WO2022206149A1 - 基于生成对抗网络的三维频谱态势补全方法及装置 - Google Patents

基于生成对抗网络的三维频谱态势补全方法及装置 Download PDF

Info

Publication number
WO2022206149A1
WO2022206149A1 PCT/CN2022/073723 CN2022073723W WO2022206149A1 WO 2022206149 A1 WO2022206149 A1 WO 2022206149A1 CN 2022073723 W CN2022073723 W CN 2022073723W WO 2022206149 A1 WO2022206149 A1 WO 2022206149A1
Authority
WO
WIPO (PCT)
Prior art keywords
network
dimensional
spectrum situation
defect
generative adversarial
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2022/073723
Other languages
English (en)
French (fr)
Inventor
黄洋
朱秋明
胡田钰
吴启晖
龚志仁
吴璇
仲伟志
毛开
张小飞
陆逸炜
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to US18/032,573 priority Critical patent/US12430817B2/en
Publication of WO2022206149A1 publication Critical patent/WO2022206149A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/382—Monitoring; Testing of propagation channels for resource allocation, admission control or handover
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—Two-dimensional [2D] image generation
    • G06T11/10—Texturing; Colouring; Generation of textures or colours
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/90—Determination of colour characteristics
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]

Definitions

  • the invention relates to the technical field of wireless communication, in particular to a three-dimensional spectrum situation completion method based on a generative confrontation network and a three-dimensional spectrum situation completion device based on a generative confrontation network.
  • Spectrum situation completion is a technology that completes the comprehensive situation of the entire electromagnetic environment in the spatial dimension through discrete and defective spectrum data (such as received signal power) obtained by sampling.
  • Applications such as spectrum access, spectrum management and control, improve the utilization of spectrum resources, and try to overcome the problem that the wireless communication system can only obtain the distribution of received signal power in a local area rather than the entire target area through limited radio monitoring equipment.
  • Limited radio monitoring equipment can only get an incomplete picture of the current state of the electromagnetic environment.
  • the UAV tracks designed by these studies are very random and complex, and the tracks in the actual measurement process are required to be consistent with the designed tracks. It is difficult to realize the actual measurement, so it is difficult for the pilot to control the UAV to achieve the track; even limited by the environmental influence, it is difficult to keep the programmed automatic piloting consistent.
  • the designed completion algorithm also needs to know several complete three-dimensional spectrum situations, and this prior information is also difficult to obtain in practical situations.
  • the traditional two-dimensional spectrum situation completion algorithm such as the inverse distance weighting (IDW) algorithm
  • IDW inverse distance weighting
  • the traditional two-dimensional spectrum situation completion algorithm can also be extended to three-dimensional, but this algorithm assumes that the spectrum data is only related to distance, which is not in line with the physical model of the spectrum situation.
  • the 3D spectrum situation completion accuracy is low when the algorithm is used. Therefore, in practice, when the UAV flies on a designated track to sample 3D spectrum situation data, the previous 3D spectrum situation completion scheme will face challenges such as difficulty in obtaining a complete 3D spectrum situation and very little effective data.
  • the State Intellectual Property Office published an invention patent with the publication number CN112819082A, titled “A Deep Learning-Based Satellite Spectrum Sensing Data Reconstruction Method", which can reconstruct low-resolution spectrum data into High-resolution spectrum data, thereby effectively reducing the amount of data transmitted from the satellite to the ground, reducing the pressure of data transmission between the satellite and the ground.
  • the data reconstruction step is to use historical high-resolution data for model training, which is difficult to apply in the absence of prior laws.
  • the purpose of the present invention is to provide a three-dimensional spectrum situation completion method and device based on generative adversarial network. Serious deficiencies, lack of complete three-dimensional spectrum situation, etc.
  • the present invention can obtain the complete three-dimensional distribution of the power spectral density through the defect rather than the complete historical or empirical spectrum data, that is, without the need for the complete historical or empirical spectrum data, and further can over-rely on the priori of the traditional spectrum situation completion method. Information, inability to achieve 3D completion, and low completion accuracy are effectively resolved.
  • an embodiment of the present invention provides a three-dimensional spectrum situation completion method based on a generative adversarial network.
  • the three-dimensional spectrum situation completion method includes:
  • the preprocessing of grayscale and coloring is carried out to obtain the three-channel defect 3D spectrum situation diagram represented by color and the defect 3D spectrum situation based on the obtained three-channel defect represented by color.
  • the three-dimensional spectrum situation map forms a training set
  • a generative adversarial network is trained, and a trained generator network in the generative adversarial network is obtained, wherein the generative adversarial network is configured with a sampling processing function, and the sampling processing function is used to: The sampled position points and the unsampled position points in the three-channel defect three-dimensional spectrum situation diagram of the generative adversarial network expressed in color, and the complementary three-dimensional color of the three-channel output by the generator network in the generative adversarial network.
  • the spectrum situation map is sampled;
  • the preprocessing of grayscale and coloring is performed to obtain the three-channel measured defect three-dimensional frequency spectrum situation map represented by color, and the three-channel measured defect three-dimensional frequency spectrum situation is obtained.
  • the measured defect three-dimensional spectrum situation map is used as the input data of the generator network, and the measured and complemented three-dimensional spectrum situation map of three channels in color is obtained through the generator network.
  • the three-channel complementary three-dimensional spectrum situation map is sampled, including:
  • the numerical matrix includes a first specified numerical column at the sampling position point and a second specified numerical column at the unsampled position point, and the first specified numerical column is different from the second specified numerical column;
  • a Hadamard product of the numerical matrix and the complementary three-dimensional spectral profile of the three channels represented in color is calculated.
  • the preprocessing of grayscale and colorization includes:
  • the defect three-dimensional spectrum situation map of one channel represented by grayscale is determined.
  • the preprocessing of grayscale and colorization also includes:
  • the input data to the generator network is configured as a three-channel defect 3D spectral map represented in color, and
  • the output data of the generator network serves as a three-channel complementary three-dimensional spectral map in color.
  • the generator network includes an input side 3D convolution layer, a downsampling module, a residual module, an upsampling module and an output side 3D convolution layer;
  • the input data of the generator network is sequentially processed by the input side 3D convolution layer, the downsampling module, the residual module, the upsampling module and the output side 3D convolution layer of the generator network;
  • the residual module of the generator network has atrous convolutional layers.
  • the input data of the discriminator network is configured as a three-channel defective three-dimensional spectral situation map represented in color, or a sampled three-channel complemented three-dimensional spectral situation map.
  • the discriminator network includes an input side 3D convolution layer, a downsampling module, a residual module and an output side 3D convolution layer;
  • the input data of the discriminator network is sequentially processed by the input side 3D convolution layer, the downsampling module, the residual module and the output side 3D convolution layer of the discriminator network;
  • the residual module of the discriminator network has atrous convolutional layers.
  • the hyperparameter set of the residual module of the discriminator network is configured to input the output data of the specified number of channels to the output-side three-dimensional convolutional layer of the discriminator network;
  • the output-side 3D convolutional layer of the discriminator network has a controlled set of hyperparameters.
  • the objective function of the generative adversarial network is configured with a gradient penalty term, configured without latent variables, and also configured as a reconstruction loss with weights,
  • the calculation of the reconstruction loss of the weight includes the calculation of the Hadamard product of the weight and the three-channel defect three-dimensional spectrum situation diagram represented by color, the Hadamard product calculation of the weight and the complement three-dimensional spectrum situation map of the three channels represented by color, And the one-norm calculation of the difference between two Hadamard products.
  • the training of the generative adversarial network includes:
  • a three-channel penalized three-dimensional spectrum situation map represented in color is calculated from the real sample and the sampled simulated sample, where,
  • the simulated sample after the sampling processing is obtained after the real sample is processed by the generator network and sampled by the sampling processing function.
  • the input data of the discriminator network is also configured as the penalty three-dimensional spectrum situation map, and the output data of the discriminator network is represented by a matrix;
  • the calculation of the gradient penalty term includes the calculation of the gradient of the output data of the discriminator network with respect to the penalized three-dimensional spectral situation map and the calculation of the two-norm of the gradient.
  • the training of the generative adversarial network further includes:
  • the gradient of the parameter of the objective function and/or the reconstruction loss is calculated.
  • An embodiment of the present invention provides a three-dimensional spectrum situation completion device based on a generative adversarial network.
  • the three-dimensional spectrum situation completion device includes:
  • the preprocessing module is used to perform grayscale and colorization preprocessing based on the defect 3D spectral situation obtained by the UAV sampling the target area, and obtain a three-channel defect 3D spectral situation map represented by color and based on the obtained color.
  • the defect three-dimensional spectrum situation map of the three channels represented forms a training set;
  • the completion module is used to perform grayscale and colorization preprocessing based on the measured defect three-dimensional spectrum situation obtained by the UAV sampling the designated area to obtain a three-channel measured defect three-dimensional spectrum situation map represented by color, and Using the three-channel measured defect three-dimensional spectrum situation map as the input data of the generator network, the three-channel measured and complemented three-dimensional frequency spectrum situation map represented by color is obtained through the generator network.
  • an embodiment of the present invention provides an electronic device, and the electronic device includes:
  • a memory connected to the at least one processor
  • the memory stores instructions that can be executed by the at least one processor, the at least one processor implements the foregoing by executing the instructions stored in the memory, and the at least one processor executes the instructions stored in the memory.
  • an embodiment of the present invention provides a computer-readable storage medium storing computer instructions, which, when the computer instructions are executed on a computer, cause the computer to execute the foregoing method.
  • FIG. 1 is a schematic diagram of a main processing flow of a three-dimensional spectrum situation completion method based on a generative adversarial network according to an embodiment of the present invention
  • FIG. 2 is a schematic diagram of an exemplary residual module of nested hole convolution according to an embodiment of the present invention
  • FIG. 3 is a schematic diagram of an output architecture of a generative adversarial network 3D-SSCGAN for 3D spectrum situation completion according to an embodiment of the present invention
  • FIG. 4 is a schematic diagram of a generator network architecture of a generative adversarial network 3D-SSCGAN for 3D spectrum situation completion according to an embodiment of the present invention
  • FIG. 5 is a schematic diagram of a discriminator network architecture of a generative adversarial network 3D-SSCGAN for 3D spectrum situation completion according to an embodiment of the present invention.
  • generative (formula) adversarial network as a new type of deep learning framework has a very broad application and development space.
  • the generative adversarial network is based on the idea of game theory, and can output generated data (or "fake" data) that can be compared to the real training data through iterative adversarial training of a pair of neural networks.
  • a pair of neural networks are the generator network and the discriminator network, respectively.
  • the generator network With the help of training data obtained from real data, the generator network generates simulated samples with the same specifications as the real samples, and the discriminator network distinguishes the simulated samples from the real samples by measuring the difference between the two. Thereby, higher-order correlations of observed or visible data can be captured without target class label information. Therefore, in the embodiment of the present invention, the generative adversarial network can be applied to the completion of the 3D spectrum situation to learn the data distribution of the real and complete 3D spectrum situation or the mapping relationship between the defective 3D spectrum situation and the complete 3D spectrum situation.
  • the spectrum situation may refer to the received signal power distribution data corresponding to the spatial location points and frequencies of the electromagnetic environment, the spectrum situation may be converted into a graph, and the spectrum situation may be a radio frequency map.
  • the designated track may be some easily realized track, such as a drone flying around at different altitudes.
  • the embodiment of the present invention provides a three-dimensional spectrum situation completion method based on generative adversarial network, and the three-dimensional spectrum situation completion method may include:
  • the foregoing method may be regarded as performing steps in a sequence of a preprocessing stage, an offline training stage, and an online deployment stage. Further, the aforementioned method first enters the preprocessing stage (which can be regarded as the first stage) to perform step S1), then enters the offline training stage (which can be regarded as the second stage) and performs step S2), or enters the online deployment stage (which can be regarded as the second stage). For the third stage), go to step S3). Specific steps are as follows:
  • the data preprocessing method based on grayscale and coloring is performed on several defective 3D spectral situations of the target area sampled by the UAV, so as to improve the training speed of the generative adversarial network and the completion of the 3D spectral situation. performance.
  • the normalized value of each defect three-dimensional spectrum situation is calculated, and the normalized value is directly used as the color value in the specified color mode, for example, the gray value or gray value is specified at this time. grade.
  • the defect three-dimensional spectrum situation diagram of one channel can be obtained by conversion, which can be represented by color channels (RGB, YVU, etc.).
  • the color in one channel is gray color, and gray color can include gray, black and white.
  • the defect three-dimensional spectrum situation map of one channel represented by gray color can be called the “grayscale map” of the defect three-dimensional spectrum situation of one channel.
  • the historical or empirical spectrum data of the defect can be obtained by sampling and collecting the target area A by the drone, and the three-dimensional spectrum situation of the defect is obtained.
  • the target area A may have a specified area size.
  • the color can be a color value or a numerical value range in a specified color mode; in order to facilitate the expression of the color information of the frequency spectrum graph obtained by conversion, the following notation is adopted in the embodiment of the present invention:
  • the defect three-dimensional spectrum situation map of one channel represented by gray color is recorded as the defect three-dimensional spectrum situation "gray map" of one channel;
  • Defect three-dimensional spectrum situation map of three channels represented in color is recorded as "color map" of defect three-dimensional spectrum situation of three channels;
  • the three-channel complementary three-dimensional spectrum situation map represented in color is recorded as the three-channel complementary three-dimensional spectral situation "color map";
  • the measured defect three-dimensional spectrum situation map of one channel represented by gray color is recorded as the measured defect three-dimensional spectrum situation "gray map" of one channel;
  • the three-channel measured and completed three-dimensional spectrum situation map represented in color is recorded as the three-channel measured and completed three-dimensional spectrum situation "color map", etc.
  • the unsampled position points in the "grayscale map" of the three-channel defected 3D spectral situation can be colored in red (or other colors other than black, white and gray) can be used to obtain the three-channel defected 3D spectral situation" color map", and abbreviated as Therefore, in the preprocessing stage, a number of "color maps" of the three-channel defect 3D spectral situation represented in color can be obtained, and several "color maps" of the three-channel defect three-dimensional spectral situation represented in color can be used as a training set. Among them, red or other colors other than black, white and gray are also the color values in the specified color mode.
  • the generative adversarial network is trained through the sampling processing, so as to realize the unsupervised learning of the mapping relationship between the defective 3D spectral situation map and the complete 3D spectral situation map.
  • the generative adversarial network used in this stage in the embodiment of the present invention may be the three-dimensional spectrum situation completion generative adversarial network (i.e. Three dimensional Spectrum Situation Completion Generative Adversarial Network, 3D-SSCGAN) proposed in the embodiment of the present invention.
  • the input data of the generator network in 3D-SSCGAN is a three-channel defective three-dimensional spectral situation "color map"
  • the output data is a three-channel complementary three-dimensional spectrum situation "color map”
  • the input of the discriminator network in 3D-SSCGAN is a three-channel defective 3D spectral situation "color map” Or the "color map” of the complete defect three-dimensional spectrum situation after sampling processing Among them, the sampling processing function F s ( ⁇ ) is a "color map” based on the defect three-dimensional spectrum situation of the three channels input at this time.
  • the output data of the discriminator network can be a matrix that can represent the discrimination result. Therefore, the real samples and simulated samples of 3D-SSCGAN are the three-channel defect three-dimensional spectrum situation "color map” and the three-channel complement defect three-dimensional spectrum situation "color map” respectively, namely, and Therefore, through this sampling process, the generator network and the discriminator network can achieve unsupervised learning through adversarial training, and then obtain a trained generator network.
  • the trained generator network can be directly deployed in the spectrum situation application of the actual three-dimensional electromagnetic spectrum space, and the defective three-dimensional spectrum situation obtained by the UAV can be accurately completed, and then the target area B (The received power distribution data of the actual measured designated area).
  • the target area B may be the target area A, or may be another area having the area size that is substantially the same as the aforementioned area size/in a scaling relationship.
  • the input data of the generator network is the preprocessed three-channel measured defect three-dimensional spectral situation "color map”
  • the output data is the three-channel measured complement three-dimensional spectral situation "color map”.
  • the three-dimensional spectrum situation completion generative adversarial network of the embodiment of the present invention namely 3D-SSCGAN
  • the generative adversarial network composed of layers has the following characteristics:
  • the residual module is introduced to increase the number of 3D convolutional layers in the generative adversarial network, so as to enhance the feature extraction ability of the generative adversarial network for the defective 3D spectrum situation, and then deal with the known defective 3D spectrum situation with very little effective data. challenge.
  • the residual module constructs the information paths between different 3D convolutional layers through residual connections to avoid network degradation caused by the increase of 3D convolutional layers.
  • Atrous convolution is introduced to increase the receptive field of the 3D convolution layer of the generative adversarial network to make up for the loss of information caused by the reduction of data size during the downsampling process, and then cooperate with the unsupervised learning realized in the offline training stage to deal with the complete 3D The challenge of the unknown spectrum situation (but to complete the missing three-dimensional spectrum situation).
  • the hole convolution layer is a neural network convolution layer with a specified hole rate.
  • the hole convolution layer when the hole rate dc > 1, the convolution operation is performed.
  • the hole rate dc 1 corresponding to the ordinary convolutional layer. Therefore, when the hole convolution is not used, it means that the hole rate of the convolution layer is all set to 1, so that the receptive field of the convolution layer cannot be enlarged, that is, it is difficult to process a wider range of spectrum situations.
  • PatchGAN patch-oriented
  • PatchGAN means that the discriminator network of the generative adversarial network no longer outputs a real number representing the discrimination result of the discriminator network on the entire input spectrum situation map, but outputs a number of blocks representing the input spectrum situation map by the discriminator network. A matrix of discrimination results. Therefore, when PatchGAN is not used, it means that the discriminator network only outputs a real number, so it is difficult to focus on the local texture information of the input spectrum situation map, but only the overall texture information.
  • PatchGAN can be used to deal with the challenge of very little effective data in the known defect three-dimensional spectrum situation, and then improve the generation ability of the generative adversarial network by improving the discrimination ability.
  • it can be implemented by the discriminator network architecture of the embodiment of the present invention and the hyperparameter set that controls the three-dimensional convolution layer on the output side of the discriminator network.
  • the weighted reconstruction loss specifically refers to:
  • the weighted reconstruction loss mainly uses the L1 loss function to evaluate the three-channel defect and complete the defect three-dimensional spectral situation "color map", that is, and difference between the two.
  • the physical meaning of the weight in the weighted reconstruction loss is: if the received power value at a certain location is larger, the location is more susceptible to the 3D spectrum situation completion problem and the corresponding application of the spectrum situation. The higher the corresponding weight value in the situation "color map”.
  • the weighted reconstruction loss V L1 (G) is calculated using the following formula:
  • the three-dimensional spectrum situation complements the generative adversarial network, that is, the training target of the 3D-SSCGAN is calculated by the following formula:
  • D is the discriminator network
  • ⁇ L1 is the reconstruction loss factor
  • ⁇ is the Wasserstein penalty factor
  • ⁇ d is the discriminator network parameter
  • 2 is the two-norm.
  • V(D,G) is the objective function that characterizes the adversarial nature of the generative adversarial network, that is, the mapping relationship between the defective 3D spectral situation and the complete 3D spectral situation is gradually learned through training based on adversarial loss, and V L1 (G ) is to further strengthen the learning ability by reconstructing the loss, so the sum of V(D, G) and V L1 (G) is the total objective function of 3D-SSCGAN.
  • the ordinary generative adversarial network has a random variable called a latent variable to learn the target distribution, but since the distribution of the three-dimensional spectral situation data is unknown, the random variable is removed from the objective function to better learn the above mapping. relation.
  • the three-channel penalized 3D spectral situation "color map" from real samples and the simulated samples after sampling can be calculated, and its corresponding distribution.
  • the discriminator network aims to discriminate the difference between the three-channel defective 3D spectral situation "color map” or the three-channel complementary 3D spectral situation "color map”, due to the introduced penalty 3D spectral situation "color map” It also needs to be input to the discriminator network, so the input data of the discriminator network can be considered as a sample here or To sum up, the input and output mechanism of the 3D spectrum situation completion generative adversarial network, that is, 3D-SSCGAN, is shown in Figure 3.
  • the three-dimensional spectrum situation complements the generative adversarial network, that is, the specific steps of the training algorithm of 3D-SSCGAN are as follows:
  • the samples in the training set are the three-channel defect three-dimensional spectral situation "color maps" obtained in the preprocessing stage.
  • the described three-dimensional spectrum situation completion generative adversarial network that is, the construction process of the generator network architecture of 3D-SSCGAN is as follows:
  • k ⁇ k represents the length of the single edge of the convolution kernel of the 3D convolution layer or 3D transposed convolution layer
  • s C represents the sliding step size of the convolution kernel
  • p C represents the number of zero-padded layers
  • d C represents the hole convolution Or the dilation rate of ordinary convolutions.
  • the 32-channel output data of the above-mentioned convolutional layer will be input to two sequentially connected three-dimensional convolutional layers whose hyperparameter set h is (4, 2, 1, 1), thereby performing two data size reductions. Small feature extraction with double the number of channels F. This part is called the downsampling module.
  • the output data of the 128 channels of the above-mentioned down-sampling module will be input into the six sequentially connected residual modules as shown in Fig. feature.
  • the hyperparameter sets h of the two three-dimensional convolutional layers in the residual module are (3,1,2r, 2r ) and (3,1,1,1) respectively, where r is the residual The index of the module.
  • each 3D convolutional layer will keep the data size and number of channels unchanged.
  • the 128-channel output data of the last residual module will be input into two sequentially connected three-dimensional transposed convolutional layers whose hyperparameter set h is (4, 2, 1, 1).
  • the 32-channel output data of the above-mentioned upsampling module will be input to a three-dimensional convolutional layer whose hyperparameter set h is (5, 1, 2, 1), so that the last data size remains unchanged and the number of channels F decreases. Feature extraction and data fine-tuning as small as 3. As a result, the generator network of 3D-SSCGAN outputs a three-channel complementary spectral situation "color map".
  • FIG. 4 the schematic diagram of the generator network architecture of the three-dimensional spectral situation completion generative adversarial network (ie 3D-SSCGAN) is shown in Figure 4.
  • each 3D convolutional layer and 3D transposed convolutional layer are activated with a linear rectifier unit (ie, ReLU function).
  • the described three-dimensional spectrum situation completion generative adversarial network that is, the construction process of the discriminator network architecture of 3D-SSCGAN is as follows:
  • the basic structure of the discriminator network is set to be the same as the first half of the generator network. Therefore, the input samples of the discriminator network will sequentially pass through a 3D convolutional layer (the set of hyperparameters h is (5, 1, 2, 1)), followed by two sequentially connected 3D convolutional layers (the set of hyperparameters h is (4, 2, 1, 1)) and three sequentially connected residual modules as shown in Figure 2 (where the hyperparameter sets h of the two three-dimensional convolution layers are (3, 1) ,4 r ,4 r ) and (3,1,1,1)) for feature extraction.
  • a 3D convolutional layer the set of hyperparameters h is (5, 1, 2, 1)
  • two sequentially connected 3D convolutional layers the set of hyperparameters h is (4, 2, 1, 1)
  • three sequentially connected residual modules as shown in Figure 2 (where the hyperparameter sets h of the two three-dimensional convolution layers are (3, 1) ,4 r ,4 r ) and (3,1,1,1)
  • the 128-channel output data of the last residual module above will be input to a 3D convolutional layer with a hyperparameter set h of (5, 1, 2, 1) to implement a block-oriented generative adversarial network structure.
  • the discriminator network of 3D-SSCGAN outputs a certain matrix representing the discrimination result. Therefore, the schematic diagram of the discriminator network architecture of the three-dimensional spectral situation completion generative adversarial network (ie 3D-SSCGAN) is shown in Figure 5, and the linear rectifier unit is still used for activation.
  • the embodiment of the present invention can complete the generative adversarial network (that is, the aforementioned three-dimensional spectrum situation) through a number of defective three-dimensional spectral situations in historical or empirical spectral data, data preprocessing operations based on grayscale and colorization, and sampling processing of sampling processing functions.
  • 3D-SSCGAN realizes unsupervised learning of the mapping relationship between the defected 3D spectrum situation and the complete 3D spectrum situation, does not require prior complete historical or empirical spectrum data, and avoids UAVs in the track specified or unmanned It is difficult to complete the 3D spectrum situation with very little valid data when specified, and can cope with the unknown complete 3D spectrum situation (but the defect 3D spectrum situation needs to be completed) and the known defective 3D spectrum situation with very little valid data. Challenges, and then deployed in actual spectrum situation applications to achieve high-precision display of received power distribution.
  • the 3D-SSCGAN embodiment of the present invention can realize the completion of the defect three-dimensional spectrum situation with the situation completion accuracy far superior to the traditional IDW method under the actual situation that the UAV is sampling with the designated or undesignated track, and can Deployment is performed in the actual three-dimensional electromagnetic spectrum space under the condition that the random distribution of the received power values in the actual deployment environment is the same as or different from the random distribution of the received power values in the training set.
  • the UAV performs flight and sampling under the specified track, which can be set as follows: the UAV circles 16 times at different heights at equal intervals of 5m with a random radius, and then passes the radio on it.
  • the monitoring equipment samples the spectrum data.
  • the radiation sources in the target area A, or the number of main users NT are randomly selected from 1 to 5, and the sampling rate ⁇ 3.8 % under the UAV's designated track.
  • test 10000 groups of defective 3D spectral situations respectively
  • the two parts of data are preprocessed based on grayscale, and the resulting sets are called training set and test set, respectively.
  • the test set is used to simulate or refer to the actual deployment environment of the proposed 3D spectrum situation completion algorithm.
  • the training data or test data can be obtained by sampling several times of the UAV in a certain 3D electromagnetic spectrum space, so it can be completely distributed in the training data, that is, training
  • 3D-SSCGAN 3D spectral situation completion generative adversarial network
  • the mean square error is used to evaluate the situational completion performance of each set of training/testing data, and the calculation formula is:
  • training error and test error refer to the average MSE of all training and testing data in the training and testing sets, respectively.
  • the results of the spectrum situation completion of the test data compare the proposed algorithm and The test error of the traditional IDW algorithm at different training rounds. It can be seen from the comparison results that regardless of the training data setting, the 3D spectral situation completion performance of the proposed algorithm gradually converges with the increase of the number of training rounds, and the test error during convergence is much smaller than that of the IDW algorithm. test error. Therefore, the situational completion performance of this algorithm is improved compared with the traditional IDW algorithm.
  • 3D-SSCGAN 3D spectral situational completion generative adversarial network
  • 3D-CompareGAN 3D-CompareGAN
  • structure in which atrous convolution and PatchGAN structures are no longer used. Therefore, the offline training and online deployment phases of the proposed algorithm are performed based on the 3D-SSCGAN and 3D-CompareGAN structures, respectively, and the average MSE of the two structures at different training epochs is compared, where the random distribution of the received power values p t_s ⁇ p d_s .
  • the proposed 3D-SSCGAN structure can effectively improve the training stability and completion performance of the 3D spectrum situation completion algorithm.
  • the embodiment of the present invention and implementations 1 and 2 all belong to the same inventive concept.
  • the embodiment of the present invention provides a three-dimensional spectrum situation completion device based on a generative adversarial network.
  • the three-dimensional spectrum situation completion device may include:
  • the preprocessing module is used to perform grayscale and colorization preprocessing based on the defect 3D spectral situation obtained by the UAV sampling the target area, and obtain a three-channel defect 3D spectral situation map represented by color and based on the obtained color.
  • the defect three-dimensional spectrum situation map of the three channels represented forms a training set;
  • the completion module is used to perform grayscale and colorization preprocessing based on the measured defect three-dimensional spectrum situation obtained by the UAV sampling the designated area to obtain a three-channel measured defect three-dimensional spectrum situation map represented by color, and Using the three-channel measured defect three-dimensional spectrum situation map as the input data of the generator network, the three-channel measured and complemented three-dimensional frequency spectrum situation map represented by color is obtained through the generator network.
  • the aforementioned modules may be implemented in digital electronic circuits, systems of integrated circuits, field programmable gate arrays (FPGA), application specific integrated circuits (ASIC), application specific standard products (ASSP), system on chip system (SoC), load programmable logic device (CPLD), computer hardware, firmware, software, and/or combinations thereof.
  • sampling processing function in the training module is configured to be specifically used for:
  • the numerical matrix includes a first specified numerical column at the sampling position point and a second specified numerical column at the unsampled position point, and the first specified numerical column is different from the second specified numerical column;
  • a Hadamard product of the numerical matrix and the complementary three-dimensional spectral profile of the three channels represented in color is calculated.
  • the preprocessing of grayscale and colorization in the preprocessing module includes:
  • the defect three-dimensional spectrum situation map of one channel represented by grayscale is determined.
  • the preprocessing of grayscale and colorization in the preprocessing module further includes:
  • the defect three-dimensional frequency spectrum situation diagram of the described one channel, and the defect three-dimensional frequency spectrum situation diagram obtained by copying twice, are expanded to the defect three-dimensional frequency spectrum situation diagram of the three channels represented by the gray color;
  • the input data to the generator network is configured as a three-channel defect 3D spectral map represented in color, and
  • the output data of the generator network serves as a three-channel complementary three-dimensional spectral map in color.
  • the generator network includes an input side 3D convolution layer, a downsampling module, a residual module, an upsampling module and an output side 3D convolution layer;
  • the input data of the generator network is sequentially processed by the input side 3D convolution layer, the downsampling module, the residual module, the upsampling module and the output side 3D convolution layer of the generator network;
  • the residual module of the generator network has atrous convolutional layers.
  • the input data of the discriminator network is configured as a three-channel defective three-dimensional spectral situation map represented in color, or a sampled three-channel complemented three-dimensional spectral situation map.
  • the discriminator network includes an input side 3D convolution layer, a downsampling module, a residual module and an output side 3D convolution layer;
  • the input data of the discriminator network is sequentially processed by the input side 3D convolution layer, the downsampling module, the residual module and the output side 3D convolution layer of the discriminator network;
  • the residual module of the discriminator network has atrous convolutional layers.
  • the hyperparameter set of the residual module of the discriminator network is configured to input the output data of the specified number of channels to the output-side three-dimensional convolutional layer of the discriminator network;
  • the output-side 3D convolutional layer of the discriminator network has a controlled set of hyperparameters.
  • the objective function of the generative adversarial network is configured with a gradient penalty term, configured without latent variables, and also configured as a reconstruction loss with weights,
  • the calculation of the reconstruction loss of the weight includes the calculation of the Hadamard product of the weight and the three-channel defect three-dimensional spectrum situation diagram represented by color, the Hadamard product calculation of the weight and the complement three-dimensional spectrum situation map of the three channels represented by color, And the one-norm calculation of the difference between the two Hadamard products.
  • the training of the generative adversarial network includes:
  • a three-channel penalized three-dimensional spectrum situation map represented in color is calculated from the real sample and the sampled simulated sample, where,
  • the simulated sample after the sampling processing is obtained after the real sample is processed by the generator network and sampled by the sampling processing function.
  • the input data of the discriminator network is also configured as the penalty three-dimensional spectrum situation map, and the output data of the discriminator network is represented by a matrix;
  • the calculation of the gradient penalty term includes the calculation of the gradient of the output data of the discriminator network with respect to the penalized three-dimensional spectral situation map and the calculation of the two-norm of the gradient.
  • the training of the generative adversarial network further includes:
  • the gradient of the parameter of the objective function and/or the reconstruction loss is calculated.
  • the embodiments of the present invention only use defects instead of complete historical or empirical spectrum data to train the generative adversarial network, which overcomes the problem that the complete three-dimensional spectrum situation is difficult to obtain accurately.
  • the embodiment of the present invention performs the actual sampling setting of the spectrum data of the three-dimensional target area based on the designated or undesignated track instead of the random chaotic track, which overcomes the excessively long flight time of the drone equipped with radio monitoring equipment. Or the deployment is difficult and other constraints.
  • the generative adversarial network can learn without any complete 3D spectrum situation. It shows the distribution of received power, that is, to further improve the completion performance of the three-dimensional spectrum situation.
  • the three-dimensional spectrum situation completion algorithm proposed in the embodiment of the present invention can solve the challenges of unknown complete 3D spectrum situation and very little effective data. In this way, the stable extraction of environmental features of the 3D electromagnetic spectrum space is strengthened, and the loss of useful information in the 3D electromagnetic spectrum space is reduced. Then, guided by the importance of the radiation source, the mapping relationship between the defect three-dimensional spectrum situation and the complete three-dimensional spectrum situation is learned.
  • the embodiment of the present invention realizes the offline training of the generative adversarial network for the completion of the three-dimensional spectrum situation in an unsupervised learning manner according to the sampling processing function. Furthermore, the algorithm can be deployed in the actual three-dimensional electromagnetic spectrum space when the random distribution of the received power value in the actual deployment environment is the same as or different from the random distribution of the received power value in the training set, so as to achieve accurate three-dimensional spectrum situation completion.
  • An embodiment of the present invention further provides an electronic device, the electronic device includes: at least one processor; a memory connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, The at least one processor implements the aforementioned method by executing the memory-stored instructions by the at least one processor executing the memory-stored instructions.
  • Embodiments of the present invention further provide a computer-readable storage medium storing computer instructions, which, when executed on a computer, cause the computer to execute the foregoing method.
  • the aforementioned storage medium may be non-transitory, and the storage medium may include: U disk, hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), flash memory (Flash memory), Various media that can store program codes, such as a magnetic disk or an optical disk.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Electromagnetism (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)
  • Testing And Monitoring For Control Systems (AREA)
  • Monitoring And Testing Of Transmission In General (AREA)

Abstract

本发明提供一种基于生成对抗网络的三维频谱态势补全方法及装置，属于无线通信技术领域。所述方法包括：基于无人机对目标区域采样获得的缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的缺损三维频谱态势图以及以该缺损三维频谱态势图形成训练集；基于训练集，训练生成对抗网络，获得生成对抗网络中训练后的生成器网络；基于无人机对实测的指定区域采样获得的实测缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的实测缺损三维频谱态势图，并将该实测缺损三维频谱态势图作为生成器网络的输入数据，经生成器网络获得以彩色表示的三通道的实测补全三维频谱态势图。本发明可用于频谱态势补全。

Description

基于生成对抗网络的三维频谱态势补全方法及装置
相关申请的交叉引用
本申请要求2021年03月30日提交的中国专利申请202110345357.3的权益，该申请的内容通过引用被合并于本文。
技术领域
本发明涉及无线通信技术领域，具体地涉及一种基于生成对抗网络的三维频谱态势补全方法和一种基于生成对抗网络的三维频谱态势补全装置。
背景技术
频谱态势补全是一种通过采样得到的离散且缺损频谱数据(如接收信号功率)来对整个电磁环境的综合形势在空间维度上进行补全的技术，该技术可以使得无线通信系统有的放矢地实现频谱接入、频谱管控等应用，提升频谱资源利用率，以尝试克服无线通信系统只能通过有限的无线电监测设备来获取局部区域而非整个目标区域的接收信号功率分布情况的问题，即尝试克服有限的无线电监测设备只能得到不完整的电磁环境当前状态的问题。
在以往的频谱态势补全方案中，大多考虑二维平面，并且存在补全精度低、鲁棒性差、环境受限大等缺点。但是，随着我国空天地信息网络一体化的不断发展，二维平面内的频谱态势补全已很难满足需求，而以多种异构无人机为载体的天地之间机动网络下的三维频谱态势补全的研究愈发受到重视，由此需要针对空天地信息网络在三维电磁频谱空间下的频谱资源利用情况进行分析和探索。
尽管已有少数研究针对三维频谱态势补全展开，但这些研究所设计的无人机航迹十分杂乱随机、复杂，并要求实际测量过程中的航迹与设计的航迹保持一致性，这在实际测量时是很难实现的，因此实际情况下飞手难以操控无人机来实现该航迹；甚至受限于环境影响，编程自动引航也很难保持一致。同时，设计的补全算法还需要已知若干完整三维频谱态势，该先验信息在实际情况中同样难以获得。需要说明的是，传统二维频谱态势补全算法，如反距离加权(IDW)算法也可以扩展至三维，但该算法假设频谱数据仅与距离相关，这与频谱态势的物理模型较为不符合，进而导致在使用该算法时三维频谱态势补全精度较低。因此，实际情况中，当无人机以指定的航迹，来飞行进而采样三维频谱态势数据时，以往的三维频谱态势补全方案将面临完整三维频谱态势难以获得、有效数据极少等挑战。
国家知识产权局于2019年05月17日公开了一件公开号为CN106682234A、名称为“一种基于空间插值的电磁频谱分布预测和动态可视化方法”的发明专利，其对来自网格化监测的数据按照时间粒度聚合，对每一个时间段内的监测数据使用Kriging空间插值算法，拟合监测站点间距离与电磁信号强度的统计关系，对未知区域的电磁频谱分布数据进行最优无偏插值预测，但在数据量不足的情况中，很难得以应用。
国家知识产权局于2019年06月28日公开了一件公开号为CN109946512A、名称为“一种改进频域插值的动态功率分析方法”的发明专利，通过使用快速傅里叶变换、计算频谱插值系数等方法降低了非同步采样或数据的非整数周期阶段引起的误差，但存在插值结果精度低的问题。
国家知识产权局于2021年05月18日公开了一件公开号为CN112819082A、名称为“一种基于深度学习的卫星频谱感知数据重构方法”的发明专利，可将低分辨率频谱数据重建为高分辨率频谱数据，从而有效降低星地传输数据量，减缓星地间数据传输压力。在该方案中，数据重建步骤是利用历史高分辨率数据进行模型训练，该方法很难在先验规律缺失的情况下得以应用。
发明内容
本发明的目的是提供一种基于生成对抗网络的三维频谱态势补全方法及装置，本发明面向无人机以指定或未指定航迹进行采样时的三维频谱态势补全场景，解决了采样数据严重不足、完整三维频谱态势缺失等问题。本发明可以通过缺损而非完整的历史或经验频谱数据，即不需要完整的历史或经验频谱数据，来得到完整的功率谱密度三维分布，进而能对传统频谱态势补全方法的过度依赖先验信息、无法实现三维补全、补全精度低下等缺陷进行有效解决。
为了实现上述目的，本发明实施例提供一种基于生成对抗网络的三维频谱态势补全方法，该三维频谱态势补全方法包括：
基于无人机对目标区域采样获得的缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的缺损三维频谱态势图以及基于获得的以彩色表示的三通道的缺损三维频谱态势图形成训练集；
基于所述训练集，训练生成对抗网络，获得所述生成对抗网络中训练后的生成器网络，其中，所述生成对抗网络被配置有采样处理函数，所述采样处理函数用于：基于输入所述生成对抗网络的以彩色表示的三通道的缺损三维频谱态势图中采样位置点和未采样位置点，对 由所述生成对抗网络中生成器网络输出的以彩色表示的三通道的补全三维频谱态势图进行采样处理；
基于无人机对实测的指定区域采样获得的实测缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的实测缺损三维频谱态势图，并将所述三通道的实测缺损三维频谱态势图作为所述生成器网络的输入数据，经所述生成器网络获得以彩色表示的三通道的实测补全三维频谱态势图。
具体的，基于输入所述生成对抗网络的以彩色表示的三通道的缺损三维频谱态势图中采样位置点和未采样位置点，对由所述生成对抗网络中生成器网络输出的以彩色表示的三通道的补全三维频谱态势图进行采样处理，包括：
确定数值矩阵，所述数值矩阵包括采样位置点处的第一指定数值列和未采样位置点处的第二指定数值列，所述第一指定数值列与所述第二指定数值列不同；
计算所述数值矩阵和以彩色表示的三通道的补全三维频谱态势图的哈达玛积。
具体的，其中灰度化和着色的预处理，包括：
基于缺损的历史或经验频谱数据，对缺损三维频谱态势进行归一化值计算；
以该归一化值确定以灰度色表示的一通道的缺损三维频谱态势图。
具体的，其中灰度化和着色的预处理，还包括：
将所述一通道的缺损三维频谱态势图复制两次；
将所述一通道的缺损三维频谱态势图，以及复制两次得到的缺损三维频谱态势图，扩充为以所述灰度色表示的三通道的缺损三维频谱态势图；
对以所述灰度色表示的三通道的缺损三维频谱态势图中未采样位置点，以指定颜色进行着色，获得以彩色表示的三通道的缺损三维频谱态势图，所述彩色包括所述灰度色和所述指定颜色。
具体的，其中，所述生成对抗网络中，
生成器网络的输入数据被配置为以彩色表示的三通道的缺损三维频谱态势图，且
所述生成器网络的输出数据作为以彩色表示的三通道的补全三维频谱态势图。
具体的，其中，所述生成对抗网络中，
所述生成器网络包括输入侧三维卷积层、下采样模块、残差模块、上采样模块和输出侧三维卷积层；
所述生成器网络的输入数据依次经所述生成器网络的输入侧三维卷积层、下采样模块、残差模块、上采样模块和输出侧三维卷积层处理；
所述生成器网络的残差模块具有空洞卷积层。
具体的，其中，所述生成对抗网络中，
鉴别器网络的输入数据被配置为以彩色表示的三通道的缺损三维频谱态势图、或者采样处理后的三通道的补全三维频谱态势图。
具体的，其中，所述生成对抗网络中，
所述鉴别器网络包括输入侧三维卷积层、下采样模块、残差模块和输出侧三维卷积层；
所述鉴别器网络的输入数据依次经所述鉴别器网络的输入侧三维卷积层、下采样模块、残差模块和输出侧三维卷积层处理；
所述鉴别器网络的残差模块具有空洞卷积层。
具体的，其中，所述生成对抗网络中，
所述鉴别器网络的残差模块的超参数集被配置为用于将指定数量的通道数的输出数据，输入至所述鉴别器网络的输出侧三维卷积层；
所述鉴别器网络的输出侧三维卷积层具有受控制的超参数集合。
具体的，其中，所述生成对抗网络中，
所述生成对抗网络的目标函数被配置为具有梯度惩罚项、被配置为不具有潜变量且还被配置为具有权重的重建损失，
所述权重的重建损失的计算包括权重与以彩色表示的三通道的缺损三维频谱态势图的哈达玛积计算、权重与以彩色表示的三通道的补全三维频谱态势图的哈达玛积计算、以及两哈达玛积之差的一范数计算。
具体的，所述训练生成对抗网络，包括：
按照配置的训练参数，从所述训练集中采样真实样本；
基于瓦瑟斯坦距离，由所述真实样本和采样处理后的模拟样本计算得到以彩色表示的三通道的惩罚三维频谱态势图，其中，
所述采样处理后的模拟样本为所述真实样本经所述生成器网络处理和所述采样处理函数采样处理后得到的。
具体的，其中，所述生成对抗网络中，
所述鉴别器网络的输入数据还被配置为所述惩罚三维频谱态势图，且所述鉴别器网络的输出数据以矩阵表示；
所述梯度惩罚项的计算包括所述鉴别器网络的输出数据关于所述惩罚三维频谱态势图的梯度的计算以及该梯度的二范数的计算。
具体的，所述训练生成对抗网络，还包括：
基于所述真实样本中采样位置点和未采样位置点的具体分布，确定采样处理函数的表达式；
通过所述表达式和所述训练集，计算所述目标函数和/或所述重建损失的参数的梯度。
本发明实施例提供一种基于生成对抗网络的三维频谱态势补全装置，该三维频谱态势补全装置包括：
预处理模块，用于基于无人机对目标区域采样获得的缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的缺损三维频谱态势图以及基于获得的以彩色表示的三通道的缺损三维频谱态势图形成训练集；
训练模块，用于基于所述训练集，训练生成对抗网络，获得所述生成对抗网络中训练后的生成器网络，其中，所述生成对抗网络被配置有采样处理函数，所述采样处理函数用于：基于输入所述生成对抗网络的以彩色表示的三通道的缺损三维频谱态势图中采样位置点和未采样位置点，对由所述生成对抗网络中生成器网络输出的以彩色表示的三通道的补全三维频谱态势图进行采样处理；
补全模块，用于基于无人机对实测的指定区域采样获得的实测缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的实测缺损三维频谱态势图，并将所述三通道的实测缺损三维频谱态势图作为所述生成器网络的输入数据，经所述生成器网络获得以彩色表示的三通道的实测补全三维频谱态势图。
再一方面，本发明实施例提供一种电子设备，该电子设备包括：
至少一个处理器；
存储器，与所述至少一个处理器连接；
其中，所述存储器存储有能被所述至少一个处理器执行的指令，所述至少一个处理器通过执行所述存储器存储的指令，所述至少一个处理器通过执行所述存储器存储的指令实现前述的方法。
又一方面，本发明实施例提供一种计算机可读存储介质，存储有计算机指令，当所述计算机指令在计算机上运行时，使得计算机执行前述的方法。
本发明实施例的其它特征和优点将在随后的具体实施方式部分予以详细说明。
附图说明
附图是用来提供对本发明实施例的进一步理解，并且构成说明书的一部分，与下面的 具体实施方式一起用于解释本发明实施例，但并不构成对本发明实施例的限制。在附图中：
图1为本发明实施例的基于生成对抗网络的三维频谱态势补全方法主要处理流程示意图；
图2为本发明实施例的一种示例性的嵌套空洞卷积的残差模块示意图；
图3为本发明实施例的三维频谱态势补全生成对抗网络3D-SSCGAN的输出架构示意图；
图4为本发明实施例的三维频谱态势补全生成对抗网络3D-SSCGAN的生成器网络架构示意图；
图5为本发明实施例的三维频谱态势补全生成对抗网络3D-SSCGAN的鉴别器网络架构示意图。
具体实施方式
以下结合附图对本发明实施例的具体实施方式进行详细说明。应当理解的是，此处所描述的具体实施方式仅用于说明和解释本发明实施例，并不用于限制本发明实施例。
伴随着大数据时代以及人工智能的快速发展，深度学习引起了学者们的广泛兴趣，也推动了无线通信与人工智能的有机融合。其中，生成(式)对抗网络作为一种新型的深度学习框架具有非常广阔的应用与发展空间。生成对抗网络基于博弈论思想，能通过对一对神经网络的迭代对抗式训练，输出可比拟真实训练数据的生成数据(或可称“伪造”数据)。其中，一对神经网络分别为生成器网络和鉴别器网络。借助于从真实数据中获得的训练数据，生成器网络生成与真实样本规格相同的模拟样本，而鉴别器网络则通过衡量模拟样本与真实样本间的差异来对二者进行区分。由此，可以在没有目标类标签信息的情况下捕捉观测到或可见数据的高阶相关性。因此，在本发明实施例中，可以将生成对抗网络应用于三维频谱态势补全中，以学习真实完整三维频谱态势的数据分布或缺损三维频谱态势到完整三维频谱态势间的映射关系。
在本发明实施例中，频谱态势可以指与电磁环境的空间位置点和频率对应的接收信号功率分布数据，频谱态势可以转换为以图的方式呈现，频谱态势图可以是射频地图。
在本发明实施例中，指定航迹可以是某些易于实现的航迹，如无人机在不同高度环绕飞行。
实施例1
本发明实施例提供了基于生成对抗网络的三维频谱态势补全方法，该三维频谱态势补 全方法可以包括：
S1)基于无人机对目标区域采样获得的缺损三维频谱态势，进行灰度化和着色的预处理，获得三通道的缺损三维频谱态势以及基于获得的三通道的缺损三维频谱态势形成训练集；
S2)基于所述训练集，训练生成对抗网络，获得所述生成对抗网络中训练后的生成器网络，其中，所述生成对抗网络被配置有采样处理函数，所述采样处理函数用于：基于输入所述生成对抗网络的三通道的缺损三维频谱态势中采样位置点和未采样位置点，对由所述生成对抗网络中生成器网络输出的三通道的补全三维频谱态势进行采样处理；
S3)基于无人机对实测的指定区域采样获得的实测缺损三维频谱态势，进行灰度化和着色的预处理，获得三通道的实测缺损三维频谱态势，并将所述三通道的实测缺损三维频谱态势作为所述生成器网络的输入，经所述生成器网络获得三通道的实测补全三维频谱态势。
在本发明实施例中，结合图1，前述方法可以视为按照预处理阶段、离线训练阶段和在线部署阶段的阶段顺序执行步骤。进一步的，前述方法先进入预处理阶段(可以视为第一阶段)进行步骤S1)，然后，进入离线训练阶段(可以视为第二阶段)进行步骤S2)，或进入线部署阶段(可以视为第三阶段)，进行步骤S3)。具体步骤如下：
在第一阶段中，该阶段对无人机采样得到的目标区域的若干缺损三维频谱态势进行基于灰度化和着色的数据预处理方法，进而来提升生成对抗网络训练速度以及三维频谱态势补全性能。
首先，在已知关于三维目标区域A的缺损的历史或经验频谱数据的情况下，对采样的数据中的每个缺损三维频谱态势
Figure PCTCN2022073723-appb-000001
进行归一化处理，即：
将缺损的历史或经验频谱数据中接收信号功率(值)最大值与该缺损三维频谱态势中的接收信号功率最大值进行比较，并将缺损的历史或经验频谱数据中接收信号功率最小值与该缺损三维频谱态势中的接收信号功率最小值进行比较；
取两最大值中的较大值以及取两最小值中的较小值，较大值和较小值作为该缺损三维频谱态势进行最大最小归一化时的基准值；
通过最大最小归一化函数，计算每个缺损三维频谱态势的归一化值，并将该归一化值直接作为在指定颜色模式下的颜色数值，例如此时指定为灰度值或灰度等级。
因此，可转换得到一通道的缺损三维频谱态势图，可以以颜色通道(RGB、YVU等)表示，此时的一通道内颜色是灰度色，灰度色可以包括灰色、黑色和白色，以灰度色表示的一通道的缺损三维频谱态势图，可以称为一通道的缺损三维频谱态势“灰度图”。需要补充说明的是，缺损的历史或经验频谱数据，可以是，无人机对目标区域A进行采样以及收集， 获得的缺损的三维频谱态势，目标区域A可以有指定的区域尺寸。在本发明实施例中，颜色可以是在指定颜色模式下的颜色数值或数值范围；为了便于转换得到的频谱态势图的颜色信息表达，本发明实施例中采用以下记法：
将以灰度色表示的一通道的缺损三维频谱态势图，记为一通道的缺损三维频谱态势“灰度图”；
将以彩色表示的三通道的缺损三维频谱态势图，记为三通道的缺损三维频谱态势“彩色图”；
将以彩色表示的三通道的补全三维频谱态势图，记为三通道的补全三维频谱态势“彩色图”；
将以灰度色表示的一通道的实测缺损三维频谱态势图，记为一通道的实测缺损三维频谱态势“灰度图”；
将以彩色表示的三通道的实测缺损三维频谱态势图，记为三通道的实测缺损三维频谱态势“彩色图”；以及
将以彩色表示的三通道的实测补全三维频谱态势图，记为三通道的实测补全三维频谱态势“彩色图”等。
其次，可以将该一通道的缺损三维频谱态势“灰度图”复制两次后，通过一通道的缺损三维频谱态势“灰度图”以及复制两次的一通道的缺损三维频谱态势“灰度图”，扩充得到三通道的缺损三维频谱态势“灰度图”。
最后，可以对该三通道的缺损三维频谱态势“灰度图”中的未采样位置点以红色(或也可以使用其他除黑白灰色以外的颜色)进行着色，得到三通道的缺损三维频谱态势“彩色图”，并简记为
Figure PCTCN2022073723-appb-000002
因此，预处理阶段下可得到若干以彩色表示三通道的缺损三维频谱态势“彩色图”，并将若干以彩色表示的三通道的缺损三维频谱态势“彩色图”作为训练集。其中，红色或其他除黑白灰色以外的颜色也是在指定颜色模式下的颜色数值。
在第二阶段中，该阶段基于上述训练集，通过进行的采样处理来对生成对抗网络进行训练，以实现缺损三维频谱态势图与完整三维频谱态势图之间映射关系的无监督学习。需要说明的是，本发明实施例中该阶段使用的生成对抗网络可以是本发明实施例提出的三维频谱态势补全生成对抗网络(即Three dimensional Spectrum Situation Completion GenerativeAdversarial Network，3D-SSCGAN)。
首先，一方面，3D-SSCGAN中的生成器网络的输入数据为三通道的缺损三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000003
而输出数据为三通道的补全三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000004
另一方面，3D-SSCGAN中的鉴别器网络的输入为三通道的缺损三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000005
或采样处理后的补全缺损三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000006
其中采样处理函数F s(·)则是根据此时输入的三通道的缺损三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000007
所展现的无人机采样情况对三通道的补全三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000008
进行采样处理，该采样处理可以是与无人机采样情况相同的采样处理。例如，在本发明实施例公开的一种示例中，输入的三通道的缺损三维频谱态势“彩色图”所展现的无人机采样情况，可以通过定义一个数值矩阵M进行确定和表示，其中采样位置点处的值为[1,1,1] T，未采样位置点处的值为[0,0,0] T(因为此时“彩色图”维度是3×N L×N W×N H，所以每个位置x∈R 3处的值都是一个包含三个元素的向量)。因此，采样处理函数的形式可以为F s(·)＝M⊙·，例如
Figure PCTCN2022073723-appb-000009
其中
Figure PCTCN2022073723-appb-000010
为三通道的补全三维频谱态势“彩色图”，⊙为Hadamand积(即逐点相乘)。在本发明实施例中，采样情况可以理解为采样位置点和未采样位置点的具体分布。
而鉴别器网络的输出数据可为可以表征鉴别结果的某矩阵。因此，3D-SSCGAN的真实样本和模拟样本分别为三通道的缺损三维频谱态势“彩色图”和三通道的补全缺损三维频谱态势“彩色图”，即分别为
Figure PCTCN2022073723-appb-000011
和
Figure PCTCN2022073723-appb-000012
因此，通过该采样处理，生成器网络和鉴别器网络可以通过对抗训练实现无监督学习，进而得到训练完成的生成器网络。
在第三阶段，该阶段可将已训练好的生成器网络直接部署于实际三维电磁频谱空间的频谱态势应用，对无人机得到的缺损三维频谱态势进行精确补全，进而得到目标区域B(实测的指定区域)的接收功率分布数据。其中，目标区域B可以是目标区域A，也可以是具有与前述的区域尺寸基本相同/呈缩放关系的区域尺寸的其他区域。而此时生成器网络的输入数据为预处理后的三通道的实测缺损三维频谱态势“彩色图”，而输出数据为三通道的实测补全三维频谱态势“彩色图”。
结合图2、图3，进一步的实施例中，在第二阶段和第三阶段中，本发明实施例的三维频谱态势补全生成对抗网络，即3D-SSCGAN，相较于由普通三维卷积层构成的生成对抗网络具有如下特征：
1)引入残差模块来增加生成对抗网络中三维卷积层的层数，以此增强生成对抗网络对于缺损三维频谱态势的特征提取能力，进而应对已知缺损三维频谱态势中有效数据极少的挑战。其中，残差模块通过残差连接来构建不同三维卷积层之间的信息通路，以避免三维卷积层增多引起的网络退化。
2)引入空洞卷积来增加生成对抗网络三维卷积层的感受野，以弥补下采样过程中因数据尺寸减小而造成的信息丢失，进而配合离线训练阶段实现的无监督学习来应对完整三维 频谱态势未知(却要对缺损三维频谱态势补全)的挑战。需要说明的是，空洞卷积将嵌套在残差模块中进行使用，如图2所示。空洞卷积层是具有指定空洞率的神经网络卷积层，空洞卷积层：在空洞率d c>1时，进行卷积运算。而普通卷积层对应的空洞率d c＝1。因此，当不使用空洞卷积时，即说明卷积层空洞率全部被设置为1，进而导致卷积层感受野无法变大，即难以对更大范围的频谱态势进行处理。
3)在生成对抗网络中，向鉴别器网络中引入面向块(PatchGAN)的机制，以提取缺损三维频谱态势中面向局部的有效数据特征。其中，PatchGAN是指生成对抗网络的鉴别器网络不再输出表征了鉴别器网络对整个输入频谱态势图的鉴别结果的实数，而是输出表征了鉴别器网络对输入频谱态势图的若干分块的鉴别结果的矩阵。因此，当不使用PatchGAN时，即说明鉴别器网络只输出一个实数，从而难以关注输入频谱态势图的局部纹理信息，而只能关注整体纹理信息。因此，PatchGAN可以用于应对已知缺损三维频谱态势中有效数据极少的挑战，进而通过提升鉴别能力来对抗提升生成对抗网络的生成能力。在本发明实施例公开的一种示例中，可以通过本发明实施例鉴别器网络架构和控制该鉴别器网络输出侧三维卷积层的超参数集实现。
4)在生成对抗网络中，向目标函数中引入带有梯度惩罚项的Wasserstein距离及参数，以缓解生成对抗网络训练过程中由有效数据极少和完整三维频谱态势未知这两大挑战而加剧的训练不稳定性。
5)在生成对抗网络中，向目标函数中引入无潜变量，以学习缺损三维频谱态势与完整三维频谱态势之间的映射关系，进而解决由完整三维频谱态势未知这一挑战而导致的三维频谱态势数据分布无法学习的问题。
6)在目标函数中引入带权重的重建损失，以体现三维频谱态势补全问题及对应频谱态势应用对辐射源的重视程度，进而应对已知缺损三维频谱态势中有效数据极少的挑战。
进一步的实施例中，在引入带权重的重建损失来应对已知缺损三维频谱态势有效数据极少挑战的特征中，所述的带权重的重建损失具体是指：
带权重的重建损失主要是通过L 1损失函数来评价三通道的缺损和补全缺损三维频谱态势“彩色图”，即
Figure PCTCN2022073723-appb-000013
和
Figure PCTCN2022073723-appb-000014
这两者之间的差异。带权重的重建损失中权重的物理意义为：若某位置处的接收功率值越大，则该位置就更易受到三维频谱态势补全问题及对应频谱态势应用的关注，因此该位置在缺损三维频谱态势“彩色图”中的对应权重值就越大。带权重的重建损失V L1(G)利用如下公式计算：
Figure PCTCN2022073723-appb-000015
式中，
Figure PCTCN2022073723-appb-000016
为经采样处理函数得到的三通道的补全缺损三维频谱态势“彩色图”，权重
Figure PCTCN2022073723-appb-000017
为归一化后的三通道的缺损三维频谱态势“彩色图”的权重，p data为真实数据分布，⊙为哈达玛(Hadamard)积，
Figure PCTCN2022073723-appb-000018
为f(x)关于p(x)的期望，G为生成器网络，θ g为生成器网络参数，||·|| 1为一范数。因此，在训练过程中，即可通过最小化V L1(G)来对生成器网络参数θ g进行更新。
进一步的实施例中，所述的三维频谱态势补全生成对抗网络，即3D-SSCGAN的训练目标利用如下公式计算：
Figure PCTCN2022073723-appb-000019
式中，D为鉴别器网络，
Figure PCTCN2022073723-appb-000020
为y关于x的梯度，λ L1为重建损失因子，λ为Wasserstein惩罚因子，θ d为鉴别器网络参数，||·|| 2为二范数。
进一步的，V(D,G)为表征生成对抗网络对抗性的目标函数，即通过基于对抗损失的训练来逐渐学习缺损三维频谱态势与完整三维频谱态势之间的映射关系，而V L1(G)则是通过重建损失来进一步加强该学习能力，所以V(D,G)和V L1(G)之和即为3D-SSCGAN的总目标函数。需要说明的是，普通的生成对抗网络存在潜变量这一随机变量来学习目标分布，但由于三维频谱态势数据分布未知，因此，在目标函数中将该随机变量去除，以更好的学习上述映射关系。
进一步的，基于瓦瑟斯坦Wasserstein距离(描述分布之间的距离)，三通道的惩罚三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000021
由真实样本
Figure PCTCN2022073723-appb-000022
和采样处理后的模拟样本
Figure PCTCN2022073723-appb-000023
计算得到，而
Figure PCTCN2022073723-appb-000024
为其对应分布。具体而言，样本
Figure PCTCN2022073723-appb-000025
由两部分相加组成，其中一部分由真实样本
Figure PCTCN2022073723-appb-000026
乘以一服从均匀分布U(0,1)的随机数ε而得到，另一部分则是由模拟样本
Figure PCTCN2022073723-appb-000027
乘以(1-ε)而得到。需要说明的是，尽管鉴别器网络旨在鉴别三通道的缺损三维频谱态势“彩色图”或三通道的补全三维频谱态势“彩色图”之间的差异，但由于所引入的惩罚三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000028
同样需要输入至鉴别器网络，因此鉴别器网络的输入数据此处可以认为是样本
Figure PCTCN2022073723-appb-000029
或
Figure PCTCN2022073723-appb-000030
综上，所述的三维频谱态势补全生成对抗网络，即3D-SSCGAN的输入输出机制如图3所示。
进一步的实施例中，所述的三维频谱态势补全生成对抗网络，即3D-SSCGAN的训练算法的具体步骤如下：
1)设置训练轮数N epoch、批训练大小M batch、训练集中的样本频谱态势个数N train、鉴别器网络较生成器网络的关替频数N dis。其中，训练集中的样本即预处理阶段得到的若干三通道的缺损三维频谱态势“彩色图”
Figure PCTCN2022073723-appb-000031
2)从训练集中采样真实样本
Figure PCTCN2022073723-appb-000032
从真实样本中获取无人机采样情况，确定函数F s(·)表达式，可以通过前述的指定数值矩阵的方式完成表达式确定，计算目标函数V(D,G)关于参数θ d的梯度，并对θ d进行更新。
3)从训练集中采样真实样本
Figure PCTCN2022073723-appb-000033
从真实样本中获取无人机采样情况，确定函数F s(·)表达式，可以通过前述的指定数值矩阵的方式完成表达式确定，计算目标函数V(D,G)和V L1(G)关于参数θ g的梯度，并对θ g进行更新。
4)循环更新，共更新N dis次。
5)循环迭代，共迭代(N train/N batch)次。
6)循环训练，共训练N epoch轮。
7)循环结束，得到训练好的生成器网络和鉴别器网络。
结合图4，进一步的实施例中，所述的三维频谱态势补全生成对抗网络，即3D-SSCGAN的生成器网络架构搭建过程具体如下：
首先，采用超参数集合h＝(k,s C,p C,d C)为(5,1,2,1)的一个三维卷积层对输入样本，即三通道的缺损三维频谱态势“彩色图”进行一次数据尺寸不变而通道数F增加为32的特征提取和处理。其中k×k表征三维卷积层或三维转置卷积层的卷积核单边长度，s C表征卷积核的滑动步长尺寸，p C表征补零层数，d C表征空洞卷积或普通卷积的空洞率。
进一步的，上述卷积层的32通道的输出数据则会输入至超参数集合h为(4,2,1,1)的两个顺序连接的三维卷积层，由此进行两次数据尺寸减小而通道数F加倍的特征提取。而该部分则被称作下采样模块。
再进一步的，上述下采样模块的128通道的输出数据则将输入至六个如图2所示的顺序连接的残差模块中，即通过连接特征通路来更好地提取极少数据中的有效特征。需要说明的是，残差模块中的两个三维卷积层的超参数集合h分别为(3,1,2 r,2 r)和(3,1,1,1)，其中r为残差模块的索引。同时，每个三维卷积层将维持数据尺寸和通道数保持不变。
再进一步的，上述最后一个残差模块的128通道的输出数据则将输入至超参数集合h 为(4,2,1,1)的两个顺序连接的三维转置卷积层，由此进行两次数据尺寸增大而通道数F减半的恢复处理，即使其输出数据的数据尺寸与输入缺损三维频谱态势“彩色图”)的尺寸保持一致。因此，该部分被称作上采样模块。
最后，上述上采样模块的32通道的输出数据则会输入至超参数集合h为(5,1,2,1)的一个三维卷积层，以进行最后一次数据尺寸不变而通道数F减小至3的特征提取和数据微调。由此，3D-SSCGAN的生成器网络则输出三通道的补全频谱态势“彩色图”。
因此，三维频谱态势补全生成对抗网络(即3D-SSCGAN)的生成器网络架构示意图如图4所示。其中的箭头表示某超参数集合h＝(k,s C,p C,d C)下的三维卷积层或三维转置卷积层，而方框则表示某模块或三维卷积层的输出数据(通道数为F)。此外，每层三维卷积层和三维转置卷积层后均使用线性整流单元(即ReLU函数)进行激活。
结合图5，进一步的实施例中，所述的三维频谱态势补全生成对抗网络，即3D-SSCGAN的鉴别器网络架构搭建过程具体如下：
首先，设定鉴别器网络的基本结构与生成器网络前半部分相同。因此，鉴别器网络的输入样本将依次通过一个三维卷积层(超参数集合h为(5,1,2,1))、由两个顺序连接的三维卷积层(超参数集合h均为(4,2,1,1))构成的下采样模块和三个如图2所示的顺序连接的残差模块(其中的两个三维卷积层的超参数集合h分别为(3,1,4 r,4 r)和(3,1,1,1))来进行特征提取。此外，上述最后一个残差模块的128通道的输出数据则将输入至超参数集合h为(5,1,2,1)的一个三维卷积层，以实现面向块的生成对抗网络结构。由此，3D-SSCGAN的鉴别器网络则输出表征鉴别结果的某矩阵。因此，三维频谱态势补全生成对抗网络(即3D-SSCGAN)的鉴别器网络架构示意图如图5所示，且仍使用线性整流单元进行激活。
本发明实施例可以通过历史或经验频谱数据中的若干缺损三维频谱态势，基于灰度化和着色的数据预处理操作和采样处理函数的采样处理，经三维频谱态势补全生成对抗网络(即前述的3D-SSCGAN)，实现对缺损三维频谱态势和完整三维频谱态势间的映射关系进行无监督学习，不需要先验的完整的历史或经验频谱数据，避免了无人机在航迹指定或未指定时所导致极少的有效数据很难补全三维频谱态势的问题，能够应对完整三维频谱态势未知(却要对缺损三维频谱态势补全)和已知缺损三维频谱态势中有效数据极少的挑战，进而于实际频谱态势应用进行部署，实现接收功率分布的高精度展现。
3D-SSCGAN本发明实施例可以在无人机以指定或未指定航迹进行采样的实际情况下，以远优于传统IDW方法的态势补全精度来实现缺损三维频谱态势的补全，并且可以在实际部署环境中的接收功率值的随机分布与训练集中的接收功率值的随机分布相同或不同的情况下于实际三维电磁频谱空间进行部署。
实施例2
本发明实施例与实施了1属于同一发明构思，本发明实施例公开了实施例1中的缺损三维频谱态势补全方法中的一些示例，可以理解的，以下示例并非是本发明实施例限定的唯一实施方式，可以基于测试、使用的效果，可以进行其他数值选择和实施操作等。下面以三维目标区域A的大小为L×W×H＝240m×240m×80m，并栅格化为N L×N W×N H＝48×48×16个栅格的情况为例。
在上述情况中，无人机以指定航迹下执飞和采样，可以设定为：无人机在等间隔为5m的不同高度以随机半径环绕飞行16圈，进而通过搭载在其上的无线电监测设备进行频谱数据的采样。同时，目标区域A内的辐射源，或说主要用户的数量N T从1～5中随机选择，而无人机指定航迹下的采样率α≈3.8％。并且，计算信号传播损耗时的参考距离d 0＝0.01km，而辐射源和搭载在无人机上的无线电监测设备的工作频率f∈(25MHz,125MHz)。
对于所提基于生成对抗网络的三维频谱态势补全算法，分别设置N train＝30000组和N test＝10000组缺损三维频谱态势
Figure PCTCN2022073723-appb-000034
作为生成对抗网络的训练数据和测试数据。将这两部分数据进行基于灰度的数据预处理，由此得到的集合分别称为训练集和测试集。需要说明的是，测试集用于模拟或代指所提三维频谱态势补全算法的实际部署环境。
进一步的，由于生成对抗网络基于缺损三维频谱态势进行无监督学习，仅需要无人机在某三维电磁频谱空间采样若干次即可得到训练数据或测试数据，因此完全可以在训练数据分布，即训练集中的接收功率值的随机分布p t_s和测试数据分布，即实际部署环境中的接收功率值的随机分布p d_s相同或不同的情况下分别进行算法的性能评估。具体而言，首先设置测试数据中路径损耗因子n＝6和噪声功率σ 2＝10 -5mw。因此，当考虑接收功率值的随机分布p t_s＝p d_s的情况时，训练数据的参数与测试数据将保持一致，即路径损耗因子n＝6和噪声功率σ 2＝10 -5mw。而当考虑接收功率值的随机分布p t_s≠p d_s时，设置训练数据中每10000组数据的参数依次为n＝8,σ 2＝10 -5、n＝4,σ 2＝10 -5和n＝4,σ 2＝10 -3。设定辐射源发射频率的范围f max-f min＝25MHz，而发射功率(单位：dBm)随机从向量[18.3,26.7,29.4,30,30,30,29.4,26.7,18.3]中进行抽取。
进一步的，关于所提出的三维频谱态势补全生成对抗网络(即3D-SSCGAN)，设置 其生成器网络和鉴别器网络的学习率为10000，批训练大小N batch＝16，训练轮数N epoch的最大值设置为100，鉴别器网络较生成器网络的关替频数N dis＝2，惩罚因子λ＝10，而λ L1＝1000。
进一步的，使用均方误差来评价每组训练/测试数据的态势补全性能，其计算公式为：
Figure PCTCN2022073723-appb-000035
其中完整三维频谱态势
Figure PCTCN2022073723-appb-000036
仅用于该均方误差的计算，而不用于所提算法的离线训练和在线部署阶段。此外，训练误差和测试误差则分别代指训练集和测试集中的所有训练数据和测试数据的平均MSE。
首先，将分别在接收功率值的随机分布p t_s和p d_s相同或不同这两种生成对抗网络训练情况下，比较所提基于生成对抗网络的三维频谱态势补全算法与传统IDW算法的补全性能，其中无人机在等间隔高度以指定航迹采样16圈，而IDW算法的幂值参数n p＝2。
从随机选取的某次测试数据下三维频谱态势的补全结果(该次测试时的分布p t_s＝p d_s)可以看出，所提算法的补全三维频谱态势与原完整三维频谱态势相比几乎没有差别，而IDW算法却与之相差较大。即在该示例上，所提算法明显比IDW算法的补全结果要表现的更好。
进一步的，根据传统IDW算法和所提算法在不同生成对抗网络训练情况(即分布p t_s＝p d_s和分布p t_s≠p d_s)下对测试数据的频谱态势补全结果，对比所提算法和传统IDW算法于不同训练轮数时的测试误差。从对比结果可以看出，无论处于何种训练数据设置情况，所提算法的三维频谱态势补全性能均随着训练轮数的增加而逐渐收敛，且收敛时的测试误差均远小于IDW算法的测试误差。因此，本算法下的态势补全性能较传统IDW算法得到了提升。进一步的，当分布p t_s＝p d_s时，算法的测试误差还低于分布p t_s≠p d_s时的对应测试误差。这说明所提算法完全可以在某频谱态势补全应用所在的三维电磁频谱空间内直接进行训练数据收集工作，并能得到更优越的性能。
进一步的，为了评估所提出的三维频谱态势补全生成对抗网络(即3D-SSCGAN)在所提算法中体现的补全性能提升和该结构的优越性，构建了一个名为3D-CompareGAN的对比结构，其中不再使用空洞卷积和PatchGAN结构。因此，分别基于3D-SSCGAN与3D-CompareGAN结构执行所提算法的离线训练和在线部署阶段，并对两种结构于不同训练轮数时的平均MSE进行比较，其中接收功率值的随机分布p t_s≠p d_s。从对比结果可以看出，无论关注于训练误差还是测试误差，基于3D-SSCGAN结构的生成对抗网络性能相对于 3D-CompareGAN结构都要更加稳定，且算法补全误差也与之更低。因此，所提3D-SSCGAN结构能有效提升三维频谱态势补全算法的训练稳定性和补全性能。
实施例3
本发明实施例与实施了1和2均属于同一发明构思，本发明实施例提供了基于生成对抗网络的三维频谱态势补全装置，该三维频谱态势补全装置可以包括：
预处理模块，用于基于无人机对目标区域采样获得的缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的缺损三维频谱态势图以及基于获得的以彩色表示的三通道的缺损三维频谱态势图形成训练集；
训练模块，用于基于所述训练集，训练生成对抗网络，获得所述生成对抗网络中训练后的生成器网络，其中，所述生成对抗网络被配置有采样处理函数，所述采样处理函数用于：基于输入所述生成对抗网络的以彩色表示的三通道的缺损三维频谱态势图中采样位置点和未采样位置点，对由所述生成对抗网络中生成器网络输出的以彩色表示的三通道的补全三维频谱态势图进行采样处理；
补全模块，用于基于无人机对实测的指定区域采样获得的实测缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的实测缺损三维频谱态势图，并将所述三通道的实测缺损三维频谱态势图作为所述生成器网络的输入数据，经所述生成器网络获得以彩色表示的三通道的实测补全三维频谱态势图。需要提出的是，在一些情况中，前述的模块可以在数字电子电路系统、集成电路系统、现场可编程门阵列(FPGA)、专用集成电路(ASIC)、专用标准产品(ASSP)、芯片上系统的系统(SoC)、负载可编程逻辑设备(CPLD)、计算机硬件、固件、软件、和/或它们的组合中实现。
具体的，所述训练模块中采样处理函数被配置为具体用于：
确定数值矩阵，所述数值矩阵包括采样位置点处的第一指定数值列和未采样位置点处的第二指定数值列，所述第一指定数值列与所述第二指定数值列不同；
计算所述数值矩阵和以彩色表示的三通道的补全三维频谱态势图的哈达玛积。
具体的，所述预处理模块中灰度化和着色的预处理，包括：
基于缺损的历史或经验频谱数据，对缺损三维频谱态势进行归一化值计算；
以该归一化值确定以灰度色表示的一通道的缺损三维频谱态势图。
具体的，所述预处理模块中灰度化和着色的预处理，还包括：
将所述一通道的缺损三维频谱态势图复制两次；
将所述一通道的缺损三维频谱态势图，以及复制两次得到的缺损三维频谱态势图，扩 充为以所述灰度色表示的三通道的缺损三维频谱态势图；
对以所述灰度色表示的三通道的缺损三维频谱态势图中未采样位置点，以指定颜色进行着色，获得以彩色表示的三通道的缺损三维频谱态势图，所述彩色包括所述灰度色和所述指定颜色。
具体的，其中，所述生成对抗网络中，
生成器网络的输入数据被配置为以彩色表示的三通道的缺损三维频谱态势图，且
所述生成器网络的输出数据作为以彩色表示的三通道的补全三维频谱态势图。
具体的，其中，所述生成对抗网络中，
所述生成器网络包括输入侧三维卷积层、下采样模块、残差模块、上采样模块和输出侧三维卷积层；
所述生成器网络的输入数据依次经所述生成器网络的输入侧三维卷积层、下采样模块、残差模块、上采样模块和输出侧三维卷积层处理；
所述生成器网络的残差模块具有空洞卷积层。
具体的，其中，所述生成对抗网络中，
鉴别器网络的输入数据被配置为以彩色表示的三通道的缺损三维频谱态势图、或者采样处理后的三通道的补全三维频谱态势图。
具体的，其中，所述生成对抗网络中，
所述鉴别器网络包括输入侧三维卷积层、下采样模块、残差模块和输出侧三维卷积层；
所述鉴别器网络的输入数据依次经所述鉴别器网络的输入侧三维卷积层、下采样模块、残差模块和输出侧三维卷积层处理；
所述鉴别器网络的残差模块具有空洞卷积层。
具体的，其中，所述生成对抗网络中，
所述鉴别器网络的残差模块的超参数集被配置为用于将指定数量的通道数的输出数据，输入至所述鉴别器网络的输出侧三维卷积层；
所述鉴别器网络的输出侧三维卷积层具有受控制的超参数集合。
具体的，其中，所述生成对抗网络中，
所述生成对抗网络的目标函数被配置为具有梯度惩罚项、被配置为不具有潜变量且还被配置为具有权重的重建损失，
所述权重的重建损失的计算包括权重与以彩色表示的三通道的缺损三维频谱态势图的哈达玛积计算、权重与以彩色表示的三通道的补全三维频谱态势图的哈达玛积计算、以及 两哈达玛积之差的一范数计算。
具体的，所述训练生成对抗网络，包括：
按照配置的训练参数，从所述训练集中采样真实样本；
基于瓦瑟斯坦距离，由所述真实样本和采样处理后的模拟样本计算得到以彩色表示的三通道的惩罚三维频谱态势图，其中，
所述采样处理后的模拟样本为所述真实样本经所述生成器网络处理和所述采样处理函数采样处理后得到的。
具体的，其中，所述生成对抗网络中，
所述鉴别器网络的输入数据还被配置为所述惩罚三维频谱态势图，且所述鉴别器网络的输出数据以矩阵表示；
所述梯度惩罚项的计算包括所述鉴别器网络的输出数据关于所述惩罚三维频谱态势图的梯度的计算以及该梯度的二范数的计算。
具体的，所述训练生成对抗网络，还包括：
基于所述真实样本中采样位置点和未采样位置点的具体分布，确定采样处理函数的表达式；
通过所述表达式和所述训练集，计算所述目标函数和/或所述重建损失的参数的梯度。
本发明实施例只利用缺损而非完整的历史或经验频谱数据来训练生成对抗网络，克服了完整三维频谱态势难以准确获得的问题。本发明实施例基于无人机以指定或未指定航迹而非随机杂乱航迹来进行三维目标区域频谱数据的实际采样设置，克服了搭载无线电监测设备的无人机所存在的飞行时间过长或部署难度较大等限制情况。
本发明实施例根据基于生成对抗网络的三维频谱态势补全算法，生成对抗网络可以在无任何完整三维频谱态势的情况下进行学习，即通过对缺损历史或经验频谱数据进行特征提取，来更实际的展现接收功率的分布情况，即进一步提升三维频谱态势的补全性能。本发明实施例在根据面向三维频谱态势补全的生成对抗网络(即3D-SSCGAN)的前提下，所提出的三维频谱态势补全算法在可以在完整三维频谱态势未知和有效数据极少的挑战下，加强对三维电磁频谱空间环境特征的稳定提取，同时减少三维电磁频谱空间有用信息的丢失。进而以辐射源的重视程度为导向，学习缺损三维频谱态势与完整三维频谱态势之间的映射关系。
本发明实施例根据采样处理函数，以无监督学习的方式实现三维频谱态势补全生成对抗网络的离线训练。进而可以在实际部署环境中的接收功率值的随机分布与训练集中的接收功率值的随机分布相同或不同的情况下于实际三维电磁频谱空间进行算法的部署，实现精确 的三维频谱态势补全。
本发明实施例还提供了电子设备，该电子设备包括：至少一个处理器；存储器，与所述至少一个处理器连接；其中，所述存储器存储有能被所述至少一个处理器执行的指令，所述至少一个处理器通过执行所述存储器存储的指令，所述至少一个处理器通过执行所述存储器存储的指令实现前述的方法。本发明实施例还提供计算机可读存储介质，存储有计算机指令，当所述计算机指令在计算机上运行时，使得计算机执行前述的方法。
以上结合附图详细描述了本发明实施例的可选实施方式，但是，本发明实施例并不限于上述实施方式中的具体细节，在本发明实施例的技术构思范围内，可以对本发明实施例的技术方案进行多种简单变型，这些简单变型均属于本发明实施例的保护范围。
另外需要说明的是，在上述具体实施方式中所描述的各个具体技术特征，在不矛盾的情况下，可以通过任何合适的方式进行组合。为了避免不必要的重复，本发明实施例对各种可能的组合方式不再另行说明。
本领域技术人员可以理解实现上述实施例方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，该程序存储在一个存储介质中，包括若干指令用以使得单片机、芯片或处理器(processor)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质可以是非瞬时的，存储介质可以包括：U盘、硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、闪存(Flash memory)、磁碟或者光盘等各种可以存储程序代码的介质。
此外，本发明实施例的各种不同的实施方式之间也可以进行任意组合，只要其不违背本发明实施例的思想，其同样应当视为本发明实施例所公开的内容。

Claims (14)

  1. 一种基于生成对抗网络的三维频谱态势补全方法，其特征在于，该三维频谱态势补全方法包括：
    基于无人机对目标区域采样获得的缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的缺损三维频谱态势图以及基于获得的以彩色表示的三通道的缺损三维频谱态势图形成训练集；
    基于所述训练集，训练生成对抗网络，获得所述生成对抗网络中训练后的生成器网络，其中，所述生成对抗网络被配置有采样处理函数，所述采样处理函数用于：基于输入所述生成对抗网络的以彩色表示的三通道的缺损三维频谱态势图中采样位置点和未采样位置点，对由所述生成对抗网络中生成器网络输出的以彩色表示的三通道的补全三维频谱态势图进行采样处理；
    基于无人机对实测的指定区域采样获得的实测缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的实测缺损三维频谱态势图，并将所述三通道的实测缺损三维频谱态势图作为所述生成器网络的输入数据，经所述生成器网络获得以彩色表示的三通道的实测补全三维频谱态势图。
  2. 根据权利要求1所述的基于生成对抗网络的三维频谱态势补全方法，其特征在于，基于输入所述生成对抗网络的以彩色表示的三通道的缺损三维频谱态势图中采样位置点和未采样位置点，对由所述生成对抗网络中生成器网络输出的以彩色表示的三通道的补全三维频谱态势图进行采样处理，包括：
    确定数值矩阵，所述数值矩阵包括采样位置点处的第一指定数值列和未采样位置点处的第二指定数值列，所述第一指定数值列与所述第二指定数值列不同；
    计算所述数值矩阵和以彩色表示的三通道的补全三维频谱态势图的哈达玛积。
  3. 根据权利要求2所述基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中灰度化和着色的预处理，包括：
    基于缺损的历史或经验频谱数据，对缺损三维频谱态势进行归一化值计算；
    以该归一化值确定以灰度色表示的一通道的缺损三维频谱态势图。
  4. 根据权利要求3所述的基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中灰度化和着色的预处理，还包括：
    将所述一通道的缺损三维频谱态势图复制两次；
    将所述一通道的缺损三维频谱态势图，以及复制两次得到的缺损三维频谱态势图，扩充为以所述灰度色表示的三通道的缺损三维频谱态势图；
    对以所述灰度色表示的三通道的缺损三维频谱态势图中未采样位置点，以指定颜色进行着色，获得以彩色表示的三通道的缺损三维频谱态势图，所述彩色包括所述灰度色和所述指定颜色。
  5. 根据权利要求4所述的基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中，所述生成对抗网络中，
    生成器网络的输入数据被配置为以彩色表示的三通道的缺损三维频谱态势图，且
    所述生成器网络的输出数据作为以彩色表示的三通道的补全三维频谱态势图。
  6. 根据权利要求5所述的基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中，所述生成对抗网络中，
    所述生成器网络包括输入侧三维卷积层、下采样模块、残差模块、上采样模块和输出侧三维卷积层；
    所述生成器网络的输入数据依次经所述生成器网络的输入侧三维卷积层、下采样模块、残差模块、上采样模块和输出侧三维卷积层处理；
    所述生成器网络的残差模块具有空洞卷积层。
  7. 根据权利要求5所述的基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中，所述生成对抗网络中，
    鉴别器网络的输入数据被配置为以彩色表示的三通道的缺损三维频谱态势图、或者采样处理后的三通道的补全三维频谱态势图。
  8. 根据权利要求7所述的基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中，所述生成对抗网络中，
    所述鉴别器网络包括输入侧三维卷积层、下采样模块、残差模块和输出侧三维卷积层；
    所述鉴别器网络的输入数据依次经所述鉴别器网络的输入侧三维卷积层、下采样模块、残差模块和输出侧三维卷积层处理；
    所述鉴别器网络的残差模块具有空洞卷积层。
  9. 根据权利要求8所述的基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中，所述生成对抗网络中，
    所述鉴别器网络的残差模块的超参数集被配置为用于将指定数量的通道数的输出数据，输入至所述鉴别器网络的输出侧三维卷积层；
    所述鉴别器网络的输出侧三维卷积层具有受控制的超参数集合。
  10. 根据权利要求1或9所述的基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中，所述生成对抗网络中，
    所述生成对抗网络的目标函数被配置为具有梯度惩罚项、被配置为不具有潜变量且还被配置为具有权重的重建损失，
    所述权重的重建损失的计算包括权重与以彩色表示的三通道的缺损三维频谱态势图的哈达玛积计算、权重与以彩色表示的三通道的补全三维频谱态势图的哈达玛积计算、以及两哈达玛积之差的一范数计算。
  11. 根据权利要求10所述基于生成对抗网络的三维频谱态势补全方法，其特征在于，所述训练生成对抗网络，包括：
    按照配置的训练参数，从所述训练集中采样真实样本；
    基于瓦瑟斯坦距离，由所述真实样本和采样处理后的模拟样本计算得到以彩色表示的三通道的惩罚三维频谱态势图，其中，
    所述采样处理后的模拟样本为所述真实样本经所述生成器网络处理和所述采样处理函数采样处理后得到的。
  12. 根据权利要求11所述基于生成对抗网络的三维频谱态势补全方法，其特征在于，其中，所述生成对抗网络中，
    所述鉴别器网络的输入数据还被配置为所述惩罚三维频谱态势图，且所述鉴别器网络的输出数据以矩阵表示；
    所述梯度惩罚项的计算包括所述鉴别器网络的输出数据关于所述惩罚三维频谱态势图的梯度的计算以及该梯度的二范数的计算。
  13. 根据权利要求11所述基于生成对抗网络的三维频谱态势补全方法，其特征在于，所述训练生成对抗网络，还包括：
    基于所述真实样本中采样位置点和未采样位置点的具体分布，确定采样处理函数的表达式；
    通过所述表达式和所述训练集，计算所述目标函数和/或所述重建损失的参数的梯度。
  14. 一种基于生成对抗网络的三维频谱态势补全装置，其特征在于，该三维频谱态势补全装置包括：
    预处理模块，用于基于无人机对目标区域采样获得的缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的缺损三维频谱态势图以及基于获得的以彩色表示的三通道的缺损三维频谱态势图形成训练集；
    训练模块，用于基于所述训练集，训练生成对抗网络，获得所述生成对抗网络中训练后的生成器网络，其中，所述生成对抗网络被配置有采样处理函数，所述采样处理函数用于：基于输入所述生成对抗网络的以彩色表示的三通道的缺损三维频谱态势图中采样位置点和未采样位置点，对由所述生成对抗网络中生成器网络输出的以彩色表示的三通道的补全三维频谱态势图进行采样处理；
    补全模块，用于基于无人机对实测的指定区域采样获得的实测缺损三维频谱态势，进行灰度化和着色的预处理，获得以彩色表示的三通道的实测缺损三维频谱态势图，并将所述三通道的实测缺损三维频谱态势图作为所述生成器网络的输入数据，经所述生成器网络获得以彩色表示的三通道的实测补全三维频谱态势图。
PCT/CN2022/073723 2021-03-30 2022-01-25 基于生成对抗网络的三维频谱态势补全方法及装置 Ceased WO2022206149A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/032,573 US12430817B2 (en) 2021-03-30 2022-01-25 Three-dimensional spectrum situation completion method and device based on generative adversarial network

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110345357.3 2021-03-30
CN202110345357.3A CN113114399B (zh) 2021-03-30 2021-03-30 基于生成对抗网络的三维频谱态势补全方法和装置

Publications (1)

Publication Number Publication Date
WO2022206149A1 true WO2022206149A1 (zh) 2022-10-06

Family

ID=76712968

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/073723 Ceased WO2022206149A1 (zh) 2021-03-30 2022-01-25 基于生成对抗网络的三维频谱态势补全方法及装置

Country Status (3)

Country Link
US (1) US12430817B2 (zh)
CN (1) CN113114399B (zh)
WO (1) WO2022206149A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114997047A (zh) * 2022-05-26 2022-09-02 电子科技大学 一种基于循环生成对抗网络的电磁频谱信息补全方法
CN121476544A (zh) * 2026-01-08 2026-02-06 广东科学技术职业学院 一种卫星观测二氧化硫垂直柱总量条带校正方法、系统及设备

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114900234B (zh) * 2022-07-14 2022-10-21 四川太赫兹通信有限公司 一种太赫兹频谱环境地图构建方法及设备
CN115913424A (zh) * 2022-11-11 2023-04-04 南京航空航天大学 一种基于条件生成对抗网络的频谱地图精确构建方法
CN117078544A (zh) * 2023-08-18 2023-11-17 秦皇岛港股份有限公司 一种基于双重约束生成对抗网络的水下图像增强方法
CN118102316B (zh) * 2024-02-02 2025-04-04 中国人民解放军国防科技大学 一种基于删除搜索的频谱数据检测方法及应用
CN120017147B (zh) * 2025-04-21 2025-07-11 北京理工大学 一种自编码器的星地信道检测带宽延展方法
CN121561611B (zh) * 2026-01-21 2026-04-21 季华实验室 一种基于组合光谱的批次识别方法、装置、设备及介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106569044A (zh) * 2016-11-02 2017-04-19 西安电子科技大学 基于沉浸式虚拟现实系统的电磁频谱态势观测方法
CN107945811A (zh) * 2017-10-23 2018-04-20 北京大学 一种面向频带扩展的生成式对抗网络训练方法及音频编码、解码方法
CN111431580A (zh) * 2020-03-24 2020-07-17 南京航空航天大学 一种基于无人机平台的频谱态势认知测绘装置及方法
US20200293875A1 (en) * 2019-03-12 2020-09-17 International Business Machines Corporation Generative Adversarial Network Based Audio Restoration

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017155634A1 (en) * 2016-03-11 2017-09-14 Origin Wireless, Inc. Methods, devices, servers, apparatus, and systems for wireless internet of things applications
CN106682234B (zh) 2017-01-17 2019-12-20 北京工业大学 一种基于空间插值的电磁频谱分布预测和动态可视化方法
CN109946512B (zh) 2019-04-17 2019-12-03 贵州电网有限责任公司 一种改进频域插值的动态功率分析方法
CN110276745B (zh) * 2019-05-22 2023-04-07 南京航空航天大学 一种基于生成对抗网络的病理图像检测算法
CN110166154B (zh) * 2019-06-13 2021-03-26 桂林电子科技大学 一种基于神经网络的软件无线电频谱监测识别的方法
US11694333B1 (en) * 2020-02-05 2023-07-04 State Farm Mutual Automobile Insurance Company Performing semantic segmentation of 3D data using deep learning
CN111444967B (zh) * 2020-03-30 2023-10-31 腾讯科技（深圳）有限公司 生成对抗网络的训练方法、生成方法、装置、设备及介质
CN112331232B (zh) * 2020-11-06 2022-08-12 杭州电子科技大学 一种结合cgan谱图去噪和双边滤波谱图增强的语音情感识别方法
CN112819082B (zh) 2021-02-09 2022-06-17 南京邮电大学 一种基于深度学习的卫星频谱感知数据重构方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106569044A (zh) * 2016-11-02 2017-04-19 西安电子科技大学 基于沉浸式虚拟现实系统的电磁频谱态势观测方法
CN107945811A (zh) * 2017-10-23 2018-04-20 北京大学 一种面向频带扩展的生成式对抗网络训练方法及音频编码、解码方法
US20200293875A1 (en) * 2019-03-12 2020-09-17 International Business Machines Corporation Generative Adversarial Network Based Audio Restoration
CN111431580A (zh) * 2020-03-24 2020-07-17 南京航空航天大学 一种基于无人机平台的频谱态势认知测绘装置及方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LIU DONG; HUANG YANG; GAO ZHEN: "Spectrum Situation Completion Based on Model-Enhanced Generative Learning", 2021 13TH INTERNATIONAL CONFERENCE ON WIRELESS COMMUNICATIONS AND SIGNAL PROCESSING (WCSP), IEEE, 20 October 2021 (2021-10-20), pages 1 - 5, XP034032771, DOI: 10.1109/WCSP52459.2021.9613512 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114997047A (zh) * 2022-05-26 2022-09-02 电子科技大学 一种基于循环生成对抗网络的电磁频谱信息补全方法
CN114997047B (zh) * 2022-05-26 2024-05-14 电子科技大学 一种基于循环生成对抗网络的电磁频谱信息补全方法
CN121476544A (zh) * 2026-01-08 2026-02-06 广东科学技术职业学院 一种卫星观测二氧化硫垂直柱总量条带校正方法、系统及设备

Also Published As

Publication number Publication date
US12430817B2 (en) 2025-09-30
CN113114399A (zh) 2021-07-13
CN113114399B (zh) 2022-02-01
US20230386098A1 (en) 2023-11-30

Similar Documents

Publication Publication Date Title
WO2022206149A1 (zh) 基于生成对抗网络的三维频谱态势补全方法及装置
CN111650453B (zh) 基于加窗特征希尔伯特图像化的电力设备诊断方法及系统
LU500265B1 (en) A Method of Upsampling of Point Cloud Based on Deep Learning
CN107610194B (zh) 基于多尺度融合cnn的磁共振图像超分辨率重建方法
CN111028327B (zh) 一种三维点云的处理方法、装置及设备
US20210020360A1 (en) Internal thermal fault diagnosis method of oil-immersed transformer based on deep convolutional neural network and image segmentation
CN116740587B (zh) 一种数据与模型双重不确定性感知的无人机航拍目标可信识别方法
CN114966312B (zh) 基于迁移卷积神经网络的配电网故障检测定位方法和系统
CN109598220B (zh) 一种基于多元输入多尺度卷积的人数统计方法
CN106355151A (zh) 一种基于深度置信网络的三维sar图像目标识别方法
CN111709318B (zh) 一种基于生成对抗网络的高分辨率遥感图像分类方法
CN110555841A (zh) 基于自注意图像融合和dec的sar图像变化检测方法
CN114283285A (zh) 交叉一致性自训练遥感图像语义分割网络训练方法及装置
CN110084181B (zh) 一种基于稀疏MobileNetV2网络的遥感图像舰船目标检测方法
CN116977872B (zh) 一种CNN+Transformer遥感图像检测方法
CN111259733A (zh) 一种基于点云图像的船舶识别方法及装置
CN114639102A (zh) 基于关键点与尺寸回归的细胞分割方法和装置
CN117671666A (zh) 一种基于自适应图卷积神经网络的目标识别方法
CN115346061B (zh) 基于多保真度贝叶斯优化的图像识别方法
JP7795843B1 (ja) グラフ構造の統合最適化に基づく産業機器の故障診断方法及びシステム
Zhou et al. Infrared small target detection via learned infrared patch-image convolutional network
Liu et al. MFPIM: a deep learning model based on multimodal fusion technology for pulsar identification
CN115293983A (zh) 一种融合多层级互补特征的自适应图像超分辨率复原方法
Hoppe et al. Uncertainty quantification for learned ISTA
LU508134B1 (en) Method and system for analyzing water hammer caused by stagnant gas by coupling real gas characteristics with clsvo

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22778337

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 18032573

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22778337

Country of ref document: EP

Kind code of ref document: A1

WWG Wipo information: grant in national office

Ref document number: 18032573

Country of ref document: US