Electromagnetic spectrum map prediction method based on knowledge and data fusion driving
Through the electromagnetic spectrum map prediction method driven by knowledge and data fusion, combined with propagation model, empirical fitting and generative artificial intelligence, the problems of data volume and terrain environment complexity in the existing technology are solved, and the complexity of the terrain environment is realized in multiple environments. Electromagnetic spectrum map prediction and regional environment identification improve prediction accuracy and applicability.
Patent Information
- Application Number
- CN202410823836.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-24
- Publication Date
- 2024-11-19
AI Technical Summary
The existing electromagnetic spectrum map construction methods are difficult to adapt to different data volumes and complex terrain environments, and it is difficult to effectively manage electromagnetic spectrum resources in practical applications, especially when spectrum resources are tight after the popularization of 5G and Internet of Things devices.
Using an electromagnetic spectrum map prediction method driven by knowledge and data fusion, by obtaining target area images and base station prior information, selecting appropriate prediction methods based on data volume, including using propagation models, empirical fitting and generative artificial intelligence, combined with volume Generate neural networks and custom neural networks to predict electromagnetic environment distribution.
Implement electromagnetic spectrum map prediction suitable for a variety of data volumes and terrain environments, can identify regional environments and generate accurate electromagnetic information distribution maps, solve the problems of small data volumes and difficult to identify propagation paths, and improve prediction accuracy and applicability. .
Smart Images

Figure CN118982601A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for constructing an electromagnetic environment map, and in particular to a method for predicting an electromagnetic spectrum map based on knowledge and data fusion drive, including three construction methods: knowledge-driven, data-driven, and knowledge and data-driven, as well as partial image recognition and generative artificial intelligence methods. Background Art
[0002] In the military field, electromagnetic space has become the sixth dimension of combat space after land, sea, air, space and the Internet. The electromagnetic spectrum can provide auxiliary decision-making information for electromagnetic situation generation, equipment deployment and operation, combat frequency, etc. With the rapid development of information technology such as 5G and the Internet of Things, the number of various radio devices has increased exponentially, the possibility of interference and abnormal frequency use has greatly increased, and spectrum resources have become increasingly tight, which will bring new challenges to the management of electromagnetic spectrum resources.
[0003] The construction of the electromagnetic spectrum map requires the collection of electromagnetic information from the real environment in advance, and then the global electromagnetic environment map is constructed using the collected information. How to perform this mapping is an important research topic. The current research methods can be divided into three methods: knowledge-driven, data-driven, and knowledge and data-driven. The knowledge-driven method uses some existing prior knowledge, such as the location of the base station (transmitting signal), the propagation model of the wireless channel, etc. to directly predict the receiving point. The data-driven method does not require relevant prior information. It directly collects electromagnetic data in the actual environment, and then uses appropriate data processing algorithms to complete the electromagnetic information of the remaining positions, and finally completes the construction of the electromagnetic spectrum map. The knowledge and data-driven method incorporates some prior knowledge into the data processing algorithm, making it more professional-oriented and achieving the complementary advantages of the two.
[0004] In actual application scenarios, there are many possible amounts of electromagnetic information data collected, and a single prediction method can only predict well when the data amount is appropriate. Therefore, different prediction methods need to be selected under different data amounts. In addition, the actual environment is complex and changeable, which also has a great interference on the prediction results. The current research on electromagnetic spectrum maps focuses more on scientific research and is difficult to move towards practical applications. Based on the above problems, the present invention proposes an electromagnetic spectrum map prediction method that can be applied to a variety of data amounts, can be applied to a variety of different terrain environments, and is equipped with an image processing function for identifying regional environments. Summary of the invention
[0005] In view of the defects and shortcomings of the above-mentioned prior art, the present invention aims to provide a method for constructing an electromagnetic environment map of a target area. After obtaining the target area image and prior information of the base station, an appropriate method is adopted according to different data amounts to predict the electromagnetic environment distribution of the entire area.
[0006] The technical solution of the present invention is an electromagnetic spectrum map prediction method based on knowledge and data fusion drive. An appropriate prediction method is selected according to the existing data volume and geographical environment information. The specific steps are as follows:
[0007] S1: Obtain image information of the area to be predicted. Divide the target area into multiple terrain blocks according to a certain ratio, use the trained convolutional neural network to identify the terrain category of each block, and label each terrain category.
[0008] S2: Determine the prediction method based on the current amount of data.
[0009] When there is no data, input the frequency, transmitting antenna height, receiving antenna height, and propagation distance, and directly use the existing propagation model for prediction. There are M propagation models corresponding to different terrains. First, determine which terrain patches the propagation path passes through to confirm which model is used for the propagation path.
[0010] When the amount of data is small, empirical fitting is used for prediction. First, find out the relevant electromagnetic physical quantities that affect the electromagnetic environment, then mathematically model these electromagnetic physical quantities according to different terrains, confirm the parameters to be fitted, and finally use the existing data to empirically correct the parameters to be fitted so that the prediction results gradually approach the actual values.
[0011] When the amount of data is large, the generative artificial intelligence AWGAN_GP is used to learn the distribution law of the data based on the real data, and to generate additional simulated data. The simulated data and real data are used together to train the custom neural network. First, a custom neural network is built, and some electromagnetic propagation knowledge is embedded in it to improve the prediction accuracy. A large amount of data is used to train the neural network, and finally the electromagnetic information of the unknown location is predicted.
[0012] Compared with the prior art, the present invention has the following beneficial effects:
[0013] 1. The present invention is applicable to electromagnetic spectrum map prediction methods under various data volumes, can be applied to various different terrain environments, and is equipped with an image processing function for identifying regional environments.
[0014] 2. The present invention can select appropriate prediction methods under different data amounts, and can use generative artificial intelligence to generate simulated data with approximate distribution based on existing data, thereby solving problems such as small data volume and difficult identification of propagation paths. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 It is a work flow chart of the present invention.
[0016] Figure 2It is a diagram of the structure of a custom neural network of the present invention.
[0017] Figure 3 It is a schematic diagram for visualizing the prediction results of the present invention. DETAILED DESCRIPTION
[0018] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.
[0019] The application scenario of the present invention belongs to the situation where the terrain area is large. An image contains a variety of terrain information, so it is necessary to complete the classification processing of different terrain areas in advance. The convolutional neural network adapts well to this demand. The convolution of the present invention includes 3 convolutional layers, 2 pooling layers and a fully connected layer. The convolutional layer is used to filter image features, the pooling layer is used for dimensionality reduction and compression, and filtering redundant information. The fully connected layer is used for the final classification result output. Figure 1 As shown in the figure, after the entire image is divided into multiple small blocks, each small block is sent to the convolutional neural network for terrain classification to facilitate subsequent model selection.
[0020] Table 1 shows the relevant physical quantities needed for prediction.
[0021]
[0022]
[0023] Select the appropriate prediction method based on the current data volume. (1) When there is no data, use the existing propagation model to directly predict. Different propagation models correspond to the five terrain types: urban, suburban, rural, mountainous, and water. The Okumura-Hata model is used for urban, suburban, and rural areas. The Egli model is used to calculate the propagation path loss in uneven areas and is applied to the propagation model in the mountains. Their mathematical expressions are as follows:
[0024] L p城 =69.55+26.16logf c -13.82logh te -α(h re )+(44.9-6.55logh te )logd
[0025]
[0026] L p村 =L p城 -40.94-4.78(logf c ) 2 +18.33logf c
[0027] L p山 =88+40logd+20logf c -20logh te -20logh re
[0028]
[0029] where f c (MHz) indicates the operating frequency, h te (m) represents the effective height of the base station antenna, h re (m) is the effective height of the mobile antenna at the receiving end, d is the transmission distance from the base station to the terminal, α(h re )(dB) is the effective antenna correction factor, and the calculation rules are as follows:
[0030]
[0031] When making predictions, both direct use of the propagation model and empirical correction require obtaining terrain information of the propagation path. First, determine the location of the base station and the terminal to determine the propagation path, and model the propagation path as a line segment with the starting point being the transmitting base station and the end point being the receiving terminal, and each small plot is a rectangular area. Then, use a graphical algorithm to determine which terrain patches the propagation path intersects with, so as to determine which types of terrain the path passes through, and select the appropriate propagation model based on the category with the largest proportion. After calculating the path loss, subtract the corresponding path loss from the transmit signal strength of each base station to obtain the received signal strength, and select the one with the largest received signal strength as the predicted value for the current location.
[0032] (2) When the amount of data is small, the prediction is made by empirical fitting, and different mathematical formulas are used for fitting different terrains. For five different terrains, namely, urban, suburban, rural, mountainous, and water surface, different mathematical expressions are used for modeling, as shown below:
[0033] L p城 =a1+a2logf c +a3logh te +a4logh re +a5logd+a6(logd×logh te )
[0034]
[0035] L p村 =a1+a2logf c +a3logh te +a4logh re+a5logd+a6(logd×logh te )
[0036] +a7(logf c ) 2
[0037] L p山 =a1+a2logf c +a3logh te +a4logh re +a5logd
[0038]
[0039] First, set the initial values of the parameters a1 to a7 to be fitted, and use the existing data to fit the various empirical formulas, so that the output of the fitted mathematical model gradually approaches the true value. After obtaining the fitting formula, use the method in (1) to complete the prediction of the target area.
[0040] (3) When the amount of data is large, deep learning is used for prediction. First, the actual data is input into the AWGAN_GP network in the data generation module to generate more simulated data with similar distribution. The AWGAN_GP network uses Wasserstein distance to replace the original JS divergence of GAN, and introduces a regularization term to satisfy Lipschitz continuity. Finally, a term is added to accelerate the convergence of the network and prevent the network from falling into the state of Nash equilibrium. The final loss function of AWGAN_GP is as follows:
[0041]
[0042] in is the Wasserstein distance, which is the loss function of the original WGAN and uses the Kantorovich-Rubinstein duality to simplify the calculation. is the gradient constraint (gradient penalty) term. The closer it is to 1, the less penalty is imposed. It is used to satisfy the Lipschiz continuity condition. It is an additional term introduced to break the Nash equilibrium.
[0043] like Figure 2As shown in the figure, the customized neural network is trained using actual data and simulated data. One part of the customized neural network is composed of a multi-layer perceptron MLP, and the other part is embedded with propagation model knowledge. The latitude and longitude, carrier frequency, time and other information of the predicted point are used as input features to predict the received signal strength RSRP and interference noise noise, and the signal to interference noise ratio SINR is calculated based on this. The neural network first separates the input information in half, one half is directly passed through the propagation model to obtain the prediction result, and the other half is predicted through the MLP, and then the prediction results of the two are weighted and summed to obtain the final result. Figure 3 It is a schematic diagram for visualizing the prediction results of the present invention.
[0044] The description of the specific embodiments is only used to help understand the method and core idea of the present invention. It should be noted that, for those of ordinary skill in the art, several improvements and modifications can be made to the present invention without departing from the principles of the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting electromagnetic spectrum maps based on knowledge and data fusion, characterized in that: The steps include: S1: Obtain image information of the area to be predicted. Divide the target area into multiple terrain blocks according to a certain ratio, use the trained convolutional neural network to identify the terrain category of each block, and label each terrain category. S2: Determine the prediction method based on the current amount of data. When there is no data, input the frequency, transmitting antenna height, receiving antenna height, and propagation distance, and directly use the existing propagation model for prediction. There are M propagation models corresponding to different terrains. First, determine which terrain patches the propagation path passes through to confirm which model is used for the propagation path. When the amount of data is small, empirical fitting is used for prediction. First, find out the relevant electromagnetic physical quantities that affect the electromagnetic environment, then mathematically model these electromagnetic physical quantities according to different terrains, confirm the parameters to be fitted, and finally use the existing data to empirically correct the parameters to be fitted so that the prediction results gradually approach the actual values. When the amount of data is large, the generative artificial intelligence AWGAN_GP is used to learn the distribution law of the data based on the real data, and to generate additional simulated data. The simulated data and real data are used together to train the custom neural network. First, a custom neural network is built, and some electromagnetic propagation knowledge is embedded in it to improve the prediction accuracy. A large amount of data is used to train the neural network, and finally the electromagnetic information of the unknown location is predicted.
2. According to claim 1, the electromagnetic spectrum map prediction method based on knowledge and data fusion drive is characterized in that: The step S1 is specifically as follows: firstly, the entire image is divided into n*m different square blocks, the image information of the small blocks is input into the trained convolutional neural network, and the soft classification task is performed. The neural network will output the probability that the small block belongs to different terrains, and the terrain with the highest probability is selected as the prediction result. Finally, the prediction result of each small block and the position information of the four vertices are recorded for subsequent use.
3. The electromagnetic spectrum map prediction method based on knowledge and data fusion drive according to claim 1 is characterized in that: When there is no data in step S2, specifically: five propagation models of different terrains are used: urban, suburban, rural, mountainous, and water surface. The urban, suburban, and rural areas use the Okumura-Hata model to calculate the loss of the transmission path. Taking the urban propagation model as an example: L p城 =69.55+26.16logf c -13.82logh te -α(h re )+(44.9-6.55logh te )log d where f c (MHz) indicates the operating frequency, h te (m) represents the effective height of the base station antenna, h re (m) is the effective height of the mobile antenna at the receiving end, d is the transmission distance from the base station to the terminal, α(h re )(dB) is the effective antenna correction factor and at different f c Under the condition of knowing the above prior information, the path loss L can be calculated by only the distance d (km) between the base station and the receiving end. p城 . Both direct use of the propagation model and empirical correction require obtaining terrain information of the propagation path. On the one hand, graphics-related algorithms are used to model the propagation path as a line segment starting from the transmitting base station and ending at the receiving terminal, and each small plot is a rectangular area. It is only necessary to determine whether the line segment intersects with the rectangle to determine whether the propagation path passes through the plot. A propagation path will pass through multiple small plots. By calculating which of the five types of plots mentioned above accounts for the largest proportion, the terrain to which the propagation path belongs is confirmed, and the corresponding propagation formula is applied.
4. The electromagnetic spectrum map prediction method based on knowledge and data fusion drive according to claim 1 is characterized in that: When the amount of data in step S2 is small, the specific content is: define different formulas to be fitted according to 5 different terrains. Taking urban terrain as an example, its propagation model is defined as: L p城 =a1+a2 logf c +a3 log h te +a4 logα(h re )+a5 log d+a6 log d×log h te Among them, a1 to a6 are the parameters to be fitted, f c (MHz) indicates the operating frequency, h te (m) represents the effective height of the base station antenna, h re (m) is the effective height of the mobile antenna at the receiving end, α(h re )(dB) is the effective antenna correction factor, and at different f c There are different calculation rules in different cases, where d (km) is the distance between the base station and the receiving end. Before fitting, the initial values of a1 to a6 are set in advance, and the formula is fitted with the existing data so that its prediction results gradually approach the true value.
5. The electromagnetic spectrum map prediction method based on knowledge and data fusion drive according to claim 1 is characterized in that: When the amount of data in step S2 is large, the specific content is: the AWGAN_GP network uses Wasserstein distance to replace the original JS divergence of GAN, and introduces a regularization term to satisfy Lipschitz continuity. Finally, a term is added to accelerate the convergence of the network and prevent the network from falling into the state of Nash equilibrium. The final loss function of AWGAN_GP is as follows: in is the Wasserstein distance, which is the loss function of the original WGAN and uses the Kantorovich-Rubinstein duality to simplify the calculation. is the gradient constraint (gradient penalty) term. The closer it is to 1, the less penalty is imposed. It is used to satisfy the Lipschiz continuity condition. It is an additional term introduced to break the Nash equilibrium.
6. The electromagnetic spectrum map prediction method based on knowledge and data fusion drive according to claim 1 is characterized in that: The specific operation of using the simulated data and the real data together for the training of the customized neural network in step S2 is as follows: the input data is divided into two halves, one part is directly outputted through the logarithmic propagation model, and the other part is predicted by the neural network through the trained multi-layer perceptron, and finally the prediction results of the two are weighted summed to obtain the final result.
Citation Information
Cited By
Data semantic dual-driven joint spectrum map construction and signal source positioning method
CN120018060A
