CN114826459A - Spectrum map accurate construction method based on cross-domain reasoning - Google Patents

Spectrum map accurate construction method based on cross-domain reasoning Download PDF

Info

Publication number
CN114826459A
CN114826459A CN202210226232.3A CN202210226232A CN114826459A CN 114826459 A CN114826459 A CN 114826459A CN 202210226232 A CN202210226232 A CN 202210226232A CN 114826459 A CN114826459 A CN 114826459A
Authority
CN
China
Prior art keywords
frequency
data
spectrum map
frequency spectrum
network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210226232.3A
Other languages
Chinese (zh)
Other versions
CN114826459B (en
Inventor
周福辉
王晨玥
徐铭
吴雨航
袁璐
吴启晖
董超
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to CN202210226232.3A priority Critical patent/CN114826459B/en
Publication of CN114826459A publication Critical patent/CN114826459A/en
Application granted granted Critical
Publication of CN114826459B publication Critical patent/CN114826459B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/391—Modelling the propagation channel
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G06T11/20—Drawing from basic elements, e.g. lines or circles
    • G06T11/206—Drawing of charts or graphs
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/309—Measuring or estimating channel quality parameters
    • H04B17/318—Received signal strength
    • H04B17/327—Received signal code power [RSCP]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/382—Monitoring; Testing of propagation channels for resource allocation, admission control or handover
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/391—Modelling the propagation channel
    • H04B17/3911—Fading models or fading generators

Abstract

The invention discloses a method for accurately constructing a frequency spectrum map based on cross-domain reasoning, which comprises the steps of collecting frequency spectrum data and preprocessing the frequency spectrum data; constructing a corresponding defect spectrum map according to the preprocessed spectrum data for each frequency band; stacking the defect frequency spectrum map of each frequency band and the blank frequency spectrum map of the target frequency band in a frequency dimension to form a cross-domain frequency spectrum map sample, and dividing the cross-domain frequency spectrum map sample into a training set sample and a test set sample; improving a self-encoder network structure according to the three-dimensional data characteristics of the cross-domain frequency spectrum map sample and training the network structure; and inputting the test set sample into the trained network structure, and performing post-processing on network output data to obtain a frequency spectrum map construction result. The method can solve the problem that the frequency spectrum map is difficult to accurately obtain due to the fact that the target frequency band has no frequency spectrum data.

Description

Spectrum map accurate construction method based on cross-domain reasoning
Technical Field
The invention belongs to the technical field of wireless communication, and particularly relates to a spectrum map accurate construction method based on cross-domain reasoning.
Background
The accurate construction of the spectrum map is an important technology in cognitive radio, the spectrum data of a target area is complemented through correlation among the spectrum data, and the complete distribution situation of radio parameters (such as received signal power, power spectral density and the like) on the geographical position is visualized, so that the use situation of spectrum resources in the target area can be represented. The frequency spectrum map can provide basis for ordered management and rapid decision-making of frequency spectrum resources, so that the communication system is assisted to realize reasonable utilization of the frequency spectrum resources.
Most of the existing frequency spectrum map construction work is to perform spatial estimation on missing data in the current frequency point frequency spectrum map by using a 'spatial interpolation' method according to known frequency spectrum data at different positions. That is, the construction of the frequency band spectrum map can be realized only when the target frequency band has certain spectrum data.
In fact, with the continuous development of various online services, such as a radio service from a text, a voice call to a video call, and then to an online game, the demand of a wireless communication system for bandwidth is increasing, a communication frequency band is expanded from a low frequency band to a high frequency band, and 5G and 6G mobile communication systems exhibit the characteristic of ultra wide band. If a site survey is conducted for all frequency points in a wide frequency band and sufficient spectrum data is obtained, a great deal of deployment and sensing cost is consumed. In practical situations, only a part of frequency bands of spectrum data are collected. Therefore, it is of great necessity and practical significance to research how to accurately construct a complete spectrum map of a target frequency band from sparse data acquired from other frequency bands.
Koya Sato, Katsuya Suto, Kei Inage et al in its published paper "Space-Frequency-interleaved Radio Map" (IEEE trans. veh. technol., vol.70, No.1, pp.714-725, jan.2021) mention the concept of Frequency-Space domain correlated spectrum Map model, but assume that the signal is a narrowband signal, and according to the principle that "propagation phenomena of similar frequencies are similar", the propagation laws of signals at different Frequency points in a narrowband are considered to be approximately the same. The work uses frequency separation, a three-dimensional spectrum map structure is directly split into a plurality of two-dimensional spectrum maps along the frequency dimension, and the two-dimensional spectrum maps are used as independent samples to be completed, so that a cross-domain structure is not actually utilized, and only the spatial correlation of data is essentially utilized. Yves Teganya, Daniel Romero et al published in the paper "Data-drive Spectrum graphics via Deep assembly Autoencoders" (IEEE icc., jul.2020), proposed a frequency-space dual interpolation method based on the conclusion that the received signal power shows strong correlation in a wide frequency domain. But the method simply fixes the frequency correlation coefficient to 1, namely directly takes the measured values of other frequency bands at the same position as the measured value of the target frequency band, and then completes the defect frequency spectrum map of the frequency band by utilizing the spatial correlation. It is seen that the frequency correlation is not effectively utilized, and the frequency domain completion precision is not high. A patent with publication number CN 112967357 a, which is proposed by national defense science and technology university of the people's liberation military in china, "a spectrum map construction method based on a convolutional neural network" discloses a spectrum map construction method based on a convolutional neural network. The method comprises the steps of firstly, carrying out primary construction on frequency spectrum maps with different spatial resolutions by a traditional distributed cluster kriging interpolation method, and then, carrying out extraction on image features by a dictionary learning model and a convolutional neural network learning model. The contradiction between the calculation complexity and the accuracy of the spectrum map is relieved to a certain extent, but the method only focuses on spatial resolution and neglects the cross-domain study of the spectrum map.
The methods ignore the frequency fading difference of signals with different frequencies in propagation, the starting point of the algorithm is still the spatial completion of data, and the frequency correlation is not utilized, so that the methods are difficult to accurately construct a frequency spectrum map under the condition that no frequency spectrum data exists in a target frequency band. The signals of 5G and 6G communication systems are mostly broadband signals. How to accurately acquire the wide-band frequency spectrum map has important significance. Therefore, it is necessary to develop a new spectrum mapping method.
Disclosure of Invention
The invention aims to solve the technical problems that aiming at the defects of the prior art, the invention provides the method for accurately constructing the frequency spectrum map based on the cross-domain reasoning, aiming at the problem that the frequency spectrum map is difficult to accurately obtain due to no frequency spectrum data of the target frequency band, the frequency-space domain double-domain model is constructed, and the correlation between frequency points and the correlation between spaces are utilized, so that the accurate construction of the frequency spectrum map of the target frequency band is realized, and the technical support is provided for the efficient management of the frequency spectrum of the 6G mobile communication system.
In order to achieve the technical purpose, the technical scheme adopted by the invention is as follows:
the method for accurately constructing the frequency spectrum map based on cross-domain reasoning comprises the following steps:
step 1, collecting frequency spectrum data and preprocessing the collected frequency spectrum data;
step 2, each frequency band respectively constructs a corresponding defect frequency spectrum map according to the preprocessed frequency spectrum data;
step 3, stacking the defect frequency spectrum map of each frequency band and the blank frequency spectrum map of the target frequency band in a frequency dimension to form a cross-domain frequency spectrum map sample and dividing the cross-domain frequency spectrum map sample into a training set sample and a test set sample;
step 4, improving a network structure of a self-encoder according to three-dimensional data characteristics of the cross-domain frequency spectrum map sample and training the network structure according to a training set sample;
and 5, inputting the test set sample into the trained network structure, and performing post-processing on network output data to obtain a frequency spectrum map construction result.
In order to optimize the technical scheme, the specific measures adopted further comprise:
the step 1 of acquiring the spectrum data specifically includes:
randomly generating a fixed number of sensing node positions in a target area, and collecting received signal data at nodes;
and uploading the acquired signal data together with the receiving position and the transmitting frequency to a central database, and calculating by the central database to generate a complete frequency spectrum map of the target frequency point to finish the acquisition of the frequency spectrum data.
In the step 1, the preprocessing of the spectrum data includes:
1) traversing the frequency spectrum data to obtain the maximum value p min 、p max ；
2) Data normalization:
Figure BDA0003535891050000031
wherein, p (f) k ) Representing frequency point f k Corresponding spectral data, p min Representing the minimum value, p, of all spectral data max Representing the maximum of all spectral data.
In the step 2, when each frequency band constructs a corresponding defect spectrum map according to the preprocessed spectrum data, the target region is rasterized, and if the grid point I is (I, j) corresponding to the grid where the spectrum data exists, the grid point I is (I, j) corresponding to the spectrum data, the target region is rasterized
Figure BDA0003535891050000032
Taking values as spectrum data values; if not, then,
Figure BDA0003535891050000033
indicating that spectral information of the grid is missing;
wherein the content of the first and second substances,
Figure BDA0003535891050000034
representing frequency point f k A corresponding defect frequency spectrum map is obtained,
Figure BDA0003535891050000035
representing the elements on grid point (i, j) in the spectrum map.
The missing spectrum map in step 2 is constructed as follows:
the grid where I ═ I, j is located has spectral data:
Figure BDA0003535891050000036
and (I, j) is located on the grid without spectrum data:
Figure BDA0003535891050000037
wherein the content of the first and second substances,
Figure BDA0003535891050000038
representing frequency point f k A corresponding defect frequency spectrum map is obtained,
Figure BDA0003535891050000039
a numerical value p (f) representing a grid point (i, j) in the defect spectrum map k ,x i,j ) Representing actual spectral data, averaged when there are multiple spectral data within a grid.
The spectrum map is represented as an N × N discrete grid model, which is obtained by uniformly dividing an actual scene, and assuming that the signal intensity in each grid is constant, the signal intensity is constant at a grid point.
In the step 3, the defect spectrum maps of the frequency bands of the same radio propagation scene and different transmission frequencies and the blank spectrum maps of the target frequency band are stacked according to the frequency size in the third dimension to form a cross-domain spectrum map sample with a three-dimensional tensor structure, so that the defect spectrum maps of different frequencies are associated in a frequency domain, and the cross-domain spectrum map sample is further divided into a training set sample and a test set sample.
The step 4 includes:
1) compressing the structural depth of the U-Net network and releasing the load of network parameters;
2) modifying the parameters and the structures of the convolutional layer and the transposed convolutional layer according to the three-dimensional data characteristics;
3) copying and splicing the characteristic graphs with the same resolution ratio between the layers corresponding to the encoder network and the decoder network;
4) the network is trained using training set samples.
The training network using the training set samples specifically includes:
first, the trainable parameters of the network are initialized randomly:
initializing a network training iteration time epoch to be 1, learning rate to be 0.0005, and taking an Adam optimization algorithm as a network training optimizer;
secondly, inputting the training set samples into a network for training in batches:
the size of each batch is adjusted according to the training effect, and the training error of each batch is propagated in a reverse gradient manner so as to optimize the model parameters;
when all batches of samples in the training set complete one forward propagation and one reverse propagation, the 1 epoch is completed;
the training error function adopts a mean square error loss function, and the calculation formula is as follows:
Figure BDA0003535891050000041
where L is the number of samples in a training batch, (M + 1). times.NxN represents the size of each training sample, P l The true result of the ith sample is shown,
Figure BDA0003535891050000042
representing the estimation result of the l sample;
thirdly, judging whether the network training is finished:
and (5) judging whether the network is trained to be convergent or not, if so, executing the step 5, otherwise, adding one to the epoch, and continuing to train the network in the second step.
The self-encoder network structure improved in the step 4 is as follows:
Figure BDA0003535891050000043
wherein epsilon θ Representing an improved self-encoder network, delta θ Representing a decoder network.
The post-processing of the network output data in the step 5 above to obtain a spectrum map construction result specifically includes:
p＝p(p max -p min )+p min
wherein p represents the frequency spectrum data constructed by the model, namely the frequency spectrum map construction result, p min Representing the minimum value, p, of all spectral data max Representing the maximum of all spectral data.
The invention has the following beneficial effects:
the method can realize the accurate construction of the frequency spectrum map of the target frequency band through the frequency-space domain dual-domain frequency spectrum map construction and the improved self-encoder network according to the frequency spectrum data of other frequency bands under the condition that the target frequency band has no frequency spectrum data, and solves the problem that the frequency spectrum map is difficult to accurately obtain due to the fact that the target frequency band has no frequency spectrum data.
1. The invention introduces a frequency correlation concept of radio propagation and a frequency-space domain dual-domain spectrum map (namely a cross-domain spectrum map sample) modeling method, and solves the problem that the existing spectrum map construction method is difficult to accurately construct a spectrum map when no spectrum data exists in a target frequency band.
2. The invention introduces a cross-domain reasoning method, completes the synchronous learning of the frequency fading characteristics and the space fading characteristics based on a neural network model, improves the construction precision of a frequency spectrum map by utilizing the frequency correlation and the space correlation, overcomes the problem that the frequency correlation is not effectively utilized because the frequency fading difference of different frequency signals during propagation is ignored in the prior art, and can be applied to an actual wireless communication system.
3. Compared with a classical deep neural network, the network framework designed by the invention compresses the network depth, simplifies the network structure, has smaller network depth and more simplified network structure, releases the parameter load of the network, overcomes the problems of more network parameters, high training complexity and easy overfitting caused by three-dimensional input data, and improves the convergence rate of the network.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a network framework diagram of the present invention;
FIG. 3 is a comparison graph of the visualization effect of the frequency spectrum map constructed by the invention and other methods;
fig. 4 is a graph comparing the performance of the present invention and other methods at different sampling rates.
Fig. 5 is a convergence diagram of the present invention under different emission source scenarios.
Detailed Description
Embodiments of the present invention are described in further detail below with reference to the accompanying drawings.
Referring to fig. 1, the method for accurately constructing a spectrum map based on cross-domain reasoning of the present invention includes the following specific steps:
step 1, collecting frequency spectrum data and preprocessing the collected frequency spectrum data;
step 1.1, collecting frequency spectrum data:
and randomly generating a fixed number of sensing node positions in the target area, and acquiring received signal data at the nodes. And uploading the signal acquisition data together with information such as the receiving position, the transmitting frequency and the like to a central database, and calculating by the central database to generate a complete frequency spectrum map of the target frequency point.
In an embodiment of the invention it is assumed that the signal source can be switched between several transmission frequencies.
The invention adopts a signal large-scale fading model P (f) k ,x i,j )＝P Tx -n f log 10 (f)-10n d log 10 (d)-L C And characterizing radio propagation rules.
Wherein, P Tx Representing the transmit power of the signal source; f. of k Representing a signal sourceThe transmit frequency of (d); n is f And n d Respectively representing a frequency loss coefficient and a path loss coefficient; d represents the distance between the transmitting end and the receiving end; l is C Is a constant for free space propagation loss.
Step 1.2, preprocessing the acquired frequency spectrum data:
1) according to the signal data obtained in the step 1.1, firstly, traversing the frequency spectrum data to obtain the minimum value p in all the data min And maximum value p max 。
2) In order to enable the neural network to better learn the distribution characteristics of data and improve the convergence effect of the network, data preprocessing is performed on input data, namely normalization processing is performed on frequency spectrum data on each grid point, so that the common problem of gradient disappearance or gradient explosion in the neural network is prevented:
Figure BDA0003535891050000061
wherein, p (f) k ) Representing the frequency f k Corresponding spectral data, p min Representing the minimum value, p, of all spectral data max Representing the maximum of all spectral data.
Step 2, each frequency band respectively constructs a corresponding defect frequency spectrum map according to the preprocessed frequency spectrum data;
the invention represents the frequency spectrum map as an NxN discretization grid model which is obtained by uniformly dividing an actual scene. The values at the grid points represent the signal strength of the grid in which they are located, assuming that the signal strength within each grid is constant. And respectively generating a corresponding two-dimensional frequency spectrum map for each frequency point.
Rasterizing the target area, and if the grid point I is (I, j) corresponding to the grid, acquiring data exists in the grid
Figure BDA0003535891050000062
Taking the value as the measured value; if not, then,
Figure BDA0003535891050000063
to represent the gridSpectral information is missing. Wherein the content of the first and second substances,
Figure BDA0003535891050000064
representing frequency point f k A corresponding defect frequency spectrum map is obtained,
Figure BDA0003535891050000065
representing the elements on grid point (i, j) in the spectrum map.
The missing spectrum map is constructed in the following specific manner:
the grid where I ═ I, j is located has spectral data:
Figure BDA0003535891050000066
and (I, j) is located on the grid without spectrum data:
Figure BDA0003535891050000067
wherein the content of the first and second substances,
Figure BDA0003535891050000068
representing frequency point f k A corresponding defect frequency spectrum map is obtained,
Figure BDA0003535891050000069
a numerical value p (f) representing a grid point (i, j) in the defect spectrum map k ,x i,j ) The actual spectral data representing the location is averaged when there are multiple spectral data within a grid.
Step 3, stacking the defect frequency spectrum map of each frequency band and the blank frequency spectrum map of the target frequency band in a frequency dimension to form a cross-domain frequency spectrum map sample and dividing the cross-domain frequency spectrum map sample into a training set sample and a test set sample;
the invention utilizes the improved self-encoder network structure to perform cross-domain reasoning of frequency and space domains according to sparse spectrum data, and realizes the accurate construction of a spectrum map through learned signal fading characteristics. The key of this step is to use frequency correlation, so it is first necessary to associate the defect spectrum maps of different frequencies in the frequency domain.
A plurality of frame frequency spectrum maps of the same radio transmission scene but different signal transmission frequencies are stacked according to the frequency in the third dimension to form a three-dimensional tensor structure, so that a plurality of independent two-dimensional spectrum maps are associated in a frequency domain. The frequency spectrum map of each measured frequency band is a defect frequency spectrum map containing a plurality of measured data, and the frequency spectrum map of the target frequency band is a blank frequency spectrum map without any frequency spectrum data.
Step 4, improving a network structure of a self-encoder according to three-dimensional data characteristics of the cross-domain frequency spectrum map sample and training the network structure according to a training set sample;
the invention designs a network structure based on a U-Net pixel-level semantic segmentation structure of an auto-encoder. The three-dimensional spectrum tensor is regarded as a three-dimensional image, and a coding process is adopted to carry out down-sampling to extract the frequency fading characteristic and the space fading characteristic of radio propagation; and then, performing upsampling by adopting a decoding process to reconstruct a complete image representing the distribution of the frequency spectrum data. Therefore, the prediction process of the spectrum map by the network is as follows:
Figure BDA0003535891050000071
wherein epsilon θ Representing a network of encoders, delta θ Representing a decoder network.
Aiming at the problems that the classical U-Net deep neural network structure is concentrated on two-dimensional data processing, and excessive parameters, high training complexity and easiness in overfitting are caused after the classical U-Net deep neural network structure is applied to three-dimensional data, the method is improved based on the U-Net structure, and a network structure suitable for the three-dimensional data structure is built. The method comprises the following steps:
1) compressing the depth of the U-Net network and releasing the parameter load of the network;
2) modifying the parameters and the structures of the convolutional layer and the transposed convolutional layer according to the three-dimensional characteristics, and simultaneously realizing the increase of a data channel and the size compression of a characteristic diagram, thereby reducing the network depth and lightening the parameter load;
3) and copying and splicing the characteristic graph with the same resolution ratio between the corresponding layers of the encoder network and the decoder network to supplement the detail characteristics.
4) Training the network by utilizing the training set sample, judging whether the training of the network is finished, if so, executing the next step, and if not, continuing the training after adding one to the training iteration times, wherein the method specifically comprises the following steps:
first, trainable parameters of the network are randomly initialized. Initializing the network training iteration frequency epoch to be 1, learning rate to be 0.0005, and using an Adam optimization algorithm as a network training optimizer.
And secondly, inputting the training set samples into the network for training in batches, adjusting the batch size according to the training effect, and performing inverse gradient propagation on the training error of each batch so as to optimize the model parameters. When all batches of samples in the training set complete one forward propagation and one backward propagation, 1 epoch is completed. The training error function adopts a commonly used mean square error loss function, and the calculation formula is as follows:
Figure BDA0003535891050000081
where L is the number of samples in a training batch, (M + 1). times.NxN represents the size of each training sample, P l The true result of the ith sample is shown,
Figure BDA0003535891050000082
the model estimation result of the l-th sample is represented.
Thirdly, judging whether the network training is finished:
and (5) judging whether the network is trained to be convergent or not, if so, executing the step 5, otherwise, adding one to the epoch, and continuing to train the network in the second step.
The network structure of the present invention is shown in fig. 2.
And 5, inputting the test set sample into the trained network structure, and performing post-processing on network output data to obtain a frequency spectrum map construction result.
And (3) carrying out data reprocessing on the output result of the network, and restoring elements in the frequency spectrum map into received signal strength with dBm as a unit:
p＝p(p max -p min )+p min
wherein p represents the spectral data of the model construction, p min Representing the minimum value, p, of all spectral data max Representing the maximum of all spectral data.
The effect of the present invention will be further explained with the simulation experiment.
1. Simulation conditions and parameter setting:
the simulation experiment of the invention is carried out on a simulation platform of Python3.6, Pytroch 1.9.1. The spectrum data used for predicting the spectrum map is acquired by sensing nodes randomly distributed in a target area, and two kinds of mining rates of 15% and 29% and two kinds of scenes of a single emission source and multiple emission sources are considered. Initializing the network training iteration frequency epoch to be 1, learning rate to be 0.0005, and using an Adam optimization algorithm as a network training optimizer. The number of samples for a single batch was 4.
2. Simulation content:
fig. 3 is a comparison graph of the visualization effect of the frequency spectrum map constructed by the invention and other methods.
Fig. 3(a) in fig. 3 is a real result of a spectrum map, fig. 3(b) is a result of a spectrum map constructed by the method of the present invention, and fig. 3(c) is a result of a spectrum map based on a conventional spatial interpolation method. The comparison shows that the frequency spectrum map constructed by the method has an obviously better effect than that of another method, and can accurately represent the overall distribution condition of frequency spectrum data in a target area and the position of a signal source.
Fig. 4 is a graph comparing the performance of the present invention and other methods at different sampling rates.
The abscissa in fig. 4 represents the different sampling rates (%), and the ordinate represents the completion error (dB). The dark columns at each sampling rate represent errors in the construction using the method of the present invention, and the white columns represent errors in the generation of a spectral map using conventional spatial interpolation. The sampling rates were 15% and 29%, respectively. By comparison, the construction performance of the invention is obviously better than that of another method. From the general trend, model performance improves as the sampling rate increases. When the sampling rate is 15%, the construction error of the method is about 2.53dB, which is about 9% lower than that of a method based on spatial interpolation; at a sampling rate of 29%, the construction error of the present invention is around 1.69dB, which is about 39% lower than the method based on spatial interpolation.
Fig. 5 is a convergence diagram of the present invention under different emission source scenarios.
The abscissa in fig. 5 represents the number of training rounds (times), and the ordinate represents the training loss. The broken line marked by a circle represents the convergence curve of the method under the scene of a single emission source, and the broken line marked by a rectangle represents the convergence curve of the method under the scene of multiple emission sources. It can be seen that the network convergence speed of the method of the invention is high, and the convergence can be achieved within 10 training periods under two scenes.
By integrating the simulation results and analysis, the method for accurately constructing the frequency spectrum map based on cross-domain reasoning can better utilize the frequency correlation and the space correlation of radio transmission, and realize the accurate construction of the frequency spectrum map when the target frequency band has no frequency spectrum data. The invention can be widely applied to actual communication scenes.
The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and embellishments within the scope of the invention may be made by those skilled in the art without departing from the principle of the invention.

Claims (10)

1. The method for accurately constructing the frequency spectrum map based on cross-domain reasoning is characterized by comprising the following steps of:
step 1, collecting frequency spectrum data and preprocessing the collected frequency spectrum data;
step 2, each frequency band respectively constructs a corresponding defect frequency spectrum map according to the preprocessed frequency spectrum data;
step 3, stacking the defect frequency spectrum map of each frequency band and the blank frequency spectrum map of the target frequency band in a frequency dimension to form a cross-domain frequency spectrum map sample and dividing the cross-domain frequency spectrum map sample into a training set sample and a test set sample;
step 4, improving a network structure of a self-encoder according to three-dimensional data characteristics of the cross-domain frequency spectrum map sample and training the network structure according to a training set sample;
and 5, inputting the test set sample into the trained network structure, and performing post-processing on network output data to obtain a frequency spectrum map construction result.
2. The method for accurately constructing a spectrum map based on cross-domain reasoning according to claim 1, wherein the step 1 collects spectrum data, and specifically comprises the following steps:
randomly generating a fixed number of sensing node positions in a target area, and collecting received signal data at nodes;
and uploading the acquired signal data together with the receiving position and the transmitting frequency to a central database, and calculating by the central database to generate a complete frequency spectrum map of the target frequency point to finish the acquisition of the frequency spectrum data.
3. The method for accurately constructing a spectrum map based on cross-domain reasoning according to claim 1, wherein in the step 1, preprocessing the spectrum data includes:
1) traversing the frequency spectrum data to obtain the maximum value p min 、p max ；
2) Data normalization:
Figure FDA0003535891040000011
wherein, p (f) k ) Representing frequency point f k Corresponding spectral data, p min Representing the minimum value, p, of all spectral data max Representing the maximum of all spectral data.
4. The method for accurately constructing a spectrum map based on cross-domain reasoning according to claim 1, wherein in the step 2, each frequency band is respectively presetWhen the processed spectrum data construct a corresponding defect spectrum map, the target area is rasterized, and if the grid point I is (I, j) corresponding to the grid, the spectrum data exists in the grid
Figure FDA0003535891040000012
Taking values as spectrum data values; if not, then,
Figure FDA0003535891040000013
indicating that spectral information of the grid is missing;
wherein the content of the first and second substances,
Figure FDA0003535891040000014
representing frequency point f k A corresponding defect frequency spectrum map is obtained,
Figure FDA0003535891040000015
representing the elements on grid point (i, j) in the spectrum map.
5. The method for accurately constructing a spectrum map based on cross-domain reasoning according to claim 1, wherein the missing spectrum map in the step 2 is constructed in the following manner:
the grid where I ═ I, j is located has spectral data:
Figure FDA0003535891040000021
and I is (I, j) located grid spectrum-free data:
Figure FDA0003535891040000022
wherein the content of the first and second substances,
Figure FDA0003535891040000023
representing frequency point f k A corresponding defect frequency spectrum map is obtained,
Figure FDA0003535891040000024
map middle grid for representing defect frequency spectrumThe value of point (i, j), p (f) k ,x i,j ) Representing actual spectral data, averaged when there are multiple spectral data within a grid.
The spectrum map is represented as an N × N discrete grid model, which is obtained by uniformly dividing an actual scene, and assuming that the signal intensity in each grid is constant, the signal intensity is constant at a grid point.
6. The method for accurately constructing a spectrum map based on cross-domain reasoning according to claim 1, wherein in step 3, the defective spectrum maps of different frequency bands of different transmission frequencies in the same radio propagation scene and the blank spectrum maps of the target frequency band are stacked according to the frequency in the third dimension to form a cross-domain spectrum map sample with a three-dimensional tensor structure, so that the defective spectrum maps of different frequencies are associated in a frequency domain, and the cross-domain spectrum map sample is further divided into a training set sample and a test set sample.
7. The method for accurately constructing a spectrum map based on cross-domain reasoning according to claim 1, wherein the step 4 comprises:
1) compressing the structural depth of the U-Net network and releasing the load of network parameters;
2) modifying the parameters and the structures of the convolutional layer and the transposed convolutional layer according to the three-dimensional data characteristics;
3) copying and splicing the characteristic graphs with the same resolution ratio between the layers corresponding to the encoder network and the decoder network;
4) the network is trained using training set samples.
8. The method for accurately constructing a spectrum map based on cross-domain reasoning according to claim 7, wherein the training of the network by using the training set samples specifically comprises:
first, the trainable parameters of the network are initialized randomly:
initializing a network training iteration time epoch to be 1, learning rate to be 0.0005, and taking an Adam optimization algorithm as a network training optimizer;
secondly, inputting the training set samples into a network for training in batches:
the size of each batch is adjusted according to the training effect, and the training error of each batch is propagated in a reverse gradient manner so as to optimize the model parameters;
when all batches of samples in the training set complete one forward propagation and one reverse propagation, the 1 epoch is completed;
the training error function adopts a mean square error loss function, and the calculation formula is as follows:
Figure FDA0003535891040000031
where L is the number of samples in a training batch, (M + 1). times.NxN represents the size of each training sample, P l The true result of the ith sample is shown,
Figure FDA0003535891040000032
representing the estimation result of the l sample;
thirdly, judging whether the network training is finished:
and (5) judging whether the network is trained to be converged, if so, executing the step (5), otherwise, adding one to the epoch, and continuing to train the network in the second step.
9. The method for accurately constructing a spectrum map based on cross-domain reasoning according to claim 1, wherein the self-encoder network structure improved in the step 4 is as follows:
Figure FDA0003535891040000033
wherein epsilon θ Representing an improved self-encoder network, delta θ Representing a decoder network.
10. The method for accurately constructing a frequency spectrum map based on cross-domain reasoning according to claim 1, wherein the step 5 is to perform post-processing on the network output data to obtain a frequency spectrum map construction result, and specifically comprises the following steps:
p＝p(p max -p min )+p min
wherein p represents the frequency spectrum data constructed by the model, namely the frequency spectrum map construction result, p min Representing the minimum value, p, of all spectral data max Representing the maximum of all spectral data.
CN202210226232.3A 2022-03-08 2022-03-08 Spectrum map accurate construction method based on cross-domain reasoning Active CN114826459B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210226232.3A CN114826459B (en) 2022-03-08 2022-03-08 Spectrum map accurate construction method based on cross-domain reasoning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210226232.3A CN114826459B (en) 2022-03-08 2022-03-08 Spectrum map accurate construction method based on cross-domain reasoning

Publications (2)

Publication Number Publication Date
CN114826459A true CN114826459A (en) 2022-07-29
CN114826459B CN114826459B (en) 2023-02-03

Family

ID=82529732

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210226232.3A Active CN114826459B (en) 2022-03-08 2022-03-08 Spectrum map accurate construction method based on cross-domain reasoning

Country Status (1)

Country Link
CN (1) CN114826459B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116663437A (en) * 2023-08-02 2023-08-29 中国人民解放军战略支援部队航天工程大学 Spectrum map construction method and system based on deep neural network

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111381499A (en) * 2020-03-10 2020-07-07 东南大学 Internet-connected aircraft self-adaptive control method based on three-dimensional space radio frequency map learning
CN112084274A (en) * 2020-08-24 2020-12-15 浙江云合数据科技有限责任公司 Data-based space-time spectrum map construction and prediction method
CN112688746A (en) * 2020-12-14 2021-04-20 中山大学 Spectrum prediction method based on space-time data
WO2022021602A1 (en) * 2020-07-27 2022-02-03 南京邮电大学 5g reference signal received power prediction method based on geographic information
CN114116209A (en) * 2021-11-12 2022-03-01 中国人民解放军国防科技大学 Spectrum map construction and distribution method and system based on deep reinforcement learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111381499A (en) * 2020-03-10 2020-07-07 东南大学 Internet-connected aircraft self-adaptive control method based on three-dimensional space radio frequency map learning
WO2022021602A1 (en) * 2020-07-27 2022-02-03 南京邮电大学 5g reference signal received power prediction method based on geographic information
CN112084274A (en) * 2020-08-24 2020-12-15 浙江云合数据科技有限责任公司 Data-based space-time spectrum map construction and prediction method
CN112688746A (en) * 2020-12-14 2021-04-20 中山大学 Spectrum prediction method based on space-time data
CN114116209A (en) * 2021-11-12 2022-03-01 中国人民解放军国防科技大学 Spectrum map construction and distribution method and system based on deep reinforcement learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
XIAOFU DU等: "Three-Dimensional_Completion_Method_with_Uniformity_Decision_Mechanism_for_Spectrum_Sensing_Data", 《2021 IEEE 21ST INTERNATIONAL CONFERENCE ON COMMUNICATION TECHNOLOGY》 *
胡田钰等: "基于生成对抗网络的三维频谱态势补", 《数据采集与处理》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116663437A (en) * 2023-08-02 2023-08-29 中国人民解放军战略支援部队航天工程大学 Spectrum map construction method and system based on deep neural network
CN116663437B (en) * 2023-08-02 2023-11-21 中国人民解放军战略支援部队航天工程大学 Spectrum map construction method and system based on deep neural network

Also Published As

Publication number Publication date
CN114826459B (en) 2023-02-03

Similar Documents

Publication Publication Date Title
CN109890043B (en) Wireless signal noise reduction method based on generative countermeasure network
CN110166154B (en) Software radio frequency spectrum monitoring and identifying method based on neural network
CN111369442B (en) Remote sensing image super-resolution reconstruction method based on fuzzy kernel classification and attention mechanism
CN109460815B (en) Monocular vision depth estimation method
CN112183742B (en) Neural network hybrid quantization method based on progressive quantization and Hessian information
CN112770325B (en) Cognitive internet of vehicles spectrum sensing method based on deep learning
CN114826459B (en) Spectrum map accurate construction method based on cross-domain reasoning
CN113838107B (en) Automatic heterogeneous image registration method based on dense connection
CN114998757A (en) Target detection method for unmanned aerial vehicle aerial image analysis
CN115100148B (en) Crop pest detection method based on lightweight convolutional neural network
CN112418049A (en) Water body change detection method based on high-resolution remote sensing image
CN112507849A (en) Dynamic-to-static scene conversion method for generating countermeasure network based on conditions
CN109448039B (en) Monocular vision depth estimation method based on deep convolutional neural network
CN114528987A (en) Neural network edge-cloud collaborative computing segmentation deployment method
CN114373104A (en) Three-dimensional point cloud semantic segmentation method and system based on dynamic aggregation
CN115913424A (en) Spectrum map accurate construction method for generating confrontation network based on conditions
CN111586151B (en) Intelligent city data sharing system and method based on block chain
CN110782396B (en) Light-weight image super-resolution reconstruction network and reconstruction method
CN110751699B (en) Color reconstruction method of optical remote sensing image based on convolutional neural network
CN113343796B (en) Knowledge distillation-based radar signal modulation mode identification method
CN116977866A (en) Lightweight landslide detection method
CN115861779A (en) Unbiased scene graph generation method based on effective feature representation
CN114049519A (en) Optical remote sensing image scene classification method
CN112598598A (en) Image reflected light removing method based on two-stage reflected light eliminating network
CN112634161B (en) Reflected light removing method based on two-stage reflected light eliminating network and pixel loss

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant