3D spectrum completion and prediction method based on auto-encoder network

Abstract

The invention discloses a 3D frequency spectrum completion and prediction method based on an auto-encoder network, and the method comprises the steps: carrying out the region segmentation of a to-be-predicted three-dimensional geographic region based on a grid quantization method, carrying out the spatial sparse sampling through employing an unmanned plane with a preset track, and obtaining the received signal intensity data of a frequency spectrum; based on a multi-scale feature fusion module and an auto-encoder network of a three-dimensional space attention mechanism, performing space completion on the acquired incomplete received signal strength data to obtain spectrum completion data; a deep learning-based spectrum prediction network extracts key features of time, space and frequency from the spectrum completion data, predicts spectrum information in a future time period, and outputs predicted spectrum data; according to the method, missing spectrum data are recovered by extracting three-dimensional space features and multi-scale feature fusion, and space, time and frequency correlation features are extracted by using a deep learning multi-branch spectrum prediction network based on a spectrum completion result, so that the spectrum prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic spectrum perception prediction, and in particular to a 3D spectrum completion and prediction method based on an autoencoder network. Background Art

[0002] With the advancement of 6G technology and the widespread deployment of integrated air, land, and sea networks, the shortage of spectrum resources and their low utilization have become a major challenge that needs to be solved urgently. Dynamic spectrum access (DSA) optimizes spectrum allocation through real-time spectrum sensing to address this problem. Compared with spectrum sensing, spectrum prediction uses historical time slot data to identify potential evolution patterns and infers future spectrum conditions from past spectrum samples, solving the timeliness and energy consumption problems of spectrum sensing. In recent years, many spectrum prediction techniques based on Bayesian reasoning, matrix completion and other methods have been proposed. And with the development of neural networks, spectrum prediction techniques based on neural networks have been applied by automatically learning and analyzing historical spectrum sensing data. Commonly used techniques include long short-term memory networks (LSTM), convolutional LSTM (ConvLSTM) and hierarchical learning systems, the latter of which usually uses fine-tuned convolutional neural networks (CNN) and gated recurrent units (GRU). In addition, CNN-based residual neural networks (ResNet) are also used in existing work to cope with complex electromagnetic spectrum composition and radio propagation environments.

[0003] However, since not all areas in the real world are accessible and the flight trajectory of spectrum data collection in actual scenarios is limited, the collected spectrum data is incomplete, which will further affect the accuracy of spectrum prediction. Therefore, in order to improve the prediction accuracy of future spectrum data, spectrum completion technology in the spatial dimension is crucial. The main idea of ​​spectrum completion is to model the correlation between the collected spectrum data to restore the missing spectrum data. In recent years, many spectrum data completion schemes have been proposed, among which the most widely used are nearest neighbor (NN) interpolation, inverse distance weighted (IDW) interpolation and Kriging method (a regression algorithm for spatial modeling and prediction of random processes / random fields based on covariance functions). Later, researchers also proposed a three-dimensional spectrum map reconstruction algorithm based on total variation (3DTV-SMR) and a spectrum map completion algorithm based on improved generative adversarial network (3D-DCRGAN).

[0004] However, the above spectrum prediction methods lack a powerful spectrum completion solution for complex scenarios with multiple radiation sources. The fuzzy spectrum completion results will affect the subsequent spectrum prediction performance. In addition, the above work fails to design the network structure in a targeted manner in the prediction stage to utilize the three-dimensional spatial dependency, time dependency, and frequency dependency. In addition, these methods are mainly designed for ground communication scenarios. With the transition from ground to integrated communication systems in the air, land, and sea, the ability to extract higher-dimensional spatial features is crucial for three-dimensional spectrum prediction scenarios in complex 3D space scenarios with multiple radiation sources.

[0005] Therefore, the prior art needs to be improved. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a 3D spectrum completion and prediction method based on an autoencoder network in response to the defects of the prior art, so as to solve the problem that the accuracy of the spectrum prediction historical data recovery scheme in complex scenarios with 3D multiple radiation sources is low, and the spectrum prediction network is difficult to specifically extract the complex spatiotemporal-frequency features in the spectrum data.

[0007] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0008] In a first aspect, the present invention provides a 3D spectrum completion and prediction method based on an autoencoder network, comprising:

[0009] The three-dimensional geographic area to be predicted is segmented based on the grid quantization method, and a drone with a preset trajectory is used for spatial sparse sampling to obtain the received signal strength data of the spectrum;

[0010] Based on the autoencoder network of the multi-scale feature fusion module and the three-dimensional spatial attention mechanism, the collected incomplete received signal strength data is spatially completed to obtain spectrum completion data;

[0011] The spectrum prediction network based on deep learning extracts the key features of time, space and frequency from the spectrum completion data, predicts the spectrum information of future time periods, and outputs the predicted spectrum data.

[0012] In one implementation, the grid quantization method is used to segment the three-dimensional geographic area to be predicted, and a drone with a preset trajectory is used to perform spatial sparse sampling to obtain received signal strength data of the spectrum, including:

[0013] Based on the grid quantization method, the three-dimensional geographic area to be predicted is segmented to obtain N L ×N W ×N H grid subspaces; where N L , N W , NH are the number of quantization intervals in the three dimensions of space respectively; each grid constitutes a regional subspace, and the interval of each grid is Δd. Assuming that the received signal strength in each grid is consistent, the grid index I = (i, j, k) represents the actual geographical location x (i,j,k) =(i+0.5)Δd,(j+0.5)Δd,(k+0.5)Δd);

[0014] And adopt the UAV of the preset trajectory to perform spatial sparse sampling to obtain the received signal strength data of the spectrum; wherein the received signal strength data is the missing spectrum data.

[0015] In one implementation, the autoencoder network based on the multi-scale feature fusion module and the three-dimensional spatial attention mechanism performs spatial completion on the collected incomplete received signal strength data to obtain spectrum completion data, including:

[0016] Extracting the spectral spatial distribution features based on 3D convolution and compressing the received signal strength data to obtain semantic features with an expanded receptive field, and allocating corresponding weights to different parts of the three-dimensional feature map extracted from the received signal strength data based on a three-dimensional spatial attention mechanism to automatically evaluate the importance of spatial features at various locations;

[0017] The hidden feature representation is expanded to a dimension consistent with the received signal strength data based on a 3D transposed convolutional layer, and feature integration is performed based on a multi-scale feature fusion module to obtain the spectrum completion data.

[0018] In one implementation, the step of performing feature integration based on a multi-scale feature fusion module to obtain the spectrum completion data includes:

[0019] In the feature integration stage, based on the multi-scale feature fusion module, dilated convolution and ordinary convolution are jointly used to expand the receptive field, and multi-scale feature fusion is performed through a parallel multi-branch structure to extract three-dimensional spatial features and obtain the spectrum completion data.

[0020] In one implementation, the prediction process of the spectrum prediction network is defined as:

[0021]

[0022] Among them, g θ represents the spectrum prediction network;

[0023] Δt represents the prediction range;

[0024] Indicates received signal strength data including multiple frequency bands and spatial positions after spectrum completion;

[0025] T represents the number of time slots of historical spectrum data.

[0026] In one implementation, the spectrum prediction network based on deep learning extracts key features of time, space and frequency from the spectrum completion data and predicts spectrum information for a future period, including:

[0027] A three-dimensional convolution structure is used to extract the three-dimensional spatial features of the spectrum completion data, and the regional spatial correlation features of different spaces are obtained by combining a three-dimensional spatial attention module and a jump connection evaluation;

[0028] Utilizing 1D convolution combined with 1D attention to extract frequency correlation features of the spectrum completion data;

[0029] Extracting time-correlation features of the spectrum completion data using a gated recurrent unit network;

[0030] The various spectral semantic feature information extracted by different branches of the spectrum prediction network are spliced ​​and fused to finally realize prediction, and the various features extracted by the multi-branch feature extraction structure are combined to realize the prediction of spectrum information in the future time period.

[0031] In a second aspect, the present invention provides a 3D spectrum completion and prediction device based on an autoencoder network, comprising:

[0032] The sampling module is used to segment the three-dimensional geographic area to be predicted based on the grid quantization method, and use a drone with a preset trajectory to perform spatial sparse sampling to obtain the received signal strength data of the spectrum;

[0033] The spectrum completion module is used to spatially complete the collected incomplete received signal strength data based on the autoencoder network of the multi-scale feature fusion module and the three-dimensional spatial attention mechanism to obtain spectrum completion data;

[0034] The spectrum prediction module is used to extract the key features of time, space and frequency of the spectrum completion data based on the spectrum prediction network of deep learning, predict the spectrum information of the future time period, and output the predicted spectrum data.

[0035] In a third aspect, the present invention provides a terminal, comprising: a processor and a memory, wherein the memory stores a 3D spectrum completion and prediction program based on an autoencoder network, and when the 3D spectrum completion and prediction program based on an autoencoder network is executed by the processor, it is used to implement the operation of the 3D spectrum completion and prediction method based on the autoencoder network as described in the first aspect.

[0036] In a fourth aspect, the present invention further provides a medium, which is a computer-readable storage medium, and the medium stores a 3D spectrum completion and prediction program based on an autoencoder network. When the 3D spectrum completion and prediction program based on an autoencoder network is executed by a processor, it is used to implement the operation of the 3D spectrum completion and prediction method based on an autoencoder network as described in the first aspect.

[0037] The present invention adopts the above technical solution to achieve the following effects:

[0038] 1) This paper proposes a framework that integrates spectrum completion and spectrum prediction, which solves the problem of missing historical spectrum data caused by the limited sampling trajectory of drones in three-dimensional scenes. Based on accurate completion results and a powerful spectrum prediction network that extracts spatiotemporal and frequency features respectively, the accuracy of spectrum prediction is effectively improved.

[0039] 2) The present invention introduces an autoencoder network equipped with a three-dimensional spatial attention mechanism and a multi-scale feature fusion module, which can improve the spectrum completion performance in complex scenes, thereby further enhancing the subsequent spectrum prediction accuracy.

[0040] 3) The present invention adopts a multi-branch network structure to extract three types of features through different network branches respectively, and then combines the correlation information extracted from the branches for prediction, so as to more effectively utilize the complex three-dimensional spatial dependency, time dependency and frequency dependency in the spectrum data, and reduce the error of spectrum prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0042] Figure 1 It is a flow chart of the 3D spectrum completion and prediction method based on the autoencoder network in the present invention.

[0043] Figure 2 It is a schematic diagram of the overall architecture of 3D spectrum completion and prediction in the present invention.

[0044] Figure 3 Schematic diagram of an encoder-based spectrum completion network with multi-scale feature fusion in the present invention.

[0045] Figure 4 It is a schematic diagram of the multi-scale feature fusion module in the completion network of the present invention.

[0046] Figure 5 It is a structural block diagram of the spectrum prediction model in the present invention.

[0047] Figure 6 It is a schematic diagram of the change of the mean square error of various completion schemes in the present invention with the number of training rounds.

[0048] Figure 7 It is a schematic diagram showing the variation of the errors of various combinations of completion prediction schemes in the present invention with the number of training rounds.

[0049] Figure 8 It is a functional principle diagram of a terminal in one implementation of the present invention.

[0050] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0052] Exemplary Methods

[0053] Existing spectrum prediction methods lack a powerful 3D spectrum completion solution. The fuzzy and inaccurate spectrum completion results will affect the subsequent spectrum prediction performance. In addition, the above work fails to design the network structure in a targeted manner in the prediction stage to utilize the three-dimensional spatial dependency, time dependency, and frequency dependency. In addition, these methods are mainly designed for ground communication scenarios. With the transition from ground to integrated air, land, and sea communication systems, the ability to extract higher-dimensional and relatively complex spatial features is crucial for three-dimensional spectrum prediction scenarios.

[0054] In view of the above technical problems, an embodiment of the present invention provides a 3D spectrum completion and prediction method based on an autoencoder network. The method performs regional segmentation on the three-dimensional geographical area to be predicted based on a grid quantization method, and uses a drone with a preset trajectory for spatial sparse sampling to obtain the RSS (Received signal strength) data of the spectrum; based on the MSFF module (Multi-scale feature fusion module) and the autoencoder network of the three-dimensional space attention mechanism, the collected incomplete RSS data is spatially completed to obtain spectrum completion data; the spectrum prediction network based on deep learning extracts the key features of time, space and frequency of the spectrum completion data, predicts the spectrum information of the future time period, and outputs the predicted spectrum data; therefore, the embodiment of the present invention restores the missing spectrum data by extracting three-dimensional spatial features and multi-scale feature fusion, and based on the results of spectrum completion, a multi-branch spectrum prediction network based on deep learning can be used to extract spatial, temporal and frequency features, thereby effectively improving the accuracy of spectrum prediction. In terms of spectrum completion, the scheme proposed in this embodiment is superior to the benchmark algorithm in terms of mean square error (MSE). More importantly, with better completion results and a carefully designed prediction network, the prediction performance is further improved.

[0055] like Figure 1 As shown, an embodiment of the present invention provides a 3D spectrum completion and prediction method based on an autoencoder network, comprising the following steps:

[0056] Step S100, segmenting the three-dimensional geographic area to be predicted based on a grid quantization method, and performing spatial sparse sampling using a drone with a preset trajectory to obtain RSS data of the spectrum.

[0057] Specifically, in an implementation of this embodiment, step S100 includes the following steps:

[0058] Step S101, segmenting the three-dimensional geographic area to be predicted based on the grid quantization method to obtain N K ×N W ×N H grid subspaces;

[0059] Step S102, using the UAV with the preset trajectory to perform spatial sparse sampling to obtain RSS data of the spectrum; wherein the RSS data is the missing spectrum data.

[0060] In this embodiment, if Figure 2As shown, in order to facilitate the training of the 3D spectrum completion and prediction network, according to the signal propagation mechanism in the actual scene, this embodiment constructs a simulation data set to realize the simulation of the actual electromagnetic spectrum situation. In the scene setting, this embodiment considers a three-dimensional geographical region of interest, namely ROI (Region of interest), and uses a grid quantization method to divide the three-dimensional geographical region to be predicted in the scene into three-dimensional regions, and each small three-dimensional grid divided constitutes an ROI subspace, and each subspace is equally spaced. And because the RSS inside the grid changes little, it is assumed that the RSS of each position inside each ROI subspace is consistent; in the radiation source setting, this embodiment considers multiple ground base station transmitters transmitting signals at a constant speed in a random walk mode, where the radiation source trajectory is continuous in time and space; in the channel modeling setting, the simulation data set constructed by this embodiment considers free space path loss and shadow to construct the signal propagation function from the radiation source to each ROI subspace; in the sampling setting, a UAV with a preset trajectory is used to sample the grid subspace, and the UAV equipped with an aerial spectrum detection device hovers in each ROI subspace to monitor the spectrum intensity.

[0061] like Figure 1 As shown, an embodiment of the present invention provides a 3D spectrum completion and prediction method based on an autoencoder network, comprising the following steps:

[0062] Step S200, based on the autoencoder network of the MSFF module and the three-dimensional spatial attention mechanism, the collected incomplete RSS data is spatially completed to obtain spectrum completion data.

[0063] In this embodiment, according to the observation of the drone, Figure 3 As shown in the figure, an autoencoder network (AE-MSFF) with MSFF module and 3D spatial attention mechanism is proposed. This can capture and integrate 3D spatial dependencies and help complete the 3D spectrogram to solve the problem of missing historical spectrum data caused by the limited sampling trajectory of unmanned aerial vehicles (UAVs) in 3D scenes.

[0064] The autoencoder neural network proposed in this embodiment is a deep neural network designed for pixel-level semantic segmentation. Its architecture consists of two main components: a contraction path (encoder) and an expansion path (decoder). This embodiment defines the encoder network as And the decoder network is Wherein, θ1 and θ2 represent the network parameters of the encoder and decoder respectively. The autoencoder network of this embodiment completes the spectrum through the following process:

[0065]

[0066] In this embodiment, the input and output of the network are slices of RSS data. Represent the normalized, incomplete three-dimensional spectrum and the complete three-dimensional spectrum respectively. The training is performed to learn the changing trend characteristics of radio propagation phenomena. Specifically, the training model acquires knowledge of the frequency attenuation characteristics in the frequency dimension and the spatial attenuation characteristics in different geographical dimensions. The goal of the training process is to ensure that the output of the autoencoder network proposed in this embodiment is as close as possible to the actual distribution of the spectrum data in the data set, which can be described as:

[0067]

[0068] Where L represents the number of training samples, l represents the lth training sample, ∥·∥ F is the Frobenius norm.

[0069] Specifically, in an implementation of this embodiment, step S200 includes the following steps:

[0070] Step S201, based on 3D convolution, extracts spectral spatial distribution features and compresses the RSS data to obtain semantic features with an expanded receptive field, and based on the three-dimensional spatial attention mechanism, allocates corresponding weights to different parts of the three-dimensional feature map extracted from the RSS data to automatically evaluate the importance of spatial features at each location.

[0071] In this embodiment, if Figure 3 As shown, the first half of the completion network uses the encoder network It is used to complete the downsampling in the network feature extraction stage. In this stage, 3D convolution is used to extract distribution features and compress the original RSS data to obtain high-level semantic features with an expanded receptive field. In addition, based on the overall autoencoder network architecture, a three-dimensional spatial attention module is proposed to enhance the network feature extraction capability. The three-dimensional spatial attention module assigns corresponding weights to different parts of the three-dimensional feature map extracted from the RSS data.

[0072] 3D Spatial Attention Module:

[0073] This embodiment mainly uses 3D convolution to extract 3D spatial dependencies in the network proposed in the feature extraction stage. On this basis, this embodiment uses a 3D spatial attention mechanism to enhance the ability to extract spatial features from spectral data, which is placed in the feature extraction stage. This mechanism dynamically assigns different weights to different parts of the 3D feature map from RSS data, allowing the neural network to adjust its focus based on the calculated importance scores of the spatial features. This adaptive focus significantly promotes continuous optimization during training and helps the model performance to be more refined.

[0074] Specifically, in an implementation of this embodiment, step S200 further includes the following steps:

[0075] Step S202: Expand the hidden feature representation to a dimension consistent with the RSS data based on a 3D transposed convolution layer, and perform feature integration based on an MSFF module to obtain the spectrum completion data.

[0076] In this embodiment, if Figure 3 As shown, the second half of the completion network uses a decoder network For upsampling and feature integration, at this stage, the advanced 3D transposed convolution layer is mainly used to expand the hidden feature representation to the same dimension as the original data, and convert the high-level information in the encoder into a more refined feature representation. Based on the overall autoencoder network architecture, an MSFF module is proposed to enhance the network feature integration capability, and feature integration is performed based on the MSFF module to obtain the spectrum completion data.

[0077] In one implementation of this embodiment, step S202 includes the following steps:

[0078] Step S202a, in the feature integration stage, based on the MSFF module, dilated convolution and ordinary convolution are jointly used to expand the receptive field, and multi-scale feature fusion is performed through a parallel multi-branch structure to extract three-dimensional spatial features and obtain the spectrum completion data.

[0079] In this embodiment, dilated convolution and ordinary convolution are used to expand the receptive field, and feature fusion is performed in a parallel multi-branch manner. These feature fusion processes are implemented through the following MSFF module and the channel attention module of MSFF.

[0080] MSFF Module:

[0081] In the feature integration stage, this embodiment proposes a MSFF module to enhance the extraction capability of spatial multi-scale features. Figure 4 As shown in the figure, the module expands the receptive field by combining dilated convolution and conventional convolution, and effectively performs multi-scale feature fusion through a parallel branch structure, thereby strengthening the feature processing capability of the feature integration stage. The basis of the MSFF module is dilated convolution, which expands the convolution kernel by filling 0 in the middle of the convolution kernel without increasing the convolution parameters. Using dilated convolution can effectively extract three-dimensional spatial features without significantly losing information or increasing parameters. However, due to the limited ability to capture details close to the radiation source under high dilation rate, the present embodiment uses the MSFF module in this network. This module ensures complete coverage of the receptive field by adjusting the dilation coefficient of each branch.

[0082] In this embodiment, a channel attention module is also set at the end of the above-mentioned MSFF module to further enhance the network feature integration capability.

[0083] Channel attention part of MSFF:

[0084] For RSS data collected by drones, feature extraction must pay attention to detailed information near the location of the radiation source, while also capturing the global trend of RSS changes. Convolutional neural networks extract target features through a layer-by-layer abstraction process. High-level networks have large receptive fields and strong semantic information representation capabilities. However, it has deficiencies in the resolution of feature maps, and the geometric information representation (especially the lack of detailed spatial geometric features) is weak. In contrast, low-level networks have relatively small receptive fields but have strong capabilities in representing geometric details. Although this network provides high resolution, its ability to represent semantic information is still limited. In dilated convolution, different dilation exponents correspond to different receptive fields, and also correspond to features of different scales. The MSFF module improves data completion efficiency by extracting and integrating features from the above levels. The multi-scale fusion scheme mentioned above in this embodiment only considers the simple superposition of features of all branches, and features of different scales are distributed in different channels. However, not all features are equally important; therefore, the channel attention module constitutes the last component of the MSFF module. Similar to spatial attention, this scheme enables the neural network to selectively focus on features of different scales on different channels during training iterations, automatically evaluating the importance of extracted features of different scales to achieve stronger spectral completion effects.

[0085] These enhanced network structures and mechanisms significantly improve the performance of spectrum completion and provide a solid foundation for accurate spectrum prediction.

[0086] like Figure 1 As shown, an embodiment of the present invention provides a 3D spectrum completion and prediction method based on an autoencoder network, comprising the following steps:

[0087] Step S300: extract key features of time, space and frequency from the spectrum completion data based on a deep learning spectrum prediction network, predict spectrum information for a future period of time, and output predicted spectrum data.

[0088] In this embodiment, based on the completion results obtained by the above-mentioned spectrum completion network, a new spectrum prediction scheme based on deep learning is proposed, aiming to predict the spectrum status in the future time period.

[0089] The spectrum prediction problem based on spectrum data in this embodiment can be expressed as follows:

[0090]

[0091] Among them, p T+Δt =[p T+1 ,...,p T+Δt ], Δt represents the prediction range, It is the RSS data containing multiple frequency bands and spatial positions after spectrum completion, and T is the number of time slots of historical spectrum data.

[0092] Assume that the neural network proposed in this embodiment is represented by g θ (·), the prediction process of the spectrum prediction network is defined as:

[0093]

[0094] The spectrum prediction solution based on deep learning proposed in this embodiment adopts a "divide and conquer" strategy. Figure 5 As shown, the present embodiment designs three different network branches respectively, which focus on the extraction of key features of time, space and frequency, so as to more thoroughly analyze the corresponding dependency information in the spectrum history data, so that the neural network can better abstract the semantic information of time, space and frequency from the completed spectrum data in the step-by-step feature extraction. By integrating various features from multiple branches and the extracted semantic information, the present embodiment can better utilize the complexity and diversity of high-dimensional spectrum data. The prediction network of the present embodiment has a powerful feature extraction capability, and combined with the excellent performance of the previous spectrum completion network, the effect of spectrum prediction is significantly enhanced. The following are the design concepts of the three feature extraction branches:

[0095] Specifically, in an implementation of this embodiment, step S300 includes the following steps:

[0096] Step S301, using a three-dimensional convolution structure to extract the three-dimensional spatial features of the spectrum completion data, and combining a three-dimensional spatial attention module and a jump connection evaluation to obtain regional spatial correlation features of different spaces.

[0097] A. Spatial feature extraction:

[0098] The historical RSS data contains many continuous spatial sampling points. The distances from the same radiation source to adjacent ROI subspaces are relatively close, which indicates that there is local spatial dependence in the historical spectrum data. Therefore, this embodiment uses a three-dimensional convolutional structure with strong local feature extraction capabilities to extract three-dimensional spatial features. In addition, this embodiment also combines a three-dimensional spatial attention module and a jump connection to evaluate the spatial features of different spatial regions, which enhances the focus of feature extraction and ensures the stability of the multi-layer network.

[0099] It is worth noting that the current spectrum prediction method converts 2D spatial data into 1D through ROI subspace indexing, but this processing scheme may not preserve the proximity of the converted subspace. This method cannot fully capture the three-dimensional spatial dependency in local areas, especially in data with spatial continuity collected by drones. Therefore, considering these factors, this embodiment uses two three-dimensional convolutional layers with a three-dimensional spatial attention mechanism for spatial feature extraction.

[0100] Specifically, in an implementation of this embodiment, step S300 further includes the following steps:

[0101] Step S302: 1D convolution combined with 1D attention is used to extract frequency correlation features of the spectrum completion data.

[0102] In this embodiment, when extracting spatial features, 1D convolution combined with 1D attention can also be applied to extract frequency features in spectrum completion data.

[0103] B. Frequency feature extraction:

[0104] Regarding frequency dependence, the path loss formula shows that adjacent frequency bands at the same location and time usually exhibit similar path losses, resulting in similar related RSS values. Therefore, the frequency dependence between frequency bands also tends to be local. Therefore, this embodiment applies 1D convolution combined with 1D attention to extract frequency features.

[0105] Specifically, in an implementation of this embodiment, step S300 further includes the following steps:

[0106] Step S303: using a GRU network (Gated recurrent unit) to extract the time correlation features of the spectrum completion data.

[0107] In this embodiment, when extracting frequency features, the time correlation features in the spectrum completion data may also be extracted through a GRU network.

[0108] C. Temporal feature extraction:

[0109] Regarding time dependency, given the superior ability of the GRU network structure in extracting time features, this embodiment uses GRU for this purpose.

[0110] Specifically, in an implementation of this embodiment, step S300 further includes the following steps:

[0111] Step S304, the various spectral semantic feature information extracted by different branches of the spectrum prediction network are spliced ​​and fused to finally achieve prediction, and the various features specifically extracted by the multi-branch feature extraction structure are combined to achieve prediction of spectrum information in the future period.

[0112] After the neural network extracts three types of associations through three branches respectively, this embodiment connects the extracted feature maps to predict the spectrum of the future period. A better feature extraction structure helps to achieve more accurate prediction.

[0113] In order to illustrate the feasibility and advantages of this embodiment, this embodiment uses the scenario introduced above and the modeled data set for experimental verification, the number of divided sub-bands F=32, and the number of transmitters NT in the training set is randomly selected from 1 to 3, that is, NT∈{1,2,3}.

[0114] Specifically, the scene and modeling process are as follows:

[0115] In the simulation process, scene modeling is required. In this embodiment, a three-dimensional geographic region (ROI) to be predicted with a size of L×W×H is considered. In order to construct a spectrum graph, the region is divided into N L ×N W ×N H grids, using the grid quantization method. Each grid constitutes a subspace, and the grid interval is Δd = 1km. Since the RSS changes little within the grid, this embodiment assumes that the RSS within each grid is consistent. Each grid index I = (i, j, k), where i ranges from 0 to L-1, j ranges from 0 to W-1, and k ranges from 0 to H-1. The index corresponds to the actual geographic location x (i,j,k) =(i+0.5)Δd, (j+0.5)Δd, (k+0.5)Δd).

[0116] NT ground base station transmitters transmit signals on K frequency bands. The index of the frequency band is defined as f∈{1,...,F} in this embodiment. The transmitter continuously transmits signals in a random walk mode. In this mode, each transmitter moves in the same direction at a constant speed for 10 seconds, and then randomly selects a new direction and speed. The speed variation range is from 1.38m / s to 2.88m / s. In the experiment, the trajectory of the transmitter is continuous in space and time. Considering the limitations of three-dimensional geographic scenes, it is difficult to collect spectrum data from sensors at fixed positions due to geographical restrictions. Therefore, an unmanned aerial vehicle (UAV) with a preset trajectory is used to spatially sample historical spectrum data. For efficient sampling, a drone equipped with an aerial spectrum detection device can hover at a precise location to detect the spectrum power of multiple bandwidths.

[0117] In this embodiment, the signal propagation function of the RSS data is The path loss model for distance d from the transmitter to the regional subspace is:

[0118] PL(d)=PL(d0)+10αlog 10 (d / d0)+X g ;

[0119] Where d0 represents the reference distance, PL(d0) represents the path loss at a distance d0 from the radiation source, α represents the path loss exponent, and X g represents shadow; the path loss exponent α is determined by the propagation environment.

[0120] Assume that the RSS of the region subspace with coordinates (i, j, k) is expressed as It is calculated by the following formula:

[0121]

[0122] Among them, σ 2 represents the variance of additive white noise, represents the transmission power of the nth transmitter, NT represents the number of radiation sources, and the indexes f, t and (i, j, k) correspond to the frequency, time and position index of the 3D area subspace respectively.

[0123] In this embodiment, the complete RSS of the multidimensional grid is defined as p f,t , the index set of the actual sampling position of the drone is Ω, so the actual sampling data is After a specific completion scheme, the reconstructed complete RSS data is:

[0124] This embodiment conducts experiments for 50 time slots, of which the first 80% are used as historical data and the last 20% are used as future data. In the spectrum sampling stage, under the restricted trajectory configuration, it is assumed that the UAV flies along 7 circular trajectories at different altitudes (except the lowest altitude) to sample the spectrum data of historical time slots. Before the sampled data is input into the neural network, data preprocessing is performed to reduce the training complexity and improve the training efficiency and learning ability of the model. First, the non-zero data in the spectrum data is normalized to the minimum-maximum value, expressed as:

[0125]

[0126] in They are the maximum and minimum RSS values ​​in the historical data slices, respectively.

[0127] After the above normalization process, a single-channel "grayscale image" is obtained. After the "grayscale image" is expanded to three channels, the unsampled grid is painted red in this embodiment to obtain the input of the neural network. The mean square error (MSE) is used to evaluate the performance of the completion and prediction algorithms.

[0128] In order to complete the RSS data, four schemes are evaluated in this embodiment: NN interpolation, IDW interpolation, a completion scheme based on an improved GAN network (3D-DCRGAN), and the spectrum completion scheme (AE-MSFF) proposed in this embodiment.

[0129] like Figure 6 As shown, the figure shows the MSE curves of the solution in this embodiment and the completion solution based on the improved GAN network in the training rounds, where the noise power is set to -50dBm. Figure 6 As shown in , with the increase of training rounds, MSE gradually stabilizes. Figure 6 Two gentle lines are introduced in , corresponding to the two interpolation methods mentioned above (i.e., NN interpolation and IDW interpolation). IDW interpolation uses distance to construct weights and usually produces smoother and more accurate results than NN interpolation, which is reflected in its lower MSE value. Figure 6 As shown, after the two spectrum completion schemes based on neural networks have reached convergence after multiple rounds of training, the MSE of the completion is significantly smaller than the two interpolation methods. This is because the appropriate neural network can capture more complex nonlinear spatial features from a large number of data sets, and the completion scheme based on neural networks is more advantageous for high-dimensional data completion with relatively complex features. In addition, compared with the two neural networks, the AE-MSFF scheme proposed in this embodiment has a better MSE than the 3D-DCRGAN scheme after the network training converges. Therefore, the addition of the 3D spatial attention mechanism and the multi-scale fusion module in this embodiment significantly enhances the network's ability to extract complex spatial features, making the scheme proposed in this embodiment the most effective.

[0130] like Figure 7 As shown, in this embodiment, the proposed spectrum prediction scheme based on the "divide and conquer" strategy is evaluated against the traditional ConvLSTM model method. Figure 7 The curves in are the MSEs of the spectrum prediction schemes using the ConvLSTM structure and the one proposed in this embodiment in the training rounds. The two spectrum prediction schemes are predicted based on the spectrum completion results of the IDW interpolation and the AE-MSFF network proposed in this embodiment. Under the premise of using the same completion scheme, the MSE of the multi-branch deep learning network proposed in this embodiment after multiple rounds of training convergence is smaller than that of the traditional ConvLSTM scheme. In addition, Figure 7As shown in the figure, under the premise of using the same prediction scheme, the AE-MSFF network and IDW interpolation proposed in this embodiment are used in the completion stage to restore the missing values ​​for subsequent predictions, and the subsequent prediction MSE based on the AE-MSFF completion scheme is also lower. Therefore, the spectrum prediction scheme proposed in this embodiment is superior to the traditional ConvLSTM method in the evaluated data set. And the use of a more advanced and accurate spectrum completion scheme can significantly improve the prediction accuracy.

[0131] This embodiment achieves the following technical effects through the above technical solution:

[0132] 1) This embodiment proposes a framework that integrates spectrum completion and spectrum prediction, which solves the problem of missing historical spectrum data caused by the limited sampling trajectory of drones in three-dimensional scenes. Based on accurate completion results and a powerful spectrum prediction network that extracts spatiotemporal and frequency features respectively, the accuracy of spectrum prediction is effectively improved.

[0133] 2) This embodiment introduces an autoencoder network equipped with a three-dimensional spatial attention mechanism and an MSFF module, which can improve the spectrum completion performance in complex scenarios, thereby enhancing the subsequent spectrum prediction accuracy.

[0134] 3) This embodiment adopts a multi-branch network structure to extract three types of features through different network branches respectively, and then combines the correlation information extracted from the branches for prediction, so as to more effectively utilize the complex three-dimensional spatial dependency, time dependency and frequency dependency in the spectrum data, and reduce the error of spectrum prediction.

[0135] Exemplary Devices

[0136] Based on the above embodiments, the present invention further provides a 3D spectrum completion and prediction device based on an autoencoder network, comprising:

[0137] The sampling module is used to segment the three-dimensional geographic area to be predicted based on the grid quantization method, and use a drone with a preset trajectory to perform spatial sparse sampling to obtain the RSS data of the spectrum;

[0138] The spectrum completion module is used for the autoencoder network based on the MSFF module and the three-dimensional spatial attention mechanism to spatially complete the collected incomplete RSS data to obtain spectrum completion data;

[0139] The spectrum prediction module is used to extract the key features of time, space and frequency of the spectrum completion data based on the spectrum prediction network of deep learning, predict the spectrum information of the future time period, and output the predicted spectrum data.

[0140] This embodiment achieves the following technical effects through the above technical solution:

[0141] 1) This embodiment proposes a framework that integrates spectrum completion and spectrum prediction, which solves the problem of missing historical spectrum data caused by the limited sampling trajectory of drones in three-dimensional scenes. Based on accurate completion results and a powerful spectrum prediction network that extracts spatiotemporal and frequency features respectively, the accuracy of spectrum prediction is effectively improved.

[0142] 2) This embodiment introduces an autoencoder network equipped with a three-dimensional spatial attention mechanism and an MSFF module, which can improve the spectrum completion performance in complex scenarios, thereby enhancing the subsequent spectrum prediction accuracy.

[0143] 3) This embodiment adopts a multi-branch network structure to extract three types of features through different network branches respectively, and then combines the correlation information extracted from the branches for prediction, so as to more effectively utilize the complex three-dimensional spatial dependency, time dependency and frequency dependency in the spectrum data, and reduce the error of spectrum prediction.

[0144] Based on the above embodiment, the present invention further provides a terminal, whose principle block diagram can be as follows: Figure 8 shown.

[0145] The terminal includes: a processor, a memory, an interface, a display screen and a communication module connected through a system bus; wherein the processor of the terminal is used to provide computing and control capabilities; the memory of the terminal includes a storage medium and an internal memory; the storage medium stores an operating system and a computer program; the internal memory provides an environment for the operation of the operating system and the computer program in the storage medium; the interface is used to connect to external devices; the display screen is used to display corresponding information; and the communication module is used to communicate with a cloud server or other devices.

[0146] When the computer program is executed by a processor, it is used to implement the operation of a 3D spectrum completion and prediction method based on an autoencoder network.

[0147] It can be understood by those skilled in the art that Figure 8 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the terminal to which the scheme of the present invention is applied. The specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0148] In one embodiment, a terminal is provided, comprising: a processor and a memory, wherein the memory stores a 3D spectrum completion and prediction program based on an autoencoder network, and the 3D spectrum completion and prediction program based on an autoencoder network is used to implement the operation of the above-mentioned 3D spectrum completion and prediction method based on an autoencoder network when the 3D spectrum completion and prediction program based on an autoencoder network is executed by the processor.

[0149] In one embodiment, a storage medium is provided, wherein the storage medium stores a 3D spectrum completion and prediction program based on an autoencoder network, and when the 3D spectrum completion and prediction program based on an autoencoder network is executed by a processor, it is used to implement the operation of the above-mentioned 3D spectrum completion and prediction method based on an autoencoder network.

[0150] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and volatile memory.

[0151] In summary, the present invention provides a 3D spectrum completion and prediction method based on an autoencoder network, comprising: performing regional segmentation on a three-dimensional geographical area to be predicted based on a grid quantization method, and performing spatial sparse sampling using a drone with a preset trajectory to obtain RSS data of the spectrum; based on an autoencoder network with an MSFF module and a three-dimensional spatial attention mechanism, spatially completing the collected incomplete RSS data to obtain spectrum completion data; using a spectrum prediction network based on deep learning to extract key features of time, space and frequency from the spectrum completion data, predicting spectrum information for future time periods, and outputting predicted spectrum data; the present invention restores missing spectrum data by extracting three-dimensional spatial features and fusing multi-scale features, and based on the results of spectrum completion, uses a multi-branch spectrum prediction network of deep learning to extract spatial, temporal and frequency correlation features, thereby effectively improving the accuracy of spectrum prediction.

[0152] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A 3D spectrum completion and prediction method based on an autoencoder network, characterized in that: include: The three-dimensional geographic area to be predicted is segmented based on the grid quantization method, and a drone with a preset trajectory is used for spatial sparse sampling to obtain the received signal strength data of the spectrum; Based on the autoencoder network of the multi-scale feature fusion module and the three-dimensional spatial attention mechanism, the collected incomplete received signal strength data is spatially completed to obtain spectrum completion data; The spectrum prediction network based on deep learning extracts the key features of time, space and frequency from the spectrum completion data, predicts the spectrum information of future time periods, and outputs the predicted spectrum data.

2. The 3D spectrum completion and prediction method based on the autoencoder network according to claim 1, characterized in that: The grid quantization method is used to segment the three-dimensional geographic area to be predicted, and a drone with a preset trajectory is used to perform spatial sparse sampling to obtain received signal strength data of the spectrum, including: Based on the grid quantization method, the three-dimensional geographic area to be predicted is segmented to obtain N L ×N W ×N H grid subspaces; where N L , N W , N H are the number of quantization intervals in the three dimensions of space respectively; each grid constitutes a regional subspace, and the interval of each grid is Δd. Assuming that the received signal strength in each grid is consistent, the grid index I = (i, j, k) represents the actual geographical location x (i,j,k) =(i+0.5)Δd,(j+0.5)Δd,(k+0.5)Δd); The drone with the preset trajectory is used to perform spatial sparse sampling to obtain received signal strength data of the spectrum; wherein the received signal strength data is the missing spectrum data.

3. The 3D spectrum completion and prediction method based on the autoencoder network according to claim 1, characterized in that: The autoencoder network based on the multi-scale feature fusion module and the three-dimensional spatial attention mechanism performs spatial completion on the collected incomplete received signal strength data to obtain spectrum completion data, including: Extracting the spectral spatial distribution features based on 3D convolution and compressing the received signal strength data to obtain semantic features with an expanded receptive field, and allocating corresponding weights to different parts of the three-dimensional feature map extracted from the received signal strength data based on a three-dimensional spatial attention mechanism to automatically evaluate the importance of spatial features at various locations; The hidden feature representation is expanded to a dimension consistent with the received signal strength data based on a 3D transposed convolutional layer, and feature integration is performed based on a multi-scale feature fusion module to obtain the spectrum completion data.

4. The 3D spectrum completion and prediction method based on the autoencoder network according to claim 3, characterized in that: The step of integrating features based on a multi-scale feature fusion module to obtain the spectrum completion data includes: In the feature integration stage, based on the multi-scale feature fusion module, dilated convolution and ordinary convolution are jointly used to expand the receptive field, and multi-scale feature fusion is performed through a parallel multi-branch structure to extract three-dimensional spatial features and obtain the spectrum completion data.

5. The 3D spectrum completion and prediction method based on the autoencoder network according to claim 1, characterized in that: The prediction process of the spectrum prediction network is defined as: Among them, g θ represents the spectrum prediction network; Δt represents the prediction range; Indicates received signal strength data including multiple frequency bands and spatial positions after spectrum completion; T represents the number of time slots of historical spectrum data.

6. The 3D spectrum completion and prediction method based on the autoencoder network according to claim 1, characterized in that: The spectrum prediction network based on deep learning extracts key features of time, space and frequency from the spectrum completion data and predicts spectrum information in future time periods, including: A three-dimensional convolution structure is used to extract the three-dimensional spatial features of the spectrum completion data, and the regional spatial correlation features of different spaces are obtained by combining a three-dimensional spatial attention module and a jump connection evaluation; Utilizing 1D convolution combined with 1D attention to extract frequency correlation features of the spectrum completion data; Extracting time-correlation features of the spectrum completion data using a gated recurrent unit network; The various spectral semantic feature information extracted by different branches of the spectrum prediction network are spliced ​​and fused to finally realize prediction, and the various features extracted by the multi-branch feature extraction structure are combined to realize the prediction of spectrum information in the future time period.

7. A 3D spectrum completion and prediction device based on an autoencoder network, characterized in that: include: The sampling module is used to segment the three-dimensional geographic area to be predicted based on the grid quantization method, and use a drone with a preset trajectory to perform spatial sparse sampling to obtain the received signal strength data of the spectrum; The spectrum completion module is used to spatially complete the collected incomplete received signal strength data based on the autoencoder network of the multi-scale feature fusion module and the three-dimensional spatial attention mechanism to obtain spectrum completion data; The spectrum prediction module is used to extract the key features of time, space and frequency of the spectrum completion data based on the spectrum prediction network of deep learning, predict the spectrum information of the future time period, and output the predicted spectrum data.

8. A terminal, characterized in that: include: A processor and a memory, wherein the memory stores a 3D spectrum completion and prediction program based on an autoencoder network, and when the 3D spectrum completion and prediction program based on an autoencoder network is executed by the processor, it is used to implement the operation of the 3D spectrum completion and prediction method based on an autoencoder network as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a 3D spectrum completion and prediction program based on an autoencoder network, and when the 3D spectrum completion and prediction program based on an autoencoder network is executed by a processor, it is used to implement the operation of the 3D spectrum completion and prediction method based on an autoencoder network as described in any one of claims 1-6.

Images