conditions. The estimation results were highly consistent with the real radio map, which verified the effectiveness of this scheme in training the radio map estimation model by introducing sparse observation data with trajectory constraint characteristics during the training phase and eliminating the inconsistency in data distribution between the training and deployment phases.

[0068] It is understandable that by using sparse observation data with trajectory constraint characteristics to train the model, the data distribution during the training phase can be kept consistent with that during the actual deployment phase, thus eliminating the impact of the aforementioned distribution offset on reconstruction performance.

[0069] In a typical embodiment, the radio map estimation model This can be implemented using an encoder-decoder convolutional neural network structure. The encoder can include multiple layers of convolution, batch normalization, and nonlinear activation layers to extract sparse observation distribution and spatial structure features from the model input tensor. Between each encoding layer, stride convolution or pooling layers can be used to gradually reduce the feature map resolution and expand the receptive field. The number of convolution channels can be set to 32, 64, 128, and 256 respectively, and the convolution kernel size can be set to 3×3.

[0070] The decoder can include upsampling layers or deconvolutional layers and establish skip connections with feature maps of the corresponding scale in the encoder to fuse shallow spatial detail information while restoring resolution; finally, it outputs the radio map estimation result of all discrete spatial units of the target region through 1×1 convolution. .

[0071] In 3D scenes, the 2D convolutions in the above structure can be replaced with 3D convolutions, or a lightweight structure combining horizontal 2D convolutions with height dimension fusion can be adopted. Furthermore, the radio map estimation model can also be replaced with other image completion or spatial interpolation models such as U-Net, RadioUNet, Transformer-based encoder structures, or diffusion reconstruction networks, as long as the model input includes sparse observations generated by trajectory constraint sampling, a sampling mask, and a passable mask; the training method disclosed herein can be applied to all of these.

[0072] Furthermore, the radio map estimation result can refer to the complete radio signal strength estimation matrix output by the radio map estimation model for all discrete spatial cells within the target area, denoted as... This reflects the spatial distribution of wireless signals within the target area.

[0073] The radio map estimation method, model training method, and apparatus of the above embodiments of this disclosure firstly employ a technique based on probability triggering to determine sampling locations from each trajectory point of the actual motion trajectory and construct a sparse observation set. This achieves accurate screening of effective observation locations in real trajectory sampling scenarios, eliminates redundancy introduced by indiscriminate processing of trajectory point signal measurements, and ensures the accuracy of sparse observation set construction under actual deployment conditions. Secondly, a technique is employed to jointly map the sparse observation set and a traversability mask to discrete spatial units, constructing a multi-channel model input tensor that includes at least sparse observation values, a sampling mask, and a traversability mask. This allows the model input to simultaneously carry three types of information: observation distribution, sampling coverage, and spatial traversability. This effectively solves the problem of confusion between missing values ​​at unsampled locations and real low signal strength, improving the model's ability to perceive spatial signal distribution. Finally, by inputting the above model input tensor into a radio map estimation model trained with sparse observation data containing trajectory constraint characteristics, the inconsistency between the sampling distribution assumptions during the training phase and the trajectory sampling characteristics during the actual deployment phase is eliminated. This achieves high-precision estimation of the radio map of the target area and effectively reduces the reconstruction error of mobile devices under real trajectory sampling conditions.

[0074] In one possible implementation of step S102 above, determining the sampling position from each trajectory point based on a probability-triggered method includes: For each trajectory point in the actual motion trajectory, random sampling decisions are made based on preset trigger probabilities; When sampling is triggered and the distance between the trajectory point and each recorded sampling position in the sparse observation set is not less than the preset minimum interval, the trajectory point is determined as the sampling position. The number of recorded sampling locations within a local spatial window centered on a trajectory point is counted. When the number of sampling locations exceeds a preset upper limit, the trigger probability of subsequent trajectory points within the local spatial window is reduced.

[0075] In this embodiment, when performing sampling decisions on each trajectory point in the actual motion trajectory, for the current trajectory point... Generate random variables that follow a uniform distribution on [0,1]. , random variable With preset trigger probability Comparisons include: when When the sampling decision is triggered, the process proceeds to the subsequent interval constraint judgment. when The current trajectory point is determined to be only a passing position and no sampling is performed.

[0076] Among them, the preset trigger probability It can be used to control sampling density. The larger the value, the higher the expected number of sampling locations triggered per unit trajectory length; The smaller the value, the higher the sampling sparsity.

[0077] Furthermore, after triggering sampling, the current trajectory point is calculated. The spatial distance between the sparse observation set and the recorded sampling locations is only when The distance between each recorded sampling location and the location is not less than the preset minimum interval. At that time, The sampling location was determined and included in the sparse observation set; if The distance to any recorded sampling location is less than If the trigger is not triggered, the current trajectory point will not be recorded as a sampling location.

[0078] Here, by introducing a preset minimum interval constraint, the sampling positions can be effectively avoided from being overly concentrated in local segments of the trajectory, thereby ensuring that the sampling positions in the sparse observation set have a certain degree of spatial dispersion.

[0079] Furthermore, after completing the interval constraint judgment, the current trajectory point... Define a local spatial window centered on the observation point and count the number of recorded sampling locations in the sparse observation set within that window.

[0080] When the number of statistics exceeds the preset upper limit, the trigger probability of subsequent trajectory points within the local spatial window is reduced to suppress further aggregation of sampling locations in the local spatial region.

[0081] Among them, the aforementioned density adaptive control mechanism can actively limit the sampling density in a local area when the trajectory of the mobile device meanders back and forth, so that the overall sampling distribution in the target area remains relatively balanced.

[0082] As an example, please refer to Figure 3 , Figure 3 The diagram illustrates a comparison of three mechanisms—random sampling, fixed-interval sampling, and random-triggered sampling—in a radio map estimation method provided by embodiments of this disclosure. Figure 3 As shown, the gray area represents impassable areas such as buildings, and the gray curve represents the movement trajectory of the mobile device.

[0083] like Figure 3 As shown in the left figure, under the traditional random sampling mechanism, the sampling points represented by the green dots are independently and randomly distributed within the target area, without any trajectory constraints. Although they have a certain degree of spatial dispersion, this does not match the actual scenario where real mobile devices can only collect data along a continuous trajectory, resulting in a significant difference between the data distribution during the training phase and the actual deployment phase.

[0084] like Figure 3 As shown in the figure, under the fixed interval sampling mechanism, the sampling points represented by the blue dots are evenly distributed along the continuous trajectory at fixed intervals d, satisfying the trajectory continuity constraint. However, the spacing between the sampling points is completely regular, forming a periodic sampling pattern, which is still different from the actual sampling pattern formed by real mobile devices due to factors such as power consumption limitations, communication frequency fluctuations, and irregular measurement times.

[0085] like Figure 3 As shown in the right figure, under the random trigger sampling mechanism proposed in this disclosure, the sampling points represented by the orange dots are determined according to the probability triggering method along the continuous trajectory. The distances d1 and d2 between adjacent sampling positions are non-uniformly distributed due to random triggering. While maintaining the continuity of the trajectory, randomness is introduced, breaking the periodicity of fixed interval sampling and more closely resembling the actual sampling pattern formed by the randomness of triggering conditions when a real mobile device performs a collection task.

[0086] For example, a preset trigger probability Values ​​between 0.05 and 0.5 can be selected; preferably, typical values ​​such as 0.1 / 0.2 / 0.3 can be chosen to achieve a reasonable balance between sampling cost and coverage density; a minimum interval is preset. The number of sampling points can be determined based on the area of ​​the target region and the desired number of sampling points to ensure sufficient spatial spacing between adjacent sampling locations.

[0087] The radio map estimation method, model training method, and apparatus of the above embodiments of this disclosure, by adjusting each trajectory point according to a preset trigger probability, By implementing random sampling decisions and combining preset minimum interval constraints with local spatial window density adaptive control mechanisms, the problem of excessive concentration of sampling positions in local areas when mobile devices travel along continuous trajectories is effectively suppressed while retaining the randomness of probability triggering. This makes the spatial distribution of the final sparse observation set more balanced within the target area, thereby improving the spatial coverage diversity of effective observation positions.

[0088] In one possible implementation of step S103 above, when mapping the sparse observation set and the passable mask to discrete spatial cells, the method further includes: For a discrete spatial unit with multiple signal measurements, the multiple signal measurements are fused according to a weighted fusion method to obtain the observation value of the discrete spatial unit; The observation values ​​of each discrete spatial unit are linearly normalized according to the preset upper and lower bounds of the signal strength and mapped to the preset normalization interval. Normalized coordinate information of each discrete spatial unit is added to the channel dimension of the model input tensor.

[0089] In this embodiment, when the mobile device repeatedly passes through the same area along its actual movement trajectory, or when multiple adjacent trajectory points belong to the same discrete spatial unit, multiple signal measurement values ​​will accumulate within that spatial unit. For a discrete spatial unit with multiple signal measurements, a weighted fusion method is used to fuse the multiple signal measurements to obtain the representative observation value of the spatial unit.

[0090] Weighted fusion methods may include, but are not limited to, distance-weighted average and time-decay-weighted average.

[0091] Preferably, in the distance-weighted averaging method, the fusion weight of each measurement value is inversely proportional to the spatial distance between the trajectory point corresponding to the measurement value and the center of the grid to which it belongs, and the measurement value closer to the grid center is given a higher weight; in the time-attenuation-weighted averaging method, the measurement value more recently acquired is given a higher weight, so as to reflect the characteristics of the signal environment changing over time.

[0092] After normalizing the weights of each measurement value, a weighted summation is performed to obtain the fused observation value of the discrete spatial unit.

[0093] Furthermore, after the observation fusion is completed, linear normalization processing can be performed on the observations of each discrete spatial unit, including: using a preset lower bound for signal strength. With the upper realm As a normalization reference range, the observations of each spatial unit are mapped to the preset normalization interval [0,1].

[0094] Here, it can be understood that by using a unified linear normalization process, the influence of differences in signal strength dimensions between different application scenarios, frequency bands, or acquisition devices on the model input is eliminated, thus making the sparse observation matrix... The numerical range remains consistent.

[0095] Furthermore, after constructing the model input tensor, normalized coordinate information of each discrete spatial unit is added to its channel dimension.

[0096] Specifically, in a two-dimensional scene, add normalization. coordinate channel and normalization Coordinate Channel In each channel, the first The values ​​at each discrete spatial unit are respectively and ,in and These represent the number of grid cells in the horizontal and vertical directions of the target area, respectively.

[0097] In a 3D scene, further normalization can be added. Coordinate Channel After adding coordinate channels, the model input tensor is expanded to... This enables the model to perceive the absolute position of each spatial unit within the target area when processing each spatial unit, effectively distinguishing regions that are spatially close but have different signal distribution patterns.

[0098] In addition, auxiliary feature channels such as distance to base station channels or building height channels can be added according to actual needs to provide the model with richer spatial prior information.

[0099] The radio map estimation method, model training method, and apparatus of the above embodiments of this disclosure firstly reduce the impact of single measurement noise on the representativeness of observations by performing weighted fusion on multiple signal measurements within the same discrete spatial unit, enabling the sparse observation set to maintain the accuracy of observations even when mobile devices repeatedly pass through the same area. Secondly, by performing linear normalization on the observations of each spatial unit, the interference of signal strength dimension differences between different deployment scenarios and acquisition devices on the consistency of model input is eliminated, improving the generalization stability of the model across different scenarios. Then, by adding a normalized coordinate channel to the model input tensor, the model can utilize spatial location information during inference, improving the estimation accuracy for wireless signal scenarios with significant spatial distribution patterns.

[0100] In one possible implementation of step S102 above, constructing the sparse observation set further includes: Acquire historical signal measurement values ​​from multiple batches of data collected during multiple flights, multiple mobile devices, or multiple time periods within the target area, and map the historical signal measurement values ​​to discrete spatial units according to their spatial location; For multiple historical signal measurements within the same discrete spatial unit, a time attenuation weight is set based on the acquisition time of each historical signal measurement, and weighted fusion is performed. The historical signal measurements corresponding to each batch of data acquisition are mapped to independent observation channels, and the historical sampling count and the time interval from the last sampling are added to the channel dimension of the model input tensor.

[0101] In this embodiment, in a real deployment scenario, the system may accumulate historical signal measurement values ​​formed by multiple flight missions, multiple mobile devices, or multiple time periods.

[0102] Preferably, the aforementioned historical signal measurements can be mapped to the corresponding discrete spatial units according to the spatial coordinates of each sampling location, thereby completing the spatial alignment of historical observation data.

[0103] Specifically, for multiple historical signal measurements accumulated within the same discrete spatial unit, a time attenuation weight can be set based on the acquisition time of each historical signal measurement, including: Historical signal measurements acquired more recently than the current time are assigned a higher fusion weight, while historical signal measurements acquired earlier have a lower weight.

[0104] By using time-attenuation weighted fusion, the fused observations of this spatial unit can better reflect the recent signal environment status, reducing the impact of inaccurate early historical data caused by changes in the signal environment over time.

[0105] Furthermore, after the historical signal measurement values ​​within each discrete spatial unit are fused, the historical signal measurement values ​​corresponding to each acquisition batch are mapped to independent observation channels and incorporated into the model input tensor, enabling the model to distinguish the coverage distribution differences of different acquisition batches within the target area.

[0106] In addition, two types of auxiliary information channels are added to the channel dimension of the model input tensor, including: One is the historical sampling count channel, which records the cumulative number of times each discrete spatial unit has been sampled in all historical acquisition batches, and can be denoted as... ; The second is the time interval channel from the last sampling, which records the time interval experienced by each discrete spatial unit from the most recent historical sampling to the current moment, and can be denoted as... .

[0107] Through the two types of auxiliary channels mentioned above, the model can distinguish between areas with stable observation records in multiple historical collections and areas that have not been updated for a long time, and adopt differentiated estimation strategies for areas with different historical sampling densities and timeliness during inference.

[0108] For example, if the system has accumulated If there are several historical data collection batches, the model input tensor can be expanded to... ,in For the first Each historical data collection batch corresponds to an observation channel. and These are the historical sampling count matrix and the time interval matrix, respectively.

[0109] The radio map estimation method, model training method, and apparatus of the above embodiments of this disclosure firstly, by mapping historical signal measurements from multiple historical acquisition batches to discrete spatial units according to their spatial locations and performing time-attenuation weighted fusion, the fused observations prioritize reflecting the recent signal environment status, effectively reducing the impact of inaccurate early historical data due to changes in the signal environment on the estimation results. Secondly, by mapping each acquisition batch to an independent observation channel and adding auxiliary channels for historical sampling counts and time intervals, the model can perceive the degree of historical observation accumulation and data timeliness of each spatial unit, gradually improving the estimation accuracy of the radio map of the target area as historical acquisition batches continue to accumulate.

[0110] refer to Figure 4 , Figure 4 This illustration shows a flowchart of a radio map estimation model training method provided in an embodiment of the present disclosure, as follows: Figure 4 As shown, the process of this method may include the following steps: Step S401: Divide the target area into multiple discrete spatial units based on the environmental information of the target area, and construct a passable mask.

[0111] In this embodiment, step S401 is executed in the same way as step S101, dividing the target area into two-dimensional scenes according to a preset spatial resolution. A discrete spatial unit, which is divided into three-dimensional scenes as follows: Each discrete spatial unit is used to construct a passable mask based on environmental information. This indicates the accessibility of each discrete spatial unit, which will not be elaborated on here.

[0112] In the training method, a passable mask can be used. It can also be used to constrain the generation range of simulated continuous trajectories, thereby ensuring that the simulated trajectory extends only within the spatial area where mobile devices are allowed to pass.

[0113] Step S402: Generate a simulated continuous trajectory that satisfies motion constraints based on a passable mask, and perform random triggered sampling at each trajectory point of the simulated continuous trajectory to construct a simulated sparse observation set.

[0114] In this embodiment, the simulated continuous trajectory can refer to a continuous spatial path generated by simulating the actual motion characteristics of a mobile device in an algorithmic manner based on the feasible spatial range defined by the passable mask. Its spatial distribution characteristics are consistent with the trajectory formed by the real mobile device when performing a task in the target area, and it has path continuity and local spatial correlation.

[0115] After the simulated continuous trajectory is generated, the same probability triggering method as in step S102 is used to randomly trigger sampling of each trajectory point of the simulated continuous trajectory. The triggered sampling trajectory points and their corresponding signal measurements are then included in the simulated sparse observation set, denoted as... .

[0116] Among them, the signal measurement values ​​in the simulation phase From a pre-prepared complete radio map sample Read from, i.e. .

[0117] Here, by introducing simulated continuous trajectories that satisfy the motion constraints of real mobile devices for sampling during the training phase, the spatial distribution characteristics of the sampling points of the simulated sparse observation set are made consistent with the sparse observation set in the actual deployment phase, thereby eliminating the distribution offset between the independent random sampling assumption and the sampling characteristics of the real trajectory.

[0118] Step S403: Map the simulated sparse observation set and the passable mask to discrete spatial cells to construct a model input tensor that contains at least the simulated sparse observations, the sampling mask, and the passable mask.

[0119] In this embodiment, step S403 can be executed in the same way as step S103, mapping the signal measurement values ​​at each sampling location in the simulated sparse observation set to the corresponding discrete spatial units to form a simulated sparse observation matrix. Simultaneous generation of sampling masks This identifies whether simulation sampling records exist in each discrete spatial unit; the simulation sparse observation matrix, sampling mask, and passability mask are concatenated along the channel dimension to form the model input tensor. .

[0120] Step S404: The model input tensor and the corresponding complete radio map are used to form training sample pairs, and the reconstruction error on the effective spatial area marked by the passable mask is used as the loss function to train the radio map estimation model.

[0121] In this embodiment, a complete radio map can refer to the real radio signal strength distribution matrix corresponding to all discrete spatial units within the target area, denoted as... It is used as a supervisory label during the training phase.

[0122] Input tensors of the model With corresponding complete radio map Constructing training sample pairs ,Will Input radio map estimation model The estimation results were obtained. With a passable mask The reconstruction error over the effective spatial region of the identifier is used as the loss function:

[0123] The summation range covers all discrete spatial units within the target region. It can be used to exclude obstacle regions or invalid regions that do not need to be estimated. It can also be further combined with gradient smoothing loss, structural similarity loss or edge preservation loss to improve map continuity.

[0124] In addition, gradient smoothing loss, structural similarity loss, or edge preservation loss can be further superimposed on the above mean square error loss to improve the spatial continuity and structural integrity of the radio map estimation results.

[0125] Preferably, for each complete radio map sample in the training set, multiple lines with different starting points and different directions can be generated to maintain their positions. and different trigger probabilities The simulation continuous trajectory is obtained, and the corresponding model input tensors are constructed respectively, so that the same complete radio map sample corresponds to multiple different sparse observation inputs in a single training iteration.

[0126] By employing the aforementioned multi-trajectory enhancement training strategy, the model is exposed to diverse trajectory sampling patterns during the training phase, including short trajectory sampling inputs with low coverage and long trajectory sampling inputs with high coverage, as well as the local clustering distribution of smooth trajectories and zigzag trajectories in different regions, thereby improving the model's generalization and adaptability to different task trajectory patterns.

[0127] Furthermore, regarding the model parameters Iterative updates complete the training.

[0128] The radio map estimation model training method of the above embodiments of this disclosure introduces simulated continuous trajectories generated based on traversable mask constraints during the training phase and constructs a simulated sparse observation set using the same probability-triggered sampling mechanism as the actual deployment phase. This ensures that the spatial distribution characteristics of the sampling points of the training data are consistent with the actual mobile device trajectory sampling, eliminating the distribution offset caused by the independent random sampling assumption in the training phase in existing methods during actual deployment. This allows the trained radio map estimation model to maintain stable reconstruction accuracy under real trajectory sampling conditions.

[0129] In one possible implementation of step S402 above, generating a simulated continuous trajectory that satisfies motion constraints based on a passable mask includes: Select the starting point of the simulation trajectory from the valid spatial area marked by the passable mask; The next direction of motion is selected from candidate directions whose angle with the current direction of motion does not exceed a preset angle threshold with a preset direction-maintaining probability; the next direction of motion is randomly selected from candidate directions that satisfy the passability constraint with a complementary probability. When all candidate directions that meet the passability constraints are unreachable, backtrack to the previous feasible position and re-execute the direction selection; when no feasible direction can be found after continuous backtracking, re-select the starting point of the simulation trajectory. Repeatedly select the direction of execution to gradually build a simulated continuous trajectory until the number of trajectory points of the simulated continuous trajectory reaches the preset trajectory length.

[0130] In this embodiment, from the passable mask The starting point of the simulation trajectory can be selected from the valid spatial area of ​​the identifier. The starting point can be from... It can be randomly selected from discrete spatial units, or a specific location can be specified according to the actual task requirements.

[0131] After determining the starting point of the simulation trajectory, candidate neighborhoods for the current position are determined based on the passability mask. In a 2D scene, the candidate neighborhood can be set to a 4-neighborhood or an 8-neighborhood centered on the current trajectory point, corresponding to movement in four directions (up, down, left, right) and eight directions (including diagonal movement), respectively. In a 3D scene, the candidate neighborhood can be expanded to a 26-neighborhood including horizontal, vertical, and diagonal movement to support multi-directional motion simulation of mobile devices in 3D space. All candidate directions must simultaneously satisfy the target position. Constraints include not moving more than the preset step length and not crossing impassable areas.

[0132] After selecting the starting point, a random walk mechanism with direction-preserving properties is used to gradually generate trajectory points. The direction-preserving probability is set to a preset value. Select the next motion direction from candidate directions whose angle with the current motion direction does not exceed a preset angle threshold; with complementary probability. The next direction of motion is randomly selected from the candidate directions that satisfy the passability constraint.

[0133] Here, the above mechanism can simulate the inertial characteristics of mobile device motion, i.e., the probability of maintaining orientation. The higher the value, the smoother the generated simulation trajectory, the lower the path bending frequency, and the closer it is to the smooth motion characteristics of actual drones or vehicles.

[0134] For example, Values ​​between 0.6 and 0.9 can be used to generate simulated trajectories that combine directional consistency with a moderate degree of randomness.

[0135] When all candidate directions that satisfy the passability constraints are unreachable, backtrack to the previous feasible trajectory point and re-execute direction selection; when no feasible direction can be found after continuous backtracking, abandon the current trajectory and reselect the starting point of the simulation trajectory to generate a new trajectory, so as to ensure that the simulation continuous trajectory can still be generated completely in the scenario of complex obstacle distribution.

[0136] Repeat the above direction selection process until the number of trajectory points in the simulated continuous trajectory reaches the preset trajectory length. .

[0137] Preset trajectory length Based on the total number of passable grid cells within the target area Compared with expected coverage Sure.

[0138] For example, Desirable ,in A value between 0.05 and 0.30 can be used to strike a balance between trajectory coverage and simulation computation cost.

[0139] The radio map estimation method, model training method, and apparatus of the above embodiments of this disclosure generate simulated continuous trajectories by employing a random walk mechanism with direction-preserving characteristics, and introduce direction-preserving probability. This approach ensures that the simulated trajectory exhibits path smoothness and spatial continuity consistent with the motion of real mobile devices, while retaining the irregularity of random direction selection. This allows the spatial distribution characteristics of the simulated trajectory to effectively approximate the actual motion trajectory of mobile devices in real deployment scenarios. Consequently, the simulated sparse observation set generated in this way maintains a high degree of consistency with the real collected data in terms of the local clustering and spatial correlation of the sampling point distribution, effectively improving the generalization performance of the model under real trajectory sampling conditions.

[0140] As an example, please continue to refer to Figure 5 , Figure 5 This illustration shows a quantitative comparison of the radio map reconstruction performance of a radio map estimation method provided in this disclosure under different training-test sampling distributions, as shown in the diagram. Figure 5 As shown in the table, the “Training → Testing” column indicates the combination of sampling methods used in the training and testing phases. “Random” indicates independent random sampling, and “ST-TBS” indicates stochastic trigger-based trajectory sampling.

[0141] Performance metrics include Root Mean Square Error (RMSE, downward arrow indicates lower is better), Peak Signal-to-Noise Ratio (PSNR, upward arrow indicates higher is better), Structural Similarity Index Measure (SSIM, upward arrow indicates higher is better), and Robustness Gain. (The upward arrow indicates that a higher value is better). The experiments were evaluated on two public datasets, RadioMapSeer and SpectrumNet. Furthermore, Figure 5 The bolded values ​​in the table represent the best results in the same dataset.

[0142] Depend on Figure 5 It can be seen that when the training and testing sampling distributions are consistent (training with random sampled data and testing with random sampled data, i.e., Random→Random), the model can achieve high reconstruction accuracy on both datasets. The root mean square error is 0.0305 on the RadioMapSeer dataset and 0.0391 on the SpectrumNet dataset.

[0143] However, when the training sampling distribution and the test sampling distribution do not match (training with randomly sampled data and testing with randomly triggered trajectory sampling data, i.e., Random→ST-TBS), the reconstruction performance on both datasets shows a catastrophic decline: the root mean square error on the RadioMapSeer dataset increases from 0.0305 to 0.0775, and the root mean square error on the SpectrumNet dataset increases from 0.0391 to 0.2632, an increase of nearly seven times. The peak signal-to-noise ratio and the structural similarity index deteriorate significantly in tandem, proving that the inconsistency between the independent random sampling assumption and the distribution of the actual trajectory constraint sampling characteristics leads to a serious collapse in reconstruction performance.

[0144] In contrast, after training with trajectory-constrained sampling data using this scheme (training with randomly triggered trajectory sampling data and testing with randomly triggered trajectory sampling data, i.e., ST-TBS→ST-TBS), the reconstruction accuracy of the model under trajectory-constrained sampling testing conditions is significantly improved, and the root mean square error on the SpectrumNet dataset is reduced to 0.0571. Meanwhile, it also maintains good generalization performance and robustness gain under random sampling testing conditions (training with randomly triggered trajectory sampling data and testing with random sampling data, i.e., ST-TBS→random). The results achieved 60.00% and 79.29% on the two datasets, respectively, validating the effectiveness of the proposed scheme in achieving high-precision radio map estimation through training-deployment distribution alignment.

[0145] As another example, please refer to [link / reference]. Figure 6 , Figure 6 This illustration shows a quantitative comparison of the root mean square error of various radio map estimation models of a radio map estimation method provided in this disclosure under different training-test sampling distribution configurations, as shown in the diagram. Figure 6 As shown, the table columns identify the following four training-test configurations: Random → Random (R → R) means training with randomly sampled data and testing with randomly sampled data; Random → Trajectory (R→T) means training with randomly sampled data and testing with trajectory-constrained sampled data; Trajectory → Random (T→R), that is, training with trajectory-constrained sampled data and testing with random sampled data; Trajectory → Trajectory (T → T) means training with trajectory-constrained sampled data and testing with trajectory-constrained sampled data.

[0146] Figure 6 All evaluations were performed on the RadioMapSeer dataset, and the performance metric was root mean square error (RMSE, the lower the better).

[0147] The models included in the comparison can be divided into two categories: spatial interpolation methods and deep learning reconstruction networks. Traditional methods can include Kriging interpolation and cubic spline interpolation; deep learning methods can include PMNet, UNet, CBAM, RadioUNet, RadioDUN, and RadioFormer.

[0148] also, Figure 6 The bolded values ​​in the column represent the best results in the same column.

[0149] Depend on Figure 6 It can be seen that when the training and testing sampling distributions are consistent (R→R or T→T), each model can achieve relatively low reconstruction errors. RadioFormer has a root mean square error of 0.0186 under the R→R configuration, and RadioUNet has a root mean square error of 0.0570 under the T→T configuration, indicating that modern deep reconstruction models have strong spatial signal reconstruction capabilities when the training and deployment distributions are matched.

[0150] However, when the training and testing sampling distributions do not match, the reconstruction performance of all model architectures deteriorates significantly, and this trend is prevalent across traditional and deep learning methods and various network architectures, exhibiting the following characteristics: Taking RadioFormer as an example, its root mean square error increased from 0.0186 to 0.0989 under the random → trajectory (R→T) configuration; RadioUNet's root mean square error increased from 0.0305 to 0.0775 under the same configuration; similar performance degradation was observed for all models under the trajectory → random (T→R) configuration.

[0151] Understandable Figure 6 The above results show that the inconsistent training-deployment sampling distribution causes a serious loss of generalization performance for all existing reconstruction network architectures. This problem is universal and consistent across models, further proving that the inconsistent distribution problem solved by this scheme is a key bottleneck affecting the practical application effect of radio map estimation.

[0152] In one embodiment, a radio map estimation apparatus is provided, which corresponds one-to-one with the radio map estimation methods described in the above embodiments. For example... Figure 7 As shown, the device includes: The mask construction module 71 is used to obtain environmental information of the target area, divide the target area into multiple discrete spatial units, and construct a passable mask based on the environmental information; wherein, the passable mask is used to identify the spatial area within the target area that mobile devices are allowed to pass through; The random sampling module 72 is used to acquire signal measurement values ​​collected at each trajectory point when the mobile device moves along the actual motion trajectory in the target area, and to determine the sampling position from each trajectory point based on the probability triggering method to construct a sparse observation set; Input building module 73 is used to map sparse observation sets and passable masks to discrete spatial units, and to build a model input tensor that contains at least sparse observations, sampling masks and passable masks; The map estimation module 74 is used to input the model input tensor into a radio map estimation model trained with sparse observation data containing trajectory constraint characteristics, and output the radio map estimation result of the target area.

[0153] In one embodiment, the random sampling module 72 is specifically used to make random sampling decisions for each trajectory point in the actual motion trajectory based on a preset trigger probability; When sampling is triggered and the distance between the trajectory point and each recorded sampling position in the sparse observation set is not less than the preset minimum interval, the trajectory point is determined as the sampling position. The number of recorded sampling locations within a local spatial window centered on a trajectory point is counted. When the number of sampling locations exceeds a preset upper limit, the trigger probability of subsequent trajectory points within the local spatial window is reduced.

[0154] In one embodiment, the input construction module 73 is further configured to fuse multiple signal measurement values ​​for a discrete spatial unit with multiple signal measurement values ​​in a weighted fusion manner to obtain the observation value of the discrete spatial unit; The observation values ​​of each discrete spatial unit are linearly normalized according to the preset upper and lower bounds of the signal strength and mapped to the preset normalization interval. Normalized coordinate information of each discrete spatial unit is added to the channel dimension of the model input tensor.

[0155] In one embodiment, the random sampling module 72 is further configured to acquire historical signal measurement values ​​of multiple batches formed by multiple flights, multiple mobile devices, or multiple time periods within the target area, and map the historical signal measurement values ​​to discrete spatial units according to spatial location; For multiple historical signal measurements within the same discrete spatial unit, a time attenuation weight is set based on the acquisition time of each historical signal measurement, and weighted fusion is performed. The historical signal measurements corresponding to each batch of data acquisition are mapped to independent observation channels, and the historical sampling count and the time interval from the last sampling are added to the channel dimension of the model input tensor.

[0156] In one embodiment, a training apparatus for a radio map estimation model is provided, which corresponds one-to-one with the training method for the radio map estimation model in the above embodiments. For example... Figure 8 As shown, the device includes: The mask construction module 81 is used to divide the target area into multiple discrete spatial units based on the environmental information of the target area and construct a passable mask; The trajectory sampling module 82 is used to generate a simulated continuous trajectory that satisfies motion constraints based on a passable mask, and to perform random triggered sampling at each trajectory point of the simulated continuous trajectory to construct a simulated sparse observation set. Input building module 83 is used to map the simulated sparse observation set and the passable mask to discrete spatial cells, and to build a model input tensor that contains at least the simulated sparse observations, the sampling mask and the passable mask; The model training module 84 is used to train the radio map estimation model by using the model input tensor and the corresponding complete radio map to form training sample pairs, and using the reconstruction error on the effective spatial area marked by the passable mask as the loss function.

[0157] In one embodiment, the trajectory sampling module 82 is specifically used to select the starting point of the simulated trajectory from the effective spatial region of the passable mask identifier; The next direction of motion is selected from candidate directions whose angle with the current direction of motion does not exceed a preset angle threshold with a preset direction-maintaining probability; the next direction of motion is randomly selected from candidate directions that satisfy the passability constraint with a complementary probability. When all candidate directions that meet the passability constraints are unreachable, backtrack to the previous feasible position and re-execute the direction selection; when no feasible direction can be found after continuous backtracking, re-select the starting point of the simulation trajectory. Repeatedly select the direction of execution to gradually build a simulated continuous trajectory until the number of trajectory points of the simulated continuous trajectory reaches the preset trajectory length.

[0158] It should be noted that the radio map estimation device provided in the above embodiments, when implementing the corresponding radio map estimation method, and / or the radio map estimation model training device, when implementing the corresponding radio map estimation model training method, are only illustrated by the division of the above-described program modules. In practical applications, the above processing can be assigned to different program modules as needed, that is, the internal structure of the above device can be divided into different program modules to complete all or part of the processing described above. Furthermore, the device provided in the above embodiments and the corresponding... Figure 1 or Figure 4 The embodiments of the methods shown belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0159] This disclosure also provides an electronic device having the above-described features. Figure 7 The radio map estimation device shown, and / or the above-mentioned Figure 8 The training apparatus shown is for the radio map estimation model.

[0160] Figure 9 A schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure is shown.

[0161] The following is a detailed reference. Figure 9 The diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present disclosure. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 901, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 902 or a program loaded from memory 908 into random access memory (RAM) 903. The RAM 903 also stores various programs and data required for the operation of the electronic device. The processor 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0162] Typically, the following devices can be connected to I / O interface 905: input devices 906 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 907 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 908 including, for example, magnetic tapes, hard disks, etc.; and communication devices 909. Communication device 909 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 9 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0163] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 909, or installed from a memory 908, or installed from a ROM 902. When the computer program is executed by the processor 901, it performs the functions defined in the radio map estimation method and / or the radio map estimation model training method of embodiments of this disclosure.

[0164] Figure 9 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0165] This disclosure also provides a computer-readable storage medium in which the methods described in this disclosure can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium may also include combinations of the above types of memory. It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the radio map estimation method and / or radio map estimation model training method shown in the above embodiments are implemented.

[0166] A portion of this disclosure can be applied to computer program products, such as computer program instructions, which, when executed by a computer, can invoke or provide methods and / or technical solutions according to this disclosure through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, and installation package files. Accordingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions; the computer compiling the instructions and then executing the corresponding compiled program; the computer reading and executing the instructions; or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0167] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A radio map estimation method, characterized in that, include: The environmental information of the target area is obtained, the target area is divided into multiple discrete spatial units, and a passability mask is constructed based on the environmental information; wherein, the passability mask is used to identify the spatial area within the target area that mobile devices are allowed to pass through; The signal measurement values ​​collected at each trajectory point when the mobile device moves along the actual motion trajectory within the target area are obtained, and the sampling position is determined from each trajectory point based on the probability triggering method to construct a sparse observation set; Map the sparse observation set and the passable mask to the discrete spatial unit to construct a model input tensor that contains at least sparse observations, a sampling mask, and the passable mask; The model input tensor is a radio map estimation model trained with sparse observation data containing trajectory constraint characteristics, and the output is the radio map estimation result of the target area.

2. The method according to claim 1, characterized in that, The method of determining the sampling position from each trajectory point based on probability triggering includes: For each trajectory point in the actual motion trajectory, a random sampling decision is made based on a preset trigger probability; When sampling is triggered and the distance between the trajectory point and each recorded sampling position in the sparse observation set is not less than the preset minimum interval, the trajectory point is determined as the sampling position. The number of recorded sampling locations within a local spatial window centered on the trajectory point is counted. When the number of sampling locations exceeds a preset upper limit, the trigger probability of subsequent trajectory points within the local spatial window is reduced.

3. The method according to claim 2, characterized in that, When mapping the sparse observation set and the passable mask to the discrete spatial unit, the method further includes: For a discrete spatial unit with multiple signal measurements, the multiple signal measurements are fused according to a weighted fusion method to obtain the observation value of the discrete spatial unit; The observation values ​​of each discrete spatial unit are linearly normalized according to the preset upper and lower bounds of the signal strength and mapped to the preset normalization interval. Normalized coordinate information of each discrete spatial unit is added to the channel dimension of the input tensor of the model.

4. The method according to claim 1, characterized in that, The construction of the sparse observation set also includes: Acquire historical signal measurement values ​​from multiple batches of data collected during multiple flights, multiple mobile devices, or multiple time periods within the target area, and map the historical signal measurement values ​​to the discrete spatial units according to their spatial locations; For multiple historical signal measurements within the same discrete spatial unit, time attenuation weights are set based on the acquisition time of each historical signal measurement, and weighted fusion is performed. The historical signal measurements corresponding to each batch of data acquisition are mapped to independent observation channels, and the historical sampling count and the time interval from the last sampling are added to the channel dimension of the input tensor of the model.

5. A method for training a radio map estimation model, characterized in that, include: The target area is divided into multiple discrete spatial units based on the environmental information of the target area, and a passable mask is constructed. Based on the passable mask, a simulated continuous trajectory that satisfies motion constraints is generated. Randomly triggered sampling is performed on each trajectory point of the simulated continuous trajectory to construct a simulated sparse observation set. The simulated sparse observation set and the passable mask are mapped to the discrete spatial unit to construct a model input tensor that includes at least the simulated sparse observation values, the sampling mask, and the passable mask; The model input tensor and the corresponding complete radio map are used to form training sample pairs, and the reconstruction error on the effective spatial area marked by the passable mask is used as the loss function to train the radio map estimation model.

6. The method according to claim 5, characterized in that, The generation of a simulated continuous trajectory satisfying motion constraints based on the passable mask includes: Select the starting point of the simulation trajectory from the valid spatial area of ​​the passable mask identifier; The next direction of motion is selected from candidate directions whose angle with the current direction of motion does not exceed a preset angle threshold with a preset direction-maintaining probability; the next direction of motion is randomly selected from candidate directions that satisfy the passability constraint with a complementary probability. When all candidate directions that satisfy the passability constraint are unreachable, backtrack to the previous feasible position and re-execute the direction selection; when no feasible direction can be found after continuous backtracking, reselect the starting point of the simulation trajectory. Repeat the direction selection process to gradually construct the simulated continuous trajectory until the number of trajectory points of the simulated continuous trajectory reaches the preset trajectory length.

7. A radio map estimation device, characterized in that, The radio map estimation method for performing any one of claims 1 to 4 includes: A mask construction module is used to acquire environmental information of a target area, divide the target area into multiple discrete spatial units, and construct a passable mask based on the environmental information; wherein, the passable mask is used to identify the spatial area within the target area that mobile devices are allowed to pass through; The random sampling module is used to acquire signal measurement values ​​collected at each trajectory point when the mobile device moves along the actual motion trajectory within the target area, and to determine the sampling position from each trajectory point based on a probability triggering method to construct a sparse observation set. An input construction module is used to map the sparse observation set and the passable mask to the discrete spatial unit, and construct a model input tensor that includes at least sparse observations, a sampling mask, and the passable mask; The map estimation module is used to input the model input tensor into a radio map estimation model trained with sparse observation data containing trajectory constraint characteristics, and output the radio map estimation result of the target area.

8. A training device for a radio map estimation model, characterized in that, A training method for executing the radio map estimation model according to any one of claims 5 to 6, comprising: The mask construction module is used to divide the target area into multiple discrete spatial units based on the environmental information of the target area, and construct a passable mask; The trajectory sampling module is used to generate a simulated continuous trajectory that satisfies motion constraints based on the passable mask, and to perform random triggered sampling at each trajectory point of the simulated continuous trajectory to construct a simulated sparse observation set. An input construction module is used to map the simulated sparse observation set and the passable mask to the discrete spatial unit, and construct a model input tensor that includes at least the simulated sparse observation values, the sampling mask and the passable mask; The model training module is used to train the radio map estimation model by using the model input tensor and the corresponding complete radio map to form training sample pairs, and using the reconstruction error on the effective spatial area marked by the passable mask as the loss function.

9. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the radio map estimation method of any one of claims 1 to 4, and / or to perform the training method of the radio map estimation model of any one of claims 5 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the radio map estimation method of any one of claims 1 to 4, and / or the training method of the radio map estimation model of any one of claims 5 to 6.

11. A computer program product, characterized in that, It includes computer instructions for causing a computer to perform the radio map estimation method of any one of claims 1 to 4, and / or to perform the training method of the radio map estimation model of any one of claims 5 to 6.

Images