A small sample automatic modulation recognition method for complex channel environment

Abstract

The application discloses a small sample automatic modulation recognition method for a complex channel environment, and fully excavates the complementarity of cross-view information by constructing a multi-view signal representation and using an independent feature encoder to learn discriminative features. On this basis, an adaptive measurement mechanism with intra-class variance perception is further introduced, and the feature dimension is dynamically reweighted according to the support set statistics, so as to suppress the unreliable dimension interference caused by noise and channel distortion. At the same time, a query-related multi-view distance attention fusion strategy is designed, and the measurement results of each view are adaptively integrated for different query samples, so as to avoid the negative transfer caused by fixed fusion. Then, the model is continuously learned on different small sample modulation recognition tasks, the loss function is optimized, the model is updated, and the optimal recognition model is obtained. Finally, the baseband signal belonging to a modulation type but the specific modulation type is unknown is input into the trained recognition model, and the modulation type is output. The application can effectively improve the small sample modulation recognition reliability in a complex wireless environment, provides a feasible and efficient solution for the modulation recognition application of an actual communication system, and provides a guarantee for subsequent demodulation and signal recovery.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and more specifically, relates to a small-sample automatic modulation recognition method for complex channel environments. Background Technology

[0002] With the proliferation of drones, low-altitude high-speed platforms, and high-speed vehicle-mounted vehicles, the number of unauthorized signal sources and interference emitters in the urban electromagnetic spectrum is increasing. These sources are exhibiting greater concealment and dynamism, posing new challenges to spectrum regulation and wireless situational awareness. The ability to efficiently and accurately sense and identify the modulation patterns of communication signals in complex environments is crucial for signal demodulation analysis and source tracing.

[0003] Automatic Modulation Recognition (AMR) is a technology that automatically identifies the modulation scheme of intercepted signals under unknown or non-cooperative communication conditions. As an intermediate process in signal detection and demodulation, it is a crucial component in the development of intelligent non-cooperative communication systems and cognitive radio systems. It plays an irreplaceable role in modern communication scenarios such as spectrum management, non-cooperative communication, electronic warfare, and wireless surveillance.

[0004] However, modulation type labels typically rely on expert analysis, which is costly and difficult to scale. Consequently, reliable labeled samples are often scarce, creating a natural small-sample learning constraint. Therefore, how to achieve reliable modulation identification using a very small number of labeled samples under complex channel conditions has become a key problem that urgently needs to be solved for AMR to be applied in engineering.

[0005] Currently, existing automatic modulation recognition methods for few-shot samples can be mainly divided into three categories. One category is based on metric learning and prototype discrimination frameworks, improving discrimination ability under few-shot conditions by refining feature representation or similarity calculation. Another category introduces meta-learning ideas into modulation recognition tasks, improving few-shot generalization ability by learning model initializations that can quickly adapt to new tasks. In addition, some studies alleviate the few-shot problem from the perspectives of data and knowledge transfer. Data-level methods mainly improve performance through data augmentation and representation preprocessing, while knowledge transfer methods mainly involve pre-training on large-scale source domain data and transferring knowledge to the target few-shot task. These methods, by constructing discriminative models with only a small number of labeled samples for each class, enable the system to achieve effective recognition in data-scarce scenarios, thus alleviating the dependence on large-scale training data to some extent.

[0006] However, most existing automatic modulation recognition methods for small samples are based on relatively ideal channel assumptions, primarily modeling the signal-to-noise ratio (SNR) as the only adverse factor, without fully considering the impact of complex channel effects on signal feature distribution. Furthermore, these methods generally employ simple distance metric strategies, typically assuming relatively stable feature distributions and limited intra-class differences. When signal features are affected by noise and channel disturbances, these assumptions often fail to hold, leading to unstable or even severely degraded recognition performance. In summary, existing automatic modulation recognition methods for small samples generally suffer from the following limitations:

[0007] (1) The verification is still carried out under ideal or simplified channel conditions. The lack of system modeling for actual propagation damage such as multipath, time-varying channels and Doppler effect leads to a significant decrease in the generalization performance of the model in complex environments.

[0008] (2) It is necessary to assume that the feature distribution is relatively stable and the intra-class differences are limited, which can easily lead to the degradation of the discrimination features under complex channel interference.

[0009] (3) Insufficient utilization of complementary information in the time and frequency domains of the signal leads to a significant decrease in performance when noise or distortion is amplified;

[0010] (4) Under small sample conditions, higher-order modulation types with similar amplitude and phase characteristics are more likely to cause inter-class overlap and confusion, making it difficult to form a stable discrimination boundary, resulting in a higher misjudgment rate.

[0011] Therefore, there is an urgent need for an automatic modulation recognition method that can achieve high accuracy and strong robustness in complex, low-sample communication environments. Summary of the Invention

[0012] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a small-sample automatic modulation recognition method for complex channel environments. This method enhances the representation capability of complex channel disturbances by introducing multi-view signal representation and independent feature encoding. Simultaneously, by introducing an adaptive metric mechanism based on intra-class variance awareness, it effectively characterizes the uncertainty and dimensionality reliability differences of feature distribution under small-sample conditions. Furthermore, a query-related multi-view attention fusion strategy is designed, enabling the model to adaptively integrate information from different views according to specific test samples. This achieves more robust discrimination decisions under various complex channel conditions, providing a solid foundation for subsequent signal demodulation and recovery.

[0013] To achieve the above-mentioned objectives, the present invention provides a small-sample automatic modulation identification method for complex channel environments, characterized by comprising the following steps:

[0014] (1) Signal reception

[0015] The receiver receives the wireless signal from the transmitter and samples it to obtain a complex baseband signal. , is represented as:

[0016]

[0017] in, and The first In-phase and quadrature components at each sampling point , This represents the total number of sampling points;

[0018] (2) Multi-view signal representation and independent feature encoding to obtain the embedding space features of the four views;

[0019] (3) Construct a small sample modulation recognition task

[0020] Using the N-way K-shot method, a few-shot modulation recognition task is constructed based on the embedding spatial features obtained in step (2): In each few-shot modulation recognition task, randomly select Each modulation type is selected. Selecting baseband signals as samples of the support set Each baseband signal is used as a sample of the query set. The samples between the query set and the support set are not repeated. All samples are used to obtain the embedding space features of the four views through step (2).

[0021] (4) Adaptive measurement based on intra-class variance perception

[0022] 4.1) For the first The first of the modulation types The first support set sample The embedded spatial features of a view after being mapped by the encoder are denoted as follows: ,in , , ;

[0023] 4.2) Calculate the first All support set samples of the modulation type in the first... Mean of embedding spatial features under each view:

[0024]

[0025] 4.3) Calculate the first Intra-class variance of all support set samples for each modulation type on each feature dimension:

[0026] ;

[0027] in, Representing embedded spatial features The Dimensional features, Represents the mean of the embedding space features The Dimensional features, , Indicates the number of dimensions;

[0028] 4.4) Construct a dimension-adaptive weight vector based on within-class variance:

[0029] ;

[0030] in, A view-dependent learnable scaling factor used to adjust the strength of the effect of variance on the weight distribution;

[0031] 4.5) Calculate the first The query sample and the first The modulation type in the first Prototype distance under each view:

[0032]

[0033] in, Indicates the first The query sample in the 1st The first view under the first view 3D embedding space features;

[0034] (5) Query-related multi-view distance attention fusion

[0035] 5.1) Constructing the first The query sample and the first View distance vector for each modulation type :

[0036]

[0037] 5.2) First, a lightweight attention network is used. For distance vector Perform mapping to obtain the distance vector. Then use The function will divide the distance vector Mapped to a probability distribution, i.e., a distance weight vector :

[0038]

[0039] in, Indicates the first The modulation type of the first The query sample in the 1st The distance weights under each view range from 0 to 1, and the sum of the distance weights under all four views is 1.

[0040] 5.3) Calculate the first The query sample and the first The final fusion distance for each modulation type prototype is:

[0041] ;

[0042] in Indicates the first The query sample in the 1st The modulation type is the first Distance weights under each view;

[0043] 5.4) Calculate the first The actual modulation type label of each query sample For the first The actual modulation type label for each modulation type The probability of:

[0044]

[0045] 5.5) Calculate the loss function for each few-sample task. :

[0046] ;

[0047] 5.6) Construct an automatic modulation recognition model based on steps (3) to (5.5);

[0048] 5.7) Construct a baseband signal training set and a baseband signal validation set to optimize the automatic modulation recognition model in multiple episodes: In each episode, first obtain the loss function based on the baseband signal training set according to steps (3) to 5.5). According to the loss function The automatic modulation recognition model is optimized, and then the loss function is obtained according to steps (3) to (5.5) based on the baseband signal validation set. And the loss function obtained in the previous round based on the baseband signal validation set. The comparison is performed. If the difference is smaller, the optimized automatic modulation recognition model is retained; otherwise, the unoptimized automatic modulation recognition model is used for the next round of optimization.

[0049] 5.8) Repeat step 5.7) for multiple epochs to obtain the trained automatic modulation recognition model;

[0050] (6) Automatic modulation and recognition of small samples

[0051] Construct a few-sample modulation recognition task based on the embedding spatial features obtained in step (2): Select Each modulation type is selected. Using baseband signals as samples in the support set, the signals belonging to... A baseband signal with one modulation type but an unknown specific modulation type is used as a query sample. The true modulation type label of the query sample is calculated according to steps (4) to (5.5). For the first The actual modulation type label for each modulation type The probability of the modulation type is the highest, and the modulation type with the highest probability is the modulation type of the baseband signal whose specific modulation type is unknown.

[0052] The objective of this invention is achieved as follows:

[0053] This invention presents a small-sample automatic modulation recognition method for complex channel environments. It constructs multi-view signal representations and employs independent feature encoders to learn discriminative features, fully leveraging the complementarity of cross-view information. Building upon this, an adaptive metric mechanism with intra-class variance awareness is introduced, dynamically reweighting feature dimensions based on support set statistics to suppress unreliable dimension interference caused by noise and channel distortion. Simultaneously, a query-related multi-view distance attention fusion strategy is designed to adaptively integrate the metric results of various views for different query samples, avoiding negative transfer caused by fixed fusion. Subsequently, the model continuously learns on different small-sample modulation recognition tasks, optimizing the loss function and updating the model to obtain the optimal recognition model. Finally, the model is assigned to the appropriate category... A baseband signal with an unknown modulation type is input into a trained modulation recognition model, which outputs the modulation type, providing a reliable input for subsequent signal demodulation and recovery. This invention effectively improves the reliability of few-sample modulation recognition in complex wireless environments, providing a feasible and efficient solution for modulation recognition applications in practical communication systems, and ensuring subsequent demodulation and signal recovery.

[0054] Meanwhile, the small-sample automatic modulation recognition method of the present invention for complex channel environments also has the following beneficial effects:

[0055] (1) The present invention can effectively cope with the challenges in complex channel environments, such as multipath effects, noise interference and Doppler frequency shift, and thus has stronger generalization and robustness in practical applications.

[0056] (2) The present invention adopts a small sample learning strategy. By making full use of the support set and query set, it can perform effective learning and reasoning in the case of scarce data, reduce the dependence on a large amount of labeled data and achieve high-precision modulation recognition.

[0057] (3) This invention combines multi-view learning technology to fuse information through different signal representations. Through this multi-view feature extraction, the model can capture detailed information of the signal from different angles, thereby significantly improving the modulation recognition accuracy under complex channel conditions.

[0058] (4) The adaptive metric based on intra-class variance and the query-related attention distance fusion strategy introduced in this invention can further improve the ability to distinguish between high-order modulation and easily confused modulation categories. In complex channels, it can effectively reduce the performance degradation caused by channel disturbances.

[0059] (5) The present invention has been verified on a dataset generated based on RadioML2016.10A. Experimental results show that it is stable under various modulation, multipath fading, Doppler frequency shift and signal-to-noise ratio conditions, and has the prospect of being applied in engineering in complex wireless communication systems. Attached Figure Description

[0060] Figure 1 This is a flowchart of a specific implementation of the small-sample automatic modulation recognition method for complex channel environments according to the present invention;

[0061] Figure 2 This is a network structure diagram of a specific embodiment of the feature encoder used in this invention;

[0062] Figure 3 This is the confusion matrix of the present invention under different N-way K-shot small sample modulation recognition task settings, where (a) is 3-way 1-shot, (b) is 3-way 3-shot, (c) is 3-way 5-shot, (d) is 5-way 1-shot, (e) is 5-way 3-shot, and (f) is 5-way 5-shot;

[0063] Figure 4 This is a graph showing the accuracy curves of different SNRs under different N-way K-shot small sample modulation recognition task settings of the present invention;

[0064] Figure 5 This is a bar chart showing the accuracy of the present invention under different channel environments with different N-way K-shot small sample modulation recognition task settings;

[0065] Figure 6This is the ablation experiment (training accuracy) under the 3-way 5-shot small sample modulation recognition task setting of this invention.

[0066] Figure 7 This is an ablation experiment (training loss) under the 3-way 5-shot small sample modulation recognition task setting of this invention.

[0067] Figure 8 This is an ablation experiment (accuracy graphs of different SNRs) under the 3-way 5-shot small sample modulation recognition task setting of this invention.

[0068] Figure 9 This is an ablation experiment (accuracy map of different channel environments) under the 3-way 5-shot small sample modulation recognition task setting of this invention. Detailed Implementation

[0069] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0070] Figure 1 This is a flowchart of a specific implementation of the small-sample automatic modulation and identification method for complex channel environments according to the present invention.

[0071] In this embodiment, as Figure 1 As shown, the small-sample automatic modulation recognition method for complex channel environments of the present invention includes the following steps:

[0072] Step S1: Signal Reception

[0073] The receiver receives the wireless signal from the transmitter and samples it to obtain a complex baseband signal. , is represented as:

[0074]

[0075] in, and The first In-phase and quadrature components at each sampling point , This represents the total number of sampling points.

[0076] In this embodiment, .

[0077] Step S2: Multi-view signal representation and independent feature encoding

[0078] In this invention, multi-view signal representation and independent feature coding are introduced to enhance the representation capability of complex channel disturbances. Specifically, the invention includes the following steps:

[0079] Step S2.1: Convert the baseband signal Perform time-domain processing to obtain the waveform view:

[0080]

[0081] in, Represents a waveform view.

[0082] Waveform views can preserve the timing structure information of a signal, which is of great importance in distinguishing modulation methods with different symbol structures.

[0083] Step S2.2: Convert the baseband signal The in-phase component I and the quadrature component Q of each IQ sampling point are mapped as complex plane coordinates to obtain a constellation view:

[0084]

[0085] in, Represents a constellation view.

[0086] Under ideal conditions, different modulation schemes correspond to different constellation point distribution structures. Although constellation points may spread or rotate under conditions such as low signal-to-noise ratio, multipath, and Doppler, the constellation view can still provide important geometric discrimination information.

[0087] Step S2.3: Convert the baseband signal Convert to polar coordinates, i.e., baseband signal The in-phase component I and quadrature component Q are converted into amplitude components and phase components:

[0088]

[0089]

[0090] Then based on the amplitude sequence With phase sequence Constructing an amplitude-phase view:

[0091]

[0092] in, This represents an amplitude-phase view.

[0093] The amplitude-phase view has better decoupling capability for amplitude perturbations and phase noise, which helps to improve the robustness of the model under channel distortion conditions.

[0094] Step S2.4: Baseband signal The in-phase component I and the quadrature component Q are respectively estimated using the Welch method:

[0095]

[0096]

[0097] in, For window function, baseband signal Divided into Segments, each segment is [length missing] , , The first The first paragraph In-phase and quadrature components at each sampling point This represents the energy normalization term of the window function. This represents a discrete frequency index.

[0098] Based on in-phase component power spectral density Power spectral density of orthogonal components Constructing a frequency domain power spectral density view:

[0099]

[0100] in, This represents a frequency domain power spectral density view.

[0101] A frequency domain power spectral density view can reflect the differences in frequency domain energy distribution of modulation methods, which is of great significance for distinguishing modulation types with similar time-domain structures. In this embodiment, the length of each segment is... ,common part.

[0102] Step S2.5: Addressing the significant differences in statistical characteristics and physical meaning between different views, for each view... An independent feature encoder was designed. ,in, This represents the set of parameters of the encoder, the feature encoder. The network structure is completely identical, with no shared parameters and weights. Each view is mapped to the feature space via a feature encoder.

[0103]

[0104] in, For the first A view Embedded space features obtained by mapping through a feature encoder.

[0105] In this embodiment, all feature encoders are completely identical in network structure, such as... Figure 2 As shown, parameters and weights are not shared.

[0106] Figure 2 This paper demonstrates a network structure for the feature encoder used in this invention, detailing each convolutional layer, batch normalization (BatchNorm), activation function (ReLU), pooling layer (MaxPool), and the dimensionality changes of the output. The input data is a view with the following shape: This is a three-channel image. After four convolutional layers, the feature map is flattened into a vector for subsequent operations.

[0107] Step S3: Construct a few-sample modulation recognition task

[0108] The N-way K-shot method is used to construct a few-shot modulation recognition task based on the embedding spatial features obtained in step S2: In each few-shot modulation recognition task, randomly select Each modulation type is selected. Selecting baseband signals as samples of the support set Each baseband signal is used as a sample of the query set. The samples between the query set and the support set are not repeated. All samples are used to obtain the embedding space features of the four views through step S2.

[0109] In this embodiment, six N-way K-shot combinations were selected because it is necessary to test the model's performance under different combinations. The N-way K-shots used for each training, validation, and testing session were consistent, namely 5-way 1-shot, 5-way 3-shot, 5-way 5-shot, 3-way 1-shot, 3-way 3-shot, and 3-way 5-shot. Regardless of the combination of NK, Q is always 15.

[0110] Step S4: Adaptive Measurement Based on Within-Class Variance Awareness

[0111] Step S4.1: For the first The first of the modulation types The first support set sample The embedded spatial features of a view after being mapped by the encoder are denoted as follows: ,in , , ;

[0112] Step S4.2: Calculate the first... All support set samples of the modulation type in the first... Mean of embedding spatial features under each view:

[0113]

[0114] The mean of the embedding space features is the first... Modulation type number The prototype vector of each view.

[0115] Step S4.3: Calculate the first... Intra-class variance of all support set samples for each modulation type on each feature dimension:

[0116] ;

[0117] in, Representing embedded spatial features The Dimensional features, Represents the mean of the embedding space features The Dimensional features, , This indicates the number of dimensions. The within-class variance characterizes the stability of the feature dimension within the current class; a larger variance indicates that the dimension is more significantly affected by noise or channel disturbances.

[0118] Step S4.4: Construct a dimension-adaptive weight vector based on within-class variance:

[0119] ;

[0120] in, This is a view-dependent learnable scaling factor used to adjust the strength of the effect of variance on the weight distribution.

[0121] In this invention, to suppress dimensions significantly affected by noise or channel disturbances and to incorporate intra-class variance information into the distance metric process, an adaptive dimension weight vector based on intra-class variance is constructed. By introducing intra-class variance information into the distance metric process, feature dimensions with smaller intra-class variance and higher stability will receive greater weights, while the influence of unstable dimensions is explicitly suppressed.

[0122] Step S4.5: Calculate the first... The query sample and the first The modulation type in the first Prototype distance under each view:

[0123]

[0124] in, Indicates the first The query sample in the 1st The first view under the first view The proposed adaptive distance metric is a class-conditionally adaptive weighted Euclidean distance, with weights dynamically generated from the statistical properties of the support set, eliminating the need for additional supervision. This adaptive distance metric effectively suppresses the interference of noise-sensitive dimensions on distance calculation, providing a more robust similarity evaluation basis for subsequent multi-view fusion and classification decisions.

[0125] Step S5: Query relevant multi-view distance attention fusion

[0126] Step S5.1: Construct the first The query sample and the first View distance vector for each modulation type :

[0127]

[0128] Step S5.2: First, use a lightweight attention network. For distance vector Perform mapping to obtain the distance vector. Then use The function will divide the distance vector Mapped to a probability distribution, i.e., a distance weight vector :

[0129]

[0130] in, Indicates the first The modulation type of the first The query sample in the 1st The distance weights for each of the four views range from 0 to 1, and the sum of the distance weights for all four views is 1.

[0131] The function maps each input value to a probability, transforming the elements of a vector into a probability distribution with values ​​ranging from 0 to 1, and the sum of all outputs being 1. This distance weight vector... The display depends on the matching relationship between the query sample and the category, which is the contribution of the model's ability to dynamically adjust each view for different query samples.

[0132] Step S5.3: Calculate the first... The query sample and the first The final fusion distance for each modulation type prototype is:

[0133] ;

[0134] in Indicates the first The query sample in the 1st The modulation type is the first Distance weights under each view.

[0135] Through this query-related fusion approach, the model can adaptively emphasize view information that is more discriminative for the current query sample.

[0136] Step S5.4: Calculate the first... The actual modulation type label of each query sample For the first The actual modulation type label for each modulation type The probability of:

[0137]

[0138] Step S5.5: Calculate the loss function for each few-sample task. :

[0139] ;

[0140] Step S5.6: Construct an automatic modulation recognition model based on steps S3 to S5.5;

[0141] Step S5.7: Construct a baseband signal training set and a baseband signal validation set to optimize the automatic modulation recognition model in multiple episodes: In each episode, first obtain the loss function based on the baseband signal training set according to steps S3~S5.5. According to the loss function The automatic modulation recognition model is optimized, and then the loss function is obtained according to steps S3 to S5.5 based on the baseband signal validation set. And the loss function obtained in the previous round based on the baseband signal validation set. The models are compared; if the difference is smaller, the optimized automodulation recognition model is retained; otherwise, the unoptimized automodulation recognition model is used for the next round of optimization.

[0142] Step S5.8: Repeat step S5.7 for multiple epochs to obtain the trained automatic modulation recognition model.

[0143] Each episode is trained based on different categories and samples to improve the model's generalization ability and accuracy in few-shot learning tasks, ultimately resulting in a trained model. In this example, the number of epochs is 50, and the number of episodes is 200.

[0144] Step S6: Automatic Modulation and Recognition of Small Samples

[0145] Construct a few-sample modulation recognition task based on the embedded spatial features obtained in step S2: Select Each modulation type is selected. Using baseband signals as samples in the support set, the signals belonging to... A baseband signal with a modulation type but an unknown specific modulation type is used as a query sample. The true modulation type label of the query sample is calculated according to steps S2 to S5.5. For the first The actual modulation type label for each modulation type The probability of the modulation type is the highest, and the modulation type with the highest probability is the modulation type of the baseband signal whose specific modulation type is unknown.

[0146] Figure 3 The visualization of the confusion matrix is ​​presented, illustrating the performance of different class recognition methods in a few-shot modulation recognition task. Each subplot represents a different N-way K-shot setting, corresponding to different numbers of classes and the number of support set samples for each class. As the number of samples per class increases (i.e., from 1-shot to 5-shot), the classification accuracy gradually improves, and misclassification significantly decreases. In the 3-way 1-shot and 5-way 1-shot settings, misclassification is more pronounced, especially when the number of classes is large (e.g., in the 5-way setting), the similarity between classes may lead to a high misclassification rate. However, in the 3-way 5-shot and 5-way 5-shot settings, due to the larger number of samples per class, the model's recognition ability is stronger, and the classification accuracy is significantly improved. In summary, increasing the number of samples and optimizing the settings helps improve the performance of modulation recognition. In complex environments, this invention can effectively improve classification accuracy and reduce errors caused by channel influences, demonstrating its robustness in practical applications.

[0147] Figure 4 The graphs show the relationship between signal-to-noise ratio (SNR) and accuracy under different N-way K-shot settings. Different curves correspond to different N-way K-shot settings, and each curve is distinguished by a different sign. As can be seen from the graphs, accuracy significantly improves with increasing SNR, especially when the SNR is greater than 0 dB, where accuracy approaches 100%. At low SNRs (e.g., -20 dB), the model performs poorly with lower accuracy because the signal may no longer possess separable features in extreme SNR environments. Increasing the number of K-shot samples effectively improves accuracy. Under the same SNR and K-shot conditions, the 3-way setting generally outperforms the 5-way setting, demonstrating that fewer classes provide higher classification performance in complex environments. Overall, this invention exhibits strong adaptability and robustness in low-sample learning and complex channel environments.

[0148] Figure 5 The accuracy of the model under various channel conditions is compared with different N-way K-shot settings. It can be seen that the model accuracy gradually improves with the increase of the number of K-shot samples, especially with the most outstanding performance in the 5-way 5-shot setting. In complex channel environments, factors such as multipath fading, Doppler shift, and random channel environments have a certain impact on accuracy, but the present invention still maintains good robustness under these conditions.

[0149] Figure 6 The classification accuracy of the models under different module combinations during training is shown. It can be observed that the classification accuracy of each model generally shows a gradual upward trend with the increase of training epochs. A very consistent pattern is observed: models that introduce MVR-IFE (Multi-View Signal Representation and Independent Feature Encoding) have significantly higher training accuracy curves than the single-view version SVR-IFE and its combinations. This indicates that compared to single-view input, multi-view representation can extract features complementary from different signal domains, thereby improving intra-class aggregation and inter-class separability, and improving the convergence characteristics of the optimization process, providing more sufficient and stable discriminative information, which is the key source of performance improvement. The model proposed in this invention (MVR-IFE+CVAM+QAMF) maintains high accuracy in the later stages of training.

[0150] Figure 7 The paper demonstrates the changes in model loss during training under different module combinations. It can be observed that the training loss of each model continuously decreases with the increase of training epochs, indicating that all models can converge stably. Models incorporating MVR-IFE (Multi-View Signal Representation and Independent Feature Encoding) exhibit significantly lower training losses. The proposed model (MVR-IFE+CVAM+QAMF) shows an even lower loss level in the later stages of training, reflecting the complementarity and synergistic effect of each module.

[0151] Figure 8The classification performance of various ablation variants under different signal-to-noise ratio (SNR) conditions is demonstrated. It can be observed that the recognition accuracy of all models increases with the gradual increase of SNR. Overall, models incorporating MVR-IFE (Multi-View Signal Representation and Independent Feature Encoding) show significantly higher performance curves than single-view routes (SVR-IFE and its combinations) across most SNR ranges, with even more pronounced advantages at low and high SNR conditions, demonstrating stronger noise resistance. This indicates that multi-view representation can still provide sufficient and stable discriminative information even under strong noise, which is a key factor in improving the model's robustness at low SNR. Based on multi-view representation, the introduction of IVAM (Intra-Class Variance Aware Adaptive Metric) and QAMF (Query-Related Multi-View Distance Attention Fusion) further enhances the model's stable gain in low and medium-high SNR ranges. The model proposed in this invention maintains the highest or near-highest recognition accuracy across the entire SNR range, demonstrating that CVAM and QAMF can further enhance discriminative capabilities in the multi-view feature space. Furthermore, it can be observed that SVR-IFE and SVR-IFE+QAMF, as well as SVR-IFE+CVAM and SVR-IFE+CVAM+QAMF, exhibit the same accuracy performance. This is because a single view does not involve adaptive fusion, so this module does not play a role.

[0152] Figure 9 The classification performance of various ablation variants under different channel conditions is demonstrated. It can be observed that the multi-view related model significantly outperforms the single-view route in all channel modes. This indicates that multi-view signal representation has stronger stability in dealing with feature distortions caused by complex channels. Under noisy, multipath, and Doppler channel conditions, the intra-class distribution of signal features is more prone to diffusion, leading to a decline in the performance of models based on fixed distance metrics. In contrast, the model incorporating CVAM achieves higher recognition accuracy under the above channel modes, indicating that the adaptive metric method with intra-class variance awareness can effectively alleviate the problem of increased intra-class differences under complex channel conditions. Meanwhile, the model incorporating QAMF brings stable performance improvements under multipath, Doppler, and random mixed channel conditions compared to the model relying solely on multi-views, indicating that the query-related multi-view fusion mechanism can dynamically adjust view weights according to the channel characteristics of different samples, thereby more effectively utilizing view information that is more critical to the current sample discrimination. In summary, the model proposed in this invention achieves optimal or near-optimal performance in all channel modes, verifying the complementarity and synergistic effect of multi-view representation, adaptive metric, and query-related fusion mechanism in complex channel environments.

[0153] In this embodiment, the present invention conducted experiments on a dataset generated based on the RadioML2016.10A data generation framework. The detailed parameters of the dataset are shown in the table.

[0154]

[0155] Table 1

[0156] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A small-sample automatic modulation recognition method for complex channel environments, characterized in that, Includes the following steps: (1) Signal reception; The receiver receives the wireless signal from the transmitter and samples it to obtain a complex baseband signal. , is represented as: ； in, and The first In-phase and quadrature components at each sampling point , This represents the total number of sampling points; (2) Multi-view signal representation and independent feature encoding to obtain the embedding space features of the four views; (3) Construct a small sample modulation recognition task; Using the N-way K-shot method, a few-shot modulation recognition task is constructed based on the embedding spatial features obtained in step (2): In each few-shot modulation recognition task, randomly select Each modulation type is selected. Selecting baseband signals as samples of the support set Each baseband signal is used as a sample of the query set. The samples between the query set and the support set are not repeated. All samples are used to obtain the embedding space features of the four views through step (2). (4) Adaptive measurement based on intra-class variance perception; 4.1) For the first The first of the modulation types The first support set sample The embedded spatial features of a view after being mapped by the encoder are denoted as follows: ,in , , ; 4.2) Calculate the first All support set samples of the modulation type in the first... Mean of embedding spatial features under each view: ； 4.3) Calculate the first Intra-class variance of all support set samples for each modulation type on each feature dimension: ； in, Representing embedded spatial features The dimensional features, Represents the mean of the embedding space features The dimensional features, , Indicates the number of dimensions; 4.4) Construct a dimension-adaptive weight vector based on within-class variance: ； in, A view-dependent learnable scaling factor used to adjust the strength of the effect of variance on the weight distribution; 4.5) Calculate the first The query sample and the first The modulation type in the first Prototype distance under each view: ； in, Indicates the first The query sample in the 1st The first view under the first view 3D embedding space features; (5) Query-related multi-view distance attention fusion; 5.1) Constructing the first The query sample and the first View distance vector for each modulation type : ； 5.2) First, a lightweight attention network is used. For distance vector Perform mapping to obtain the distance vector. Then use The function will divide the distance vector Mapped to a probability distribution, i.e., a distance weight vector : ； in, Indicates the first The modulation type of the first The query sample in the 1st The distance weights under each view range from 0 to 1, and the sum of the distance weights under all four views is 1. 5.3) Calculate the first The query sample and the first The final fusion distance for each modulation type prototype is: ； in Indicates the first The query sample in the 1st The modulation type is the first Distance weights under each view; 5.4) Calculate the first The actual modulation type label of each query sample For the first The actual modulation type label for each modulation type The probability of: ； 5.5) Calculate the loss function for each few-sample task. : ； 5.6) Construct an automatic modulation recognition model based on steps (3) to (5.5); 5.7) Construct a baseband signal training set and a baseband signal validation set to optimize the automatic modulation recognition model in multiple episodes: In each episode, first obtain the loss function based on the baseband signal training set according to steps (3) to 5.5). According to the loss function The automatic modulation recognition model is optimized, and then the loss function is obtained according to steps (3) to (5.5) based on the baseband signal validation set. And the loss function obtained in the previous round based on the baseband signal validation set. The comparison is performed. If the difference is smaller, the optimized automatic modulation recognition model is retained; otherwise, the unoptimized automatic modulation recognition model is used for the next round of optimization. 5.8) Repeat step 5.7) for multiple epochs to obtain the trained automatic modulation recognition model; (6) Automatic modulation and recognition of small samples; Construct a few-sample modulation recognition task based on the embedding spatial features obtained in step (2): Select Each modulation type is selected. Using baseband signals as samples in the support set, the signals belonging to... A baseband signal with one modulation type but an unknown specific modulation type is used as a query sample. The true modulation type label of the query sample is calculated according to steps (4) to (5.5). For the first The actual modulation type label for each modulation type The probability of the modulation type is the highest, and the modulation type with the highest probability is the modulation type of the baseband signal whose specific modulation type is unknown.

2. The small-sample automatic modulation recognition method for complex channel environments according to claim 1, characterized in that, The aforementioned multi-view signal representation and independent feature encoding yield the following embedding space features for the four views: 2.1) Transmit the baseband signal Perform time-domain processing to obtain the waveform view: ； in, Represents a waveform view; 2.2) Baseband signal The in-phase component I and the quadrature component Q of each IQ sampling point are mapped as complex plane coordinates to obtain a constellation view: ； in, Represents a constellation view; 2.3) Baseband signal Convert to polar coordinates, i.e., baseband signal The in-phase component I and quadrature component Q are converted into amplitude components and phase components: ； ； Then based on the amplitude sequence With phase sequence Constructing an amplitude-phase view: ； in, Indicates amplitude-phase view; 2.4) Baseband signal The in-phase component I and the quadrature component Q are respectively estimated using the Welch method: ； ； in, For window function, baseband signal Divided into Segments, each segment is [length missing] , , The first The first paragraph In-phase and quadrature components at each sampling point This represents the energy normalization term of the window function. Indicates a discrete frequency index; Based on in-phase component power spectral density Power spectral density of orthogonal components Constructing a frequency domain power spectral density view: ； in, This represents a frequency domain power spectral density view. 2.5) For each view An independent feature encoder was designed. ,in, This represents the set of parameters of the encoder, the feature encoder. The network structure is completely identical, with no shared parameters and weights. Each view is mapped to the feature space via a feature encoder. ； in, For the first A view Embedded space features obtained by mapping through a feature encoder.

Images