An automatic modulation recognition method based on time convolution and attention collaborative modeling

Abstract

The application provides an automatic modulation recognition method based on time sequence convolution and attention collaborative modeling, which is used for solving the problems of insufficient time sequence feature expression of modulation signals, difficult modeling of long-range dependence and difficult self-adaptive adjustment of feature reliability in a complex wireless channel environment. The method takes a baseband I / Q time sequence signal as input, and realizes collaborative extraction of local and global time sequence features through three modules of time sequence feature enhancement, mixed time sequence modeling and time sequence correlation modeling, in combination with a deep separable convolution, a multi-scale dilated convolution, a multi-head self-attention and a channel weighting mechanism, and introduces time difference feature enhancement to enhance dynamic perception capability. Experiments show that the method has high and stable recognition accuracy on RadioML2016.10A and RadioML2016.10B data sets, and especially shows stronger robustness under low signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the fields of communication technology and deep learning, specifically to an automatic modulation recognition method based on temporal convolution and attention co-modeling. Background Technology

[0002] With the continuous growth in the number of wireless communication devices, limited spectrum resources are becoming increasingly scarce, and signal perception and management in complex electromagnetic environments face higher demands. Automatic modulation identification (AMI) technology, as a crucial link connecting signal detection and signal processing, plays a key role in applications such as spectrum monitoring, wireless management, and intelligent reception, and is an important technological foundation for achieving efficient spectrum utilization and intelligent communication systems. In real-world communication scenarios, wireless channel conditions are complex and variable, and signal waveforms and modulation methods exhibit diverse and dynamic characteristics. This makes modulation identification tasks more challenging under the influence of noise interference, channel fading, and other factors. How to accurately characterize the temporal characteristics of signals in complex environments and effectively distinguish different modulation methods has become an urgent problem to be solved in the field of automatic modulation identification.

[0003] Existing automatic modulation recognition methods often focus on extracting temporal structures at a single scale or in a single form during the temporal feature modeling process. This makes it difficult to simultaneously consider both the fine-grained features of the signal within a local time range and the overall structural characteristics during continuous time evolution, thus limiting their ability to represent complex modulation signals. Furthermore, in complex channel environments, different feature channels contribute differently to the modulation recognition results. Existing methods typically lack adaptive adjustment mechanisms for feature reliability, easily introducing redundant or unstable features, thereby affecting the robustness and stability of the recognition results. Therefore, it is essential to construct an automatic modulation recognition method that can hierarchically characterize local and continuous temporal structures, effectively model long-term temporal correlations, and comprehensively handle feature reliability and temporal dynamic changes, all while utilizing only the received baseband signal. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic modulation recognition method based on temporal convolution and attention co-modeling. By integrating local continuous structures and global temporal correlations during the temporal feature modeling process, and processing the changing relationships of temporal features before global feature aggregation, the method improves the ability to express the temporal characteristics of modulation signals, thereby enhancing the overall accuracy of modulation recognition and strengthening the model's adaptability and stability under different channel conditions and noise environments.

[0005] An automatic modulation recognition method based on temporal convolution and attention co-modeling, the construction method comprising the following steps: Step 1: Obtain I / Q signal samples from the wireless communication system and process the I / Q signal samples to obtain basic timing features; Step 2: Input the basic time series features into the time series feature enhancement module to perform local time series feature enhancement and channel information interaction on the basic time series features to obtain a more robust local time series feature representation; Step 3: Input the local temporal feature representation into the hybrid temporal feature modeling module, and mine the continuous time-dependent features in the baseband temporal signal through multi-scale local temporal feature extraction and continuous temporal modeling; Step 4: Input the features after continuous time series modeling into the time series correlation modeling module. Model the correlation between features at different times through the self-attention mechanism to realize cross-time step information interaction, thereby effectively capturing long-range dependencies. Further perform channel weighted adjustment to enhance the feature channels that are discriminative for modulation recognition. Step 5: The temporal feature enhancement module, the hybrid temporal feature modeling module, and the temporal correlation modeling module are stacked alternately, and scale matching is performed on the feature representation during the stacking process to obtain multi-level temporal feature representations; Step 6: Perform time difference processing on the multi-level temporal feature representation to characterize the temporal change relationship of the temporal features and use it for time difference feature enhancement. Then, perform global feature aggregation on the differentially processed features and output the recognition result of the modulation mode based on the aggregated features.

[0006] Furthermore, the hybrid temporal feature modeling in step 3 includes: Three temporal convolution operations with different kernel scales are employed to extract local temporal features at different time scales. The temporal features obtained at different scales are then fused to form the multi-scale local temporal feature representation. The multi-scale local temporal features are processed by depthwise separable temporal convolutions with different dilation rates, and multi-scale continuous temporal modeling is achieved by layer-by-layer residual stacking. This allows for the characterization of the temporal dependencies between adjacent and cross-time scales in the baseband temporal signal without changing the feature time length.

[0007] Furthermore, the temporal correlation modeling in step 4 includes: By using attention-weighted modeling to model the correlation between features at different times, information interaction across time steps is achieved, thereby effectively capturing long-range dependencies.

[0008] The attention-weighted modeling satisfies the following relationship:

[0009] ; ; in, This represents the input time-series feature matrix. These are the trainable feature mapping weight matrices; These represent the query matrix, key matrix, and value matrix obtained through linear mapping, respectively. This represents the temporal feature representation after multi-head self-attention weighted fusion; The above attention calculation process can be executed in parallel in multiple attention subspaces, and the outputs of each subspace are concatenated and linearly mapped to form the final feature representation; The fused temporal feature channels are adaptively adjusted to enhance the feature channels that are discriminative for modulation scheme identification; The channel weighted adjustment satisfies the following relationship: ; in, This represents the temporal features after fusion; Represents the channel weight vector; This indicates a channel-by-channel weighted operation.

[0010] Furthermore, the alternating stacking process in step 5 includes: By combining the channel-adjusted temporal feature representation with convolutional feature extraction, and through multiple feature extractions and temporal modeling updates, a multi-level temporal feature representation is formed.

[0011] Furthermore, the time difference processing in step 6 includes: Temporal difference processing is performed before global feature aggregation to model the temporal variation relationships of multi-level temporal feature representations, thereby highlighting the change information of features at adjacent time points.

[0012] The time difference processing satisfies the following relationship: ;in, The temporal characteristics at time t are represented. This represents the corresponding time difference feature.

[0013] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention combines multi-scale temporal convolution with continuous temporal modeling mechanism to perform hierarchical modeling of baseband I / Q temporal signals in wireless communication system. It can fully extract the local structural features and continuous time-dependent features of modulation signals at different time scales. Compared with modulation recognition methods that rely only on a single convolutional structure or recurrent network, this invention improves the ability to express the temporal characteristics of complex modulation signals while maintaining the stability of the model structure.

[0014] (2) This invention introduces a hybrid temporal feature modeling method. First, it extracts and fuses local temporal features at different time scales through multi-scale convolution. Then, it combines depthwise separable convolution with different dilation rates and layer-by-layer residual superposition to achieve multi-scale continuous temporal modeling. This effectively enhances the model's ability to model temporal dependencies at different time scales and avoids the information decay problem that is easy to occur when modeling traditional temporal models. This allows the model to more comprehensively depict the overall change law of the baseband modulation signal in the time dimension while maintaining the feature time length.

[0015] (3) The present invention further introduces a channel weighting adjustment mechanism after the fusion of temporal features to adaptively adjust the importance of different feature channels, thereby enhancing the response of feature channels that are discriminative to modulation mode discrimination, which helps to suppress the interference of redundant features and noise features on the recognition results and improve the discriminability and stability of feature representation in the modulation recognition process.

[0016] (4) Before global feature aggregation, the present invention introduces time difference processing on multi-level time sequence features to characterize the feature changes at adjacent time moments and highlight the transient information of the modulation signal to enhance the dynamic perception capability of the model. The time difference operation is used to characterize the relationship between the feature changes at adjacent time moments, so that the difference features and the original features form a complementary time sequence representation, thereby improving the accuracy of modulation mode recognition. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an automatic modulation recognition method based on temporal convolution and attention collaborative modeling according to the present invention. Figure 2 This is a schematic diagram of the overall network architecture of the present invention; Figure 3 This is a schematic diagram of the Hybrid Temporal Feature Modeling Module (HTCM) of the present invention; Figure 4 This is a line graph showing the accuracy of the signal-to-noise ratio obtained by the method of the present invention on the test set RML2016.10A. Figure 5 This is the confusion matrix obtained by testing the method of the present invention on the test set. Figure 5 (a) and Figure 5 (b) RML2016.10A and RML2016.10B respectively; Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of protection of this application.

[0019] like Figures 1-2 As shown, this invention provides an automatic modulation recognition method based on temporal convolution and attention co-modeling, specifically including the following steps: Step 1: Obtain the dataset and divide it into training, validation, and test sets; In this embodiment, the publicly available wireless communication modulation signal datasets RadioML2016.10A and RadioML2016.10B are used as example data sources. RadioML2016.10A contains baseband I / Q signal samples corresponding to 11 digital modulation methods, covering signal-to-noise ratio conditions from -20dB to 18dB, with 220,000 signal samples. RadioML2016.10B contains baseband I / Q signal samples corresponding to 10 digital modulation methods, covering signal-to-noise ratio conditions from -20dB to 18dB, with 1,200,000 signal samples. The I / Q signals in the datasets, which are in the shape of 2x128, are used as model inputs. The datasets are divided into training set, validation set, and test set in a ratio of 6:2:2.

[0020] Step 2: Build the network model; A network model based on the collaborative modeling of temporal convolution and attention is constructed, including a temporal feature enhancement module, a hybrid temporal feature modeling module, a temporal correlation modeling module, and a temporal difference feature enhancement module. These modules work together to form an end-to-end deep learning model. The temporal feature enhancement module adopts a lightweight depthwise separable convolution and residual connection architecture, which consists of a 1D deep convolutional layer, a normalization layer, two 1D pointwise convolutional layers, a GELU activation function, a random depth layer, and optional layer scaling parameters. The deep convolutional layer is set with the same number of groups as the number of input channels, and equal-length padding is used to ensure that the dimensions of the input and output features remain unchanged. The Hybrid Temporal Feature Modeling (HTCM) module is composed of a local representation submodule and an dilated temporal convolutional network submodule connected in series, such as... Figure 3 As shown, the local representation submodule uses multiple convolutional kernel scales to process the input features in parallel to obtain local temporal features at different time scales. Let the set of convolution kernel scales be: ;

[0021] Where K represents the set of convolutional kernel scales, used to characterize the local temporal feature extraction methods at different time scales; The corresponding multi-scale convolution output is: ; in This represents the intermediate temporal feature tensor obtained after processing by the previous temporal feature enhancement module. This represents a one-dimensional temporal convolution operator with a kernel size of k, used for local temporal feature extraction of input features along the time dimension; The features obtained at different scales are then fused to obtain local multi-scale features: ; Here, Fuse() represents the feature fusion operation, which is used to fuse local temporal features obtained at different scales. The fusion method is summation and averaging.

[0022] Dilated temporal convolutional network submodules stack multiple depthwise separable convolutional units with different dilation rates. Let the set of dilation rates be: ; Where D represents the set of dilation rates, which is used to control the size of the receptive field of the dilated convolution in the time dimension; No. The calculation process of each dilated convolutional unit is represented as follows: ; in, Indicates the first Temporal feature representation of the input of each dilated convolutional unit; Indicates the expansion rate A one-dimensional depthwise separable dilated convolution operator is used to extract local and cross-timescale temporal features along the time dimension while keeping the number of feature channels constant. This represents a one-dimensional pointwise convolution operator, used to perform channel-dimensional information interaction and fusion on features after depth convolution. This represents a non-linear mapping function used to perform feature transformation on the convolution result. The non-linear mapping function includes normalization, GELU activation function, and Dropout operation.

[0023] The temporal correlation modeling module consists of an AMC multi-head self-attention submodule and a channel gating submodule connected in series and combined with residual connections.

[0024] Among them, the AMC multi-head self-attention submodule generates query feature Q, key feature K and value feature V through linear mapping and divides them into multiple attention heads; In each attention head, attention weights between different temporal positions are calculated through scaled dot product operations, and feature values ​​are weighted and fused based on the attention weights to obtain a temporal feature representation that integrates multiple temporal information.

[0025] The attention-weighted modeling satisfies the following relationship: ; ; in, This represents the input time-series feature matrix. These are the trainable feature mapping weight matrices; These represent the query matrix, key matrix, and value matrix obtained through linear mapping, respectively. This represents the temporal feature representation after multi-head self-attention weighted fusion; The channel gating submodule first calculates the mean of the temporal dimension, then generates channel weights through two linear layers and a ReLU-activated MLP. After Sigmoid activation, the features are adjusted by channel dimension weighting. Finally, the module superimposes the attention-processed features with the original features through random depth residuals, keeping the input and output feature dimensions [B,T,C] unchanged. By using the channel gating submodule to adaptively weight and adjust the feature channel weights, the channel features that are discriminative for modulation identification are enhanced, and invalid channel interference is suppressed. The channel weighted adjustment satisfies the following relationship: ; in, This represents the temporal features after fusion; Represents the channel weight vector; ⊙ represents the channel-by-channel weighting operation; Finally, the original feature information is preserved through residual connections, which improves the model's ability to represent the temporal correlation features of the modulated signal and provides more discriminative temporal features for subsequent recognition tasks. The temporal difference feature enhancement module, with an optional enabled temporal difference branch as its core, first performs adjacent time-series difference operations along the time dimension on the temporal features of shape [B,C,T] to characterize the dynamic changes of the temporal features. Its temporal difference processing satisfies the following relationship: ; in, The temporal characteristics at time t are represented. This represents the corresponding time difference feature.

[0026] The difference features of [B,C,T-1] are obtained through the above difference operation. After zero-padding to align the temporal length of the original features, they are transformed by the projection layer and then weighted with the learnable scaling coefficients and superimposed on the original features. The features with fused differential information are then fed into a global average pooling layer to obtain global features of [B,C]. Classification prediction values ​​are generated by the classification head. The module supports returning classification results and enhanced temporal features. By introducing temporal difference operations to characterize the dynamic changes of temporal features, the differential features are weighted and fused into the original features to enhance the expressive power of temporal features, making the classification prediction results generated by the model more accurate.

[0027] Step 3: Set the parameters required for training, and train and optimize the model using the training and validation sets; In this embodiment, a label-smoothing cross-entropy loss function is used as the classification loss function, with the label smoothing coefficient set to 0.1 to alleviate overfitting and improve generalization ability in low signal-to-noise ratio scenarios. The AdamW optimization algorithm is selected as the optimizer, with an initial learning rate of 0.004. Simultaneously, an adaptive learning rate adjustment strategy based on validation loss is introduced. When the validation loss no longer decreases after several consecutive training rounds, the learning rate is decayed according to a preset ratio, with a learning rate decay factor set to 0.5 and a minimum learning rate limit of 5 × 10⁻⁻⁶. 5 .

[0028] Step 4: Determine if training is complete; During model training, the completion of training is determined by the performance changes on the validation dataset. If the model's validation performance does not improve significantly after 10 consecutive training cycles, the training process is terminated early. Alternatively, the model training is terminated when the number of training cycles reaches the preset maximum number of training cycles.

[0029] Step 5: Test the final modulation recognition model using the test set, and plot the accuracy curve and confusion matrix; Figure 4 This presents a comparison of modulation recognition accuracy of the method of this invention under different signal-to-noise ratio (SNR) environments on the RML2016.10a dataset test set conditions. Figure 4 It can be seen that the method proposed in this invention exhibits relatively stable recognition performance and outperforms the comparative methods across the entire signal-to-noise ratio range. In the low signal-to-noise ratio range (e.g., −20dB to −6dB), the recognition accuracy of the method proposed in this invention is significantly higher than that of the comparative methods such as FEA-T, MCLDNN, MCDformer, PETCGDNN, and AWN. This indicates that the temporal correlation modeling and feature enhancement mechanism adopted in this invention can effectively extract discriminative modulation features under strong noise interference and has better robustness. As the signal-to-noise ratio gradually increases, the recognition accuracy of each method shows an upward trend, but the method proposed in this invention still maintains a leading advantage in the medium-to-high signal-to-noise ratio range.

[0030] Figure 5 (a) and Figure 5 (b) are schematic diagrams showing the modulation recognition confusion matrices obtained by the method of the present invention on the test sets divided from the RML2016.10a and RML2016.10b datasets under full signal-to-noise ratio conditions; Figure 5 (a) and Figure 5 (b) It can be seen that the method of the present invention can achieve relatively accurate identification on most modulation types under different dataset conditions. The overall classification results are mainly concentrated in the diagonal region of the confusion matrix, indicating that the present invention has good modulation discrimination ability. Some modulation methods still exhibit a certain degree of confusion in low signal-to-noise ratio environments, especially between higher-order modulation methods such as QAM16 and QAM64, and between analog modulation types such as AM-DSB and WBFM. This confusion mainly stems from the similar signal characteristics of related modulation methods in noisy environments, which easily leads to cross-discrimination.

[0031] exist Figure 5 (a) and Figure 5 (b) shows that 8PSK, BPSK, CPFSK, GFSK, PAM4, QPSK and some QAM modulation methods can maintain a high correct recognition probability, indicating that the method of the present invention has good adaptability under different channel environments and sample distribution conditions. The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

[0032] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. An automatic modulation recognition method based on temporal convolution and attention co-modeling, the construction method comprising the following steps: Step 1: Obtain I / Q signal samples from the wireless communication system and process the I / Q signal samples to obtain basic timing features; Step 2: Input the basic time series features into the time series feature enhancement module to perform local time series feature enhancement and channel information interaction on the basic time series features to obtain a more robust local time series feature representation; Step 3: Input the local temporal feature representation into the hybrid temporal feature modeling module, and mine the continuous time-dependent features in the baseband temporal signal through multi-scale local temporal feature extraction and continuous temporal modeling; Step 4: Input the features after continuous time series modeling into the time series correlation modeling module. Model the correlation between features at different times through the self-attention mechanism to realize cross-time step information interaction, thereby effectively capturing long-range dependencies. Further perform channel weighted adjustment to enhance the feature channels that are discriminative for modulation recognition. Step 5: The temporal feature enhancement module, the hybrid temporal feature modeling module, and the temporal correlation modeling module are stacked alternately, and scale matching is performed on the feature representation during the stacking process to obtain multi-level temporal feature representations; Step 6: Perform time difference processing on the multi-level temporal feature representation to characterize the temporal change relationship of the temporal features and use it for time difference feature enhancement. Then, perform global feature aggregation on the differentially processed features and output the recognition result of the modulation mode based on the aggregated features.

2. Further, the hybrid temporal feature modeling in step 3 includes: Three temporal convolution operations with different kernel sizes are used to extract local temporal features at different time scales, and the temporal features obtained at different scales are fused to form the multi-scale local temporal feature representation. Multi-scale local temporal features are processed by depthwise separable temporal convolutions with different dilation rates, and multi-scale continuous temporal modeling is achieved by layer-by-layer residual superposition. This allows us to characterize the temporal dependencies between adjacent and cross-time scales in the baseband temporal signal without changing the feature time length.

3. Further, the temporal correlation modeling in step 4 includes: By using attention-weighted modeling to model the correlation between features at different times, information interaction across time steps is achieved, thereby effectively capturing long-range dependencies.

4. Among them, The attention-weighted modeling satisfies the following relationship: ; Where X represents the input time-series feature matrix, , , These are the trainable feature mapping weight matrices; These represent the query matrix, key matrix, and value matrix obtained through linear mapping, respectively. This represents the temporal feature representation after multi-head self-attention weighted fusion; The above attention calculation process can be executed in parallel in multiple attention subspaces, and the outputs of each subspace are concatenated and linearly mapped to form the final feature representation; The fused temporal feature channels are adaptively adjusted to enhance the feature channels that are discriminative for modulation scheme identification; The channel weighted adjustment satisfies the following relationship: ;in, This represents the temporal features after fusion; Represents the channel weight vector; This indicates a channel-by-channel weighted operation.

5. Further, the alternating stacking process in step 5 includes: By combining the channel-adjusted temporal feature representation with convolutional feature extraction, and through multiple feature extractions and temporal modeling updates, a multi-level temporal feature representation is formed.

6. Further, the time difference processing in step 6 includes: Temporal difference processing is performed before global feature aggregation to model the temporal variation relationships of multi-level temporal feature representations, thereby highlighting the change information of features at adjacent time points.

7. The time difference processing satisfies the following relationship: ;in, The temporal characteristics at time t are represented. This represents the corresponding time difference feature.

Images