An improved tri-training-based AMC scene migration method
By generating pseudo-labels and augmenting the dataset through an improved tri-training network, the cross-scene recognition challenge of communication signal modulation types under varying drone scenarios was solved, improving recognition accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2022-11-10
- Publication Date
- 2026-06-30
AI Technical Summary
In communication systems, especially in the changing scenarios of drone applications, it is difficult to accurately classify the modulation type of communication signals, resulting in poor recognition performance of communication signals across different scenarios.
An improved tri-training network is used to generate pseudo-labels and augment the dataset, and three classifiers are used for training to achieve automatic modulation classification across scenarios.
It improves the recognition accuracy and robustness of communication signals across different scenarios and enhances the generalization ability of the classifier.
Smart Images
Figure CN115659146B_ABST
Abstract
Description
Technical Field
[0001] This invention pertains to the type identification of communication signals in the field of signal processing, specifically involving an AMC (Automatic Modulation Classification) scene transfer method based on an improved tri-training (semi-supervised learning algorithm). Background Technology
[0002] In communication systems, signals are transmitted based on different modulation types, and the receiving end needs to know the modulation method in order to perform operations such as signal demodulation and information acquisition. In non-cooperative communication, the modulation method adopted by the transmitting end is unpredictable, and the modulation type needs to be determined based on the signal itself.
[0003] Therefore, automatic modulation recognition has high research and application value in the field of signal processing, and deep neural networks based on data-driven and feedback optimization are a research hotspot in this direction. However, the communication environment is complex and ever-changing, especially in the time-varying scenarios of UAV applications. Different terrains cause waveform distortion and aberration, and the distribution of communication signal data shifts, making it difficult to correctly identify and classify communication signals across different scenarios. Summary of the Invention
[0004] To address the problem of poor communication signal recognition performance under changing communication scenarios, this invention designs an improved tri-training network based on the Resent network. When the scenario changes, the improved asymmetric tri-training network is used for AMC scenario migration, resulting in a significant improvement in performance.
[0005] The AMC scene transfer method based on improved tri-training has the following specific steps:
[0006] Step 1: For actual UAV flight scenarios in the open sea and mountainous areas, based on large-scale fading models and small-scale fading models, set different air-to-ground channel model parameters and use simulation software to generate simulation datasets for each scenario.
[0007] Step 2: Measure the similarity of the generated simulation datasets for different scenarios based on cosine similarity and maximum mean difference. Based on the similarity measurement results, set the simulation data of the open sea scenario as the source domain data and the simulation data of the hill and mountain scenario as the target domain data.
[0008] The source domain data contains labels, and the dataset is denoted as . The target domain data is unlabeled, and the dataset is denoted as [database name missing]. The dimensions are all 1024*2; m and n are the sample sizes of the source domain data and the target domain data set, respectively.
[0009] Step 3: Transfer unlabeled target domain data Input the classifier to generate pseudo-label target domain data Combine labeled source domain data The classifier is trained to obtain a new classifier F. t ;
[0010] (x i ,y i ) in x i It is the i-th source domain data, y i Source domain data x i The tag; x j It is the j-th target domain data; The target domain data x j pseudo-tags;
[0011] The specific steps are as follows:
[0012] Step 301, using source domain data X s Pre-train three classifiers F1, F2, and F3;
[0013] Among them, F1, F2 and F3 all contain a shared convolutional layer F;
[0014] Step 302: Randomly sample a portion of the unlabeled target domain data. For the current single target domain data, input it into any two pre-trained classifiers F. i (1≤i≤3) and F j In the range (1≤j≤3), the output predicted labels are consistent and at least one is greater than a certain threshold. The pseudo-label that is greater than the certain threshold is valid and is saved to the pseudo-label target domain dataset.
[0015] Step 303, transfer the pseudo-label target domain dataset Add to source domain dataset X s In the above, let's call it the augmented dataset. Repeated random sampling yielded three amplified datasets L. 12 L 13 L 23 ;
[0016] Step 304: Traverse the unlabeled target domain data X t For a single target domain data set, the data is input into three pre-trained classifiers F1, F2, and F3, respectively, resulting in three corresponding pseudo-labels. If the predicted labels output by the three classifiers are consistent and at least one is greater than a specific threshold, the pseudo-label exceeding the threshold is considered valid and saved to the target domain dataset X containing the pseudo-labels. t l .
[0017] Step 305, using the three augmented datasets L 12 L 13 L23 Update the three classifiers separately, while using the target domain dataset X containing pseudo-labels. t l Training a new classifier F t .
[0018] Data set L ij Used for training classifier F i and F j ;
[0019] Step 306: Return to step 301 and continuously resample the target domain data to update the augmented dataset and the target domain dataset X containing pseudo-labels. t l The four classifiers are updated repeatedly until all classifiers converge.
[0020] Step 4: Based on classifier F t The target domain data is identified and classified, which means that cross-scenario identification and classification is achieved.
[0021] The present invention has the following advantages:
[0022] 1) An AMC scene transfer method based on improved tri-training, which generates different scene datasets based on simulation software modeling, facilitating theoretical analysis and mathematical deduction;
[0023] 2) An AMC scene transfer method based on improved tri-training introduces an asymmetric tri-training network model to minimize the divergence between the source and target domains to obtain domain-invariant features, thereby improving the recognition effect and robustness of communication modulation signals in cross-scene situations;
[0024] 3) An AMC scene transfer method based on improved tri-training. Asymmetric tri-training only has two classifiers to generate pseudo-labels. This invention adds a classifier to this, using three classifiers to obtain three augmented datasets, reducing the dependence on the initial two classifiers. Attached Figure Description
[0025] Figure 1 This is a flowchart of an AMC scene transfer method based on improved tri-training according to the present invention;
[0026] Figure 2 This invention improves the tri-training network pre-training to obtain a new classifier F. t Flowchart;
[0027] Figure 3 This is a schematic diagram of the improved tri-training network structure constructed in this invention;
[0028] Figure 4 It is the residual module of ResNet used in the improved tri-training network of this invention;
[0029] Figure 5 It is a residual unit in the ResNet residual module of this invention;
[0030] Figure 6 These are comparison images showing the effects of scene migration before and after using AMC in this invention. Detailed Implementation
[0031] The specific implementation method of the present invention will be further described in detail below with reference to the accompanying drawings.
[0032] This invention primarily studies automatic signal modulation and classification across different communication scenarios, aiming to improve the generalization ability of classifiers and enhance the robustness of cross-scenario signal recognition. Based on whether the source and target domains are labeled, it can be categorized into: inductive transfer learning with a small number of labeled samples in the target domain, transductive transfer learning with labeled samples only in the source domain, and unsupervised transfer learning with no labels in either the source or target domain. This invention belongs to transductive transfer learning, aiming to extract domain-invariant features by minimizing the divergence between domains through an improved tri-training network.
[0033] The AMC scene transfer method based on improved tri-training, such as Figure 1 As shown, the specific steps are as follows:
[0034] Step 1: For the actual flight scenarios of UAVs in the open sea and mountainous areas, based on the large-scale fading model and the small-scale fading model, set different air-to-ground channel model parameters, use simulation software to generate simulation datasets for each scenario, and conduct migration feasibility analysis.
[0035] Step 2: Measure the similarity of the generated simulation datasets for different scenarios based on cosine similarity and maximum mean difference. Based on the similarity measurement results, set the simulation data of the open sea scenario as the source domain data and the simulation data of the hill and mountain scenario as the target domain data.
[0036] The source domain data contains labels, and the dataset is denoted as . The target domain data is unlabeled, and the dataset is denoted as [database name missing]. The dimensions are all 1024*2; m and n are the sample sizes of the source domain data and the target domain data set, respectively.
[0037] For simulation data and After calculating the amplitudes based on the IQ and IQ paths, we obtain corresponding 1024*1 dimensional vectors a and b. The cosine similarity is then calculated by taking the cosine of the included angle θ, as shown in the following formula:
[0038]
[0039] Where a = [a1, a2, ... a l b = [b1, b2, ... b] l The length of l is 1024.
[0040] The formula for the maximum mean difference is as follows:
[0041]
[0042] Where k() is the kernel function; H indicates that the maximum mean difference is measured by mapping the data to a high-dimensional space using the kernel function.
[0043] Step 3: Transfer unlabeled target domain data Input the classifier to generate pseudo-label target domain data Combine labeled source domain data The classifier is trained to obtain a new classifier F. t ;
[0044] (x i ,y i ) in x i It is the i-th source domain data, y i Source domain data x i The tag; x j It is the j-th target domain data; The target domain data x j pseudo-tags;
[0045] An improved tri-training network was constructed, setting different scenario data as source and target domain data. Three classifiers were pre-trained using the source domain data (labeled). Based on these classifiers, pseudo-labels were generated using the target domain data (unlabeled). Finally, formal training was performed using the source domain data and the pseudo-labeled target domain data. Figure 2 As shown, the specific steps are as follows:
[0046] Step 301, using source domain data X s Pre-train three classifiers F1, F2, and F3;
[0047] Among them, F1, F2 and F3 all contain a shared convolutional layer F;
[0048] Step 302: Randomly sample a portion of the unlabeled target domain data. For the current single target domain data, input it into any two classifiers F. i (1≤i≤3) and F jIn the range (1≤j≤3), the output predicted labels are consistent and at least one is greater than a certain threshold. The pseudo-label that is greater than the certain threshold is valid and is saved to the pseudo-label target domain dataset.
[0049] Step 303, transfer the pseudo-label target domain dataset Add to source domain dataset X s In the above, let's call it the augmented dataset. Repeated random sampling yielded three amplified datasets L. 12 L 13 L 23 ;
[0050] Step 304: Traverse the unlabeled target domain data X t For a single target domain data set, the data is input into three pre-trained classifiers F1, F2, and F3, respectively, resulting in three corresponding pseudo-labels. If the predicted labels output by the three classifiers are consistent and at least one is greater than a specific threshold, the pseudo-label exceeding the threshold is considered valid and saved to the target domain dataset X containing the pseudo-labels. t l .
[0051] Step 305, using the three augmented datasets L 12 L 13 L 23 Update the three classifiers separately, while using the target domain dataset X containing pseudo-labels. t l Training a new classifier F t .
[0052] Data set L ij Used for training classifier F i and F j ;F t Includes a shared convolutional layer F;
[0053] Step 306: Return to step 301 and continuously resample the target domain data to update the augmented dataset and the target domain dataset X containing pseudo-labels. t l The four classifiers are updated repeatedly until all classifiers converge.
[0054] Step 4: Based on classifier F t The target domain data is identified and classified, which means that cross-scenario identification and classification is achieved.
[0055] Example:
[0056] Step 1: Construct a dataset of ocean and mountain / hill scenes.
[0057] 1) Based on different scenarios, we studied large-scale fading models and small-scale fading air-to-ground channel models, and set different air-to-ground channel parameters based on measured data. We also used simulation software to generate simulation datasets, which included 11 modulation schemes with signal-to-noise ratios ranging from -20dB to 30dB. Each sample had 1024 sampling points and two channels (I and Q).
[0058] The modulation schemes are: "BPSK", "QPSK", "8PSK", "16PSK", "16APSK", "32APSK", "64APSK", "16QAM", "32QAM", "GMSK", "OQPSK".
[0059] There are 3016 of each modulation scheme, the dataset size is 33176, and there are 116 of each modulation scheme and each SNR.
[0060] 2) Scenario transferability analysis
[0061] The generated simulation datasets for different scenarios are measured based on two metrics: cosine similarity and maximum mean difference.
[0062] The dataset for the offshore scene is denoted as The dataset for the hill and mountain scene is denoted as The dimension is 1024*2; m and n are the sample sizes of the datasets for each scenario.
[0063] Cosine similarity measures the difference between two samples by calculating the cosine of the angle between them. This is relevant to simulation data. and After calculating the amplitudes of the IQ and IQ channels, we obtain the corresponding 1024*1 dimensional vectors a and b, respectively. The cosine similarity is obtained by calculating the cosine value of the included angle θ.
[0064] The maximum mean difference is found by finding a mapping function that maps variables to a high-dimensional space. The upper bound of the expected difference between the two random variables after mapping is the maximum mean difference, defined by the following formula:
[0065]
[0066] Where φ is the kernel function, and H represents the maximum mean difference, which is measured by mapping the data to the Reproducing Hilbert space (RKHS) using φ(). Expanding this, it can be expressed as:
[0067]
[0068] Where k() is the kernel function; H indicates that the maximum mean difference is measured by mapping the data to a high-dimensional space using the kernel function.
[0069] The similarity measurement results between the source domain data and the target domain data are shown in Table 1:
[0070] Table 1
[0071]
[0072] Scenario transferability analysis: Table 1 shows that the similarity metrics are positively correlated with cosine similarity and negatively correlated with the maximum mean difference. Cosine similarity differences are small, indicating a weak ability to reflect differences in data distribution. However, the maximum mean difference is relatively large, thus it is a better measure of similarity between domains. Furthermore, Table 1 data shows that data distribution differences are small within the same scenario, but significant differences exist between different scenarios.
[0073] Therefore, the data for offshore scenes and hilly / mountainous scenes can be set as different domain data.
[0074] Step 2: Build an improved tri-training network
[0075] Different scenario data are designated as source and target domain data. Tagned offshore scenario simulation data is designated as the source domain, denoted as . Unlabeled hill and mountain range scene simulation data is set as the target domain, denoted as . Three classifiers are pre-trained based on the input source domain data. Then, pseudo-labels are generated based on the input target domain data, denoted as...
[0076] The overall network structure is as follows Figure 3 As shown, F represents the output shared feature network. Classifiers F1, F2, and F3 classify based on the features generated by F, and the prediction results are used to provide pseudo-labels. Classifiers F1, F2, and F3 are trained based on source domain data and target domain data containing pseudo-labels. t Training is performed on target domain data containing pseudo-labels.
[0077] First, using source domain data X s Three classifiers F1, F2, and F3 are pre-trained; a portion of the target domain data is randomly sampled, and two classifiers F1, F2, and F3 are used. i and F j Label it, for classifier F i and F j The pseudo-label is valid only if the two classifiers predict the same label and at least one of them is greater than a specific threshold (set to 0.9). The pseudo-label is then used to predict the label using classifiers F1 and F2 to obtain the target domain dataset containing the pseudo-label. Add to source domain dataset X s Obtain the augmented dataset L 12 .
[0078] Similarly, for classifiers F1 and F3, and classifiers F2 and F3, repeat the above operations to obtain three augmented datasets L. 12 L 13 L 23 ;
[0079] Three classifiers are used to label the target domain data. A pseudo-label is considered valid if all three classifiers predict the same label and at least one label is greater than a specific threshold (set to 0.9), resulting in a target domain dataset X containing pseudo-labels. t l .
[0080] Then, using three augmented datasets L 12 L 13 L 23 Update these three classifiers, where L ij Training classifiers F, F i and F j Using the target domain dataset X containing pseudo-labels t l Training a new classifier F t ;
[0081] Finally, the target domain data is continuously resampled, and the three augmented datasets and the target domain dataset with pseudo-labels are updated using the updated three classifiers. The training is repeated until the model converges.
[0082] The shared network layer F adopts the ResNet network structure, and the model structure is shown in Table 2. It includes six residual blocks, and the residual block structure is as follows: Figure 4 As shown, each residual block includes a 1×1 linear convolution, two residual units, and a max pooling layer, where the residual units are as follows: Figure 5 As shown, F1, F2, F3 and F t It consists of two fully connected layers with 128 neurons each, and the last layer is a softmax layer.
[0083] Table 2
[0084]
[0085]
[0086] Step 3: Based on classifier F t The target domain data is identified and classified, which means that cross-scenario identification and classification is achieved.
[0087] The effectiveness of the improved tri-training network model in this application was verified by training and simulation testing.
[0088] The test environment was Windows 10, Python 3.6, and TensorFlow 1.8. The source domain data was a distant ocean scene, and the target domain data was a hilly mountain scene. It included 11 modulation schemes, with a signal-to-noise ratio (SNR) ranging from -20dB to 30dB. Each sample had 1024 sampling points, both I and Q channels, and the dataset size was 33176 data points. Each modulation scheme had 116 SNR values. The results are shown in Table 3.
[0089] Table 3
[0090]
[0091] like Figure 6 As shown in Table 4, the overall accuracy based on the ResNet network is about 37.1%, and the accuracy is about 45% when the signal-to-noise ratio is above 0 dB. In contrast, the overall accuracy of the tri-training model based on the ResNet structure is improved to about 46.2%, and the signal-to-noise ratio is improved to about 60.5% when it is above 0 dB.
[0092] Table 4
[0093]
Claims
1.A method for AMC scenario migration based on improved tri-training, characterized in that, The specific steps are as follows: First, for the actual flight scenarios of UAVs in the open sea and mountainous areas, based on large-scale fading models and small-scale fading models, different air-to-ground channel model parameters are set, and simulation datasets for each scenario are generated using simulation software. The source domain data contains labels, and the data set is denoted as The target domain data is unlabeled, and the data set is denoted as Both have a dimension of 1024*2; and are the sample sizes of the source domain data and the target domain data set, respectively; Then, the generated simulation datasets of different scenarios are similarized based on cosine similarity and maximum mean difference. Based on the similarity measurement results, the simulation data of the open sea scenario is set as the source domain data, and the simulation data of the hill and mountain scenario is set as the target domain data. The formula for calculating the cosine similarity of simulation data is as follows: in and simulation data and The 1024*1 dimensional vector obtained after calculating the amplitude based on the IQ two paths; , Length is ; The formula for the maximum mean difference is as follows: in, It is a kernel function; The maximum mean difference is measured by mapping the data to a high-dimensional space using a kernel function. Unlabeled target domain data Input the classifier to generate pseudo-label target domain data Combined with labeled source domain data Train the classifier to obtain a new classifier. ; middle It is the first Data from each source domain, Source domain data Tags; It is the first Data for each target domain; Target domain data pseudo-tags; The specific steps are as follows: Step 301, using source domain data Pre-trained three classifiers , and ; Step 302: Randomly sample a portion of the unlabeled target domain data. For the current single target domain data, input it into any two pre-trained classifiers. and In the process, the output predicted labels are consistent and at least one is greater than a certain threshold. The pseudo-labels that are greater than the certain threshold are valid and are saved to the pseudo-label target domain dataset. ; Step 303, transfer the pseudo-label target domain dataset Add to source domain dataset In the above, let's call it the augmented dataset. Three augmented datasets were obtained through repeated random sampling. , , ; Step 304: Traverse the unlabeled target domain data For the current single target domain data, input three pre-trained classifiers respectively. , and In the process, three corresponding pseudo-label data are obtained. When the predicted labels output by the three classifiers are consistent and at least one of them is greater than a certain threshold, the pseudo-label that is greater than the certain threshold is valid and saved to the target domain dataset containing pseudo-labels. ; Step 305, using the three augmented datasets , , Update the three classifiers separately, while utilizing the target domain dataset containing pseudo-labels. Training a new classifier ; Step 306: Return to step 301 and continuously resample the target domain data to update the augmented dataset and the target domain dataset containing pseudo-labels. And repeatedly update the four classifiers until all classifiers converge; Finally, based on the classifier The target domain data is identified and classified, which means that cross-scenario identification and classification is achieved.