Dual-channel-based signal comparison prediction semi-supervised automatic modulation identification method

Abstract

The invention discloses a signal comparison prediction semi-supervised automatic modulation identification method based on two channels, and relates to the technical field of radio signal modulation identification, and the method comprises the steps: firstly carrying out the self-supervised pre-training through deeply utilizing the information of an I / Q sequence and an A / P sequence of unmarked data, effectively improving the generalization capability of a model, and carrying out the self-supervised pre-training of the information of the I / Q sequence and the A / P sequence of the unmarked data; and then a good semi-supervised recognition effect can be obtained by using a small number of marked samples for fine tuning, and the problem of high dependence of an automatic modulation recognition algorithm based on deep learning on a large amount of marked data is solved. Under the condition of the same number of marked samples, the recognition accuracy of the method is superior to that of an existing supervised automatic modulation recognition method.

Description

Technical Field

[0001] The invention belongs to the technical field of radio signal modulation recognition, and in particular relates to a dual-channel signal comparison prediction semi-supervised automatic modulation recognition method. Background Art

[0002] In recent years, with the rapid development of communication technology, accurate identification of the modulation scheme of radio signals plays a key role in communication reconnaissance, electronic countermeasures, satellite communications, and spectrum monitoring. In non-cooperative scenarios, it is difficult to obtain various prior information of the received signal. In order to further analyze and process the signal, it is necessary to accurately identify the modulation scheme of complex and diverse radio signals. With the rapid development of deep learning, a large number of automatic modulation recognition methods based on deep learning have emerged, which has improved the accuracy of modulation recognition.

[0003] Existing deep learning methods can be roughly divided into three categories: modulation recognition methods based on convolutional neural networks (CNN), modulation recognition methods based on recurrent neural networks (RNN), and modulation recognition methods based on hybrid models. CNN networks can better extract the spatial correlation characteristics of signals and are the earliest deep learning networks used in automatic modulation recognition. Since radio signals have strong temporal correlation characteristics, many scholars have used RNN to achieve modulation recognition. In addition, more and more scholars have begun to study automatic modulation recognition methods based on hybrid models of the two.

[0004] However, most of the existing deep learning methods are implemented under supervised conditions, that is, most of them are supervised Automatic Modulation Recognition (AMR) methods. The AMR method requires a large amount of reliable labeled data as training samples to ensure model performance. However, obtaining a large amount of valid labeled data under non-cooperative communication conditions requires a lot of professional time and energy, and the efficiency is extremely low, which limits the application of the AMR method in non-cooperative scenarios. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a semi-supervised automatic modulation recognition method based on dual-channel signal comparison prediction. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] The present invention provides a dual-channel signal comparison prediction semi-supervised automatic modulation recognition method, comprising:

[0007] Step 1: obtaining an I / Q sequence and an A / P sequence of a first radio modulation signal according to the first radio modulation signal, wherein the first radio modulation signal is a tag-free radio modulation signal;

[0008] Step 2: inputting the I / Q sequence and the A / P sequence of the first radio modulation signal into a preset self-supervised contrastive prediction encoder for self-supervised pre-training to obtain a trained self-supervised contrastive prediction encoder;

[0009] Step 3: constructing a modulation recognition model based on the trained self-supervised contrast prediction encoder, and training the modulation recognition model using a second radio modulation signal to obtain a trained modulation recognition model, wherein the second radio modulation signal is a radio modulation signal with a true category label;

[0010] Step 4: Using the trained modulation recognition model to realize modulation recognition of the radio signal to be tested.

[0011] In one embodiment of the present invention, step 1 comprises:

[0012] Step 1.1: obtaining an I / Q sequence of the first radio modulation signal according to the first radio modulation signal;

[0013] Step 1.2: Extract amplitude information and phase information from the I / Q sequence of the first radio modulation signal to form the A / P sequence.

[0014] In one embodiment of the present invention, the self-supervised contrastive prediction encoder comprises a cascaded dual-channel input network, an autoregressive model and a fully connected layer unit, wherein the dual-channel input network comprises a first encoder, a second encoder and a channel dimension fusion unit;

[0015] The first encoder is used to extract characteristic information of the I / Q sequence;

[0016] The second encoder is used to extract characteristic information of the A / P sequence;

[0017] The channel dimension fusion unit is used to fuse the feature information of the I / Q sequence and the feature information of the A / P sequence in the channel dimension to obtain a potential representation sequence;

[0018] The autoregressive model is used to extract the historical coding information of the radio modulation signal according to the potential representation sequence before the current moment;

[0019] The fully connected layer unit is used to obtain a predicted potential representation sequence with a preset prediction step size according to the historical coding information.

[0020] In one embodiment of the present invention, step 2 comprises:

[0021] Step 2.1: Inputting the I / Q sequence and the A / P sequence of the first radio modulation signal into the self-supervised contrast prediction encoder to obtain a predicted potential representation sequence;

[0022] Step 2.2: According to the constructed self-supervised prediction loss function, calculate the contrast loss between the predicted potential representation sequence of the preset prediction step size and the potential representation sequence of the preset prediction step size obtained according to the dual-channel input network, and train the self-supervised contrast prediction encoder based on the contrast loss back-propagation, and obtain a trained self-supervised contrast prediction encoder after reaching a preset number of training times.

[0023] In one embodiment of the present invention, the self-supervised prediction loss function is expressed as:

[0024]

[0025] In the formula, K is the preset prediction step size, Indicates the average value, W k is the time series prediction projection head corresponding to each moment in the preset prediction step, T represents the matrix transpose, Represents historical coding information, t is the current time, is a potential of a preset prediction step length obtained according to the first radio modulation signal

[0026] Represents a sequence, Representation and The potential representation sequence with different values, Sequence of predicted latent representations with a preset prediction step size constitute a negative sample pair.

[0027] In one embodiment of the present invention, step 3 comprises:

[0028] Step 3.1: constructing the modulation recognition model, wherein the modulation recognition model includes a cascade of trained self-supervised contrastive prediction encoders and a classifier, wherein the classifier includes two cascaded fully connected layers and a batch normalization layer;

[0029] Step 3.2: Obtaining an I / Q sequence and an A / P sequence of the second radio modulation signal according to the second radio modulation signal;

[0030] Step 3.3: inputting the I / Q sequence and the A / P sequence of the second radio modulation signal into the modulation recognition model, wherein the classifier is used to map the predicted potential representation sequence of the second radio modulation signal output by the trained self-supervised contrastive prediction encoder to a target category, so as to obtain the predicted modulation category of the second radio modulation signal;

[0031] Step 3.4: Train the modulation recognition model by back-propagation according to the classification loss between the predicted modulation category and the true category label, and obtain a trained modulation recognition model after reaching a preset number of training times.

[0032] In one embodiment of the present invention, step 4 comprises:

[0033] Step 4.1: Processing the radio signal to be tested to obtain an I / Q sequence and an A / P sequence of the radio signal to be tested;

[0034] Step 4.2: Input the I / Q sequence and A / P sequence of the radio signal to be tested into the trained modulation recognition model to obtain the predicted modulation category of the radio signal to be tested.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The semi-supervised automatic modulation recognition method based on dual-channel signal comparison prediction of the present invention firstly performs self-supervised pre-training by deeply utilizing the above information of the I / Q sequence and amplitude phase (A / P) sequence of unlabeled data, effectively improving the generalization ability of the model, and then uses a small amount of labeled samples for fine-tuning to obtain a good semi-supervised recognition effect, solving the high dependence of the automatic modulation recognition algorithm based on deep learning on a large amount of labeled data. Under the same number of labeled samples, the recognition accuracy of the method of the present invention is better than that of the existing supervised automatic modulation recognition method.

[0037] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following specifically cites a preferred embodiment and describes it in detail with the accompanying drawings as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of a dual-channel signal comparison prediction semi-supervised automatic modulation recognition method provided by an embodiment of the present invention;

[0039] Figure 2 It is a structural block diagram of a self-supervised contrast prediction encoder provided by an embodiment of the present invention.

[0040] Figure 3 This is a diagram of recognition accuracy under different signal-to-noise ratios with different numbers of labeled samples provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a dual-channel signal comparison prediction semi-supervised automatic modulation recognition method proposed by the present invention in combination with the accompanying drawings and specific implementation methods.

[0042] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0043] In the first aspect, the embodiment of the present invention provides a semi-supervised automatic modulation recognition method based on dual-channel signal comparison prediction, see Figure 1 , Figure 1 1 is a flow chart of a method for semi-supervised automatic modulation recognition based on dual-channel signal comparison prediction provided by an embodiment of the present invention. Figure 1 As shown, the dual-channel signal comparison prediction semi-supervised automatic modulation recognition method of this embodiment includes:

[0044] Step 1: Obtain an I / Q sequence and an A / P sequence of the first radio modulation signal according to a first radio modulation signal, where the first radio modulation signal is a tag-free radio modulation signal.

[0045] Optionally, the first radio modulation signal may be a radio modulation signal received in a non-cooperative communication scenario, which may be expressed as:

[0046]

[0047] Where s(t) is the transmitted signal, h(t) is the impulse response of the transmission channel, * represents the convolution operation, n(t) is the noise signal, Δf and They are respectively the phase offset caused by the additional carrier frequency offset and phase jitter during the transmission process.

[0048] In an optional embodiment, step 1 includes:

[0049] Step 1.1: Obtain an I / Q sequence of the first radio modulation signal according to the first radio modulation signal.

[0050] In this embodiment, the received radio modulation signal is converted into an I / Q sequence after frequency conversion. The I / Q sequence is usually stored in the form of 2×L, which can be expressed as r(t)=[r I (t); r Q (t)].

[0051] Step 1.2: Extract amplitude information and phase information from the I / Q sequence of the first radio modulation signal to form an A / P sequence.

[0052] In this embodiment, the A / P sequence, i.e., the amplitude-phase sequence, is stored in the form of 2×L, which can be expressed as r′(t)=[r A (t); r P (t)].

[0053] Step 2: Input the I / Q sequence and the A / P sequence of the first radio modulation signal into a preset self-supervised contrastive prediction encoder for self-supervised pre-training to obtain a trained self-supervised contrastive prediction encoder.

[0054] See also Figure 2 , Figure 2 is a structural block diagram of a self-supervised contrast prediction encoder provided by an embodiment of the present invention. Figure 2 As shown, the self-supervised contrastive prediction encoder of this embodiment includes a cascaded dual-channel input network, an autoregressive model and a fully connected layer unit, wherein the dual-channel input network includes a first encoder, a second encoder and a channel-dimensional fusion unit.

[0055] In this embodiment, the first encoder is used to extract characteristic information of the I / Q sequence; and the second encoder is used to extract characteristic information of the A / P sequence.

[0056] Optionally, the first encoder and the second encoder are both two-layer convolutional neural networks, and the convolution kernel sizes are (1, 8) and (2, 7) respectively.

[0057] In this embodiment, the channel dimension fusion unit is used to fuse the characteristic information of the I / Q sequence and the characteristic information of the A / P sequence in the channel dimension to obtain a potential representation sequence. That is, the input radio modulation signal can be mapped into a potential representation sequence F through the dual-channel input network. r .

[0058] In this embodiment, the autoregressive model is used to extract the historical coding information of the radio modulation signal according to the latent representation sequence before the current moment.

[0059] The autoregressive model (Gate Recurrent Unit, GRU) can extract the historical coding information of the radio modulation signal to predict the future coding information. In this embodiment, the GRU aggregates the potential representation sequence before the current moment And output the historical encoding information based on the latent space, which can be expressed as:

[0060]

[0061] Among them, f Rn Represents the autoregressive model GRU, and t is the current time.

[0062] In this embodiment, the fully connected layer unit is used to obtain a predicted potential representation sequence with a preset prediction step size according to historical coding information.

[0063] It can be understood that the prediction result of each step in the latent space can be achieved through different linear transformations. Then, the predicted latent representation sequence with a preset prediction step size can be expressed as:

[0064]

[0065] Among them, f Ln represents the fully connected layer unit, W k is the time series prediction projection head corresponding to each moment in the preset prediction step, W k ∈[W1,W2,…W K ], K is the preset prediction step size, and the number of projection heads is the same as the preset prediction step size of the signal.

[0066] In an optional embodiment, step 2 includes:

[0067] Step 2.1: Input the I / Q sequence and the A / P sequence of the first radio modulation signal into a self-supervised contrast prediction encoder to obtain a predicted potential representation sequence;

[0068] Step 2.2: According to the constructed self-supervised prediction loss function, calculate the contrast loss between the predicted potential representation sequence of the preset prediction step and the potential representation sequence of the preset prediction step obtained according to the dual-channel input network, and train the self-supervised contrastive prediction encoder based on the contrastive loss back-propagation. After reaching the preset number of training times, the trained self-supervised contrastive prediction encoder is obtained.

[0069] In this embodiment, by comparing the predicted potential representation sequence with the preset prediction step length The potential representation sequence with the preset prediction step length obtained from the two-channel input network Similarity, approximate indirect estimation and maximize the context information m t and predict the next step r t+k The mutual information between them is used to complete the self-supervised pre-training by back-propagation with the help of contrastive loss.

[0070] But directly modeling p(r t+k ,mt ) Extract m t and r t+k The computational cost of the mutual information between is very high. This embodiment proposes to use the prediction signal r t+k and the current context information are encoded into a compact distributed vector, expressed as:

[0071]

[0072] In the concrete representation L(r t+k ,m t ), a simple log-bilinear model is used for optimization:

[0073]

[0074] Therefore, the self-supervised prediction loss function constructed in this embodiment is expressed as:

[0075]

[0076] In the formula, K is the preset prediction step size, Indicates the average value, W k is the time series prediction projection head corresponding to each moment in the preset prediction step, T represents the matrix transpose, Represents historical coding information, t is the current time, is a potential representation sequence of a preset prediction step length obtained according to the first radio modulation signal, Representation and The potential representation sequence with different values, Sequence of predicted latent representations with a preset prediction step size constitute a negative sample pair.

[0077] Step 3: construct a modulation recognition model based on the trained self-supervised contrast prediction encoder, and train the modulation recognition model using the second radio modulation signal to obtain a trained modulation recognition model, where the second radio modulation signal is a radio modulation signal with a true category label.

[0078] In an optional embodiment, step 3 includes:

[0079] Step 3.1: Construct a modulation recognition model, which includes a cascade of trained self-supervised contrastive prediction encoders and classifiers, wherein the classifier includes two cascaded fully connected layers and a batch normalization layer.

[0080] Step 3.2: Obtain an I / Q sequence and an A / P sequence of the second radio modulation signal according to the second radio modulation signal.

[0081] Step 3.3: Input the I / Q sequence and A / P sequence of the second radio modulation signal into the modulation recognition model. The classifier is used to map the predicted potential representation sequence of the second radio modulation signal output by the trained self-supervised comparative prediction encoder to the target category to obtain the predicted modulation category of the second radio modulation signal.

[0082] In this embodiment, the classifier maps the input data to the target category and determines which modulation category the radio signal belongs to has the highest probability, which can be expressed as:

[0083]

[0084] Among them, M represents the modulation type, N is the type of modulation type, and D Lr represents the data set consisting of all labeled signals, and α∈[0,1] represents the proportion of labeled signals used.

[0085] Step 3.4: Train the modulation recognition model by back-propagating the classification loss between the predicted modulation category and the true category label, and obtain the trained modulation recognition model after reaching the preset number of training times.

[0086] In this embodiment, by making full use of contrast prediction coding, unlabeled modulated signal samples are pre-trained for self-supervision to extract the characterization information of the signal. Based on the contrast learning in the modulation recognition model, the context of one time point in the sample is established, and each negative sample contains only the features of one time point. The signal does not need to be data enhanced, and the overhead of the pre-training stage is reduced while extracting robust modulation signal characterization information. Satisfactory results can be obtained by fine-tuning with only a small number of labeled samples. With the same number of labeled samples, the modulation recognition accuracy is significantly improved.

[0087] It can be understood that after obtaining the trained modulation recognition model, the modulation recognition model can be tested through the test set, that is, the samples in the test set are input into the trained modulation recognition model for classification and recognition, and the output is the predicted label of the input data. The predicted label and the true label of the sample are compared to obtain the classification and recognition result, and the recognition results of all samples are statistically analyzed to obtain the classification accuracy of the modulation recognition model.

[0088] Step 4: Use the trained modulation recognition model to realize modulation recognition of the radio signal to be tested.

[0089] In an optional embodiment, step 4 includes:

[0090] Step 4.1: Process the radio signal to be tested to obtain an I / Q sequence and an A / P sequence of the radio signal to be tested;

[0091] Step 4.2: Input the I / Q sequence and A / P sequence of the radio signal to be tested into the trained modulation recognition model to obtain the predicted modulation category of the radio signal to be tested.

[0092] The semi-supervised automatic modulation recognition method based on dual-channel signal comparison prediction of the present invention firstly performs self-supervised pre-training by deeply utilizing the above information of the I / Q sequence and amplitude phase (A / P) sequence of unlabeled data, effectively improving the generalization ability of the model, and then uses a small amount of labeled samples for fine-tuning to obtain a good semi-supervised recognition effect, solving the high dependence of the automatic modulation recognition algorithm based on deep learning on a large amount of labeled data. Under the same number of labeled samples, the recognition accuracy of the method of the present invention is better than that of the existing supervised automatic modulation recognition method.

[0093] Furthermore, the effect of the dual-channel signal comparison prediction semi-supervised automatic modulation recognition method of this embodiment is described through simulation experiments.

[0094] 1. Open Source Dataset Experiment

[0095] The experiment uses the public dataset RML2016.10a, which includes 220,000 modulated signals with signal-to-noise ratios (SNRs) ranging from -20dB to +18dB, and considers 11 commonly used modulations: WBFM, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, QPSK, and 8PSK. Each signal in the dataset has 128 I / Q samples in complex form and is generated in a harsh simulated propagation environment that is corrupted by AWGN, multipath fading, sampling rate offset, and center frequency offset to simulate the actual environment.

[0096] The experiment divides the dataset into three parts in a ratio of 6:2:2: unlabeled training set, test set, and validation set. All unlabeled signals in the training set will be used for self-supervised pre-training. In the semi-supervised classification stage, N labeled signals are randomly extracted from the unlabeled dataset. The number of samples selected for each modulation type at each signal-to-noise ratio is kept consistent. In order to eliminate the influence of individual differences of samples on the classification results, all experimental results are averaged over ten Monte Carlo experiments. The Adam optimizer is used in all experiments. In the unsupervised pre-training stage, a total of 240 rounds of training are performed. The learning rate is set to 0.00025 and the batch size of gradient update is 128 to avoid local optimality. When the validation loss does not improve within 30 epochs, the training process is stopped. In the fine-tuning stage, a total of 120 rounds of training are performed. In order to achieve better fine-tuning effect, the learning rates of 0.0005 and 0.001 are used for the model and classifier, respectively.

[0097] In order to analyze the effective and full utilization of the annotation resources by the modulation recognition model, the fine-tuning performance of the modulation recognition model was experimentally analyzed using different numbers of labeled samples N = 2, 5, 10 (1%), 20, 50, 100 (10%). The analysis results are shown in the figure. Figure 3 When N=100 (10%), the semi-supervised classification recognition accuracy of the modulation recognition model basically reaches more than 83% when SNR≥0dB.

[0098] 2. Model performance analysis experiment

[0099] In order to evaluate the effectiveness of the modulation model of the present invention, mainstream supervised models (CNN1, CNN2, ResNet, CLDNN, MCLDNN) were selected for comparison with the proposed method of the present invention. By testing on labeled data sets of different sizes, the recognition results are shown in Table 1, showing the significant advantages of the modulation model of the present invention under different labeled sample sizes. When the labeled samples are extremely low (N=2, 0.2%), its recognition accuracy at the full signal-to-noise ratio is significantly improved compared with the supervised model, approximately 31.51%-20.15%. As the sample size continues to increase, the performance gap narrows, and there is still an improvement of 24.51%-3.97% under the full signal-to-noise ratio condition. It highlights the effective use of unlabeled data by the method of the present invention.

[0100] Table 1

[0101]

[0102]

[0103] 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 that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant is intended to cover non-exclusive inclusion, so that the article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the article or device including the elements. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The orientation or position relationship indicated by "up", "down", "left", "right", etc. is based on the orientation or position relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention.

[0104] 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 dual-channel signal comparison prediction semi-supervised automatic modulation recognition method, characterized in that: include: Step 1: obtaining an I / Q sequence and an A / P sequence of a first radio modulation signal according to the first radio modulation signal, wherein the first radio modulation signal is a radio modulation signal without a tag; Step 2: inputting the I / Q sequence and the A / P sequence of the first radio modulation signal into a preset self-supervised contrastive prediction encoder for self-supervised pre-training to obtain a trained self-supervised contrastive prediction encoder; Step 3: constructing a modulation recognition model based on the trained self-supervised contrast prediction encoder, and training the modulation recognition model using a second radio modulation signal to obtain a trained modulation recognition model, wherein the second radio modulation signal is a radio modulation signal with a true category label; Step 4: Using the trained modulation recognition model to realize modulation recognition of the radio signal to be tested.

2. The method for semi-supervised automatic modulation recognition based on dual-channel signal comparison prediction according to claim 1 is characterized in that: The step 1 comprises: Step 1.1: obtaining an I / Q sequence of the first radio modulation signal according to the first radio modulation signal; Step 1.2: Extract amplitude information and phase information from the I / Q sequence of the first radio modulation signal to form the A / P sequence.

3. The method for semi-supervised automatic modulation recognition based on dual-channel signal comparison prediction according to claim 1 is characterized in that: The self-supervised contrastive prediction encoder includes a cascaded dual-channel input network, an autoregressive model, and a fully connected layer unit, wherein the dual-channel input network includes a first encoder, a second encoder, and a channel dimension fusion unit; The first encoder is used to extract characteristic information of the I / Q sequence; The second encoder is used to extract characteristic information of the A / P sequence; The channel dimension fusion unit is used to fuse the feature information of the I / Q sequence and the feature information of the A / P sequence in the channel dimension to obtain a potential representation sequence; The autoregressive model is used to extract the historical coding information of the radio modulation signal according to the potential representation sequence before the current moment; The fully connected layer unit is used to obtain a predicted potential representation sequence with a preset prediction step size according to the historical coding information.

4. The method for semi-supervised automatic modulation recognition based on dual-channel signal comparison prediction according to claim 3 is characterized in that: The step 2 comprises: Step 2.1: Inputting the I / Q sequence and the A / P sequence of the first radio modulation signal into the self-supervised contrast prediction encoder to obtain a predicted potential representation sequence; Step 2.2: According to the constructed self-supervised prediction loss function, calculate the contrast loss between the predicted potential representation sequence of the preset prediction step size and the potential representation sequence of the preset prediction step size obtained according to the dual-channel input network, and train the self-supervised contrast prediction encoder based on the contrast loss back-propagation, and obtain a trained self-supervised contrast prediction encoder after reaching a preset number of training times.

5. The method for semi-supervised automatic modulation recognition based on dual-channel signal comparison prediction according to claim 4 is characterized in that: The self-supervised prediction loss function is expressed as: In the formula, K is the preset prediction step size, Indicates the average value, W k is the time series prediction projection head corresponding to each moment in the preset prediction step, T represents the matrix transpose, Represents historical coding information, t is the current time, is a potential representation sequence of a preset prediction step length obtained according to the first radio modulation signal, Representation and The potential representation sequence with different values, Sequence of predicted latent representations with a preset prediction step size constitute a negative sample pair.

6. The method for semi-supervised automatic modulation recognition based on dual-channel signal comparison prediction according to claim 3 is characterized in that: The step 3 comprises: Step 3.1: constructing the modulation recognition model, wherein the modulation recognition model includes a cascade of trained self-supervised contrastive prediction encoders and a classifier, wherein the classifier includes two cascaded fully connected layers and a batch normalization layer; Step 3.2: Obtaining an I / Q sequence and an A / P sequence of the second radio modulation signal according to the second radio modulation signal; Step 3.3: inputting the I / Q sequence and the A / P sequence of the second radio modulation signal into the modulation recognition model, wherein the classifier is used to map the predicted potential representation sequence of the second radio modulation signal output by the trained self-supervised contrastive prediction encoder to a target category, so as to obtain the predicted modulation category of the second radio modulation signal; Step 3.4: Train the modulation recognition model by back-propagation according to the classification loss between the predicted modulation category and the true category label, and obtain a trained modulation recognition model after reaching a preset number of training times.

7. The method for semi-supervised automatic modulation recognition based on dual-channel signal comparison prediction according to claim 6 is characterized in that: The step 4 comprises: Step 4.1: Processing the radio signal to be tested to obtain an I / Q sequence and an A / P sequence of the radio signal to be tested; Step 4.2: Input the I / Q sequence and A / P sequence of the radio signal to be tested into the trained modulation recognition model to obtain the predicted modulation category of the radio signal to be tested.

Images