Multi-modulation signal robust identification method based on time-frequency diagram and data enhancement

Abstract

The invention provides a multi-modulation signal robust identification method based on a time-frequency diagram and data enhancement. Original baseband signals of the traditional modulation mode and the novel modulation mode are generated and preprocessed; performing data set division according to a double sampling mode; performing STFT on the data to generate a time-frequency graph; performing enhancement operations of time axis offset, frequency axis offset, noise injection and tiny rotation on the time-frequency graph of the training set; a ResNet-18 model adaptive to single channel input is constructed; and training the model and carrying out a performance test of unknown signal modulation mode identification. According to the method, the novel data set comprising a traditional modulation mode and an emerging modulation technology is constructed, the defect that most of the current modulation data sets are the traditional modulation mode is overcome, a data basis closer to a future communication scene is provided for modulation recognition in a complex scene, multi-dimensional enhancement is performed on data, data diversity is improved, and the method is suitable for being applied to modulation recognition in a complex scene. The model generalization ability is significantly improved, and the identification precision of the modulation signal is also effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication signal processing, in particular to modulation signal recognition technology, and further relates to image classification technology based on deep learning and data enhancement. BACKGROUND

[0002] In the field of communication, modulation recognition technology identifies the modulation parameters and methods used by the transmitted signal, providing key information for subsequent signal detection and demodulation. As a prerequisite and foundation for signal analysis, modulation recognition technology has been widely used in military and civilian fields, such as radio spectrum detection, cognitive radio, signal interception, target locking and positioning, etc. However, in recent years, wireless communication technology has continued to iterate and upgrade, signal modulation methods have become increasingly complex and diverse, and channel environments have been increasingly affected by noise, interference and other factors, leading to serious challenges in modulation signal recognition.

[0003] Traditional modulation recognition technology is divided into maximum likelihood estimation and feature extraction-based recognition. Maximum likelihood estimation recognition is based on strict mathematical derivation, which maximizes the likelihood function to achieve accurate prediction. However, this method relies too much on prior information and has high computational complexity, limiting its use in certain scenarios. Feature extraction-based recognition requires manual extraction of useful features from the signal, such as high-order statistics, constellation information, and cyclic spectrum, and then uses a classifier for recognition. Compared with maximum likelihood estimation, this method has lower computational complexity, but the final recognition result relies heavily on the professional level and recognition experience of manual feature extraction, which cannot guarantee the accuracy of recognition.

[0004] In recent years, with the continuous development of machine learning in various fields, deep learning, as a branch of machine learning, has also been widely applied. Combining deep learning with modulation recognition has become the main trend. The advantage of using deep learning for recognition is that it automatically extracts signal features by training neural networks, without relying on signal prior knowledge and human experience, and can capture more complex and subtle changes in the signal, especially in low signal-to-noise ratio environments, showing stronger robustness and generalization ability. Currently, Convolutional Neural Network (CNN) has been widely used in image classification and recognition due to its strong image feature extraction ability; Recurrent Neural Network (RNN) is good at capturing temporal features and is suitable for processing language and sequence signals. Compared with traditional methods, deep learning-based modulation recognition can extract more complex features and achieve more stable high recognition accuracy in multi-interference scenarios.

[0005] However, the existing deep learning-based modulation recognition method still has significant shortcomings:

[0006] 1. Under low signal-to-noise ratio conditions, the signal is affected by noise and interference, and the neural network recognition performance is greatly degraded, making it difficult to meet the application requirements of complex channels;

[0007] 2. The training process is prone to overfitting due to insufficient data diversity, which limits the model generalization ability;

[0008] 3. Most of the current modulation recognition datasets are for traditional modulation methods, and the adaptability to new modulation methods in emerging communication scenarios such as 5G / 6G is poor, making it difficult to capture their unique signal characteristics. SUMMARY

[0009] To solve the above problems, the present application provides a multi-modulation signal robust recognition method based on time-frequency graph and data enhancement. This method combines signal time-frequency conversion, data enhancement expansion, and deep learning classification to achieve high-precision recognition of multiple types of modulation methods in complex channel environments. Specifically, generate original signals covering traditional and new modulation methods and preprocess to obtain sample signals; divide the sample signals into datasets; perform Short-Time Fourier Transform (STFT) on the sample signals in each dataset to generate corresponding two-dimensional time-frequency graphs; perform multi-dimensional data enhancement on the time-frequency graphs of the training set to improve data diversity; construct a ResNet-18 adapted to single-channel input as a network model for training to identify the modulation method of unknown signals.

[0010] A multi-modulation signal robust recognition method based on time-frequency graph and data enhancement, comprising the following steps:

[0011] Step 1: Generate original baseband signals of traditional modulation methods (such as BPSK, QPSK, 16QAM, etc.) and new modulation methods (such as OTFS, UFMC, etc.), and add noise, introduce multipath fading, etc. Interference to simulate actual complex channel environment, and finally normalize to obtain sample IQ signals. Sample signals include signal sequences, modulation method labels, and signal-to-noise ratio labels;

[0012] Step 2: Divide the sample signals into datasets and divide them into training set, test set, and validation set in the ratio of 6:2:2;

[0013] Step 3: Use STFT to generate time-frequency graphs for sample signals in each dataset;

[0014] Step 4: Perform multi-dimensional data enhancement on the time-frequency graphs of the training set, including time axis shift, frequency axis shift, Gaussian noise injection, and small angle rotation, etc.

[0015] Step five: build a ResNet-18 network model suitable for single-channel input, and introduce a Dropout layer to suppress overfitting.

[0016] Step six: use the enhanced training set to train the network model, use the validation set to adjust the network's hyperparameters, and finally use the test set to evaluate the network performance.

[0017] Further, in step 1, the traditional modulation modes include BPSK, QPSK, GFSK, 16QAM, and 64QAM, which are typical modulation modes in most current modulation data sets;

[0018] The new modulation modes include Universal Filtered Multi-Carrier (UFMC), Filtered Orthogonal Frequency Division Multiplexing (F-OFDM), Filter Bank Multi-Carrier-Offset Quadrature Amplitude Modulation (FBMC-OQAM), Generalized Frequency Division Multiplexing (GFDM), Orthogonal Frequency Division Multiplexing with Index Modulation (OFDM-IM), and Orthogonal Time Frequency Space (OTFS). These modulation modes are cutting-edge technologies in 5G and Beyond 5G communication, and their specific characteristics are as follows:

[0019] UFMC: By allocating and modulating the transmitted signal to multiple independent subcarriers, the symbol rate on the subchannel is reduced, effectively suppressing system interference and improving spectral efficiency and transmission reliability.

[0020] F-OFDM: The entire communication frequency band is divided into multiple subbands, each subband is filtered to reduce out-of-band attenuation and interference, and the parameters within the subband can be flexibly configured to adapt to various application scenarios.

[0021] FBMC-OQAM: Using a prototype filter with good time-frequency focusing, the signal on each subcarrier is individually shaped, and the in-band interference is eliminated through offset quadrature, with low out-of-band leakage, suitable for non-orthogonal transmission scenarios.

[0022] GFDM: Employs a two-dimensional time-frequency data modulation scheme, which precisely manages the allocation of each subcarrier and sub-symbol, and can flexibly respond to different quality of service requirements;

[0023] OFDM-IM: By transmitting additional index information through the active state of subcarriers, it reduces system energy consumption while improving spectrum efficiency;

[0024] OTFS: Modulates signals in the time-delay-Doppler domain, making more efficient use of pilot resources;

[0025] The preprocessing operations performed on the raw IQ signal are as follows:

[0026] First, additive white Gaussian noise (AWGN) is added to the original IQ signal s(t) to obtain a signal with AWGN:

[0027] x(t) = s(t) + n(t)

[0028] Next, based on the superimposed AWGN, three types of interference—multipath fading, Doppler shift, and carrier frequency offset—are introduced to simulate the real channel environment:

[0029] Multipath fading occurs when, in complex communication environments, the receiver may receive signals from different paths, which then superimpose. The mathematical expression is:

[0030]

[0031] Where N is the number of multipaths, α n (t) is the time-varying amplitude attenuation coefficient of the signal along different paths, τ n (t) represents the transmission delay along different paths, f c It is the carrier frequency. These are Doppler phase shifts along different paths;

[0032] Doppler shift refers to the frequency shift of a signal caused by high-speed movement, and its expression is:

[0033]

[0034] in,

[0035]

[0036] f d θ is the Doppler frequency shift, v is the relative velocity between the transmitter and receiver, c is the speed of light, and θ is the angle between the direction of motion and the direction of signal incidence.

[0037] Carrier frequency offset, expressed as:

[0038] x cfo (t) = x(t)e j2πΔft Where Δf is the random frequency offset;

[0039] Let x'(t) represent the signal after adding interference, and finally normalize the energy of x'(t) to calculate the total energy of the signal:

[0040]

[0041] Where L is the length of the signal sequence, divide each sample in the signal sequence by the square root of the energy, and get the energy-normalized signal sequence:

[0042]

[0043] Where ε = 10 -10 , eliminate the difference in signal amplitude and power, and ensure that all sample energies are consistent.

[0044] Further, in step 2, the data set is divided by a double stratified sampling strategy, taking modulation mode and signal-to-noise ratio as stratification basis, to ensure that the data in the training set, test set and validation set are uniformly distributed under different modulation modes and signal-to-noise ratios, and the interference scene is randomly allocated, which not only ensures the balanced distribution of core features, but also retains the randomness of interference occurrence.

[0045] Further, in step 3, the STFT uses a Hamming window as the window function, and the window length is dynamically adapted to the signal length, selecting the smaller value of the preset value and the length of a single signal sample, and the overlap ratio of adjacent windows is set to 50%.

[0046] Further, in step 4, in order to enhance the robustness of the model to sudden changes in complex channel environment, multi-dimensional data augmentation is performed on the training data, including: to simulate the different step characteristics of different signals in time, a random time shift is performed on the time-frequency graph; to simulate the occurrence of frequency selective fading in the channel, a random frequency shift is performed on the time-frequency graph; random noise is added to the time-frequency graph to enhance the noise resistance of the model; to simulate the random occurrence of small distortions in the signal during transmission, a small-angle random rotation is performed on the time-frequency graph, and the above enhancement operations are randomly added to the training data set to increase data diversity, introduce more features, and avoid overfitting of the model due to excessive reliance on original training data.

[0047] Further, in step 5, a ResNet-18 network suitable for single-channel input is constructed, which is composed of multiple convolution layers cascaded, can deeply extract subtle features of the time-frequency graph, and solves the gradient disappearance problem of the network structure deep through multiple residual structures; a Dropout layer is introduced in the fully connected layer, and the output of part of the neurons is randomly shielded during the training process, so that the model is forced to learn more generalizable features.

[0048] Further, in step 6, the ResNet-18 is trained using the enhanced training data set, and the model parameters are optimized using the cross-entropy loss function, and the cross-entropy loss function is L:

[0049] L=-∑y i log(p i )

[0050] Wherein, y i is the true label, and p i is the predicted probability;

[0051] The Adam optimizer is selected to update the network parameters, and the learning rate is dynamically updated by combining the ReduceLROnPlateau learning rate scheduling; the hyperparameters are adjusted using the validation set, such as the selection of the optimizer and the learning rate scheduling strategy, to ensure that the model has good performance on the validation set; finally, the trained network model is tested using the test set, and the performance indicators include the modulation mode recognition accuracy, the confusion matrix, etc., to comprehensively verify the generalization ability and recognition accuracy of the model on unknown samples.

[0052] Advantages

[0053] The present application constructs a modulation data set containing traditional and new modulation modes, multiple interference scenarios and multiple signal-to-noise ratio levels, which makes up for the current modulation data set which is mostly traditional modulation modes, adds emerging modulation technologies such as OTFS, FBMC-OQAM, F-OFDM, etc. of 5G and Beyond 5G, provides a more abundant and more close-to-future-communication-scenario data basis for modulation recognition research in complex scenarios, and significantly improves the generalization ability of the network model through the enhancement operation on the training data set, and realizes the deep extraction of time-frequency features while effectively enhancing the model robustness by matching the ResNet-18 network, which is suitable for more complex channel environments. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is the overall flow chart of the signal modulation recognition method of the present application;

[0055] Figure 2 It is the time-frequency graph comparison of BPSK and QPSK with and without data enhancement under 0dB provided by the examples of the present application;

[0056] Figure 3 Figure 1 is a schematic diagram of the loss and average recognition accuracy of the training data and validation data provided by the present application with and without data enhancement, wherein subgraph (a) is a schematic diagram of the loss value, and subgraph (b) is a schematic diagram of the average accuracy;

[0057] Figure 4 Figure 2 is a schematic diagram of the average recognition accuracy of the test data provided by the present application with and without data enhancement with respect to the change in signal-to-noise ratio;

[0058] Figure 5 Figure 3 is a schematic diagram of the confusion matrix of the network model classification ability test provided by the present application, wherein the signal-to-noise ratios of subgraphs (a) and (b) are -10 dB and 0 dB, respectively;

[0059] Figure 6 Figure 4 is a schematic diagram of the network model recognition accuracy for each modulation mode provided by the present application. DETAILED DESCRIPTION

[0060] The present application is further described below in conjunction with the accompanying drawings and examples, but should not be limited in scope by the present application.

[0061] As shown in Figure 1 :

[0062] A multi-modulation signal robust recognition method based on time-frequency graph and data enhancement, comprising the following steps:

[0063] Step 1: Generate original baseband signals of traditional modulation modes (such as BPSK, QPSK, 16QAM, etc.) and new modulation modes (such as OTFS, UFMC, etc.), and perform noise addition, introduce multipath fading and other interference to simulate actual complex channel environment, and finally perform normalization processing to obtain sample IQ signals. The sample signal includes a signal sequence and modulation mode label and signal-to-noise ratio label;

[0064] Step 2: Divide the sample signal into a training set, a test set, and a validation set in a ratio of 6:2:2.

[0065] Step 3: Use STFT to generate a time-frequency graph for the sample signal in each data set.

[0066] Step 4: Perform multi-dimensional data enhancement on the time-frequency graph of the training set, including time axis shift, frequency axis shift, Gaussian noise injection, and small angle rotation, etc.

[0067] Step 5: Construct a ResNet-18 network model suitable for single-channel input, and suppress overfitting by introducing a Dropout layer.

[0068] Step six: train the network model with the enhanced training set, adjust the network's hyperparameters with the validation set, and finally evaluate the network performance with the test set.

[0069] Further, in step 1, the traditional modulation modes include BPSK, QPSK, GFSK, 16QAM, and 64QAM, which are typical modulation modes in most current modulation data sets;

[0070] The new modulation modes include Universal Filtered Multi-Carrier (UFMC), Filtered Orthogonal Frequency Division Multiplexing (F-OFDM), Filter Bank Multi-Carrier-Offset Quadrature Frequency Division Multiplexing (FBMC-OQAM), Generalized Frequency Division Multiplexing (GFDM), Index Modulation Based Orthogonal Frequency Division Multiplexing (OFDM-IM), and Orthogonal Time Frequency Space Modulation (OTFS), which are all cutting-edge technologies in 5G and Beyond5G communication;

[0071] The preprocessing operation on the original IQ signal is as follows:

[0072] First, add Additive White Gaussian Noise (AWGN) to the original IQ signal s(t) to obtain the signal with AWGN:

[0073] x(t)=s(t)+n(t)

[0074] Next, on the basis of superimposing AWGN, introduce three kinds of interference, namely multipath fading, Doppler shift, and carrier frequency offset, to simulate the real channel environment:

[0075] Multipath fading is when the communication environment is complex, the receiving end may receive signals from different paths and superimpose them, and the mathematical expression is:

[0076]

[0077] Where N is the number of multipath, α n (t) is the amplitude time-varying attenuation coefficient of different path signals, τ n (t) is the transmission delay of different paths, f c is the carrier frequency, is the Doppler phase shift of different paths;

[0078] Doppler shift refers to the frequency shift of the signal caused by high-speed movement, and the expression is:

[0079]

[0080] Where,

[0081]

[0082] f d is the Doppler shift, v is the relative motion speed between the sending end and the receiving end, c is the speed of light, and θ is the included angle between the motion direction and the signal incidence direction;

[0083] The carrier frequency offset is expressed as:

[0084] x cfo (t) = x(t)e j2πΔft where Δf is the random frequency offset;

[0085] Let x'(t) represent the signal after adding the interference, and finally perform energy normalization processing on x'(t) to calculate the total energy of the signal as:

[0086]

[0087] where L is the length of the signal sequence, and each sample in the signal sequence is divided by the square root of the energy to obtain the energy-normalized signal sequence as:

[0088]

[0089] where ε = 10 -10 , eliminating the difference in signal amplitude and power to ensure that all sample energies are consistent.

[0090] Further, in step 2, the data set division adopts a double stratified sampling strategy, taking the modulation mode and the signal-to-noise ratio as the stratification basis to ensure that the proportion distribution of the data in the training set, the test set, and the validation set under different modulation modes and signal-to-noise ratios remains consistent and uniform, and the interference scene is randomly allocated, which not only ensures the balanced distribution of core features but also retains the randomness of the occurrence of interference.

[0091] As Figure 2 shown:

[0092] Further, in step 3, the STFT adopts a Hamming window as the window function, and the window length is dynamically adapted to the signal length, selecting the smaller value of the preset value and the length of a single signal sample, and the adjacent window overlap ratio is set to 50%.

[0093] Further, in step 4, in order to enhance the robustness of the model to the mutation of the complex channel environment, multi-dimensional data enhancement is performed on the training data, including: in order to simulate the different synchronization characteristics of different signals in time, the time-frequency graph is randomly offset in time; in order to simulate the frequency selective fading of the channel, the time-frequency graph is randomly offset in frequency; random noise is added to the time-frequency graph to enhance the noise resistance of the model; in order to simulate the random micro-distortion of the signal in the transmission process, the time-frequency graph is randomly rotated by a small angle, and the above enhancement operations are randomly added to the training data set to increase the diversity of the data, introduce more features, and avoid overfitting of the model due to excessive dependence on the original training data.

[0094] Further, in step 5, a ResNet-18 network suitable for single-channel input is constructed, which is composed of multiple convolution layers to extract fine features of the time-frequency graph in depth, and multiple residual structures to solve the gradient disappearance problem of deep network structure; a Dropout layer is introduced in the fully connected layer to randomly mask the output of part of the neurons during training, forcing the model to learn more generalizable features.

[0095] Further, in step 6, the enhanced training data set is used to train the ResNet-18, and the Adam optimizer is selected to update the network parameters, and the learning rate is dynamically updated by combining the ReduceLROnPlateau learning rate scheduling; the hyperparameters are adjusted using the validation set, such as the optimizer, learning rate scheduling strategy, etc., to ensure that the model performs well on the validation set; finally, the test set is used to test the trained network model, and the performance indicators include modulation mode recognition accuracy, confusion matrix, etc., to comprehensively verify the generalization ability and recognition accuracy of the model on unknown samples.

[0096] Embodiment 1:

[0097] 1. In this embodiment, the experimental environment is as follows:

[0098] Hardware environment: a computing device equipped with an NVIDIA RTX 3090 graphics card (24 GB of video memory) is used;

[0099] Software environment: MATLAB 2023b is used for signal sample generation, and model training is implemented based on Python language and Pytorch framework.

[0100] 2. The specific parameters used in the signal samples generated in the simulation experiment of the present application are as follows:

[0101] Table 1. Modulation parameter table of 11 modulation modes

[0102]

[0103] According to the modulation parameters set in Table 1, each modulation mode generates 1000 original baseband signal samples, and the signal length is set to 4096, including IQ two-way signals.

[0104] The signal-to-noise ratio interval is set to -10dB to 20dB, with an interval of 2dB, and the noise is AWGN. All samples are interfered by AWGN. Among them, 1500 noisy signal samples are generated under each signal-to-noise ratio from -10dB to 0dB, and 1000 noisy signal samples are generated under each signal-to-noise ratio from 2dB to 20dB. Random interference is loaded in all samples to ensure the diversity of interference scenarios. The interference and its proportion include: 30% of all data samples only add AWGN to simulate a more ideal channel environment; 40% of the samples add multipath interference, with a multipath number of 2 / 4 / 6 paths and a time delay of 0 to 5us to simulate the case of more obstructions in urban environment; 15% of the samples add Doppler shift, with a frequency shift of 100 / 500Hz to simulate high-speed moving scenarios; the remaining samples add random frequency offset, with a frequency offset of 200 / 500Hz, corresponding to the frequency synchronization deviation scenario of the transmitting and receiving ends.

[0105] Combined with 11 modulation modes, a total of 209000 noisy samples are generated, including different channel environments and different noise intensities.

[0106] 3. The data set is divided by stratified sampling, and the division ratio is training set: test set: validation set = 6:2:2.

[0107] The signals of each data set are converted into time-frequency by STFT to generate time-frequency graphs, where the preset value of the sliding window is 256. When the signal length is less than 256, the window length is taken as the signal length; otherwise, 256 is taken. The window overlap degree is set to 50%. After time-frequency conversion, the image is normalized and size adjusted. The samples of the training set are subjected to random data enhancement operation, and the specific operation strategy and parameters are:

[0108] Time dimension enhancement: time offset is set to (-0.1, 0.1);

[0109] Frequency dimension enhancement: frequency offset is set to (-0.1, 0.1);

[0110] Noise enhancement: random injection of noise of different intensities in the time-frequency graph, with a noise standard deviation range of (0, 0.015);

[0111] Rotation enhancement: randomly rotate the time-frequency graph slightly, with a rotation angle of (-3°, 3°);

[0112] Each sample is independently data augmented, and the probability of data augmentation is set to 0.5. The final generated image has a size of (128, 128).

[0113] 4. The input of ResNet-18 is modified from three channels to single-channel image input, and the convolution kernel parameters (kernel_size = 7, padding = 3) remain unchanged; the original maximum pooling layer MaxPool2d is replaced with an identity mapping Identity to slow down the feature map size reduction speed; a Dropout layer is added to the fully connected layer to enhance the model generalization ability, and the p value is set to 0.5; the final output of the fully connected layer is 11, corresponding to 11 modulation modes, and Softmax is used as the activation function to obtain the classification probability.

[0114] 5. The training parameters are as follows:

[0115] Batch size: 128;

[0116] Each training input dimension (batch size, channel number, image height, image width): (128, 1, 128, 128);

[0117] Training rounds: 50 rounds;

[0118] Optimizer and parameters: Adam, learning rate 1e-4, weight decay coefficient 1e-5, β1 = 0.9, β2 = 0.999;

[0119] Learning rate scheduling: ReduceLROnPlateau is used, and the learning rate is reduced by 0.5 proportionally when the validation loss does not decrease for 5 consecutive rounds;

[0120] Early stopping mechanism: when the current validation loss is less than 0.003 from the historical best validation loss, the early stopping mechanism is added by one, and the training stops when the early stopping is triggered for 25 consecutive rounds.

[0121] 6. In this example, the experimental results are as follows:

[0122] As Figure 3 (a) The loss value comparison diagram shows that when there is no data augmentation, the validation loss decreases in the first 10 rounds, and then increases significantly, and the difference between the training loss and the validation loss is greater than 0.3; after introducing data augmentation, the training loss and the validation loss decrease steadily, and the difference does not exceed 0.1, and the model generalization ability is significantly improved;

[0123] As Figure 4As shown, under different signal-to-noise ratios, the classification accuracy with data enhancement is higher than that without data enhancement, and when there is data enhancement, the classification accuracy gradually improves from 82.5% to 94% in the low signal-to-noise ratio (-10dB to 0dB), and the overall recognition accuracy is better than that without data enhancement, which further illustrates that improving data diversity can effectively improve the generalization ability of the model;

[0124] As Figure 5 As shown in (b), under 0dB, the model's prediction of BPSK, FBMC-OQAM, F-OFDM, GFDM, GFSK, OFDM-IM, OTFS, and UFMC is almost 0 except for the diagonal element, and the recognition accuracy is accurate, while 16QAM, 64QAM, and QPSK have certain misjudgment, mainly because their constellation structures are similar, and 16QAM and 64QAM belong to quadrature amplitude modulation, and there is certain similarity in time-frequency characteristics, and in a noisy environment, it is more likely to cause misjudgment.

[0125] As Figure 6 As shown, BPSK, FBMC-OQAM, F-OFDM, OTFS, and UFMC modulation methods have recognition accuracy of 90% or above when the signal-to-noise ratio is greater than or equal to -6dB, and the accuracy does not fluctuate significantly with the change of signal-to-noise ratio, indicating that these modulation methods have strong anti-noise ability, and the recognition accuracy of QPSK, 16QAM, and 64QAM also increases with the increase of signal-to-noise ratio, and finally the recognition accuracy of 16QAM and 64QAM exceeds 70%, and the recognition accuracy of QPSK exceeds 90%, further verifying that the present application is adaptive to different modulation methods.

Claims

1. A robust identification method for multi-modulation signals based on time-frequency graphs and data augmentation, comprising the following steps: Step 1: Generate the original baseband signals of traditional modulation methods and new modulation methods, and add noise, multipath fading, Doppler frequency shift and carrier frequency offset interference to simulate the actual complex channel environment. Finally, perform normalization processing to obtain sample IQ signals. The sample signals include signal sequences, modulation method labels and signal-to-noise ratio labels. Step 2: The sample signal is divided into a training set, a test set, and a validation set using double stratified sampling, with a ratio of 6:2:

2. Step 3: Use STFT on the sample signals in each dataset to generate a time-frequency plot; Step 4: Perform multi-dimensional data augmentation on the time-frequency plot of the training set, specifically including time axis shift, frequency axis shift, Gaussian noise injection, and small-angle rotation operation; Step 5: Construct a ResNet-18 network model adapted for single-channel input, and introduce a Dropout layer with a p-value of 0.5; Step 6: Train the adapted ResNet-18 network model using the enhanced training set, adjust the network's hyperparameters using the validation set, evaluate the network performance using the test set, and output the final modulation scheme recognition result.

2. The method according to claim 1, characterized in that, In step 1, the traditional modulation methods include BPSK, QPSK, GFSK, 16QAM, and 64QAM; The novel modulation methods mentioned include Universal Filtered Multi-Carrier (UFMC), Filtered Orthogonal Frequency Division Multiplexing (F-OFDM), Filter Bank Multi-Carrier-Offset Quadrature Amplitude Modulation (FBMC-OQAM), Generalized Frequency Division Multiplexing (GFDM), Orthogonal Frequency Division Multiplexing with Index Modulation (OFDM-IM), and Orthogonal Time Frequency Space (OTFS). In step 1, the preprocessing operations performed on the original IQ signal are as follows: First, additive white Gaussian noise (AWGN) is added to the original IQ signal s(t) to obtain a signal with AWGN: x(t) = s(t) + n(t) Next, based on the superimposed AWGN, three types of interference—multipath fading, Doppler shift, and carrier frequency offset—are introduced to simulate the real channel environment: The mathematical expression for multipath fading is: Where N is the number of multipaths, α n (t) is the time-varying amplitude attenuation coefficient of the signal along different paths, τ n (t) represents the transmission delay along different paths, f c It is the carrier frequency. These are Doppler phase shifts along different paths; The expression for Doppler frequency shift is: in, f d θ is the Doppler frequency shift, v is the relative velocity between the transmitter and receiver, c is the speed of light, and θ is the angle between the direction of motion and the direction of signal incidence. Carrier frequency offset, expressed as: x cfo (t)=x(t)e j2πΔft Where Δf is the random frequency offset; Let x′(t) represent the signal after interference is added. Finally, perform energy normalization on x′(x) and calculate the total signal energy as follows: Where L is the length of the signal sequence, and dividing each sample in the signal sequence by the square root of its energy yields the energy-normalized signal sequence as follows: Where ε = 10 -10 This eliminates differences in signal amplitude and power, ensuring that all samples have consistent energy.

3. The method according to claim 1, characterized in that, In step 2, the dataset partitioning adopts a dual stratified sampling strategy, using modulation method and signal-to-noise ratio as the basis for stratification, to ensure that the proportion of data in the training set, test set, and validation set remains consistent and uniform under different modulation methods and signal-to-noise ratios, while interference scenarios are randomly assigned.

4. The method according to claim 1, characterized in that, In step 3, the STFT uses a Hamming window as the window function, and the window length is dynamically adapted to the signal length. The preset window length value is set to 256, and the window length is selected as the smaller value between the preset value and the length of a single signal sample. The overlap ratio of adjacent windows is set to 50%.

5. The method according to claim 1, characterized in that, In step 4, multi-dimensional data augmentation is performed on the time-frequency graph of the training data. The augmentation operations include random shifting in time, random shifting in frequency, random noise injection, and random rotation at small angles.

6. The method according to claim 1, characterized in that, In step 5, a ResNet-18 network adapted for single-channel input is constructed. This network consists of multiple convolutional layers cascaded together and multiple residual structures. Dropout layers are introduced into the fully connected layers to randomly mask the output of some neurons during training.

7. The method according to claim 1, characterized in that, In step 6, ResNet-18 is trained using the enhanced training dataset, and the model parameters are optimized using the cross-entropy loss function, L: L=-Σy i log(p i ) Among them, y i For real labels, p i To predict probabilities; The Adam optimizer is used to update the network parameters using gradients, while the learning rate is dynamically updated using ReduceLROnPlateau learning rate scheduling. The hyperparameters are adjusted using a validation set. Finally, the performance of the trained model is verified using a test set.

Images