Modulation signal classification method

Abstract

The invention relates to a modulation signal classification method, and the method employs a trained deep learning model to carry out the recognition of a large number of inputted unlabeled modulation signals and labeled modulation signals, and the trained deep learning model is obtained through employing the labeled data and unlabeled data. A deep learning model is subjected to multiple different initialization processing to obtain a first model and a second model which are different, and the first model and the second model are subjected to multiple times of iterative training to obtain the multi-label data based on loss values corresponding to multiple pieces of labeled data and loss values corresponding to multiple pieces of unlabeled data in each iteration process. And jointly adjusting parameters of the first model and the second model to be trained, and determining loss values corresponding to the plurality of pieces of unlabeled data by using the pseudo labels corresponding to the plurality of pieces of unlabeled data. When the method is applied to a modulation signal classification scene, the generalization ability and recognition precision of the model in a complex signal environment can be enhanced, the probability of misrecognition is reduced, the recognition accuracy and reliability are improved, and the wireless communication transmission quality and stability are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of modulation signal classification, and in particular relates to a modulation signal classification method. Background Art

[0002] The rapid development of information technology has made wireless communications more and more demanding, so the utilization and management of wireless spectrum has become crucial. Full wireless spectrum sensing technology is a technology based on software-defined radio that can monitor the wireless spectrum in real time and evaluate channel quality and signal strength. This technology needs to be combined with related technical tasks such as identification and classification to efficiently monitor and manage the spectrum. Among them, modulation signal classification, as one of the key tasks of wireless communication systems, has become an important means to ensure the quality and stability of wireless communication transmission. In addition, modulation signal classification can also be used in wireless communications in areas such as network security, spectrum allocation, and signal anti-interference.

[0003] However, actual communications involve various complex interferences and channel fading, which can have a serious impact on the structure and characteristics of the signal, making signal modulation recognition technology face great challenges. In order to improve recognition accuracy and reliability, expand the scope of application, and reduce computational complexity, researchers in various countries are actively researching and exploring, striving to achieve more efficient, intelligent, and reliable automatic modulation recognition technology in the future to meet people's strict requirements for communication quality and stability.

[0004] In the early stage, automatic modulation recognition mainly classified signal features manually based on decision boundaries, etc. The recognition effect was extremely dependent on the selected features, and manual feature selection was very complicated. When the channel is affected by noise, multipath fading, etc., it will greatly affect the recognition efficiency. Although the machine learning methods that emerged later have the advantages of high classification efficiency and high performance compared to traditional methods, the classification effect still depends on the experience of experts to a certain extent. In short, the existing modulation signal classification methods have low recognition accuracy, especially in complex signal scenarios with sparse recognition labels. The accuracy is greatly reduced, there is a risk of error propagation, and it cannot be used normally. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a modulation signal classification method. The technical problem to be solved by the present invention is achieved by the following technical solution:

[0006] The present invention provides a modulation signal classification method, the method comprising: obtaining multiple modulation signals; the multiple modulation signals include: M labeled modulation signals and N unlabeled modulation signals; M and N are both positive integers, and M is less than N; inputting the M labeled modulation signals and the N unlabeled modulation signals into a trained deep learning model for identification, and obtaining multiple identification results corresponding to one another; wherein the trained deep learning model is obtained by performing multiple iterative training on a first model and a second model using multiple labeled data and multiple unlabeled data; wherein the first model and the second model are obtained by performing different initialization processing on the deep learning model; in each iterative training process, based on the loss values ​​corresponding to the multiple labeled data and the loss values ​​corresponding to the multiple unlabeled data, jointly adjusting the parameters of the first model and the second model to be trained, and obtaining the first model and the second model for the next iterative training; wherein the loss values ​​corresponding to the multiple unlabeled data are determined based on the pseudo labels corresponding to the multiple unlabeled data, and the pseudo labels corresponding to the multiple unlabeled data are obtained by converting the prediction scores of the multiple unlabeled data in the first model and the second model to be trained that are greater than a preset confidence.

[0007] Compared with the prior art, the present invention has the following beneficial effects: in view of the low recognition accuracy of the existing modulation signal classification method, especially in complex signal scenarios with sparse recognition labels, the accuracy is greatly reduced, there is a risk of error propagation, and it cannot be used normally. The present invention provides a modulation signal classification method, which uses a trained deep learning model to identify a large number of input unlabeled modulation signals and labeled modulation signals, wherein the trained deep learning model is obtained by performing multiple iterative training on a first model and a second model with differences obtained by performing multiple different initialization processing on the deep learning model using labeled data and unlabeled data, and in each iteration process, based on the loss values ​​corresponding to multiple labeled data and the loss values ​​corresponding to multiple unlabeled data, the parameters of the first model and the second model to be trained are jointly adjusted, wherein the loss values ​​corresponding to the multiple unlabeled data are determined using pseudo labels corresponding to the multiple unlabeled data; this method can make the prediction of the data different without introducing additional disturbance to the data, and the model parameters are updated using the sum of the loss values ​​of the two, so as to realize the consistency constraint training of the two models. When this method is applied to modulation signal classification scenarios, especially in complex modulation signal classification scenarios with sparse identification labels, it can enhance the generalization ability and recognition accuracy of the model in complex signal environments and reduce the probability of misidentification; it can also improve the accuracy and reliability of identifying modulation signal types, and ensure the quality and stability of wireless communication transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1It is a schematic diagram of the training process of the deep learning model provided by the present invention;

[0009] Figure 2 is a schematic diagram of unit connection of a base model provided in an embodiment of the present invention;

[0010] Figure 3 This is an example diagram of deep learning model training provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0012] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0013] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification.

[0014] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art may understand and implement other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0015] A modulation signal classification method provided by the present invention is now described in conjunction with the accompanying drawings. The method includes: obtaining multiple modulation signals; the multiple modulation signals include: M labeled modulation signals and N unlabeled modulation signals; M and N are both positive integers, and M is less than N; inputting the M labeled modulation signals and the N unlabeled modulation signals into a trained deep learning model for identification, and obtaining multiple identification results corresponding to each other; wherein the trained deep learning model is obtained by using multiple labeled data and multiple unlabeled data to perform multiple iterative training on the first model and the second model; wherein the first model and the second model are obtained by performing different initialization processing on the deep learning model; in each iterative training process, based on the loss values ​​corresponding to the multiple labeled data and the loss values ​​corresponding to the multiple unlabeled data, the parameters of the first model and the second model to be trained are jointly adjusted to obtain the first model and the second model for the next iterative training; wherein the loss values ​​corresponding to the multiple unlabeled data are determined based on the pseudo labels corresponding to the multiple unlabeled data, and the pseudo labels corresponding to the multiple unlabeled data are obtained by converting the prediction scores of the multiple unlabeled data in the first model and the second model to be trained that are greater than the preset confidence.

[0016] Figure 1 Schematic diagram of the training process of the deep learning model provided by the present invention. Figure 1 As shown in the figure, the trained deep learning model is obtained through the following steps:

[0017] S1. Perform different initialization processes on the deep learning models to obtain a first model and a second model.

[0018] Here, when processing radio time series sample data, the deep learning model can achieve efficient and accurate modulation mode recognition by relying only on a small amount of labeled data and a large amount of unlabeled data. The deep learning model in the present invention is a base model, which includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a first timing layer, a second timing layer, a first fully connected layer, a second fully connected layer, a first random drop layer, a second random drop layer and a softmax layer. Figure 2 Schematic diagram of the connection of the base model units provided in the embodiment of the present invention. Figure 2As shown in the figure, the first convolutional layer, the second convolutional layer and the third convolutional layer, as the input layers of the base model, are all connected to the input signal; the second convolutional layer and the third convolutional layer are both connected to the fourth convolutional layer, the fourth convolutional layer and the first convolutional layer are both connected to the fifth convolutional layer, the fifth convolutional layer is connected to the first temporal layer and the second temporal layer in sequence, the second temporal layer is connected to the first fully connected layer and the first random drop layer in sequence, the first random drop layer is connected to the second fully connected layer and the second random drop layer in sequence, the second random drop layer is connected to the softmax layer, and the softmax layer is used as the output layer of the base model.

[0019] Here, each convolutional layer specifically includes: convolutional layer, batch normalization layer, ReLU activation layer. The number of convolution kernels in the convolutional layer is 50, 50, 50, 50, 100, the size of the convolution kernel is 2*8, 1*8, 1*8, 1*8, 2*5, the stride is 1, and the padding is 0; the hidden layer dimension of the first sequential layer is 128, the hidden layer dimension of the second sequential layer is 128, the output dimension of the first fully connected layer is 128, and the output dimension of the second fully connected layer is 11; the optimizer uses Adam.

[0020] Here, the initialization processing includes: Xavier parameter initialization processing and Kaiming parameter initialization processing; step one includes: performing Xavier parameter initialization processing on the deep learning model to obtain a first model; performing Kaiming parameter initialization processing on the deep learning model to obtain a second model.

[0021] After obtaining the first model and the second model, multiple iterations of the two models are started to obtain a trained deep learning model. Steps 2 to 5 are specific iterative training processes.

[0022] S2. In the i-th iteration, based on the acquired multiple labeled data, the i-th first recognition model and the i-th second recognition model, obtain the i-th first loss value and the i-th second loss value; wherein, the i-th first recognition model is obtained by performing i-1 iterations on the first model, and the i-th second recognition model is obtained by performing i-1 iterations on the second model; i is a positive integer.

[0023] Exemplarily, in the first iteration, the first first recognition model is the first model, and the second first recognition model is the second model.

[0024] Here, the labeled data can be a modulated signal carrying a label. Using the labeled data to obtain the corresponding loss value can play a role in supervised training. Specifically, step two includes: performing multiple signal flipping processes on multiple labeled data to obtain multiple second flipped data; inputting the multiple second flipped data into the i-th first recognition model and the i-th second recognition model respectively, and using the cross entropy loss function to calculate the i-th first loss value and the i-th second loss value.

[0025] Specifically, multiple signal flipping processes are performed on multiple labeled data to obtain multiple second flipped data, including: flipping multiple labeled data by 90 degrees, 180 degrees, and 270 degrees respectively to obtain multiple second flipped data, and the number of multiple second flipped data is four times the number of multiple labeled data. After obtaining the multiple second flipped data, in each iteration, the multiple second flipped data are directly input into the i-th first recognition model and the i-th second recognition model respectively, and the corresponding loss value is calculated using the cross entropy loss function. It should be noted that the data content of the multiple second flipped data does not change during each iteration.

[0026] Here, the i-th first loss value and the i-th second loss value can be expressed as:

[0027]

[0028]

[0029] Among them, L1 refers to the i-th first loss value, L2 refers to the i-th second loss value, refers to the prediction score of multiple second flipped data in the i-th first recognition model, It refers to the prediction score of multiple second flip data in the i-th second recognition model. Refers to the true one-hot encoding labels corresponding to multiple second flipped data.

[0030] S3. In the i-th iteration, based on the acquired multiple unlabeled data, the i-th first recognition model and the i-th second recognition model, obtain the i-th third loss value and the i-th fourth loss value.

[0031] Specifically, step three includes: in the i-th iteration, randomly flipping multiple unlabeled data to obtain first flipped data; adding noise processing to multiple unlabeled data to obtain first noise data; adding noise processing to multiple unlabeled data to obtain first noise data; based on the first flipped data, the first noise data, the i-th first recognition model and the i-th second recognition model, obtaining the i-th third loss value and the i-th fourth loss value.

[0032] Here, the first flip data is the weakly enhanced data Xunlabel_weak , the first noise data is the strong enhancement data X unlabel_strong .

[0033] Here, in the i-th iteration process, when performing random flipping on the multiple unlabeled data, the flipping angle includes but is not limited to: 60°, 120°, 180°, 240° and 300°. Each time the random flipping is performed, the first flipped data is obtained. And, when adding noise to the multiple unlabeled data, the noise data is Gaussian white noise data.

[0034] Here, based on the first flipped data, the first noise data, the i-th first recognition model and the i-th second recognition model, the i-th third loss value and the i-th fourth loss value are obtained, including: inputting the first noise data into the i-th first recognition model to obtain a one-to-one corresponding i-th first prediction score; determining the i-th first pseudo-label according to the first flipped data and the i-th second recognition model; based on the mean square error loss function, using the i-th first prediction score and the i-th first pseudo-label, obtaining the i-th third loss value; inputting the first noise data into the i-th second recognition model to obtain a one-to-one corresponding i-th second prediction score; determining the i-th second pseudo-label according to the first flipped data and the i-th first recognition model; based on the mean square error loss function, using the i-th second prediction score and the i-th second pseudo-label, obtaining the i-th fourth loss value.

[0035] Here, the i-th first pseudo label is determined according to the first flipped data and the i-th second recognition model, including: inputting the first flipped data into the i-th second recognition model to obtain a one-to-one corresponding i-th third prediction score; when the i-th third prediction score is greater than a preset confidence level, converting the i-th third prediction score into the i-th first pseudo label.

[0036] Here, the i-th second pseudo-label is determined according to the first flipped data and the i-th first recognition model, including: inputting the first flipped data into the i-th first recognition model to obtain a one-to-one corresponding i-th fourth prediction score; when the i-th fourth prediction score is greater than a preset confidence level, converting the i-th fourth prediction score into the i-th second pseudo-label.

[0037] By introducing confidence judgment, that is, using the score output by the model, the matching degree between the prediction score corresponding to the first noise data and the pseudo label corresponding to the first flipped data is estimated, so as to convert the prediction score greater than the preset confidence into a high-quality pseudo label. unlabel_weak And strong enhancement data X unlabel_strongImposing consistency, that is, applying mean square error loss to the two, can make the corresponding outputs of the i-th first recognition model and the i-th second recognition model as similar as possible when they input the same data, thereby efficiently utilizing limited labeled data and a large amount of unlabeled data, enhancing the generalization ability and recognition accuracy of the deep learning model in a complex signal environment, reducing the risk of error propagation, and reducing dependence on the initial model.

[0038] It should be noted that when the prediction score corresponding to the first flipped data is less than or equal to the preset confidence, the corresponding pseudo label is not generated, and the corresponding loss value is not calculated. Specifically, assuming that in the i-th iteration, when the i-th third prediction score is less than or equal to the preset confidence, the corresponding i-th third loss value is not calculated; or, when the i-th fourth prediction score is less than or equal to the preset confidence, the corresponding i-th fourth loss value is not calculated.

[0039] Here, the i-th third loss value and the i-th fourth loss value can be expressed as:

[0040]

[0041]

[0042] Among them, L3 refers to the third loss value of the i-th order, It refers to the i-th first prediction score corresponding to the first noise data in the i-th first recognition model. refers to the i-th second prediction score corresponding to the first flipped data in the i-th second recognition model; L4 refers to the i-th fourth loss value, It refers to the i-th third prediction score corresponding to the first noise data in the i-th second recognition model. It refers to the i-th fourth prediction score corresponding to the first flipped data in the i-th first recognition model.

[0043] S4. Determine whether i is less than the preset number of iterations; if so, execute step five; if not, select the model with the highest recognition accuracy between the i-th first recognition model and the i-th second recognition model as the trained deep learning model.

[0044] Here, selecting the model with the highest recognition accuracy rate from the i-th first recognition model and the i-th second recognition model as the trained deep learning model includes: obtaining test set data; inputting the test set data into the i-th first recognition model and the i-th second recognition model respectively to obtain the first recognition result and the second recognition result; judging whether the accuracy rate of the first recognition result is greater than the accuracy rate of the second recognition result; if so, taking the i-th first recognition model as the trained deep learning model; if not, taking the i-th second recognition model as the trained deep learning model. Exemplarily, the preset number of iterations is 200.

[0045] In a possible implementation, whether the model has converged can also be determined by determining whether the loss values ​​output by the i-th first recognition model and the i-th second recognition model no longer decrease.

[0046] S5. Based on the i-th total loss value, update the parameter values ​​in the i-th first recognition model and the i-th second recognition model to generate the i+1-th first recognition model and the i+1-th second recognition model, let i=i+1 in steps 2 to 4, and use the i+1-th first recognition model and the i+1-th second recognition model to perform the i+1-th iteration; the i-th total loss value is determined using the i-th first loss value, the i-th second loss value, the i-th third loss value and the i-th fourth loss value.

[0047] Here, the i-th total loss value is the sum of the i-th first loss value, the i-th second loss value, the i-th third loss value, and the i-th fourth loss value. The i-th total loss value can be expressed as: L 总 =L1+L2+L3+L4.

[0048] It should be noted that, in the case where the i-th third loss value L3 and / or the i-th fourth loss value L4 is missing in step 3, the i-th total loss value can be expressed as: L 总 =L1+L2, or, L 总 =L1+L2+L3, or, L 总 =L1+L2+L4.

[0049] Here, when i is less than the preset number of iterations, it means that the model has not reached the convergence state. At this time, based on the backpropagation algorithm, the i-th total loss value is used to update the model parameters and execute the i+1-th iteration until i is greater than or equal to the preset number of iterations.

[0050] Figure 3 is an example diagram of deep learning model training provided by an embodiment of the present invention. Figure 3 As shown in the figure, the RML2016.10a dataset is used to train the preprocessed base model, where 10% of the data is selected as labeled training data, 70% as unlabeled training data, 10% as validation set, and 10% as test set. The labeled training data is input into Model1 and Model2 respectively to obtain two cross entropy loss values; Model1 is used to refer to the first recognition model corresponding to multiple iterations, and Model2 is used to refer to the second recognition model corresponding to multiple iterations. It should be noted that in each iteration, the model parameters of Model1 and Model2 are different; the unlabeled training data is randomly flipped and noise processed respectively to obtain weakly enhanced data X unlabel_weakAnd strong enhancement data X unlabel_strong , using the weakly enhanced data X in Model1 unlabel weak The corresponding labels greater than the threshold, and the strong enhancement data X in Model2 unlabel_strong Corresponding prediction scores, calculate a mean square error loss; use the weakly enhanced data X in Model2 unlabel_weak The corresponding labels greater than the threshold, and the strong enhancement data X in Model1 unlabel_strong Corresponding prediction scores, calculate another mean square error loss; determine whether the current number of iterations is less than the preset number of iterations. If it is greater than or equal to the preset number of iterations, stop the iteration, and use the test set and validation set to determine the current Model1 and Model2 with the highest classification accuracy as the trained deep learning model. If it is less than the preset number of iterations, use two cross entropy loss values ​​and two mean square error losses to update the model parameters, and continue to use labeled training data and unlabeled training data for training until a trained deep learning model is reached.

[0051] In view of the problem that the existing modulation signal classification methods have low recognition accuracy, especially in complex signal scenarios with sparse recognition labels, the accuracy is greatly reduced, there is a risk of error propagation, and the method cannot be used normally, the present invention provides a modulation signal classification method, which uses a trained deep learning model to identify a large number of input unlabeled modulation signals and labeled modulation signals, wherein the trained deep learning model is obtained by performing multiple iterative training on a first model and a second model with differences obtained by performing multiple different initialization processing on the deep learning model using labeled data and unlabeled data, and in each iteration process, based on loss values ​​corresponding to multiple labeled data and loss values ​​corresponding to multiple unlabeled data, the parameters of the first model and the second model to be trained are jointly adjusted, wherein the loss values ​​corresponding to the multiple unlabeled data are determined using pseudo labels corresponding to the multiple unlabeled data; this method can make the prediction of the data different without introducing additional disturbance to the data, and the model parameters are updated using the sum of the loss values ​​of the two, so as to realize the consistency constraint training of the two models. When this method is applied to modulation signal classification scenarios, especially in complex modulation signal classification scenarios with sparse identification labels, it can enhance the generalization ability and recognition accuracy of the model in complex signal environments and reduce the probability of misidentification; it can also improve the accuracy and reliability of identifying modulation signal types, and ensure the quality and stability of wireless communication transmission.

[0052] In order to verify the reliability of the deep learning model trained by the present invention for the classification of modulation signals, simulation software is used to perform simulation experiments. In this experiment, the data set uses the RML2016.10a data set for training, which contains 11 different types of modulation signals, 8 of which are digital modulation signals, and these 8 digital modulation signals are: (8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK), and 3 are analog modulation signals (AM-DSB, AM-SSB, WBFM). Each modulation signal contains 20 (-20:2:18dB) signal-to-noise ratios, and a single modulation signal has 1000 data for a single signal-to-noise ratio, for a total of 220,000 data. Among them, 10% of the data is selected as labeled training data, 70% as unlabeled training data, 10% as a validation set, and 10% as a test set.

[0053] The deep learning model trained by the present invention and the conventional base model are simulated, and the accuracy is shown in Table 1. It should be noted that the deep learning model trained by the present invention and the conventional base model have the same architecture, and only the model training method is different. It can be seen from the accuracy in Table 1 that the deep learning model trained by the present invention has a higher recognition accuracy than the conventional base model in complex signal classification scenarios with sparse recognition labels, and significantly reduces the risk of error propagation.

[0054] Table 1

[0055] Model Name Accuracy Conventional base model 1 56.79％ Conventional base model 2 57.75％ The base model obtained by training in the present invention 59.52％

[0056] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A modulation signal classification method, characterized in that: The method comprises: Acquire multiple modulation signals; the multiple modulation signals include: M tagged modulation signals and N untagged modulation signals; M and N are both positive integers, and M is less than N; Input the M labeled modulation signals and the N unlabeled modulation signals into a trained deep learning model for recognition, and obtain a plurality of one-to-one corresponding recognition results; wherein the trained deep learning model is obtained by performing multiple iterative training on the first model and the second model using multiple labeled data and multiple unlabeled data; wherein the first model and the second model are obtained by performing different initialization processing on the deep learning model; in each iterative training process, based on the loss values ​​corresponding to the multiple labeled data and the loss values ​​corresponding to the multiple unlabeled data, the parameters of the first model and the second model to be trained are jointly adjusted to obtain the first model and the second model for the next iterative training; wherein the loss values ​​corresponding to the multiple unlabeled data are determined based on the pseudo-labels corresponding to the multiple unlabeled data, and the pseudo-labels corresponding to the multiple unlabeled data are obtained by converting the prediction scores of the multiple unlabeled data in the first model and the second model to be trained that are greater than a preset confidence.

2. The modulation signal classification method according to claim 1, characterized in that: The trained deep learning model is obtained by training in the following steps: S1. Perform different initialization processes on the deep learning models to obtain the first model and the second model; S2. In the i-th iteration, based on the acquired multiple labeled data, the i-th first recognition model and the i-th second recognition model, obtain the i-th first loss value and the i-th second loss value; wherein the i-th first recognition model is obtained by performing i-1 iterations on the first model, and the i-th second recognition model is obtained by performing i-1 iterations on the second model; i is a positive integer; S3. In the i-th iteration, based on the acquired multiple unlabeled data, the i-th first recognition model and the i-th second recognition model, obtain an i-th third loss value and an i-th fourth loss value; S4, determine whether i is less than a preset number of iterations; if so, execute step five; if not, select the model with the highest recognition accuracy among the i-th first recognition model and the i-th second recognition model as the trained deep learning model; S5. Based on the i-th total loss value, update the parameter values ​​in the i-th first recognition model and the i-th second recognition model to generate the i+1-th first recognition model and the i+1-th second recognition model, let i=i+1 in steps 2 to 4, and use the i+1-th first recognition model and the i+1-th second recognition model to perform the i+1-th iteration; the i-th total loss value is determined using the i-th first loss value, the i-th second loss value, the i-th third loss value and the i-th fourth loss value.

3. The modulation signal classification method according to claim 2, characterized in that: The step three comprises: In the i-th iteration, the plurality of unlabeled data are randomly flipped to obtain first flipped data; Adding noise processing to the plurality of unlabeled data to obtain first noise data; Based on the first flip data, the first noise data, the i-th first recognition model and the i-th second recognition model, the i-th third loss value and the i-th fourth loss value are obtained.

4. The modulation signal classification method according to claim 3, characterized in that: The obtaining the i-th third loss value and the i-th fourth loss value based on the first flip data, the first noise data, the i-th first recognition model and the i-th second recognition model comprises: Inputting the first noise data into the i-th first recognition model to obtain a one-to-one corresponding i-th first prediction score; Determine an i-th first pseudo label according to the first flipped data and the i-th second recognition model; Based on a mean square error loss function, using the i-th first prediction score and the i-th first pseudo label, obtaining the i-th third loss value; Inputting the first noise data into the i-th second recognition model to obtain a one-to-one corresponding i-th second prediction score; Determine an i-th second pseudo label according to the first flipped data and the i-th first recognition model; Based on the mean square error loss function, the i-th fourth loss value is obtained using the i-th second prediction score and the i-th second pseudo label.

5. The modulation signal classification method according to claim 4, characterized in that: The step of determining the i-th first pseudo label according to the first flipped data and the i-th second recognition model includes: Inputting the first flipped data into the i-th second recognition model to obtain a one-to-one corresponding i-th third prediction score; When the i-th third prediction score is greater than the preset confidence, the i-th third prediction score is converted into the i-th first pseudo label.

6. The modulation signal classification method according to claim 4, characterized in that: The determining the i-th second pseudo label according to the first flipped data and the i-th first recognition model comprises: Inputting the first flipped data into the i-th first recognition model to obtain a one-to-one corresponding i-th fourth prediction score; When the i-th fourth prediction score is greater than the preset confidence, the i-th fourth prediction score is converted into the i-th second pseudo label.

7. The modulation signal classification method according to claim 2, characterized in that: The second step comprises: Performing multiple signal inversion processes on the multiple labeled data to obtain multiple second inversion data; The multiple second flipped data are respectively input into the i-th first recognition model and the i-th second recognition model, and the i-th first loss value and the i-th second loss value are calculated using the cross entropy loss function.

8. The modulation signal classification method according to claim 2, characterized in that: The step of selecting the model with the highest recognition accuracy rate from among the i-th first recognition model and the i-th second recognition model as the trained deep learning model includes: Get the test set data; Inputting the test set data into the i-th first recognition model and the i-th second recognition model respectively to obtain a first recognition result and a second recognition result; Determining whether the accuracy of the first recognition result is greater than the accuracy of the second recognition result; If so, use the i-th first recognition model as the trained deep learning model; If not, the i-th second recognition model is used as the trained deep learning model.

9. The modulation signal classification method according to claim 2, characterized in that: The initialization process includes: xavier parameter initialization process and kaiming parameter initialization process; the step 1 includes: Performing Xavier parameter initialization processing on the deep learning model to obtain the first model; Perform kaiming parameter initialization processing on the deep learning model to obtain the second model.

10. The modulation signal classification method according to claim 1, characterized in that: The deep learning model is a base model, which includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a first sequential layer, a second sequential layer, a first fully connected layer, a second fully connected layer, a first random drop layer, a second random drop layer and a softmax layer; The first convolutional layer, the second convolutional layer and the third convolutional layer, as the input layers of the base model, are all connected to the input signal; the second convolutional layer and the third convolutional layer are both connected to the fourth convolutional layer, the fourth convolutional layer and the first convolutional layer are both connected to the fifth convolutional layer, the fifth convolutional layer is connected to the first temporal layer and the second temporal layer in sequence, the second temporal layer is connected to the first fully connected layer and the first random drop layer in sequence, the first random drop layer is connected to the second fully connected layer and the second random drop layer in sequence, the second random drop layer is connected to the softmax layer, and the softmax layer is used as the output layer of the base model.

Images