A signal intelligent modulation recognition method and system fusing time-frequency features

Abstract

The application discloses a signal intelligent modulation recognition method and system fusing time-frequency characteristics. The method extracts time domain characteristics and frequency domain characteristics in a radio signal in a complex baseband form. The radio signal, the time domain characteristics and the frequency domain characteristics are fused to obtain multi-channel fusion characteristics. A convolution attention network model containing a stem embedding layer, multiple cascaded feature extraction stages and a classification head is trained. The multi-channel fusion characteristics are input into the trained convolution attention network model. The stem embedding layer extracts features from the multi-channel fusion characteristics to obtain first extraction features. The multiple cascaded feature extraction stages extract deep features from the first extraction features to obtain second extraction features. The second extraction features are input into the classification head for signal modulation recognition to obtain a signal modulation recognition result corresponding to the radio signal. The application can improve the accuracy of signal modulation recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a signal intelligent modulation recognition method and system fusing time-frequency features. BACKGROUND

[0002] Automatic modulation recognition (AMR) technology is mainly divided into traditional methods based on prior rules / features and data-driven methods based on deep learning. Among them, the traditional method includes recognition based on likelihood decision and recognition based on feature parameters. The latter usually first performs feature transformation on the signal, such as time-frequency analysis, constellation mapping, and statistical feature extraction (such as high-order cumulants), etc. The extracted features are then input into a machine learning classifier (such as KNN, SVM, and random forest, etc.) to complete the modulation category decision. This feature engineering method relies on expert experience, and the feature selection may not be complete, which is difficult to represent the subtle differences of the signal, so the overall recognition performance is often not as good as the deep neural network model.

[0003] However, the existing deep neural network model needs to learn complex modulation features such as time-frequency distribution and phase pattern from the original I / Q data, which usually requires a large amount of training data and a deep network structure to fully generalize, resulting in high training cost. It is also not realistic to deploy such a model on a platform with limited computing resources (such as embedded devices). When the signal signal-to-noise ratio decreases, the network has difficulty in effectively extracting discriminative features, and the performance drops sharply, which is particularly evident in models lacking domain priori. And some signal modulation recognition methods based on deep neural network model only use manually extracted features as input, which limits the precision and adaptability of the model.

[0004] Therefore, the existing signal modulation recognition method has a relatively low accuracy in signal modulation recognition. SUMMARY

[0005] The present application aims to provide a signal intelligent modulation recognition method and system fusing time-frequency features, which can improve the accuracy of signal modulation recognition.

[0006] In a first aspect, an embodiment of the present application provides a signal intelligent modulation recognition method fusing time-frequency features, which comprises:

[0007] extracting time-domain features and frequency-domain features in a radio signal in complex baseband form;

[0008] fusing the radio signal, the time-domain features and the frequency-domain features to obtain multi-channel fusion features;

[0009] training a convolutional attention network model containing a stem embedding layer, a plurality of cascaded feature extraction stages and a classification head;

[0010] input the multi-channel fusion feature into the trained convolutional attention network model, and perform feature extraction on the multi-channel fusion feature through the stem embedding layer to obtain first extracted features;

[0011] perform deep feature extraction on the first extracted features through the plurality of cascaded feature extraction stages to obtain second extracted features;

[0012] input the second extracted features into the classification head for signal modulation identification to obtain a signal modulation identification result corresponding to the radio signal.

[0013] Compared with the prior art, the first aspect of the present application has the following beneficial effects:

[0014] The method extracts the time domain features and frequency domain features in the radio signal in complex baseband form, fuses the radio signal, the time domain features and the frequency domain features to obtain multi-channel fusion features, injects the time domain features and the frequency domain features and other physical priors, and aligns and fuses the radio signal by time to obtain the multi-channel fusion features as model input, which can reduce the learning difficulty of the convolutional attention network model, accelerate the convergence of the model, and make the convolutional attention network model maintain higher discriminant stability under medium and low signal-to-noise ratio (SNR) conditions. The convolutional attention network model including the stem embedding layer, the plurality of cascaded feature extraction stages and the classification head is trained, the multi-channel fusion feature is input into the trained convolutional attention network model, the stem embedding layer is used to extract features from the multi-channel fusion feature to obtain first extracted features, the plurality of cascaded feature extraction stages are used to perform deep feature extraction on the first extracted features to obtain second extracted features, and the second extracted features are input into the classification head for signal modulation identification to obtain a signal modulation identification result corresponding to the radio signal. By guiding the convolutional attention network model to learn with artificial features (i.e. time domain features and frequency domain features), the artificial features provide direct physical significance information, and the convolutional attention network model focuses on key signal patterns accordingly, which can improve the accuracy of signal modulation identification and suppress noise interference.

[0015] In a second aspect, the embodiments of the present application further provide a signal intelligent modulation identification system fusing time-frequency features, and the system comprises:

[0016] a feature extraction unit configured to extract time domain features and frequency domain features in a radio signal in complex baseband form;

[0017] a feature fusion unit configured to fuse the radio signal, the time domain features and the frequency domain features to obtain multi-channel fusion features;

[0018] The model training unit is configured to train a convolutional attention network model including a stem embedding layer, a plurality of cascaded feature extraction stages, and a classification head.

[0019] The first feature extraction unit is configured to input the multi-channel fusion features into the trained convolutional attention network model, perform feature extraction on the multi-channel fusion features through the stem embedding layer, and obtain first extracted features.

[0020] The second feature extraction unit is configured to perform deep feature extraction on the first extracted features through the plurality of cascaded feature extraction stages, and obtain second extracted features.

[0021] The signal modulation identification unit is configured to input the second extracted features into the classification head to perform signal modulation identification, and obtain a signal modulation identification result corresponding to the radio signal.

[0022] In a third aspect, an electronic device is provided, which includes at least one control processor and a memory in communication connection with the at least one control processor; the memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor to enable the at least one control processor to perform the signal intelligent modulation identification method of fusing time-frequency features as described above.

[0023] In a fourth aspect, a computer readable storage medium is provided, which stores computer executable instructions for causing a computer to perform the signal intelligent modulation identification method of fusing time-frequency features as described above.

[0024] It can be understood that the beneficial effects of the second aspect to the fourth aspect compared with the related art are the same as the beneficial effects of the first aspect compared with the related art, and reference can be made to the related description in the first aspect, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0025] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0026] Figure 1 is a flowchart of an embodiment of the signal intelligent modulation identification method of fusing time-frequency features provided by the present application;

[0027] Figure 2 is a feature extraction and fusion flowchart in the best embodiment of the signal intelligent modulation identification method of fusing time-frequency features provided by the present application;

[0028] Figure 3is a convolution attention network model structure schematic diagram in the best embodiment of the signal intelligent modulation recognition method provided in the application that fuses time-frequency features;

[0029] Figure 4 is a stem embedding layer structure schematic diagram in the best embodiment of the signal intelligent modulation recognition method provided in the application that fuses time-frequency features;

[0030] Figure 5 is a deep separable convolution structure schematic diagram in the best embodiment of the signal intelligent modulation recognition method provided in the application that fuses time-frequency features;

[0031] Figure 6 is a convolution embedding layer structure schematic diagram in the best embodiment of the signal intelligent modulation recognition method provided in the application that fuses time-frequency features;

[0032] Figure 7 is a convolution double attention module structure schematic diagram in the best embodiment of the signal intelligent modulation recognition method provided in the application that fuses time-frequency features;

[0033] Figure 8 is a classification recognition accuracy comparison result schematic diagram of different models in each signal-to-noise ratio in the best embodiment of the signal intelligent modulation recognition method provided in the application that fuses time-frequency features;

[0034] Figure 9 is an ablation experiment recognition accuracy comparison result schematic diagram in each signal-to-noise ratio in the best embodiment of the signal intelligent modulation recognition method provided in the application that fuses time-frequency features;

[0035] Figure 10 is a structure schematic diagram of an embodiment of the signal intelligent modulation recognition system provided in the application;

[0036] Figure 11 is a structure schematic diagram of an embodiment of the electronic device provided in the application. DETAILED DESCRIPTION

[0037] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numbers represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary, only for explaining the present application, and cannot be understood as limiting the present application.

[0038] In the description of the present application, if there is a description to first, second, etc., it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or implicitly indicating the sequence of indicated technical features.

[0039] In the description of the present application, it should be understood that the orientation description, such as the orientation or position relationship indicated by up, down, etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0040] In the description of the present application, it should be noted that, unless otherwise explicitly limited, the words such as setting, installing, connecting, etc. should be understood in a broad sense, and the person skilled in the art can reasonably determine the specific meaning of the above words in the present application in combination with the specific content of the technical solution.

[0041] Firstly, some terms involved in the present application are analyzed:

[0042] Automatic Modulation Recognition (AMR): refers to automatically identifying the modulation mode adopted by the received signal at the receiving end.

[0043] Dual-Attention Mechanism: a mechanism that combines two types of attention. In the present embodiment, it refers to applying "time sequence self-attention" (extracting global time dependence) and "channel attention" (extracting the importance of each channel) at the same time to improve feature extraction effect. In the present embodiment, channel attention and other ideas are also used in residual blocks and feature fusion.

[0044] Feature Engineering: refers to the process of analyzing, processing and feature construction on original data before model training. By extracting statistical quantities, time domain features, frequency domain features or morphological features, etc., the separability of data and the learning efficiency of the model are improved. In the present embodiment, time domain and frequency domain statistical features of I / Q signals are extracted through feature engineering, and deep network learning features are combined to achieve better classification effect.

[0045] Short-time Fourier Transform (STFT): a method of converting a signal from time domain to time-frequency domain, which cuts a long signal into short segments and does Fourier transform on each segment to obtain a spectrum with frequency varying with time. In the present embodiment, STFT with multiple window lengths is implemented to obtain multi-scale frequency domain features.

[0046] Label Smoothing: a regularization technique to improve cross-entropy loss. By "smoothing" the one-hot distribution of the true label, i.e. assigning a very small probability to non-target classes, the model's overconfidence is avoided, and the generalization ability is improved. Label smoothing is often used in classification tasks to reduce the risk of overfitting.

[0047] Instantaneous frequency: a quantity that describes the instantaneous oscillation rate of a signal at a certain time. Mathematically, it is the derivative of the signal phase with respect to time, reflecting the local frequency change of the signal. Instantaneous frequency can be used to analyze frequency-modulated signals (such as FSK, FM), revealing the frequency structure characteristics of the signal over time.

[0048] Amplitude envelope: refers to the outer envelope of the signal amplitude change over time, reflecting the trend of signal energy change in the time dimension. The amplitude envelope is often obtained by Hilbert transform, which can be used to describe the intensity change of amplitude-modulated signals (such as AM) or fading signals, and is an important index in time-domain feature extraction.

[0049] To solve the problems existing in the prior art, the present application provides a signal intelligent modulation recognition method and system fusing time-frequency features.

[0050] Reference Figure 1 The signal intelligent modulation recognition method fusing time-frequency features provided by the embodiments of the present application has the flowchart as shown in the figure. The signal intelligent modulation recognition method fusing time-frequency features is applied to an electronic device, which can be a server or a mobile terminal, etc. As shown in the figure, the signal intelligent modulation recognition method fusing time-frequency features can include the following steps: Figure 1

[0051] Step S101, extract the time-domain features and frequency-domain features in the complex baseband form of the radio signal;

[0052] Step S102, fuse the radio signal, the time-domain features and the frequency-domain features to obtain multi-channel fusion features;

[0053] Step S103, train a convolution attention network model containing a stem embedding layer, a plurality of cascaded feature extraction stages and a classification head;

[0054] Step S104, input the multi-channel fusion features into the trained convolution attention network model, and extract features of the multi-channel fusion features through the stem embedding layer to obtain first extraction features;

[0055] Step S105, perform deep feature extraction on the first extraction features through the plurality of cascaded feature extraction stages to obtain second extraction features;

[0056] Step S106, input the second extraction features into the classification head for signal modulation recognition to obtain a signal modulation recognition result corresponding to the radio signal.

[0057] ​In the embodiment, by extracting the time domain features and the frequency domain features in the radio signal in complex baseband form, fusing the radio signal, the time domain features and the frequency domain features, obtaining the multi-channel fusion features, by injecting the time domain features and the frequency domain features and the radio signal, aligning and fusing them into the multi-channel fusion features as the model input, the learning difficulty of the convolutional attention network model can be reduced, the model convergence can be accelerated, and the convolutional attention network model can maintain higher discriminant stability under the condition of medium and low signal-to-noise ratio (SNR). The convolutional attention network model including a stem embedding layer, a plurality of cascaded feature extraction stages and a classification head is trained, the multi-channel fusion features are input into the trained convolutional attention network model, the stem embedding layer is used to extract features of the multi-channel fusion features to obtain first extraction features, the plurality of cascaded feature extraction stages are used to extract deep features of the first extraction features to obtain second extraction features, the second extraction features are input into the classification head for signal modulation recognition to obtain a signal modulation recognition result corresponding to the radio signal. By guiding the convolutional attention network model to learn through artificial features (i.e. time domain features and frequency domain features), the artificial features provide direct physical meaning information, and the convolutional attention network model focuses on key signal patterns faster accordingly, so that the accuracy of signal modulation recognition can be improved and noise interference can be suppressed.

[0058] The time domain features can include amplitude envelope and instantaneous frequency.

[0059] The frequency domain features can include front half spectrum average amplitude sequence and rear half spectrum average amplitude sequence.

[0060] The plurality of cascaded feature extraction stages can be that the output result of a previous feature extraction stage is used as the input of a next cascaded feature extraction stage, so that more abstract features are extracted continuously.

[0061] The classification head can be used to calculate the scores of each modulation category, and can adopt a network structure known to those skilled in the art, or can be a classification head composed of global pooling and multilayer perceptron (i.e. two-layer perceptron).

[0062] The trained convolutional attention network model including the stem embedding layer, the plurality of cascaded feature extraction stages and the classification head can be that a training data set is obtained, a cross-entropy loss function is used to supervise the training of the convolutional attention network model, and the trained convolutional attention network model is obtained.

[0063] In some embodiments, the time domain features and the frequency domain features in the radio signal in complex baseband form are extracted, including:

[0064] The amplitude envelope in the radio signal in complex baseband form is extracted, and the amplitude envelope is the amplitude of the radio signal.

[0065] calculating the instantaneous frequency based on the in-phase component and the quadrature component in the complex baseband form of the radio signal;

[0066] taking the amplitude envelope and the instantaneous frequency as the time-domain features;

[0067] performing a short-time Fourier transform on the radio signal in the complex baseband form to obtain a complex-valued spectrum matrix;

[0068] calculating the average amplitude of the first half spectrum and the average amplitude of the second half spectrum based on the complex-valued spectrum matrix;

[0069] mapping the average amplitude of the first half spectrum and the average amplitude of the second half spectrum back to discrete sequences to obtain a first half spectrum average amplitude sequence and a second half spectrum average amplitude sequence;

[0070] taking the first half spectrum average amplitude sequence and the second half spectrum average amplitude sequence as the frequency-domain features.

[0071] In this embodiment, the time-domain features and the frequency-domain features in the radio signal in the complex baseband form are extracted, which lays a good data foundation for subsequent better guiding the convolution attention network model to learn, because the artificial features (i.e., the time-domain features and the frequency-domain features) provide direct physical meaning information, the convolution attention network model focuses on the key signal patterns faster accordingly, which can improve the accuracy of signal modulation recognition and suppress noise interference.

[0072] The first half spectrum and the second half spectrum can be dividing the spectrum of each time frame into two parts, i.e., the first half spectrum and the second half spectrum.

[0073] In some embodiments, the stem embedding layer includes a one-dimensional convolution, a batch normalization, an activation function, and a depth separable convolution, and the stem embedding layer is used to extract features from the multi-channel fused features to obtain first extracted features, including:

[0074] The one-dimensional convolution is used to preliminarily extract features and upgrade channels from the multi-channel fused features to obtain initial features;

[0075] The batch normalization and the activation function are used to process the initial features to obtain processed features;

[0076] The depth separable convolution is used to extract features and enhance nonlinearity from the processed features to obtain the first extracted features.

[0077] In this embodiment, the depth separable convolution is used to extract features and enhance nonlinearity from the processed features, which can encode the original multi-channel sequence into a higher-dimensional feature representation, and provide a suitable feature space for the subsequent multiple cascaded feature extraction stages.

[0078] In some embodiments, the first extracted feature is subjected to deep feature extraction through a plurality of cascaded feature extraction stages to obtain a second extracted feature, including:

[0079] The plurality of cascaded feature extraction stages are divided into a first feature extraction stage, a second feature extraction stage, and a third feature extraction stage in cascade, wherein the first feature extraction stage includes an identity mapping and two convolutional double attention modules, the second feature extraction stage includes a convolutional embedding layer and two convolutional double attention modules, and the third feature extraction stage includes a convolutional embedding layer and two convolutional double attention modules;

[0080] The first extracted feature is subjected to feature extraction through the identity mapping and the two convolutional double attention modules in the first feature extraction stage to obtain a first feature extraction stage output result;

[0081] The first feature extraction stage output result is subjected to feature extraction through the convolutional embedding layer and the two convolutional double attention modules in the second feature extraction stage to obtain a second feature extraction stage output result;

[0082] The second feature extraction stage output result is subjected to feature extraction through the convolutional embedding layer and the two convolutional double attention modules in the third feature extraction stage to obtain the second extracted feature.

[0083] In the present embodiment, by dividing the plurality of cascaded feature extraction stages into the first feature extraction stage, the second feature extraction stage, and the third feature extraction stage in cascade, the three feature extraction stages can gradually perform temporal downsampling and feature dimensionality increase, so that the network model can extract more abstract and larger receptive field features at each level while maintaining computational efficiency. In combination with the deep neural network structure of "convolutional extraction + double attention fusion", the network model effectively learns the discrimination features of the modulated signal.

[0084] In some embodiments, the first extracted feature is subjected to feature extraction through the identity mapping and the two convolutional double attention modules in the first feature extraction stage to obtain a first feature extraction stage output result, including:

[0085] The result of the identity mapping of the first extracted feature is input into two cascaded convolutional double attention modules for feature extraction to obtain the first feature extraction stage output result, wherein each convolutional double attention module includes a plurality of depth separable convolutions, a temporal dimension self-attention mechanism, a channel dimension attention mechanism, and a feedforward network, and the feature extraction process of each convolutional double attention module includes:

[0086] After the identity-mapped result is subjected to batch normalization processing, a batch-normalized feature is obtained;

[0087] The batch-normalized features are extracted by multiple depth separable convolutions to obtain query features, key features and value features.

[0088] Based on the query features, the key features and the value features, a time dimension self-attention mechanism is used to calculate a time attention output.

[0089] The batch-normalized features are processed by a channel dimension attention mechanism to obtain a channel attention output.

[0090] The time attention output and the channel attention output are fused to obtain a fused attention result.

[0091] The fused attention result is connected in residual connection with the batch-normalized features after a Dropout operation to obtain a residual connection result.

[0092] The residual connection result is input into a feedforward network after batch normalization to obtain a feedforward network output result.

[0093] The feedforward network output result is connected in residual connection with the residual connection result after a Dropout operation to obtain a convolutional double-attention module output result.

[0094] In this embodiment, the batch-normalized features are extracted by multiple depth separable convolutions to obtain query features, key features and value features, and then based on the query features, the key features and the value features, a time dimension self-attention mechanism is used to calculate a time attention output, and the batch-normalized features are processed by a channel dimension attention mechanism to obtain a channel attention output. Thus, this embodiment adopts a double-attention structure combining time self-attention and channel attention, wherein the time self-attention mechanism reduces the influence of irrelevant disturbances on feature extraction, and the channel attention enhances the main mode features and suppresses redundant information, so that the network model can assign higher weights to high discriminative time segments and high contribution channels in the training process, thereby laying a good data foundation for improving the accuracy of signal modulation recognition in the later stage.

[0095] In some embodiments, the channel dimension attention mechanism includes a global average pooling and a channel attention gate unit, and the batch-normalized features are processed by the channel dimension attention mechanism to obtain a channel attention output, including:

[0096] The batch-normalized features are globally averaged in the time dimension to calculate the average value of each channel in the entire time sequence to obtain a channel-level feature vector.

[0097] The channel-level feature vector is input into the channel attention gate unit to obtain a channel attention weight.

[0098] The channel attention weight is expanded in the time dimension to obtain an expanded attention weight;

[0099] The expanded attention weight is multiplied element by element with the batch-normalized feature to obtain a channel attention output.

[0100] In the embodiment, the batch-normalized feature is processed through the channel dimension attention mechanism to dynamically measure the importance of different feature channels, thereby realizing adaptive feature selection, dynamically highlighting effective channels and suppressing redundant features, and having a significant effect, especially in a multi-source feature fusion scenario.

[0101] The channel attention gate unit can include two fully connected layers and a nonlinear activation function.

[0102] In some embodiments, the classification head includes global pooling and a multilayer perceptron, and the second extracted feature is input into the classification head for signal modulation recognition to obtain a signal modulation recognition result corresponding to the radio signal, including:

[0103] The second extracted feature is processed by global pooling to obtain a pooled global feature vector;

[0104] The pooled global feature vector is input into the multilayer perceptron for signal modulation recognition to obtain a prediction score of each modulation category;

[0105] The prediction score of each modulation category is processed by a Softmax function to obtain a probability distribution, and a modulation category corresponding to a maximum probability value in the probability distribution is taken as the signal modulation recognition result corresponding to the radio signal.

[0106] In the embodiment, the second extracted feature is input into the classification head for signal modulation recognition to obtain a signal modulation recognition result corresponding to the radio signal. Since the second extracted feature contains rich detailed features, after global pooling, multilayer perceptron processing, and Softmax function processing, the accuracy of signal modulation recognition can be improved.

[0107] The global pooling can be generalized average pooling to better preserve peak information and enhance the response to key time segments.

[0108] For the convenience of those skilled in the art, a set of best embodiments is provided as follows:

[0109] In the prior art, since the open source dataset RML2016.10a was published in 2016 and the convolutional neural network (CNN) was successfully implemented for end-to-end modulation recognition, AMR based on deep learning has become a research hotspot. A large number of subsequent works try to use more complex network structures (such as deep CNN, recurrent neural network RNN and Transformer, etc.) to improve the recognition accuracy. However, the performance improvement of such models often depends on stacking more layers or increasing the parameter size, resulting in increased computational overhead and storage requirements. For example, RNN sequence models cannot be calculated in parallel, and their inference efficiency is low in long sequences; CNN increases the network size, which may cause parameter explosion, and it is challenging to deploy on resource-limited devices. In order to balance accuracy and efficiency, lightweight convolution and Transformer combined models have begun to appear. Some research has proposed a MCLDNN model that uses pure I / Q signals as input, that is, local features are extracted by convolution and global dependencies are captured by self-attention, achieving high-precision modulation classification. On the open source dataset RML2016.10a, the highest recognition accuracy of the model after standard training is about 92.18% (under high signal-to-noise ratio conditions). At the same time, the parameter size of the model is only about 63,000, which is much lower than the parameter size of the traditional CNN model under the condition of close accuracy. This method can adapt to variable length and incomplete signal input, and has representative progress in current deep learning AMR technology. However, the prior art has the following disadvantages:

[0110] Disadvantage one: pure end-to-end deep neural network is used for modulation recognition, which has high requirements for large datasets and model capacity.

[0111] Deep models need to learn complex modulation features such as time-frequency distribution and phase pattern from raw I / Q signal data, which usually requires a large amount of training data and a deep network structure to fully generalize, resulting in high training costs. It is also not practical to deploy such models on platforms with limited computing resources, such as embedded devices. When the signal-to-noise ratio is low, the network has difficulty extracting discriminative features, and the performance drops sharply, which is particularly evident in models without domain prior.

[0112] Disadvantage two: only artificial feature extraction methods are used, and the accuracy and adaptability of the model are limited.

[0113] The artificially designed feature dimension is relatively low and the information is lossy, for example, high-order cumulants or cyclic spectrum only capture some aspects of the signal characteristics, when the channel condition is complex or the signal is distorted, relying solely on these features may not be enough to distinguish all modulation types. At the same time, different features have different sensitivities to different modulation signals, and fixed feature combinations are difficult to maintain stable and effective in all scenarios. This leads to the performance of traditional methods often lags behind data-driven deep learning methods. The fundamental reason is that artificial features cannot fully describe the details of the signal, and there is a lack of important mechanisms to adaptively adjust the features according to the data.

[0114] Disadvantage three: the prior art lacks an effective solution to combine artificial features with the advantages of deep networks.

[0115] Although intuitively combining "expert knowledge" and "data-driven learning" can complement each other, directly splicing artificial features and raw data may cause problems such as inconsistent feature scales and redundant information interference, leading to unstable or suboptimal deep network training. Therefore, a special fusion strategy and supporting model structure are needed to fully utilize the auxiliary information of artificial features. Currently, there are few documents that use artificial time / frequency domain features and end-to-end deep networks to improve AMR performance, and how to realize the cooperative work of the two is a problem to be solved.

[0116] In view of the above deficiencies of the prior art, the purpose of the embodiment is to improve the accuracy and robustness of automatic modulation recognition, and reduce the dependence of the model on large-scale data and ultra-deep networks. The embodiment realizes high-precision recognition of radio signal modulation mode by fusing artificially extracted signal features and convolutional attention network model. Specifically, the embodiment hopes to introduce domain prior knowledge by using artificial features to assist deep neural networks to learn the distinguishing patterns of signals more effectively, so as to achieve performance comparable to more complex models under a relatively simple model structure.

[0117] To achieve the above-mentioned purpose, the solution path of the embodiment includes the following key steps: first, the I / Q data of the original received signal is preprocessed and feature engineered to extract artificial features reflecting the time domain and frequency domain properties of the signal; then, these artificial features are fused with the original I / Q sequence as the multi-channel input of the deep neural network; finally, a deep network combining convolution and attention mechanism is designed to extract features and make classification decisions on the fused multi-channel data. Specifically, the preprocessing part uses signal processing methods to obtain features such as amplitude envelope, instantaneous frequency and time-frequency energy distribution, and splices them with the original I / Q signal to form an enhanced input signal. Subsequently, the convolutional neural network extracts local pattern (such as waveform shape, burst spike, etc.) features, and the self-attention mechanism captures global correlation (such as symbol sequence pattern, periodic structure, etc.), which complement each other. Through end-to-end training, the model automatically learns how to balance the use of artificial features and original data, ensuring that the introduction of artificial features can improve the performance of modulation recognition from a causal point of view. In summary, the embodiment path guides the deep network model learning with artificial features: artificial features provide direct physical meaning information, and the network model focuses on key signal patterns faster accordingly; this can improve the accuracy of signal modulation recognition and suppress noise interference. The scheme balances between accuracy and model complexity, achieving an innovative technical breakthrough.

[0118] In view of the three core shortcomings of existing automatic modulation recognition technology, i.e., "heavy learning burden (end-to-end model learns prior from zero from pure I / Q data)", "rigid feature utilization (artificial features and original data are difficult to cooperate, inconsistent scales and redundant interference)", and "one-sided representation (only local convolution or only global attention, lacking unified modeling across scales)", the embodiment aims to provide an automatic modulation recognition method of convolutional attention network model fused with time-frequency features, which aims to:

[0119] 1. Solve the problem of "heavy learning burden": existing end-to-end AMR methods use pure I / Q as the only input, lacking explicit physical time-frequency prior and structural hints. To overcome the problem of slow convergence and instability at medium and low SNR caused by the high dependence of pure end-to-end model on data size and network depth when discovering time-frequency structure from original I / Q signal, the embodiment injects amplitude envelope, instantaneous frequency and segmented band energy as physical prior on the input side, and fuses them with the original I / Q signal as multi-channel input according to time alignment, reducing the learning difficulty, accelerating the convergence, and maintaining higher discrimination stability under medium and low SNR conditions.

[0120] 2. Solve the problem of "feature utilization rigidity": break the single path limitation of "only using artificial features" or "only using original I / Q" and the redundant interference of simple splicing, realize the dynamic adaptive utilization of features. Through channel attention gate (dynamically weighting multiple source channels) and channel-by-channel standardization / alignment, the network automatically highlights effective channels and suppresses irrelevant or noise channels during training, realizing the synergistic gain of manual priori and data-driven representation.

[0121] 3. Solve the problem of "one-sided representation": transcend the one-sided representation of "only local convolution" or "only global attention", and unify the modeling of local patterns and long-range dependencies. Adopt a hierarchical structure of phased convolution embedding (including down-sampling and up-channel), time self-attention, residual feedforward, and GeM adaptive pooling to simultaneously capture short-range details and long-range dependencies with small parameter size and controlled computational complexity, improving the ability to distinguish easily confused modulations (such as QAM16 / 64 and AM-DSB / SSB) and overall recognition accuracy.

[0122] The technical solutions of the embodiments specifically include the following contents:

[0123] The automatic modulation recognition method of the convolution attention network model fusing manual time-frequency features proposed in the embodiment has the core of multi-source fusion input of I / Q original channels and artificial time / frequency channels, realizes automatic modulation recognition through "convolution-double attention" hierarchical feature extraction and adaptive aggregation, and systematically solves the shortcomings of the prior art. The implementation principle can be decomposed into the following 2 interrelated innovation modules:

[0124] 1. Input signal and artificial feature extraction.

[0125] The object processed by the embodiment is a complex baseband form of radio signal (i.e. input signal), that is, a discrete sequence containing in-phase component I and quadrature component Q. Let the original radio signal be represented as , the sampling length is (for example, in the embodiment ), , and is the imaginary unit. The embodiment first calculates two types of artificial features: time domain features and frequency domain features for each signal, aiming to describe the modulation mode of the signal from different angles.

[0126] (1) Time domain feature extraction: for the input signal , two features of amplitude envelope and instantaneous frequency are extracted. The amplitude envelope is defined as the modulus of the complex amplitude of the signal, and the calculation formula is:

[0127] (1);

[0128] where , and amplitude of a signal sample (i.e. modulus of the complex amplitude of the signal), representing the branch of the shunt, representing the branch of the shunt. In the present embodiment, the global time index (sample index) within a single signal sequence is referred to uniformly, which can be a positive integer.

[0129] The instantaneous frequency reflects the rate of change of the signal phase with time. First, the complex signal phase of the signal is calculated, and the phase sequence is unwrapped to obtain continuous phase .

[0130] (2);

[0131] wherein, is the continuous phase. It should be noted that the unwrapping process in the present embodiment uses a technique known to those skilled in the art, which is not described in detail in the present embodiment.

[0132] In implementation, the phase difference is normalized by the sampling interval , and divided by to perform amplitude compression, so that the result of the adjacent difference approximation of the instantaneous frequency falls within the range of about . For the problem that the sequence start cannot be differentiated, the present embodiment uses the value of to fill the preposition in implementation, so as to ensure well-defined. Finally, an instantaneous frequency sequence with the same length as the original signal is obtained. The amplitude and the instantaneous frequency described above jointly constitute the time-domain feature vector, which represents the change of the amplitude and the frequency of the signal with time by using 2 real value channels.

[0133] (2) Frequency domain feature extraction: in order to obtain the spectral distribution characteristics of the signal, the present embodiment performs a short-time Fourier transform (STFT) on the signal. The STFT performs a local Fourier transform on the signal by sliding a time window, and its definition is as follows:

[0134] (3);

[0135] wherein, is the window length, in sample points, is the number of FFT points per frame / frequency discrete length (i.e. the length of the STFT frequency axis), is the frequency bin index, is the frame shift / step size, in "sample points", the start position of the adjacent two frames is different sample points, is the time frame index, represents the STFT complex spectrum value of the frame in the frequency bin, is the original sequence length, that is, the number of sample points, is a complex baseband discrete sequence (composed of I / Q) with a length of , let , r represents the index within each frame, r = 0, 1,..., L-1, represents the global sample corresponding to the offset within the frame, that is, the signal sample at the offset within the frame starts from the global position , , , is a window function with a length of , and in this embodiment, a Hamming window is used as an analysis window in the STFT, where 0.54 and 0.46 are standard coefficients of the Hamming window, which are classical engineering values. Through the STFT, a complex spectrum matrix is obtained, and the amplitude thereof is taken to obtain a time-frequency amplitude matrix . Next, the spectrum of each time frame is divided into two parts, the front half spectrum and the back half spectrum, and the average amplitude of the two parts of the frequency band is calculated. Specifically, the frequency range is divided into two parts: is the front half spectrum, is the back half spectrum. The average amplitude of the front half spectrum of each frame is calculated as , and the average amplitude of the back half spectrum is calculated as . In this way, two amplitude curves and varying with time frames are obtained. Considering that the length of the original signal sampling sequence is points, and the number of time frames obtained by the STFT may not be consistent with , this embodiment re-maps and back to discrete sequences and with a length of by interpolation resampling. When the number of STFT frames is small, linear interpolation is used for expansion; if the number of frames is more than This allows for equal-interval sampling, ensuring that the resampled sequence has the same length as the original signal. The final result is a frequency domain feature vector containing the average amplitude sequence of the first half of the spectrum. and the second half-spectrum average amplitude sequence One channel each.

[0136] (3) Feature fusion and preprocessing: such as Figure 2 As shown, the time-domain features (2 channels) and frequency-domain features (2 channels) obtained above are concatenated with the original I and Q signals (2 channels) along the channel dimension to form a dimension of The feature tensor. That is, each signal sample is now represented by a 6-position aligned one-dimensional sequence, representing: , Amplitude Instantaneous frequency First half-spectrum average amplitude sequence and the second half-spectrum average amplitude sequence The concatenated multi-channel data undergoes further channel-by-channel normalization preprocessing: the mean is subtracted from each of the six channels of each sample, and then divided by the standard deviation, to eliminate differences in the dimensions and ranges of different features and prevent any one channel from dominating network updates during subsequent training. The normalized 6-channel feature tensor is denoted as... , ( For batch size, The signal sampling length, Indicates all shapes as The set of real matrices can be used as input to a convolutional attention network model.

[0137] 2. Design of Convolutional Attention Network Model Architecture.

[0138] The deep network proposed in this embodiment adopts an architecture that combines convolutional neural networks with the Transformer self-attention mechanism to handle multi-channel fusion features. Automatic modulation classification is performed. Structurally, the model includes an input layer, a stem embedding layer, three feature extraction stages, and a classification head. The classification head consists of global pooling and a multilayer perceptron (i.e., a two-layer perceptron). Figure 3 As shown, the network processes signals layer by layer from top to bottom and left to right. Convolutional layers are used to extract local pattern features, and self-attention layers are used to model global correlations. The main components are described below:

[0139] (1) Convolution stem layer (i.e., stem embedding layer): Considering that the original number of input channels is 6, the input is first processed by a convolutional "stem" subnetwork for preliminary feature extraction and channel dimensionality increase. The stem maps the channel number from 6 to 32 by a one-dimensional convolutional layer (convolution kernel length 7, input channel 6, output channel 32). Then, batch normalization and SiLU activation function are sequentially applied, followed by a depth separable convolution DSConv1d to further extract features and enhance nonlinearity, obtaining the first extracted feature, as shown in Figure 4 . The depth separable convolution DSConv1d consists of two one-dimensional convolutions (the first convolution kernel length is kernel, the input channel number and the output channel number are both c, and the second convolution kernel length is 1, the input channel number and the output channel number are both c), a batch normalization, and a SiLU activation function, as shown in Figure 5 . The role is to encode the original multi-channel sequence into a higher-dimensional feature representation, providing a suitable feature space for the subsequent Transformer module. After the stem layer processing, the output tensor (i.e., the first extracted feature) has a dimension of , where is the batch size, is the initial sequence length (not down-sampled).

[0140] (2) Multi-stage convolution-attention feature extraction (i.e., three feature extraction stages): After the stem, the network main body is composed of a number of convolution-attention extraction stages concatenated together. Each stage includes an optional convolution embedding (down-sampling) layer ConvEmbed and two convolutional dual attention blocks (Convolutional Dual Attention Block, CDA Block) stacked in order.

[0141] The convolution embedding (down-sampling) layer ConvEmbed is shown in Figure 6 , which consists of 4 Conv1d layers, 2 batch normalization layers, and 2 SiLU activation functions. First, the input feature is processed by a Conv1d layer with a convolution kernel size of 7, an input channel number of , and an output channel number of to extract local temporal features; then, a Conv1d layer with a convolution kernel size of 1, an input channel number of , and an output channel number of is used to complete the channel transformation. Next, the feature is processed by a batch normalization and a SiLU activation function to enhance the non-linear expression ability. Then, a Conv1d layer with a convolution kernel size of 3, an input channel number of , and an output channel number of Conv1d layer for feature integration.

[0142] Specific stage configuration:

[0143] The first stage input is 32 channels, and the first extracted features are not immediately down-sampled (the embedding layer is an identity mapping), and are directly input to two convolutional dual attention modules; in the second stage, the channel number is increased from 32 to 64 through a ConvEmbed embedding layer (parameters are 64, and the time length is halved to extract more abstract features, and then input to two convolutional dual attention modules. In the third stage, the channel number is increased from 64 to 128 through a ConvEmbed embedding layer (parameters are 128, and the time length is halved to further extract more abstract features, and then input to two convolutional dual attention modules. Through this step-by-step time domain down-sampling and feature dimension increasing, the network can extract more abstract and larger receptive field features while maintaining computational efficiency.

[0144] The convolutional dual attention module CDA Block (Convolutional Dual Attention Block) is as shown in Figure 7 . Let the convolutional dual attention module input tensor ( be the batch size, be the channel number, be the current sequence length), first do batch normalization (norm1) on the convolutional dual attention module input tensor (that is, any input to the convolutional dual attention module), to obtain the batch-normalized feature , and then pass it through three groups of depth separable convolution DSConv1d to extract query feature vector , key feature vector and value feature vector three feature vectors. The three groups of convolution kernel parameters are independent but have the same structure to extract information within the local neighborhood of each channel, as shown in Figure 7 . This step obtains tensor . Using convolution instead of traditional linear layers produces , , which fully utilizes the local correlation and shift-invariance of the signal, helping to reduce the model parameter amount and improve the robustness to noise.

[0145] Temporal dimension self-attention mechanism: the above , , The correlation between different positions in the global modeling time series is applied to self-attention calculation. , , Therefore, the transpose of is , that is, each row corresponds to the feature of a time step, and MT represents the transpose operation; keep as , and get the attention of , here is only different in dimension convention but mathematically equivalent in the standard Attention formula. Calculate the attention weight matrix:

[0146] (4);

[0147] Wherein, is the dimension of the key feature vector, is a scaling factor, represents the attention weight matrix of the information at time position to position . Then, transpose to , and multiply it with the attention weight matrix: , to get the output fused with global time dependence (i.e., time attention output) . This realizes the self-attention of the time dimension: the feature vector at each time is updated as the weighted sum of all times, thereby capturing long-term correlation (e.g., the association between the symbols before and after the cycle). The Softmax of the attention weight ensures that it is mainly contributed by the positions with high correlation, and the weighted result can highlight the time sequence segment that is most useful for current discrimination.

[0148] Channel dimension attention (gate) mechanism: in addition to time correlation, the module of the embodiment introduces a channel attention mechanism to dynamically measure the importance of different feature channels. The specific method is to do batch normalization on the features of the input of the convolutional double attention module, do global average pooling in the time dimension, calculate the average value of each channel over the entire time sequence, and get the channel-level feature vector :

[0149] (5);

[0150] This vector represents the statistical mean of each channel in the time dimension, that is, the global feature summary at the channel level.

[0151] Subsequently, input After passing through the channel attention gating unit (composed of two fully connected (i.e., linear layers) and nonlinear activation functions), the channel weight vector is obtained :

[0152] (6);

[0153] wherein, and represent the weight matrices of the two linear layers, and represent the bias vectors, represents the SiLU activation function, represents the Sigmoid activation function for limiting the output to between 0 and 1, the channel weight vector , each represents the importance weight of the i-th channel, and when the feature corresponding to a certain channel contributes more to classification, its weight approaches 1; conversely, if the channel is useless or has a large amount of noise, the weight tends to 0.

[0154] Next, the channel attention weight is expanded (copied T times) in the time dimension to obtain the expanded weight vector , and is multiplied element-wise with the batch-normalized feature :

[0155] (7);

[0156] wherein, represents element-wise broadcast multiplication (broadcasted along the time dimension).

[0157] Therefore, the channel gating mechanism can achieve adaptive feature selection, dynamically highlighting effective channels and suppressing redundant features, and has a significant effect, especially in the multi-source feature fusion scenario.

[0158] Residual fusion and feedforward network: the time attention output is added to the channel attention output to obtain the fused attention result :

[0159] (8);

[0160] This result integrates the global dependencies in the time dimension and the importance evaluation in the channel dimension. Subsequently, a Dropout operation is applied to prevent overfitting, and the result is added to the input feature of the convolutional dual attention module through a residual connection:

[0161] (9);

[0162] Next, The result of the residual connection (i.e., the residual connection result) is sent to a feed-forward network (FFN) after being batch normalized (Norm2). The FFN is composed of a deep separable convolution (DSConv1d) and a one-dimensional convolution, and cooperates with batch normalization and a SiLU activation function to complete nonlinear transformation and local feature enhancement:

[0163] (10);

[0164] The final output of the convolutional dual attention module CDA Block The output result of the convolutional dual attention module is obtained through a second residual connection:

[0165] (11);

[0166] The residual structure ensures the integrity of the backbone information during cross-layer transmission. Even if the attention module output is weak or produces negative gradients, the original feature flow will not be damaged, thereby significantly improving the training stability and convergence speed of the network.

[0167] At this point, the complete calculation process of a convolutional dual attention module (CDA Block) is completed. The module integrates local convolution feature extraction, temporal self-attention, channel gating attention, and feed-forward update mechanism, and is functionally equivalent to a convolutional Transformer encoder unit. In the network of the present embodiment, two such structures are connected in series at each stage to gradually refine deep features.

[0168] Global pooling and output layer: After processing by the three stages (Stage 1 to Stage 3), the time sequence length of the input global-pooled feature (i.e., the second extracted feature) has been gradually down-sampled (e.g., from 128 points to 32 points), while the channel number has been increased to the highest The tensor size obtained at this time (i.e., the size of the second extracted feature) is Next, the time dimension feature is aggregated by the global pooling layer to obtain a fixed-length channel-level global feature vector. In order to better preserve peak information and enhance the response to key time segments, the global pooling of the present embodiment preferably uses generalized mean pooling (GeM). In order to ensure the expression is rigorous, the present embodiment explicitly introduces the sample index s=0,1,...,B-1 (B is the batch size) in the pooling formula to reflect the tensor structure of batch processing. represents the feature value of the c-th channel in the s-th sample in the batch at the t-th time position. The generalized mean pooling (GeM) is defined as:

[0169] (12);

[0170] in, In the s-th sample, the first... The feature values ​​of each channel at time position t. This represents the pooling output of the c-th channel of the s-th sample. After calculating for all channels, the global feature vector of the s-th sample is obtained. The pooling results of all samples are stacked according to the batch dimension to form the overall feature matrix. . These are learnable pooling parameters used to adaptively adjust between average pooling and max pooling. It degenerates into ordinary average pooling when This approximates max pooling. By... GeM’s adaptive learning automatically balances between averaging and max pooling, which can enhance discriminative peak features while preserving the overall trend.

[0171] The global feature vector obtained after pooling is input into a multilayer perceptron, which includes linear transformation, nonlinear activation, and Dropout regularization operations to further extract discriminative features and output scores for each modulation category. Finally, a normalized probability distribution is obtained through the Softmax function, and the network output is a vector of length equal to the number of modulation categories, representing the confidence level of the input signal belonging to each modulation type.

[0172] During the training phase, the cross-entropy loss function is used, which can be combined with a label smoothing strategy to improve the model's generalization ability; during the inference phase, the category with the highest confidence (i.e., the modulation category corresponding to the maximum probability value in the probability distribution) is selected as the recognition result (i.e., the signal modulation recognition result).

[0173] Global feature vector after pooling Feature mapping and classification output are performed using a multilayer perceptron. This multilayer perceptron integrates linear transformation, nonlinear activation, dropout, and final class mapping. Its computational expression is as follows:

[0174] (13);

[0175] in, , For classification layer weights, , For bias, For the number of categories, This represents the SiLU activation function. This represents a random deactivation operation, used for regularization. The prediction score of each category is obtained. Supervised learning is performed using a cross-entropy loss function during training, and the category corresponding to the maximum probability score is taken as the final recognition result during inference.

[0176] In summary, the technical scheme of the embodiment realizes effective learning of modulation signal discrimination features through multi-channel input of "I / Q raw signals and manual features", combined with a deep neural network structure of "convolution extraction and double attention fusion". Due to the use of lightweight convolution and attention modules (such as replacing ordinary convolution with depth separable convolution), the parameter quantity and computational complexity of the model are controlled. For example, in the implementation described above, the total parameter quantity is about 63,000, which is very lightweight in the same depth model. This means that the embodiment method has the potential to be deployed in embedded or real-time systems. The network model uses a self-attention mechanism to reduce the impact of irrelevant disturbances on feature extraction, and channel attention enhances the main mode features and suppresses redundant information. The overall process realizes the organic integration of manual prior knowledge and data-driven learning, significantly improving the performance and applicability of automatic modulation recognition.

[0177] The embodiment method has the following three essential differences compared with existing technologies represented by "raw I / Q end-to-end learning" (such as directly learning features from I / Q signals using only CNN / Transformer without explicitly introducing physical priori):

[0178] 1. Different starting points: from "cold start" to "warm start with physical priori".

[0179] Existing technology: Directly facing raw I / Q, it needs to find time-frequency structure and invariance in a vast parameter space, and it is highly sensitive to data size and network depth, especially slow convergence and unstable representation under medium and low signal-to-noise ratio (SNR) conditions.

[0180] Embodiment method: Inject interpretable physical priori such as amplitude envelope, instantaneous frequency, and segmented band power into the input side, and fuse it with the raw I / Q signal as a multi-channel input according to time alignment. It is equivalent to providing structured priori and reasonable normalization at the starting point, so that the network starts learning from a state with basic knowledge, converges faster, and is more stable under medium and low SNR.

[0181] 2. Different guidance strategies: from "uniform processing" to "attention-driven key learning".

[0182] Existing technology: Conventional convolution or pure self-attention treats time segments / channels equally, lacking adaptive emphasis mechanisms for key moments (such as sudden phase transitions / bandwidth mutations) and key channels (such as priori channels more sensitive to certain modulations).

[0183] Embodiment method: Convolutionized ​ / The dual attention structure of time self-attention and channel attention gives higher weights to high discriminative time slices and high contribution channels during the training process; and the GeM learnable pooling is introduced to adaptively select the aggregation method between "average-maximum" to realize dynamic focusing and importance-based allocation of learning resources at the process level.

[0184] 3. Different knowledge contents: from "pure statistical representation" to "dual-source representation of physical priori and data-driven".

[0185] Prior art: mainly rely on the network to automatically learn statistical related representation from data, weak in explanation, when the training data is limited or the signal-to-noise ratio is low, the feature is unstable or overfitting to noise.

[0186] The method of the embodiment: in the content level, the interpretable time-frequency physical features and the high-order statistical representation of the deep network are fused in the same domain, which not only retains the interpretability and priori constraint, but also gives full play to the fitting ability of the deep nonlinear model to the complex discriminant boundary; cooperate with the unified shape alignment and training set statistical standardization strategy, improve the consistency and reproducible experiment of cross-SNR and cross-scene.

[0187] Summary: the embodiment is not a single-point fine-tuning of the prior art, but a systematic AMR framework constructed from three dimensions of starting point (priori enhancement), process (attention guidance) and content (dual-source representation). It changes the originally "passive and uniform" end-to-end fitting into a "priori enhancement and dynamic focus" learning process; without relying on special hardware, it improves the convergence efficiency, low-SNR robustness and high-SNR discriminability in a pure software way (the corresponding effects have been quantitatively verified in the embodiment).

[0188] In order to better illustrate the technical solutions of the embodiment, the following experiments are carried out in the embodiment:

[0189] To verify the effectiveness of the technical solutions of the present embodiment, the following specific embodiments are given in combination with the disclosed data set RML2016.10a. RML2016.10a provides a radio signal modulation data set, containing 11 modulation categories, including digital modulation (such as BPSK, QPSK, 8PSK, 16QAM, 64QAM, CPFSK and PAM4) and analog modulation (such as AM-DSB, AM-SSB, WBFM and GFSK) and the like, each modulation signal is collected under a series of different signal-to-noise ratios (SNR) conditions. Specifically, the SNR value range is -20dB to 18dB, every 2dB, a total of 20 grades; each (modulation category, SNR) combination contains 1000 independent signal samples, each sample length is 128 data points. The experiment first divides the data set into training set, validation set and test set. The division method follows the common practice: randomly select 600 samples for each modulation category and each SNR for training, 200 for validation and 200 for testing, accounting for 60%, 20% and 20% respectively. In this way, the distribution of each modulation category and each SNR condition in the training, validation and test sets is consistent and independent without overlap.

[0190] According to the process of the present embodiment, artificial features are first extracted from each I / Q signal of the training set and fused into 6-channel input. The training stage simultaneously trains the model parameters of the present embodiment on all SNR training data to enhance the adaptability of the present embodiment model to different signal-to-noise ratio conditions. In the validation stage, the performance of the present embodiment model under each SNR is evaluated respectively, and the total accuracy of the validation set is used as the early stopping and model selection basis. At the same time, the accuracy change of the high signal-to-noise ratio (for example, 18dB) validation set is specially monitored, which is used to guide the learning rate scheduling (when the 18dB validation set accuracy tends to be saturated, the learning rate is reduced). The present embodiment uses the Adam optimizer for 400 rounds of training, the initial learning rate is set to 0.002, and the learning rate is multiplied by 0.5 to decay every time the validation set performance does not improve for 50 rounds, to refine the convergence. The loss function uses cross-entropy loss with label smoothing (smoothing factor 0.1), and the batch size is 128. During the training process, the model does not appear overfitting on the validation set, and the model parameters with the highest validation set accuracy are finally selected for test evaluation.

[0191] On the test set, the classification accuracy of the method of the embodiment under the condition of high signal-to-noise ratio 18dB reaches 94.23%, that is, about 94% of the 44000 18dB test samples are correctly classified. This performance exceeds the level of the prior art pure deep learning model on the same data set (92.18%). More importantly, due to the fusion of artificial features, the method of the embodiment shows a more stable performance decline curve under medium and low signal-to-noise ratio: for example, more than 60% of the recognition rate is still achieved under the condition of 0dB, while some pure CNN models may be less than 50%. The test results show that most modulation types can achieve higher recognition accuracy than CNN and MCLDNN and other pure deep learning methods under medium and high SNR (>0dB); under very low SNR (such as below-10dB), the performance decreases, but is still significantly better than the baseline model without fused features. The experimental results are shown in Figure 8 The method of the embodiment (i.e., the method) is compared with the recognition accuracy (ACC) of the network models such as CNN, LSTM, GRU, ResNet1D and MCLDNN in the prior art. As can be seen, the method of the embodiment has good recognition ability for different signal modes while maintaining high overall accuracy. The above test results prove that the deep learning strategy of the embodiment with fused features helps to improve the performance of automatic modulation recognition and has significant advantages compared with traditional schemes.

[0192] In order to verify the effectiveness of different modules of the method, the four key components of “prior feature fusion”, “time attention”, “channel attention” and “GeM pooling” are ablated one by one under the premise of strictly maintaining consistent data division and training process, as shown in Table 1. Taking the 18dB single-point performance as an example, removing any component will cause a significant decrease in accuracy: -2.69% for removing prior, -2.23% for removing time attention, -2.02% for removing channel attention, and -2.28% for replacing GeM with average pooling; the complete model is 94.23%, which is significantly better than each ablation variant (see Table 1 for details). This shows that the four components are necessary and complementary to the final performance.

[0193] The recognition accuracy (ACC) of the ablation experiment under different signal-to-noise ratios is shown in Figure 9 From the trend of the whole SNR, the positive effect of each component is not only reflected in high SNR, but also in medium and low SNR intervals, and the overall curve is relatively smoother than the ablation variants; as can be seen from Table 1, the four modules are not mutually replaced but mutually complementary: the prior feature improves the separability baseline, the double attention improves the key alignment and noise suppression, and the GeM is responsible for information aggregation, and the three are superimposed to be better than any one of the modules.

[0194] Table 1 Ablation experiment results

[0195]

[0196] With reference Figure 10 The embodiment of the present application also provides a signal intelligent modulation recognition system fusing time-frequency features, which comprises a feature extraction unit 100, a feature fusion unit 200, a model training unit 300, a first feature extraction unit 400, a second feature extraction unit 500 and a signal modulation recognition unit 600, wherein:

[0197] The feature extraction unit 100 is used for extracting time-domain features and frequency-domain features in a radio signal in a complex baseband form.

[0198] The feature fusion unit 200 is used for fusing the radio signal, the time-domain features and the frequency-domain features to obtain multi-channel fusion features.

[0199] The model training unit 300 is used for training a convolution attention network model containing a stem embedding layer, a plurality of cascaded feature extraction stages and a classification head.

[0200] The first feature extraction unit 400 is used for inputting the multi-channel fusion features into the trained convolution attention network model, extracting features of the multi-channel fusion features through the stem embedding layer to obtain first extraction features.

[0201] The second feature extraction unit 500 is used for extracting deep features of the first extraction features through the plurality of cascaded feature extraction stages to obtain second extraction features.

[0202] The signal modulation recognition unit 600 is used for inputting the second extraction features into the classification head to perform signal modulation recognition, so as to obtain a signal modulation recognition result corresponding to the radio signal.

[0203] It should be noted that since the signal intelligent modulation recognition system fusing time-frequency features in the embodiment and the signal intelligent modulation recognition method fusing time-frequency features described above are based on the same inventive concept, the corresponding contents in the method embodiment are also applicable to the system embodiment, which will not be described in detail here.

[0204] With reference Figure 11 The embodiment of the present application also provides an electronic device, which comprises:

[0205] At least one memory;

[0206] At least one processor;

[0207] At least one program;

[0208] The program is stored in the memory, and the processor executes the at least one program to implement the signal intelligent modulation recognition method fusing time-frequency features described above in the embodiment of the present disclosure.

[0209] The electronic device can be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (PDA), a vehicle-mounted computer, and the like.

[0210] The electronic device of the embodiments of the present application is described in detail below.

[0211] The processor 1600 can be implemented in a general-purpose central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), or one or more integrated circuits, and is configured to execute related programs to implement the technical solutions provided by the embodiments of the present disclosure.

[0212] The memory 1700 can be implemented in a read only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1700 can store an operating system and other application programs. When the technical solutions provided by the embodiments of the present disclosure are implemented by software or firmware, the related program codes are stored in the memory 1700 and are called and executed by the processor 1600 to implement the signal intelligent modulation recognition method of fusion time-frequency features.

[0213] The input / output interface 1800 is configured to realize information input and output.

[0214] The communication interface 1900 is configured to realize the communication interaction between the device and other devices. The communication can be realized by a wired manner (for example, a USB, a network cable, etc.) or a wireless manner (for example, a mobile network, WIFI, Bluetooth, etc.).

[0215] The bus 2000 is configured to transmit information between the components (for example, the processor 1600, the memory 1700, the input / output interface 1800, and the communication interface 1900) of the device.

[0216] The processor 1600, the memory 1700, the input / output interface 1800, and the communication interface 1900 are connected to each other through the bus 2000 to realize the communication connection between them inside the device.

[0217] The embodiment of the present disclosure further provides a storage medium, which is a computer readable storage medium, and stores computer executable instructions for causing a computer to execute the signal intelligent modulation identification method of the fused time-frequency feature.

[0218] The memory, as a non-transitory computer readable storage medium, can be used to store non-transitory software programs and non-transitory computer executable programs. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory remotely arranged relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0219] The embodiments described in the embodiments of the present disclosure are for more clearly illustrating the technical solutions of the embodiments of the present disclosure, and do not constitute a limitation on the technical solutions provided by the embodiments of the present disclosure. Those skilled in the art can know that, with the evolution of technology and the appearance of new application scenarios, the technical solutions provided by the embodiments of the present disclosure are also applicable to similar technical problems.

[0220] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present disclosure, and can include more or fewer steps than those shown in the figures, or combine certain steps, or different steps.

[0221] The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separate, that is, can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments of the present disclosure.

[0222] Those skilled in the art can understand that all or some steps in the above disclosed method, the functions of the modules / units in the system and the device can be implemented as software, firmware, hardware and their appropriate combinations.

[0223] The terms "first", "second", "third", "fourth", and the like in the description of this application and in the claims hereof, if any, are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so termed herein is solely for the convenience of the reader and does not limit the scope of the application. It is also to be understood that the description and examples in this application are intended to cover all possible combinations where any of the several elements can represent one or more elements.

[0224] It should be understood that, in the application, "at least one" means one or more, "multiple" means two or more. "And / or" is used to describe the relationship between associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0225] In several embodiments provided in the application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0226] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on multiple network units. According to actual needs, some or all of the units can be selected to achieve the purpose of the embodiment of the application.

[0227] In addition, each function unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software function unit.

[0228] If the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions for causing an electronic device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, and various program storage media. The embodiments of the present application are described in detail above in combination with the drawings, but the present application is not limited to the above embodiments, and various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the purpose of the present application.

Claims

1. A signal intelligent modulation recognition method integrating time-frequency features, characterized in that, The method includes: Extracting time-domain and frequency-domain features from complex baseband radio signals, including: Extract the amplitude envelope from the complex baseband form of the radio signal, where the amplitude envelope is the amplitude of the radio signal; The instantaneous frequency is calculated based on the in-phase and quadrature components in the complex baseband form of the radio signal; The amplitude envelope and the instantaneous frequency are used as time-domain features; A short-time Fourier transform is performed on the complex baseband form of the radio signal to obtain the complex-valued spectral matrix; Based on the complex-valued spectral matrix, calculate the average amplitude of the first half of the spectrum and the average amplitude of the second half of the spectrum; The average amplitude of the first half-spectrum and the average amplitude of the second half-spectrum are mapped back to discrete sequences to obtain the average amplitude sequence of the first half-spectrum and the average amplitude sequence of the second half-spectrum. The average amplitude sequence of the first half-spectrum and the average amplitude sequence of the second half-spectrum are used as frequency domain features; The radio signal, the time-domain features, and the frequency-domain features are fused to obtain a multi-channel fused feature; A trained convolutional attention network model is constructed, comprising a stem embedding layer, multiple cascaded feature extraction stages, and a classification head. The stem embedding layer includes one-dimensional convolution, batch normalization, activation functions, and depthwise separable convolution. The first feature extraction stage includes identity mapping and two convolutional dual-attention modules. The second feature extraction stage includes a convolutional embedding layer and two convolutional dual-attention modules. The third feature extraction stage includes a convolutional embedding layer and two convolutional dual-attention modules. Each convolutional dual-attention module includes multiple depthwise separable convolutions, a temporal self-attention mechanism, a channel-dimensional attention mechanism, and a feedforward network. The channel-dimensional attention mechanism includes global average pooling and channel attention gating units. The classification head includes global pooling and a multilayer perceptron. The multi-channel fused features are input into the trained convolutional attention network model, and the multi-channel fused features are extracted through the STEM embedding layer to obtain the first extracted features; The first extracted feature is subjected to deep feature extraction through the multiple cascaded feature extraction stages to obtain the second extracted feature; The second extracted feature is input into the classification head for signal modulation recognition to obtain the signal modulation recognition result corresponding to the radio signal.

2. The intelligent modulation and recognition method for signals that integrates time-frequency features according to claim 1, characterized in that, The step of extracting features from the multi-channel fused features through the stem embedding layer to obtain the first extracted features includes: The initial features are obtained by performing preliminary feature extraction and channel dimensionality upscaling on the multi-channel fused features through the one-dimensional convolution. The initial features are subjected to batch normalization and activation function processing to obtain the processed features; The processed features are extracted and nonlinearly enhanced by the depthwise separable convolution to obtain the first extracted features.

3. The intelligent modulation and recognition method for signals that integrates time-frequency features according to claim 1, characterized in that, The process of performing deep feature extraction on the first extracted features through the multiple cascaded feature extraction stages to obtain the second extracted features includes: The multiple cascaded feature extraction stages are divided into a first feature extraction stage, a second feature extraction stage, and a third feature extraction stage. The first extracted features are extracted by identity mapping and two convolutional dual attention modules in the first feature extraction stage to obtain the output result of the first feature extraction stage. The output of the first feature extraction stage is obtained by extracting features from the output of the second feature extraction stage through the convolutional embedding layer and two convolutional dual attention modules in the second feature extraction stage. The second extracted features are obtained by extracting features from the output of the second feature extraction stage through the convolutional embedding layer and two convolutional dual attention modules in the third feature extraction stage.

4. The intelligent modulation and recognition method for signals that integrates time and frequency features according to claim 3, characterized in that, The first feature extraction stage extracts features by using the identity mapping and two convolutional dual attention modules in the first feature extraction stage, and obtains the output result of the first feature extraction stage, including: The result of performing an identity mapping on the first extracted features is input into two cascaded convolutional dual-attention modules for feature extraction, resulting in the output of the first feature extraction stage. The feature extraction process of each convolutional dual-attention module includes: After batch normalization of the result of the identity mapping, the batch normalized features are obtained. The multiple depthwise separable convolutions are used to extract features from the batch-normalized features to obtain query features, key features, and value features. Based on the query features, the key features, and the value features, the time-dimensional self-attention mechanism is used to calculate the time attention output; The batch-normalized features are processed by the channel-dimensional attention mechanism to obtain the channel attention output; The temporal attention output and the channel attention output are fused to obtain the fused attention result; The fused attention result is subjected to Dropout operation and then residually connected to the batch-normalized features to obtain the residual connection result. The residual connection results are batch normalized and then input into the feedforward network to obtain the feedforward network output results; The output of the feedforward network is subjected to Dropout operation and then residually connected with the residual connection result to obtain the output of the convolutional dual attention module.

5. The intelligent modulation and recognition method for signals that integrates time-frequency features according to claim 4, characterized in that, The step of processing the batch-normalized features through the channel-dimensional attention mechanism to obtain the channel attention output includes: The batch-normalized features are subjected to global average pooling in the time dimension, and the average value of each channel over the entire time series is calculated to obtain the channel-level feature vector. The channel-level feature vector is input into the channel attention gating unit to obtain the channel attention weights; The channel attention weights are extended in the time dimension to obtain the extended attention weights; The expanded attention weights are multiplied element-wise with the batch-normalized features to obtain the channel attention output.

6. The intelligent modulation recognition method for signals that integrates time-frequency features according to claim 1, characterized in that, The step of inputting the second extracted feature into the classification head for signal modulation recognition to obtain the signal modulation recognition result corresponding to the radio signal includes: The second extracted feature is subjected to global pooling to obtain the pooled global feature vector. The pooled global feature vector is input into the multilayer perceptron for signal modulation recognition, and prediction scores for each of the multiple modulation categories are obtained. The prediction scores of each of the multiple modulation categories are processed by the Softmax function to obtain a probability distribution. The modulation category corresponding to the maximum probability value in the probability distribution is taken as the signal modulation identification result corresponding to the radio signal.

7. A signal intelligent modulation recognition system integrating time-frequency features, characterized in that, The system includes: The feature extraction unit is used to extract time-domain and frequency-domain features from radio signals in complex baseband form, including: Extract the amplitude envelope from the complex baseband form of the radio signal, where the amplitude envelope is the amplitude of the radio signal; The instantaneous frequency is calculated based on the in-phase and quadrature components in the complex baseband form of the radio signal; The amplitude envelope and the instantaneous frequency are used as time-domain features; A short-time Fourier transform is performed on the complex baseband form of the radio signal to obtain the complex-valued spectral matrix; Based on the complex-valued spectral matrix, calculate the average amplitude of the first half of the spectrum and the average amplitude of the second half of the spectrum; The average amplitude of the first half-spectrum and the average amplitude of the second half-spectrum are mapped back to discrete sequences to obtain the average amplitude sequence of the first half-spectrum and the average amplitude sequence of the second half-spectrum. The average amplitude sequence of the first half-spectrum and the average amplitude sequence of the second half-spectrum are used as frequency domain features; The feature fusion unit is used to fuse the radio signal, the time-domain features, and the frequency-domain features to obtain multi-channel fused features; The model training unit is used to train a convolutional attention network model that includes a stem embedding layer, multiple cascaded feature extraction stages, and a classification head. The stem embedding layer includes one-dimensional convolution, batch normalization, activation functions, and depthwise separable convolution. The first feature extraction stage includes identity mapping and two convolutional dual-attention modules. The second feature extraction stage includes a convolutional embedding layer and two convolutional dual-attention modules. The third feature extraction stage includes a convolutional embedding layer and two convolutional dual-attention modules. Each convolutional dual-attention module includes multiple depthwise separable convolutions, a temporal self-attention mechanism, a channel-dimensional attention mechanism, and a feedforward network. The channel-dimensional attention mechanism includes global average pooling and a channel attention gating unit. The classification head includes global pooling and a multilayer perceptron. The first feature extraction unit is used to input the multi-channel fused features into the trained convolutional attention network model, and extract features from the multi-channel fused features through the STEM embedding layer to obtain the first extracted features; The second feature extraction unit is used to perform deep feature extraction on the first extracted feature through the multiple cascaded feature extraction stages to obtain the second extracted feature. The signal modulation recognition unit is used to input the second extracted features into the classification head to perform signal modulation recognition and obtain the signal modulation recognition result corresponding to the radio signal.

8. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform the intelligent modulation identification method for integrating time-frequency features as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the intelligent modulation and recognition method for integrating time-frequency features as described in any one of claims 1 to 6.

Images