A modulation mode identification method, system and related components

Abstract

The present application discloses a modulation mode recognition method, system and related components, the method includes: obtaining data to be recognized; inputting the data to be recognized into a recognition network to obtain a recognition result; wherein the training process of the recognition network includes: initializing the network parameters of the recognition network; obtaining non-balanced training samples corresponding to different modulation modes and with different prior probabilities; looping through the following steps until the preset iteration requirements are met, and determining the recognition network according to the network parameters: inputting the non-balanced training samples into the recognition network and obtaining an output feature vector; calculating the weighted loss function of the output feature vector and the non-balanced training samples; optimizing the network parameters with the purpose of minimizing the weighted loss function. The recognition network used in the present application is more suitable for the recognition network in the actual environment of non-balanced probability, and has excellent classification performance and higher recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the field of signal modulation for wireless communications, and in particular to a modulation mode identification method, system and related components. Background Art

[0002] At present, the rapid development of communication technology and the growing demand for communication services require wireless communication technology and equipment to have the ability to quickly perceive the complex electromagnetic environment around them, so as to achieve efficient use of the spectrum. Most of the existing spectrum sensing technologies use the statistical characteristics of the signal to perceive the spectrum status, and can only determine whether there is a signal occupying the channel, but cannot distinguish between useful signals and illegal interference signals, and cannot monitor and manage interference signals. Spectrum analysis is an analytical method that determines the type of modulation mode and signal format of the spectrum sensed signal based on spectrum sensing.

[0003] Signal modulation recognition is a key component of spectrum analysis, which refers to determining the modulation mode of the received signal when the signal information is unknown. Traditional modulation recognition methods are divided into two categories: modulation recognition methods based on likelihood ratios and modulation recognition methods based on statistical features. Both methods are model-driven modulation recognition methods, and the probability of misclassification is high when facing complex new environments.

[0004] A new modulation recognition method, modulation recognition method based on deep neural network, has attracted wide attention due to its robustness to changes in complex communication environment and tolerance to noise interference signals. However, the existing modulation recognition method based on deep neural network, as a data-driven classifier, has the premise that the prior probability of each modulation mode is the same, which is quite different from the situation in actual environment where the probability of occurrence of different modulation modes is uneven. The accuracy of the current modulation recognition method based on deep neural network is difficult to guarantee.

[0005] Therefore, how to provide a solution to the above technical problems is a problem that those skilled in the art need to solve. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a modulation mode identification method, system and related components with non-balanced prior probability and high identification accuracy. The specific scheme is as follows:

[0007] A modulation mode identification method, comprising:

[0008] Obtaining data to be identified;

[0009] Inputting the data to be identified into the identification network to obtain the identification result;

[0010] The training process of the recognition network includes:

[0011] Initialize the network parameters of the recognition network;

[0012] Obtaining non-balanced training samples corresponding to different modulation modes and having different prior probabilities;

[0013] The following steps are executed in a loop:

[0014] Inputting the unbalanced training sample into the recognition network and obtaining an output feature vector;

[0015] Calculating a weighted loss function between the output feature vector and the unbalanced training sample, wherein the weighted loss function is specifically a distance between a predicted value of the output feature vector and a true label of the unbalanced training sample;

[0016] Optimizing the network parameters with the goal of minimizing the weighted loss function;

[0017] When the preset iteration requirement is met, the loop is exited, and the identification network is determined according to the network parameters.

[0018] Preferably, the process of calculating the weighted loss function of the output feature vector and the unbalanced training sample includes:

[0019] According to the first formula, the weighted loss function of the output feature vector and the unbalanced training sample is calculated, and the first formula is:

[0020]

[0021] Where UWL is the weighted loss function, the correction coefficient is the predicted value of sample m when category k in the output feature vector, y km is the true label of sample m when category k is in the unbalanced training sample, ε is a preset smoothing factor, Mb is the total number of samples of a certain category of the unbalanced training samples, and K is the total number of categories of the unbalanced training samples.

[0022] Preferably, the process of optimizing the network parameters for the purpose of minimizing the weighted loss function includes:

[0023] With the purpose of minimizing the weighted loss function, the Adam optimization algorithm is used to optimize the network parameters.

[0024] Preferably, the preset iteration requirement is specifically:

[0025] The number of iterations reaches the preset number;

[0026] Or, the prediction accuracy reaches a preset percentage;

[0027] Or, the weighted loss function reaches a preset function value.

[0028] Preferably, the recognition network comprises a representation module, a classification module and an output module which are connected in sequence.

[0029] Preferably, the process of inputting the unbalanced training sample into the recognition network and obtaining the output feature vector includes:

[0030] The unbalanced training samples are input into the recognition network to obtain a normalized output feature vector.

[0031] Accordingly, the present application also discloses a modulation mode identification system, comprising:

[0032] An acquisition module, used for acquiring data to be identified;

[0033] An action module, used for inputting the data to be identified into the identification network to obtain the identification result;

[0034] The recognition system further includes a recognition network training module, and the recognition network training module includes:

[0035] An initialization unit, used for initializing network parameters of the identification network;

[0036] A sample acquisition unit, used to acquire non-balanced training samples corresponding to different modulation modes and having different prior probabilities;

[0037] A training unit is used to cyclically execute the following steps: input the unbalanced training sample into the recognition network and obtain an output feature vector; calculate a weighted loss function between the output feature vector and the unbalanced training sample, wherein the weighted loss function is specifically the distance between the predicted value of the output feature vector and the true label of the unbalanced training sample; optimize the network parameters with the purpose of minimizing the weighted loss function; exit the loop when the preset iteration requirement is met, and determine the recognition network according to the network parameters.

[0038] Accordingly, the present application also discloses a modulation mode identification device, comprising:

[0039] Memory for storing computer programs;

[0040] A processor is used to implement the steps of the modulation mode identification method as described in any one of the above items when executing the computer program.

[0041] Correspondingly, the present application also discloses a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the modulation mode identification method as described in any one of the above items are implemented.

[0042] The present application discloses a modulation mode recognition method, including: obtaining data to be recognized; inputting the data to be recognized into a recognition network to obtain a recognition result; wherein the training process of the recognition network includes: initializing the network parameters of the recognition network; obtaining non-balanced training samples corresponding to different modulation modes and with different prior probabilities; looping the following steps: inputting the non-balanced training samples into the recognition network and obtaining an output feature vector; calculating the weighted loss function of the output feature vector and the non-balanced training sample, wherein the weighted loss function is specifically the distance between the predicted value of the output feature vector and the true label of the non-balanced training sample; optimizing the network parameters for the purpose of minimizing the weighted loss function; when the preset iteration requirements are met, exiting the loop, and determining the recognition network according to the network parameters. The recognition network used in the present application is a recognition network that is more adaptable to the actual environment of non-balanced probabilities by setting non-balanced training samples corresponding to different modulation modes and different prior probabilities, and adopting a weighted loss function to train and feedback-adjust the recognition network, and finally obtaining a recognition network that is more adaptable to the actual environment of non-balanced probabilities, and has excellent classification performance and higher recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0044] Figure 1 A flowchart of the steps of a training process of a recognition network in an embodiment of the present invention;

[0045] Figure 2 A structural distribution diagram of a modulation mode identification network in an embodiment of the present invention;

[0046] Figure 3 The structure distribution diagram of a modulation mode identification system in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] A new modulation recognition method, modulation recognition method based on deep neural network, has attracted wide attention due to its robustness to changes in complex communication environment and tolerance to noise interference signals. However, the existing modulation recognition method based on deep neural network, as a data-driven classifier, has the premise that the prior probability of each modulation mode is the same, which is quite different from the situation in actual environment where the probability of occurrence of different modulation modes is uneven. The accuracy of the current modulation recognition method based on deep neural network is difficult to guarantee.

[0049] This application sets unbalanced training samples corresponding to different modulation modes and different prior probabilities, and uses a weighted loss function to train and feedback-adjust the recognition network, ultimately obtaining a recognition network that is more adaptable to the actual environment of unbalanced probabilities, with excellent classification performance and higher recognition accuracy.

[0050] The embodiment of the present invention discloses a modulation mode identification method, comprising:

[0051] Obtaining data to be identified;

[0052] Inputting the data to be identified into the identification network to obtain the identification result;

[0053] The training process of the recognition network is shown in Figure 1 As shown, including:

[0054] S1: Initialize the network parameters of the recognition network;

[0055] S2: Obtaining non-balanced training samples corresponding to different modulation modes and with different prior probabilities;

[0056] The following steps are executed in a loop:

[0057] S3: Input the unbalanced training sample into the recognition network and obtain the output feature vector;

[0058] S4: Calculate the weighted loss function of the output feature vector and the unbalanced training sample, where the weighted loss function is specifically the distance between the predicted value of the output feature vector and the true label of the unbalanced training sample;

[0059] S5: Optimize network parameters with the goal of minimizing the weighted loss function;

[0060] S6: When the preset iteration requirements are met, exit the loop and determine the identification network based on the network parameters.

[0061] Specifically, the recognition network in this embodiment is generally a global pooling-based convolutional neural network (GPCN), which usually includes a representation module, a classification module and an output module connected in sequence, such as Figure 2 As shown in the structure of . Usually, the input signal is input to the recognition network, and first passes through the convolution layer and the strided convolution layer in the representation module to extract shallow features, then passes through the global average pooling layer in the classification module to extract and classify the features in depth, and finally passes through the flattening layer and the Softmax activation function layer of the output module to finally realize the classification of the modulation recognition signal. Among them, the strided convolution layer is optimized from the pooling layer to avoid the discarding of unknown information by the pooling layer; the global average pooling layer is optimized from the fully connected layer to avoid the complex calculation amount brought by the fully connected layer. The strided convolution layer and the global pooling layer can avoid the loss of signal features and further reduce the computational complexity of the recognition network. In addition, in order to balance the input and output of each layer, the output of each layer can be batch normalized, so step S3 is specifically reflected in inputting the unbalanced training sample into the recognition network to obtain a normalized output feature vector.

[0062] Furthermore, the unbalanced training samples obtained in step S2 are in the form of sample sets, which can be specifically represented as follows:

[0063] (Ω x ,y)=(x (1) ,y (1) ),(x (2) ,y (2) ),…,(x (m) ,y (m) )…,(x (M) ,y (M) );

[0064] Among them, Ω x and y are the sample set and the true label set marked with true labels, respectively. (m) represents the mth sample, and M represents the number of training samples. is a one-hot vector, which can be expressed as:

[0065]

[0066] Where k = 1, 2, ..., K, K is the total number of modulation methods in the training sample set, k * Represents the true label of the mth sample.

[0067] Before step S3, the network parameters of the recognition network, ie, the weights, are initialized, generally randomly initialized to values ​​close to 0.

[0068] Then the contents of steps S3-S5 are executed in a loop:

[0069] Among them, step S3 is a forward propagation algorithm, which starts from inputting the unbalanced training sample and propagates forward, passes through each module in turn, and finally outputs a normalized output vector feature:

[0070]

[0071] in, In the assumption that H k The nonlinear transformation under Represents the extracted deep representation features under this assumption.

[0072] Based on this, the conditional probability is expressed as H k : and

[0073] Specifically, the process of calculating the weighted loss function of the output feature vector and the unbalanced training sample in step S4 includes:

[0074] According to the first formula, the weighted loss function of the output feature vector and the unbalanced training sample is calculated. The first formula is:

[0075]

[0076] Among them, UWL is a weighted loss function, which can also be called an unbalanced weighted loss function (Unbalanced Weighted Loss), and the correction coefficient is the predicted value of sample m when category k is in the output feature vector, y km is the true label of sample m when category k is in the unbalanced training sample, ε is the preset smoothing factor, Mb is the total number of samples in a certain category of the unbalanced training samples, and K is the total number of categories of the unbalanced training samples.

[0077] It is understandable that the weighted loss function in this embodiment is actually designed from the perspective of the F1 score in statistics to balance the contribution of the imbalanced categories to the objective function. It directly acts on the imbalanced weighted coefficient to improve the overall classification accuracy of the model.

[0078] Specifically, the unbalanced weighted coefficient (UWC) is used to measure the similarity between the predicted value and the label, which is expressed as:

[0079]

[0080] in, is a batch of predicted probabilities, y is the true label, and the coefficient 2 is repeated calculation of the denominator and y Consider the setting.

[0081] Furthermore, the unbalanced weighted loss function corresponding to the unbalanced weighted coefficient is expressed as:

[0082]

[0083] To avoid When both y and y are 0, the denominator is 0. Adding the smoothing factor ε, the loss function becomes:

[0084]

[0085] After correcting the contributions of different modulation modes, the weighted loss function cited in this embodiment is finally obtained, that is, the first formula.

[0086] Furthermore, step S5 actually executes the back propagation algorithm to achieve the effect of optimizing the network parameters. Specifically, the weighted loss function is minimized, and the partial derivative of the weighted loss function is obtained, and the Adam optimization algorithm is used to finally optimize the network parameters.

[0087] Furthermore, the preset iteration requirement in step S6 is specifically:

[0088] The number of iterations reaches the preset number;

[0089] Or, the prediction accuracy reaches a preset percentage;

[0090] Or, the weighted loss function reaches a preset function value.

[0091] Usually, as long as the number of iterations is reached or the loss function does not change within a certain range, it can be considered that the optimal network parameters of the recognition network have been obtained:

[0092] Further, after the recognition network is determined, the recognition network can be tested by testing the test data set (Ω' x ,y)

[0093] Input into the current recognition network, execute the forward propagation algorithm, and then obtain the conditional probability of each sample under different modulation modes, and select the modulation mode H corresponding to the maximum probability k As the category label of the sample, that is:

[0094]

[0095] Finally, the predicted category label is compared with the true label to calculate the prediction accuracy.

[0096] The prediction accuracy can be used to evaluate the accuracy of the current recognition network. If the prediction accuracy is low, the network parameters of the test network can be corrected again based on the test results.

[0097] The present application discloses a modulation mode recognition method, including: obtaining data to be recognized; inputting the data to be recognized into a recognition network to obtain a recognition result; wherein the training process of the recognition network includes: initializing the network parameters of the recognition network; obtaining non-balanced training samples corresponding to different modulation modes and with different prior probabilities; looping the following steps: inputting the non-balanced training samples into the recognition network and obtaining an output feature vector; calculating the weighted loss function of the output feature vector and the non-balanced training sample, wherein the weighted loss function is specifically the distance between the predicted value of the output feature vector and the true label of the non-balanced training sample; optimizing the network parameters for the purpose of minimizing the weighted loss function; when the preset iteration requirements are met, exiting the loop, and determining the recognition network according to the network parameters. The recognition network used in the present application is a recognition network that is more adaptable to the actual environment of non-balanced probabilities by setting non-balanced training samples corresponding to different modulation modes and different prior probabilities, and adopting a weighted loss function to train and feedback-adjust the recognition network, and finally obtaining a recognition network that is more adaptable to the actual environment of non-balanced probabilities, and has excellent classification performance and higher recognition accuracy.

[0098] Correspondingly, the present application also discloses a modulation mode identification system, see Figure 3 As shown, including:

[0099] Acquisition module 1, used to acquire data to be identified;

[0100] Action module 2, used for inputting the data to be identified into the identification network to obtain the identification result;

[0101] The recognition system further includes a recognition network training module 3, wherein the recognition network training module 3 includes:

[0102] An initialization unit 31, used to initialize network parameters of the identification network;

[0103] A sample acquisition unit 32, used to acquire non-balanced training samples corresponding to different modulation modes and having different prior probabilities;

[0104] The training unit 33 is used to cyclically execute the following steps: input the unbalanced training sample into the recognition network and obtain an output feature vector; calculate a weighted loss function between the output feature vector and the unbalanced training sample, wherein the weighted loss function is specifically the distance between the predicted value of the output feature vector and the true label of the unbalanced training sample; optimize the network parameters with the purpose of minimizing the weighted loss function; exit the loop when the preset iteration requirements are met, and determine the recognition network according to the network parameters.

[0105] This application sets unbalanced training samples corresponding to different modulation modes and different prior probabilities, and uses a weighted loss function to train and feedback-adjust the recognition network, ultimately obtaining a recognition network that is more adaptable to the actual environment of unbalanced probabilities, with excellent classification performance and higher recognition accuracy.

[0106] In some specific embodiments, the process of the training unit 33 calculating the weighted loss function of the output feature vector and the unbalanced training sample includes:

[0107] According to the first formula, the weighted loss function of the output feature vector and the unbalanced training sample is calculated, and the first formula is:

[0108]

[0109] Where UWL is the weighted loss function, the correction coefficient is the predicted value of sample m when category k in the output feature vector, y km is the true label of sample m when category k is in the unbalanced training sample, ε is a preset smoothing factor, Mb is the total number of samples of a certain category of the unbalanced training samples, and K is the total number of categories of the unbalanced training samples.

[0110] In some specific embodiments, the training unit 33 optimizes the network parameters for the purpose of minimizing the weighted loss function, including:

[0111] With the purpose of minimizing the weighted loss function, the Adam optimization algorithm is used to optimize the network parameters.

[0112] In some specific embodiments, the preset iteration requirement in the training unit 33 is specifically:

[0113] The number of iterations reaches the preset number;

[0114] Or, the prediction accuracy reaches a preset percentage;

[0115] Or, the weighted loss function reaches a preset function value.

[0116] In some specific embodiments, the recognition network includes a representation module, a classification module and an output module connected in sequence.

[0117] In some specific embodiments, the training unit 33 inputs the unbalanced training sample into the recognition network and obtains the output feature vector, including:

[0118] The unbalanced training samples are input into the recognition network to obtain a normalized output feature vector.

[0119] Accordingly, the embodiment of the present application further discloses a modulation mode identification device, including:

[0120] Memory for storing computer programs;

[0121] A processor is used to implement the steps of the modulation mode identification method as described in any one of the above items when executing the computer program.

[0122] Correspondingly, an embodiment of the present application further discloses a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the modulation mode identification method as described in any one of the above items are implemented.

[0123] For details about the modulation mode identification method, please refer to the relevant description in the above embodiment, which will not be repeated here.

[0124] Among them, the modulation mode identification device and the readable storage medium in this embodiment have the same technical effects as the modulation mode identification method in the above embodiment, and will not be repeated here.

[0125] Finally, 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 "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0126] The above is a detailed introduction to a modulation mode identification method, system and related components provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for a general technician in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A modulation mode identification method, characterized in that: include: Obtaining data to be identified; Inputting the data to be identified into the identification network to obtain the identification result; The training process of the recognition network includes: Initialize the network parameters of the recognition network; Obtaining non-balanced training samples corresponding to different modulation modes and having different prior probabilities; The following steps are executed in a loop: Inputting the unbalanced training sample into the recognition network and obtaining an output feature vector; Calculating a weighted loss function between the output feature vector and the unbalanced training sample, wherein the weighted loss function is specifically a distance between a predicted value of the output feature vector and a true label of the unbalanced training sample; Optimizing the network parameters with the goal of minimizing the weighted loss function; When the preset iteration requirement is met, exit the loop and determine the identification network according to the network parameters; The process of calculating the weighted loss function of the output feature vector and the unbalanced training sample includes: According to the first formula, the weighted loss function of the output feature vector and the unbalanced training sample is calculated, and the first formula is: ； Where UWL is the weighted loss function, the correction coefficient ; is the predicted value of sample m when category k in the output feature vector, is the true label of sample m when category k is in the unbalanced training sample, is the preset smoothing factor, is the total number of samples of a certain category of the unbalanced training samples, and K is the total number of categories of the unbalanced training samples.

2. The modulation mode identification method according to claim 1, characterized in that: The process of optimizing the network parameters for the purpose of minimizing the weighted loss function includes: With the purpose of minimizing the weighted loss function, the Adam optimization algorithm is used to optimize the network parameters.

3. The modulation mode identification method according to claim 1, characterized in that: The preset iteration requirements are specifically: The number of iterations reaches the preset number; Or, the prediction accuracy reaches a preset percentage; Or, the weighted loss function reaches a preset function value.

4. The modulation mode identification method according to any one of claims 1 to 3, characterized in that: The recognition network includes a representation module, a classification module and an output module which are connected in sequence.

5. The modulation mode identification method according to claim 4, characterized in that: The process of inputting the unbalanced training sample into the recognition network and obtaining an output feature vector includes: The unbalanced training samples are input into the recognition network to obtain a normalized output feature vector.

6. A modulation mode identification system, characterized in that: include: An acquisition module, used for acquiring data to be identified; An action module, used for inputting the data to be identified into the identification network to obtain the identification result; The recognition system further includes a recognition network training module, and the recognition network training module includes: An initialization unit, used for initializing network parameters of the identification network; A sample acquisition unit, used to acquire non-balanced training samples corresponding to different modulation modes and having different prior probabilities; A training unit is used to cyclically execute the following steps: input the unbalanced training sample into the recognition network and obtain an output feature vector; calculate a weighted loss function between the output feature vector and the unbalanced training sample, wherein the weighted loss function is specifically the distance between the predicted value of the output feature vector and the true label of the unbalanced training sample; optimize the network parameters with the purpose of minimizing the weighted loss function; exit the loop when the preset iteration requirement is met, and determine the recognition network according to the network parameters; The process of calculating the weighted loss function of the output feature vector and the unbalanced training sample includes: According to the first formula, the weighted loss function of the output feature vector and the unbalanced training sample is calculated, and the first formula is: ； Where UWL is the weighted loss function, the correction coefficient ; is the predicted value of sample m when category k in the output feature vector, is the true label of sample m when category k is in the unbalanced training sample, is the preset smoothing factor, is the total number of samples of a certain category of the unbalanced training samples, and K is the total number of categories of the unbalanced training samples.

7. A modulation mode identification device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the modulation mode identification method as claimed in any one of claims 1 to 5 when executing the computer program.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the modulation mode identification method according to any one of claims 1 to 5 are implemented.

Images