A spectrum semantic communication system for sparse data completion and emitter localization

Abstract

The present invention discloses a spectrum semantic communication system for sparse data completion and radiation source positioning, including four basic modules: semantic extraction, channel encoding and decoding, semantic recovery, and task completion. The system aims to cope with the challenges posed by large-scale spectrum data to the transmission performance of traditional communication systems and alleviate the problem of spectrum scarcity. First, the semantic extraction module extracts discrete semantic information from sparse spectrum data to reduce the extracted features as much as possible without affecting the task accuracy. Secondly, in order to combine the semantic communication system with the traditional digital communication system, the discrete spectrum semantics are converted into a bit stream through the channel encoding and decoding module and restored at the receiving end. Then, in order to complete the subsequent tasks, the semantic recovery module designs a fully connected network and uses nonlinear regression to restore continuous spectrum data. Finally, in order to complete the radiation source positioning task, the sparse spectrum map is completed into a complete spectrum map through an automatic encoder, and a convolutional network is designed to output the positioning result. Simulation results show that under the premise of greatly reducing the amount of data transmission, the scheme can achieve similar task completion effects compared with the traditional scheme at a higher signal-to-noise ratio, and is far better than the traditional scheme at a lower signal-to-noise ratio, thereby improving the transmission efficiency of spectrum data.

Description

Technical Field

[0001] The present invention relates to a spectrum semantic communication system for sparse data completion and radiation source positioning, and belongs to the intersection field of artificial intelligence and communication technology. Background Art

[0002] Radio spectrum is a strategic resource widely used in social and national defense development. It needs to be used reasonably to realize its huge resource value. With the rapid development of new technologies such as the Internet of Things, the huge social and economic value of radio spectrum has become increasingly apparent, but it has also led to an increasingly prominent contradiction between spectrum demand and supply. The problem of spectrum scarcity has become one of the most challenging problems in wireless communications. As a database integration technology, spectrum mapping has important applications in wireless network planning and resource optimization. It can intuitively display the distribution of different radio parameters (such as received signal strength and power spectrum density) in a geographical area, which helps to alleviate the problem of spectrum scarcity. In order to construct spectrum maps and realize real-time spectrum resource management, the data generated by large-scale simulation experiments and high-density spectrum detectors is growing exponentially. However, due to the limited transmission resources and computing power, traditional communication systems are difficult to efficiently transmit large-scale spectrum data, resulting in low utilization of spectrum resources.

[0003] Weaver, a well-known information science expert, believes that the current communication system only stays at the technical level, but it is possible to develop towards the semantic level. With the changes in artificial intelligence technology, the communication system is moving towards the inherent direction of intelligence. The shift to the semantic paradigm will fundamentally change the status quo of the communication system. Through semantic understanding and generation, the amount of abstract semantic concept data can be much smaller than the amount of precise feature data, thereby greatly improving transmission efficiency.

[0004] The representation of semantic information is an indispensable part of semantic communication. Different definitions of semantic information will have a significant impact on the semantic extraction of data, encoding schemes, and the execution of downstream tasks. Inspired by the source-channel joint coding achieved by deep learning, H. Xie, Z. Qin, et al. designed the DeepSC semantic communication system in "Deep Learning Enabled Semantic Communication Systems" (IEEE Transactions on Signal Processing). The system adopts a transformer-based approach to transmit text and considers semantic encoding as a mapping from high-dimensional source vectors to low-dimensional encoded transmission vectors. D. Huang, et al. designed a coarse-to-fine image semantic coding model in the paper "Deep Learning-Based Image Semantic Coding for Semantic Communications" (IEEE Global Communications Conference), in which the base layer of the image retains semantic information and the enhancement layer restores fine details. In another paper by the author, "Towards Semantic Communications: Deep Learning-Based Image Semantic Coding" (IEEE Journal on Selected Areas in Communications), different entities generated by semantic segmentation are defined as new representation units as semantic concepts. On this basis, a novel reinforcement learning-based semantic bit allocation model is designed, whose reward is the improvement of rate, semantics, and perceptual performance after encoding a semantic concept using adaptive quantization levels. Z. Zhang et al. introduced a multi-level semantic feature extractor in the paper "Wireless Transmission of Images With The Assistance of Multi-level Semantic Information" (2022 International Symposium on Wireless Communication Systems (ISWCS)) to utilize high-level semantic information including text semantic information and segmentation semantic information, as well as low-level semantic information such as local details of the image.In the paper “Task-Oriented Image Transmission for Scene Classification in Unmanned Aerial Systems” (IEEE Transations on Communications), X. Kang et al. divide the image into 4×4 sub-images in equal proportion and define each sub-image as a semantic block. The semantic information extracted in the above literature is incompatible with digital communication and must be converted into bits before it can be transmitted through the channel. To address this problem, Q. Fu et al. proposed a transceiver for extracting multi-scale semantic features of images in “Vector Quantized Semantic Communication System” (IEEE Wireless Communications Letters), and introduced a multi-scale semantic embedding space for semantic feature quantization to make the data compatible with digital communication systems. Although semantic communication for images has been widely studied, there is currently a lack of applications for spectral data. Visualizing spectral data is similar to images, but it faces unique challenges due to the nature of the input data itself and the expected usage scenarios. On the one hand, images have fixed channel numbers and discrete value ranges, while spectral data are represented as floating-point values ​​with a dynamic range. On the other hand, spectral data is usually transmitted for the purpose of related task analysis and has strict quality requirements. Summary of the invention

[0005] The purpose of the present invention is to propose a spectrum semantic communication system for sparse data completion and radiation source positioning in response to the problem of large-scale spectrum data transmission. The method uses semantic extraction and recovery steps based on non-uniform quantization algorithm and neural network to improve communication efficiency as much as possible without affecting task accuracy. The spectrum semantic communication system proposed by the present invention can adapt to the characteristics of spectrum data, greatly reduce the amount of data transmission, and is not inferior to traditional solutions in terms of robustness, completion and positioning effects.

[0006] The technical solution adopted by the present invention to solve its technical problems is: a spectrum semantic communication system for sparse data completion and radiation source positioning, which is composed of four basic modules: semantic extraction, channel encoding and decoding, semantic recovery and task completion, and is intended to improve data transmission efficiency and accuracy. The semantic extraction module is used to extract discrete semantic information of sparse spectrum data to minimize the number of extracted features without affecting the task accuracy. In order to combine the semantic communication system with the traditional digital communication system, the channel encoding and decoding module converts the discrete spectrum semantics into a bit stream and performs a recovery operation at the receiving end. In order to complete subsequent tasks, the semantic recovery module adopts a fully connected network design and uses a nonlinear regression method to restore continuous spectrum data. The task completion module focuses on the radiation source positioning task, completes the sparse spectrum map into a complete spectrum map through an automatic encoder, and uses a convolutional network to analyze the complete spectrum map to output the positioning result. The system comprehensively uses the above key technologies to provide a comprehensive and efficient solution for the implementation of the spectrum semantic communication system.

[0007] The present invention also provides a method for implementing a spectrum semantic communication system for sparse data completion and radiation source positioning, the method comprising the following steps:

[0008] Step 1: Get the dataset and preprocess it to build the model to be used later. First, we need to construct a dataset suitable for the spectral semantic communication system based on the mathematical model. Then, we preprocess the dataset, including normalization and other operations, to ensure the quality and consistency of the data so that it can be input into the neural network. Next, we will build an appropriate model, which will be used for semantic recovery and task completion in subsequent steps.

[0009] Step 2: Semantic extraction. Use a specific algorithm to extract discrete semantic information from sparse spectral data. By analyzing the features and patterns of spectral data, it is possible to capture useful information and minimize the number of extracted features without affecting the accuracy of the task. This can effectively reduce the burden of data transmission and improve communication efficiency.

[0010] Step 3: Channel coding and decoding and communication. Channel coding and decoding technology is used to convert discrete spectrum semantics into bit streams and transmit them during the communication process. Finally, channel decoding is performed at the receiving end.

[0011] Step 4: Semantic restoration. A fully connected network is used to perform semantic restoration. Through nonlinear regression analysis, we are able to restore the continuous spectrum data so that it can better reflect the characteristics and patterns of the original data. This can improve the accuracy and completeness of the data and provide strong support for the completion of subsequent tasks.

[0012] Step 5: Task completion: Complete tasks such as sparse data completion and radiation source location according to specific needs and goals.

[0013] Step 5-1: Sparse data completion. Through the completion method of the autoencoder, we can complete the sparse spectrum map into a complete spectrum map. This step is very critical for the subsequent radiation source location task, because only a complete spectrum map can provide accurate radiation source location information.

[0014] Step 5-2: Radiation source location. Using the designed convolutional network, the radiation source location task can be performed based on the complete spectrum map. By analyzing the features and patterns in the spectrum map, we can accurately determine the location of the radiation source and provide important positioning information for related applications.

[0015] Step 6: Test the system effect. Test and evaluate the entire system to verify its performance and effect.

[0016] Beneficial effects:

[0017] 1. In order to extract the semantic representation of spectrum data, the present invention discretizes the spectrum data and only retains 4 levels of spectrum intensity. In addition, a spectrum semantic extraction iterative optimization algorithm based on Lloyd-max is proposed by adopting a non-uniform quantization method to minimize the gap between the extracted spectrum data and the source spectrum data. It can greatly reduce the resources occupied by channel transmission while retaining the necessary spectrum situation information.

[0018] 2. The present invention designs a fully connected network to recover spectrum data at the receiving end using nonlinear regression. Subsequent tasks cannot be completed by relying solely on spectrum semantic information. Therefore, an intelligent completion method based on a neural network is proposed to map the discrete data with only 4 semantic quantization levels into a series of continuous values ​​at the receiving end to improve task completion.

[0019] 3. Considering that only sparse spectrum data can be sampled in reality, the present invention proposes a self-encoder-based completion solution, and then uses a convolutional neural network to realize radiation source positioning based on a complete spectrum map. This improves the completion performance far superior to other solutions, thereby improving the accuracy of the positioning task. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a system flow chart of the present invention.

[0021] Figure 2 It is a schematic diagram of the visualization effect of comparing the real map, the sampling map, the reconstructed sampling map after semantic extraction and restoration, the reconstructed map and the positioning result when the number of sampling points is 154 in the present invention.

[0022] Figure 3 This is a graph showing the convergence of the fully connected network used for semantic recovery in the present invention.

[0023] Figure 4 The figure is a comparison chart of the task completion effect between the present invention and the traditional communication system.

[0024] Logo Description: Figure 4 (a) is a line graph of the sparse data completion task of the spectrum semantic communication system proposed in the present invention and the traditional 32-bit bit coding communication system, i.e., the root mean square error RMSE under different signal-to-noise ratio conditions, Figure 4 (b) is a line graph of the radiation source positioning task of the spectrum semantic communication system proposed in the present invention and the traditional 32-bit bit coding communication system, that is, the distance between the actual and predicted radiation sources under different signal-to-noise ratio conditions.

[0025] Figure 5 This is a comparison chart of the root mean square error (RMSE) performance of the automatic encoder completion method proposed in the present invention and other completion methods under different numbers of sampling points. DETAILED DESCRIPTION

[0026] The invention will be described in further detail below in conjunction with the accompanying drawings.

[0027] It should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present invention.

[0028] like Figure 1 As shown, the present invention provides a method for implementing a spectrum semantic communication system for sparse data completion and radiation source positioning, the method comprising the following steps:

[0029] Step 1: Obtain the data set and preprocess it to build the model to be used later.

[0030] Consider the urban area of ​​interest In order to input the data into the neural network for processing, the collected raw spectrum data is gridded, that is, the x-axis and y-axis are defined on Γ and divided into a rectangular grid of A×B, which contains points ξ uniformly distributed along the x-axis and y-axis respectively. a，b , where a=1,…,A,b=1,…,B,assuming a certain frequency band F, there are N radiation sources α at Γ n , n = 1, ..., N, the grid point where the drone flies is β m , m = 1, ..., M. In the following discussion, we will only study the special case of a single frequency f∈F to process the power spectrum density map, so the complete spectrum map can be expressed as The power spectral density at the UAV sampling point can be expressed as

[0031]

[0032] Among them, η n (f) represents the emission power spectrum density of the nth radiation source, H n (β, f) represents the frequency response of the channel between the nth radiation source and the drone. v(β, f) represents the noise and interference of the model. The sampled power spectral density map is expressed as is a mask matrix whose value can be set to 1 or 0, indicating whether the grid point is sampled. The power spectral density map of the entire sampling can also be defined as Ψ = {Ψ S ,Ψ T}, where the power spectrum density set of the sampling points is recorded as And T is the set of power spectral densities of the points to be completed. Before inputting the spectrum map into the framework, the data is first standardized.

[0033] Step 2: Semantic extraction.

[0034] At the semantic level, we first need to extract the semantic information of the source information and express it in a certain form. This invention defines a semantic quantization array To map the spectrum map into a spectrum semantic map. Mapping the input data into a limited number of semantic quantization arrays can intuitively reflect the strength of the power spectrum density to guide subsequent semantic recovery and greatly reduce the transmission cost.

[0035] To reduce the spectrum semantic map In order to improve the subsequent recovery performance, an iterative optimization algorithm for spectrum semantic extraction based on Lloyd-max is proposed. First, c needs to be initialized. Then, each sampling point data in Ψ is assigned the c closest to its value. i Next, in order to minimize the error between data points, the present invention selects MSE as the measurement indicator, which is expressed as follows:

[0036]

[0037] in, and are the real value and semantic value of the sampling point respectively. Finally, the key step is to update the semantic quantization array c. The specific process is to traverse each semantic quantization level and check whether the input sample data exists in the current semantic quantization interval (l i , l i+1 ). If yes, then l will be calculated i+1 The average of the left and right sample data is used to update l i If not, li+1 The default setting will be l i and l i+2 Finally, the allocation and update steps are repeated until the termination condition is met, that is, the maximum number of iterations is reached or the loss change of consecutive iterations is minimal. This algorithm can more accurately represent the spectral semantics while reducing the computational complexity and preserving the original distribution of the data as much as possible.

[0038] Step 3: Channel encoding, decoding and communication.

[0039] The discrete semantic information is converted into a bit stream using a coding algorithm. This coding technology can improve the reliability and error correction capability of data transmission, thereby reducing errors and losses that may occur in channel transmission. The bit stream after channel coding is transmitted to the receiving end. During the process, it will be affected by the physical noise in the channel.

[0040] During the transmission process, we need to ensure the integrity and stability of the data to ensure the quality and reliability of communication. Finally, the transmitted bit stream is decoded using the corresponding decoding algorithm. Through decoding, the original discrete semantic information can be restored and prepared for subsequent semantic recovery and task completion.

[0041] Step 4: Semantic restoration.

[0042] After semantic extraction, the spectrum data is expressed as a series of discrete values ​​representing intensity, which is subject to precision loss caused by semantic noise. Then, it is encoded into a bit stream suitable for transmission in the channel, which is further affected by physical noise. After the bit stream is restored at the receiving end, a semantic recovery method for the above two types of noise is urgently needed. Therefore, the present invention proposes an intelligent semantic recovery scheme based on a fully connected network, which uses nonlinear regression to map data with only L semantic quantization levels into a series of continuous values. In order to fully realize the accuracy of the prediction and measure the gap between the predicted value and the actual target, the loss function is set to:

[0043]

[0044] in, and represents the true value and semantic restoration value of the mth sampling point. The network is trained under ideal channel conditions to verify the robustness of semantic restoration.

[0045] Step 5: Mission accomplished.

[0046] In the present invention, the ultimate task of the receiving end is to achieve more accurate radiation source positioning. Prior to this, the sparse spectrum map needs to be completed.

[0047] Step 5-1: Sparse data completion.

[0048] The completion problem can be viewed as a radio propagation prediction problem. Traditional methods solve it through model-based methods, including empirical models, main path models, and more complex ray tracing algorithms. These data-driven methods require knowing the number, location, and power of all transmitters, and are susceptible to noise and interference during the measurement process. Therefore, for the scenario setting of this article, a data-driven spectrum map completion method was selected. Specifically, the generative model based on the autoencoder structure is able to use data sets from past measurements in different environments to learn the spatial structure of related propagation phenomena, such as shadows, reflections, and diffraction. A fully convolutional autoencoder was selected, which consists of a 13-layer encoder and a 13-layer decoder. The autoencoder is trained so that, which forces the encoder to The information in is compressed into N in the latent variable λ λ variables, and the decoder restores the complete spectrum map based on it. In order to guide the autoencoder to complete the position, a mask matrix is ​​added As an additional input dimension. The training loss of the network is defined as:

[0049]

[0050] Among them, ψ i and is the i-th power spectral density value of the real map and the reconstructed map.

[0051] Step 5-2: Locate the radiation source.

[0052] Traditional methods for locating radiation sources include mathematical modeling, maximum likelihood estimation, and statistical techniques, but they are often constrained by prior knowledge, model limitations, and computational complexity. Neural networks have the characteristics of automatic feature learning, nonlinearity, generalization, and speed. Among them, convolutional neural networks have the potential to improve the accuracy and speed of radiation source localization tasks by effectively learning complex spatial patterns. This paper proposes a simple convolutional neural network, which includes batch normalization, convolutional layers with maximum pooling functions, and fully connected layers with activation functions. The training loss of the network is defined as:

[0053]

[0054] Among them, α i and is the predicted position of the i-th source on the real map and the reconstructed map, and N is the number of sources.

[0055] Step 6: Test the system performance.

[0056] To test the model effect, the present invention randomly selects 90% of the data as the training set and the remaining 10% of the data as the test set. The calculation formula of the root mean square error RMSE is as follows:

[0057]

[0058] The effects of the present invention are further described in detail below in conjunction with simulation experiments, specifically including:

[0059] 1. Simulation hardware conditions

[0060] The simulation experiment of the present invention is carried out on the simulation platform of Python 3.6 and TensorFlow 1.15. The computer CPU model is E5-2680 v4, the number is 5, and the GPU model is NVIDIA Geforce RTX 3060. The video memory is 12.6GB.

[0061] 2. Simulation system parameters

[0062] The present invention uses simulated maps generated by mathematical models. The dataset is considered to be realistic enough to describe the structure of propagation phenomena, including path loss, shadowing, and noise. The dataset consists of square maps with a side length of 100m, whose frequency is set to 1400MHz, bandwidth is set to 5MHz, and noise is set to zero. In order to ensure the standardized format of the spectrum map, a grid resolution of 3m is set, and the map is converted into a 32×32×1 tensor format. In addition, the number of radiation sources is set to 2. The proposed framework consists of an iterative algorithm and three neural networks. Among them, the threshold of the Lloyd-Max quantizer is set to 0.0001, and the maximum number of iterations is set to 20. The semantic restoration network is a fully connected neural network for nonlinear regression. The dataset consists of 12,800 maps, and the number of iterations and batch size are set to 100 and 64 respectively. The completion network is an autoencoder structure, and its dataset consists of 40,000 maps. The number of iterations and batch size are set to 100 and 64 respectively. The localization network is a convolutional neural network, the dataset has 128,000 maps, and the number of iterations and batch size are set to 120 and 128 respectively. The above neural networks are trained using the Adam optimizer and the Leaky ReLU activation function, with a training set and validation set split ratio of 9:1. If the validation set loss does not improve, the learning rate will be reduced by 10% after 8 training rounds. The initial learning rate is set to 0.001.

[0063] 3. Simulation content

[0064] Figure 2The visualization results of the output of each module of the proposed framework are shown under the condition of a Gaussian white noise channel signal-to-noise ratio of 18dB. The horizontal and vertical axes represent the x-axis and y-axis of the region of interest, respectively, and the color of the pixel represents the intensity of the PSD. The data in the figure uses the logarithmic unit dBmW (abbreviated as dBm) to avoid performance loss caused by uneven data distribution. Specifically, the first figure shows the real map, the second figure shows the sparse map generated by drone sampling, and the third figure shows the map after semantic extraction, channel transmission and semantic restoration. Compared with the second figure, it is affected by semantic noise and physical noise, resulting in certain distortion. The last figure shows the complete map after completion, which is very similar to the original image. The black mark on the figure represents the real radiation source position, and the white mark represents the predicted radiation source position, which almost coincides with each other. It can be seen that the proposed framework can achieve ideal completion and positioning effects under good channel conditions.

[0065] Figure 3 The convergence of the semantic restoration network during training and testing is demonstrated. As can be seen from the figure, the validation set loss is slightly greater than the training set loss, but both are steadily decreasing, and the change in loss value reaches a stable state around 50 iterations. The results show that the proposed network can fit the data distribution well, map discrete data back to the corresponding continuous value, and achieve the expected semantic restoration effect.

[0066] Figure 4 A comparison chart showing the task completion of the spectrum semantic communication system proposed in the present invention and the traditional 32-bit bit coding communication system is shown. Figure 4 (a) Comparison of the completion effects of sparse data under different signal-to-noise ratio conditions, specifically shown in terms of root mean square error (RMSE); Figure 4 (b) is a comparison of the completion effect of locating the radiation source under different signal-to-noise ratio conditions, specifically showing the distance between the real radiation source and the predicted radiation source. When the channel conditions are good, the proposed system can achieve similar performance to the traditional system while greatly reducing the transmission bits. When there is more noise in the channel, the proposed system completes the task much better than the traditional communication system, reflecting the good robustness of the system.

[0067] Figure 5 The root mean square error (RMSE) of the proposed autoencoder completion method is shown compared with competing algorithms, where the parameters of these competitors are tuned to approximately produce the best performance. These state-of-the-art alternatives include (i) a regularization parameter of 10 -5 Kriging interpolation completion method; (ii) the regularization parameter is 10 -5 The multi-kernel function completion method of (iii) the regularization parameter is 3×10 -1Gaussian process regression completion method; (iv) k-nearest neighbor interpolation completion method as the baseline. As can be seen from the figure, for spectral data affected by semantic noise, when the number of sampling points is between 25 and 175, the performance of the adopted method is better than other competitors. Especially when the number of sampling points is small, the performance can be twice that of other methods. In actual situations, drones are more restricted and can often only collect fewer data points, so the autoencoder is the best choice for the completion method.

[0068] Based on the above simulation results and analysis, the spectrum semantic communication system proposed in the present invention can reasonably extract spectrum semantics and restore them to achieve the goal of reducing the amount of data transmission without affecting the accuracy of task completion. Under the premise of reducing the amount of data transmission, this solution can achieve consistent or even better task indicators than the baseline, and has good robustness, which can greatly alleviate the pressure of large data transmission, allowing the present invention to be better applied in actual recommendation scenarios.

[0069] In this field, the above is only a preferred example of the present invention and is not intended to limit the present invention. Any modification, replacement, improvement and polishing made by any technician in this field within the spirit and concept of the present invention should be covered within the protection scope of the present invention.

Claims

1. A spectrum semantic communication system for sparse data completion and radiation source positioning, characterized in that: The system includes a semantic extraction module, a channel encoding and decoding module, a semantic recovery module and a task completion module; The semantic extraction module is used to extract discrete semantic information from sparse spectrum data to minimize the number of extracted features without affecting the task accuracy. In order to make the semantic communication system compatible with the traditional digital communication system, the channel encoding and decoding module converts discrete spectrum semantics into bit streams and performs recovery operations at the receiving end. In order to complete subsequent tasks, the semantic recovery module adopts a fully connected network design and uses a nonlinear regression method to restore continuous spectrum data. The task completion module focuses on the radiation source positioning task, specifically completing the sparse spectrum map into a complete spectrum map through an automatic encoder, and using a convolutional network to analyze the complete spectrum map to output the positioning result. The implementation method of the spectrum semantic communication system for sparse data completion and radiation source positioning includes the following steps: Step 1: Obtain the data set and preprocess it to establish the model to be used later; First, it is necessary to construct a data set suitable for the spectrum semantic communication system based on the mathematical model. Then, the data set is preprocessed to ensure the quality and consistency of the data so that it can be input into the neural network. Next, an appropriate model will be established, which will be used for semantic recovery and task completion in the subsequent steps; Step 2: Semantic extraction: Use a specific algorithm to extract discrete semantic information from sparse spectrum data. By analyzing the features and patterns of spectrum data, useful information can be captured and the number of extracted features can be reduced as much as possible without affecting the task accuracy. This can effectively reduce the burden of data transmission and improve communication efficiency. Step 3: Channel coding and decoding and communication: Use channel coding and decoding technology to convert discrete spectrum semantics into bit streams, transmit them during the communication process, and finally perform channel decoding at the receiving end; Step 4: Semantic restoration: Use a fully connected network to perform semantic restoration operations. Through nonlinear regression analysis, continuous spectrum data can be restored so that it can better reflect the characteristics and patterns of the original data. This can improve the accuracy and completeness of the data and provide strong support for the completion of subsequent tasks. Step 5: Task completion: Complete the sparse data completion and radiation source positioning tasks according to specific needs and goals; Step 5-1: Sparse data completion: Through the completion method of the automatic encoder, the sparse spectrum map can be completed into a complete spectrum map; Step 5-2: Radiation source positioning: Using the designed convolutional network, the radiation source positioning task can be performed based on the complete spectrum map. By analyzing the features and patterns in the spectrum map, the location of the radiation source can be accurately determined, providing important positioning information for related applications.

2. A spectrum semantic communication system for sparse data completion and radiation source positioning according to claim 1, characterized in that: The method step 1 comprises: Consider the urban area of ​​interest In order to input the data into the neural network for processing, the collected raw spectrum data is gridded, that is, the x-axis and y-axis are defined on Γ and divided into a rectangular grid of A×B, which contains points ξ uniformly distributed along the x-axis and y-axis respectively. a，b , where a=1,...,A, b=1,...,B, A and B are the number of grid points of length and width of the map segmentation respectively; Assuming a certain frequency band F, there are N radiation sources α at Γ n , n = 1, ..., N, the grid point where the drone flies is β m , m = 1, ..., M, where M is the number of sampling points; the power spectrum density map is processed for the special case of a single frequency f∈F, so the complete spectrum map can be expressed as The power spectral density at the UAV sampling point can be expressed as Among them, η n (f) represents the emission power spectrum density of the nth radiation source, H n (β, f) represents the frequency response of the channel between the nth radiation source and the UAV; v(β, f) represents the noise and interference of the model; the sampled power spectral density map is expressed as is a mask matrix whose value can be set to 1 or 0, indicating whether the grid point is sampled; the power spectrum density map of the entire sampling can also be defined as Ψ={Ψ S ,Ψ T }, where the power spectrum density set of the sampling points is recorded as And T is the set of power spectral densities of the points to be completed; before the spectrum map is input into the framework, the data is first standardized.

3. The spectrum semantic communication system for sparse data completion and radiation source positioning according to claim 1, characterized in that: Step 2 of the method comprises: At the semantic level, define a semantically quantified array To map the spectrum map into a spectrum semantic map, and map the input data into a limited number of semantic quantization arrays, which can intuitively reflect the strength of the power spectrum density to guide subsequent semantic recovery and greatly reduce the transmission cost; To reduce the spectrum semantic map In order to improve the subsequent recovery performance, an iterative optimization algorithm for spectrum semantic extraction based on Iloyd-max is proposed. First, c needs to be initialized. Then, each sampling point in Ψ is assigned the c closest to its value. i ; Then, in order to minimize the error between data points, MSE is selected as the measurement indicator, which is expressed as follows, in, and are the true value and semantic value of the sampling point, respectively, and M is the number of sampling points; finally, the key step is to update the semantic quantization array c; the specific process is to traverse each semantic quantization level and check whether the input sample data exists in the current semantic quantization interval (l i , l i+1 ); if yes, calculate l i+1 The average of the left and right sample data is used to update l i If not, l i+1 The default setting will be l i and l i+2 The average value of; finally, the assignment and update steps are repeated until the termination condition is met, that is, the maximum number of iterations is reached or the loss change of consecutive iterations is minimal; this algorithm can more accurately represent the spectral semantics while reducing the computational complexity and preserving the original distribution of the data as much as possible.

4. The spectrum semantic communication system for sparse data completion and radiation source positioning according to claim 1, characterized in that: Step 3 of the method comprises: The discrete semantic information is converted into a bit stream using a coding algorithm. This coding technology can improve the reliability and error correction capability of data transmission, thereby reducing errors and losses that may occur in channel transmission. The bit stream after channel coding is transmitted to the receiving end. During the process, it will be affected by the physical noise in the channel. During the transmission process, it is necessary to ensure the integrity and stability of the data to ensure the quality and reliability of communication; finally, the transmitted bit stream is decoded using the corresponding decoding algorithm; through decoding, the original discrete semantic information can be restored and prepared for subsequent semantic recovery and task completion.

5. The spectrum semantic communication system for sparse data completion and radiation source positioning according to claim 3, characterized in that: Step 4 of the method comprises: An intelligent semantic restoration scheme based on a fully connected network is proposed. Nonlinear regression is used to map the data with only L semantic quantization levels into a series of continuous values. In order to fully realize the accuracy of prediction and measure the gap between the predicted value and the actual target, the loss function is set as: in, and represents the true value and semantic restoration value of the mth sampling point; the network is trained under ideal channel conditions to verify the robustness of semantic restoration.

6. The spectrum semantic communication system for sparse data completion and radiation source positioning according to claim 1, characterized in that: The method step 5-1 comprises: A data-driven spectrum map completion method was selected; specifically, a generative model based on an autoencoder structure was able to use a dataset of past measurements in different environments to learn the spatial structure of relevant propagation phenomena, including shadows, reflections, and diffraction; a fully convolutional autoencoder was selected, which consists of a 13-layer encoder and a 13-layer decoder; the autoencoder was trained so that the encoder The information in is compressed into N in the latent variable λ λ variables, and the decoder restores the complete spectrum map based on it; in order to guide the autoencoder to complete the position, a mask matrix is ​​added As an additional input dimension; the training loss of the network is defined as: Among them, ψ i and is the ith power spectral density value of the real map and the reconstructed map, A and B are the length and width of the grid points of the map segmentation, respectively.

7. The spectrum semantic communication system for sparse data completion and radiation source positioning according to claim 1, characterized in that: The method step 5-2 comprises: A simple convolutional neural network is proposed. The network structure includes batch normalization, a convolutional layer with a maximum pooling function, and a fully connected layer with an activation function. The training loss of the network is defined as: Among them, α i and is the predicted position of the i-th source on the real map and the reconstructed map, and N is the number of sources.

Images