Electromagnetic environment reconstruction method based on depth compressed sensing and related equipment

Abstract

The invention provides an electromagnetic environment reconstruction method based on deep compressed sensing and related equipment. The method comprises the following steps: acquiring electromagnetic environment sample data, determining a data set from the electromagnetic environment sample data, and performing compression sampling processing on the data set to obtain sample data; performing preliminary compression reconstruction through a gradient acceleration optimization algorithm, and performing gradient calculation on the current estimation value according to the sample data to obtain the current gradient of the current estimation value; further compressing and reconstructing by extruding and exciting a dense convolutional network, and updating the current estimated value according to the current gradient to obtain an updated estimated value; and in response to determining that the number of updating times reaches a preset updating time threshold value, taking a deep compression reconstruction model formed by the gradient acceleration optimization algorithm and the extrusion excitation dense convolutional network as a trained target electromagnetic environment reconstruction model, and reconstructing the discretely distributed electromagnetic environment sampling data into a continuously distributed frequency spectrum information distribution diagram.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data processing, and in particular, to an electromagnetic environment reconstruction method and related devices based on deep compressive sensing. Background Art

[0002] With the development of intelligence and digitalization, the analysis of the electromagnetic environment has important research significance. However, limited by physical obstacles and economic costs, it is often difficult to comprehensively collect electromagnetic environment data. Electromagnetic environment reconstruction can convert discrete electromagnetic environment sampling data into a continuous spectrum information distribution map.

[0003] In view of this, how to quickly and accurately convert discrete electromagnetic environment sampling data into a continuous spectrum information distribution map has become an urgent technical problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of the present disclosure is to propose an electromagnetic environment reconstruction method and related devices based on deep compressive sensing to solve or partially solve the above technical problems.

[0005] Based on the above purpose, a first aspect of the present disclosure proposes an electromagnetic environment reconstruction method based on deep compressive sensing, and the method includes:

[0006] Obtain electromagnetic environment sample data, determine a data set from the electromagnetic environment sample data, and perform compressive sampling processing on the data set to obtain sample data;

[0007] Perform preliminary compressive reconstruction through a gradient acceleration optimization algorithm, and calculate the current gradient of the current estimated value according to the sample data to obtain the current gradient of the current estimated value;

[0008] Perform further compressive reconstruction through a squeeze-and-excitation dense convolutional network, and update the current estimated value according to the current gradient to obtain an updated estimated value;

[0009] In response to determining that the number of updates reaches a preset update number threshold, use the deep compressive reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the target electromagnetic environment reconstruction model that has been trained.

[0010] Based on the same inventive concept, a second aspect of the present disclosure proposes an electromagnetic environment reconstruction device based on deep compressive sensing, including:

[0011] An acquisition module, configured to obtain electromagnetic environment sample data, determine a data set from the electromagnetic environment sample data, and perform compressive sampling processing on the data set to obtain sample data;

[0012] The first update module is configured to perform preliminary compression and reconstruction through a gradient acceleration optimization algorithm, and calculate the current gradient of the current estimated value based on the sample data for the current estimated value;

[0013] The second update module is configured to further compress and reconstruct through a squeeze-and-excitation dense convolutional network, and update the current estimated value according to the current gradient to obtain an updated estimated value;

[0014] The generation module is configured to, in response to determining that the number of updates reaches a preset update number threshold, use the deep compression and reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the target electromagnetic environment reconstruction model that has been trained.

[0015] Based on the same inventive concept, a third aspect of the present disclosure proposes an electronic device, including a memory, a processor, and a computer program stored on the memory and executable by the processor. When the processor executes the computer program, the method described above is implemented.

[0016] Based on the same inventive concept, a fourth aspect of the present disclosure proposes a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the method described above.

[0017] As can be seen from the above, the electromagnetic environment reconstruction method and related devices provided by the present disclosure. Electromagnetic environment sample data is acquired, and a data set is determined from the electromagnetic environment sample data. The data set is subjected to compressive sampling processing to obtain sample data. Preliminary compression and reconstruction are performed through a gradient acceleration optimization algorithm, and the current gradient of the current estimated value is calculated based on the sample data for the current estimated value, which can move faster in the direction of the optimal solution, improving the convergence speed and stability of the algorithm. Further compression and reconstruction are performed through a squeeze-and-excitation dense convolutional network, and the current estimated value is updated according to the current gradient to obtain an updated estimated value. Unfolding the iterative update step into a deep neural network can improve the reconstruction accuracy and indirectly reduce the dependence of the network model on the scale of training data. When the number of updates reaches the preset update number threshold, the deep compression and reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network is used as the target electromagnetic environment reconstruction model that has been trained. In this way, based on the target electromagnetic environment reconstruction model, the discretely distributed electromagnetic environment sampling data can be reconstructed into a continuously distributed spectrum information distribution map. Description of the Drawings

[0018] To more clearly illustrate the technical solutions in the present disclosure or related technologies, the following will briefly introduce the drawings required for use in the embodiments or related technology descriptions. Obviously, the drawings in the following descriptions are only the embodiments of the present disclosure. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 Flowchart of the electromagnetic environment reconstruction method based on deep compressive sensing according to an embodiment of the present disclosure;

[0020] Figure 2 Schematic diagram of the network architecture of the electromagnetic environment reconstruction method based on deep compressive sensing according to an embodiment of the present disclosure;

[0021] Figure 3 Schematic diagram of the squeeze-and-excitation dense convolutional neural network according to an embodiment of the present disclosure;

[0022] Figure 4 Schematic diagram of the network architecture of the squeeze-and-excitation module according to an embodiment of the present disclosure;

[0023] Figure 5 Schematic diagram of the RMSE error of different algorithms according to an embodiment of the present disclosure varying with the sampling rate CSRatio;

[0024] Figure 6 Schematic diagram of the RMSE error of different algorithms according to an embodiment of the present disclosure varying with the number of training epochs Epoch;

[0025] Figure 7 Flowchart of the electromagnetic environment reconstruction method based on deep compressive sensing according to an embodiment of the present disclosure;

[0026] Figure 8 Schematic diagram of the structure of the electromagnetic environment reconstruction device based on deep compressive sensing according to an embodiment of the present disclosure;

[0027] Figure 9 Schematic diagram of the structure of the electronic device according to an embodiment of the present disclosure. Detailed implementation manners

[0028] To make the objectives, technical solutions, and advantages of the present disclosure clearer and more understandable, the following further elaborates on the present disclosure in detail with reference to specific embodiments and the accompanying drawings.

[0029] It should be noted that unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present disclosure should have the ordinary meanings understood by those of ordinary skill in the field to which the present disclosure belongs. The "first", "second" and similar terms used in the embodiments of the present disclosure do not denote any order, quantity or importance, but are only used to distinguish different components. Words such as "including" or "comprising" mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right", etc. are only used to represent relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0030] Based on the description of the background technology, with the development of intelligence, digitalization and socialization, the number of electromagnetic devices is increasing day by day, the electromagnetic spectrum signals are becoming increasingly dense, the spectrum resources are in a tense situation, the electromagnetic noise and interference are increasing day by day, the electromagnetic environment is becoming more and more complex, which has a serious impact on the normal operation order and radio monitoring management. The research on the electromagnetic environment is an important topic in the electromagnetic field. In the civilian field, the analysis of the electromagnetic environment is helpful for interference detection and management, electromagnetic pollution monitoring and prevention, spectrum resource management, combating illegal communication facilities, and creating a good social atmosphere; in the military field, the analysis of the electromagnetic environment helps the commanders to master the state and trend of the electromagnetic environment, grasp the electromagnetic force distribution of both sides of the battlefield, and can give an effective assessment of the threat level of the electromagnetic environment situation, which is helpful to improve the effect of electronic countermeasures. Therefore, the research on the utilization and analysis of the electromagnetic environment has important theoretical significance and military application value. The major premise for utilizing and analyzing the electromagnetic environment information is to obtain accurate and reliable electromagnetic environment field data. However, limited by physical obstacles (such as building occlusion, complex terrain, etc.) and economic costs, it is often difficult to achieve comprehensive data collection.

[0031] Electromagnetic environment reconstruction is a comprehensive task integrating multiple fields such as data collection, signal processing, mathematical modeling and algorithm optimization. The core lies in converting discrete electromagnetic data into a continuous and accurate spectrum information distribution map. However, in the actual environment, due to the influence of complex factors such as buildings and terrain, the collection of electromagnetic data often only covers a limited area, and the data points are discretely distributed. Therefore, how to use fewer sampling points to restore the complete high-resolution electromagnetic field distribution and fill in the missing electromagnetic field data has become a challenging topic.

[0032] The sparse reconstruction methods of the electromagnetic environment are mainly divided into three categories: spatial interpolation, parameter construction, and compressive sensing. Before the 21st century, the mainly used sparse reconstruction methods were spatial interpolation method and parameter construction method. The spatial interpolation method calculates the values of unknown field points through appropriate mathematical formulas based on the data of known field points; the parameter construction method estimates the distribution of the electromagnetic field based on physical or statistical models through a large amount of prior information; compressive sensing samples the signal at non-uniform intervals to achieve the purpose of using a small number of non-linear observations to recover sparse or compressible signals.

[0033] The spatial interpolation method does not consider the propagation characteristics of electromagnetic waves. It may not be the best method to approximate the trend of the electromagnetic environment, and it often ignores the details in the distribution, making it difficult to achieve accurate reconstruction. The parameter construction method requires a large amount of prior information to ensure the reconstruction accuracy, which is difficult to obtain in the actual sampling of the electromagnetic environment field. To solve the above shortcomings, a compressive sensing algorithm with a smaller demand for the number of sampling points and a more accurate reconstruction effect can be adopted.

[0034] In recent years, deep learning algorithms have achieved good research results in many fields. Some scholars have seized the opportunity of the development of deep learning and applied depth information to compressive sensing. Most of the current image compressive sensing reconstruction algorithms are divided into three categories: model-driven iterative optimization methods, data-driven neural network methods, and model-driven deep unfolding methods. The model-driven iterative optimization methods usually use mathematical models to describe the underlying structure of the original signal, and then use iterative optimization techniques to find the best solution that satisfies the metric constraints, such as the Iterative Shrinkage-Thresholding Algorithm (ISTA); the data-driven neural network methods use deep neural networks to learn the mapping between the compressed measurements and the original images. These methods usually include training a neural network with a large dataset of image examples to learn the latent patterns and correlations in the data; the model-driven deep unfolding methods connect the iterative optimization methods with the neural network methods. In the model-driven deep unfolding network, each iteration of the optimization algorithm is mapped to a layer of the reconstruction network, converting the algorithm parameters (such as model parameters and regularization coefficients) into corresponding network parameters, and using the trained network to process the signal, which is equivalent to performing multiple iterative optimizations. This network structure effectively overcomes the problem of the lack of interpretability of the data-driven neural network methods.

[0035] Model-driven iterative optimization methods show strong interpretability based on well-studied mathematical models and inherent image properties in most cases. However, they inevitably suffer from high computational complexity. To accurately recover the original signal, usually hundreds of iterations need to be performed, such as the ISTA algorithm. In addition, these methods require manual parameter adjustment by humans and it is difficult to achieve the optimal. Data-driven neural network methods are trained as black boxes and it is difficult to explain how the network performs image reconstruction, without sufficient theoretical guarantee. Model-driven deep unfolding methods start from traditional mathematical modeling, unfold the iterative process of traditional compressive sensing algorithms with neural networks for learning, and the network parameters are self-updated and optimized during the training process. This method can not only utilize image prior knowledge but also make the algorithm interpretable, achieving a good balance between computational complexity and theoretical interpretability.

[0036] In current model-driven deep unfolding methods, the ISTA-Net model combines the traditional iterative shrinkage threshold algorithm (ISTA) with a deep learning architecture, replaces the linear shrinkage threshold step in the ISTA algorithm with a non-linear function, and uses a convolutional neural network (CNN) to replace the iterative process, greatly improving the efficiency of signal recovery. However, in the CNN, the connections between layers are relatively limited. Usually, the output of the previous layer is used as the input of the next layer, and the information transmission path is relatively single. Information loss may occur during the information transmission process due to the increase in the number of network layers. The ISTA-Net+ model is improved and optimized based on ISTA-Net, introduces skip connections, further enhances the performance and adaptability of the network, can more effectively handle complex inverse problems and a wider type of signal recovery tasks, and improves the problem of information loss in the CNN, but the learning of features is still limited. With the increase in the number of network layers, overfitting and gradient vanishing problems may occur. The ISTA-Net++ network model proposes a dynamic unfolding strategy and a cross-block strategy, processes multi-scale sampling rate information through 3 fully connected layers, and fuses the obtained parameters into the reconstruction network to increase network flexibility, but focuses on learning residuals, is relatively limited in terms of feature diversity, and requires a large increase in parameters to learn new features. Integrating a customized ISTA-based transformer backbone with a CNN, a transformer-based hybrid architecture (called TransCS) is proposed for high-performance CS reconstruction, but the utilization of spatial structure information of data such as images is relatively indirect, and problems such as gradient explosion or vanishing may occur. Compressive sensing algorithms combined with deep learning can design measurement matrices with smaller dimensions and self-adaptability, have a faster sampling speed, and higher algorithm reconstruction accuracy, and are involved in fields such as image acquisition, image and video coding, wireless remote monitoring, image super-resolution, and medical magnetic resonance imaging.

[0037] As described above, how to quickly and accurately convert discrete electromagnetic environment sampling data into a continuous spectrum information distribution map has become an important research issue.

[0038] Based on the above description, as Figure 1 shown, the electromagnetic environment reconstruction method based on deep compressed sensing proposed in this embodiment includes:

[0039] Step 101: Obtain electromagnetic environment sample data, determine a data set from the electromagnetic environment sample data, and perform compressed sampling processing on the data set to obtain sample data.

[0040] In specific implementation, discretize the physical space environment, convert the continuous electromagnetic data distributed in space into multiple discrete points in a grid, and use an empirical model to generate complete electromagnetic data as the data set of the algorithm.

[0041] Compressed sensing (CS for short) is a signal sampling and processing theory that combines signal sampling and compression. By performing non-uniform sampling on the signal, it aims to recover a sparse or compressible signal using a small number of non-linear observations.

[0042] Step 102: Perform preliminary compressed reconstruction through a gradient acceleration optimization algorithm, and calculate the current gradient of the current estimated value according to the sample data to obtain the current gradient of the current estimated value.

[0043] In specific implementation, incorporate the Nesterov Accelerated Gradient (NAG) theory to optimize the gradient update step of the ISTA algorithm.

[0044] The ISTA algorithm is a proximal gradient algorithm. Its core idea is to solve an optimization problem through iteration, gradually approaching the optimal solution while ensuring the sparsity of the signal. NAG is an optimization algorithm mainly used to accelerate the convergence speed of the gradient descent method.

[0045] Step 103: Further perform compressed reconstruction through a squeeze-and-excitation dense convolutional network, and update the current estimated value according to the current gradient to obtain an updated estimated value.

[0046] In specific implementation, a Squeeze-and-Excitation Net (SENet) and a DenseNet are introduced. The iterative update process of the estimated value in the ISTA algorithm is expanded into a deep learning network, which is used as the Squeeze-and-Excitation Dense Block in the algorithm. The iterative update process of the gradient adopts the gradient acceleration optimization algorithm optimized by the traditional compressive sensing in step 102.

[0047] SENet and DenseNet can perform multi-scale learning of the important information of each feature channel, strengthen the useful features according to the importance of each feature, weaken the unimportant features, and recalibrate the feature responses of the channels in an adaptive manner to enhance the ability to extract data features.

[0048] Step 104: In response to determining that the number of updates reaches a preset update number threshold, the deep compression reconstruction model composed of the gradient acceleration optimization algorithm and the Squeeze-and-Excitation Dense Network is used as the target electromagnetic environment reconstruction model that has been trained.

[0049] In specific implementation, the number of updates of the gradient and the estimated value is recorded. When the number of updates reaches the preset update number threshold, the gradient acceleration optimization algorithm and the Squeeze-and-Excitation Dense Network are used as the target electromagnetic environment reconstruction model. In this way, based on the target electromagnetic environment reconstruction model, the discretely distributed electromagnetic environment sampling data can be reconstructed into a continuously distributed spectrum information distribution map.

[0050] Specifically, set the hyperparameters during model training, mainly including the number of iterations, the number of convolutional block layers, the loss function, the optimizer, the learning rate, etc. The model is trained by calculating the root mean square error between the model output and the real input data to obtain the target electromagnetic environment reconstruction model.

[0051] The present disclosure unfolds a traditional compressed sensing reconstruction algorithm into a multi-layer neural network for learning, thereby improving the reconstruction accuracy. The optimization of traditional compressed sensing reconstruction algorithms requires manual adjustment of parameters by humans, making it difficult to achieve optimality; moreover, the required number of iterations is too high, the computational time complexity is high, and excessive computational resources are consumed. The introduction of deep learning has gradually networked traditional algorithms. Deep learning has the ability to perform multi-task parallel processing and powerful feature extraction capabilities. Using a deep learning reconstruction network can effectively reduce the computational complexity, reduce the computational time, and improve the reconstruction accuracy. Among them, the deep networking method of traditional algorithms starts from traditional mathematical modeling, uses a neural network to simulate the iterative process of traditional compressed sensing algorithms, maps each iteration to a network layer, and obtains parameters through end-to-end network training. These parameters are self-updated and optimized during the training process without manual setting. The algorithm has higher reconstruction accuracy and faster speed. This method can not only utilize image prior knowledge but also make the algorithm interpretable.

[0052] In addition, in order to improve the convergence speed and stability of the algorithm, the NAG theory is incorporated into this solution. NAG is an optimization algorithm technique mainly used to accelerate the convergence speed of the gradient descent method. Traditional gradient descent updates parameters according to the gradient at the current position, while NAG considers the gradient situation at a position ahead in the current gradient direction, utilizes the concept of momentum, estimates a future position for the current position, and calculates the gradient at this estimated future position, so that the algorithm can move faster towards the direction of the optimal solution. Compared with the ordinary gradient descent method, it can achieve a similar or better optimization effect with fewer iterations and is relatively more robust to the selection of the learning rate. Applied in a deep neural network, NAG can help optimize the weight parameters of the network faster and improve the convergence speed and stability.

[0053] Through the above embodiments, electromagnetic environment sample data is obtained, a data set is determined from the electromagnetic environment sample data, and the data set is subjected to compressive sampling processing to obtain sample data. Preliminary compressed reconstruction is performed through a gradient acceleration optimization algorithm. The current gradient of the current estimated value is calculated based on the sample data, and it can move faster towards the direction of the optimal solution, improving the convergence speed and stability of the algorithm. Further compressed reconstruction is performed through a squeeze-and-excitation dense convolutional network. The current estimated value is updated based on the current gradient to obtain an updated estimated value. Unfolding the iterative update step into a deep neural network can improve the reconstruction accuracy and indirectly reduce the dependence of the network model on the scale of training data. When the number of updates reaches the preset update number threshold, the deep compressed reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network is used as the target electromagnetic environment reconstruction model that has been trained. In this way, based on the target electromagnetic environment reconstruction model, the discretely distributed electromagnetic environment sampling data can be reconstructed into a continuously distributed spectrum information distribution map.

[0054] In some embodiments, step 102 includes:

[0055] Step 1021, initialize the sampled data in the sample data to obtain an initial estimate value, and use the initial estimate value as the current estimate value.

[0056] Step 1022, perform preliminary compression and reconstruction through a gradient acceleration optimization algorithm, calculate the current gradient of the current estimate value according to the current estimate value and the previous estimate value,

[0057] z (t) = q (t) - ρΦ T (Φq (t) - y)

[0058] where z (t) is the current gradient of the t-th iteration, ρ is the step size, Φ is the sampling matrix of the sample data, Φ T is the transpose matrix of the sampling matrix, y is the sampled data in the sample data, q (t) is the estimate coefficient constructed based on the current estimate value and the previous estimate value,

[0059]

[0060] where x (t-1) is the current estimate value of the (t - 1)-th iteration, x (t-2) is the previous estimate value of the (t - 2)-th iteration, λ t is the regularization parameter of the t-th iteration, λ t-1 is the regularization parameter of the (t - 1)-th iteration,

[0061] Specifically, the ISTA algorithm is a proximal gradient algorithm, and its core idea is to solve an optimization problem through iteration, while ensuring the sparsity of the signal and gradually approaching the optimal solution. The basic steps include: initialization, select an initial estimate value x0; iterative update, in each iteration step, construct the gradient z of x by adding a positive semi-definite quadratic form, and the specific expression is shown in formula (1), and then update x using the soft threshold method, that is, formula (2).

[0062] z (t) = x (t-1) - ρΦ T (Φx (t-1) - y) (1)

[0063]

[0064] where t is the number of iterations, ρ is the step size, Φ is the sampling matrix of the sample data, ΦT is the transpose matrix of the sampling matrix, y is the sampled data in the sample data, λ is the regularization parameter (usually a constant greater than 0), and Ψ is the sparse matrix.

[0065] Among them, ‖x‖2 represents the l2 norm (Euclidean norm) of the vector x. If the vector x = (x1, x2, …, x n ) is an n-dimensional vector, then the calculation formula for the l2 norm of the vector x is which is the square of the l2 norm of the vector x, that is ‖x‖1 usually represents the l1 norm of the vector x. If the vector x = (x1, x2, …, x n ) is an n-dimensional vector, then the calculation formula for the l1 norm of the vector x is ‖x‖1 = |x1| + |x2| + … + |x n |.

[0066] NAG is an optimization algorithm mainly used to accelerate the convergence speed of the gradient descent method. Traditional gradient descent updates parameters according to the gradient at the current position, while NAG considers the gradient situation at the leading position in the current gradient direction, utilizes the concept of momentum, estimates a future position at the current position, and calculates the gradient at this estimated future position, so that the algorithm can move faster towards the optimal solution direction. This disclosure applies the NAG theory to the ISTA algorithm to optimize formula (1) to obtain formula (3), where

[0067]

[0068] z (t) = q (t) - ρΦ T (Φq (t) - y) (3)

[0069] Through the above solution, in the process of updating the current gradient through the gradient acceleration optimization algorithm to obtain the updated gradient, according to the current estimated value x at the (t - 1)-th iteration (t-1) and the previous estimated value x at the (t - 2)-th iteration (t-2) to update the current gradient to obtain the updated gradient. In this way, considering the current estimated value and the previous estimated value comprehensively, it can move faster towards the optimal solution direction, can achieve the optimization effect with fewer iteration times, and improve the convergence speed and stability.

[0070] In some embodiments, step 103 includes:

[0071] Step 1031, input the current gradient into the Dense Convolutional Network DenseNet to learn the target features of each layer.

[0072] Step 1032: After performing an activation operation on each dense convolutional block of the Dense Convolutional Network (DenseNet), connect a squeeze-and-excitation module. Through the channel attention mechanism (SENet), strengthen the target features according to the importance of each feature and recalibrate the feature responses of the channels in an adaptive manner.

[0073] Step 1033: Update the current estimated value through multiple squeeze-and-excitation dense convolutional learning processes to obtain an updated estimated value.

[0074] In specific implementation, the present disclosure innovatively introduces SENet and DenseNet, expands the iterative update process of the estimated value in the ISTA algorithm, i.e., formula (2), into a deep learning network as the squeeze-and-excitation dense convolutional block (Squeeze-and-Excitation Dense Block) in the algorithm, and the update process of the gradient z adopts the optimized formula (3) of traditional compressive sensing. The network can perform multi-scale learning of the important information of each feature channel, strengthen the useful features according to the importance of each feature, weaken the unimportant features, and recalibrate the feature responses of the channels in an adaptive manner to enhance the ability to extract data features.

[0075] Figure 2 It is a schematic diagram of the network architecture of the electromagnetic environment reconstruction method based on deep compressive sensing according to the embodiments of the present disclosure. As Figure 2 shown, first, obtain a complete electromagnetic data sample x, then sample the electromagnetic data sample x to obtain a sampled data y = Ax, and obtain an initial estimated value x (0) , and input the initial estimated value x (0) into the NSED-ISTANet model of the present disclosure for T iterations to obtain a final estimated value x (T) . Among them, in each iteration, first construct an updated gradient z for the input current estimated value x (t-1) through formula (3), and then input the updated gradient z (t) into the squeeze-and-excitation dense convolutional block (Squeeze-and-Excitation Dense Block). The updated gradient z (t) is input into the squeeze-and-excitation dense convolutional block (Squeeze-and-Excitation Dense Block), and the updated gradient z (t)First, it goes through a layer of concat to concatenate with the previous layers (connecting multiple tensors along a specific dimension to form a new, larger tensor, which is the structure of DenseNet here). Then, it performs a 3×3 convolution. After the convolution, it passes through a ReLU activation layer. Finally, it undergoes a squeeze-and-excitation operation through the SEBlock to obtain F1. This process is looped 6 times to get F6, and F6 is post-processed (including a layer of concat and a 1×1 CNN convolution with a kernel size of 1 to adjust the output channel number and size to make the front and back sizes and channel numbers consistent) to obtain the updated estimated value x (t) The above iterative process is repeated T times to obtain the final estimated value x (t) 。

[0076] Figure 3 is a schematic diagram of the squeeze-and-excitation dense convolutional neural network according to an embodiment of the present disclosure. As Figure 3 shown, each layer in the squeeze-and-excitation dense convolutional neural network includes: a channel attention module (SEBlock), an activation layer (ReLU), a 3×3 Conv, and feature fusion (Concat).

[0077] Figure 4 is a schematic diagram of the network architecture of the squeeze-and-excitation module according to an embodiment of the present disclosure. As Figure 4 shown, SENet is a novel channel attention mechanism that obtains the importance of each feature channel through compression and excitation operations, and then strengthens useful features and weakens unimportant features according to the importance of the features, enabling the model to adaptively recalibrate the importance of channel features to improve the performance of the network. The main steps are: compression, compressing the features from the spatial dimension through global pooling to convert each two-dimensional feature channel into a real number, noting that the output dimension needs to match the number of input feature channels; excitation, adopting a threshold mechanism, following the process of fully connected layer (FC) → activation layer (ReLU) → fully connected layer (FC) → activation layer (Sigmoid) to learn the correlation information between channels; reweighting, after feature selection, obtaining the weights of the excitation output (the larger the weight, the higher the importance of the feature channel), then weighting according to the channels, multiplying with the previous features, and finally recalibrating the features on each channel dimension.

[0078] DenseNet is an innovative convolutional neural network architecture, mainly characterized by establishing dense connections. DenseNet mainly consists of convolutional layers, dense blocks, and transition layers. Each convolutional layer is connected to the layers in the dense block in a feed-forward manner, and then activation mapping operations are performed on each layer in the network and passed as input between layers. In traditional convolutional neural networks, such as ResNet which adopts residual connections, each layer is only connected to a few layers (usually adjacent layers). While in DenseNet, as Figure 3 shown, each layer is directly connected to all the previous layers. This connection method enables the network to make full use of the feature information of the previous layers, utilize both low-level and high-level features simultaneously, promotes the reuse of features and the flow of information, alleviates the problems of gradient disappearance, gradient explosion, and network degradation, thereby enhancing the network's learning ability and feature representation ability, and having fewer parameters than ResNet.

[0079] The squeeze-and-excitation dense convolutional neural network is a deep model with multiple hidden layers. After the activation operation of each dense convolutional block, a squeeze-and-excitation module is connected. Among them, through the dense convolutional network DenseNet, the target features of each previous layer can be learned. Through the channel attention mechanism SENet, useful features can be strengthened according to the importance of each feature and the feature responses of channels can be recalibrated adaptively. Through multiple squeeze-and-excitation dense convolutional learning, the current estimated value can be updated according to the current gradient to obtain an updated estimated value.

[0080] The squeeze-and-excitation dense convolutional block of the present disclosure mainly includes 6 convolutional blocks and 1 post-processing layer (including one layer of concat splicing and one CNN convolution with a kernel size of 1×1, aiming to adjust the output channel number and size to make the front and back sizes and channel numbers consistent). The final output is x obtained after one iteration. Each convolutional block is designed to add a squeeze-and-excitation module after each dense block, mainly including one layer of concat splicing (connecting multiple data sequences (such as arrays, tensors, etc.) together along a specific dimension to form a new and larger data structure), one CNN convolution with a kernel size of 3×3, one ReLu activation layer, and the squeeze-and-excitation module.

[0081] Through the above solution, the squeeze excitation dense convolution block uses SENet and DenseNet to expand the iterative update step of ISTA into a deep neural network. Through compression and excitation operations, the importance of each feature channel is obtained, and then, according to the importance of the features, useful features are strengthened while unimportant features are weakened, enabling the model to adaptively recalibrate the importance of channel features. Through the dense connections of each convolutional layer, the network can make full use of the feature information of the previous layers, and at the same time utilize low-level and high-level features, promoting the reuse of features and the flow of information, and can more effectively capture important features of data from multiple scales and channels, thereby improving the reconstruction accuracy and indirectly reducing the dependence of the network model on the scale of training data.

[0082] In some embodiments, step 101 includes:

[0083] Step 1011, determining a target electromagnetic region and performing a partitioning process on the target electromagnetic region to obtain grid cells.

[0084] Step 1012, predicting the electromagnetic signal intensity of each grid cell.

[0085] Step 1013, determining a first grid for signal acquisition from the grid cells according to a preset sampling rate, and taking the other grids in the grid cells except the first grid as second grids.

[0086] Step 1014, collecting the electromagnetic signal intensity within the first grid, and setting the electromagnetic signal intensity within the second grid to a preset value to obtain electromagnetic environment sample data.

[0087] Step 1015, dividing the electromagnetic environment sample data into a training data set and a test data set according to a preset ratio.

[0088] Specifically, when implemented, the physical space environment is discretized, the continuous electromagnetic data distributed in space is converted into multiple discrete points in a grid, and an empirical model is used to generate complete electromagnetic data as the data set of the algorithm.

[0089] First, the target electromagnetic region with size (x, y) is evenly divided into N = N x *N y grid cells, where N represents the total number of grids, and each grid cell represents a discrete spatial sampling point.

[0090] Then, construct the complete electromagnetic environment data. Considering that there are K emission sources in the target area, set the number, position, and power of the emission sources, use the Hata empirical model (a calculation model for predicting radio wave propagation loss in wireless communication) to calculate the received signal strength at each grid point, and use the Gudmundson model (a model for describing channel characteristics in wireless communication, which has significant characteristics in considering shadow fading and small-scale fading) to superimpose the shadow effect, so as to generate the electromagnetic environment data of each grid point. Subsequently, according to the spatial position relationship, these electromagnetic environment data are corresponded with the coordinates of the grid points one by one to construct the complete electromagnetic environment data.

[0091] Finally, construct a data set for the algorithm. Through the above two steps, 88,912 different complete electromagnetic environment data are produced as the data set, and the data set is divided into a training set and a test set according to the ratio of 8:2. The number of samples in the training set is 71,129, and the number of samples in the test set is 17,783. On this basis, randomly sample the original data and construct a measurement matrix M according to the sampling point positions. The data at the sampling points represent the data that can be obtained through electromagnetic data detection equipment in the real environment, and the value is the electromagnetic signal strength at the sampling points. The data that are not sampled represent the missing data that cannot be detected due to physical environment limitations, and the value is 0. In order to better verify the effectiveness of the algorithm, the sampling rates adopted in this disclosure are 1%, 4%, 10%, 25%, 40%, and 50% respectively.

[0092] For example, if the preset sampling rate is 10% and the number of grids is 100, then collect the electromagnetic signal strengths in 10 grids, and set the electromagnetic signal strengths in 90 grids to 0 to obtain the electromagnetic environment sample data.

[0093] Through the above solution, by dividing the target electromagnetic area into grid units and sampling the grid units according to the preset sampling rate, the electromagnetic environment sample data can be accurately obtained. Dividing the electromagnetic environment sample data into a training data set and a test data set can use the training data set to train the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network, and use the test data set to test the trained gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network. The gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network that meet the test conditions are used as the target electromagnetic environment reconstruction model, so that the obtained target electromagnetic environment reconstruction model is more accurate.

[0094] In some embodiments, after step 103, it further includes:

[0095] Step 103A, determine the test data set from the electromagnetic environment sample data.

[0096] Step 103B: Obtain the reconstructed electromagnetic signal intensity according to the test data set by means of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network.

[0097] Step 103C: Obtain the loss value between the reconstructed electromagnetic signal intensity and the predicted electromagnetic signal intensity based on the loss function.

[0098] Step 103D: In response to determining that the loss value is less than a preset loss threshold, use the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the trained target electromagnetic environment reconstruction model.

[0099] In specific implementation, the reconstructed electromagnetic signal intensity is obtained according to the test data set in the electromagnetic environment sample data by means of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network.

[0100] Among them, the reconstructed electromagnetic signal intensity is the electromagnetic signal intensity reconstructed by means of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network. The predicted electromagnetic signal intensity is the complete electromagnetic signal intensity generated in Step 101.

[0101] Obtain the loss value between the reconstructed electromagnetic signal intensity and the predicted electromagnetic signal intensity based on the loss function. When the loss value is less than the preset loss threshold, it indicates that the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network meet the convergence condition, and use the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the trained target electromagnetic environment reconstruction model.

[0102] Through the above solution, obtain the loss function based on the reconstructed electromagnetic signal intensity and the predicted electromagnetic signal intensity. When the loss function is less than the preset loss threshold, use the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the target electromagnetic environment reconstruction model. In this way, it is possible to accurately determine according to the loss function that the trained gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network meet the conditions, and use the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the target electromagnetic environment reconstruction model, so that the obtained target electromagnetic environment reconstruction model is more accurate.

[0103] In some embodiments, Step 103D includes:

[0104] Step 103Da: Obtain the loss value between the reconstructed electromagnetic signal intensity and the predicted electromagnetic signal intensity based on the loss function,

[0105]

[0106] where Loss is the loss value, N is the number of samples in the test data set, Y i is the i-th reconstructed electromagnetic signal intensity, and X i is the i-th predicted electromagnetic signal intensity.

[0107] In specific implementation, hyperparameters during model training are set, mainly including the number of iterations, the number of convolutional block layers, loss function, optimizer, learning rate, etc. The number of iterations refers to Figure 2 the value of T in Figure 3 ; the number of convolutional block layers refers to the value of C in -4 .

[0108]

[0109] The model is trained by calculating the root mean square error between the model output and the real input data to obtain an electromagnetic reconstruction model. First, the training data is randomly shuffled and batched through the DataLoader in PyTorch, and then each sample (the data of each batch is regarded as a sample) is sent into the model for training in turn. Each time during training, the root mean square error between the model output and the real input data is calculated, and through backpropagation, the model continuously learns and optimizes automatically to update the model parameters. Finally, the model weights with the minimum loss value are obtained as the electromagnetic reconstruction model.

[0110] Through the above solution, according to the loss function, it can be accurately determined that the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network after training meet the conditions, and the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network are used as the target electromagnetic environment reconstruction model, so that the obtained target electromagnetic environment reconstruction model is more accurate.

[0111] In some embodiments, after step 104, it further includes:

[0112] Step 104A, obtaining electromagnetic environment sampling data with discrete distribution.

[0113] Step 104B, through the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network in the target electromagnetic environment reconstruction model, performing reconstruction processing on the electromagnetic environment sampling data to obtain a spectrogram information distribution map with continuous distribution.

[0114] In specific implementation, electromagnetic environment reconstruction is performed according to the trained model. This step requires generating a brand-new electromagnetic data and sending it into the trained model for electromagnetic environment reconstruction. The new data is different from the training data and the test data, so as to effectively evaluate the performance of the model.

[0115] The present disclosure is applicable to model training in different scenarios and tasks, including image classification, image segmentation, language text processing and generation, etc. According to different task objectives, fine-tuning the model architecture, backpropagation logic, input and output data formats, and parameters designed by the present invention can achieve the expected goals.

[0116] Although the present invention has been described in connection with specific embodiments of the present disclosure, many alternatives, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art based on the foregoing description. For example, after obtaining the electromagnetic training data set, replacing the iterative update process of the ISTA algorithm with a simple convolutional network or a basic structure based on a convolutional network, and adding or replacing some feature extraction and enhancement structures, such as dilated convolution, residual connection, etc., can all achieve the reconstruction of electromagnetic data. However, these solutions do not consider the sample data volume and model convergence speed during model training, and their reconstruction accuracy cannot reach the optimal value.

[0117] Through the above solution, the electromagnetic environment reconstruction algorithm based on deep compressive sensing can perform sufficient feature extraction on data in the case of a small sample data volume and complete the task of electromagnetic environment reconstruction.

[0118] To study the reconstruction accuracy of the NSED-ISTANet algorithm proposed in the present disclosure relative to other model-driven deep unfolding algorithms, ISTA-Net and ISTA-Net++ are respectively used to conduct comparative simulations with the algorithm proposed in the present disclosure. Figure 5 It is a schematic diagram of the RMSE error of different algorithms in the embodiments of the present disclosure varying with the sampling rate CSRatio. As Figure 5 shown, the ordinate represents the RMSE error between the reconstructed value and the true value, and the abscissa represents different sampling rates. Using a data set of 88,912 data for training, at the same sampling rate, the RMSE of the algorithm proposed in the present disclosure is reduced by approximately 45.43% - 98.82% compared to ISTA-Net; and by approximately 12.28% - 95.70% compared to ISTA-Net++. It can be preliminarily proved that the research method of the present disclosure has higher-performance electromagnetic data reconstruction accuracy. At the same time, in the case of a low data volume of the data set, it can fully capture and learn data features, overcome the defect of requiring a large amount of data for training, greatly reduce the cost of electromagnetic data acquisition, and save the resources consumed by hardware training.

[0119] To improve the convergence speed of the model, the present disclosure incorporates NAG into the ISTA gradient update process. To verify the rationality of such a design, corresponding simulation experiments are designed. When the sampling rate is 25% in both cases, the ISTA Net, ISTA Net ++, and the research method of the present disclosure are trained for 200 rounds respectively, and the RMSE of each round of training is recorded. When the RMSE tends to be stable, it is regarded as the convergence of the algorithm. Figure 6 It is a schematic diagram showing the change of the RMSE error of different algorithms in the embodiments of the present disclosure with the number of training epochs Epoch. As Figure 6 shown, the electromagnetic environment reconstruction method based on deep compressive sensing converges around the 20th round, while the ISTA Net and ISTA Net ++ converge around the 40th round, which can preliminarily prove that the research method of the present disclosure has a faster model convergence speed.

[0120] Through the above embodiments, electromagnetic environment sample data is obtained, and a neural network data set is determined from the electromagnetic environment sample data. Through the gradient acceleration optimization algorithm, preliminary compression and reconstruction are carried out. According to the data set, the current gradient of the current estimated value is calculated, and it can move faster in the direction of the optimal solution, improving the convergence speed and stability of the algorithm. Through the squeeze-and-excitation dense convolutional network, further compression and reconstruction are carried out. According to the current gradient, the current estimated value is updated to obtain an updated estimated value. Unfolding the iterative update step into a deep neural network can improve the reconstruction accuracy and indirectly reduce the dependence of the network model on the scale of training data. When the number of updates reaches the preset update number threshold, the deep compression reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network is used as the target electromagnetic environment reconstruction model that has completed training. In this way, based on the target electromagnetic environment reconstruction model, the discretely distributed electromagnetic environment sampling data can be reconstructed into a continuously distributed spectrum information distribution map.

[0121] It should be noted that the embodiments of the present disclosure can also be further described in the following ways:

[0122] Figure 7 It is a flowchart of the electromagnetic environment reconstruction method based on deep compressive sensing in the embodiments of the present disclosure. As Figure 7 shown, the electromagnetic environment reconstruction method based on deep compressive sensing includes:

[0123] Step 201, discretize the physical space environment, convert the continuous electromagnetic data distributed in space into multiple discrete points in the grid, and use the empirical model to generate complete electromagnetic data as the data set of the algorithm. It further includes the following specific steps:

[0124] Step 2011, first, evenly divide the target electromagnetic region with size (x, y) into N = N x *N yN grid cells, where N represents the total number of grids, and each grid cell represents a discrete spatial sampling point.

[0125] Step 2012. Then, construct the complete electromagnetic environment data. Considering that there are K emission sources in the target area, set the number, position, and power of the emission sources, use the Hata empirical model (a calculation model for predicting radio wave propagation loss in wireless communication) to calculate the received signal strength at each grid point, and use the Gudmundson model to superimpose the shadow effect, so as to generate the electromagnetic environment data of each grid point. Subsequently, construct the complete electromagnetic environment data by corresponding these electromagnetic environment data with the coordinates of the grid points according to the spatial position relationship.

[0126] Step 2013. Finally, construct a data set for the algorithm. Make 88,912 different complete electromagnetic environment data as the data set through Steps 2011 and 2022, and divide the data set into a training set and a test set according to the ratio of 8:2. The number of samples in the training set is 71,129, and the number of samples in the test set is 17,783. On this basis, randomly sample the original data and construct a measurement matrix M according to the sampling point positions. The data of the sampling points represent the data that can be obtained through electromagnetic data detection equipment in the real environment, and the value is the electromagnetic signal strength of the sampling point. The data that is not sampled represents the missing data that cannot be detected due to physical environment limitations, and the value is 0. In order to better verify the effectiveness of the algorithm, the sampling rates adopted in this disclosure are 1%, 4%, 10%, 25%, 40%, and 50% respectively.

[0127] Step 202. Combine the traditional compressive sensing algorithm with the deep neural network to build the NSED-ISTANet neural network framework, and set the hyperparameters during model training, mainly including the number of iterations, the number of convolutional block layers, the loss function, the optimizer, the learning rate, etc. The NSED-ISTANet neural network architecture is shown in Figure 2 as shown.

[0128] Step 2021. The present invention innovatively incorporates the NAG theory to optimize the gradient update step of the ISTA algorithm.

[0129] The ISTA algorithm is a kind of proximal gradient algorithm. The core idea is to solve an optimization problem through iteration, and gradually approach the optimal solution while ensuring the sparsity of the signal. The basic steps include: initialization, select an initial estimate value x0; iterative update, in each step of iteration, construct the gradient z of x by adding a positive semi-definite quadratic form, and the specific expression is shown in formula (1), and then use the soft threshold method, that is, formula (2), to update x.

[0130] z (t) = x (t-1) - ρΦT (Φx (t-1) -y)(1)

[0131]

[0132] where \(t\) is the number of iterations, \(\rho\) is the step size, \(\Phi\) is the sampling matrix, \(\Phi\) T is the transpose of the sampling matrix, \(y\) is the sampled data, \(\lambda\) is a constant, and \(\psi\) is the sparse matrix.

[0133] NAG is an optimization algorithm technology mainly used to accelerate the convergence rate of the gradient descent method. Traditional gradient descent updates parameters according to the gradient at the current position, while NAG considers the gradient situation at the forward position in the current gradient direction, utilizes the concept of momentum, estimates a future position in the current position, and calculates the gradient at this estimated future position, so that the algorithm can move faster towards the optimal solution direction. The present disclosure applies the NAG theory to the ISTA algorithm, optimizes formula (1), and obtains formula (3), where

[0134]

[0135] z (t) =q (t) -\rho\Phi T (\Phi q (t) -y)(3)

[0136] In step 2022, the present disclosure innovatively introduces SENet and DenseNet, expands the iterative update process of the estimated value in the ISTA algorithm, that is, formula (2), into a deep learning network as the Squeeze-and-Excitation Dense Block in the algorithm, and the update process of the gradient \(z\) adopts the optimized formula (3) of traditional compressive sensing. The network can perform multi-scale learning of the important information of each feature channel, strengthen the useful features according to the importance of each feature, weaken the unimportant features, and recalibrate the feature response of the channel in an adaptive manner to enhance the ability to extract data features.

[0137] SENet is a novel channel attention mechanism, and the specific network architecture is as Figure 4As shown. The importance of each feature channel is obtained through compression and excitation operations. Then, according to the importance of the features, useful features are strengthened while unimportant features are weakened, enabling the model to adaptively recalibrate the importance of channel features to improve the performance of the network. The main steps are as follows: Compression, where the features are compressed from the spatial dimension through global pooling, converting each two-dimensional feature channel into a real number, noting that the output dimension needs to match the number of input feature channels; Excitation, using a threshold mechanism, following the process of fully connected layer (FC) → activation layer (ReLU) → fully connected layer (FC) → activation layer (Sigmoid) to learn the correlation information between channels; Reweighting, after feature selection, the weights of the excitation output are obtained (the larger the weight, the higher the importance of the feature channel), then weighted according to the channels and multiplied by the previous features, and finally the features on each channel dimension are recalibrated.

[0138] DenseNet is an innovative convolutional neural network architecture, mainly characterized by establishing dense connections. DenseNet mainly includes convolutional layers, dense blocks, and transition layers. Each convolutional layer is connected to the layers in the dense block in a feed-forward manner, then an activation mapping operation is performed on each layer in the network, and then it is passed as input between layers. In traditional convolutional neural networks, such as ResNet using residual connections, each layer is only connected to a few layers (usually adjacent layers). In DenseNet, each layer is directly connected to all the previous layers. This connection method enables the network to make full use of the feature information of the previous layers, simultaneously utilize low-level and high-level features, promote the reuse of features and the flow of information, alleviate the problems of gradient vanishing, gradient explosion, and network degradation, thereby enhancing the network's learning ability and feature representation ability, and having fewer parameters than ResNet.

[0139] The squeeze-and-excitation dense convolutional neural network is a deep model with multiple hidden layers. After the activation operation of each dense convolutional block, a squeeze-and-excitation module is connected. Among them, through the dense convolutional network DenseNet, the target features of each previous layer can be learned. Through the channel attention mechanism SENet, useful features can be strengthened according to the importance of each feature and the feature responses of the channels can be adaptively recalibrated. Through multiple squeeze-and-excitation dense convolutional learning, the current estimated value can be updated according to the current gradient to obtain an updated estimated value.

[0140] The squeeze-and-excitation dense convolution block of the present disclosure mainly includes 6 convolution blocks and 1 post-processing layer (including one layer of concat splicing and one CNN convolution with a kernel size of 1×1, aiming to adjust the number of output channels and dimensions to make the front and back dimensions and the number of channels consistent). The final output is x obtained after one iteration. Each convolution block is designed to add a squeeze-and-excitation module after each dense block, mainly including one layer of concat splicing (connecting multiple data sequences (such as arrays, tensors, etc.) together along a specific dimension to form a new and larger data structure), one CNN convolution with a kernel size of 3×3, one ReLu activation layer, and a squeeze-and-excitation module.

[0141] Step 1023, set the hyperparameters during model training, mainly including the number of iterations, the number of convolution block layers, the loss function, the optimizer, the learning rate, etc. The number of iterations refers to Figure 2 the value of T in Figure 3 ; the number of convolution block layers refers to -4 .

[0142]

[0143] Step 203, train the model by calculating the root mean square error between the model output and the real input data to obtain the electromagnetic reconstruction model. First, randomly shuffle and batch the training data through the DataLoader in PyTorch, and then send each sample (each batch of data is used as a sample) into the model for training in turn. Each time during training, calculate the root mean square error between the model output and the real input data, and let the model continuously learn and optimize automatically through backpropagation to update the model parameters. Finally, obtain the model weights with the minimum loss value as the electromagnetic reconstruction model.

[0144] Step 204, perform electromagnetic environment reconstruction according to the trained model. This step requires generating a brand-new electromagnetic data and sending it into the trained model for electromagnetic environment reconstruction. The new data is different from the training data and the test data, so as to effectively evaluate the performance of the model.

[0145] Through the above embodiments, the electromagnetic environment reconstruction algorithm based on deep compressive sensing can extract sufficient features from data under the condition of a small amount of sample data and complete the task of electromagnetic environment reconstruction. To address the problem of slow algorithm convergence speed, the present disclosure incorporates the NAG ISTA gradient update step, which improves the convergence speed and stability of the algorithm while making full use of image prior knowledge. By estimating a future position for the current position and calculating the gradient at this estimated future position, the algorithm can move faster towards the direction of the optimal solution. Compared with the original ISTA gradient update formula, it can achieve a similar or better optimization effect with fewer iterations. To address the problems of high complexity and high number of iterations in traditional compressive sensing, difficult parameter adjustment, and the large number of model parameters and insufficient feature acquisition in existing model-driven deep unfolding methods, the present disclosure proposes a squeeze-and-excitation dense convolutional block. The squeeze-and-excitation dense convolutional block uses SENet and DenseNet to unfold the iterative update step of ISTA into a deep neural network. Through compression and excitation operations, the importance of each feature channel is obtained, and then, according to the importance of the features, useful features are strengthened while unimportant features are weakened, enabling the model to adaptively recalibrate the importance of channel features. Through the dense connection of each convolutional layer, the network can make full use of the feature information of the previous layers, and at the same time utilize low-level and high-level features, promoting the reuse of features and the flow of information, and can more effectively capture important features of data from multiple scales and channels, thereby improving the reconstruction accuracy and indirectly reducing the dependence of the network model on the scale of training data.

[0146] It should be noted that the method of the embodiments of the present disclosure can be executed by a single device, such as a computer or a server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate with each other to complete it. In such a distributed scenario, one of the multiple devices can only execute one or more steps of the method of the embodiments of the present disclosure, and these multiple devices will interact with each other to complete the described method.

[0147] It should be noted that some embodiments of the present disclosure have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the above embodiments and still achieve the desired results. Additionally, the processes depicted in the figures do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0148] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present disclosure also provides an electromagnetic environment reconstruction device based on deep compressive sensing.

[0149] Reference Figure 8 , the electromagnetic environment reconstruction device based on deep compressive sensing includes:

[0150] An acquisition module 301, configured to acquire electromagnetic environment sample data, determine a data set from the electromagnetic environment sample data, and perform compressive sampling processing on the data set to obtain sample data;

[0151] A first update module 302, configured to perform preliminary compressive reconstruction through a gradient acceleration optimization algorithm, and calculate a current gradient of the current estimated value according to the sample data;

[0152] A second update module 303, configured to further compress and reconstruct through a squeeze-and-excitation dense convolutional network, and update the current estimated value according to the current gradient to obtain an updated estimated value;

[0153] A generation module 304, configured to, in response to determining that the number of updates reaches a preset update number threshold, use the deep compressive reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the target electromagnetic environment reconstruction model that has been trained.

[0154] In some embodiments, the first update module 302 includes:

[0155] An initialization unit, configured to initialize the sampled data in the sample data to obtain an initial estimated value, and use the initial estimated value as the current estimated value;

[0156] A gradient update unit, configured to perform preliminary compressive reconstruction through a gradient acceleration optimization algorithm, and calculate a current gradient of the current estimated value according to the current estimated value and the previous estimated value,

[0157] z (t) = q (t) - ρΦ T (Φq (t) - y)

[0158] where z (t) is the current gradient of the t-th iteration, ρ is the step size, Φ is the sampling matrix of the sample data, Φ T is the transpose matrix of the sampling matrix, y is the sampled data in the sample data, q (t) is the estimated value coefficient constructed based on the current estimated value and the previous estimated value,

[0159]

[0160] where x (t-1) is the current estimated value of the (t - 1)-th iteration, x (t-2)is the previous estimated value for the (t - 2)-th iteration, λ t is the regularization parameter for the t-th iteration, λ t-1 is the regularization parameter for the (t - 1)-th iteration,

[0161] In some embodiments, the second update module 303 includes:

[0162] A target feature learning unit configured to input the current gradient into a Dense Convolutional Network (DenseNet) to learn the target features of each layer;

[0163] A target feature enhancement unit configured to connect a squeeze-and-excitation module after the activation operation in each dense convolutional block of the Dense Convolutional Network (DenseNet), and enhance the target features according to the importance of each feature through a channel attention mechanism (SENet) and recalibrate the feature responses of the channels in an adaptive manner;

[0164] An estimated value update unit configured to update the current estimated value through multiple squeeze-and-excitation dense convolutional learning to obtain an updated estimated value.

[0165] In some embodiments, the acquisition module 301 includes:

[0166] A partitioning processing unit configured to determine a target electromagnetic region and perform partitioning processing on the target electromagnetic region to obtain grid cells;

[0167] A prediction module configured to predict the electromagnetic signal intensity of each grid cell;

[0168] A grid determination unit configured to determine a first grid for signal acquisition from the grid cells according to a preset sampling rate, and use the other grids in the grid cells except the first grid as a second grid;

[0169] An acquisition unit configured to acquire the electromagnetic signal intensity within the first grid, set the electromagnetic signal intensity within the second grid to a preset value, and obtain electromagnetic environment sample data;

[0170] A classification unit configured to divide the electromagnetic environment sample data into a training data set and a test data set according to a preset ratio.

[0171] In some embodiments, after further compressing and reconstructing through a squeeze-and-excitation dense convolutional network and updating the current estimated value according to the current gradient to obtain an updated estimated value, the apparatus further includes:

[0172] A test data set determination module configured to determine a test data set from the electromagnetic environment sample data;

[0173] A test module, configured to obtain the reconstructed electromagnetic signal strength according to the test data set through the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network;

[0174] A loss value determination module, configured to obtain the loss value between the reconstructed electromagnetic signal strength and the predicted electromagnetic signal strength based on a loss function;

[0175] A model determination module, configured to, in response to determining that the loss value is less than a preset loss threshold, use the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the target electromagnetic environment reconstruction model.

[0176] In some embodiments, the loss value determination module includes:

[0177] A loss function determination unit, configured to obtain the loss value between the reconstructed electromagnetic signal strength and the predicted electromagnetic signal strength based on a loss function,

[0178]

[0179] where Loss is the loss value, N is the number of samples in the test data set, Y i is the i-th reconstructed electromagnetic signal strength, and X i is the i-th predicted electromagnetic signal strength.

[0180] In some embodiments, after, in response to determining that the number of update times reaches a preset update times threshold, using the deep compression reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the trained target electromagnetic environment reconstruction model, the apparatus further includes:

[0181] A sampling data acquisition module, configured to acquire electromagnetic environment sampling data with a discrete distribution;

[0182] A reconstruction module, configured to perform reconstruction processing on the electromagnetic environment sampling data through the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network in the target electromagnetic environment reconstruction model to obtain a continuously distributed spectrum information distribution map.

[0183] For the convenience of description, when describing the above apparatus, various modules are described separately according to their functions. Of course, when implementing the present disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0184] The apparatus in the above embodiments is used to implement the corresponding electromagnetic environment reconstruction method based on deep compressive sensing in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated herein.

[0185] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present disclosure further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the method for electromagnetic environment reconstruction based on deep compressive sensing described in any one of the above embodiments is implemented.

[0186] Figure 9 FIG. shows a more specific schematic diagram of the hardware structure of the electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. Among them, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are communicatively connected to each other inside the device through the bus 1050.

[0187] The processor 1010 may be implemented in a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0188] The memory 1020 may be implemented in the form of a ROM (Read Only Memory), a RAM (Random Access Memory), a static storage device, a dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When implementing the technical solutions provided in the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 1020 and are called and executed by the processor 1010.

[0189] The input / output interface 1030 is used to connect to an input / output module to implement information input and output. The input / output module may be configured as a component in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Among them, the input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0190] The communication interface 1040 is used to connect to a communication module (not shown in the figure) to achieve communication and interaction between this device and other devices. The communication module can achieve communication through a wired method (such as USB (Universal Serial Bus), network cable, etc.) or through a wireless method (such as mobile network, WIFI (Wireless Fidelity), Bluetooth, etc.).

[0191] The bus 1050 includes a path for transmitting information between various components of the device (such as the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0192] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, those skilled in the art can understand that the above device may also only include the components necessary to implement the solution of the embodiments of this specification, and does not necessarily include all the components shown in the figure.

[0193] The electronic device in the above embodiments is used to implement the corresponding electromagnetic environment reconstruction method based on deep compressive sensing in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0194] Based on the same inventive concept, corresponding to the method in any of the above embodiments, the present disclosure also provides a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to cause the computer to execute the electromagnetic environment reconstruction method based on deep compressive sensing as described in any of the foregoing embodiments.

[0195] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0196] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the electromagnetic environment reconstruction method based on deep compressive sensing described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0197] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present application also provides a computer program product, including computer program instructions. When the computer program instructions run on a computer, the computer is caused to execute the electromagnetic environment reconstruction method based on deep compressive sensing described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0198] It can be understood that before using the technical solutions of the various embodiments in the present disclosure, the types, usage scopes, usage scenarios, etc. of the personal information involved will be informed to the user in an appropriate manner, and the user's authorization will be obtained.

[0199] For example, in response to receiving an active request from the user, a prompt message is sent to the user to clearly prompt the user that the operation requested by the user will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, an application program, a server, or a storage medium that performs the operations of the technical solutions of the present disclosure according to the prompt message.

[0200] As an optional but non-limiting implementation manner, the manner of sending a prompt message to the user in response to receiving an active request from the user may be, for example, in the form of a pop-up window, and the prompt message may be presented in text in the pop-up window. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0201] It can be understood that the above process of notifying and obtaining the user's authorization is only illustrative and does not limit the implementation manner of the present disclosure. Other manners that meet relevant laws and regulations can also be applied to the implementation manner of the present disclosure.

[0202] Those of ordinary skill in the art should understand that: the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present disclosure is limited to these examples; under the concept of the present disclosure, the technical features between the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of the embodiments of the present disclosure as described above, which are not provided in detail for the sake of brevity.

[0203] In addition, for simplicity of explanation and discussion, and so as not to make the embodiments of the present disclosure difficult to understand, well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Further, the devices may be shown in block diagram form in order to avoid making the embodiments of the present disclosure difficult to understand, and this also takes into account the fact that details of the implementation of these block diagram devices are highly dependent on the platform on which the embodiments of the present disclosure are to be implemented (i.e., these details should be entirely within the understanding of those skilled in the art). In cases where specific details (such as circuits) are set forth to describe exemplary embodiments of the present disclosure, it will be apparent to those skilled in the art that the embodiments of the present disclosure may be practiced without these specific details or with variations of these specific details. Accordingly, these descriptions should be considered illustrative rather than restrictive.

[0204] Although the present disclosure has been described in connection with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art in light of the foregoing description. For example, other memory architectures (such as dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0205] Embodiments of the present disclosure are intended to cover all such alternatives, modifications, and variations that fall within the broad scope of the present disclosure. Accordingly, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of the present disclosure shall be included within the protection scope of the present disclosure.

Claims

1. An electromagnetic environment reconstruction method based on deep compressive sensing, characterized in that, The method includes: Obtaining electromagnetic environment sample data, determining a data set from the electromagnetic environment sample data, and performing compressive sampling processing on the data set to obtain sample data; Performing preliminary compressive reconstruction through a gradient acceleration optimization algorithm, and calculating a current gradient of the current estimated value based on the sample data; Performing further compressive reconstruction through a squeeze-and-excitation dense convolutional network, and updating the current estimated value based on the current gradient to obtain an updated estimated value; In response to determining that the number of updates reaches a preset update number threshold, using the deep compressive reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the target electromagnetic environment reconstruction model that has been trained.

2. The method according to claim 1, characterized in that, The performing preliminary compressive reconstruction through a gradient acceleration optimization algorithm, and calculating a current gradient of the current estimated value based on the sample data includes: Initializing the sampled data in the sample data to obtain an initial estimated value, and using the initial estimated value as the current estimated value; Performing preliminary compressive reconstruction through a gradient acceleration optimization algorithm, and calculating a current gradient of the current estimated value based on the current estimated value and the previous estimated value, z (t) = q (t) - ρΦ T (Φq (t) - y) where z (t) is the current gradient of the t-th iteration, ρ is the step size, Φ is the sampling matrix of the sample data, and Φ T is the transpose matrix of the sampling matrix, y is the sampled data in the sample data, and q (t) is the estimated value coefficient constructed based on the current estimated value and the previous estimated value. where x (t-1) is the current estimate of the (t - 1)-th iteration, x (t-2) is the previous estimate of the (t - 2)-th iteration, λ t is the regularization parameter of the t-th iteration, λ t-1 is the regularization parameter of the (t - 1)-th iteration, 3. The method according to claim 1, wherein The performing further compressive reconstruction through a squeeze-and-excitation dense convolutional network, and updating the current estimated value based on the current gradient to obtain an updated estimated value includes: Inputting the current gradient into a dense convolutional network DenseNet to learn the target features of each layer; Connecting a squeeze-and-excitation module after performing an activation operation on each dense convolutional block of the dense convolutional network DenseNet, and strengthening the target features according to the importance of each feature through a channel attention mechanism SENet and recalibrating the feature responses of the channels in an adaptive manner; Performing multiple squeeze-and-excitation dense convolutional learning to update the current estimated value to obtain an updated estimated value.

4. The method according to claim 1, wherein The obtaining electromagnetic environment sample data includes: Determining a target electromagnetic region, and performing division processing on the target electromagnetic region to obtain grid cells; Predicting the electromagnetic signal intensity of each grid cell; Determining a first grid for signal collection from the grid cells according to a preset sampling rate, and using the other grids in the grid cells except the first grid as a second grid; Collecting the electromagnetic signal intensity within the first grid, and setting the electromagnetic signal intensity within the second grid to a preset value to obtain electromagnetic environment sample data; Dividing the electromagnetic environment sample data into a training data set and a test data set according to a preset ratio.

5. The method according to claim 1, wherein After the performing further compressive reconstruction through a squeeze-and-excitation dense convolutional network, and updating the current estimated value based on the current gradient to obtain an updated estimated value, it further includes: Determining a test data set from the electromagnetic environment sample data; Obtaining a reconstructed electromagnetic signal intensity according to the test data set through the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network; Obtaining a loss value between the reconstructed electromagnetic signal intensity and the predicted electromagnetic signal intensity based on a loss function; In response to determining that the loss value is less than a preset loss threshold, the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network are used as the trained target electromagnetic environment reconstruction model.

6. The method according to claim 5, wherein The obtaining of the loss value of the reconstructed electromagnetic signal intensity and the predicted electromagnetic signal intensity based on the loss function includes: Obtaining the loss value of the reconstructed electromagnetic signal intensity and the predicted electromagnetic signal intensity based on the loss function. Where Loss is the loss value, N is the number of samples in the test data set, Y i is the intensity of the i-th reconstructed electromagnetic signal, and X i is the intensity of the i-th predicted electromagnetic signal.

7. The method according to claim 1, wherein After, in response to determining that the number of update times reaches a preset update times threshold, using the deep compression reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the trained target electromagnetic environment reconstruction model, it further includes: Obtaining discretely distributed electromagnetic environment sampling data; Performing reconstruction processing on the electromagnetic environment sampling data through the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network in the target electromagnetic environment reconstruction model to obtain a continuously distributed spectrum information distribution map.

8. An electromagnetic environment reconstruction device based on deep compressive sensing, characterized in that, Including: An acquisition module configured to acquire electromagnetic environment sample data, determine a data set from the electromagnetic environment sample data, and perform compressive sampling processing on the data set to obtain sample data; A first update module configured to perform preliminary compression reconstruction through a gradient acceleration optimization algorithm, and calculate the current gradient of the current estimated value according to the sample data; A second update module configured to further compress and reconstruct through a squeeze-and-excitation dense convolutional network, and update the current estimated value according to the current gradient to obtain an updated estimated value; A generation module configured to, in response to determining that the number of update times reaches a preset update times threshold, use the deep compression reconstruction model composed of the gradient acceleration optimization algorithm and the squeeze-and-excitation dense convolutional network as the trained target electromagnetic environment reconstruction model.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the program, it implements the method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the method according to any one of claims 1 to 7.

Images