A robust modulation recognition method for channel-varying scenarios
Patent Information
- Application Number
- CN202611459663.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-20
- Publication Date
- 2026-10-27
AI Technical Summary
本申请首先结合无线通信信号波形先验和盲同步信号处理领域知识，针对性消除定时偏差、频偏相偏等易剥离的、确定性的信号畸变，规避基础同步误差对后续特征提取的干扰；然后利用多调制类型假设的信道盲均衡，对未知调制信号下的多径衰落、盲同步残余畸变等复合信道影响进行自适应补偿，弱化信道域偏移带来的负面影响；最后通过对多调制假设的均衡后IQ信号进行循环谱域和星座域的特征融合提取，挖掘调制假设匹配和失配下均衡后IQ信号的差异化波形特征，并依托深度分类网络实现高精度、强鲁棒性的调制识别。相较于现有技术，本申请无需依赖海量全场景信道下的信号训练数据，也无需复杂的域泛化约束设计，可有效适配开放、动态多变的真实无线通信信道环境。
Smart Images
Figure CN122994128A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of communication signal processing technology, and in particular to a robust modulation recognition method for scenarios with variable channel conditions. Background Technology
[0002] In non-cooperative communication scenarios, the receiving end typically lacks prior information such as the modulation scheme used by the transmitting end. Therefore, modulation identification technology is needed to determine the modulation type of the received signal, providing a foundation for subsequent signal demodulation, communication detection, and related signal processing. Existing modulation identification methods mainly include traditional identification methods based on artificial features and data-driven deep learning identification methods. Traditional identification methods typically extract manually designed features such as higher-order statistics, instantaneous features, and cyclostationary features, and combine them with a classifier to determine the modulation type. Deep learning identification methods, on the other hand, utilize neural networks to automatically extract features from the received signal and perform classification based on the extracted features.
[0003] However, in real-world non-cooperative communication environments, received signals can be affected by various factors such as time-varying fading, multipath propagation, Doppler shift, noise, and nonlinear distortion. Different channel conditions can lead to significantly different waveform characteristics and data distributions for signals of the same modulation type. For identification methods that rely on data augmentation, their performance is usually related to the degree to which the training data covers the actual channel conditions. However, the channel states in real-world communication environments are highly diverse and time-varying, making it difficult to fully cover potential channel changes with limited measured or simulated data. For identification methods that employ domain adaptation, feature alignment, or invariant feature learning, it is usually necessary to set corresponding network structures, training constraints, or loss functions for different data distributions. Their model design and training processes are complex, and even under channel conditions where no training is involved or where the channel differs significantly from the training channel, the extracted features may still be significantly affected by channel factors, and the feature distinguishability between different modulation types may decrease. Therefore, existing modulation identification methods still suffer from insufficient cross-channel generalization ability and poor stability of identification results in complex and dynamically changing channel environments. Summary of the Invention
[0004] To overcome the shortcomings of insufficient modulation recognition capabilities in the prior art, this invention provides a robust modulation recognition method for scenarios with variable channel conditions.
[0005] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows: A robust modulation identification method for scenarios with varying channel characteristics includes: The IQ signal to be identified is received and blind synchronization preprocessing independent of modulation type is performed to obtain the synchronized IQ signal. For multiple candidate modulation types in the candidate modulation type set, the linear equalizer of the synchronized IQ signal under any candidate modulation type assumption is processed, the linearly equalized signal is constellation quantization constellation constraint is performed on the standard constellation set corresponding to the candidate modulation type, and the waveform is reconstructed based on the equivalent channel impulse response to obtain the blindly equalized IQ signal. For any candidate modulation type, the blindly equalized IQ signal is extracted and merged with the high-order cyclic cumulant features and constellation diagram features respectively; the merged features corresponding to multiple candidate modulation types are fused to obtain multi-branch merged features. The multi-branch merging features are input into a modulation classification network containing a deep residual convolutional neural network and a fully connected classifier to perform modulation classification and discrimination, thereby determining the modulation type of the IQ signal to be identified.
[0006] As a preferred embodiment, the blind synchronization preprocessing of the received IQ signal to be identified, independent of modulation type, includes: Blind frequency offset estimation and frequency offset compensation are performed on the IQ signal to be identified to obtain the frequency offset compensated IQ signal; Blind symbol rate estimation is performed on the frequency offset compensated IQ signal, and downsampling is performed based on the estimated symbol rate to obtain an IQ signal with a single symbol sampling rate. Blind symbol timing synchronization is performed on the IQ signal at the single symbol sampling rate to obtain the timing-upgraded IQ signal; Blind phase bias estimation and phase bias correction are performed on the timing-post IQ signal to obtain the synchronized IQ signal; Specifically, the blind frequency offset estimation adopts blind frequency offset estimation based on joint peak search of higher-order cyclic cumulants of the signal, the blind symbol rate estimation adopts symbol rate blind estimation based on cyclic spectrum peak detection, the blind symbol timing synchronization adopts Gardner timing synchronization, and the blind phase offset estimation adopts phase offset estimation based on nonlinear minimum mean square criterion.
[0007] As a preferred approach, the steps of linear equilibrium processing include: The linearly equalized IQ signal is determined by performing a convolution operation between the synchronized IQ signal and the linear equalizer tap vector corresponding to any candidate modulation type, and its expression is as follows:
[0008] in, The signal after linear equalization For length is The The candidate modulation type corresponds to the tap vector of the linear equalizer. gather; The synchronized IQ signal; For in the complex field On E 3D vector space.
[0009] As a preferred embodiment, the steps of constellation quantization constraints include: For the linearly equalized signal, the distance between each IQ sample point and each standard constellation point in the standard constellation set corresponding to the candidate modulation type is determined, and the standard constellation point with the smallest distance is determined as the constellation quantization result of the corresponding IQ sample point. Based on the quantization result and the linearly equalized signal, a direct approximation is constructed to obtain the quantized constrained signal, whose expression is as follows:
[0010]
[0011] in, The signal after quantization constraints; This represents the gradient blocking operation, applied to variables of arbitrary vector form. ,satisfy ; Candidate modulation type p The total number of corresponding standard constellation points, Candidate modulation type p The quantitative results Candidate modulation type p The quantization result of the kth sampling point Candidate modulation type p No. k The linearly equalized signal at each sampling point; Candidate modulation type p The first in the corresponding standard constellation point set i A standard constellation point; In order to be with the first k The index of the standard constellation point that minimizes the Euclidean distance between signals after linear equalization.
[0012] As a preferred option, the waveform reconstruction process also involves gradient backpropagation iterative optimization based on the loss function to obtain the tap vector estimate of the linear equalizer and the vector estimate of the equivalent channel impulse response under the current candidate modulation type assumption. The loss function includes a first error between the reconstructed waveform signal and the synchronized IQ signal, and a second error corresponding to the constellation nearest neighbor quantization error, and its expression is as follows:
[0013] in, Let be the loss function, where Candidate modulation type p The equivalent channel impulse response vector, To quantify the loss weights.
[0014] As a preferred embodiment, the extraction step of the higher-order cyclic cumulant feature includes: For any candidate modulation type, after blind equalization of the IQ signal, the higher-order cyclic cumulants are extracted according to the order of the higher-order cyclic cumulants and the number of corresponding conjugate terms under multiple cyclic harmonic orders. The extracted higher-order cyclic cumulants are moduloed, and the moduli of each higher-order cyclic cumulant are arranged according to the combination relationship between the cyclic harmonic order and the order of the higher-order cyclic cumulant and the number of conjugate terms, forming a higher-order cyclic cumulant feature matrix corresponding to the candidate modulation type, the expression of which is as follows:
[0015] in, It is the set of cyclic harmonic orders; For any A set of higher-order cyclic cumulants extracted from first-order cyclic harmonics; for under cyclic harmonics Accumulated amount of higher-order cycles; The order of the cumulant in a higher-order cycle; This is the set of orders of the extracted higher-order cyclic cumulants; The number of terms in the product conjugate of the higher-order cyclic cumulants; For the order of cumulant in higher-order cycles The set of the number of terms of the selected conjugate terms. It is a high-order cyclic cumulant characteristic matrix.
[0016] As a preferred embodiment, the steps for extracting the constellation diagram features include: Divide the two-dimensional complex plane region within a preset boundary radius into multiple grid cells at equal intervals; For any candidate modulation type, after blind equalization of the IQ signal, the number of IQ samples falling into each grid cell in the blind equalization IQ signal is counted and normalized to obtain the initial density matrix of the constellation diagram corresponding to the candidate modulation type. The initial density matrix of the constellation graph is input into the constellation graph feature encoder, which includes multiple parallel convolutional branches. Each parallel convolutional branch uses multiple two-dimensional convolutional layers to encode the initial density matrix of the constellation graph and uses convolutional kernels of different sizes to extract the constellation graph density features. Two-dimensional adaptive average pooling is performed on the constellation graph density features extracted by the parallel convolution branch to obtain a constellation graph density feature matrix with the same two-dimensional size as the higher-order cyclic cumulant feature matrix. The constellation graph density feature matrices output from multiple parallel convolution branches are concatenated along the channel dimension to obtain the constellation graph density feature tensor corresponding to the candidate modulation type.
[0017] As a preferred embodiment, the step of extracting and merging the high-order cyclic cumulant features and constellation graph features separately includes: For any candidate modulation type, the corresponding high-order cyclic cumulant feature matrix and constellation diagram density feature tensor are concatenated along the feature channel dimension to obtain the merged feature corresponding to the candidate modulation type. The fusion of merged features corresponding to multiple candidate modulation types includes: The merged features corresponding to multiple candidate modulation types are stacked according to the branch dimension corresponding to the candidate modulation type to form the multi-branch merged feature.
[0018] As a preferred embodiment, the deep residual convolutional neural network performs multi-level residual feature extraction and spatial dimensionality reduction on the multi-branch merging features to obtain a fused feature vector; The fully connected classifier performs modulation classification based on the fused feature vector; The deep residual convolutional neural network includes a residual feature extraction module and a directional downsampling module with stride convolution. The residual feature extraction module includes a main branch and a bypass connection. The main branch sequentially performs convolution, batch normalization, activation, convolution, and batch normalization on the input features, and then adds the output of the main branch to the output of the bypass connection before performing activation. The directional downsampling module performs downsampling along the corresponding feature dimension through stride convolution.
[0019] As a preferred embodiment, the candidate modulation type set includes multiple digital modulation standards with a standard constellation set, wherein the multiple digital modulation standards belong to at least one modulation type among M-PSK, M-QAM, M-ASK and M-APSK.
[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This application first combines prior knowledge of wireless communication signal waveforms and blind synchronization signal processing to specifically eliminate easily separable, deterministic signal distortions such as timing deviation, frequency offset, and phase offset, avoiding interference from fundamental synchronization errors on subsequent feature extraction. Then, it utilizes channel blind equalization with multiple modulation type assumptions to adaptively compensate for the combined channel effects of multipath fading and residual distortion in blind synchronization under unknown modulation signals, mitigating the negative impact of channel domain offset. Finally, it extracts features from the equalized IQ signals under multiple modulation assumptions by fusing cyclic spectral and constellation domains, uncovering differentiated waveform features of the equalized IQ signals under modulation assumption matching and mismatch, and achieving high-precision, robust modulation recognition based on a deep classification network. Compared to existing technologies, this application does not rely on massive amounts of signal training data across all channel scenarios, nor does it require complex domain generalization constraint design, effectively adapting to open, dynamically changing real-world wireless communication channel environments. Attached Figure Description
[0021] Figure 1 This is a flowchart of a robust modulation identification method for scenarios with varying channels, as described in Example 1. Figure 2 Here is a flowchart of the blind synchronization preprocessing process in Example 1; Figure 3 This is a diagram of the modulation identification module based on multi-path parallel blind equalization in Example 1; Figure 4 This is a schematic diagram of the blind equalization process in Example 1; Figure 5 This is a schematic diagram of the feature extraction and classification process in Example 1; Figure 6 This is a comparison chart of the average recognition accuracy of different methods in Example 2; Figure 7 This is the confusion matrix of the identification results of the seven modulation types in this application at a signal-to-noise ratio of 12dB in Example 2. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0024] Example 1 This embodiment proposes a robust modulation identification method for scenarios with variable channel conditions, such as... Figure 1 The diagram shows a flowchart of a robust modulation identification method for scenarios with varying channel conditions.
[0025] This embodiment proposes a robust modulation identification method for scenarios with varying channel characteristics, which includes: S1. Receive the IQ signal to be identified and perform blind synchronization preprocessing independent of modulation type to obtain the synchronized IQ signal; S2. For multiple candidate modulation types in the candidate modulation type set, the linear equalizer of the synchronized IQ signal under any candidate modulation type assumption is processed, the linearly equalized signal is constellation quantization constrained according to the standard constellation set corresponding to the candidate modulation type, and the waveform is reconstructed based on the equivalent channel impulse response to obtain the blindly equalized IQ signal. S3. For the blindly equalized IQ signal corresponding to any candidate modulation type, extract the high-order cyclic cumulant feature and constellation diagram feature respectively and merge them; fuse the merged features corresponding to multiple candidate modulation types to obtain multi-branch merged features. S4. Input the multi-branch merging features into a modulation classification network containing a deep residual convolutional neural network and a fully connected classifier to perform modulation classification and discrimination, and determine the modulation type of the IQ signal to be identified.
[0026] In this embodiment, relying on the idea of combining prior knowledge of communication signal processing with the feature mining capabilities of deep learning, the traditional paradigm of simply relying on data expansion or network feature constraints is abandoned. Through the signal processing analysis logic of "first distortion correction, then channel compensation, and finally feature fusion classification", the various distortion effects brought about by the variable channel are stripped away, and the inherent essential features of the modulation signal are accurately preserved. This solves the problem of poor generalization and weak stability of modulation recognition under complex channels from the root.
[0027] In an optional embodiment, the receiving of the IQ signal to be identified, performing blind synchronization preprocessing independent of modulation type, includes: Blind frequency offset estimation and frequency offset compensation are performed on the IQ signal to be identified to obtain the frequency offset compensated IQ signal; Blind symbol rate estimation is performed on the frequency offset compensated IQ signal, and downsampling is performed based on the estimated symbol rate to obtain an IQ signal with a single symbol sampling rate. Blind symbol timing synchronization is performed on the IQ signal at the single symbol sampling rate to obtain the timing-upgraded IQ signal; Blind phase bias estimation and phase bias correction are performed on the timing-post IQ signal to obtain the synchronized IQ signal; Specifically, the blind frequency offset estimation adopts blind frequency offset estimation based on joint peak search of higher-order cyclic cumulants of the signal, the blind symbol rate estimation adopts symbol rate blind estimation based on cyclic spectrum peak detection, the blind symbol timing synchronization adopts Gardner timing synchronization, and the blind phase offset estimation adopts phase offset estimation based on nonlinear minimum mean square criterion.
[0028] like Figure 2 The diagram shown is a flowchart of blind synchronization preprocessing.
[0029] More specifically, the length is Synchronized IQ signal , can be represented as:
[0030] in, This represents the convolution operation. This represents the transpose of a vector. It represents the complex field consisting of all complex numbers. It is a length of The complex transmitted signal vector, where elements Independent of each other and derived from the modulation type Corresponding standard constellation set ; Belongs to the candidate modulation type set , This represents the total number of candidate modulation types.
[0031] As an example, the transmitted signal vector of QPSK modulation At this time, the modulation type ,and , ,in It is a virtual part unit. It is the equivalent channel impulse response vector of the transmitter shaping filter, multipath channel, and blind synchronization preprocessing, etc. It is the total number of taps in the equivalent channel. The mean is The variance is The complex Gaussian noise vector.
[0032] In this embodiment, fully blind synchronization of carrier frequency, symbol rate, timing phase and carrier phase is achieved without relying on modulation type and prior parameters, which can adapt to complex electromagnetic environments.
[0033] like Figure 3The diagram shows a modulation identification module based on multi-path parallel blind equalization. The modulation identification module includes a parallel blind equalization submodule with multiple modulation hypotheses and a deep modulation classification submodule based on fusion features. In the parallel blind equalization submodule with multiple modulation hypotheses, the synchronized IQ signals are blindly equalized in multiple parallel branches corresponding one-to-one with each candidate modulation type in the candidate modulation type set, obtaining blindly equalized IQ signals under each candidate modulation type hypothesis. In the deep modulation classification submodule based on fusion features, high-order cyclic cumulant features and constellation diagram features are extracted from the multi-path blindly equalized IQ signals and then merged into multiple branches. The resulting multi-branch merged features are input into the modulation classification network for modulation classification and discrimination to determine the modulation type of the IQ signal to be identified.
[0034] In an optional embodiment, the linear equalization process includes the following steps: The linearly equalized IQ signal is determined by performing a convolution operation between the synchronized IQ signal and the linear equalizer tap vector corresponding to any candidate modulation type, and its expression is as follows:
[0035] in, The signal after linear equalization For the first of length , The candidate modulation type corresponds to the tap vector of the linear equalizer. gather; The synchronized IQ signal; For in the complex field On E 3D vector space.
[0036] Further optionally, the constellation quantization constraint step includes: For the linearly equalized signal, the distance between each IQ sample point and each standard constellation point in the standard constellation set corresponding to multiple candidate modulation types is determined, and the standard constellation point with the smallest distance is determined as the constellation quantization result of the corresponding IQ sample point. Based on the quantization result and the linearly equalized signal, a direct approximation is constructed to obtain the quantized constrained signal, whose expression is as follows:
[0037]
[0038] in, The signal after quantization constraints; This represents the gradient blocking operation, applied to variables of arbitrary vector form. ,satisfy ; Candidate modulation type p The total number of corresponding standard constellation points, Candidate modulation type p The quantitative results Candidate modulation type p The quantization result of the kth sampling point Candidate modulation type p No. k The linearly equalized signal at each sampling point; Candidate modulation type p The first in the corresponding standard constellation point set i A standard constellation point; In order to be with the first k The index of the standard constellation point that minimizes the Euclidean distance between signals after linear equalization.
[0039] In this embodiment, gradient flow in the quantization process is ensured by constructing a direct approximation.
[0040] In an optional embodiment, the waveform reconstruction process also involves gradient backpropagation iterative optimization based on the loss function to obtain the tap vector estimate of the linear equalizer and the vector estimate of the equivalent channel impulse response under the current candidate modulation type assumption. The loss function includes a first error between the reconstructed waveform signal and the synchronized IQ signal, and a second error corresponding to the constellation nearest neighbor quantization error, and its expression is as follows:
[0041] in, Let be the loss function, where Candidate modulation type p The equivalent channel impulse response vector, To quantify the loss weights, the default value is... .
[0042] More specifically, by minimizing the loss function Through gradient backpropagation and iterative optimization, the solution is continuously updated and obtained to obtain the th... Estimation of equalizer tap vectors in branches Estimation of the equivalent channel impulse response vector ,Right now:
[0043] Finally, the first Equalized IQ signal of the branch .
[0044] like Figure 4 The diagram shown is a schematic of the blind balancing process.
[0045] In an optional embodiment, the extraction step of the higher-order cyclic cumulant feature includes: For any candidate modulation type, after blind equalization of the IQ signal, the higher-order cyclic cumulants are extracted according to the order of the higher-order cyclic cumulants and the number of corresponding conjugate terms under multiple cyclic harmonic orders. The extracted higher-order cyclic cumulants are moduloed, and the moduli of each higher-order cyclic cumulant are arranged according to the combination relationship between the cyclic harmonic order and the order of the higher-order cyclic cumulant and the number of conjugate terms, forming a higher-order cyclic cumulant feature matrix corresponding to the candidate modulation type, the expression of which is as follows:
[0046] in, It is the set of cyclic harmonic orders; For any A set of higher-order cyclic cumulants extracted from first-order cyclic harmonics; for under cyclic harmonics Accumulated amount of higher-order cycles; The order of the cumulant in a higher-order cycle; This is the set of orders of the extracted higher-order cyclic cumulants; The number of terms in the product conjugate of the higher-order cyclic cumulants; For the order of cumulant in higher-order cycles The set of the number of terms of the selected conjugate terms. It is a high-order cyclic cumulant characteristic matrix.
[0047] As an example, the network parameters of the constellation graph feature encoder, the deep residual convolutional neural network, and the fully connected classifier used to generate the constellation graph features are determined using a supervised learning method. The supervised learning uses blindly equalized IQ signals with true modulation type labels as training samples, performs high-order cyclic cumulant (HOC) feature extraction, constellation graph feature extraction, and multi-branch feature fusion on the training samples, and iteratively updates the network parameters of the constellation graph feature encoder, the deep residual convolutional neural network, and the fully connected classifier based on the cross-entropy loss between the modulation classification result obtained by the fully connected classifier and the true modulation type label.
[0048] As another example, the default value is... , , , , .at this time, HOC feature matrix ,Right now
[0049] In an optional embodiment, the step of extracting the constellation map features includes: Divide the two-dimensional complex plane region into multiple grid cells at equal intervals within a preset boundary radius; For any candidate modulation type, after blind equalization of the IQ signal, the number of IQ samples falling into each grid cell in the blind equalization IQ signal is counted and normalized to obtain the initial density matrix of the constellation diagram corresponding to the candidate modulation type. The initial density matrix of the constellation graph is input into the constellation graph feature encoder, which includes multiple parallel convolutional branches. Each parallel convolutional branch uses multiple two-dimensional convolutional layers to encode the initial density matrix of the constellation graph and uses convolutional kernels of different sizes to extract the constellation graph density features. Two-dimensional adaptive average pooling is performed on the constellation graph density features extracted by the parallel convolution branch to obtain a constellation graph density feature matrix with the same two-dimensional size as the higher-order cyclic cumulant feature matrix. The constellation graph density feature matrices output from multiple parallel convolution branches are concatenated along the channel dimension to obtain the constellation graph density feature tensor corresponding to the candidate modulation type.
[0050] As an example, ( Take 2) and divide the two-dimensional complex plane into equal intervals and uniformly. grid cells ( ,like ). Statistical analysis of the number of cells falling into each grid cell. The number of elements, and normalized to Later obtained the first Initial density matrix of the constellation diagram with branches , its first Line 1 The column elements represent the density of sampling points falling into the corresponding grid cell. After passing through the constellation graph feature encoder, we obtain multiple constellation graph density feature matrices. Constellation graph density feature tensor The number of channels Specifically, Input to There are 1 parallel convolutional branch, each branch employing multiple layers ( Feature encoding is performed using two-dimensional convolutional layers with kernels of different sizes (e.g., 1×3, 3×1, 3×3, etc.), followed by two-dimensional adaptive average pooling layers. Pooling mapping is A feature matrix of dimension 1. The outputs of each branch are concatenated along the channel dimension to obtain the constellation graph density feature tensor. .
[0051] In an optional embodiment, the step of extracting and merging the higher-order cyclic cumulant features and constellation graph features respectively includes: For any candidate modulation type, the corresponding high-order cyclic cumulant feature matrix and constellation diagram density feature tensor are concatenated along the feature channel dimension to obtain the merged feature corresponding to the candidate modulation type. The fusion of merged features corresponding to multiple candidate modulation types includes: The merged features corresponding to multiple candidate modulation types are stacked according to the branch dimension corresponding to the candidate modulation type to form the multi-branch merged feature.
[0052] Alternatively, the feature matrix of the higher-order cyclic cumulants of each branch can be further... With constellation graph density feature tensor By concatenating along the feature channel dimension, the merged feature matrix of this branch is obtained. All The merged feature matrices of each branch are stacked according to the branch dimension to form a multi-branch merged feature tensor. This is used as the input for the subsequent modulation classification network.
[0053] like Figure 5 The diagram shown illustrates the feature extraction and classification process.
[0054] In an optional embodiment, the deep residual convolutional neural network performs multi-level residual feature extraction and spatial dimensionality reduction on the multi-branch merging features to obtain a fused feature vector; The fully connected classifier performs modulation classification based on the fused feature vector; The deep residual convolutional neural network includes a residual feature extraction module and a directional downsampling module with stride convolution. The residual feature extraction module includes a main branch and a bypass connection. The main branch sequentially performs convolution, batch normalization, activation, convolution, and batch normalization on the input features, and then adds the output of the main branch to the output of the bypass connection before performing activation. The directional downsampling module performs downsampling along the corresponding feature dimension through stride convolution.
[0055] In this embodiment, the main branch sequentially performs Conv2d-BN-ReLU-Conv2d-BN operations, adds them to the output of the bypass connection, and then activates them through ReLU to obtain the residual block output. When the input and output dimensions are consistent, the bypass connection is an identity mapping. When the dimensions are inconsistent, 1×1 convolution is used for dimension alignment. This indicates a directional downsampling module with stride convolution, consisting of... It consists of a layer and a BN+ReLU layer, used to increase the number of channels from Become and along The corresponding feature dimensions are downsampled by a factor of two, while maintaining... The corresponding feature dimensions remain unchanged. The Flatten layer represents using PyTorch's standard torch.nn.Flatten operation to flatten the 128×1×1 features into a 128-dimensional vector.
[0056] Finally, the output of the modulation classification network is used. Determine the modulation type ,in .
[0057] In an optional embodiment, the candidate modulation type set includes multiple digital modulation standards with a standard constellation set, wherein the multiple digital modulation standards belong to at least one modulation type among M-PSK, M-QAM, M-ASK and M-APSK.
[0058] Example 2 This embodiment demonstrates the application of the robust modulation identification method for channel-variable scenarios proposed in Embodiment 1 in a multipath Rayleigh fading channel scenario simulating dynamic channel changes, and compares it with existing commonly used technologies.
[0059] First, ideal signals with different modulations are generated, as shown in Table 1. Then, AWGN and multipath Rayleigh fading channel scenarios are simulated, i.e., distortion / error caused by different channel fading factors in actual communication is added to the ideal signals to obtain training signal samples. The channel distortion factors and parameter values of the training signal samples are shown in Table 2. Finally, differentiated test signal samples for multipath Rayleigh fading channels are generated, as shown in Table 3. These test signal samples are used to simulate and test the modulation recognition accuracy performance of each method in a multipath Rayleigh fading channel scenario with dynamic channel changes.
[0060] Table 1 Ideal signal parameters
[0061] Table 2 Channel distortion factors and parameter values for training signal samples
[0062] Table 3 Channel distortion factors and parameter values for test signal samples
[0063] In the simulation, the parameter values of each module of the method in this application are shown in Table 4.
[0064] Table 4. Parameter values of the method in this application during simulation.
[0065] The performance of the proposed method was compared and verified through simulation with typical deep learning modulation recognition methods such as high-order cumulant-SVM, constellation graph-CNN, and IQ-CNN. Figure 5 The image shows a comparison of the average recognition accuracy of different methods; Figure 6 It can be seen that under dynamic channel conditions across scenarios (where the channel parameters of the test set and the training set are significantly different), the recognition performance of the method of this invention is significantly better than that of other comparative methods across all test signal-to-noise ratio (SNR) ranges. In the low SNR range of 2–10 dB, the average recognition accuracy of the method of this invention increases significantly with the increase of SNR, with the recognition accuracy increasing from 63% to 86%, and the accuracy is at least 24% better than other comparative methods; when the SNR continues to rise to 13 dB and above, the average recognition accuracy of the method in this embodiment stably exceeds 95%, which is at least 20% better than other methods.
[0066] like Figure 7 The figure shows the confusion matrix of the identification results of the seven modulation types in this application at a signal-to-noise ratio of 12dB.
[0067] Depend on Figure 7 It can be seen that the method of this application has a good differentiation effect on all 7 modulation types. Only a few modulation types with similar constellation structures are misjudged. The overall accuracy of the identification of various modulation samples is relatively high, which verifies the reliable modulation differentiation ability of this method under medium signal-to-noise ratio variable channels.
[0068] Example 3 This embodiment proposes a storage medium storing computer-readable instructions, wherein when the computer-readable instructions are executed by a processor, they implement the steps of a robust modulation identification method for channel-variable scenarios proposed in Embodiment 1.
[0069] By way of example, the storage medium includes, but is not limited to, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks or optical disks, and other media capable of storing program code.
[0070] By way of example, the instructions, programs, code sets, or instruction sets may be implemented using conventional programming languages.
[0071] By way of example, the processor includes, but is not limited to, smartphones, personal computers, servers, network devices, etc., for executing all or part of the steps of the robust modulation identification method for channel-variable scenarios described in Example 1.
[0072] The terminology used in the accompanying drawings is for illustrative purposes only.
[0073] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A robust modulation identification method for scenarios with varying channel characteristics, characterized in that, include: The IQ signal to be identified is received and blind synchronization preprocessing independent of modulation type is performed to obtain the synchronized IQ signal. For multiple candidate modulation types in the candidate modulation type set, the linear equalizer of the synchronized IQ signal under any candidate modulation type assumption is processed, the linearly equalized signal is constellation quantization constellation constraint is performed on the standard constellation set corresponding to the candidate modulation type, and the waveform is reconstructed based on the equivalent channel impulse response to obtain the blindly equalized IQ signal. For any candidate modulation type, the blindly equalized IQ signal is extracted and merged with the high-order cyclic cumulant features and constellation diagram features respectively; the merged features corresponding to multiple candidate modulation types are fused to obtain multi-branch merged features. The multi-branch merging features are input into a modulation classification network containing a deep residual convolutional neural network and a fully connected classifier to perform modulation classification and discrimination, thereby determining the modulation type of the IQ signal to be identified.
2. The robust modulation recognition method for channel-variable scenarios according to claim 1, characterized in that, The received IQ signal to be identified undergoes blind synchronization preprocessing independent of modulation type, including: Blind frequency offset estimation and frequency offset compensation are performed on the IQ signal to be identified to obtain the frequency offset compensated IQ signal; Blind symbol rate estimation is performed on the frequency offset compensated IQ signal, and downsampling is performed based on the estimated symbol rate to obtain an IQ signal with a single symbol sampling rate. Blind symbol timing synchronization is performed on the IQ signal at the single symbol sampling rate to obtain the timing-upgraded IQ signal; Blind phase bias estimation and phase bias correction are performed on the timing-post IQ signal to obtain the synchronized IQ signal; Specifically, the blind frequency offset estimation adopts blind frequency offset estimation based on joint peak search of higher-order cyclic cumulants of the signal, the blind symbol rate estimation adopts symbol rate blind estimation based on cyclic spectrum peak detection, the blind symbol timing synchronization adopts Gardner timing synchronization, and the blind phase offset estimation adopts phase offset estimation based on nonlinear minimum mean square criterion.
3. The robust modulation recognition method for channel-variable scenarios according to claim 1, characterized in that, The steps of linear equalization processing include: The linearly equalized IQ signal is determined by performing a convolution operation between the synchronized IQ signal and the linear equalizer tap vector corresponding to any candidate modulation type, and its expression is as follows: in, The signal after linear equalization For length is The The candidate modulation type corresponds to the tap vector of the linear equalizer. A set; The synchronized IQ signal; For in the complex field On E 3D vector space.
4. The robust modulation identification method for channel-variable scenarios according to claim 3, characterized in that, The steps for constellation quantification constraints include: For the linearly equalized signal, the distance between each IQ sample point and each standard constellation point in the standard constellation set corresponding to the candidate modulation type is determined, and the standard constellation point with the smallest distance is determined as the constellation quantization result of the corresponding IQ sample point. Based on the quantization result and the linearly equalized signal, a direct approximation is constructed to obtain the quantized constrained signal, whose expression is as follows: in, The signal after quantization constraints; This represents the gradient blocking operation, applied to variables of arbitrary vector form. ,satisfy ; Candidate modulation type p The total number of corresponding standard constellation points, Candidate modulation type p The quantitative results Candidate modulation type p No. k Quantization results of each sampling point Candidate modulation type p No. k The signal after linear equalization at each sampling point K The number of sampling points; Candidate modulation type p The first in the corresponding standard constellation point set i A standard constellation point; In order to be with the first k The index of the standard constellation point that minimizes the Euclidean distance between signals after linear equalization.
5. The robust modulation identification method for channel-variable scenarios according to claim 4, characterized in that, During the waveform reconstruction process, gradient backpropagation iterative optimization is also performed based on the loss function to obtain the tap vector estimate of the linear equalizer and the vector estimate of the equivalent channel impulse response under the current candidate modulation type assumption. The loss function includes a first error between the reconstructed waveform signal and the synchronized IQ signal, and a second error corresponding to the constellation nearest neighbor quantization error, and its expression is as follows: in, Let be the loss function, where Candidate modulation type p The equivalent channel impulse response vector, To quantify the loss weights.
6. The robust modulation identification method for channel-variable scenarios according to any one of claims 1 to 5, characterized in that, The extraction steps for the higher-order cyclic cumulant features include: For any candidate modulation type, after blind equalization of the IQ signal, the higher-order cyclic cumulants are extracted according to the order of the higher-order cyclic cumulants and the number of corresponding conjugate terms under multiple cyclic harmonic orders. The extracted higher-order cyclic cumulants are moduloed, and the moduli of each higher-order cyclic cumulant are arranged according to the combination relationship between the cyclic harmonic order and the order of the higher-order cyclic cumulant and the number of conjugate terms, forming a higher-order cyclic cumulant feature matrix corresponding to the candidate modulation type, the expression of which is as follows: in, It is the set of cyclic harmonic orders; For any A set of higher-order cyclic cumulants extracted from first-order cyclic harmonics; for under cyclic harmonics Accumulated amount of higher-order cycles; The order of the cumulant in a higher-order cycle; This is the set of orders of the extracted higher-order cyclic cumulants; The number of terms in the product conjugate of the higher-order cyclic cumulants; For the order of cumulant in higher-order cycles The set of the number of terms of the selected conjugate terms. It is a high-order cyclic cumulant characteristic matrix.
7. The robust modulation identification method for channel-variable scenarios according to claim 6, characterized in that, The steps for extracting the constellation diagram features include: Divide the two-dimensional complex plane region into multiple grid cells at equal intervals within a preset boundary radius; For any candidate modulation type, after blind equalization of the IQ signal, the number of IQ samples falling into each grid cell in the blind equalization IQ signal is counted and normalized to obtain the initial density matrix of the constellation diagram corresponding to the candidate modulation type. The initial density matrix of the constellation graph is input into the constellation graph feature encoder, which includes multiple parallel convolutional branches. Each parallel convolutional branch uses multiple two-dimensional convolutional layers to encode the initial density matrix of the constellation graph and uses convolutional kernels of different sizes to extract the constellation graph density features. Two-dimensional adaptive average pooling is performed on the constellation graph density features extracted by the parallel convolution branch to obtain a constellation graph density feature matrix with the same two-dimensional size as the higher-order cyclic cumulant feature matrix. The constellation graph density feature matrices output from multiple parallel convolution branches are concatenated along the channel dimension to obtain the constellation graph density feature tensor corresponding to the candidate modulation type.
8. The robust modulation identification method for channel-variable scenarios according to claim 7, characterized in that, The step of extracting and merging the high-order cyclic cumulant features and constellation graph features separately includes: For any candidate modulation type, the corresponding high-order cyclic cumulant feature matrix and constellation diagram density feature tensor are concatenated along the feature channel dimension to obtain the merged feature corresponding to the candidate modulation type. The fusion of merged features corresponding to multiple candidate modulation types includes: The merged features corresponding to multiple candidate modulation types are stacked according to the branch dimension corresponding to the candidate modulation type to form the multi-branch merged feature.
9. The robust modulation identification method for channel-variable scenarios according to claim 8, characterized in that, The deep residual convolutional neural network performs multi-level residual feature extraction and spatial dimensionality reduction on the multi-branch merged features to obtain a fused feature vector. The fully connected classifier performs modulation classification based on the fused feature vector; The deep residual convolutional neural network includes a residual feature extraction module and a directional downsampling module with stride convolution. The residual feature extraction module includes a main branch and a bypass connection. The main branch sequentially performs convolution, batch normalization, activation, convolution, and batch normalization on the input features, and then adds the output of the main branch to the output of the bypass connection before performing activation. The directional downsampling module performs downsampling along the corresponding feature dimension through stride convolution.
10. The robust modulation identification method for channel-variable scenarios according to claim 1, characterized in that, The candidate modulation type set includes multiple digital modulation standards with standard constellation sets, wherein the multiple digital modulation standards belong to at least one modulation type among M-PSK, M-QAM, M-ASK and M-APSK.