Unmanned aerial vehicle group spectrum situation map construction method, device, equipment, medium and product

Abstract

The invention discloses an unmanned aerial vehicle group frequency spectrum situation map construction method and device, equipment, a medium and a product, and relates to the field of wireless communication. The unmanned aerial vehicle group spectrum situation map construction method comprises the following steps: constructing a stack generative adversarial network model; constructing a training set; training the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model; and complementing the local spectrum situation map acquired by the unmanned aerial vehicle group based on the trained stacked generative adversarial network model. According to the method, the complete spectrum situation of the task area is accurately predicted by using the local power spectrum information sensed by the unmanned aerial vehicle, and reliable information support is provided for accurate electromagnetic spectrum management and control.

Description

Technical Field

[0001] The present application relates to the field of wireless communications, and in particular to a method, device, equipment, medium and product for constructing a spectrum situation map of a drone swarm. Background Art

[0002] With the rapid development of the wireless communication industry, providing communication users with high-quality and diverse wireless data transmission services has made social production and life safer and more convenient. However, with the increase in frequency-using devices, the probability of communication blocking has increased sharply. How to reasonably allocate spectrum resources and realize spectrum sharing in a radio environment with tight spectrum resources to ensure that each frequency-using unit obtains a safe and reliable wireless communication transmission link for data and control signal transmission has become an urgent problem to be solved. Common methods for achieving spectrum sharing include: dynamic spectrum access, power control, task path optimization, etc., and these methods require the use of spectrum detection or spectrum database to obtain spectrum situation information in the area where the device is located.

[0003] The regional spectrum situation directly reflects the time domain characteristics, field strength, pulse density and overall complexity of the signal in the environment, and obtains the global electromagnetic spectrum situation in the mission area, such as Figure 1 As shown, the parameter information such as power spectrum density required to construct the global spectrum situation can be obtained by widely and reasonably deploying electromagnetic spectrum monitoring equipment or communication units with spectrum sensing capabilities.

[0004] Using high-precision spectrum situation maps as information support for spectrum management and control can effectively guarantee the formulation of frequency use plans and the scheduling of frequency use equipment. However, in actual application scenarios, due to the limited number of spectrum monitoring equipment and the difficulty in rapid deployment, it takes a lot of time to obtain a complete spectrum situation map. If there are dynamic signal sources in the area, the accuracy will be greatly reduced. It is still difficult to quickly build a spectrum situation map that can meet the needs of spectrum management and control. Because drones are highly dynamic and inexpensive. They can reach the target location in a short time and establish a communication connection with the control node. A swarm of spectrum monitoring drones can be used to sample spectrum status information at multiple spatial locations in the mission area.

[0005] The main difficulty in using spectrum monitoring drones to obtain spectrum data and construct regional spectrum situation maps is how to combine the measured power spectrum density data of known locations and use the correlation of wireless transmission in the spatial domain, spectrum, and time domain to predict the power spectrum density of undetected locations. In past studies, interpolation methods, tensor completion, and basis expansion algorithms of two-dimensional spectrum maps are often used to construct spatial spectrum maps. The methods for estimating incomplete locations using locally known spectrum information mainly include: Inverse Distance Weight (IDW) and Kriging interpolation. These algorithms estimate the power spectrum density of incomplete locations by the distance between the interpolated location and the transmitter, and the influence of the power spectrum density value of the detectable location on the undetected location. This estimation only considers the numerical relationship and is independent of the physical characteristics of electromagnetic field propagation in space, so it is difficult to determine the rationality of the estimation result. The power spectrum density of the incomplete location is estimated by the basis expansion algorithm, which directly uses the approximate radio propagation function as the prior information of the radio environment and uses the sparsity of the spectrum and space to establish the basis expansion model. This method is only effective in specific mission environments. In the case of complex and changeable radio environments in the model and when drones acquire spectrum status data, the number and positions of perceived drones are random and changeable, and the estimated results may deviate significantly from the actual situation.

[0006] Generative Adversarial Networks (GANs) can use real data to train the generator (G) and the discriminator (D), and adjust the parameters of the neural network in the confrontation between the two to achieve infinite fitting of complex functions, and can mine implicit radio propagation features in a large amount of pre-acquired real data. For example, in the application scenario of cognitive radio, the spectrum state is perceived by the secondary user, and the incomplete spectrum map is converted into an RGB map. The completed spectrum map is obtained through the proposed map estimation GAN (MEGAN). This method has a better completion effect than the interpolation method. However, when the data incompleteness is high, the predicted spectrum map is partially blurred, and there is still room for improvement in the performance of radio propagation details. For another example, the distribution of radio parameters such as the signal power spectrum density obtained by secondary users in cognitive radio is used to construct a three-dimensional spectrum situation using a three-dimensional U-shaped conditional generation adversarial network (3D-UCGAN). However, this method requires more spectrum monitoring units to sample the three-dimensional spatial power spectrum density, and it is difficult to deploy such a large number of sub-UAVs in a UAV swarm to achieve the required sampling rate.

[0007] How to use discrete spectrum situation data to predict and complete all locations in the mission area that have not obtained spectrum status information, so as to construct a spectrum situation map of the mission area that is close to the actual spectrum status, has become a technical problem that needs to be solved urgently. Summary of the invention

[0008] The purpose of this application is to provide a method, device, equipment, medium and product for constructing a spectrum situation map of a drone swarm, which can accurately predict the complete spectrum situation of the mission area by using the local power spectrum information perceived by the drone, and provide reliable information support for precise electromagnetic spectrum management and control.

[0009] To achieve the above objectives, this application provides the following solutions:

[0010] In a first aspect, the present application provides a method for constructing a spectrum situation map of a drone swarm, and the method for constructing a spectrum situation map of a drone swarm includes:

[0011] Build a stacked generative adversarial network model;

[0012] Constructing a training set; the training set includes a plurality of training samples, the input in the training samples is an incomplete spectrum situation map obtained by masking the complete spectrum situation map according to a preset sampling rate, and the label in the training samples is the complete spectrum situation map;

[0013] Training the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model;

[0014] The local spectrum situation map obtained by the drone swarm is completed based on the trained stacked generative adversarial network model.

[0015] In a second aspect, the present application provides a device for constructing a spectrum situation map of a drone swarm, wherein the device for constructing a spectrum situation map of a drone swarm applies the above-mentioned method for constructing a spectrum situation map of a drone swarm, and the device for constructing a spectrum situation map of a drone swarm comprises:

[0016] Model building module, used to build stacked generative adversarial network models;

[0017] A training set construction module is used to construct a training set; the training set includes a plurality of training samples, the input in the training samples is an incomplete spectrum situation map obtained by masking the complete spectrum situation map according to a preset sampling rate, and the label in the training samples is the complete spectrum situation map;

[0018] A model training module, used to train the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model;

[0019] The completion module is used to complete the local spectrum situation map obtained by the drone swarm based on the trained stacked generative adversarial network model.

[0020] In a third aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned method for constructing a spectrum situation map of a drone swarm.

[0021] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for constructing a spectrum situation map of a drone swarm.

[0022] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned method for constructing a spectrum situation map of a drone swarm.

[0023] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0024] The present application provides a method, device, equipment, medium and product for constructing a spectrum situation map of a drone swarm. The present application applies a stacked generative adversarial network to the problem of completing the incomplete spectrum situation map obtained by a spectrum monitoring drone swarm. It can extract a small amount of situation information contained in the incomplete spectrum map, construct a complete spectrum situation map that is close to the actual spectrum situation distribution, and accurately predict the power spectrum density of unknown locations in the absence of prior information such as the radio wave propagation model of the mission area, providing reliable information support for precise electromagnetic spectrum control. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0026] Figure 1 A schematic diagram of obtaining spectrum situation information through a spectrum detection unit or other sensing device in the background technology of this application;

[0027] Figure 2 This is an application environment diagram of a method for constructing a spectrum situation map of a drone swarm in one embodiment of the present application;

[0028] Figure 3 A flowchart of a method for constructing a spectrum situation map of a drone swarm provided in one embodiment of the present application;

[0029] Figure 4 A schematic diagram of the structure of a stacked generative adversarial network model provided in one embodiment of the present application;

[0030] Figure 5 A schematic diagram of the network structure of a generator provided in one embodiment of the present application;

[0031] Figure 6 A schematic diagram of the network structure of a discriminator provided in one embodiment of the present application;

[0032] Figure 7 A schematic diagram of spatial spectrum situational awareness of a spectrum monitoring drone swarm provided in one embodiment of the present application;

[0033] Figure 8 A spectrum situation map of different frequencies and times obtained by using spectrum monitoring drone group detection provided by an embodiment of the present application;

[0034] Fig. 9 A flowchart of a stacked generative adversarial network model training process provided in one embodiment of the present application;

[0035] Fig.10 A schematic diagram of a simulation of a global spectrum situation map of multiple signal sources provided in an embodiment of the present application;

[0036] Fig.11 A spectrum situation map with different numbers of signal sources and a schematic diagram of algorithm generation results provided in an embodiment of the present application;

[0037] Fig.12 A graph showing the mean square error of a test set during the iterative training process of a stacked generative adversarial network prediction and completion algorithm provided in one embodiment of the present application;

[0038] Fig.13 A mean square error curve diagram of prediction and completion results with different sampling rates provided in an embodiment of the present application;

[0039] Fig.14 A comparison chart of algorithm parameters and prediction completion effects with different numbers of stacked layers provided in an embodiment of the present application;

[0040] Fig.15 A schematic diagram of an incomplete spectrum situation map and algorithm prediction completion results of different detection modes of a drone swarm provided in an embodiment of the present application;

[0041] Fig.16 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0043] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0044] The method for constructing a drone swarm spectrum situation map provided in the embodiment of the present application can be applied to Figure 2 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the local spectrum situation map obtained by the training set or the drone swarm to the server 104, and the server 104 receives the local spectrum situation map obtained by the training set or the drone swarm. The server 104 trains the stacked generative adversarial network model based on the training set or the local spectrum situation map obtained by the drone swarm, and obtains a trained stacked generative adversarial network model; based on the trained stacked generative adversarial network model, the local spectrum situation map obtained by the drone swarm is completed. The server 104 can feedback the completed result to the terminal 102. In addition, in some embodiments, the method for constructing the spectrum situation map of the drone swarm can also be implemented by the server 104 or the terminal 102 alone. For example, the terminal 102 can train the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model; and complete the local spectrum situation map obtained by the drone swarm based on the trained stacked generative adversarial network model. The server 104 can also obtain the training set or the local spectrum situation map obtained by the drone swarm from the data storage system, and train the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model; and complete the local spectrum situation map obtained by the drone swarm based on the trained stacked generative adversarial network model.

[0045] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0046] In an exemplary embodiment, Figure 3 As shown, a method for constructing a spectrum situation map of a drone swarm is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or can be executed by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 2 The server 104 in the example is used as an example to illustrate the method, which includes the following steps 201 to 204. Among them:

[0047] Step 201, construct a stacked generative adversarial network model.

[0048] Step 202, construct a training set; the training set includes multiple training samples, the input in the training samples is an incomplete spectrum situation map obtained by masking the complete spectrum situation map according to a preset sampling rate, and the label in the training samples is the complete spectrum situation map.

[0049] Step 203: Train the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model.

[0050] Step 204: Complete the local spectrum situation map obtained by the drone swarm based on the trained stacked generative adversarial network model.

[0051] By implementing the above steps 201 to 204, it is possible to extract a small amount of situation information contained in the incomplete spectrum map, construct a complete spectrum situation map that is close to the actual spectrum situation distribution, and accurately predict the power spectrum density of an unknown location in the absence of prior information such as the radio wave propagation model of the mission area. Specifically, the following technical effects can be achieved:

[0052] (1) To address the problem of spectrum information acquisition in wide-area coverage, a global spectrum situation completion system model is established using a swarm of electromagnetic spectrum monitoring drones. This model uses a small number of drones to perform spectrum perception and a generative adversarial network to achieve prediction and completion, ultimately obtaining an accurate and complete spectrum situation.

[0053] (2) In order to address the problems of limited number of sensing drones and sparse regional power spectral density data, which lead to increased errors in the algorithm's prediction and completion results and poor detail performance, the stacked generative adversarial network method is used to enhance the quality of the completed spectrum situation map.

[0054] (3) Through a large number of simulation experiments, it is verified that the algorithm proposed in this application can use the spectrum data of the ubiquitous perception of drone swarms to complete the complete spectrum situation in the region. By comparing the completion effect and parameter quantity of the adversarial network generated by each stacking layer at different sampling rates, the more appropriate number of stacking layers in different situations is analyzed. Finally, through a variety of incomplete sample tests, it is proved that this application can complete the incomplete spectrum situation obtained under various sampling methods, further reducing the demand for the number of drones.

[0055] In step 201 of the above implementation, in order to achieve better results in the spectrum map prediction and completion problem, it is necessary to optimize and improve the generative adversarial network algorithm. General optimization methods include improving the network structure, preprocessing the input data, adding an attention mechanism to improve the feature extraction capability, etc. In this implementation, the image quality of the prediction and completion result is improved by stacking the generator network and the discriminator network structure in the generative adversarial network. Using stacked generative adversarial networks to solve known spectrum situation information The problem that the spectrum situation information {φ′(x′,y′,f)} of the incomplete position predicted by the basic generative adversarial network has a large error is caused. The incomplete spectrum situation map structure obtained by the algorithm and the drone swarm quickly constructs a regional spectrum situation map with lower error and more accurate details. The structure of the stacked generative adversarial network model provided in this embodiment is as follows: Figure 4 shown.

[0056] The first-layer generative adversarial network (Stage-ⅠGAN) of the stacked structure randomly extracts the incomplete spectrum situation map from the incomplete image set and converts it into a 3-channel tensor with a spatial size of 512×512. The incomplete spectrum situation map is input into the generator of the first-layer generative adversarial network to obtain the prediction completion result. The "sample-label" consisting of the predicted completion result and the complete spectrum situation map corresponding to the real distribution is input into the discriminator to obtain the discrimination result. In order to correct and supplement the prediction completion result of stage-1GAN, the second-layer generative adversarial network structure (Stage-ⅡGAN) introduces the conditional generative adversarial network (CGAN) structure, and inputs the spectrum situation information obtained by spectrum sensing as additional conditional information into the network. Therefore, stage-2GAN predicts the completion result of the generator of the first layer. After being spliced ​​with the incomplete spectrum situation map that was originally used as the first-layer input, it is converted into a 6-channel tensor with a spatial size of 512×512 and input into the second-layer generator. The final prediction completion result of the algorithm is obtained through the prediction of the second-layer generator. For the convenience of representation, in this implementation, the final output result of the stacked generative adversarial network is simplified as

[0057] This application optimizes the generator structure in the stacked generative adversarial network algorithm to adapt to the spectrum situation map completion problem. The generator network structure is as follows Figure 5 As shown, each generator consists of three convolutional layers and three transposed convolutional layers. Figure 5 Conv(k,s,p) and TransConv(k,s,p) respectively represent the convolution layer and the transposed convolution layer and the hyper parameters involved in constructing the layer: convolution kernel size k, step size s, and zero padding number p. In addition to the first-layer generator that only converts the incomplete spectrum situation map into a 3-channel feature input, the generator stacked on this basis merges the output results of the previous layer with the incomplete spectrum situation map into a 6-channel feature map input network.

[0058] The discriminator network structure is as follows Figure 6 As shown in the figure, it consists of 3 convolutional layers and a fully connected layer. The hyperparameter settings of the convolutional layer are the same as those of the generator. The 3-channel feature maps of the predicted generated image and the real spectrum situation are used as the discriminator input, and the corresponding discriminator network discrimination results are output.

[0059] In step 202 of the above implementation, in a complex electromagnetic environment with multiple heterogeneous signal sources, the spatial spectrum situation is obtained by superimposing the wireless transmissions performed simultaneously by all relevant heterogeneous networks in the region. Therefore, the mapping relationship between spatial position, frequency and time and power spectrum density is expressed as φ(x, y, f, t), where x and y are the position coordinates of the spectrum monitoring unit in region S when sensing the spatial spectrum situation, and f and t represent the detection frequency and detection time of spectrum sensing, respectively. The global spectrum situation is a collection of the situations of each coordinate point, so the spatial situation in the region can be expressed as: P = {φ i (x i ,y i ,f,t)} i＝1,2,3... In this embodiment, since the dynamic construction of the regional spectrum situation over time is not considered, the time variable t in the spectrum situation is ignored in subsequent studies.

[0060] When spectrum detection equipment performs spectrum sensing, the spectrum situation in the adjacent area is slightly different, so it can be considered that the smaller range around the detection location has the same power spectrum density as the detection point. Based on this feature, each task area S with a length and width of L×W is divided into N l ×N wOn this basis, φ(x,y,f) is used to represent the position of the object at (x,y) (where x∈[1,2,...,N l ]; y∈[1,2,...,N m ]) The gridded spectrum situation map represents the spatial spectrum situation, which is not only closer to the actual situation of spectrum sensing UAVs obtaining spectrum situation, but also helps to divide the incomplete position and known part of the spectrum situation information in the mission area when completing the incomplete spectrum situation map.

[0061] In order to obtain reliable electromagnetic spectrum information in the mission area, such as Figure 7 As shown in the figure, a drone swarm consisting of M perception drones is deployed to a "data blank" mission area S with T heterogeneous radio transmitters and without any prior spectrum situation information. The perception drones can reach the designated location to perform spectrum detection and upload detection data under the control of the cluster head drone, and collaborate to complete the electromagnetic spectrum situation sampling in the S area. In this process, Indicates that it is located The local power spectrum obtained by UAV i (i=1,2,...,M) is given by Indicates at location The power spectrum of the transmitter j (j = 1, 2, ..., T) at the location of the drone is the result of the superposition of the radio signals emitted by all relevant transmitters and Gaussian noise. Therefore, the relationship between the perceived local power spectrum and the transmitter power spectrum can be expressed as:

[0062]

[0063] in, Indicates the transmitter position (p j ,q j ) to the position of the sensing drone i (x i ,y i )’s propagation path gain; is a correlation function, which is affected by the distance between transmitter j and sensing drone i and the transmitter’s transmission strategy. If the signal power of the spectrum f at the location of drone i is affected by transmitter j, then Otherwise the related function In the formula, σ 2 is the additive white Gaussian noise in this area.

[0064] In order to give full play to the perception capability of the drone swarm, at most one perception drone is deployed in each grid. By visualizing the power spectrum data, a similar Figure 8The spatial spectrum situation map of area S is shown. Due to the limitation of the number of drones, there are still many areas where spectrum sensing has not been performed, and only incomplete spectrum situation maps to be completed can be obtained. The power spectrum in the grid without spectrum sensing is defined as {φ′(x′,y′,f)}. Since the detected position and the unknown part are continuous in space, there is an inevitable connection between the spectrum situation of the detected position and the spectrum situation of the unsensed area, and a series of complex relationships such as the propagation and attenuation of the electromagnetic field in space can be uniformly expressed as:

[0065]

[0066] in, is a relationship function, through which the power spectrum of an unknown location can be calculated from a known power spectrum. To accurately predict the spectrum situation in an unknown area, it is necessary to master an accurate relationship function, but this function contains many factors such as path loss and channel gain. It takes a huge amount of time to derive and calculate the relationship function to complete the incomplete spectrum situation map, and it is difficult to ensure the accuracy of the relationship function. The spectrum situation map prediction and completion algorithm based on the stacked generative adversarial network can learn from the real data accumulated in the past and mine the relationship functions hidden in a large amount of data. This enables low-error, high-definition spectrum situation map completion.

[0067] As an artificial intelligence algorithm, deep learning can use the pre-acquired complete spectrum situation map as a real sample data set and transform the random latent variables z~p through training. z (z) is transformed into the real sample s~p real (s) The same sample s′. The generative adversarial network (GANs) algorithm in deep learning is selected to realize the prediction and completion of the incomplete spectrum situation map obtained by the sensing drone swarm. It consists of the generator network G(z,θ g ) and the discriminator network D(x,θ d ), when training with the complete spectrum situation map dataset, by continuously optimizing the generator network parameters θ g Make the generator output the result Data distribution The distribution p close to the actual spectrum situation map real (s). The discriminator network and the generator network are trained alternately. The discriminator updates the parameters θ in this process. d Improve the ability to distinguish between the complete spectrum situation map s input or the spectrum situation map generated by the generator Whether it obeys the true distribution p real (s). After multiple iterations, the discriminator gradually grasps the real data distribution p real(s), improving the ability to distinguish real samples from deceptive samples. The generator is competing with the discriminator, and the distribution of the prediction completion result data gradually converges to the distribution of the real data, so as to obtain a prediction generation situation map that is closer to the actual spectrum situation. Finally, the generator and the discriminator reach a dynamic balance, and the prediction completion result approaches The discriminator cannot distinguish the authenticity of the input content. At this time, the generator can generate a result G(z,θ) that is almost the same as the real data s. g ).

[0068] In order to use the incomplete spectrum situation map obtained by the drone swarm as known prior information, directly generate a complete global spectrum situation map through the generative adversarial network algorithm to achieve the prediction and completion of the incomplete part, it is necessary to Fig. 9 The training is performed under the algorithm structure shown in the figure, and the network parameters θ of the generator and the discriminator are updated using the gradient descent method under the guidance of the objective function. g ,θ g .

[0069] In order to use the pre-acquired complete spectrum situation map to train the network, a masking function c(s) is added to preprocess the training data: the complete spectrum situation map retains the image of some grids according to a certain sampling rate α, and artificially loses the image data in the grid with a probability of (1-α), simulating the situation where some locations cannot obtain spectrum information through ubiquitous perception drones. α is directly related to the number of drones in the drone swarm and the amount of effective spectrum situation information obtained, that is: the more drones there are, the more complete the regional spectrum situation information obtained by perception, and the closer the α value is to 1. During training, the incomplete spectrum situation map s′=c(s) processed by the masking function is used as the input of the generator. The incomplete spectrum map provides spectrum situation information for some locations. As a generator, it generates the completion result G(s′,θ g ) prior information. When training the discriminator, the completion result G(s′,θ g ) and the corresponding actual spectrum situation map s form a "sample-label" pair and input the discriminator D(x,θ d ) is used for judgment, and it is expected that the actual spectrum situation map will be judged to be consistent with the real distribution, that is: D(s i ,θ d )=1; the prediction completion result G(s′,θ g ) is accurately identified as a false distribution D(G(s i ′,θ g ),θ d )=0.

[0070] In the training process of step 203 of the above embodiment, a suitable loss function is needed to guide the generator network parameters θ g and the discriminator network parameters θd Reverse update. For the generator to maximize the discriminator output result D(G(s i ′,θ g ),θ d ) is the goal, that is, to generate results that are similar to the real distribution to deceive the discriminator. The discriminator needs to minimize the gap with the expected discrimination result to accurately distinguish between real data and false data. In summary, the objective function of the algorithm can be expressed as:

[0071]

[0072] There is usually instability in the training process of generative training. This is because the randomly initialized neural network parameters θ make the distribution of the predicted completion results almost overlap with the real data. When the parameters are updated in reverse, the gradient explodes and the network model cannot be effectively optimized. In addition, the discriminator is generally easier to train, which causes the discriminator to accurately determine that the generator's predicted completion results are false data after fewer iterative training, making it difficult for the generator to continue to optimize and update parameters. By borrowing WGAN-GP instead of the general GAN ​​structure and adding Wasserstein distance as the cost function, problems such as unstable training process can be solved. The objective function of the generative adversarial network is:

[0073]

[0074] Among them, G represents the generator, D represents the discriminator, V(G,D) is the target function of the generative adversarial network, s represents the complete spectrum situation map, and p real (s) represents the distribution of spectrum situation data in the complete spectrum situation map, For D(s,θ d ) takes the expected value, D(s,θ d ) represents the output obtained by inputting the complete spectrum situation map into the discriminator, θ d represents the parameters of the discriminator, is the prediction completion result generated by the generator, represents the spectrum situation data distribution of the prediction completion results generated by the generator, Denotes -D(G(s′,θ g ),θ d ) takes the expected value, G(s′,θ g ) represents the prediction completion result obtained by inputting the incomplete spectrum situation map s′ into the generator, θ g is the parameter of the generator, D(G(s′,θ g ),θ d ) represents G(s′,θ g ) is the output obtained by inputting the discriminator, β represents the penalty coefficient, represents the sample of the gradient penalty term, which is obtained by multiplying the complete spectrum situation map by the random number x and the predicted completion result by (1-x), x~U(0,1), where U(0,1) is a function that takes a random number in the range of 0 to 1. Express Take the expected value, Represents the output obtained by inputting the gradient penalty sample into the discriminator.

[0075] Based on the above-mentioned generative adversarial network objective function, the specific implementation method of step 203 in the above-mentioned embodiment is:

[0076] Based on the training set, a generative adversarial network objective function is adopted to train the generative adversarial network to obtain a trained generative adversarial network.

[0077] Based on the training set, the conditional generative adversarial network is trained using the trained generative adversarial network and the generative adversarial network objective function to obtain a trained conditional generative adversarial network.

[0078] Connect the trained GAN and the trained CGN to get the trained stacked GAN model.

[0079] As the algorithm is iterated, the network parameters are gradually optimized. Make the prediction complete Distribution The distribution p that approaches the real spectrum situation real (s).

[0080] In step 204 of the above implementation, the prediction and completion of the incomplete spectrum situation map obtained by the sensing drone group is realized. Specifically, the information in the incomplete image is used as the algorithm input to directly output the global spectrum situation map. This process completes the incomplete position through the implicit model obtained by neural network training (i.e., the trained stacked generative adversarial network model). The incomplete image is downsampled through the generator convolution layer, and the number of feature map channels is increased to extract the spatial spectrum situation feature information. Then, the high-dimensional feature map is gradually restored to three dimensions using transposed convolution and recombined into a JPG image.

[0081] The stacked generative adversarial network structure built by the above method can use the incomplete spectrum situation map as the algorithm input and directly obtain the predicted and completed global spectrum situation map. The algorithm cannot directly copy the situation image converted from this part of information to the corresponding position of the generated result, but uses it as input condition information to guide the generation of the prediction map. In order to reduce the error generated in the algorithm prediction process, it is only necessary to use the predicted spectrum situation information {φ′(x′,y′,f)} to supplement the vacant position of the original incomplete spectrum situation map. This method of splicing the prediction results is closer to practical application and reduces the error caused by unnecessary prediction.

[0082] In order to verify the effect of the above implementation, the present application also provides the following simulation experiment.

[0083] The effectiveness of the above method is verified by a simulation experiment based on the prediction and completion of the incomplete spectrum situation map based on the stacked generative adversarial network. The simulation experiment uses python 3.6 and builds the network model under the pytorch 1.12.2 framework, and accelerates the algorithm training and testing through the Nvidia GeForce 3090 graphics processor.

[0084] The experiment requires obtaining a complete spectrum situation map dataset used for network training. The WinProp2020 electromagnetic simulation software is used to simulate the radio environment with multiple radio transmitter mission scenarios. The spectrum situation map obtained is as follows: Fig.10 Other electromagnetic simulation related parameters are shown in Table 1.

[0085] In the simulation environment, the three-dimensional spatial position, transmission power and frequency of each transmitter are variable, and an omnidirectional antenna is used to broadcast radio signals to the surrounding area. Each complete spectrum situation map is a global sampling of the signal power size of a fixed detection frequency point at a specified height at a certain moment.

[0086] Table 1 Spectrum situation map simulation parameters

[0087] index Parameter Value Area range (km) 10×10 Sampling height (m) 15 Transmit power(W) 0.1～10 Sampling frequency (GHz) 2.4 Simulation accuracy (m) 20 Image resolution (PPI) 512×512

[0088] In order to meet the needs of algorithm training, 3000 complete spectrum situation maps were constructed using electromagnetic simulation software in this simulation experiment. During algorithm training, 10 complete spectrum situation maps {s1, s2, ..., s 10} to form a batch, and then retain the image of some grids at a sampling rate of 50% through the masking function c(·) to simulate the lack of spectrum perception in this area by the ubiquitous sensing drone swarm, resulting in the loss of spectrum situation information. The embodiment of the present application uses adaptive moment estimation (Adaptive moment estimation, Adam) and sets the learning rate to lr = 0.0001 to achieve reverse update of network parameters. In addition, 200 complete spectrum situation maps are taken as test sets for alternating testing during the training process to avoid overfitting in training.

[0089] first, Fig.11 The prediction and completion results of the first-level generator G1 and the second-level generator G2 of the incomplete spectrum situation map with different numbers of signal sources are demonstrated when the sampling rate is 30% using the stacked generative adversarial network prediction and completion algorithm proposed in this application. From the generation results, it can be seen that although the first-level generator G1 has achieved the prediction and completion of most of the incomplete positions, the prediction effect of some positions is poor. The predicted image near the emission source position is relatively blurred and the details of the spectrum situation map are not fully restored. The final output obtained by further prediction and completion by the second-level generator G2 Compared with the first-level results, it is clearer and almost consistent with the real spectrum situation map.

[0090] In order to more accurately evaluate the prediction and completion effect of the algorithm, the mean squared error (MSE) is used to accurately calculate the gap between the predicted and completed spectrum situation map and the actual spectrum situation map. The mean squared error calculation formula is:

[0091]

[0092] First, the mean square error of the prediction and completion results of the incomplete spectrum situation map in the test set by stacking different levels of generative adversarial networks during the training process is compared to verify whether the stacking method effectively reduces the prediction error of the algorithm completion results. By calculating the MSE value between the prediction and completion results of the incomplete spectrum situation map and the real spectrum situation map of the algorithm and GAN provided in this application at different sampling rates, the influence of the sampling rate on the prediction and completion effect of the algorithm is further tested when the spectrum monitoring drone group has different numbers of perception drones.

[0093] Fig.12 As shown in the figure, as the number of iterative training increases, the second stage of the stacked generative adversarial network (stage-2GAN) has a lower mean square error in the interleaved test, and the completion result is more accurate than the first stage (stage-1GAN). Fig.13The mean square error between the completion results and the real situation map of the three completion methods at a sampling rate of 10% to 90% is demonstrated. Among them, the first method is the stacked generative adversarial network algorithm proposed in this application to directly output the completion result, the second method is to use the generation result corresponding to the incomplete position and the incomplete spectrum situation map to obtain a global spectrum situation map, and the third method uses a non-stacked general GAN ​​to achieve prediction completion. It can be seen that the use of a stacked generative adversarial network has a better completion effect, and a completion result with a smaller error can be obtained at each sampling rate. Splicing the perceived incomplete spectrum situation with the predicted completion part can further reduce the error caused by the unnecessary prediction process.

[0094] In order to test the relationship between the number of stacking layers of the stacked generative adversarial network algorithm and the prediction completion effect, the optimal number of stacking layers for different sampling rates is obtained. Fig.14 The structural similarity (SSIM) of the prediction results of the algorithm and the real data of the three-layer stacked generative adversarial network at sampling rates of 20%, 50%, and 80% is compared. It can be seen that increasing the number of stacked layers can improve the prediction accuracy and the quality of the generated image to a certain extent, but it will also cause the number of algorithm parameters to increase exponentially. Therefore, when the sampling rate is low or the accuracy requirement is high, the number of algorithm stacking layers can be appropriately increased. When the sampling rate is high or the accuracy requirement is low, fewer stacking layers should be used to reduce the complexity of the algorithm.

[0095] The above experiment simulates the "dispersed" deployment of perception drones in the mission area, and obtains an incomplete spectrum situation map by sensing the radio signal power at a certain time and at a specified detection frequency point to test the prediction and completion effect of the algorithm provided in this application. Considering that the spectrum situation changes slowly, the drone perception time can be appropriately extended. Each drone can perceive a larger range of spectrum situation information around its location or perceive the electromagnetic environment on the flight trajectory during movement, so as to further reduce the number of drones required to obtain the spectrum situation. Fig.15 The incomplete spectrum situation maps obtained under two perception modes are shown. From the generated results, it can be seen that the algorithm provided by this application can also predict and complete the incomplete positions well, among which, Fig.15 The left picture in (a) and (b) is the incomplete spectrum situation map obtained by through sampling. Fig.15 The right picture in (a) and (b) is a schematic diagram of the completion result corresponding to the left picture. Fig.15 The left picture in (c) and (d) is the incomplete spectrum situation map obtained by hovering sampling. Fig.15 The right pictures in (c) and (d) are schematic diagrams of the completion results corresponding to the left pictures.

[0096] Based on the above simulation experiments, it can be known that the above method considers using spectrum monitoring drone swarms in the spatial spectrum situation awareness of the mission area, using the high dynamics of drones to quickly obtain power spectrum density information in some areas, and using the stacked generative adversarial network algorithm to complete and accurately predict the complete electromagnetic spectrum situation. Through simulation experiments, the prediction and completion effect of the algorithm was verified and the optimal number of stacking layers for different degrees of incompleteness was evaluated. In a variety of incomplete sample completion experiments, it was proved that the method provided in this application can effectively predict and complete the incomplete spectrum situation map obtained by the drone swarm under different sampling methods. Later, we considered fusing three-dimensional terrain data, and directly predicting the spectrum situation map of the mission area based only on information such as transmitter location and transmission power, to provide accurate spectrum situation information for spectrum management and control.

[0097] In an exemplary embodiment, a device for constructing a spectrum situation map of a drone swarm is provided, comprising:

[0098] Model building module for building stacked generative adversarial network models.

[0099] A training set construction module is used to construct a training set; the training set includes multiple training samples, the input in the training samples is an incomplete spectrum situation map obtained by masking the complete spectrum situation map according to a preset sampling rate, and the label in the training samples is the complete spectrum situation map.

[0100] The model training module is used to train the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model.

[0101] The completion module is used to complete the local spectrum situation map obtained by the drone swarm based on the trained stacked generative adversarial network model.

[0102] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig.16As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data determined by complementarity. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for constructing a spectrum situation map of a drone swarm is implemented.

[0103] Those skilled in the art will understand that Fig.16 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0104] In an exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps in the above-mentioned embodiment of the method for constructing a spectrum situation map of a drone swarm are implemented.

[0105] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, which, when executed by a processor, implements the steps in the above-mentioned method for constructing a spectrum situation map of a drone swarm.

[0106] In an exemplary embodiment, a computer program product is provided, including a computer program, which, when executed by a processor, implements the steps in the above-mentioned method for constructing a spectrum situation map of a drone swarm.

[0107] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0108] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, etc., but is not limited thereto.

[0109] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for constructing a spectrum situation map of drone swarms, It is characterized in that The method for constructing a spectrum situation map of a drone swarm comprises: Build a stacked generative adversarial network model; Constructing a training set; the training set includes a plurality of training samples, the input in the training sample is an incomplete spectrum situation map obtained by masking the complete spectrum situation map according to a preset sampling rate, and the label in the training sample is the complete spectrum situation map; Training the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model; The local spectrum situation map obtained by the drone swarm is completed based on the trained stacked generative adversarial network model.

2. The method for constructing a spectrum situation map of a drone swarm according to claim 1, It is characterized in that The stacked generative adversarial network model includes: a generative adversarial network of the first layer and a conditional generative adversarial network of the second layer; The generative adversarial network and the conditional generative adversarial network both include a generator and a discriminator; The input of the generator in the generative adversarial network is a target spectrum situation map; the target spectrum situation map is an incomplete spectrum situation map or a partial spectrum situation map; The input of the generator in the conditional generative adversarial network is a spliced ​​spectrum situation map; the spliced ​​spectrum situation map is spliced ​​by the first prediction completion result output by the generator in the generative adversarial network and the target spectrum situation map.

3. The method for constructing a spectrum situation map of a drone swarm according to claim 2, It is characterized in that The generator includes three convolutional layers and three transposed convolutional layers, and the discriminator includes three convolutional layers and three fully connected layers.

4. The method for constructing a spectrum situation map of a drone swarm according to claim 2, It is characterized in that Training the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model specifically includes: Based on the training set, a generative adversarial network objective function is used to train the generative adversarial network to obtain a trained generative adversarial network; Based on the training set, training the conditional generative adversarial network using the trained generative adversarial network and the generative adversarial network objective function to obtain a trained conditional generative adversarial network; Connect the trained GAN and the trained CGN to get the trained stacked GAN model.

5. The method for constructing a spectrum situation map of a drone swarm according to claim 4, It is characterized in that The objective function of the generative adversarial network is: Among them, G represents the generator, D represents the discriminator, V(G,D) represents the target function of the generative adversarial network, s represents the complete spectrum situation map, and p real (s) represents the distribution of spectrum situation data in the complete spectrum situation map, For D(s,θ d ) takes the expected value, D(s,θ d ) represents the output obtained by inputting the complete spectrum situation map into the discriminator, θ d represents the parameters of the discriminator, is the prediction completion result generated by the generator, represents the spectrum situation data distribution of the prediction completion results generated by the generator, Denotes -D(G(s′,θ g ),θ d ) takes the expected value, G(s′,θ g ) represents the prediction completion result obtained by inputting the incomplete spectrum situation map s′ into the generator, θ g is the parameter of the generator, D(G(s′,θ g ),θ d ) represents G(s′,θ g ) is the output obtained by inputting the discriminator, β represents the penalty coefficient, represents the sample of the gradient penalty term, which is obtained by multiplying the complete spectrum situation map by the random number x and the predicted completion result by (1-x), x~U(0,1), where U(0,1) is a function that takes a random number in the range of 0 to 1. Express Take the expected value, Represents the output obtained by inputting the gradient penalty sample into the discriminator.

6. A device for constructing a spectrum situation map of a drone swarm, It is characterized in that The drone swarm spectrum situation map construction device applies the drone swarm spectrum situation map construction method according to any one of claims 1 to 5, and the drone swarm spectrum situation map construction device comprises: Model building module, used to build stacked generative adversarial network models; A training set construction module is used to construct a training set; the training set includes a plurality of training samples, the input in the training samples is an incomplete spectrum situation map obtained by masking the complete spectrum situation map according to a preset sampling rate, and the label in the training samples is the complete spectrum situation map; A model training module, used to train the stacked generative adversarial network model based on the training set to obtain a trained stacked generative adversarial network model; The completion module is used to complete the local spectrum situation map obtained by the drone swarm based on the trained stacked generative adversarial network model.

7. A computer device, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for constructing a spectrum situation map of a drone swarm according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the method for constructing a spectrum situation map of a drone swarm described in any one of claims 1 to 5 is implemented.

9. A computer program product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the method for constructing a spectrum situation map of a drone swarm described in any one of claims 1 to 5 is implemented.

Images