Sparse sampling-based spectrum situation awareness model updating method and device

Abstract

The application discloses a sparse sampling-based spectrum situation awareness model updating method and device, which comprises the following steps: obtaining a building layout map and a sparse received signal strength sampling map of a target environment to form an original two-channel input; constructing four enhanced feature maps through an enhanced physical propagation coding module; splicing the original two-channel input and the four enhanced feature maps to form a six-channel input; feeding the six-channel input into a fine-tuning model; the backbone network parameters of the fine-tuning model are frozen, and a multi-scale propagation perception nonlinear low-rank adaptive branch is arranged in the bypass of the backbone network; based on the supervision condition of the target environment, a corresponding loss function is constructed, and only the parameters of the multi-scale propagation perception nonlinear low-rank adaptive branch and the newly added trainable parameters related to the six-channel input expansion are updated to obtain an updated spectrum situation awareness model. Therefore, by using the application, the adaptation accuracy and robustness of the spectrum situation awareness model to the target environment can be significantly improved under the condition of only a small amount of sparse sampling data and extremely low calculation cost.

Description

Technical Field

[0001] This application relates to the field of spectrum model updating technology, and in particular to a method and apparatus for updating a spectrum situation awareness model based on sparse sampling. Background Technology

[0002] A spectrum situation map, also known as a radio map, is a data representation of the spatial distribution of radio spectrum-related parameters within a target geographical area. In wireless communication systems, constructing a spectrum situation map provides a direct understanding of signal strength distribution, interference areas, coverage blind spots, and spectrum resource availability, offering crucial information for dynamic spectrum access, interference management, network planning, base station deployment, and spectrum regulation. In practical applications, limited by sensor deployment costs, systems typically obtain received signal strength samples from only a few locations. Therefore, it is necessary to infer a high-resolution spectrum situation map of the entire area based on these sparse samples and environmental information—a problem of spectrum situation awareness based on sparse sampling. With the development of deep learning technology, spectrum situation awareness models based on U-Net, Transformer, Generative Adversarial Networks, and diffusion models have been extensively studied and have demonstrated superior performance compared to traditional interpolation methods.

[0003] Among related technologies, the closest existing technologies mainly fall into three categories. The first category is radio map estimation methods based on interpolation or physical modeling, such as Kriging interpolation, nearest neighbor interpolation, and inverse distance weighted interpolation. The basic process involves collecting received signal strength values ​​at several locations in the target area, constructing a spatial interpolation model, and then interpolating the unmeasured grid points to obtain a complete radio map. The second category is radio map estimation methods based on deep learning or generative artificial intelligence. These methods use networks such as CNN, U-Net, Transformer, GAN, and diffusion models to build a large number of training samples in the source environment, training the deep model to learn the nonlinear mapping from building maps and sparsely sampled maps to a complete radio map. The third category is model update methods based on transfer learning or efficient parameter fine-tuning, such as LoRA, Adapter, and BitFit. The basic idea is to fine-tune the pre-trained model using target environment data to adapt the model to the target environment.

[0004] However, the aforementioned existing technologies all have significant shortcomings. First, traditional interpolation methods mainly infer unknown areas based on the spatial distance or statistical correlation between measurement points, without fully considering building layout and propagation physics. This makes it difficult to recover the true propagation field structure in densely built-up, heavily obstructed urban areas. Second, pre-trained deep models perform well in the source environment, but when the building layout, propagation conditions, or sampling patterns of the target environment change, the model cannot automatically adapt to the new environment, easily leading to problems such as hotspot location shifts, insufficient signal peak estimation, and incorrect recovery of shadow areas, resulting in distorted spectrum situation maps. Third, retraining or full-parameter fine-tuning requires a large number of target environment samples and computational resources, while in actual spectrum mapping tasks, only a small number of sparse samples can be collected from the target environment, making large-scale training impossible. Fourth, existing radio map model update methods are usually model-specific algorithms, making it difficult to uniformly adapt to various model structures such as CNN, Transformer, and GAN. Fifth, standard LoRA, Adapter, BitFit and other parameter-efficient fine-tuning methods do not explicitly model path loss, multi-scale propagation, building occlusion and sampling coverage confidence. Under ultra-sparse sampling conditions, they are prone to overfitting a small number of sample points or producing non-physical prediction results. Summary of the Invention

[0005] This application provides a method and apparatus for updating a spectrum situational awareness model based on sparse sampling. To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0006] In a first aspect, embodiments of this application provide a method for updating a spectrum situation awareness model based on sparse sampling, the method comprising: The building layout map and sparse received signal strength sampling map of the target environment are obtained to form the original two-channel input; By enhancing the physical propagation coding module, four enhanced feature maps are obtained by constructing a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field based on the original two-channel input. The original two-channel input is concatenated with four enhanced feature maps to form a six-channel input. The six-channel input is fed into the model to be fine-tuned; the backbone network parameters of the model to be fine-tuned are frozen, and a multi-scale propagation sensing nonlinear low-rank adaptation branch is set in the bypass of the backbone network. The six-channel input is processed by the pre-sequence layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network. The input of the multi-scale propagation perceptual nonlinear low-rank adaptation branch is the input features. The target environment adaptation increment is generated by the multi-scale propagation perceptual nonlinear low-rank adaptation branch and added to the output of the layer of the backbone network to obtain the output feature map. Construct the corresponding loss function based on the supervision conditions of the target environment; Freeze the parameters of the backbone network; based on the output feature map and loss function, update the parameters of the multi-scale propagation perception nonlinear low-rank adaptation branch to obtain a spectrum situational awareness model adapted to the target environment.

[0007] Optionally, the building layout map and the sparse received signal strength sampling map correspond to multiple grid positions after the target area is discretized, and the non-zero positions in the sparse received signal strength sampling map correspond to measurement points; Based on the original two-channel input, a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field are constructed to obtain four enhanced feature maps, including: For each grid location, calculate its Euclidean distance to the nearest measurement point; Construct the original path loss value based on the Euclidean distance; The original path loss values ​​are normalized and reversed to obtain the logarithmic distance path loss field. For each grid location, an inverse distance weighted interpolation field is constructed based on the measurements of all measurement points and the inverse distance weights; The building density field is obtained by performing local average pooling on the building layout map. For each grid location, an exponentially decaying propagation confidence field is constructed based on its average distance to the nearest multiple measurement points.

[0008] Optionally, the model to be fine-tuned is constructed according to the following steps: Establish a backbone network; A multi-scale propagation sensing nonlinear low-rank adaptation branch is set as a bypass in at least one layer of the backbone network to obtain the model to be fine-tuned. The input of the multi-scale propagation sensing nonlinear low-rank adaptation branch is the input feature of the backbone network layer, and the output of the branch is added to the output of the backbone network layer to obtain the final output of the layer.

[0009] Optional, multi-scale propagation sensing nonlinear low-rank adaptation branches include: Low-rank dimensionality reduction mapping is used to compress the input features of this layer of the backbone network into a low-dimensional feature space to obtain low-rank features. Multi-scale parallel branches are set on a low-dimensional feature space, including an identity mapping branch and at least two deep convolution branches with different dilation rates; wherein the dilation rates of the at least two deep convolution branches are 2 and 4, respectively. The identity mapping branch is used to preserve the original low-rank features and near-field details, the deep convolution branch with a dilation rate of 2 is used to capture the medium-range propagation effect, and the deep convolution branch with a dilation rate of 4 is used to capture the spatial decay trend and long-distance propagation correlation over a larger range. A learnable gating unit is used to generate normalized weights for three branches, and to perform a weighted summation of the outputs of the identity mapping branch and at least two deep convolutional branches with different dilation rates based on the normalized weights to obtain multi-scale fusion features.

[0010] Optionally, the six-channel input is processed through the pre-processing layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network, and the input of the multi-scale propagation perceptual nonlinear low-rank adaptation branch is the input features; the target environment adaptation increment is generated through the multi-scale propagation perceptual nonlinear low-rank adaptation branch and added to the output of the layer of the backbone network to obtain the output feature map, including: In the multi-scale propagation perception nonlinear low-rank adaptation branch, the input features are sequentially subjected to low-rank dimensionality reduction, multi-scale parallel branch processing, and learnable gated weighted fusion to obtain multi-scale fused features. A propagation-aware nonlinear function is applied to the multi-scale fused features. This propagation-aware nonlinear function is used to perform logarithmic compression on large-amplitude feature responses while preserving the positive and negative signs. The features processed by the propagation sensing nonlinear function are subjected to low-rank dimensionality upsizing to obtain the target environment adaptation increment. The target environment adaptation increment is multiplied by a scaling factor and then added to the output of the corresponding layer of the backbone network to obtain the final output feature map of that layer.

[0011] Optionally, based on the supervision conditions of the target environment, a corresponding loss function is constructed, including: When the target environment has complete spectral situation map labels, the dense reconstruction loss is calculated in non-building areas under fully supervised conditions; or... When the target environment only has sparse received signal strength sampling points, an ultra-sparse supervision condition is adopted to divide the sparse received signal strength sampling points into a set of known points and a set of query points, and a frozen teacher model is introduced; wherein, the teacher model is a frozen copy of the pre-trained model of the source environment, and its input is the original two-channel input; Construct a joint loss function, which includes: a sparse supervision loss computed only on the query point set, a knowledge distillation loss used to constrain the spatial structure of the model to be fine-tuned to be consistent with the teacher model, and a total variation regularization loss used to constrain the continuity of the output space of the model to be fine-tuned.

[0012] Optionally, the parameters of the multi-scale propagation perception nonlinear low-rank adaptation branch are updated to obtain a spectral situational awareness model adapted to the target environment, including: Under fully supervised conditions, dense reconstruction loss is calculated based on the output feature map and the label of the complete spectral situation map, and dense reconstruction loss is used as the optimization objective. Under ultra-sparse supervision, the joint loss function is calculated based on the output feature map, the output of the frozen teacher model, and the set of query points in the sparse received signal strength sampling points, and the joint loss function is used as the optimization objective. During the optimization process, the parameters of the backbone network are kept frozen, and only the parameters of the multi-scale propagation sensing nonlinear low-rank adaptation branch and the newly added trainable parameters related to the six-channel input expansion are updated. After optimization, a spectrum situational awareness model adapted to the target environment is obtained.

[0013] Optionally, after updating the parameters of the multi-scale propagation sensing nonlinear low-rank adaptation branch to obtain the spectral situational awareness model adapted to the target environment, the following steps are also included: Obtain a new sparse received signal strength sampling map of the target environment; Based on the new sparse received signal strength sampling map and building layout map, the corresponding logarithmic distance path loss field, inverse distance weighted interpolation field, building density field and propagation confidence field are constructed through the enhanced physical propagation coding module. The building layout map, the new sparse received signal strength sampling map and the four enhanced feature maps are then spliced ​​together to form a new six-channel input. The new six-channel input is fed into a spectrum situation awareness model adapted to the target environment, and a complete spectrum situation map under the target environment is output.

[0014] Optionally, the knowledge distillation loss is the gradient distillation loss, used to constrain the gradient distribution of the model to be fine-tuned to be consistent with that of the teacher model in the output space.

[0015] Secondly, embodiments of this application provide a spectrum situation awareness model update device based on sparse sampling, the device comprising: The original two-channel input generation module is used to obtain the building layout map and sparse received signal strength sampling map of the target environment, which constitute the original two-channel input; The enhanced feature map generation module is used to construct four enhanced feature maps based on the original two-channel input, namely, a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field, through the enhanced physical propagation coding module. The feature concatenation module is used to concatenate the original two-channel input with four enhanced feature maps to form a six-channel input. The feature input module is used to feed the six-channel input into the model to be fine-tuned; the backbone network parameters of the model to be fine-tuned are frozen, and a multi-scale propagation sensing nonlinear low-rank adaptation branch is set in the bypass of the backbone network. The branch processing module is used to process the six-channel input through the preceding layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network. The input of the multi-scale propagation perceptual nonlinear low-rank adaptation branch is the input features. The target environment adaptation increment is generated through the multi-scale propagation perceptual nonlinear low-rank adaptation branch and added to the output of the layer of the backbone network to obtain the output feature map. The loss function construction module is used to construct the corresponding loss function based on the supervision conditions of the target environment; The model generation module is used to freeze the parameters of the backbone network; based on the output feature map and loss function, the parameters of the multi-scale propagation perception nonlinear low-rank adaptation branch are updated to obtain a spectrum situational awareness model adapted to the target environment.

[0016] The technical solutions provided in this application embodiment may include the following beneficial effects: In this embodiment, on the one hand, by introducing an enhanced physical propagation coding module at the input end, four enhanced feature maps are constructed in addition to the original building layout map and sparse sampling map: a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field. This provides the model with explicit spatial priors related to radio propagation physics and sampling layout. Compared with the existing technology that relies solely on sparse sampling points and building maps as input, this application enables the model to directly utilize information such as distance attenuation trends, measurement extrapolation trends, local occlusion backgrounds, and sampling coverage when fine-tuning in the target environment. It does not require relearning propagation laws entirely from a small number of samples, thereby significantly improving the estimation accuracy of the spectral situation map in the target environment. This effectively overcomes the problems of hotspot location shifts, insufficient signal peak estimation, and incorrect recovery of shadow areas caused by distribution shifts in the target environment in existing pre-trained models.

[0017] On the other hand, this application freezes the pre-trained backbone network parameters and introduces only a small number of trainable parameters for target environment adaptation through a multi-scale propagation perceptual nonlinear low-rank adaptation branch, eliminating the need for full parameter retraining or fine-tuning of the entire model. Compared to existing retraining methods that require a large number of target environment samples and computational resources, this application significantly reduces the number of training parameters, computational overhead, and storage requirements, and can still efficiently complete model updates even in actual spectral mapping scenarios where only a small number of sparse samples can be collected in the target environment. Simultaneously, because the backbone parameters remain frozen, the model can better retain the general propagation knowledge learned in the source environment, reducing the risk of catastrophic forgetting and achieving low-cost, high-efficiency target environment adaptation.

[0018] On the other hand, this application introduces multi-scale parallel branches and a propagation-aware nonlinear function. The multi-scale parallel branches, through identity mapping and deep convolutional branches with different dilation rates, can simultaneously capture near-field local details, mid-range diffraction effects, and long-range path loss trends, enabling the model to adaptively select appropriate scale combinations based on different spatial locations and propagation states. The propagation-aware nonlinear function performs sign-preserving logarithmic compression on large-amplitude feature responses, mitigating the excessive influence of extreme responses on the low-rank update direction. Compared to the single linear low-rank correction of standard LoRA, this application better conforms to the characteristics of radio propagation. With only a small increase in trainable parameters, the model simultaneously adapts to near-range details, mid-range diffraction, and long-range large-scale attenuation patterns, overcoming the deficiency of standard parameter-efficient fine-tuning methods that lack propagation-aware capabilities.

[0019] On the other hand, under ultra-sparse supervision conditions where the target environment cannot provide complete spectral situation map labels and only a very small number of sparse sampling points are available, effective structural and smoothing constraints are provided for unlabeled regions. Specifically, freezing the gradient distillation provided by the teacher model allows the student model to inherit the spatial structure and variation trends of the pre-trained model in the source environment in unlabeled regions; total variation regularization loss suppresses high-frequency noise and unreasonable abrupt changes in open propagation regions; and sparse supervision loss constrains labeled sampling points. The synergistic effect of these three factors enables the model to obtain more reliable complete spectral situation map estimation results under ultra-sparse supervision conditions, avoiding non-physical oscillations in unobserved regions caused by the model only fitting a small number of measurement points. This solves the problem of unreliable model updates under ultra-sparse sampling conditions in existing technologies.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] Figure 1 This is a flowchart illustrating a sparse sampling-based spectrum situation awareness model update method provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a multi-scale propagation sensing nonlinear low-rank adaptation module provided in an embodiment of this application; Figure 3 This is a schematic diagram of the overall process of a fine-tuning architecture for a spectrum situation awareness model oriented towards ultra-sparse supervision, provided in an embodiment of this application. Figure 4 This is a schematic diagram of the structure of a spectrum situation awareness model update device based on sparse sampling provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a device provided in an embodiment of this application. Detailed Implementation

[0023] The following description and accompanying drawings fully illustrate specific embodiments of this application to enable those skilled in the art to practice them.

[0024] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0025] In the following description, when referring to the accompanying drawings, the same numbers in different drawings denote the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0026] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0027] The following will be combined with the appendix Figure 1 -Appendix Figure 3 This application provides a detailed description of the sparse sampling-based spectrum situation awareness model update method provided in its embodiments. This method can be implemented using a computer program and can run on a sparse sampling-based spectrum situation awareness model update device based on the von Neumann architecture. This computer program can be integrated into applications or run as a standalone utility application.

[0028] Please see Figure 1 This document provides a flowchart illustrating a method for updating a spectrum situation awareness model based on sparse sampling, as described in an embodiment of this application. Figure 1 As shown, the method in this application embodiment may include the following steps: S101, Obtain the building layout map and sparse received signal strength sampling map of the target environment to form the original two-channel input; In some embodiments of this application, two types of raw input information are obtained from the target environment. The first type of information is a building layout map of the target environment, used to describe buildings, walls, obstacles, or other spatial structures within the target environment that may affect radio propagation. The second type of information is a sparse received signal strength sampling map, which consists of sparse received signal strength sampling points collected by sensors in the target environment, used to represent radio signal measurements at a small number of known locations in the target environment. The building layout map is denoted as... The sparse received signal strength sampling pattern is denoted as Thus, the original two-channel input is obtained. ,in For batch size.

[0029] S102, by enhancing the physical propagation coding module, four enhanced feature maps are constructed based on the original two-channel input: logarithmic distance path loss field, inverse distance weighted interpolation field, building density field, and propagation confidence field. Among them, the building layout map and the sparse received signal strength sampling map correspond to multiple grid positions after the target area is discretized, and the non-zero positions in the sparse received signal strength sampling map correspond to measurement points.

[0030] For example, let the target area be... Discretize it into There are 1 grid points, and the grid position is denoted as . The layout of buildings or obstacles can be represented by a binary graph. It means that, among them This indicates that the location is a building or obstacle. This represents the non-building area. Signal strength received inside buildings is generally not considered a valid evaluation metric, but buildings can affect surrounding radio propagation through obstruction, reflection, and attenuation. Let the set of available measurement points be... ,in The received signal strength at the m-th measurement point is The sparse measurements are embedded into a sparse received signal strength sampling map. :

[0031] The goal of the spectrum situational awareness model is to use building maps and sparse sampling map Output a complete spectrum situation map ,Right now:

[0032] in It should be as close as possible to the actual spectrum situation. Especially in non-building areas Inside.

[0033] In some embodiments of this application, the specific process of constructing a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field based on the original two-channel input to obtain four enhanced feature maps includes: for each grid location, calculating the Euclidean distance between it and the nearest measurement point; constructing the original path loss value based on the Euclidean distance; normalizing and inversely processing the original path loss value to obtain the logarithmic distance path loss field; for each grid location, constructing the inverse distance weighted interpolation field based on the measured values ​​of all measurement points and the inverse distance weight; performing local average pooling on the building layout map to obtain the building density field; and for each grid location, constructing an exponentially decaying propagation confidence field based on the average distance to the nearest multiple measurement points.

[0034] For example, after obtaining the original two-channel input, this application further utilizes the Enhanced Physical Propagation Coding Module (EPPE) to construct four additional feature channels. For each input sample, it first starts from the sparse sampling map. Determine all known measurement points. If And if the value is valid, then These are considered known measurement points. If the number of known measurement points N is too large, a maximum upper limit can be set for the number of sampling points, for example, 200. Measurement points not exceeding this upper limit can be randomly selected from the set of measurement points to participate in the EPPE feature construction, thereby reducing computational complexity. If N=0, then all four EPPE feature channels are set to zero.

[0035] For any grid location Calculate its distance to the m-th sampling point Euclidean distance: The following four enhanced feature maps are then constructed: The first enhanced feature map is used to construct a logarithmic distance path loss field. This feature channel is used to describe the distance decay trend of each location relative to the nearest sampling point. The distance to the nearest measurement point is defined as: ; Based on the logarithmic distance path loss model, the original path loss field is constructed as follows: ; Where n is the path loss exponent, which is taken as n=2 in one embodiment. Let: ; in To prevent extremely small constants from being divided by zero, the path loss field is normalized and reversed to obtain the first enhanced feature map: The larger the value of this feature map, the closer the current location is to the known measurement point, and the stronger the measurement support.

[0036] The second enhanced feature map constructs an inverse distance-weighted interpolation field. This feature channel is used to form a continuous, coarse spatial extrapolation field based on sparse measurement points. For each measurement point, the inverse distance weights are defined as follows: ; in To prevent extremely small constants from being divided by zero, a second enhanced feature map is constructed based on the measured values ​​and weights of all measurement points:

[0037] This feature map enables sparse sampled values ​​to form a smooth extrapolation trend in space, providing preliminary spectral distribution information supported by actual measurement data for subsequent models.

[0038] The third enhanced feature map constructs a building density field. This feature channel is used to describe the local density of buildings or obstacles around each location. (Building layout map) Perform local average pooling, for example using The window shows:

[0039] This feature map can reflect the local occlusion environment, enabling the model to perceive the propagation differences between densely built-up areas and open areas.

[0040] The fourth enhanced feature map constructs a propagation confidence field. This feature channel describes the degree of support for that location from nearby measurement points. The diagonal length of the region is defined as follows:

[0041] For each position Select the one with the smallest distance to all measurement points. Calculate the average distance. ,in The propagation confidence field is defined as:

[0042] The larger the value of the feature map, the more measurement points are present near the current location; the smaller the value, the farther the location is from the measurement points, and the more the model's inference about the location should rely on propagation priors and the model's learning ability.

[0043] S103 concatenates the original two-channel input with four enhanced feature maps to form a six-channel input; In some embodiments of this application, the four enhanced feature maps generated by the EPPE module are concatenated to obtain a four-channel enhanced physical propagation code:

[0044] Then, the original two-channel input was... With four-channel enhanced coding By concatenating the inputs, we obtain a six-channel input:

[0045] Since pre-trained spectral situation awareness models typically only receive two-channel input, this application expands the first input layer of the model from two channels to six channels. During expansion, the weights corresponding to the original two channels are copied from the pre-trained model, and the weights corresponding to the four newly added EPPE channels are initialized to zero. In this way, in the initial stage of model updates, the model still maintains the original input response consistent with the pre-trained model, while allowing the newly added EPPE channels to be gradually utilized by the model in subsequent fine-tuning.

[0046] S104, the six-channel input is sent to the model to be fine-tuned; the backbone network parameters of the model to be fine-tuned are frozen, and a multi-scale propagation sensing nonlinear low-rank adaptation branch is set in the bypass of the backbone network. Typically, with the development of deep learning and generative artificial intelligence, an increasing number of models are being used for radio map estimation, such as U-Net-based convolutional models, Transformer-based models, generative adversarial network models, and diffusion models. These models are usually pre-trained on source domain datasets to learn the propagation patterns in the source environment. The pre-training process can be represented as:

[0047] in For the source environment dataset, These are the model parameters obtained through pre-training. During actual deployment, the model needs to be applied to the target environment. The target environment and the source environment may differ significantly in building layout, propagation conditions, emission source distribution, frequency, and sampling mode. If a pre-trained model is used directly... Model performance will degrade. Therefore, it is necessary to update the model using a small amount of new data from the target environment to adapt it to the target environment. To reduce update costs, a reasonable approach is to freeze the pre-trained parameters. Only a small number of trainable parameters are introduced. The updated model is:

[0048] in The goal of model updating is to find a model suitable for the target environment. .

[0049] In some embodiments of this application, a backbone network is established; a multi-scale propagation sensing nonlinear low-rank adaptation branch is set as a bypass in at least one layer of the backbone network to obtain the model to be fine-tuned; wherein, the input of the multi-scale propagation sensing nonlinear low-rank adaptation branch is the input feature of the layer of the backbone network, and the output of the branch is added to the output of the layer of the backbone network to obtain the final output of the layer.

[0050] Specifically, the multi-scale propagation perception nonlinear low-rank adaptation branch includes: a low-rank dimensionality reduction mapping, used to compress the input features of this layer of the backbone network into a low-dimensional feature space to obtain low-rank features; a multi-scale parallel branch, set on the low-dimensional feature space, including an identity mapping branch and at least two deep convolutional branches with different dilation rates; wherein the dilation rates of the at least two deep convolutional branches are 2 and 4, respectively, the identity mapping branch is used to preserve the original low-rank features and near-field details, the deep convolutional branch with a dilation rate of 2 is used to capture the propagation effect over a medium range, and the deep convolutional branch with a dilation rate of 4 is used to capture the spatial decay trend and long-distance propagation correlation over a larger range; and a learnable gating unit, used to generate normalized weights for the three branches, and to perform a weighted summation of the outputs of the identity mapping branch and the at least two deep convolutional branches with different dilation rates according to the normalized weights to obtain multi-scale fused features.

[0051] For example Figure 2 As shown, Figure 2 This is a schematic diagram of a multi-scale propagation-aware nonlinear low-rank adaptation module provided in this application. This module is a trainable component used to achieve target environment adaptation based on a frozen backbone network. First, the layer input is divided into two paths: one path inputs to the multi-scale propagation-aware nonlinear low-rank adaptation branch, and the other path inputs to the frozen backbone network. In the multi-scale propagation-aware nonlinear low-rank adaptation branch, low-rank features are first obtained through low-rank dimensionality reduction mapping. Then, the low-rank features are processed by three parallel sub-branches: ① Identity mapping, directly passing the input features; ② Depthwise separable convolution (dilation rate = 2), extracting mid-scale propagation features; ③ Depthwise separable convolution (dilation rate = 4), extracting large-scale propagation features. The outputs of the three sub-branches generate gating parameters through learnable scale gating. These gating parameters are then normalized into weights using a normalized exponential function. The weights are then used to weight and sum the outputs of the three branches, achieving weighted fusion of multi-scale features. The weighted summed features are input into a propagation-aware nonlinear (PAN) module (which performs sign-preserving logarithmic compression on the features to mitigate the impact of extreme responses). This is then processed by LoRA-B low-rank upscaling mapping, and the update magnitude is adjusted using ×α / r (scaling operation, where α is the scaling factor and r is the low-rank). Finally, the outputs of the multi-scale propagation-aware nonlinear low-rank adaptation branch and the output of the frozen backbone network are added element-wise to obtain the layer output.

[0052] For example Figure 2 As shown, the pre-trained backbone retains the general spectral propagation representation learned in the source environment, and its main parameters remain frozen; the multi-scale propagation-aware nonlinear low-rank adaptation is a trainable branch used to learn incremental corrections of the target environment relative to the source environment. For any layer to be adapted in the model, let the input of that layer be... The original frozen trunk layer is denoted as... Its output is:

[0053] At the same time, input The data is fed into a multi-scale propagation perceptual nonlinear low-rank adaptation branch. This branch first performs low-rank dimensionality reduction using the LoRA-A low-rank dimensionality reduction mapping (LoRA-A) to obtain low-rank features: ; in Representing a LoRA-A mapping, it can typically be derived from... Convolution is implemented. Subsequently, low-rank features are processed. Construct a multi-scale parallel branch, comprising three branches: an identity mapping branch, a depthwise convolution branch with a dilation rate of 2, and a depthwise convolution branch with a dilation rate (i.e., expansion rate) of 4. The three branches output: , , .in, This represents a depthwise convolution with an inflation rate of 2. This represents a deep convolution with a dilation rate of 4. The identity mapping branch is used to preserve the original low-rank features and near-field details; the deep convolution branch with a dilation rate of 2 is used to capture medium-range propagation effects; and the deep convolution branch with a dilation rate of 4 is used to capture larger-scale spatial decay trends and long-distance propagation associations.

[0054] Next, learnable scale-gated parameters are introduced. The normalized weights of the three branches are obtained through Softmax (normalized exponential function). ,in The outputs of the three branches are weighted and summed according to their respective branch weights to obtain the multi-scale fusion feature: ; This weighted fusion mechanism enables the model to adaptively select the combination ratio between local details, mesoscale diffraction, and large-scale path loss based on the propagation characteristics of different regions in the target environment.

[0055] S105, the six-channel input is processed through the pre-processing layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network, and the input of the multi-scale propagation perception nonlinear low-rank adaptation branch is the input features; the target environment adaptation increment is generated through the multi-scale propagation perception nonlinear low-rank adaptation branch and added to the output of the layer of the backbone network to obtain the output feature map. In some embodiments of this application, the six-channel input is processed by the pre-processing layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network, and the input of the multi-scale propagation perception nonlinear low-rank adaptation branch is the input features. The specific process of generating the target environment adaptation increment through the multi-scale propagation perception nonlinear low-rank adaptation branch and adding it with the output of the corresponding layer of the backbone network to obtain the output feature map includes: in the multi-scale propagation perception nonlinear low-rank adaptation branch, the input features are sequentially subjected to low-rank dimensionality reduction, multi-scale parallel branch processing, and learnable gated weighted fusion to obtain multi-scale fused features; a propagation perception nonlinear function is applied to the multi-scale fused features, which is used to perform logarithmic compression on large feature responses while retaining positive and negative signs; the features processed by the propagation perception nonlinear function are subjected to low-rank dimensionality increase to obtain the target environment adaptation increment; the target environment adaptation increment is multiplied by a scaling factor and added with the output of the corresponding layer of the backbone network to obtain the final output feature map of that layer.

[0056] In one possible implementation, for example Figure 2 As shown, the multi-scale fusion features are obtained by weighted summation. Subsequently, this application introduces propagation-sensory nonlinearity (PAN). For For any scalar element t in the PAN, PAN is defined as: ; This function can logarithmically compress large-amplitude feature responses while preserving the sign of the signal. Since spectral patterns and path loss-related features typically have a large dynamic range, directly using linear low-rank branches may lead to a few extreme responses dominating model updates. PAN can alleviate this problem, allowing the adaptation branch to more stably learn the differences in target environment propagation.

[0057] After PAN processing, the feature input LoRA-B low-rank up-dimensional mapping (LoRA-B) is up-dimensionalized to obtain the low-rank adaptation branch output: ; in This indicates a LoRA-B mapping, where the number of output channels is the same as that of the frozen backbone layer. Then, the output of the low-rank adaptor branch is added to the output of the frozen backbone to obtain the final output of this layer. ; in This is the scaling factor. This refers to the number of low-rank channels. During fine-tuning, the main trunk is frozen. No update is performed; trainable parameters such as LoRA-A, the depthwise convolutional branch, the Scale Gate, and the PAN-enhanced LoRA-B in MSPAN-LoRA are included in the update.

[0058] S106, Construct the corresponding loss function based on the supervision conditions of the target environment; Typically, in the target environment, the data available for model updates falls into two categories: the first is fully supervised conditions, where a relatively complete true spectral situation map can be obtained in the target environment. In this case, dense reconstruction loss can be used to directly constrain the model output. The second type is ultra-sparse supervision conditions, where a complete spectral situation map cannot be obtained in the target environment, and only a small number of received signal strength sampling points are available. In this case, the sampling points can be divided into a set of known points. and query point set Known points are used to construct the input sparse sampling map, while query points provide the optimization objective. Since the supervision signal exists only at a very small number of query points, and most of the complete map is unlabeled, model updates under this condition are significantly more difficult.

[0059] In some embodiments of this application, the specific process of constructing the corresponding loss function according to the supervision conditions of the target environment includes: when the target environment has complete spectral situation map labels, a fully supervised condition is adopted to calculate the dense reconstruction loss in the non-building area; or, when the target environment only has sparse received signal strength sampling points, an ultra-sparse supervision condition is adopted to divide the sparse received signal strength sampling points into a set of known points and a set of query points, and a frozen teacher model is introduced; wherein, the teacher model is a frozen copy of the pre-trained model of the source environment, and its input is the original two-channel input; a joint loss function is constructed, which includes: a sparse supervision loss calculated only on the query point set, a knowledge distillation loss used to constrain the consistency of the spatial structure between the model to be fine-tuned and the teacher model, and a total variation regularization loss used to constrain the continuity of the output space of the model to be fine-tuned.

[0060] Among them, the knowledge distillation loss is the gradient distillation loss, which is used to constrain the gradient distribution of the model to be fine-tuned to be consistent with that of the teacher model in the output space.

[0061] In one possible implementation, this application supports two types of target environment supervision conditions: fully supervised conditions and ultra-sparse supervised conditions.

[0062] Under full supervision, the target environment can provide a complete and realistic spectral situation map. At this point, the model to be fine-tuned with MSPAN-LoRA is based on the six-channel input. Output Predicted Spectrum Situation Chart and in non-building areas The dense reconstruction loss can be calculated using the mean squared error as the fully supervised loss.

[0063] Under these conditions, the model primarily relies on EPPE and MSPAN-LoRA to update the target environment.

[0064] Under ultra-sparse supervision, the target environment cannot provide a complete and realistic spectral situation map, but only a very small number of received signal strength sampling points. In this case, the sampling points are divided into a set of known points. and query point set The set of known points is used to construct the sparse sampling map in the model input, and the set of query points is used to calculate the supervision loss. Since the supervision signal exists only in a small number of query points, most of the model output region does not have direct label constraints, so a KDTV module needs to be introduced further.

[0065] In another possible implementation, under ultra-sparse supervision, this application uses a frozen teacher model to provide a dense structural reference for the model to be fine-tuned. The teacher model is a frozen copy of the pre-trained model from the source environment, and its input is the original two-channel input. The output is a teacher-predicted spectrum trend chart: ; in This indicates that the teacher model is frozen. The teacher model does not participate in parameter updates and is only used to provide a reference for spatial structure and change trends in unlabeled areas.

[0066] The model to be fine-tuned serves as the student model, with its input being a six-channel augmented input. The output is a student-predicted spectrum trend chart: ; in This represents the frozen pre-trained backbone parameters. This indicates the newly added trainable adaptation parameters.

[0067] Under ultra-sparse supervision, this application constructs a joint loss function consisting of sparse supervision loss, total variation regularization loss, and knowledge distillation loss. The total loss is:

[0068] in and These are non-negative weighting coefficients.

[0069] First, we construct a sparse supervised loss. Let... For the query point mask, when the position Belongs to the query point set hour ,otherwise The sparse supervision loss is:

[0070] This loss is calculated only at the query sampling point to ensure that the student model outputs true measurements consistent with the target environment.

[0071] Secondly, construct the total variation regularization loss. Define the difference between the student model output in the vertical and horizontal directions as:

[0072]

[0073] set up Let be the set of edges where two adjacent positions in the vertical direction both belong to non-building regions. Let be the set of edges where two adjacent positions in the horizontal direction both belong to non-building regions. The total variation regularization loss is:

[0074] This loss is used to suppress high-frequency noise and unreasonable abrupt changes in unlabeled areas, so that the predicted spectral situation map maintains spatial continuity that conforms to the characteristics of the propagation field.

[0075] Finally, a knowledge distillation loss is constructed. Considering that the teacher model is pre-trained from the source environment and may have numerical biases in the target environment, this application preferably uses gradient distillation, rather than forcing the student model to output numerical values ​​exactly the same as the teacher model. Let... and Let represent the spatial gradients in the vertical and horizontal directions, respectively. Then, the knowledge distillation loss is:

[0076] This loss allows the student model to inherit the spatial structure and trends of the teacher model in unlabeled regions, such as occlusion edges, distance decay trends, and large-scale propagation distribution, while still being able to be corrected based on the query points in the target environment.

[0077] S107, freeze the backbone network parameters; based on the output feature map and loss function, update the parameters of the multi-scale propagation perception nonlinear low-rank adaptation branch to obtain a spectrum situational awareness model adapted to the target environment.

[0078] In some embodiments of this application, the specific process of updating the parameters of the multi-scale propagation sensing nonlinear low-rank adaptation branch to obtain a spectrum situation awareness model adapted to the target environment includes: under fully supervised conditions, calculating the dense reconstruction loss based on the output feature map and the labels of the complete spectrum situation map, and using the dense reconstruction loss as the optimization objective; under ultra-sparse supervised conditions, calculating the joint loss function based on the output feature map, the output of the frozen teacher model, and the query point set in the sparse received signal strength sampling points, and using the joint loss function as the optimization objective; during the optimization process, keeping the backbone network parameters frozen, only updating the parameters of the multi-scale propagation sensing nonlinear low-rank adaptation branch, as well as the newly added trainable parameters related to the six-channel input extension; after optimization, a spectrum situation awareness model adapted to the target environment is obtained.

[0079] In one possible implementation, under fully supervised conditions, the dense reconstruction loss is minimized; under ultra-sparse supervised conditions, the joint loss is minimized. During training, the pre-trained backbone model parameters Keep frozen, only update newly added adaptation parameters. These trainable parameters mainly include the LoRA-A parameters in MSPAN-LoRA, convolutional branch parameters at different scales, scale gate parameters, LoRA-B parameters, and necessary parameters related to the six-channel input expansion.

[0080] After optimization, the target environment adaptation parameters are obtained:

[0081] in Under fully supervised conditions, it is expressed as Under ultra-sparse supervised conditions, it represents .

[0082] Finally, the updated spectrum situational awareness model is obtained: .

[0083] Furthermore, after obtaining the spectrum situation awareness model adapted to the target environment, the model is deployed in the actual scenario to obtain a new sparse received signal strength sampling map of the target environment. Based on the new sparse received signal strength sampling map and the building layout map, the corresponding logarithmic distance path loss field, inverse distance weighted interpolation field, building density field and propagation confidence field are constructed through the enhanced physical propagation coding module. The building layout map, the new sparse received signal strength sampling map and the four enhanced feature maps are then concatenated to form a new six-channel input. The new six-channel input is fed into the spectrum situation awareness model adapted to the target environment to output a complete spectrum situation map of the target environment.

[0084] In one possible implementation, during the subsequent inference phase, for new sparsely sampled inputs in the target environment, the system still constructs a six-channel enhanced input following the above process and feeds it into the updated model, outputting a complete spectrum situation map. This output can be used for spectrum situation awareness, radio map construction, dynamic spectrum access, interference management, and wireless network planning in the target environment.

[0085] For example Figure 3 As shown, Figure 3 This is a schematic diagram of the overall process of a fine-tuning architecture for a spectrum situational awareness model under ultra-sparse supervision, as provided in this application. It aims to achieve efficient target environment adaptation under extremely low sampling conditions by integrating physical propagation priors and knowledge distillation mechanisms. First, the original input (2 channels) includes channel 0 (building layout map) and channel 1 (sparse received signal strength sampling points). This input is processed in parallel via two paths: one path enters the enhanced physical propagation encoding module (EPPE, 4 channels), which generates channels 2 (logarithmic distance path loss field), 3 (inverse distance weighted interpolation field), 4 (building density field), and 5 (propagation confidence field) through physical modeling, and concatenates these with the original input to form a 6-channel input; the other path directly inputs to the teacher model (pre-trained and with parameters frozen) to provide knowledge guidance under ultra-sparse supervision. Subsequently, the 6-channel input is fed into the fine-tuning model, which consists of a frozen backbone network and a multi-scale propagation-aware nonlinear low-rank adaptation branch (MSPAN-LoRA, trainable). MSPAN-LoRA includes LoRA-A low-rank dimensionality reduction mapping, multi-scale convolutional branches, scale-gated weighting, propagation-aware nonlinear activation, and LoRA-B low-rank dimensionality increase mapping. The final output is added to the frozen backbone network features to form the layer output. Simultaneously, the teacher model output and the fine-tuning model output are fed into the loss calculation module, which integrates three types of losses: supervised loss. (Based on sparse sampling points), L1 total variation loss (Smooth spatial structure), knowledge distillation loss (Transferring teacher model space knowledge), the total loss is defined as... Finally, based on the total loss, the model is updated, with only the MSPAN-LoRA branch parameters updated via backpropagation, while the main network remains frozen. This achieves high-precision adaptation of the target environment's spectral situation map with a minimal number of parameters under ultra-sparse supervision. This architecture effectively overcomes the problems of overfitting, poor generalization, and high computational cost of traditional fine-tuning methods in sparse sampling scenarios by explicitly introducing a physical propagation field, dynamic multi-scale feature fusion, and knowledge distillation constraints. It also possesses good engineering deployability and environmental transfer capabilities.

[0086] In this embodiment, on the one hand, by introducing an enhanced physical propagation coding module at the input end, four enhanced feature maps are constructed in addition to the original building layout map and sparse sampling map: a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field. This provides the model with explicit spatial priors related to radio propagation physics and sampling layout. Compared with the existing technology that relies solely on sparse sampling points and building maps as input, this application enables the model to directly utilize information such as distance attenuation trends, measurement extrapolation trends, local occlusion backgrounds, and sampling coverage when fine-tuning in the target environment. It does not require relearning propagation laws entirely from a small number of samples, thereby significantly improving the estimation accuracy of the spectral situation map in the target environment. This effectively overcomes the problems of hotspot location shifts, insufficient signal peak estimation, and incorrect recovery of shadow areas caused by distribution shifts in the target environment in existing pre-trained models.

[0087] On the other hand, this application freezes the pre-trained backbone network parameters and introduces only a small number of trainable parameters for target environment adaptation through a multi-scale propagation perceptual nonlinear low-rank adaptation branch, eliminating the need for full parameter retraining or fine-tuning of the entire model. Compared to existing retraining methods that require a large number of target environment samples and computational resources, this application significantly reduces the number of training parameters, computational overhead, and storage requirements, and can still efficiently complete model updates even in actual spectral mapping scenarios where only a small number of sparse samples can be collected in the target environment. Simultaneously, because the backbone parameters remain frozen, the model can better retain the general propagation knowledge learned in the source environment, reducing the risk of catastrophic forgetting and achieving low-cost, high-efficiency target environment adaptation.

[0088] On the other hand, this application introduces multi-scale parallel branches and a propagation-aware nonlinear function. The multi-scale parallel branches, through identity mapping and deep convolutional branches with different dilation rates, can simultaneously capture near-field local details, mid-range diffraction effects, and long-range path loss trends, enabling the model to adaptively select appropriate scale combinations based on different spatial locations and propagation states. The propagation-aware nonlinear function performs sign-preserving logarithmic compression on large-amplitude feature responses, mitigating the excessive influence of extreme responses on the low-rank update direction. Compared to the single linear low-rank correction of standard LoRA, this application better conforms to the characteristics of radio propagation. With only a small increase in trainable parameters, the model simultaneously adapts to near-range details, mid-range diffraction, and long-range large-scale attenuation patterns, overcoming the deficiency of standard parameter-efficient fine-tuning methods that lack propagation-aware capabilities.

[0089] On the other hand, under ultra-sparse supervision conditions where the target environment cannot provide complete spectral situation map labels and only a very small number of sparse sampling points are available, effective structural and smoothing constraints are provided for unlabeled regions. Specifically, freezing the gradient distillation provided by the teacher model allows the student model to inherit the spatial structure and variation trends of the pre-trained model in the source environment in unlabeled regions; total variation regularization loss suppresses high-frequency noise and unreasonable abrupt changes in open propagation regions; and sparse supervision loss constrains labeled sampling points. The synergistic effect of these three factors enables the model to obtain more reliable complete spectral situation map estimation results under ultra-sparse supervision conditions, avoiding non-physical oscillations in unobserved regions caused by the model only fitting a small number of measurement points. This solves the problem of unreliable model updates under ultra-sparse sampling conditions in existing technologies.

[0090] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0091] Please see Figure 4 This illustration shows a schematic diagram of a sparse sampling-based spectrum situation awareness model update device provided in an exemplary embodiment of this application. This sparse sampling-based spectrum situation awareness model update device can be implemented as all or part of a device through software, hardware, or a combination of both. The device 1 includes a raw two-channel input generation module 10, an enhanced feature map generation module 20, a feature stitching module 30, a feature input module 40, a branch processing module 50, a loss function construction module 60, and a model generation module 70.

[0092] The original two-channel input generation module 10 is used to acquire the building layout map and sparse received signal strength sampling map of the target environment to form the original two-channel input; The enhanced feature map generation module 20 is used to construct a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field based on the original two-channel input through the enhanced physical propagation coding module, thereby obtaining four enhanced feature maps. The feature stitching module 30 is used to stitch the original two-channel input with four enhanced feature maps to form a six-channel input. The feature input module 40 is used to feed the six-channel input into the model to be fine-tuned; the backbone network parameters of the model to be fine-tuned are frozen, and a multi-scale propagation sensing nonlinear low-rank adaptation branch is set in the bypass of the backbone network. The branch processing module 50 is used to process the six-channel input through the preceding layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network. The input of the multi-scale propagation perceptual nonlinear low-rank adaptation branch is the input features. The target environment adaptation increment is generated through the multi-scale propagation perceptual nonlinear low-rank adaptation branch and added to the output of the layer of the backbone network to obtain the output feature map. The loss function construction module 60 is used to construct the corresponding loss function based on the supervision conditions of the target environment; The model generation module 70 is used to freeze the parameters of the backbone network; based on the output feature map and loss function, the parameters of the multi-scale propagation perception nonlinear low-rank adaptation branch are updated to obtain a spectrum situational awareness model adapted to the target environment.

[0093] It should be noted that the sparse sampling-based spectrum situation awareness model update device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the sparse sampling-based spectrum situation awareness model update method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the sparse sampling-based spectrum situation awareness model update device and the sparse sampling-based spectrum situation awareness model update method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0094] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0095] In this embodiment, on the one hand, by introducing an enhanced physical propagation coding module at the input end, four enhanced feature maps are constructed in addition to the original building layout map and sparse sampling map: a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field. This provides the model with explicit spatial priors related to radio propagation physics and sampling layout. Compared with the existing technology that relies solely on sparse sampling points and building maps as input, this application enables the model to directly utilize information such as distance attenuation trends, measurement extrapolation trends, local occlusion backgrounds, and sampling coverage when fine-tuning in the target environment. It does not require relearning propagation laws entirely from a small number of samples, thereby significantly improving the estimation accuracy of the spectral situation map in the target environment. This effectively overcomes the problems of hotspot location shifts, insufficient signal peak estimation, and incorrect recovery of shadow areas caused by distribution shifts in the target environment in existing pre-trained models.

[0096] On the other hand, this application freezes the pre-trained backbone network parameters and introduces only a small number of trainable parameters for target environment adaptation through a multi-scale propagation perceptual nonlinear low-rank adaptation branch, eliminating the need for full parameter retraining or fine-tuning of the entire model. Compared to existing retraining methods that require a large number of target environment samples and computational resources, this application significantly reduces the number of training parameters, computational overhead, and storage requirements, and can still efficiently complete model updates even in actual spectral mapping scenarios where only a small number of sparse samples can be collected in the target environment. Simultaneously, because the backbone parameters remain frozen, the model can better retain the general propagation knowledge learned in the source environment, reducing the risk of catastrophic forgetting and achieving low-cost, high-efficiency target environment adaptation.

[0097] On the other hand, this application introduces multi-scale parallel branches and a propagation-aware nonlinear function. The multi-scale parallel branches, through identity mapping and deep convolutional branches with different dilation rates, can simultaneously capture near-field local details, mid-range diffraction effects, and long-range path loss trends, enabling the model to adaptively select appropriate scale combinations based on different spatial locations and propagation states. The propagation-aware nonlinear function performs sign-preserving logarithmic compression on large-amplitude feature responses, mitigating the excessive influence of extreme responses on the low-rank update direction. Compared to the single linear low-rank correction of standard LoRA, this application better conforms to the characteristics of radio propagation. With only a small increase in trainable parameters, the model simultaneously adapts to near-range details, mid-range diffraction, and long-range large-scale attenuation patterns, overcoming the deficiency of standard parameter-efficient fine-tuning methods that lack propagation-aware capabilities.

[0098] On the other hand, under ultra-sparse supervision conditions where the target environment cannot provide complete spectral situation map labels and only a very small number of sparse sampling points are available, effective structural and smoothing constraints are provided for unlabeled regions. Specifically, freezing the gradient distillation provided by the teacher model allows the student model to inherit the spatial structure and variation trends of the pre-trained model in the source environment in unlabeled regions; total variation regularization loss suppresses high-frequency noise and unreasonable abrupt changes in open propagation regions; and sparse supervision loss constrains labeled sampling points. The synergistic effect of these three factors enables the model to obtain more reliable complete spectral situation map estimation results under ultra-sparse supervision conditions, avoiding non-physical oscillations in unobserved regions caused by the model only fitting a small number of measurement points. This solves the problem of unreliable model updates under ultra-sparse sampling conditions in existing technologies.

[0099] This application also provides a computer-readable medium storing program instructions thereon, which, when executed by a processor, implement the sparse sampling-based spectrum situation awareness model update method provided in the above-described method embodiments.

[0100] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the sparse sampling-based spectrum situation awareness model update method of the above-described method embodiments.

[0101] Please see Figure 5 This is a schematic diagram of the structure of a device provided in an embodiment of this application. Figure 5 As shown, device 1000 may include: at least one processor 1001, at least one network interface 1004, user interface 1003, memory 1005, and at least one communication bus 1002.

[0102] The communication bus 1002 is used to realize the connection and communication between these components.

[0103] The user interface 1003 may include a display screen and a camera. Optionally, the user interface 1003 may also include a standard wired interface and a wireless interface.

[0104] The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0105] The processor 1001 may include one or more processing cores. The processor 1001 connects to various parts within the device 1000 using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1005, and by calling data stored in the memory 1005. Optionally, the processor 1001 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 1001 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 1001 and may be implemented as a separate chip.

[0106] The memory 1005 may include random access memory (RAM) or read-only memory. Optionally, the memory 1005 may include a non-transitory computer-readable storage medium. The memory 1005 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 1005 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 5 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for updating a spectrum situation awareness model based on sparse sampling.

[0107] exist Figure 5 In the device 1000 shown, the user interface 1003 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 1001 can be used to call the application program for updating the sparse sampling-based spectrum situational awareness model stored in the memory 1005, and specifically perform the following operations: The building layout map and sparse received signal strength sampling map of the target environment are obtained to form the original two-channel input; By enhancing the physical propagation coding module, four enhanced feature maps are obtained by constructing a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field based on the original two-channel input. The original two-channel input is concatenated with four enhanced feature maps to form a six-channel input. The six-channel input is fed into the model to be fine-tuned; the backbone network parameters of the model to be fine-tuned are frozen, and a multi-scale propagation sensing nonlinear low-rank adaptation branch is set in the bypass of the backbone network. The six-channel input is processed by the pre-sequence layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network. The input of the multi-scale propagation perceptual nonlinear low-rank adaptation branch is the input features. The target environment adaptation increment is generated by the multi-scale propagation perceptual nonlinear low-rank adaptation branch and added to the output of the layer of the backbone network to obtain the output feature map. Construct the corresponding loss function based on the supervision conditions of the target environment; Freeze the parameters of the backbone network; based on the output feature map and loss function, update the parameters of the multi-scale propagation perception nonlinear low-rank adaptation branch to obtain a spectrum situational awareness model adapted to the target environment.

[0108] In one embodiment, when processor 1001 constructs a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field based on the original two-channel input to obtain four enhanced feature maps, it specifically performs the following operations: For each grid location, calculate its Euclidean distance to the nearest measurement point; Construct the original path loss value based on the Euclidean distance; The original path loss values ​​are normalized and reversed to obtain the logarithmic distance path loss field. For each grid location, an inverse distance weighted interpolation field is constructed based on the measurements of all measurement points and the inverse distance weights; The building density field is obtained by performing local average pooling on the building layout map. For each grid location, an exponentially decaying propagation confidence field is constructed based on its average distance to the nearest multiple measurement points.

[0109] In one embodiment, when the processor 1001 constructs the model to be fine-tuned, it specifically performs the following operations: Establish a backbone network; A multi-scale propagation sensing nonlinear low-rank adaptation branch is set as a bypass in at least one layer of the backbone network to obtain the model to be fine-tuned. The input of the multi-scale propagation sensing nonlinear low-rank adaptation branch is the input feature of the backbone network layer, and the output of the branch is added to the output of the backbone network layer to obtain the final output of the layer.

[0110] In one embodiment, when the processor 1001 processes the six-channel input through the pre-processing layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network, and the input of the multi-scale propagation perceptual nonlinear low-rank adaptation branch is the input features; and generates the target environment adaptation increment through the multi-scale propagation perceptual nonlinear low-rank adaptation branch and adds it to the output of the backbone network layer to obtain the output feature map, the processor 1001 specifically performs the following operations: In the multi-scale propagation perception nonlinear low-rank adaptation branch, the input features are sequentially subjected to low-rank dimensionality reduction, multi-scale parallel branch processing, and learnable gated weighted fusion to obtain multi-scale fused features. A propagation-aware nonlinear function is applied to the multi-scale fused features. This propagation-aware nonlinear function is used to perform logarithmic compression on large-amplitude feature responses while preserving the positive and negative signs. The features processed by the propagation sensing nonlinear function are subjected to low-rank dimensionality upsizing to obtain the target environment adaptation increment. The target environment adaptation increment is multiplied by a scaling factor and then added to the output of the corresponding layer of the backbone network to obtain the final output feature map of that layer.

[0111] In one embodiment, when the processor 1001 constructs the corresponding loss function based on the supervision conditions of the target environment, it specifically performs the following operations: When the target environment has complete spectral situation map labels, the dense reconstruction loss is calculated in non-building areas under fully supervised conditions; or... When the target environment only has sparse received signal strength sampling points, an ultra-sparse supervision condition is adopted to divide the sparse received signal strength sampling points into a set of known points and a set of query points, and a frozen teacher model is introduced; wherein, the teacher model is a frozen copy of the pre-trained model of the source environment, and its input is the original two-channel input; Construct a joint loss function, which includes: a sparse supervision loss computed only on the query point set, a knowledge distillation loss used to constrain the spatial structure of the model to be fine-tuned to be consistent with the teacher model, and a total variation regularization loss used to constrain the continuity of the output space of the model to be fine-tuned.

[0112] In one embodiment, when the processor 1001 executes the parameter update of the multi-scale propagation perception nonlinear low-rank adaptation branch to obtain a spectral situational awareness model adapted to the target environment, it specifically performs the following operations: Under fully supervised conditions, dense reconstruction loss is calculated based on the output feature map and the label of the complete spectral situation map, and dense reconstruction loss is used as the optimization objective. Under ultra-sparse supervision, the joint loss function is calculated based on the output feature map, the output of the frozen teacher model, and the set of query points in the sparse received signal strength sampling points, and the joint loss function is used as the optimization objective. During the optimization process, the parameters of the backbone network are kept frozen, and only the parameters of the multi-scale propagation sensing nonlinear low-rank adaptation branch and the newly added trainable parameters related to the six-channel input expansion are updated. After optimization, a spectrum situational awareness model adapted to the target environment is obtained.

[0113] In one embodiment, after the processor 1001 executes the parameter update of the multi-scale propagation perception nonlinear low-rank adaptation branch to obtain the spectral situational awareness model adapted to the target environment, it also performs the following operations: Obtain a new sparse received signal strength sampling map of the target environment; Based on the new sparse received signal strength sampling map and building layout map, the corresponding logarithmic distance path loss field, inverse distance weighted interpolation field, building density field and propagation confidence field are constructed through the enhanced physical propagation coding module. The building layout map, the new sparse received signal strength sampling map and the four enhanced feature maps are then spliced ​​together to form a new six-channel input. The new six-channel input is fed into a spectrum situation awareness model adapted to the target environment, and a complete spectrum situation map under the target environment is output.

[0114] In this embodiment, on the one hand, by introducing an enhanced physical propagation coding module at the input end, four enhanced feature maps are constructed in addition to the original building layout map and sparse sampling map: a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field. This provides the model with explicit spatial priors related to radio propagation physics and sampling layout. Compared with the existing technology that relies solely on sparse sampling points and building maps as input, this application enables the model to directly utilize information such as distance attenuation trends, measurement extrapolation trends, local occlusion backgrounds, and sampling coverage when fine-tuning in the target environment. It does not require relearning propagation laws entirely from a small number of samples, thereby significantly improving the estimation accuracy of the spectral situation map in the target environment. This effectively overcomes the problems of hotspot location shifts, insufficient signal peak estimation, and incorrect recovery of shadow areas caused by distribution shifts in the target environment in existing pre-trained models.

[0115] On the other hand, this application freezes the pre-trained backbone network parameters and introduces only a small number of trainable parameters for target environment adaptation through a multi-scale propagation perceptual nonlinear low-rank adaptation branch, eliminating the need for full parameter retraining or fine-tuning of the entire model. Compared to existing retraining methods that require a large number of target environment samples and computational resources, this application significantly reduces the number of training parameters, computational overhead, and storage requirements, and can still efficiently complete model updates even in actual spectral mapping scenarios where only a small number of sparse samples can be collected in the target environment. Simultaneously, because the backbone parameters remain frozen, the model can better retain the general propagation knowledge learned in the source environment, reducing the risk of catastrophic forgetting and achieving low-cost, high-efficiency target environment adaptation.

[0116] On the other hand, this application introduces multi-scale parallel branches and a propagation-aware nonlinear function. The multi-scale parallel branches, through identity mapping and deep convolutional branches with different dilation rates, can simultaneously capture near-field local details, mid-range diffraction effects, and long-range path loss trends, enabling the model to adaptively select appropriate scale combinations based on different spatial locations and propagation states. The propagation-aware nonlinear function performs sign-preserving logarithmic compression on large-amplitude feature responses, mitigating the excessive influence of extreme responses on the low-rank update direction. Compared to the single linear low-rank correction of standard LoRA, this application better conforms to the characteristics of radio propagation. With only a small increase in trainable parameters, the model simultaneously adapts to near-range details, mid-range diffraction, and long-range large-scale attenuation patterns, overcoming the deficiency of standard parameter-efficient fine-tuning methods that lack propagation-aware capabilities.

[0117] On the other hand, under ultra-sparse supervision conditions where the target environment cannot provide complete spectral situation map labels and only a very small number of sparse sampling points are available, effective structural and smoothing constraints are provided for unlabeled regions. Specifically, freezing the gradient distillation provided by the teacher model allows the student model to inherit the spatial structure and variation trends of the pre-trained model in the source environment in unlabeled regions; total variation regularization loss suppresses high-frequency noise and unreasonable abrupt changes in open propagation regions; and sparse supervision loss constrains labeled sampling points. The synergistic effect of these three factors enables the model to obtain more reliable complete spectral situation map estimation results under ultra-sparse supervision conditions, avoiding non-physical oscillations in unobserved regions caused by the model only fitting a small number of measurement points. This solves the problem of unreliable model updates under ultra-sparse sampling conditions in existing technologies.

[0118] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program for updating the sparse sampling-based spectrum situation awareness model can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium for storing the program for updating the sparse sampling-based spectrum situation awareness model can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0119] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A method for updating a spectrum situation awareness model based on sparse sampling, characterized in that, The method includes: The building layout map and sparse received signal strength sampling map of the target environment are obtained to form the original two-channel input; By enhancing the physical propagation coding module, four enhanced feature maps are obtained by constructing a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field based on the original two-channel input. The original two-channel input is concatenated with the four enhanced feature maps to form a six-channel input. The six-channel input is fed into the model to be fine-tuned; the backbone network parameters of the model to be fine-tuned are frozen, and a multi-scale propagation sensing nonlinear low-rank adaptation branch is set in the bypass of the backbone network. The six-channel input is processed by the pre-sequence layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network. The input of the multi-scale propagation perceptual nonlinear low-rank adaptation branch is the input features. The target environment adaptation increment is generated by the multi-scale propagation perceptual nonlinear low-rank adaptation branch and added to the output of the layer of the backbone network to obtain the output feature map. Construct the corresponding loss function based on the supervision conditions of the target environment; Freeze the parameters of the backbone network; based on the output feature map and the loss function, update the parameters of the multi-scale propagation perception nonlinear low-rank adaptation branch to obtain a spectrum situational awareness model adapted to the target environment.

2. The method according to claim 1, characterized in that, The building layout map and the sparse received signal strength sampling map correspond to multiple grid positions after the target area is discretized, and the non-zero positions in the sparse received signal strength sampling map correspond to measurement points; Based on the original two-channel input, a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field are constructed to obtain four enhanced feature maps, including: For each grid location, calculate its Euclidean distance to the nearest measurement point; Construct the original path loss value based on the Euclidean distance; The original path loss value is normalized and reversed to obtain the logarithmic distance path loss field; For each grid location, an inverse distance weighted interpolation field is constructed based on the measurements of all measurement points and the inverse distance weights; The building layout map is subjected to local average pooling to obtain the building density field; For each grid location, an exponentially decaying propagation confidence field is constructed based on its average distance to the nearest multiple measurement points.

3. The method according to claim 1, characterized in that, The model to be fine-tuned is constructed according to the following steps: Establish a backbone network; A multi-scale propagation sensing nonlinear low-rank adaptation branch is set as a bypass in at least one layer of the backbone network to obtain the model to be fine-tuned; wherein, the input of the multi-scale propagation sensing nonlinear low-rank adaptation branch is the input feature of the layer of the backbone network, and the output of the branch is added to the output of the layer of the backbone network to obtain the final output of the layer.

4. The method according to claim 3, characterized in that, The multi-scale propagation sensing nonlinear low-rank adaptation branch includes: Low-rank dimensionality reduction mapping is used to compress the input features of this layer of the backbone network into a low-dimensional feature space to obtain low-rank features. A multi-scale parallel branch is set on the low-dimensional feature space, including an identity mapping branch and at least two deep convolution branches with different dilation rates; wherein the dilation rates of the at least two deep convolution branches are 2 and 4, respectively; the identity mapping branch is used to preserve the original low-rank features and near-field details; the deep convolution branch with a dilation rate of 2 is used to capture medium-range propagation effects; and the deep convolution branch with a dilation rate of 4 is used to capture larger-range spatial decay trends and long-distance propagation associations. A learnable gating unit is used to generate normalized weights for three branches, and to perform a weighted summation of the outputs of the identity mapping branch and at least two deep convolutional branches with different dilation rates based on the normalized weights to obtain multi-scale fusion features.

5. The method according to claim 1, characterized in that, The target environment adaptation increment is generated through the multi-scale propagation perception nonlinear low-rank adaptation branch, and added to the output of that layer of the backbone network to obtain the output feature map, including: In the multi-scale propagation sensing nonlinear low-rank adaptation branch, the input features are sequentially subjected to low-rank dimensionality reduction, multi-scale parallel branch processing, and learnable gated weighted fusion to obtain multi-scale fused features. A propagation-aware nonlinear function is applied to the multi-scale fused features, which is used to perform logarithmic compression on large-amplitude feature responses while preserving the positive and negative signs. The features processed by the propagation sensing nonlinear function are subjected to low-rank dimensionality upsizing to obtain the target environment adaptation increment. The target environment adaptation increment is multiplied by a scaling factor and then added to the output of the corresponding layer of the backbone network to obtain the final output feature map of that layer.

6. The method according to claim 1, characterized in that, The step of constructing the corresponding loss function based on the supervision conditions of the target environment includes: When the target environment has complete spectral situation map labels, the dense reconstruction loss is calculated in non-building areas under fully supervised conditions; or... When the target environment only has sparse received signal strength sampling points, an ultra-sparse supervision condition is adopted to divide the sparse received signal strength sampling points into a set of known points and a set of query points, and a frozen teacher model is introduced; wherein, the teacher model is a frozen copy of the source environment pre-trained model, and its input is the original two-channel input; Construct a joint loss function, which includes: a sparse supervision loss computed only on the query point set, a knowledge distillation loss used to constrain the spatial structure of the model to be fine-tuned to be consistent with that of the teacher model, and a total variation regularization loss used to constrain the output space continuity of the model to be fine-tuned.

7. The method according to claim 6, characterized in that, Update the parameters of the multi-scale propagation sensing nonlinear low-rank adaptation branch to obtain a spectrum situational awareness model adapted to the target environment, including: Under fully supervised conditions, dense reconstruction loss is calculated based on the output feature map and the complete spectral situation map label, and the dense reconstruction loss is used as the optimization objective. Under ultra-sparse supervision, a joint loss function is calculated based on the output feature map, the output of the frozen teacher model, and the set of query points in the sparse received signal strength sampling points, and the joint loss function is used as the optimization objective. During the optimization process, the parameters of the backbone network are kept frozen, and only the parameters of the multi-scale propagation sensing nonlinear low-rank adaptation branch and the newly added trainable parameters related to the six-channel input expansion are updated. After optimization, a spectrum situational awareness model adapted to the target environment is obtained.

8. The method according to claim 1, characterized in that, After updating the parameters of the multi-scale propagation sensing nonlinear low-rank adaptation branch to obtain the spectral situational awareness model adapted to the target environment, the method further includes: Obtain a new sparse received signal strength sampling map of the target environment; Based on the new sparse received signal strength sampling map and the building layout map, the corresponding logarithmic distance path loss field, inverse distance weighted interpolation field, building density field and propagation confidence field are constructed through the enhanced physical propagation coding module. The building layout map, the new sparse received signal strength sampling map and the four enhanced feature maps are then spliced ​​together to form a new six-channel input. The new six-channel input is fed into the spectrum situation awareness model adapted to the target environment, and a complete spectrum situation map under the target environment is output.

9. The method according to claim 6, characterized in that, The knowledge distillation loss is a gradient distillation loss, used to constrain the gradient distribution of the model to be fine-tuned to be consistent with that of the teacher model in the output space.

10. A spectral situation awareness model update device based on sparse sampling, characterized in that, The device includes: The original two-channel input generation module is used to obtain the building layout map and sparse received signal strength sampling map of the target environment, which constitute the original two-channel input; The enhanced feature map generation module is used to construct a logarithmic distance path loss field, an inverse distance weighted interpolation field, a building density field, and a propagation confidence field based on the original two-channel input through the enhanced physical propagation coding module, thereby obtaining four enhanced feature maps. The feature stitching module is used to stitch the original two-channel input with the four enhanced feature maps to form a six-channel input; The feature input module is used to send the six-channel input to the model to be fine-tuned; the backbone network parameters of the model to be fine-tuned are frozen, and a multi-scale propagation sensing nonlinear low-rank adaptation branch is set in the bypass of the backbone network. The branch processing module is used to process the six-channel input through the preceding layer of the backbone network to obtain the input features of the current layer to be adapted in the backbone network. The input of the multi-scale propagation perceptual nonlinear low-rank adaptation branch is the input features. The target environment adaptation increment is generated through the multi-scale propagation perceptual nonlinear low-rank adaptation branch and added to the output of the layer of the backbone network to obtain the output feature map. The loss function construction module is used to construct the corresponding loss function based on the supervision conditions of the target environment; The model generation module is used to freeze the parameters of the backbone network; based on the output feature map and the loss function, it updates the parameters of the multi-scale propagation perception nonlinear low-rank adaptation branch to obtain a spectrum situational awareness model adapted to the target environment.

Images