An AMC scene migration method based on adversarial domain adaptation

Abstract

The application discloses an AMC scene migration method based on an adversarial domain adaptation, and belongs to the field of signal processing. Specifically, first, according to the flight characteristics of a UAV, a common sea and shore water area air-ground channel scene signal dataset of the UAV is simulated and constructed, wherein the sea scene is taken as a source domain, and the shore water area scene is taken as a target domain. Then, an adversarial domain adaptation model is established, signal features are extracted from generated signals via a feature extractor of the model, modulation features are obtained by an automatic encoder, and the modulation features are aligned by weighting two different features and using an adversarial training mode, so that the feature extractor extracts domain-invariant modulation features, and modulation mode recognition of the target domain signal under cross-scenarios is realized. The method of the application aims at the problem of performance decline of modulation mode recognition caused by domain difference, is based on the theory of transfer learning, minimizes the difference between domains by using the adversarial domain adaptation, and enhances the robustness of the modulation signal recognition method.

Description

Technical Field

[0001] This invention belongs to the field of signal processing, specifically relating to an AMC (Automatic Modulation Classification) scene transfer method based on adversarial domain adaptation (semi-supervised learning algorithm). Background Technology

[0002] With the development and popularization of wireless communication, various modulation methods for communication signals are constantly emerging.

[0003] Modulation and demodulation (MDM) technology, as a crucial component of physical layer communication technology, is essential for achieving efficient channel utilization, reducing system power consumption, and improving communication quality. AMC technology can effectively identify the modulation scheme of a signal, providing a basis for subsequent signal analysis and processing, and plays a key role in military, intelligence, and security fields. Therefore, the research and application of this technology are of great significance and have broad application prospects in the current and future fields of wireless communication.

[0004] In recent years, with the continuous development of deep learning methods, data-driven modulation scheme classification (AMC) techniques have achieved superior performance in many fields. Given sufficient signal sample data, this technique can automatically extract complex signal features and patterns, thereby efficiently improving the classification accuracy of modulation scheme classification methods. However, with the continuous expansion of communication application scenarios, the distribution of UAV communication signal data is affected by the differences in wireless channel environments under new communication scenarios, leading to changes in its distribution and thus reducing the classification performance of deep learning-based AMC methods.

[0005] To address this issue, transfer learning has been introduced into the field of modulation scheme classification. Transfer learning relaxes the requirement that the original scene signal data and the new scene signal data follow the same distribution, and can effectively complete the modulation scheme classification task even when they are inconsistent. Summary of the Invention

[0006] To address the issue of performance degradation in UAV debugging mode classification across different communication scenarios, this invention proposes an AMC scene transfer method based on the adversarial domain adaptation model ISRA. By aligning the modulation features of the source and target domains using an adversarial transfer strategy, a feature extractor is trained to extract domain-invariant features, thereby achieving modulation feature transfer from UAV offshore communication scenarios to shore-based water communication scenarios and improving the model's recognition performance in shore-based water scenarios.

[0007] The specific steps of the AMC scene migration method based on adversarial domain adaptation are as follows:

[0008] Step 1: Based on the large-scale fading model and the small-scale fading channel model, generate AMC simulation signal datasets for offshore scenarios and shore-based UAV air-to-ground channel scenarios according to different air-to-ground channel model parameters.

[0009] Step 2: Perform channel model analysis on signal datasets under different scenarios, using AMC simulation signal data from the offshore scenario in the dataset as the source domain. AMC simulation signal data of shore-based water scenarios as the target domain

[0010] Where, n s Let n be the number of labeled signal samples in the source domain. t The sample size of the unlabeled signal in the target domain; the i-th source domain input data contains the original signal x. i Modulation mode label y i and signal-to-noise ratio (SNR) i The j-th target domain input data contains the original signal x. j and signal-to-noise ratio (SNR) j .

[0011] Step 3: Input the labeled source domain data and the unlabeled target domain data into the feature extractor F in the backbone network, respectively, and train it to extract domain-invariant features; at the same time, train its label classifier to converge.

[0012] First, all N signals from the source domain data and the target domain data are input into the backbone network containing the feature classifier and the label classifier in batches. The feature extractor outputs the signal features f. Each signal feature is passed through an autoencoder to output a modulation feature containing modulation information. The modulation features of the N signals are merged into the total modulation feature Z.

[0013] Then, for each signal feature f, calculate its weight with the N modulation features respectively; obtain an N×N weight value matrix; then sum the weights column-wise in the source and target domains respectively to obtain the weight value of each modulation feature.

[0014] Finally, the backpropagation error is calculated using the weighted average of the modulation features. The parameters of the feature extractor are then automatically updated using the backpropagation algorithm to ensure that the extracted features contain modulation information and have domain invariance. At the same time, the label classifier is automatically updated to achieve convergence.

[0015] Step 4: After the label classifier converges, the feature extractor and label classifier in the backbone network are used to identify and classify the modulation mode of the target domain signal data, thus realizing the identification and classification of signal modulation mode across scenarios.

[0016] The present invention has the following advantages:

[0017] 1) An AMC scene migration method based on adversarial domain adaptation addresses the challenges of observing the modulation scheme of acquired signals and the high cost of obtaining labels by relaxing the requirements for target domain signal data. Target domain signal data does not require labeling, and a large amount of unlabeled data can be efficiently utilized to improve AMC task performance in new scenarios.

[0018] 2) An AMC scene transfer method based on adversarial domain adaptation introduces the adversarial domain adaptation method into the AMC field. By aligning modulation-related features in the source and target domains, the difference in signal data distribution caused by different communication scenarios is reduced, thereby improving the robustness and generalization of cross-scenario communication signal modulation mode recognition.

[0019] 3) An AMC scene migration method based on adversarial domain adaptation is proposed. The adversarial domain adaptation method is introduced, and the backbone feature extractor structure is designed in combination with the characteristics of communication signal data to make it more suitable for the signal data structure and facilitate the extraction of modulation mode features. Attached Figure Description

[0020] Figure 1 This is a flowchart of an AMC scene migration method based on adversarial domain adaptation according to the present invention;

[0021] Figure 2 This is a schematic diagram of the adversarial domain adaptation network structure constructed in this invention;

[0022] Figure 3 This is a flowchart of the training process for the adversarial domain adaptation network of this invention;

[0023] Figure 4 This is a schematic diagram of the weighted modulation features of the automatic encoder of the present invention;

[0024] Figure 5 This is a comparison chart of the adversarial domain adaptation method and the baseline method under different data volumes. Detailed Implementation

[0025] The specific implementation method of the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] To improve the classification performance of communication signal modulation schemes when communication scenarios change, this invention proposes an AMC scene migration method based on adversarial domain adaptation, such as... Figure 1 As shown, the specific steps are as follows:

[0027] Step 1: Based on the large-scale fading model and the small-scale fading channel model, considering factors such as shadow fading, multipath fading and Doppler shift, set parameters such as shadow fading, maximum path delay and maximum Doppler shift according to different air-to-ground channel models to generate AMC simulation signal datasets for offshore scenarios and shore-based water UAV air-to-ground channel scenarios.

[0028] Step 2: Use the AMC simulation signal data of the offshore scenario in the dataset as the source domain. AMC simulation signal data of shore-based water scenarios as the target domain

[0029] Where, n s Let n be the number of labeled signal samples in the source domain. t The sample size of the unlabeled signal in the target domain; the i-th source domain input data contains the original signal x. i Modulation mode label y i and signal-to-noise ratio (SNR) i The j-th target domain input data contains the original signal x. j and signal-to-noise ratio (SNR) j The number of sampling points for both the source and target domain signals is the same, 1024.

[0030] Step 3: Input the labeled source domain data and the unlabeled target domain data into the feature extractor F in the backbone network, respectively, and train it to extract domain-invariant features; at the same time, train its label classifier to converge.

[0031] The specific steps are as follows:

[0032] Step 301: In one iteration, for all N signals of source domain data and target domain data in a batch, they are sequentially input into the backbone network. The feature extractor of the backbone network outputs the signal feature f of each signal, and they are aligned row by row to form the signal feature set as follows:

[0033]

[0034] Where n1 is the amount of source domain signal data input to the backbone network in one iteration, n2 is the amount of target domain signal data input to the backbone network in one iteration, and m is the length of the signal feature;

[0035] The backbone network includes feature classifiers and label classifiers;

[0036] Step 302: Simultaneously, each signal is sequentially input into the automatic encoder, and the modulation features of each signal containing modulation mode information are output; the modulation features of all N signals are merged into the total modulation feature Z.

[0037] Step 303: For each signal, calculate the weights w between the signal feature f and the N modulation features in the total modulation features Z. N,N This yields an N×N weight matrix:

[0038]

[0039] Each row in the matrix represents the weight between the corresponding signal feature and N modulation features.

[0040] Step 304: Sum the weight values ​​in the weight matrix column-wise in both the source and target domains to obtain:

[0041]

[0042] in, This represents the sum of the weights of the j-th modulation feature corresponding to the signal features of each signal in the source domain. The sum of the weight values ​​of the j-th modulation feature corresponding to the signal feature of each signal in the target domain;

[0043] Step 305: Average the sum of the above weight values ​​to obtain... Used to calculate backpropagation error.

[0044] The backpropagation error is composed of the loss function L1 for signal modulation label classification, the loss function L2 for the autoencoder, the loss function L3 for the domain discriminator D, and the loss function L4 for the source and target domain modulation feature weights.

[0045] Specifically as follows:

[0046] L1 = L s1 +L t1 =L cross-entropy (y s ,y s ')+L entropy (y t ')

[0047] L2 = L s2 +L t2 =L mse (f s ,f s ')+L mse (f t ,f t ')

[0048] L3 = L s3 +L t3 =L BCE (d s ,d s ')+L BCE (d t ,d t ')

[0049]

[0050] Where L s1 L represents the source domain label modulation classification loss. t1 L represents the target domain label modulation classification loss. cross-entropy Indicates the calculation of y s 'and ys The cross-entropy loss function between them, y s ' represents the classification label of the source domain signal modulation mode by the label discriminator, y s L represents the true label of the source domain data. entropy Indicates the calculation of y t Entropy is used as the loss function; y t 'Represents the classification label of the target domain signal modulation mode by the label discriminator;

[0051] L s2 L represents the loss function representing the source domain signal features and reconstructed features of the input autoencoder. t2 L represents the loss function representing the target domain signal features and reconstructed features input to the autoencoder. mse This indicates the calculation of the mean square error between f' and f; f s The feature f represents the source-domain input autoencoder. s 'Represents the reconstruction feature of the source domain of the autoencoder output, f t The feature f represents the input of the target domain autoencoder. t 'Represents the reconstructed features of the target domain output by the autoencoder;

[0052] L s3 Loss when performing domain discrimination on source domain signals, L t3 L represents the loss when performing domain discrimination on the target domain signal. BCE Indicates the calculation of d s 'and d s Binary cross-entropy loss between, d s 'Represents the discrimination of the source domain signal by the domain discriminator, d s For the source domain signal, 0 represents that the signal comes from the source domain, and 1 represents that the signal comes from the target domain; similarly, d t 'Represents the discrimination of the target domain signal by the domain discriminator, d t The domain label for the target domain signal.

[0053] This represents the average weight of each modulation feature in the source domain with respect to each signal feature; therefore, this value has the same dimension as the number of modulation features. This represents the average weight of each modulation feature in the target domain for each signal feature.

[0054] Step 306: The backpropagation algorithm is used to automatically update the feature extractor and label classifier. The next iteration is performed, and the process returns to step 301. Signal features and modulation features are continuously extracted from the source and target domain signals and their weights are calculated. The feature extractor and label classifier are updated repeatedly until the features extracted by the feature extractor contain modulation information and have domain invariance, and the label classifier converges.

[0055] Step 4: After the label classifier converges, the feature extractor and label classifier in the backbone network are used to identify and classify the modulation mode of the target domain signal data, thus realizing the identification and classification of signal modulation mode across scenarios.

[0056] Example:

[0057] The first step is to generate AMC simulation signal datasets for offshore scenarios and shore-based UAV air-to-ground channel scenarios based on large-scale fading models and small-scale fading channel models, according to different air-to-ground channel model parameters.

[0058] Each signal comprises three parts: signal, signal-to-noise ratio (SNR), and modulation scheme. The SNR ranges from -20dB to 30dB. By controlling the SNR, its impact on signal modulation recognition performance can be studied. Each original signal passing through the channel contains 1024 sampling points, divided into I and Q channels, represented as (1024*2). Both the source and target domains have six modulation schemes: BPSK, QPSK, 16APSK, 16QAM, GMSK, and OQPSK. Each modulation scheme corresponds to a number, so 0 to 5 represent these six modulation schemes respectively.

[0059] The second step involves channel model analysis of signal datasets under different scenarios, as follows: In the offshore scenario, the surrounding environment is open and unobstructed, with virtually no multipath effect. Signal energy is mainly attenuated due to distance, climate changes, and shadow fading. Because the UAV flies at a relatively high speed in this scenario, the Doppler effect is significant, and the signal waveform does not change much except for noise, indicating good channel conditions. In the shore-based water scenario, there are scattered buildings and trees around the antenna, and reflections from the ground or objects near the transmitter can cause multipath effects. Due to the high flight speed, the Doppler effect is significant, resulting in greater signal distortion.

[0060] Based on the analysis results, the generated AMC signal data from the offshore scene was used as the source domain. shore-based waters as the target domain n s Let n be the number of labeled signal samples in the source domain. t This represents the number of unlabeled samples in the target domain. The number of sampling points in both the source and target domains is the same, 1024. The source domain input data includes the original signal, modulation label, and signal-to-noise ratio (SNR), while the target domain input data includes the original signal and SNR. That is, the source domain consists of labeled data, denoted as: The target domain data is unlabeled data, denoted as: Each signal has a length of (1024*2). In addition, the signal-to-noise ratio (SNR) information of the signals is hidden during training. Training is performed with approximately equal numbers of signals for each SNR. The test signal data is classified according to the SNR only when testing the model, and the trend of changes in accuracy and SNR is observed.

[0061] The third step is to build an adversarial domain adaptation model based on the proposed adversarial domain adaptation method.

[0062] The model consists of two main parts: a backbone network (ResNet) containing a feature extractor and a label classifier, and an adversarial network that performs adversarial training and alignment on the features f extracted from the source and target domains.

[0063] A schematic diagram of the adversarial domain adaptation model is shown below. Figure 2 As shown, scenario one represents the source domain, and scenario two represents the target domain. Signal data is input into the network, and a feature extractor extracts domain-invariant features. The feature maps are then passed through a label classifier and a neighborhood discriminator, respectively. The loss functions of each are calculated separately, and the results are backpropagated back to the feature extractor to achieve the training effect. During backpropagation training, the gradient of the discriminator's loss passes through a gradient reversal layer (GRL). The main function of the GRL is to add a negative constant to the backpropagation gradient, thereby maximizing the error of the neighborhood discriminator. That is, the neighborhood discriminator cannot distinguish whether the signal comes from the source domain or the target domain. This training method ensures that the features extracted by the feature extractor contain information that can distinguish the modulation method, but do not contain any domain information; in other words, the extracted features are similar in the source and target domains. The flowchart of the adversarial domain adaptation network training is shown below. Figure 3 As shown.

[0064] The specific steps are as follows:

[0065] First, the source domain data X s and target domain data X t The input consists of a backbone network containing a feature classifier and a label classifier. The feature extractor outputs signal features f, which are then processed by an autoencoder to output modulation features containing modulation scheme information. The modulation features Z of the source domain signal are then processed... s Modulation characteristics Z of the target domain signal t They are combined into a total modulation feature Z.

[0066] Then, for each signal feature f, the weight w between it and the total modulation feature Z is calculated by the weight calculation module. N,N Then, the weights are summed column-wise in both the source and target domains to obtain the weight value for each modulation feature Z. The autoencoder modulation feature weighting method is as follows: Figure 4 As shown.

[0067] Finally, the loss function of the feature extractor F, the loss function between the autoencoder input f and the autodecoder output f', the loss function of the domain discriminator D, and the loss function of the source and target domain modulation feature weights are calculated. The parameters of the feature extractor are then modified by the backpropagation algorithm so that the extracted features contain modulation information and have domain invariance.

[0068] The fourth step is to identify and classify the modulation scheme of the target domain data based on the backbone network ResNet, that is, to achieve modulation scheme identification across scenarios.

[0069] A network model is established for the adversarial domain adaptation method proposed in this application. The effectiveness of the method is verified by training and simulation testing on the model.

[0070] The test environment was a Windows 10 system, Python 3.6, and PyTorch 1.2 framework. The source domain data was from offshore scenes, and the target domain data was from shore-based water scenes. It included 16 modulation schemes with a signal-to-noise ratio of -20dB to 30dB. Each sample had 1024 sampling points, with both I and Q channels. There were 2000 samples for each modulation scheme, for a total of 24,000 signals in both the source and target domains.

[0071] Furthermore, the classic ResNet network was chosen as the baseline method to demonstrate the effectiveness of the proposed method. The baseline model was tested on a Windows 10 system using Python 3.7 and the TensorFlow 2.1 framework. The source and target domain data were consistent with those of the adversarial domain network model.

[0072] This invention also investigated the impact of data volume on recognition accuracy. The data volume for each modulation method was increased to 5000 records, and the experimental results were observed under otherwise unchanged conditions. The results are as follows: Figure 5 As shown in Table 1, the accuracy (%) under different signal-to-noise ratios is as follows.

[0073] Table 1:

[0074]

[0075] Based on the experimental results, 1) In terms of recognition accuracy: the accuracy of both ResNet and the adversarial domain adaptation model in recognizing signal modulation methods increases with the increase of signal-to-noise ratio, and reaches a peak at around 4dB; under the same signal-to-noise ratio conditions, the signal recognition accuracy of the adversarial domain adaptation model is significantly higher than that of ResNet, especially under high signal-to-noise ratio conditions.

[0076] 2) Regarding signal data requirements: Within the model, increasing the amount of training data significantly improves recognition accuracy, especially under high signal-to-noise ratio (SNR) conditions, achieving an accuracy improvement of 8%–10%. The adversarial domain adaptation model, with 2000 data points per modulation scheme, achieves the accuracy of ResNet with 5000 data points per modulation scheme. Under low SNR conditions, the recognition accuracy of the adversarial domain adaptation model is approximately 4% higher than that of ResNet. Furthermore, with increasing data volume, the increase in recognition accuracy of the adversarial domain adaptation model is lower than that of ResNet, indicating that the method proposed in this invention has a weaker dependence on data and is more advantageous when data is insufficient.

Claims

1. A method for AMC scene migration based on adversarial domain adaptation, characterized in that, The specific steps are as follows: Step 1: Based on the large-scale fading model and the small-scale fading channel model, generate AMC simulation signal datasets for offshore scenarios and shore-based UAV air-to-ground channel scenarios according to different air-to-ground channel model parameters; Step 2: Perform channel model analysis on signal datasets under different scenarios, using AMC simulation signal data from the offshore scenario in the dataset as the source domain. AMC simulation signal data of shore-based water scenarios as the target domain ; in, The number of labeled signals in the source domain. The sample size of the unlabeled signal in the target domain; the first Each source domain input data contains the original signal. Modulation method label and signal-to-noise ratio , No. The target domain input data includes the original signal. and signal-to-noise ratio ; Step 3: Input the labeled source domain data and the unlabeled target domain data into the feature extractor in the backbone network, respectively. Train it to extract domain-invariant features; at the same time train its label classifier to converge. First, combine all the source domain data and target domain data. Each signal is input in batches into the backbone network containing the feature classifier and label classifier, and the feature extractor outputs the signal features. Each signal feature is processed by an automatic encoder, which outputs a modulation feature containing modulation scheme information; The modulation features of individual signals are combined into a total modulation feature. ; Then, for each signal feature Calculate its relationship with The weights among the modulation features are obtained; The weight matrix is ​​obtained; then the weights are summed column-wise in both the source and target domains to obtain the weight value of each modulation feature. Finally, the backpropagation error is calculated using the weighted average of the modulation features. The parameters of the feature extractor are then automatically updated using the backpropagation algorithm to ensure that the extracted features contain modulation information and have domain invariance. At the same time, the label classifier is automatically updated to achieve convergence. The backpropagation error is classified by the loss function of the signal modulation mode label. Loss function of an auto encoder Domain discriminator loss function and the loss function for the modulation feature weights of the source and target domains. It consists of the following components: in This represents the source domain label modulation classification loss. This represents the target domain label modulation classification loss. Indicates calculation and The cross-entropy loss function between them The label represents the classification label used by the label discriminator to determine the modulation method of the source domain signal. Represents the true label of the source domain data. Indicates calculation The entropy is used as the loss function; The label represents the classification label used by the label discriminator to determine the modulation method of the target domain signal; The loss function represents the source domain signal features and reconstructed features input to the autoencoder. The loss function represents the target domain signal features and reconstructed features input to the autoencoder. Indicates calculation and The mean square error between them; Features representing the source-domain input autoencoder Represents the reconstruction features of the source domain of the autoencoder output. Features representing the input of the target domain autoencoder Represents the reconstructed features of the target domain output by the autoencoder; Loss when performing domain discrimination on source domain signals. This represents the loss when performing domain discrimination on the target domain signal. Indicates calculation and Binary cross-entropy loss between them The representation domain discriminator's discrimination of source domain signals. For the source domain signal, 0 represents that the signal comes from the source domain, and 1 represents that the signal comes from the target domain; similarly, The domain discriminator's discrimination of the target domain signal. The domain label for the target domain signal; This represents the average weight of each modulation feature in the source domain with respect to each signal feature; therefore, this value has the same dimension as the number of modulation features. This represents the average weight of each modulation feature in the target domain for each signal feature; Step 4: After the label classifier converges, the feature extractor and label classifier in the backbone network are used to identify and classify the modulation mode of the target domain signal data, thus realizing the identification and classification of signal modulation mode across scenarios.

2. The AMC scene migration method based on adversarial domain adaptation as described in claim 1, characterized in that, In step one, shadow fading, multipath fading, and Doppler shift factors are considered, and shadow fading, maximum path delay, and maximum Doppler shift parameters are set according to different air-to-ground channel models.

3. The AMC scene migration method based on adversarial domain adaptation as described in claim 1, characterized in that, The source domain and target domain The number of signal sampling points is the same, 1024.

4. The AMC scene migration method based on adversarial domain adaptation as described in claim 1, characterized in that, Each signal feature output by the feature extractor Aligned by row, the signal feature set is as follows: in, The amount of source domain signal data input to the backbone network for one iteration. The amount of target domain signal data input to the backbone network for one iteration. The length of the signal feature.

5. The AMC scene migration method based on adversarial domain adaptation as described in claim 1, characterized in that, Each signal feature Calculate and respectively Weights among modulation features ,composition Weight matrix: Each row in the matrix represents the weight between the corresponding signal feature and the N modulation features.

6. The AMC scene migration method based on adversarial domain adaptation as described in claim 1, characterized in that, The summation of the weight value columns in the weight value matrix in both the source and target domains yields: in, The signal characteristics of each signal in the source domain correspond to the first... The sum of the weight values ​​of each modulation feature, The signal features representing each signal in the target domain correspond to the first... The sum of the weight values ​​of each modulation feature.

Images