A wireless spectrum state image classification method based on small sample learning

Abstract

The present application relates to a kind of wireless spectrum state image classification method based on small sample learning, belong to cognitive radio field.The method includes: the training set support set S tr With training set query set Q tr Input to embedding module, obtain the tensor set of each category;In the training set support set S tr Tensor set, the class prototype of each category of the training set support set S tr It is calculated by the K tensor closest to each category;The class prototype is input into measurement module with the training set query set Q tr Tensor set, similarity measurement is carried out, and small sample image classifier is obtained.The method used in the present application is different from the category used in the classification task of traditional machine learning algorithm and test, and such setting is conducive to training a small sample image classifier that can extract sample general features and specific features, solves the dependence problem of traditional machine learning-based spectrum state perception algorithm to data, improves the perception accuracy in blind perception, fast perception;Meanwhile, several feature vectors are used to represent image samples, and local features are used to classify images, eliminate the problem of intra-class difference and background confusion.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication, particularly the field of spectrum sensing technology in cognitive radio, and relates to a wireless spectrum state image classification method based on few-shot learning. Background Technology

[0002] The development of mobile internet and the widespread adoption of wireless smart devices have led to an increasing number of users accessing the internet. Simultaneously, the bandwidth demands of individual users are also growing. Basic voice services can no longer meet the diverse communication and entertainment needs, while the scarcity of spectrum resources remains a reality. Theoretically, spectrum resources are unlimited, but limitations imposed by antenna size, electromagnetic wave characteristics, equipment performance, and transmission power strictly restrict the available spectrum resources to a certain range. These frequency bands have already been largely allocated, with the International Telecommunication Union (ITU) planning available spectrum resources ranging from 9kHz to 275GHz. Research by the U.S. Federal Communications Commission (FCC) has indicated that a large amount of allocated spectrum resources are largely idle in both time and space, with the average spectrum utilization rate at any given time and location not exceeding 5%. A major factor contributing to the spectrum shortage is the static allocation policy. While this allocation method can reduce mutual interference between communication systems in different frequency bands, the static allocation strategy is inflexible. Licensed users may not use or fully utilize their spectrum for a considerable period, while other radio users are unable to utilize that frequency band. This has created a paradoxical situation of increasingly scarce spectrum resources and unsatisfactory utilization of existing spectrum. However, this paradox has also provided an opportunity for the emergence of new technologies, one of which is the dynamic allocation of spectrum—Cognitive Radio (CR).

[0003] Cognitive radio spectrum sensing technology has been extensively and deeply studied. To ensure minimal interference to licensed users, the entire spectrum sensing process requires the sensing algorithm to quickly and accurately perceive the operating status of licensed users, which poses a significant challenge to the robustness of the algorithm and the performance of the hardware.

[0004] Machine learning-based spectrum sensing schemes leverage machine learning's powerful data classification capabilities to increase the opportunistic access possibilities of unauthorized users in cognitive radio networks (LFBs) while minimizing interference from unauthorized users to authorized users' receivers. However, existing machine learning-based spectrum sensing schemes rely on a large amount of spectrum data collected by a sufficient number of receivers over a sufficient period of time within cognitive radio networks. In practice, however, issues frequently arise such as receiver malfunctions leading to data loss, complex and variable wireless environments hindering the deployment of numerous spectrum observation devices, and limited permissible spectrum observation windows. Directly applying this spectrum observation data may impact model training and performance. Furthermore, the uneven distribution of spectrum state categories in complex and variable wireless spectrum environments results in a severe lack of spectrum data within a limited timeframe. Considering the overfitting problem of traditional machine learning algorithms when data volume is insufficient and the requirement for plug-and-play blind sensing, a new spectrum sensing method is urgently needed to address these issues. Summary of the Invention

[0005] To address the limitations of traditional machine learning-based spectrum state awareness algorithms in terms of data volume, this invention provides a wireless spectrum state image classification method based on few-shot learning. This method enables fast and accurate spectrum state awareness even with very limited data. Furthermore, this invention solves the problem of large intra-class variance in traditional few-shot learning algorithms by correcting the class prototype and improving classification accuracy.

[0006] Firstly, a wireless spectrum state image classification method based on few-shot learning is provided, the method comprising: classifying the training set and the support set S tr With training set query set Q ir The input is fed into the embedding module to obtain tensor sets for each category; in the training set support set S tr The tensor set calculates the training set support set S using the K closest tensors in each category. tr The class prototypes of each category; and the class prototypes are compared with the training set query set Q. tr The tensor set is input into the metric module to perform similarity measurement, resulting in a small sample image classifier.

[0007] Furthermore, in the training set support set S tr With training set query set Q tr Before being input into the embedding module, the method further includes: processing the training set support set S. tr and training set query set Q ir Then, perform interpolation to complete the result.

[0008] Furthermore, the embedding module includes: a fully convolutional neural network without fully connected layers, wherein the output of the fully convolutional neural network is a tensor of dimensions h×w×d.

[0009] Furthermore, the training set support set S tr The tensor set calculates the training set support set S using the K closest tensors in each category. tr The class prototypes for each category include: treating the tensors in the tensor set as feature vectors of length d (h×w dimensions), in the category Find the K nearest feature vectors to the given element; obtain the category based on the K feature vectors. The class prototype.

[0010] Furthermore, the step of matching the class prototype with the training set query set Q... lr The tensor set input metric module performs similarity measurement, including: calculating the training set query set Q. lr The sum of similarities between the tensor set and the class prototype; the sum of similarities is used as the training set query set Q. lr Samples and Categories The similarity between them.

[0011] Furthermore, after obtaining the few-sample image classifier, the method further includes: converting the test set support set S... te and test set query set Q le The data is input into the small sample image classifier, which performs the classification task and outputs a wireless spectrum status label.

[0012] Secondly, a wireless spectrum state image classification device based on few-sample learning is provided, comprising:

[0013] Processing unit, used to process training set and support set S tr With training set query set Q ir The input is fed into the embedding module to obtain tensor sets for each category;

[0014] The processing unit is also used in the training set support set S tr The tensor set calculates the training set support set S using the K closest tensors in each category. tr The class prototypes for each category;

[0015] The processing unit is further configured to match the class prototype with the training set query set Q. lr The tensor set is input into the metric module to perform similarity measurement, resulting in a small sample image classifier.

[0016] Thirdly, this application provides an electronic device comprising: a processor and a memory coupled to the processor, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor reads the computer instructions from the memory, the electronic device causes the electronic device to execute the wireless spectrum state image classification method based on few-sample learning as described in the first aspect.

[0017] Fourthly, this application provides a computer storage medium including computer instructions that, when executed on a terminal, cause the terminal to perform the wireless spectrum state image classification method based on few-shot learning as described in the first aspect.

[0018] Fifthly, this application provides a computer program product that, when the computer program product is executed computationally, causes the computer to perform the wireless spectrum state image classification method based on few-shot learning as described in the first aspect.

[0019] The method used in this invention differs from traditional machine learning algorithms in that the categories used for training and testing in classification tasks are different. This setup is beneficial for training a small-sample image classifier that can extract both general and specific features of the samples. It solves the problem of data dependence in traditional machine learning-based spectrum state perception algorithms and improves the perception accuracy in blind and rapid perception. At the same time, by using several feature vectors to represent image samples and using local features to represent images for classification, the problems of intra-class differences and background confusion are eliminated.

[0020] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0022] Figure 1 This is a communication scenario diagram after the application system model is constructed in the embodiments of this application;

[0023] Figure 2 A flowchart illustrating a wireless spectrum state image classification method based on few-shot learning, provided for embodiments of this application;

[0024] Figure 3 Specific training and testing flowcharts provided for embodiments of this application;

[0025] Figure 4 The wireless spectrum state sensing device provided in the embodiments of this application;

[0026] Figure 5 The electronic device described in this application is an embodiment of the present application. Detailed Implementation

[0027] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0028] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images, and should not be construed as limiting the invention. It is understandable that some well-known structures and their descriptions may be omitted in the drawings for those skilled in the art.

[0029] It should be noted that, as Figure 1 The scenario in which this invention is applied is a macro region, in which there are M authorized user transmitters and several unauthorized user receivers.

[0030] It should also be noted that, Figure 1 There are M unauthorized users and 3 authorized users in the region. The geographical area is divided into Q = O × P grids of equal area. Each grid is further divided into two different regions based on its distance from the nearest active authorized user transmitter. Region 1 is defined as the green region: a circle with radius D centered on the active authorized user transmitter. p The authorized user receiver distribution area. When an unauthorized user receiver is located in this area, it is strictly prohibited from accessing the LFB for communication under any conditions. Area 2 is defined as the blue area: with the active authorized user transmitter as the center and a radius of D. s The area outside the circle. Unlicensed user receivers located in this area are sufficiently far from licensed user receivers, and the interference they generate is sufficiently small, allowing unlicensed user receivers to freely access the LFB for communication. The operating status of licensed user transmitters in the LBF band can be marked as On or Off. This represents the working status of the N authorized user transmitters within the CCRN during the l-th sensing, where This indicates that the authorized user's transmitter is in an active transmission state; otherwise... This indicates that the authorized user transmitter is in an idle state. Therefore, there are a total of 2 such transmitters in the network. N A scenario involving joint transmission by authorized user transmitters is specifically represented as follows:

[0031] It should also be noted that the system model of this invention is as follows:

[0032] At time i, assuming the target unauthorized user receiver is located in grid q, the observed signal of the m-th receiver can be represented as:

[0033]

[0034] Among them, P n X represents the transmit power of the authorized user's transmitter. n (i) represents the i-th transmission signal of the n-th authorized user, N m (i) represents a mean of 0 and a variance of 0. Additive white Gaussian noise, h n，m The channel gain from the nth licensed user transmitter to the mth unlicensed user receiver can be expressed as:

[0035]

[0036] Among them, S n =[S n，x S n，y ] T R represents the position of the nth authorized user transmitter in the CCRN. m =[R m，x R m，y ] T Let |d| represent the location of the m-th unauthorized user receiver, ||.|| be the Euclidean distance, and PL(d) = d. -α It is the path power loss from the nth licensed user transmitter to the mth unlicensed user receiver. It is the shadowing fading relative to the signal of the nth licensed user transmitter and This is the multipath fading factor. For the m-th unlicensed user receiver, its acquired spectral observation energy is:

[0037]

[0038] Where w is the bandwidth of the observed signal, and τ is the sensing time. Therefore, the energy values ​​corresponding to all Q grids can be expressed in matrix form:

[0039]

[0040] in, The spectral observation energy value obtained by the unlicensed user receiver at the (o, p) grid location of the CCRN network during the lth sensing is determined by (3).

[0041] Within a certain time period, acquire the spectral observation energy matrix and corresponding state labels under different signal-to-noise ratios to obtain the dataset Θ={Y1,Y2,…,Y T} and its corresponding tag set

[0042] For obtaining a sparse low-rank energy observation dataset Θ={Y1,Y2,…,Y l Interpolation completion is performed. Considering the accuracy of interpolation and the handling of outliers, commonly used interpolation algorithms are the Inverse Distance Weighted (IDW) method and the Kriging method. This application does not restrict the specific interpolation algorithm.

[0043] Optionally, the IDW method is used to calculate the interpolation points. With the m-th sample point The distance between them gives the weight ω of the sample point to the interpolation point. m The interpolation point values ​​can be represented by a weighted average of the sample points as follows:

[0044]

[0045] In equation (5), the weight ω m It can be represented as:

[0046]

[0047]

[0048] Where p represents the degree to which the weight changes with distance, and is usually taken as -2.

[0049] Optionally, the estimated value at the interpolation point using the Kriging interpolation method is expressed as:

[0050]

[0051] The Kriging variance at a given location can be expressed as:

[0052]

[0053] in, For sample points Interpolation point The values ​​of the variograms are given by L, where L is a Lagrange multiplier.

[0054] The weight ω in equation (8)m The variance can be solved using Kriging variance:

[0055]

[0056] This yields a new spectral state dataset. and its corresponding status label set

[0057] It should also be noted that the aforementioned spectrum state dataset and the corresponding tag set Random sampling N tr ×N S The training set and its corresponding labels constitute the training set and support set. Random sampling N tr ×N Q The training set query set consists of 1 sample and its corresponding label.

[0058] The spectrum state dataset and the corresponding tag set Random sampling N te ×N S The test set support set consists of 10 samples and their corresponding labels. Random sampling N te ×N Q The test set query set consists of 1 sample and its corresponding label.

[0059] Figure 2 A flowchart illustrating a wireless spectrum state image classification method based on few-shot learning, as provided in this application embodiment, is shown below. Figure 2 As shown, the method is as follows:

[0060] S201: Transfer the training set and support set S tr With training set query set Q tr The input is fed into the embedding module to obtain tensor sets for each category;

[0061] S202: In the training set support set S tr The tensor set calculates the training set support set S using the K closest tensors in each category. tr The class prototypes for each category;

[0062] S203: Match the class prototype with the training set query set Q tr The tensor set is input into the metric module to perform similarity measurement, resulting in a small sample image classifier.

[0063] It should also be noted that the embedding module includes: a fully convolutional neural network without fully connected layers, the output of which is a tensor of dimensions h×w×d, the fully convolutional neural network contains three convolutional blocks, each of which consists of a convolutional layer, a 2×2 max pooling layer, a batch normalization layer, and a ReLU activation function layer; the fully convolutional neural network is named Conv64F because there are 64 filters of size 3×3 in each convolutional layer.

[0064] It should also be noted that, in order to correct the category prototype, this invention obtains a series of tensor sets for each category from the image through an embedding module. The algorithm searches for the K closest tensors to compute the class prototype.

[0065]

[0066] It should also be noted that the step of matching the class prototype with the training set query set Q... ir Tensor sets are input into the measurement module for similarity measurement:

[0067] Calculate the training set query set Q ir The sum of similarities between the tensor set and the class prototype is used as the training set query set Q. ir Samples and Categories The similarity between them.

[0068]

[0069] It should also be noted that this application uses an N-way K-shot task for training. Specifically, during the training process, each training episode samples a training task with different compositions. Therefore, the training process will contain different combinations of categories overall. The learning process minimizes the loss, and training stops when the difference between two iterations is less than a predefined threshold.

[0070] Optionally, the iterative algorithm includes stochastic gradient descent and batch gradient descent. This application does not limit the specific iterative optimization algorithm.

[0071] Optionally, use the stochastic gradient descent algorithm to minimize the class. negative log probability In each iteration, the loss function on a single training data point is randomly optimized.

[0072] It should also be noted that after obtaining the few-sample image classifier, the method further includes: processing the test set support set S. teand test set query set Q te The data is input into the few-sample image classifier, which performs a classification task and outputs a wireless spectrum status label. Figure 3 A flowchart illustrating the specific training and testing process.

[0073] Because few-shot learning enables models to extract both general and specific image features during N-way K-shot training, it allows for rapid learning with a small number of samples when encountering new categories. Therefore, it can achieve rapid perception even with minimal samples in blind sensing. Furthermore, by using several feature vectors to represent image samples and local features to represent images for classification, it eliminates intra-class discrepancies and background confusion, providing reliable analysis for subsequent dynamic spectral decision-making.

[0074] When using integrated units, Figure 4 A schematic diagram of one of the embodiments described above is shown. The device includes a processing unit 501.

[0075] Processing unit 501 is used to process the training set and support set S tr With training set query set Q lr The input is fed into the embedding module to obtain tensor sets for each category;

[0076] Processing unit 501 is further configured to, in the training set support set S tr The tensor set calculates the training set support set S using the K closest tensors in each category. tr The class prototypes for each category;

[0077] Processing unit 501 is further configured to match the class prototype with the training set query set Q lr The tensor set is input into the metric module to perform similarity measurement, resulting in a small sample image classifier.

[0078] It should also be noted that, for the sake of convenience and brevity, the specific working process of the wireless spectrum state image classification device based on few-sample learning described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0079] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 5 As shown, the device may include a processor 601 and a memory 602, the memory 602 being coupled to the processor 601. The memory is used to store computer program code, the computer program code including computer instructions. When the processor reads the computer instructions from the memory, the electronic device executes the wireless spectrum state image classification method based on few-shot learning provided in the embodiment.

[0080] This application also provides a computer-readable storage medium that may include a computer program or instructions that, when executed on a computer, cause the computer to perform the wireless spectrum state image classification method based on few-shot learning described in the above embodiments.

[0081] This application provides a computer program product, including a computer program or instructions, which, when run on a computer, cause the computer to execute the wireless spectrum state image classification method based on few-shot learning described in the above embodiments.

[0082] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented using software programs, the above embodiments can appear, in whole or in part, as a computer program product, which includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated.

[0083] It should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0084] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A wireless spectrum state image classification method based on small sample learning, characterized in that, The method comprises: dividing a geographical area into Q grids; obtaining a wireless spectrum observation energy matrix and a corresponding state label under different signal-to-noise ratios in a certain time for the geographical area, to obtain a wireless spectrum state data set and a corresponding label set; Randomly sampling the spectrum state data set and the corresponding label set respectively, a training set support set S tr , a training set query set Q tr , a test set support set S te , and a test set query set Q te may be obtained. The training set support set S tr With the training set query set Q tr Input into the embedding module to obtain a set of tensors for each category; The embedding module comprises a fully convolutional neural network without a full connection layer, and the fully convolutional neural network comprises three convolutional blocks, each of which is composed of a convolutional layer, a 2x2 maximum pooling layer, a batch normalization layer and an activation function ReLU layer. In the training set support set S tr In the training set support set S tr In the training set support set S The protoype is compared to the training set query set Q tr The tensor set input metric module performs similarity measurement to obtain a small sample image classifier.

2. The method of claim 1, wherein, In the case of a training set query set Q tr With the training set query set Q tr The method further comprises, prior to input to the embedding module: S = {x | x e X and x e S} (1) tr Q = {x | x e X and x e Q} (2) tr and interpolating the missing values.

3. The method of claim 1, wherein, The embedding module comprises a fully convolutional neural network without a full connection layer, and the fully convolutional neural network comprises three convolutional blocks, each of which is composed of a convolutional layer, a 2x2 maximum pooling layer, a batch normalization layer and an activation function ReLU layer.

4. The method of claim 1, wherein, The training set support set S tr The training set support set S is calculated by the K tensors closest in distance in each category in the tensor set tr The class prototype of each category of the training set support set S includes: The embedding module comprises a fully convolutional neural network without a full connection layer, and the fully convolutional neural network comprises three convolutional blocks, each of which is composed of a convolutional layer, a 2x2 maximum pooling layer, a batch normalization layer and an activation function ReLU layer. The tensors in the tensor set are regarded as feature vectors with a length of d and a size of h x w dimensions, and K feature vectors closest to the tensors in the class l are found; 5. The method of claim 1, wherein, said protoype to said training set query set Q tr tensor set input metric module, performing similarity metric, comprising: computing the training set query set Q tr a sum of similarities between the tensor set and the class prototypes According to the sum of the similarities as the training set query set Q tr Similarity between sample and class `.

6. The method of claim 1, wherein, According to the K feature vectors, a class prototype of the class l is obtained. supporting set S for the test set te and test set query set Q te inputting into the small sample image classifier, performing a classification task using the classifier, and outputting a classification result as a wireless spectrum status label.

7. A wireless spectrum state image classification apparatus based on small sample learning, characterized by, After obtaining the small sample image classifier, the method further comprises: a tensor set processing unit for processing the training set support set S tr with the training set query set Q tr input to the embedding module to obtain the tensor set of each class; The class prototype processing unit is further configured to compute a support set S of the training set tr of the training set S by K tensors closest in distance in each class tr class prototypes of respective classes of the training set S The similarity measurement processing unit is further configured to compare the class prototype with the training set query set Q tr The tensor set input measurement module is configured to perform similarity measurement; and when data is injected into the processing unit, the image classification device is caused to perform the wireless spectrum state image classification method based on small sample learning according to any one of claims 1-6.

8. An electronic device, comprising: comprises: comprises:

9. A computer-readable storage medium, characterized in that, A processor and a memory, the memory is coupled with the processor, the memory is used for storing computer program code, the computer program code comprises computer instructions, when the processor reads the computer instructions from the memory, so that the electronic equipment executes the wireless spectrum state image classification method based on small sample learning as claimed in any one of claims 1-6.

10. A computer program product, characterised in that, Comprise computer instructions, when the computer instructions run on the terminal, make the terminal execute the wireless spectrum state image classification method based on small sample learning as claimed in any one of claims 1-6. When the computer program product is calculated, make the computer execute the wireless spectrum state image classification method based on small sample learning as claimed in any one of claims 1-6.

Images