Joint spectrum map construction and signal source positioning method driven by data semantics

Abstract

The application discloses a data semantic double-driven joint spectrum map construction and signal source positioning method, comprising: collecting spectrum data and preprocessing; establishing a three-dimensional spectrum map representation model of joint frequency-space domain; creating a binary city map and a binary sampling location map; training set training network; calculating total loss; updating weight parameters; judging whether the current iteration number reaches the set maximum number; inputting the verification set data into the neural network model for spectrum map construction and signal source positioning; and obtaining the complete spectrum map and the grid position of the signal source in the target area after processing the output result. The binary city map and the binary sampling location map are introduced as semantic knowledge to extract spatial information affecting signal propagation, a multi-task learning joint training framework sharing the bottom layer is designed, spectrum map construction and signal source positioning are simultaneously realized, and the spectrum map construction and signal source positioning precision in the complex urban environment are significantly improved.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and in particular relates to a data and semantic dual-driven joint spectrum map construction and signal source positioning method. Background Art

[0002] With the rapid development of wireless communication technology and the continuous increase in the number of wireless devices in urban environments, signal source positioning has become increasingly important in modern cities. Complex electromagnetic environments, high-density building obstructions, and multipath effects often result in a lack of line-of-sight between user equipment (UE) and Global Positioning System (GPS) satellites, thus affecting positioning effectiveness, especially in dense urban areas. To address the above challenges, positioning methods based on spectrum maps have gradually become a research hotspot. Spectrum map construction is an effective technology that can characterize the distribution and utilization status of spectrum resources at different frequencies in an area of ​​interest. By constructing a high-precision spectrum map, the propagation characteristics of electromagnetic signals can be more intuitively reflected, thereby improving the accuracy of signal source positioning.

[0003] In recent years, deep learning has promoted the development of spectrum map construction and signal source positioning by learning the propagation rules of signals in an end-to-end manner directly from the original spectrum data. Wang X, Wang X, Mao S et al. proposed an indoor positioning method based on spectrum map in their paper "Indoor Radio Map Construction and Localization With Deep Gaussian Processes" (IEEE Internet Things J., vol. 7, no. 11, pp. 11238-11249, 2020). This method first uses Deep Map to generate an accurate and complete spectrum map, and then achieves positioning through deep Gaussian process. Experiments have shown that the positioning accuracy of this method in indoor scenes is significantly better than that of traditional methods. However, due to the wide range of urban environments and limited sampling data, the positioning accuracy of this method in large-scale outdoor scenes is limited. To this end, Yapar C, Levie R, Kutyniok G, et al. designed a positioning network LocUNet specifically for urban environments in their paper "Real-Time Outdoor Localization Using Radio Maps: A Deep Learning Approach" (IEEE Trans. Wireless Commun., vol. 22, no. 12, pp. 9703-9717, 2023). By using the estimated complete spectrum map and the RSS measurement value of the user to be located, the method can achieve the positioning of users in dense urban scenes. However, this method relies heavily on the accuracy of the spectrum map construction. If the spectrum map has large errors, the target position cannot be accurately located.

[0004] Patent application number CN202310813202.7 proposes a method for jointly solving radiation source location and spectrum mapping based on terrain. This method optimizes radiation source locations and propagation model parameters through joint iterations, obtaining the final result after iterative convergence. Although this method achieves the joint optimization of spectrum mapping and radiation source location, it fails to effectively cope with dynamically changing environments, is computationally complex, and suffers from computational resource limitations.

[0005] In summary, most deep learning methods follow a "build first, then localize" strategy, focusing on high-precision spectrum mapping for sparse data. Consequently, localization performance is highly dependent on the accuracy of spectrum map construction. Furthermore, existing methods often cannot effectively cope with dynamically changing environments and the need for real-time updates. Furthermore, with limited computing resources, the spectrum map construction and localization processes can be computationally complex. Therefore, the development of new spectrum map construction and emitter localization methods is urgently needed. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned existing technologies and provide a data-semantics dual-driven joint spectrum map construction and signal source localization method. Based on a joint training framework, spectrum map construction and signal source localization are simultaneously realized, and city maps and sampling location maps are introduced as semantic knowledge. This not only significantly improves the construction accuracy of spectrum maps in complex urban environments and reduces the positioning error of signal sources, but also maintains a low network complexity, and has broad application potential in actual communication scenarios.

[0007] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0008] The data-semantics-driven joint spectrum map construction and signal source localization method includes:

[0009] Step 1: Multiple sensing nodes randomly deployed in the target area collect spectrum data and perform data normalization preprocessing;

[0010] Step 2: spatially discretize the target area and establish a three-dimensional spectrum map representation model in the frequency-space domain as spectrum data;

[0011] Step 3: Create a binary city map and a binary sampling location map as semantic knowledge;

[0012] Step 4: Using the training set consisting of semantic knowledge and spectrum data to train the neural network model for spectrum map construction and signal source location;

[0013] Step 5: Calculate the loss of the spectrum map construction task and the signal source localization task respectively, and sum them up by weight parameters to obtain the total loss of the neural network model;

[0014] Step 6: Use a dynamic adjustment mechanism to update the weight parameters according to the weight adjustment factor and loss threshold;

[0015] Step 7: Determine whether the current number of iterations has reached the set maximum number of iterations. If so, proceed to step 8. If not, increase the number of iterations by one and return to step 4 to continue training.

[0016] Step 8: Input the validation set data into the neural network model used for spectrum map construction and signal source location to obtain the output result;

[0017] Step 9: Process the output result to obtain a complete spectrum map and the grid position of the signal source in the target area.

[0018] To optimize the above technical solutions, specific measures taken also include:

[0019] Step 2 above includes:

[0020] The target area is spatially discretized and a two-dimensional spectrum map completion model is established. This model completes the grid values ​​in the target area after spatial discretization based on the collected data and unmonitored frequencies to form a corresponding two-dimensional spectrum map, thereby obtaining a three-dimensional spectrum map representation model of the target area in the frequency-space domain. The specific details are as follows:

[0021] (1) If there are m sensing nodes in the grid I = (i, j), the grid is assigned the normalized average value of the spectrum data collected by these m sensing nodes. Otherwise, the grid is assigned 0, and the frequency f that can be monitored in the target area is obtained. k Incomplete spectrum S on k ;

[0022] (2) setting all the values ​​of the spectrum map grids on the undetected monitoring frequency f0 to 0, forming a blank two-dimensional spectrum map E0 on the undetected frequency f0, which serves as the target spectrum map to be inferred;

[0023] (3) The frequency f that can be monitored in the target area k Incomplete spectrum S on k The blank two-dimensional spectrum map E0 on the unmonitored frequency f0 is stacked in ascending order according to the frequency dimension to obtain a frequency-space domain joint three-dimensional spectrum map representation model of the target area.

[0024] Step 3 above creates a binary city map as follows:

[0025]

[0026] where Z i,j represents the value of the binary city map at the grid I = (i, j) in the target area after spatial discretization;

[0027] Create a binary sample location map as follows:

[0028]

[0029] Among them, M i,j Represents the value of the binary sampling position map at grid I = (i, j) in the target area after spatial discretization.

[0030] The above step 4 is as follows:

[0031] (1) Feature fusion of semantic knowledge and spectral data to obtain data semantic fusion features;

[0032] (2) The data semantic fusion features are input into the spectrum map construction network DSD-UNet and the signal source localization network DSD-ResNet of the neural network model respectively. DSD-UNet completes the missing spectrum map to reflect the distribution of signal strength in the entire area; the features of the frequency-space domain combined three-dimensional spectrum map representation model extracted by the spectrum map construction network DSD-UNet are input into the signal source localization network DSD-ResNet. DSD-ResNet locates the signal source to grasp the precise position of different signal sources in the area.

[0033] The above (1) is specifically as follows: the frequency-space domain joint three-dimensional spectrum map representation model, binary city map, and binary sampling location map are divided into three channels and input into the shared layer of the neural network model, and the underlying features are extracted through the shared layer to realize the feature fusion of semantic knowledge and spectrum data.

[0034] The loss calculation formula for the spectrum map construction task described in step 5 above is as follows:

[0035]

[0036] Where K represents the total number of frequency points; N×N represents the total number of spectrum map grids; E k,i,j Indicates frequency f k The received signal strength at each grid point on the spectrum map; P k,i,j Indicates frequency f k The received signal strength of the corresponding grid point on the real spectrum map; represents the loss of the spectrum map construction task.

[0037] The calculation formula for the loss of the signal source localization task described in step 5 above is as follows:

[0038]

[0039] Among them, p t Indicates the confidence of the model prediction; α t represents the sample weight hyperparameter; γ represents the focusing parameter; represents the loss of the signal source localization task.

[0040] The total loss of the neural network model described in step 5 above is:

[0041]

[0042] in, Represents the total loss of the neural network model; represents the weight parameter of the spectrum map construction task, Represents the weight parameter of the signal source localization task.

[0043] The updating formula of the weight parameter described in step 6 above is as follows:

[0044]

[0045] in, Represents the weight parameter of the spectrum map construction task; represents the updated spectrum map construction task weight parameter; Represents the weight parameter of the signal source localization task; represents the updated weight parameter of the signal source localization task; β represents the weight adjustment factor; represents the loss of the spectrum map construction task; represents the loss of the signal source localization task; loss threshold for the spectrum map construction task; Loss threshold for the signal source localization task.

[0046] In step 9 above, the data output by the spectrum map construction network of the neural network model is denormalized to obtain a complete spectrum map:

[0047] P=(P max -P min )·P out +P min

[0048] Among them, P represents the spectrum data after denormalization, P min represents the minimum value of the spectrum data output by the spectrum map construction network, P max represents the maximum value of the spectrum data output by the spectrum map construction network, P out Represents the spectrum data output by the spectrum map construction network;

[0049] The signal source position mask output by the signal source localization network of the neural network model is processed to extract the area marked as the signal source in the mask. The coordinates corresponding to the area are the grid positions of the signal source in the target area.

[0050] The present invention has the following beneficial effects:

[0051] First, to address the problems of low spectrum map construction accuracy and large signal source positioning error caused by limited data in urban environments, the present invention introduces a binary city map and a binary sampling position map as semantic knowledge, and proposes a data-semantics dual-driven method to extract spatial information that affects signal propagation, which is beneficial to the neural network reasoning of the intrinsic mechanism of spatial domain signal propagation.

[0052] Second, the present invention designs a neural network framework for joint spectrum map construction and signal source positioning driven by both data and semantics. Based on the multi-task learning idea of ​​shared underlying layers, the spectrum map construction and signal source positioning tasks are collaboratively optimized, realizing information sharing and mutual promotion between different tasks. By jointly training and collaboratively optimizing the two tasks, the accuracy of spectrum map construction in complex urban environments is significantly improved, and the positioning error of signal sources is reduced, especially in low sampling density and multiple signal source scenarios.

[0053] Third, in the network framework designed by the present invention, three-dimensional convolution is used to construct a frequency-space domain joint three-dimensional spectrum map representation model, so that the network can achieve complete spectrum map construction without sampling all frequencies of the spectrum data, thereby improving the integrity of the spectrum map construction.

[0054] Fourth, the present invention uses MSE and FocalLoss as loss functions for spectrum mapping and signal source localization, respectively, and achieves collaborative optimization of multi-task learning by dynamically adjusting weight parameters. Specifically, a dynamic adjustment mechanism updates weight parameters in real time based on weight adjustment factors and loss thresholds, ensuring that the model can quickly adapt to more difficult tasks during training while maintaining balanced optimization of the two tasks, thereby improving model training efficiency and performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of the method of the present invention;

[0056] Figure 2 It is a network structure framework diagram of the present invention;

[0057] Figure 3 This is a comparison chart of the spectrum map visualization effects constructed by the present invention and other methods;

[0058] Figure 4 This is a comparison chart of spectrum map construction errors in different scenarios between the present invention and other methods;

[0059] Figure 5 This is a comparison chart of signal source positioning errors between the present invention and other methods in different scenarios. Figure 5 (a) is a comparison chart of signal source positioning errors of the present invention and other methods at different sampling densities in a multi-signal source scenario. Figure 5 (b) is a comparison chart of the signal source positioning errors of the present invention and other methods at different sampling densities in a single signal source scenario. DETAILED DESCRIPTION

[0060] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0061] The data-semantics dual-driven joint spectrum map construction and signal source positioning method of the present invention addresses the problems of low spectrum map construction accuracy and large signal source positioning errors caused by limited data in urban environments. It innovatively introduces binary city maps and binary sampling position maps as semantic knowledge, and proposes a data-semantics dual-driven method to extract spatial information that affects signal propagation. On this basis, a data-semantics dual-driven joint spectrum map construction and signal source positioning framework is designed. Through joint training and collaborative optimization of the two tasks, the accuracy of spectrum map construction and signal source positioning in complex urban environments is significantly improved, especially in low sampling density and multiple signal source scenarios. Figure 1-Figure 2 The specific steps of the method of the present invention are described as follows.

[0062] Step 1: spectrum data acquisition and preprocessing.

[0063] Step 1: Select the target area And randomly deploy N R The number of sensing nodes is determined by the sampling density. The collected spectrum data is uploaded to the central database for preprocessing.

[0064] The second step is to perform maximum and minimum normalization on the collected spectrum data, and the spectrum data will be normalized to [0,1]. The spectrum data is normalized to:

[0065]

[0066] Among them, P represents the real spectrum data, P min Represents the minimum value of all spectrum data, P max represents the maximum value of all spectrum data, and P′ represents the normalized spectrum data.

[0067] Step 2: Establish a three-dimensional spectrum map representation model in the frequency-space domain.

[0068] The first step is to target the area Perform spatial discretization;

[0069] The second step is to build a two-dimensional spectrum map completion model;

[0070] If there are m sensing nodes in the grid I = (i, j), the received signal strength value of the grid is the average of the received signal strengths of these m sensing nodes. Otherwise, the received signal strength value at the grid is set to 0. That is, if there are m sensing nodes in the grid I = (i, j), the grid is assigned the average of the normalized spectrum data collected by these m sensing nodes, otherwise, the grid is assigned 0, and the frequency f that can be monitored in the target area is obtained. k Incomplete spectrum S on k .

[0071] The third step is to establish a three-dimensional spectrum map representation model that combines frequency and space domains.

[0072] In the embodiment of the present invention, due to the complex propagation characteristics of the urban environment and the performance limitations of the sampling receiver, some frequency bands may not be monitored, and the sensing node cannot obtain spectrum data on all frequencies. Therefore, the values ​​of the spectrum map on the unmonitored frequency f0 are all set to 0, forming a blank two-dimensional spectrum map, and the spectrum map on this frequency is the target spectrum map to be inferred. Then, the frequency f0 that can be monitored in the target area is k The incomplete spectrum S k The blank spectrum map E0 at the unmonitored frequency f0 is stacked in ascending order according to the frequency dimension. Thus, a three-dimensional spectrum map representation model combining frequency and space domains can be obtained.

[0073] Step 3: Create a binary city map and a binary sampling location map as semantic knowledge.

[0074] The first step is to create a binary city map according to the following rules

[0075]

[0076] Where I = (i, j) represents the position coordinates of the grid within the target area, Z i,j Represents the value of the binary city map at grid I = (i, j).

[0077] The second step is to create a binary sampling location map according to the following rules

[0078]

[0079] Where I = (i, j) represents the position coordinates of the grid in the target area, M i,j Represents the value of the binary sampling position map at grid I = (i, j).

[0080] Step 4: Input the training set data into the network model.

[0081] The first step is to extract data semantic fusion features;

[0082] The frequency-space joint three-dimensional spectrum map representation model, binary city map, and binary sampling location map are divided into three channels and input into the shared layer of the network. The underlying features are extracted through the shared layer to achieve the feature fusion of semantic knowledge and spectrum data.

[0083] In the second step, the data semantic fusion features are input into the spectrum map construction network and the signal source localization network respectively;

[0084] A shared underlying multi-task learning framework was designed. Spectrum mapping and signal source localization share the same underlying features, and are then split into two networks to perform the two distinct tasks of spectrum mapping and signal source localization. DSD-UNet completes the missing spectrum map to reflect the distribution of signal strength throughout the area; DSD-ResNet locates the signal source. The features of the frequency-space joint three-dimensional spectrum map representation model extracted by the spectrum mapping network DSD-UNet are fed into the signal source localization network DSD-ResNet. By inputting the features extracted by DSD-UNet, DSD-ResNet can leverage the signal propagation patterns learned from the spectrum mapping task to compensate for insufficient data and more accurately determine the locations of different signal sources within the area.

[0085] In the third step, some features of the spectrum map construction network are integrated into the signal source localization network to alleviate the impact of severe lack of spectrum data in low sampling density scenarios.

[0086] In the embodiments of the present invention, due to the limited amount of available spectrum data and severe data omissions, even with the help of semantic knowledge to assist neural network reasoning, it is still difficult to effectively address the problem of feature loss and resolution degradation caused by insufficient data, which in turn affects the precise positioning of signal sources. To this end, some features from the spectrum mapping network are integrated into the signal source localization network to mitigate the impact of severe spectrum data omissions in low sampling density scenarios.

[0087] Step 5: Calculate the network loss.

[0088] Spectrum map construction is a regression problem. The MSE is used to measure the difference between the completed spectrum map and the true spectrum map. The calculation formula is as follows:

[0089]

[0090] Where K represents the total number of frequency points; N×N represents the total number of grids in the spectrum map; K×N×N represents the total number of samples, including the number of all frequencies and their corresponding grid points; E k,i,j Indicates frequency f k The received signal strength at each grid point on the spectrum map; P k,i,j Indicates frequency f k The received signal strength of the corresponding grid point on the real spectrum map; represents the loss of the spectrum map construction task.

[0091] Since the signal source localization task outputs the signal source position mask, Focal Loss is used as the objective function, and the calculation formula is as follows:

[0092]

[0093] Among them, p t Indicates the confidence of the model prediction: for the signal source location, the probability of the prediction being positive is p; for the non-signal source location, the probability is 1-p; α t represents the sample weight hyperparameter, which is used to adjust the weight of positive and negative samples. γ represents the focus parameter, which is used to reduce the impact of negative samples on the loss; represents the loss of the signal source localization task.

[0094] After calculating the spectrum map construction error and signal source positioning error Finally, the total loss function of the network is designed as:

[0095]

[0096] in, represents the total loss of the network; represents the loss of the spectrum map construction task; represents the loss of the spectrum map construction task; weight parameter and Control the contribution of completion error and localization error to the total loss respectively.

[0097] When all batches of data in the training data are fully back-propagated, it is considered an epoch.

[0098] Step 6: Input the test set data into the network and update the weight parameters according to the test results.

[0099] In the overall loss of the network, the weight parameter and It is not fixed during the training process, but is adjusted through error feedback. Initially, the two tasks are assigned the same weight. In the subsequent training process, the threshold in the current window is obtained by recording the historical test errors of each task and calculating the average error within the sliding window. When the error of a task exceeds its threshold, it means that the task is currently difficult to optimize and the weight parameter of the task needs to be adjusted by multiplying it by the weight adjustment factor β. On the contrary, if the error of the task is lower than the threshold, the weight remains unchanged. The weight update method is as follows:

[0100]

[0101] in, Represents the weight parameter of the spectrum map construction task; represents the updated spectrum map construction task weight parameter; represents the loss of the spectrum map construction task; Represents the weight parameter of the signal source localization task; Represents the updated weight parameter of the signal source localization task; represents the loss of the signal source localization task; β represents the weight adjustment factor.

[0102] Through this dynamic adjustment mechanism, the model can automatically balance the losses of the two tasks of spectrum completion and radiation source localization during training, enabling the network to quickly adapt to more difficult tasks in the early stages of training, and maintain a balance between the two tasks in the later stages, thereby improving the efficiency and effectiveness of joint training.

[0103] Step 7: Determine whether the network training is complete.

[0104] Determine whether the current epoch has reached the set maximum epoch. If so, proceed to step 8. If not, increase the epoch by one and return to step 4 to continue training the network.

[0105] Step 8: Input the validation set data into the network.

[0106] Step 9: Process the output results to obtain a complete spectrum map and signal source location.

[0107] Denormalize the data output by the spectrum map construction network to obtain a complete spectrum map:

[0108] P=(P max -P min )·P out +P min

[0109] Among them, P represents the spectrum data after denormalization, P min represents the minimum value of the spectrum data output by the spectrum map construction network, P max represents the maximum value of the spectrum data output by the spectrum map construction network, P out Represents the spectrum data output by the spectrum map construction network.

[0110] The signal source location mask output by the signal source localization network is processed to extract the area marked as the signal source in the mask. The coordinates corresponding to this area are the grid position of the signal source within the target area.

[0111] The effects of the present invention will be further described below in conjunction with simulation experiments.

[0112] 1. Simulation conditions and parameter settings:

[0113] The simulation experiments of the present invention were carried out on a simulation platform of Python 3.10 and Pytorch 1.12.1.

[0114] The computer CPU model is Intel Core i9, and it is equipped with an independent graphics card model Nvidia GeForce RTX 3090.

[0115] The sensing nodes are randomly deployed in the target area according to the sampling density of 5%, 10%, 15%, and 20%, and the number of signal sources can be single or multiple.

[0116] The simulation experiment covers eight scenarios in total, including four sampling densities and two numbers of signal sources.

[0117] In addition, in each scenario, the training set contains 20,000 samples, the test set contains 1,000 samples, and the validation set contains 200 samples.

[0118] The maximum number of iterations for network training is 50, the initial learning rate is set to 0.0003, the Adam optimization algorithm is used as the network training optimizer, and the sample size of each batch is 64.

[0119] 2. Simulation content:

[0120] Figure 3 The comparison of spectrum map visualization effects constructed by the present invention and other methods is shown in the scenario with a sampling density of 5% and multiple signal sources. Figure 3 The image in the upper left corner is the actual spectrum map. Figure 4 The figure in the upper right corner is the spectrum map estimated by the proposed method. Figure 4 The figure in the lower left corner is the spectrum map estimated by CGAN. Figure 4 The spectrum map obtained by Kriging estimation is shown in the lower right corner. This comparison shows that the spectrum map constructed by the present invention is closer to the true spectrum map, demonstrating a more accurate fit in terms of the spatial distribution and signal intensity variations of the spectrum map. Compared with traditional methods, the present invention demonstrates significant advantages in recovering details in the spectrum map, demonstrating the superiority of the proposed model in constructing and visualizing spectrum maps.

[0121] Figure 4 The comparison of spectrum map construction errors between the present invention and other methods in different scenarios is shown. Figure 4The horizontal axis represents the sampling density, and the vertical axis represents the spectrum map construction error. The broken line marked with a square represents the spectrum map construction error of the method of the present invention under different sampling density conditions in a multiple signal source scenario; the broken line marked with a diamond represents the spectrum map construction error of the CGAN method under different sampling density conditions in a multiple signal source scenario; the broken line marked with an inverted triangle represents the spectrum map construction error of the Kriging method under different sampling density conditions in a multiple signal source scenario; the broken line marked with a circle represents the spectrum map construction error of the method of the present invention under different sampling density conditions in a single signal source scenario; the broken line marked with a star represents the spectrum map construction error of the CGAN method under different sampling density conditions in a single signal source scenario; the broken line marked with an x ​​represents the spectrum map construction error of the Kriging method under different sampling density conditions in a single signal source scenario.

[0122] Figure 4 The results show that with the increase of sampling density, the construction error of each method tends to decrease. This is because a higher sampling density can provide more spectral data, thereby more accurately learning the signal propagation characteristics. In addition, the accuracy of the method of the present invention is improved by about 25.92% compared with the CGAN method and by about 43.87% compared with the Kriging method. In particular, in the multi-signal source scenario with a sampling density of 5%, the RMSE of the method of the present invention is 2.17dB lower than that of the CGAN method and 3.97dB lower than that of the Kriging method. These results fully demonstrate the effectiveness of the method of the present invention in various sampling density and signal source number scenarios, especially in low sampling density and multi-signal source scenarios.

[0123] Figure 5 The comparison of signal source positioning errors between the present invention and other methods in different scenarios is demonstrated. Figure 5 (a) shows the signal source positioning errors of the present invention and other methods at different sampling densities in a multi-signal source scenario. Figure 5 (b) shows the signal source positioning errors of the present invention and other methods at different sampling densities in a single signal source scenario.

[0124] Figure 5The horizontal axis represents the sampling density, and the vertical axis represents the average positioning error. The results show that the method of the present invention is significantly better than other methods in positioning effect. The KNN method and the K-means method perform poorly in positioning accuracy. Although LocUNet can have a low positioning error in single signal source and high sampling density scenarios, the effect decreases under low sampling density. In contrast, the method of the present invention can achieve a positioning error close to 0 in single signal source scenarios and low sampling density scenarios, and maintain stable positioning performance in multi-signal source scenarios. It can accurately distinguish the positions of multiple signal sources and achieve ultra-low error positioning. Compared with the positioning ambiguity or error amplification problems often faced by traditional methods in complex scenarios, the proposed method demonstrates excellent robustness and positioning accuracy. The experimental results further verify the adaptability of the present invention in complex environments and its significant positioning performance advantages.

[0125] Based on the simulation results and analysis above, the proposed data- and semantics-driven joint spectrum map construction and signal source localization method outperforms existing methods in terms of spectrum map detail recovery, spectrum map construction accuracy, and signal source localization error, particularly in low sampling density and multi-signal source scenarios. This method not only significantly improves spectrum map precision and signal source localization accuracy, but also maintains low network complexity, giving it broad application potential in practical communication scenarios.

[0126] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A data-semantics-driven joint spectrum map construction and signal source localization method, characterized by: include: Step 1: Multiple sensing nodes randomly deployed in the target area collect spectrum data and perform data normalization preprocessing; Step 2: spatially discretize the target area and establish a three-dimensional spectrum map representation model in the frequency-space domain as spectrum data; Step 3: Create a binary city map and a binary sampling location map as semantic knowledge; Step 4: Use the training set consisting of semantic knowledge and spectrum data to train the neural network model for spectrum map construction and signal source location, as follows: (1) Feature fusion of semantic knowledge and spectral data is performed to obtain data semantic fusion features. Specifically, the frequency-space domain joint three-dimensional spectrum map representation model, binary city map, and binary sampling location map are divided into three channels and input into the shared layer of the neural network model. The underlying features are extracted through the shared layer to achieve feature fusion of semantic knowledge and spectral data. (2) The data semantic fusion features are input into the spectrum map construction network DSD-UNet and the signal source localization network DSD-ResNet of the neural network model respectively. DSD-UNet completes the missing spectrum map to reflect the distribution of signal strength in the entire area. The features of the frequency-space domain combined three-dimensional spectrum map representation model extracted by the spectrum map construction network DSD-UNet are input into the signal source localization network DSD-ResNet to locate the signal source, thereby grasping the position of different signal sources in the area. Step 5: Calculate the loss of the spectrum map construction task and the loss of the signal source localization task respectively, and sum them up by weight using the weight parameter to obtain the total loss of the neural network model. The loss of the spectrum map construction task is calculated as follows: Where K represents the total number of frequency points; N×N represents the total number of spectrum map grids; E k,i,j Indicates frequency f k The received signal strength at each grid point on the spectrum map; P k,i,j Indicates frequency f k The received signal strength of the corresponding grid point on the real spectrum map; represents the loss of the spectrum map construction task; The loss calculation formula of the signal source localization task is as follows: Among them, p t Indicates the confidence of the model prediction; α t represents the sample weight hyperparameter; γ represents the focusing parameter; represents the loss of the signal source localization task; The total loss of the neural network model is: in, Represents the total loss of the neural network model; represents the weight parameter of the spectrum map construction task, Represents the weight parameter of the signal source localization task; Step 6: Use a dynamic adjustment mechanism to update the weight parameters according to the weight adjustment factor and the loss threshold. The update formula of the weight parameters is as follows: in, Represents the weight parameter of the spectrum map construction task; represents the updated spectrum map construction task weight parameter; Represents the weight parameter of the signal source localization task; represents the updated weight parameter of the signal source localization task; β represents the weight adjustment factor; represents the loss of the spectrum map construction task; represents the loss of the signal source localization task; loss threshold for the spectrum map construction task; Loss threshold for the signal source localization task; Step 7: Determine whether the current number of iterations has reached the set maximum number of iterations. If so, proceed to step 8. If not, increase the number of iterations by one and return to step 4 to continue training. Step 8: Input the validation set data into the neural network model used for spectrum map construction and signal source location to obtain the output result; Step 9, after processing the output results, obtain a complete spectrum map and the grid position of the signal source in the target area; Step 9 denormalizes the data output by the spectrum map construction network of the neural network model to obtain a complete spectrum map: P=(P max -P min )·P out +P min Among them, P represents the spectrum data after denormalization, P min represents the minimum value of the spectrum data output by the spectrum map construction network, P max represents the maximum value of the spectrum data output by the spectrum map construction network, P out Represents the spectrum data output by the spectrum map construction network; The signal source position mask output by the signal source localization network of the neural network model is processed to extract the area marked as the signal source in the mask. The coordinates corresponding to the area are the grid positions of the signal source in the target area.

2. The data-semantics dual-driven joint spectrum map construction and signal source localization method according to claim 1 is characterized by: The step 2 includes: The target area is spatially discretized and a two-dimensional spectrum map completion model is established. This model completes the grid values ​​in the target area after spatial discretization based on the collected data and unmonitored frequencies to form a corresponding two-dimensional spectrum map, thereby obtaining a three-dimensional spectrum map representation model of the target area in the frequency-space domain. The specific details are as follows: (1) If there are m sensing nodes in the grid I = (i, j), the grid is assigned the normalized average value of the spectrum data collected by these m sensing nodes. Otherwise, the grid is assigned 0, and the frequency f that can be monitored in the target area is obtained. k Incomplete spectrum S on k ; (2) setting all the values ​​of the spectrum map grids on the unmonitored frequency f0 to 0, forming a blank two-dimensional spectrum map E0 on the unmonitored frequency f0 as the target spectrum map to be inferred; (3) The frequency f that can be monitored in the target area k Incomplete spectrum S on k The blank two-dimensional spectrum map E0 on the unmonitored frequency f0 is stacked in ascending order according to the frequency dimension to obtain a frequency-space domain joint three-dimensional spectrum map representation model of the target area.

3. The data-semantics dual-driven joint spectrum map construction and signal source localization method according to claim 1 is characterized in that: Step 3 creates a binary city map as follows: where Z i,j represents the value of the binary city map at the grid I = (i, j) in the target area after spatial discretization; Create a binary sample location map as follows: Among them, M i,j Represents the value of the binary sampling position map at grid I = (i, j) in the target area after spatial discretization.

Images