Training method, device and equipment of spectrum sensing model and spectrum sensing method

Abstract

The invention discloses a training method, device and equipment of a spectrum sensing model and a spectrum sensing method. The training method comprises the following steps: acquiring a plurality of to-be-sensed signal training samples in radio wave propagation; inputting the plurality of training samples into a feature extraction network of a preset model to obtain feature matrixes corresponding to the plurality of training samples; screening the plurality of feature matrixes according to the IoU coefficient to obtain a target sample; according to the target sample, training a regional suggestion network of a preset model by using a smooth loss function to obtain an RPN suggestion; training a classification network of a preset model by using a generative adversarial network according to the RPN suggestion and the feature matrix; and when the preset training stop condition is not satisfied, returning to the step of inputting the plurality of training samples into the preset model until the preset training stop condition is satisfied, and obtaining a trained spectrum sensing model. According to the embodiment of the invention, the finally trained spectrum sensing model is good in performance and reliable, and can be used in a complex radio frequency environment.

Description

Technical Field

[0001] The present application belongs to the field of cognitive radio technology, and in particular, relates to a spectrum sensing model training method, device, equipment and spectrum sensing method. Background Art

[0002] Spectrum sensing is crucial for allocating spectrum resources for communications and organizing monitoring activities. Spectrum sensing technology can be used to implement cognitive radio, which can dynamically identify and utilize idle spectrum resources, thereby improving the radio system's perception of the surrounding electromagnetic environment and spectrum utilization efficiency.

[0003] With the development of machine learning technology, machine learning is increasingly used in spectrum sensing. However, in the radio wave propagation environment, there are problems such as a small number of samples available for deep learning or frequent changes in the electromagnetic environment. When the environment changes, the spectrum sensing model trained with old samples cannot be reliably used in the new environment, resulting in the inability of the spectrum sensing model to quickly achieve spectrum sensing in this scenario, and the sensing performance is greatly reduced. Summary of the invention

[0004] The embodiments of the present application provide a spectrum sensing model training method, apparatus, device and spectrum sensing method, which are used to at least solve the problem in the related art that when the electromagnetic environment changes, the spectrum sensing model trained with old samples cannot be reliably used in the new environment, resulting in low perception performance.

[0005] In a first aspect, an embodiment of the present application provides a method for training a spectrum sensing model, including:

[0006] Obtaining multiple training samples of signals to be sensed in radio wave propagation;

[0007] Inputting the plurality of the to-be-sensed signal training samples into a feature extraction network of a preset model to obtain feature matrices corresponding to the plurality of the to-be-sensed signal training samples;

[0008] Screen multiple feature matrices according to the IoU coefficient to obtain the target sample;

[0009] According to the target sample, a region proposal network of the preset model is trained using a smooth loss function to obtain an RPN proposal;

[0010] According to the RPN suggestion and the feature matrix, a classification network of the preset model is trained using a generative adversarial network;

[0011] If the preset training stop condition is not met, return to the step of inputting the plurality of the to-be-perceived signal training samples into the feature extraction network of the preset model until the preset training stop condition is met to obtain a trained spectrum perception model.

[0012] In a second aspect, an embodiment of the present application provides a spectrum sensing method, including:

[0013] Acquire the signal to be sensed in radio wave propagation;

[0014] Inputting the signal to be sensed into a pre-trained spectrum sensing model, wherein the spectrum sensing model includes a feature extraction network, a region proposal network, a pooling region, a classification network, a first fully connected layer, and a second fully connected layer;

[0015] Performing dimensionality reduction processing on the signal to be sensed through the feature extraction network to obtain a feature matrix, wherein the dimension of the feature matrix is ​​determined according to the dimensionality reduction coefficient of the feature extraction network;

[0016] Converting the feature matrix into a classification matrix and a regression matrix through the region proposal network;

[0017] Performing window processing on the classification matrix and the regression matrix respectively through the pooling region to obtain a first vector corresponding to the classification matrix and a second vector corresponding to the regression matrix, wherein the sizes of the first vector and the second vector are both preset sizes;

[0018] Flattening the first vector and the second vector respectively through the classification network to obtain a flattened first vector and a flattened second vector;

[0019] The flattened first vector is classified through the first fully connected layer, the flattened second vector is regressed through the second fully connected layer, and a judgment is made according to a preset threshold to obtain a spectrum sensing result.

[0020] In a third aspect, an embodiment of the present application provides a training device for a spectrum sensing model, the device comprising:

[0021] An acquisition module, used to acquire a plurality of training samples of signals to be sensed in radio wave propagation;

[0022] An input module, used to input the plurality of the to-be-sensed signal training samples into a feature extraction network of a preset model to obtain feature matrices corresponding to the plurality of the to-be-sensed signal training samples;

[0023] The screening module is used to screen multiple feature matrices according to the IoU coefficient to obtain the target sample;

[0024] A first training module is used to train the region proposal network of the preset model according to the target sample using a smooth loss function to obtain an RPN proposal;

[0025] A second training module is used to train the classification network of the preset model using a generative adversarial network according to the RPN suggestion and the feature matrix;

[0026] The third training module is used to return to the step of inputting the plurality of training samples of the signal to be sensed into the feature extraction network of the preset model when the preset training stop condition is not met, until the preset training stop condition is met, so as to obtain a trained spectrum sensing model.

[0027] In a fourth aspect, an embodiment of the present application provides an electronic device, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the steps of the spectrum sensing model training method as described in any one of the embodiments of the first aspect are implemented.

[0028] In a fifth aspect, an embodiment of the present application provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the steps of the spectrum sensing model training method described in any one of the embodiments of the first aspect are implemented.

[0029] The training method, device, equipment and spectrum sensing method of the spectrum sensing model of the embodiment of the present application construct a fast CNN spectrum sensing architecture based on the acquired training samples of the signals to be sensed in multiple radio wave propagations, and optimize the feature extraction network, region proposal network and classification network architecture therein; the input signal samples are screened according to the IoU coefficient, and the screened samples are used to train the region proposal network to generate RPN suggestions, and then the RPN suggestions are used to train the classification network, and a generative adversarial network is introduced in the training of the classification network, and the spectrum data enhancement mechanism and the domain adaptation mechanism are trained to adapt the training data to the dynamic changes of the spectrum; the spectrum sensing model finally trained can be used for spectrum sensing of one-dimensional signals, the model has good performance and reliability, improves the accuracy of spectrum sensing, reduces the complexity of spectrum sensing, and can be used in complex radio frequency environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solution of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0031] Figure 1 It is a flowchart of a method for training a spectrum sensing model provided in an embodiment of the present application;

[0032] Figure 2 is a schematic diagram of a data enhancement mechanism provided in an embodiment of the present application;

[0033] Figure 3 is a schematic diagram of a domain adaptation mechanism provided in an embodiment of the present application;

[0034] Figure 4 It is a flowchart of a spectrum sensing method provided by an embodiment of the present application;

[0035] Figure 5 It is a structural schematic diagram of a training device for a spectrum sensing model provided in an embodiment of the present application;

[0036] Figure 6 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application.

[0037] Reference numerals:

[0038] A spectrum sensing model training device 500, an acquisition module 501, an input module 502, a screening module 503, a first training module 504, a second training module 505, a third training module 506,

[0039] Electronic device 600 , processor 601 , memory 602 , communication interface 603 , bus 610 . DETAILED DESCRIPTION

[0040] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by illustrating the examples of the present application.

[0041] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0042] Spectrum sensing is crucial for allocating spectrum resources for communications and organizing monitoring activities. Spectrum sensing technology can be used to implement cognitive radio, which can dynamically identify and utilize idle spectrum resources, thereby improving the radio system's perception of the surrounding electromagnetic environment and spectrum utilization efficiency.

[0043] However, with the increasing frequency congestion and the increasing complexity of device usage patterns, spectrum sensing has become increasingly difficult. In traditional spectrum sensing technology, energy-based detection methods are insensitive to weak signals in low signal-to-noise ratio situations, cannot distinguish different signal types, and are not suitable for complex multi-signal environments.

[0044] With the development of machine learning technology, machine learning is increasingly used in spectrum sensing. However, in the radio wave propagation environment, there are problems such as a small number of samples available for deep learning or frequent changes in the electromagnetic environment. When the environment changes, the spectrum sensing model trained with old samples cannot be reliably used in the new environment, resulting in the inability of the spectrum sensing model to quickly achieve spectrum sensing in this scenario, and the sensing performance is greatly reduced.

[0045] In order to solve the problems of related technologies, the embodiments of the present application provide a spectrum sensing model training method, apparatus, device and spectrum sensing method.

[0046] The following describes in detail the training method of the spectrum sensing model provided by the embodiment of the present application through specific embodiments and application scenarios in conjunction with the accompanying drawings.

[0047] Figure 1 FIG. 1 is a flow chart of a spectrum sensing model training method 100 according to an embodiment of the present application. Figure 1 As shown, the spectrum sensing model training method 100 may specifically include the following steps:

[0048] S101, obtaining a plurality of training samples of signals to be sensed in radio wave propagation;

[0049] S102, inputting the plurality of the to-be-sensed signal training samples into a feature extraction network of a preset model to obtain feature matrices corresponding to the plurality of the to-be-sensed signal training samples;

[0050] S103, screening multiple feature matrices according to the IoU coefficient to obtain target samples;

[0051] S104, according to the target sample, using a smooth loss function to train the region proposal network of the preset model to obtain an RPN proposal;

[0052] S105, training the classification network of the preset model using a generative adversarial network according to the RPN suggestion and the feature matrix;

[0053] S106. If the preset training stop condition is not met, return to the step of inputting the plurality of training samples of the signal to be sensed into the feature extraction network of the preset model until the preset training stop condition is met to obtain a trained spectrum sensing model.

[0054] Therefore, based on the obtained training samples of the signals to be sensed in multiple radio wave propagations, a fast CNN spectrum sensing architecture was constructed, and the feature extraction network, region proposal network and classification network architecture were optimized; the input signal samples were screened according to the IoU coefficient, and the screened samples were used to train the region proposal network to generate RPN suggestions, and then the RPN suggestions were used to train the classification network, and the generative adversarial network was introduced in the training of the classification network, and the spectrum data enhancement mechanism and domain adaptation mechanism were trained to make the training data adapt to the dynamic changes of the spectrum; the spectrum sensing model finally trained can be used for spectrum sensing of one-dimensional signals, the model has good performance and reliability, improves the accuracy of spectrum sensing, reduces the complexity of spectrum sensing, and can be used in complex RF environments.

[0055] The specific implementation methods of the above steps are introduced below.

[0056] It should be noted that the preset model in the embodiment of the present application includes a convolutional neural network (CNN), and in particular, a fast CNN. Optionally, the specific structure of the preset model may include: a feature extraction network, a region proposal network (RPN), a pooling region, a classification network (classifier), and a fully connected layer.

[0057] Among them, the feature extraction layer of the preset model can be implemented using a fully convolutional network to map each feature to a region in the input; for the region proposal network, a convolutional layer can be selected as an intermediate layer to further extract features and set feature depth. For example but not limited to: the region proposal network uses a 3×1 convolutional layer as an intermediate layer, and its output is divided into 2 separate 1×1 convolutional layers, which are used for classification and regression respectively, that is, the output is an n×k classification matrix and an n×2k regression matrix.

[0058] It should be noted that the training samples of the signals to be sensed in the radio wave propagation obtained in S101 include but are not limited to: signal energy values, signal-to-noise ratio data, and category labels, etc., which are not limited in this embodiment.

[0059] In some embodiments, in S102, the dimension of the feature matrix is ​​less than the dimension of the training sample of the signal to be sensed, that is, the feature matrix after dimension reduction is obtained by performing dimensionality reduction processing on the training sample of the signal to be sensed through the feature extraction network, and the dimension of the feature matrix is ​​determined according to the dimensionality reduction coefficient of the feature extraction network. Specifically, for the dimensionality reduction coefficient, in order to ensure that the feature matrix is ​​directly mapped to the input, the dimensionality reduction coefficient must be an integer divisible by the input size to avoid truncation, otherwise truncation will result in inaccurate mapping of the position between the feature and the signal space. In addition, the dimensionality reduction coefficient can be determined according to the actual application requirements to make a trade-off between increasing the feature granularity, the number of anchor frames and the complexity. For example, a VGG network with a dimensionality reduction coefficient of 16 is selected as the feature extraction layer.

[0060] In some embodiments, an intermediate data set (ie, target samples) is required to train the region proposal network, which includes the classification and regression targets of the anchor points. It should be understood that since the dimension of the feature matrix is ​​known, the anchor point position can also be found from it.

[0061] Specifically, in S103, for each feature, the anchor point position can be screened by the IoU coefficient, and the IoU coefficient calculation formula (1) of regions A and B is as follows:

[0062]

[0063] Furthermore, when the IoU coefficient is greater than a first preset value, it is determined that the region A and region B corresponding to the feature matrix are positively overlapped; when the IoU coefficient is less than a second preset value, it is determined that the region A and region B corresponding to the feature matrix are negatively overlapped; when the IoU coefficient is less than or equal to the first preset value and greater than or equal to the second preset value, it is determined that the overlap between the region A and region B corresponding to the feature matrix is ​​fuzzy; when the region A and region B corresponding to the feature matrix are positively overlapped or negatively overlapped, the feature matrix is ​​determined as a target sample.

[0064] Taking the first preset value of 0.8 and the second preset value of 0.2 as an example, if the IoU coefficient is greater than 0.8, it is considered a positive overlap; if the IoU coefficient is less than 0.2, it is considered a negative overlap; if the IoU is between 0.2 and 0.8, the overlap is considered fuzzy and cannot be used for training. In this way, the target sample is obtained through fuzzy classification, which helps to ensure that the region proposal network learns on the samples with confidence.

[0065] Further, in some embodiments, in S104, for the region proposal network regression, a smooth loss function may be used, and formula (2) is as follows:

[0066]

[0067] Among them, ti represents the predicted regression target, Indicates overlap with actual ground conditions.

[0068] As an optional embodiment, the specific implementation of S105 may include: training a generative adversarial network according to the RPN suggestion and the feature matrix to obtain a synthetic sample; and training a classification network of the preset model using the series connection of the signal training sample to be sensed and the synthetic sample. The generative adversarial network includes a generator network and a discriminator network.

[0069] In a specific implementation, according to the RPN recommendation and the feature matrix, the generator network and the discriminator network are trained by error back propagation to obtain synthetic samples generated by the generator network.

[0070] In specific implementation, the goal of the generator network and the discriminator network is to solve the maximum and minimum value problem. Formula (3) is as follows:

[0071]

[0072] Among them, z represents the noise input to the generator G; Pz represents the model distribution of the generator G; G(z) represents the output of the generator G; Pdata represents the distribution of the input data r; the discriminator D is used to distinguish between real samples and generated samples (i.e., synthetic samples).

[0073] In addition, in some optional embodiments, in order to solve the gradient vanishing problem, the objectives of the discriminator and the generator are separated into the following two items, namely, formulas (4) and (5):

[0074]

[0075]

[0076] It should be noted that the label L(r) is added to the conditional GAN ​​model to represent the generated data instead of the real data. That is, for the actual calculation of formulas (4) and (5), D(r) and G(z) are replaced by D(r, L(r)) and G(z, L(r)) respectively. In this way, the discriminator and generator are trained by error back propagation.

[0077] Optionally, in the conditional GAN, a deep learning structure is used, which is equivalent to a three-layer neural network (feedforward neural network), each layer has multiple (e.g., 100) neurons for the discriminator and the generator. In the hidden layer, leaky-ReLU is used as the activation function; the output layer uses sigmoid as the activation function.

[0078] Among them, leaky-ReLU performs the following operation of formula (6):

[0079] f(x)=max(αx,x) (6)

[0080] Therefore, first, the data augmentation mechanism is trained. Figure 2 , is the principle diagram of the data enhancement mechanism. Figure 2 As shown in Figure 1, the generator network maps the noise source to the input space, and the discriminator and generator are trained through error back propagation; then the synthetic samples generated by the generator and the real data samples are connected in series to train the classifier. In other words, the upper branch of the discriminator is the real training sample, and the lower branch of the discriminator is the synthetic sample.

[0081] Specifically, the conditional GAN ​​is trained using real training samples, and the generator learns to synthesize new data samples; the classifier is trained using enhanced data: GAN is used to generate new synthetic samples T', and the classifier C' is trained using the concatenation of real data samples T and synthetic data samples T'.

[0082] Then, train the domain adaptation model. Figure 3 , is the principle diagram of the domain adaptation mechanism. Figure 3 As shown in the figure, a bidirectional GAN ​​is used to learn the inverse mapping from random noise to real samples, and the new samples of the new environment and the inverse mapping are used as the input of the discriminator to train the CGAN. After the CGAN training is completed, the output of the generator is used to train a new classifier, and the new classifier is used to annotate samples in the new environment.

[0083] In specific implementation, the sample T1 in the environment E1 is input to the autoencoder Enc, and the random noise z is input to the generator Discriminator In the upper branch, [T1, Enc(T1)] is taken as the true sample, and in the lower branch, As a false sample, and trained by back propagation. When the GAN training converges, the encoder will learn the inverse mapping from the noise domain to the sample domain in the environment E1. In this way, the inverse mapping is obtained, that is,

[0084] Furthermore, when the environment changes from E1 to E2, a new CGAN is trained, and the new samples in environment E2 are used as real input. Specifically, the inverse mapping Enc(T1) of the bidirectional GAN ​​is used to train the CGAN, and the labels in environment E1 are carried. After the CGAN training is completed, the classifier C'2 is trained with the samples G(Enc(T1)) in the adaptive domain, and the classifier C'2 can be used to label the new samples in environment E2.

[0085] In this way, by introducing the generative adversarial network in the classifier training process, the training data enhancement mechanism adds synthetic data samples to the existing training data set in a given spectrum environment; the training domain adaptation mechanism generates synthetic training data for new spectrum environments. Thus, adversarial learning is used to generate high-fidelity synthetic training data and train new classifiers for new spectrum conditions. In turn, the accuracy of spectrum perception in new environments can be improved.

[0086] It can be seen that the preset training stop condition in S106 includes the regression target of the region proposal network training and the convergence condition of the classification network training. During the iterative training process, it is determined whether the preset training stop condition is met. If not, it returns to execute S102, that is, the steps of training the region proposal network and the steps of training the classification network are performed alternately until the preset training stop condition is met, the entire model training is completed, and a spectrum sensing model is obtained.

[0087] In addition, the present application also provides a spectrum sensing method 400. Figure 4 FIG. 4 is a flow chart showing a spectrum sensing method 400 according to an embodiment of the present application. Figure 4 As shown, the spectrum sensing method 400 may specifically include the following steps: S401 to S407.

[0088] S401: Acquire a signal to be sensed during radio wave propagation.

[0089] Specifically, the signal to be sensed in radio wave propagation may include: signal energy value, signal-to-noise ratio data, etc., which is not limited in this embodiment.

[0090] S402: Input the signal to be sensed into a pre-trained spectrum sensing model, where the spectrum sensing model includes a feature extraction network, a region proposal network, a pooling region, a classification network, a first fully connected layer, and a second fully connected layer.

[0091] S403: Performing dimensionality reduction processing on the signal to be sensed through the feature extraction network to obtain a feature matrix, wherein the dimension of the feature matrix is ​​determined according to the dimensionality reduction coefficient of the feature extraction network.

[0092] Optionally, a VGG network is used to map each feature to a region in the input. To avoid truncation, the dimension reduction factor needs to be an integer divisible by the input size. For example, a VGG network with a dimension reduction factor of 16 is selected as the feature extraction layer.

[0093] S404: Convert the feature matrix into a classification matrix and a regression matrix through the region proposal network.

[0094] Optionally, a convolutional layer is selected as the middle layer of the region proposal network to further extract features and set the feature depth. For example, if a 3×1 convolutional layer is selected as the middle layer, its output is divided into 2 separate 1×1 convolutional layers, one for classification and one for regression, that is, the output is an n×k classification matrix and an n×2k regression matrix.

[0095] S405. Perform window processing on the classification matrix and the regression matrix respectively through the pooling region to obtain a first vector corresponding to the classification matrix and a second vector corresponding to the regression matrix, wherein the sizes of the first vector and the second vector are both preset sizes.

[0096] Optionally, the pooling region is windowed around each region of interest (ROI) and then resized into a vector of 1 × 7. This standard size enables it to be processed in a fully connected layer.

[0097] S406: Flatten the first vector and the second vector respectively through the classification network to obtain a flattened first vector and a flattened second vector.

[0098] S407: Classify the flattened first vector through the first fully connected layer, regress the flattened second vector through the second fully connected layer, and make a judgment according to a preset threshold to obtain a spectrum sensing result.

[0099] In this way, the features are flattened by the classification network of the spectrum sensing model, and then processed by two fully connected layers, that is, the output of the classification network is passed to two different fully connected layers for classification and regression. Finally, according to the output of the classification network, the decision threshold is set to further improve the performance of spectrum sensing, obtain the final output result, and realize spectrum sensing.

[0100] It can be seen that the trained spectrum sensing model can use fast CNN for one-dimensional spectrum sensing, improve the accuracy of spectrum sensing, reduce the complexity of spectrum sensing, and can be used in complex RF environments.

[0101] It should be noted that the above describes some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0102] Based on the same technical concept, corresponding to the spectrum sensing model training method of any of the above embodiments, the present application also provides a spectrum sensing model training device 500.

[0103] like Figure 5 As shown, the spectrum sensing model training device 500 may include:

[0104] An acquisition module 501 is used to acquire a plurality of training samples of signals to be sensed in radio wave propagation;

[0105] An input module 502 is used to input the plurality of the to-be-sensed signal training samples into a feature extraction network of a preset model to obtain feature matrices corresponding to the plurality of the to-be-sensed signal training samples;

[0106] A screening module 503 is used to screen multiple feature matrices according to the IoU coefficient to obtain a target sample;

[0107] A first training module 504 is used to train the region proposal network of the preset model according to the target sample using a smooth loss function to obtain an RPN suggestion;

[0108] A second training module 505 is used to train the classification network of the preset model using a generative adversarial network according to the RPN suggestion and the feature matrix;

[0109] The third training module 506 is used to return to the step of inputting the plurality of the to-be-perceived signal training samples into the feature extraction network of the preset model when the preset training stop condition is not met, until the preset training stop condition is met, to obtain a trained spectrum perception model.

[0110] In some embodiments, the second training module 505 is specifically used to train the generative adversarial network according to the RPN recommendation and the feature matrix to obtain synthetic samples; and train the classification network of the preset model using the series connection of the training samples of the signal to be sensed and the synthetic samples.

[0111] In some embodiments, the generative adversarial network includes a generator network and a discriminator network. The second training module 505 is further specifically used to train the generator network and the discriminator network through error back propagation according to the RPN suggestion and the feature matrix to obtain synthetic samples generated by the generator network.

[0112] In some embodiments, the IoU coefficient is:

[0113]

[0114] In some embodiments, the screening module 503 is specifically used to determine that the region A and region B corresponding to the feature matrix are positively overlapped when the IoU coefficient is greater than a first preset value; determine that the region A and region B corresponding to the feature matrix are negatively overlapped when the IoU coefficient is less than a second preset value; determine that the overlap between the region A and region B corresponding to the feature matrix is ​​fuzzy when the IoU coefficient is less than or equal to the first preset value and greater than or equal to the second preset value; and determine the feature matrix as a target sample when the region A and region B corresponding to the feature matrix are positively overlapped or negatively overlapped.

[0115] In some embodiments, the dimension of the feature matrix is ​​smaller than the dimension of the training sample of the signal to be sensed, and the dimension of the feature matrix is ​​determined according to the dimension reduction coefficient of the feature extraction network.

[0116] In some embodiments, the intermediate layer of the region proposal network is a convolutional layer.

[0117] It should be noted that, for the convenience of description, the above device is described in various modules according to their functions. Of course, when implementing this application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0118] The apparatus of the above embodiment is used to implement the training method of the corresponding spectrum sensing model in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0119] Based on the same technical concept, corresponding to the training method of the spectrum sensing model in any of the above embodiments, the present application also provides an electronic device.

[0120] Figure 6 A more specific schematic diagram of the hardware structure of an electronic device provided by this embodiment is shown.

[0121] The electronic device 600 may include a processor 601 and a memory 602 storing computer program instructions.

[0122] Specifically, the processor 601 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0123] The memory 602 may include a large capacity memory for data or instructions. By way of example and not limitation, the memory 602 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. In appropriate cases, the memory 602 may include a removable or non-removable (or fixed) medium. In appropriate cases, the memory 602 may be inside or outside the integrated gateway disaster recovery device. In a specific embodiment, the memory 602 is a non-volatile solid-state memory.

[0124] In certain embodiments, the memory may include a read-only memory (ROM), a random access memory (RAM), a magnetic disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical or other physical / tangible memory storage device. Thus, typically, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., a memory device) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to one aspect of the present application.

[0125] The processor 601 implements any one of the spectrum sensing model training methods in the above embodiments by reading and executing computer program instructions stored in the memory 602 .

[0126] In some examples, the electronic device 600 may further include a communication interface 603 and a bus 610. Figure 6 As shown, the processor 601, the memory 602, and the communication interface 603 are connected via a bus 610 and communicate with each other.

[0127] The communication interface 603 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiments of the present application.

[0128] Bus 610 includes hardware, software or both, and the parts of online data flow billing equipment are coupled to each other. For example, but not limitation, bus 610 may include accelerated graphics port (AGP) or other graphics bus, enhanced industrial standard architecture (EISA) bus, front-end bus (FSB), hypertransport (HT) interconnection, industrial standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnection (PCI) bus, PCI-Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations. In appropriate cases, bus 610 may include one or more buses. Although the present application embodiment describes and shows a specific bus, the present application considers any suitable bus or interconnection.

[0129] Exemplarily, the electronic device 600 may be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA).

[0130] Based on the same technical concept, corresponding to the training method of the spectrum sensing model of any of the above embodiments, the present application also provides a non-transitory computer-readable storage medium. The computer-readable storage medium stores computer program instructions; when the computer program instructions are executed by the processor, any of the spectrum sensing model training methods in the above embodiments is implemented. Examples of computer-readable storage media include non-transitory computer-readable storage media, such as portable disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, etc.

[0131] Based on the same technical concept, corresponding to the training method of the spectrum sensing model in any of the above embodiments, the present application also provides a computer program product, which includes computer program instructions. In some embodiments, the computer program instructions can be executed by one or more processors of a computer so that the computer and / or the processor execute the training method of the spectrum sensing model. Corresponding to the execution subject corresponding to each step in each embodiment of the training method of the spectrum sensing model, the processor that executes the corresponding step may belong to the corresponding execution subject.

[0132] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0133] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present application are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.

[0134] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiment, or in a different order from the embodiment, or several steps can be performed simultaneously.

[0135] The above reference is according to the method of the embodiment of the present application, the flow chart of the device (system) and the computer program product and / or the block diagram described various aspects of the present application.It should be understood that each square box in the flow chart and / or the block diagram and the combination of each square box in the flow chart and / or the block diagram can be realized by computer program instructions.These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device to produce a machine so that these instructions executed by the processor of the computer or other programmable data processing device enable the realization of the function / action specified in one or more square boxes of the flow chart and / or the block diagram.Such a processor can be but is not limited to a general-purpose processor, a special-purpose processor, a special application processor or a field programmable logic circuit.It can also be understood that each square box in the block diagram and / or the flow chart and the combination of the square boxes in the block diagram and / or the flow chart can also be realized by the dedicated hardware that performs the specified function or action, or can be realized by the combination of dedicated hardware and computer instructions.

[0136] The above is only a specific implementation of the present application. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the protection scope of the present application is not limited to this. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this application, and these modifications or replacements should be included in the protection scope of this application.

Claims

1. A method for training a spectrum sensing model, characterized in that: include: Obtaining multiple training samples of signals to be sensed in radio wave propagation; Inputting the plurality of the to-be-sensed signal training samples into a feature extraction network of a preset model to obtain feature matrices corresponding to the plurality of the to-be-sensed signal training samples; Screen multiple feature matrices according to the IoU coefficient to obtain the target sample; According to the target sample, a region proposal network of the preset model is trained using a smooth loss function to obtain an RPN proposal; According to the RPN suggestion and the feature matrix, a classification network of the preset model is trained using a generative adversarial network; If the preset training stop condition is not met, return to the step of inputting the plurality of the to-be-perceived signal training samples into the feature extraction network of the preset model until the preset training stop condition is met to obtain a trained spectrum perception model.

2. The method according to claim 1, characterized in that: The method of training the classification network of the preset model by using a generative adversarial network according to the RPN suggestion and the feature matrix includes: Training a generative adversarial network according to the RPN suggestion and the feature matrix to obtain a synthetic sample; The classification network of the preset model is trained by using the series connection of the signal training sample to be sensed and the synthetic sample.

3. The method according to claim 2, characterized in that The generative adversarial network includes a generator network and a discriminator network; The step of training a generative adversarial network according to the RPN suggestion and the feature matrix to obtain a synthetic sample includes: According to the RPN proposal and the feature matrix, the generator network and the discriminator network are trained by error back propagation to obtain synthetic samples generated by the generator network.

4. The method according to claim 1, characterized in that: The IoU coefficient is: The method of screening multiple feature matrices according to the IoU coefficient to obtain target samples includes: When the IoU coefficient is greater than the first preset value, it is determined that the region A and region B corresponding to the feature matrix are positively overlapped; when the IoU coefficient is less than the second preset value, it is determined that the region A and region B corresponding to the feature matrix are negatively overlapped; when the IoU coefficient is less than or equal to the first preset value and greater than or equal to the second preset value, it is determined that the overlap between the region A and region B corresponding to the feature matrix is ​​fuzzy; When the region A and the region B corresponding to the feature matrix are positively overlapped or negatively overlapped, the feature matrix is ​​determined as a target sample.

5. The method according to claim 1, characterized in that The dimension of the feature matrix is ​​smaller than the dimension of the training sample of the signal to be sensed, and the dimension of the feature matrix is ​​determined according to the dimension reduction coefficient of the feature extraction network.

6. The method according to claim 1, characterized in that The middle layer of the region proposal network is a convolutional layer.

7. A spectrum sensing method, characterized in that: include: Acquire the signal to be sensed in radio wave propagation; Inputting the signal to be sensed into a pre-trained spectrum sensing model, wherein the spectrum sensing model includes a feature extraction network, a region proposal network, a pooling region, a classification network, a first fully connected layer, and a second fully connected layer; Performing dimensionality reduction processing on the signal to be sensed through the feature extraction network to obtain a feature matrix, wherein the dimension of the feature matrix is ​​determined according to the dimensionality reduction coefficient of the feature extraction network; Converting the feature matrix into a classification matrix and a regression matrix through the region proposal network; Performing window processing on the classification matrix and the regression matrix respectively through the pooling region to obtain a first vector corresponding to the classification matrix and a second vector corresponding to the regression matrix, wherein the sizes of the first vector and the second vector are both preset sizes; Flattening the first vector and the second vector respectively through the classification network to obtain a flattened first vector and a flattened second vector; The flattened first vector is classified through the first fully connected layer, the flattened second vector is regressed through the second fully connected layer, and a judgment is made according to a preset threshold to obtain a spectrum sensing result.

8. A training device for a spectrum sensing model, characterized in that: The device comprises: An acquisition module, used to acquire a plurality of training samples of signals to be sensed in radio wave propagation; An input module, used to input the plurality of the to-be-sensed signal training samples into a feature extraction network of a preset model to obtain feature matrices corresponding to the plurality of the to-be-sensed signal training samples; The screening module is used to screen multiple feature matrices according to the IoU coefficient to obtain the target sample; A first training module is used to train the region proposal network of the preset model according to the target sample using a smooth loss function to obtain an RPN proposal; A second training module is used to train the classification network of the preset model using a generative adversarial network according to the RPN suggestion and the feature matrix; The third training module is used to return to the step of inputting the plurality of training samples of the signal to be sensed into the feature extraction network of the preset model when the preset training stop condition is not met, until the preset training stop condition is met, so as to obtain a trained spectrum sensing model.

9. An electronic device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; when the processor calls the computer program instructions, the training method of the spectrum sensing model as described in any one of claims 1-6 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are called by the processor, the training method of the spectrum sensing model according to any one of claims 1 to 6 is implemented.

Images