Dual-mode communication signal identification method based on deep learning

Abstract

The invention discloses a dual-mode communication signal identification method based on deep learning, relates to the technical field of communication signal processing, and aims to solve the problems that a simple neural network structure is generally adopted in a traditional deep learning algorithm, the attention of a model to important features is insufficient, gradient disappearance is likely to occur and the like. The method comprises the following steps: S1, signal acquisition and preprocessing are carried out; s1.1, receiving a signal; receiving a dual-mode communication signal from the intelligent traffic communication system; in an intelligent traffic communication scene, communication signals come from vehicle networking communication among vehicles, communication between the vehicles and road infrastructures, communication between a traffic management center and traffic nodes and the like; s1.2, carrying out adaptive sampling; an advanced self-adaptive sampling technology is used for sampling; the technology dynamically adjusts the sampling frequency and the number of sampling points according to the characteristics of signals. The method has the advantage of improving the processing capability and the recognition precision of the model on complex features.

Description

Technical Field

[0001] The present invention relates to the technical field of communication signal processing, and more specifically, to a dual-mode communication signal recognition method based on deep learning. Background Art

[0002] With the continuous development of communication technology, various communication scenarios are becoming increasingly complex, and the need to identify dual-mode communication signals is becoming increasingly urgent. In many fields such as mobile communications, satellite communications, wireless communications, Internet of Things communications, underwater communications, military communications, intelligent transportation communications, industrial automation communications, and aerospace communications, accurate identification of dual-mode communication signals is of great significance for ensuring the stable operation of communication systems and improving communication efficiency and security.

[0003] At present, when using deep learning algorithms to identify dual-mode communication signals, in terms of building deep learning models, traditional methods usually use simple neural network structures, which lack the ability to process complex features and the introduction of special structures. For example, without structures such as attention mechanisms and residual networks, the model does not pay enough attention to important features, and is prone to problems such as gradient disappearance, which affects the performance and recognition accuracy of the model. In the process of model training and optimization, traditional methods often use a single loss function and fixed hyperparameters, and cannot be adaptively adjusted according to different signal data. This makes the model training effect poor and the generalization ability limited. In view of this, we propose a dual-mode communication signal recognition method based on deep learning. Summary of the invention

[0004] The purpose of the present invention is to provide a dual-mode communication signal recognition method based on deep learning, so as to solve the technical problems that traditional deep learning algorithms usually adopt a simple neural network structure, the model does not pay enough attention to important features, and is prone to gradient disappearance, which affects the performance of the model and the recognition accuracy.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: a dual-mode communication signal recognition method based on deep learning, comprising the following steps:

[0006] S1, signal acquisition and preprocessing;

[0007] S1.1, signal reception;

[0008] receiving a dual-mode communication signal from an intelligent transportation communication system;

[0009] In the intelligent transportation communication scenario, communication signals come from vehicle-to-vehicle communication, communication between vehicles and road infrastructure, and communication between the traffic management center and various traffic nodes.

[0010] S1.2, adaptive sampling;

[0011] Use advanced adaptive sampling technology for sampling;

[0012] This technology dynamically adjusts the sampling frequency and number of sampling points according to the characteristics of the signal;

[0013] Let the original signal be x(t), where t represents the time variable and the sampling frequency is f s , the number of sampling points is N, then the sampled signal can be expressed as:

[0014] x(n)=x(nT s )

[0015] in is the sampling period, n represents the index of the discrete time point;

[0016] S1.3, noise removal;

[0017] Adopt adaptive noise removal algorithm to remove noise;

[0018] Among them, the adaptive noise removal algorithm can dynamically adjust the noise removal parameters according to the characteristics of the signal;

[0019] For mean filtering, let the signal length be N and the filter window size be M. Then the filtered signal y(n) is

[0020]

[0021] in Here i is the index used to traverse the signal points in the filter window, and y(n) represents the signal value at time point n after mean filtering;

[0022] S1.4, signal enhancement;

[0023] Signal enhancement is performed by signal amplification, filtering, etc. Signal amplification can be performed by linear amplification or nonlinear amplification. If the amplification factor is k, the amplified signal is y(n)=kx(n);

[0024] Here k represents the amplification factor, which determines the degree of signal enhancement;

[0025] The filter adopts high-pass filter, and its transfer function is

[0026]

[0027] where f c is the cut-off frequency, f represents the frequency variable;

[0028] S1.5, normalization processing;

[0029] Perform normalization processing to make the signal data in a uniform numerical range to obtain high-quality signal data;

[0030] The normalization method selects maximum and minimum normalization. Let the original signal be x(n) and the maximum value be x max , the minimum value is x min , then the normalized signal is

[0031]

[0032] Here x max and x min They represent the maximum and minimum values ​​in the original signal, respectively, and y(n) is the signal value at time point n after normalization;

[0033] S2, time domain feature extraction;

[0034] S2.1 Basic time domain feature extraction;

[0035] Perform time domain analysis on the preprocessed signal data to extract the time domain characteristics of the signal's waveform, amplitude, period, and phase, reflecting the signal's changing characteristics in the time dimension;

[0036] Waveform features can include detailed features of peaks, troughs, rising edges, and falling edges. By analyzing these features, we can understand the morphological changes of the signal.

[0037] The amplitude feature can calculate the maximum, minimum, mean, and variance statistics of the signal. Suppose the signal is x(n) and its length is N, then the mean is

[0038]

[0039] The variance is

[0040]

[0041] Here μ represents the mean of the signal, reflecting the average level of the signal;

[0042] σ 2 It represents the variance, which reflects the discreteness of the signal;

[0043] The periodic characteristics are determined by the autocorrelation function method. The autocorrelation function is

[0044]

[0045] When the autocorrelation function reaches a peak, the corresponding time delay τ is the period of the signal;

[0046] Here τ represents the time delay, R(τ) represents the autocorrelation function value, which reflects the similarity of the signal under different time delays; the phase feature is extracted by the Hilbert transform method. Let the signal be x(t), and its Hilbert transform is

[0047]

[0048] Then the instantaneous phase of the signal is

[0049]

[0050] Here y(t) represents the signal after Hilbert transform, Indicates the instantaneous phase, reflecting the phase information of the signal at different time points;

[0051] S2.2, extended time domain analysis;

[0052] Short-time Fourier transform (STFT) analysis method can also be used to extract richer time domain features. Short-time Fourier transform can divide the signal into short time periods and perform Fourier transform on each time period. Suppose the signal is x(t) and the window function is g(t), then the short-time Fourier transform is

[0053]

[0054] Here t represents the time variable, f represents the frequency variable, g(t) represents the window function, and STFT(t, f) represents the result after short-time Fourier transform, which reflects the local characteristics of the signal at different times and frequencies;

[0055] The autocorrelation function can be used to detect the periodicity and similarity of signals, and the cross-correlation function can be used to compare the similarity and time delay between two signals. Suppose the two signals are x(t) and y(t), then the cross-correlation function is

[0056]

[0057] Here τ represents the time delay, R xy (τ) represents the cross-correlation function value, which reflects the similarity of two signals at different time delays;

[0058] S2.3, deep learning time domain feature extraction;

[0059] A time domain feature extraction method based on deep learning is introduced, and the autoencoder is used to reconstruct the signal and extract more representative time domain features;

[0060] The autoencoder automatically learns the potential feature representation of the signal through unsupervised learning;

[0061] Assume that the input signal is x, the encoder function of the autoencoder is f(x), and the decoder function is g(f(x)). The goal of the autoencoder is to minimize the reconstruction error L(x, g(f(x))), using the mean square error loss function

[0062]

[0063] Here x represents the input signal, f(x) represents the encoder function, g(f(x)) represents the decoder function, and L(x, g(f(x))) represents the reconstruction error, which reflects the autoencoder's ability to reconstruct the input signal;

[0064] S3, frequency domain feature extraction;

[0065] S3.1, basic frequency domain transform;

[0066] Use Fourier transform to convert time domain signal into frequency domain signal;

[0067] Fourier transform is the most commonly used frequency domain transformation method, which can convert the signal from the time domain to the frequency domain. Let the signal be x(t), and its Fourier transform is

[0068]

[0069] in

[0070]

[0071] Here n represents the index of the discrete time point, k represents the index of the discrete frequency point, and X(k) represents the result after discrete cosine transform, which reflects the energy distribution of the signal at different frequencies;

[0072] Wavelet transform can perform multi-resolution analysis on signals and is suitable for processing non-stationary signals. Suppose the signal is x(t) and the wavelet function is ψ(t), then the wavelet transform is:

[0073]

[0074] Here t represents the time variable, a is the scale parameter, which determines the width of the wavelet function, b is the translation parameter, which determines the position of the wavelet function, and WT(a, b) represents the result after wavelet transformation, which reflects the local characteristics of the signal at different scales and positions;

[0075] S3.2, multi-level frequency domain analysis;

[0076] First, use Fourier transform to get the basic frequency distribution of the signal, and then use wavelet transform to perform more detailed analysis on specific frequency intervals;

[0077] Different wavelet basis functions can be selected according to the characteristics of the signal, such as Haar wavelet and Daubechies wavelet, to adapt to different signal characteristics;

[0078] S3.3, high-order frequency domain statistics;

[0079] Introducing high-order statistics of frequency domain features, such as calculating the skewness and kurtosis statistics of frequency domain signals, which can provide more information about the frequency distribution of the signal;

[0080] Assume the frequency domain signal is X(f), the mean is μx, and the standard deviation is σx, then the skewness is

[0081]

[0082] The kurtosis is

[0083]

[0084] Here μx represents the mean of the frequency domain signal, σx represents the standard deviation, S represents the skewness, which reflects the distribution symmetry of the frequency domain signal, and K represents the kurtosis, which reflects the peak degree of the distribution of the frequency domain signal;

[0085] S4, build a deep learning model;

[0086] S4.1, Model structure selection;

[0087] Design a deep neural network model that can simultaneously receive time domain features and frequency domain features as input and output the classification results of communication signals;

[0088] The model structure can include convolutional neural network (CNN), recurrent neural network (RNN), or a combination thereof;

[0089] CNN is good at processing local features of images and signals, while RNN is good at processing the temporal dependencies of sequence data;

[0090] The appropriate model structure can be selected according to the characteristics of the signal and the task requirements;

[0091] For example, for signals with spatially local features, CNN can be selected;

[0092] For signals with time series characteristics, RNN or CNN-RNN combination can be selected;

[0093] Let the input signal be x, and the convolutional layer output of CNN be

[0094] y conv =W*X+b

[0095] Here W is the convolution kernel, which determines the weight of the convolution operation, X represents the input feature map, b is the bias term, and y conv Represents the output result of the convolutional layer;

[0096] The hidden state update formula of RNN is

[0097] h t =tanh(W xh x t +W hh h t-1 +b h )

[0098] Where W xh is the weight matrix input to the hidden layer, W hh is the weight matrix from hidden layer to hidden layer, b h is the bias term;

[0099] Here x t represents the input signal at time point t, h t represents the hidden state at time point t;

[0100] S4.2, introduce special structures;

[0101] Introducing the attention mechanism and residual network structure to improve the model's ability to process complex features and recognition accuracy;

[0102] The attention mechanism allows the model to focus on the important parts of the signal and ignore the unimportant parts;

[0103] Residual networks can solve the gradient vanishing problem in deep neural networks and improve the training effect of the model;

[0104] Different types of attention mechanisms can be used, such as spatial attention, channel attention, etc., and selected according to the characteristics of the signal;

[0105] The residual network can adopt different residual block structures, such as basic residual block, bottleneck residual block, etc., to adapt to different task requirements;

[0106] Let the input feature be X, and the output of the attention mechanism be

[0107] Y=Att(X)

[0108] Where Att is the attention function;

[0109] The output of the residual network is

[0110] Y=X+F(X)

[0111] Where F is the residual function;

[0112] Here X represents the input feature, Y represents the output result of the attention mechanism or residual network, Att(X) represents the processing result of the attention function on the input feature, and F(X) represents the processing result of the residual function on the input feature;

[0113] S4.3, multimodal fusion;

[0114] Adopt multimodal fusion technology to better integrate time domain features and frequency domain features;

[0115] Use feature splicing and attention weighting methods;

[0116] Feature splicing is to directly splice the time domain features and frequency domain features together as the input of the model;

[0117] Attention weighting is to perform weighted fusion of time domain features and frequency domain features by learning an attention weight matrix, highlighting important features and suppressing unimportant features;

[0118] Let the time domain feature be T, the frequency domain feature be F, and the feature after feature concatenation be Y = [T; F];

[0119] The feature after attention weighting is

[0120] Y=αT+(1-α)F

[0121] Where α is the attention weight;

[0122] Here T represents the time domain features, F represents the frequency domain features, Y represents the fused features, [T; F] represents the splicing of the time domain features and the frequency domain features together, and α represents the attention weight, which determines the relative importance of the time domain features and the frequency domain features.

[0123] S5, model training and optimization;

[0124] S5.1. Data preparation;

[0125] Using a large amount of labeled bimodal communication signal data, including its time domain and frequency domain features;

[0126] Divide the data into training set, validation set and test set;

[0127] The training set is used to train the model, the validation set is used to adjust the model's hyperparameters, and the test set is used to evaluate the model's performance;

[0128] S5.2, parameter adjustment;

[0129] The deep learning model is trained by adjusting model parameters, selecting appropriate optimizers, and designing reasonable loss functions to improve the recognition accuracy and generalization ability of the model;

[0130] Model parameters include convolution kernel size, number of layers, and number of hidden layer nodes;

[0131] The optimizer selects an optimization algorithm with adaptive learning rate, such as AdamW optimizer, which can automatically adjust the learning rate according to the gradient of each parameter to speed up the convergence of the model;

[0132] The loss function can be a combination of cross entropy loss and mean square error loss. Let the true label be y and the predicted label be The loss function is

[0133] L=αL ce +βL mse

[0134] in is the cross entropy loss; is the mean square error loss;

[0135] Here y represents the true label, represents the predicted label, L represents the loss function, which reflects the difference between the model prediction result and the true label, L ce represents the cross entropy loss, which measures the difference in probability distribution in the classification task, L mse represents the mean square error loss, which measures the numerical difference in the regression task. α and β represent weight coefficients, which are used to adjust the relative importance of the cross entropy loss and the mean square error loss in the total loss function.

[0136] S5.3. Hyperparameter adjustment;

[0137] Compare the performance of different hyperparameter combinations through experiments and select the optimal hyperparameter settings;

[0138] Hyperparameters include learning rate, batch size, and number of training rounds;

[0139] Grid search and random search methods can be used to adjust hyperparameters;

[0140] S6, signal recognition and classification;

[0141] S6.1, signal preprocessing and feature extraction;

[0142] After preprocessing, time domain feature extraction and frequency domain feature extraction of the dual-mode communication signal to be identified, a high-quality feature representation of the signal is obtained;

[0143] S6.2, model input and output;

[0144] Input into the trained deep learning model;

[0145] The model outputs the classification result of the communication signal according to the input feature vector, realizing the accurate recognition of the dual-mode communication signal;

[0146] S6.3, Result evaluation and feedback;

[0147] To evaluate the recognition results, the accuracy, recall rate, and F1 value indicators can be used for evaluation;

[0148] Assume that the true positive example is TP, the true negative example is TN, the false positive example is FP, and the false negative example is FN, then the accuracy is

[0149]

[0150] The recall rate is

[0151]

[0152] The F1 value is

[0153]

[0154] in

[0155]

[0156] Here TP represents true positive examples, that is, the number of samples correctly classified as positive;

[0157] TN represents true negative examples, that is, the number of samples correctly classified as negative;

[0158] FP stands for false positive, that is, the number of samples that are incorrectly classified as positive;

[0159] FN represents false negative examples, that is, the number of samples that are misclassified as negative;

[0160] The accuracy reflects the overall accuracy of the classification results, the recall reflects the ability to identify positive samples, and the F1 value comprehensively considers the accuracy and recall, and is a balanced evaluation indicator.

[0161] Preferably, in the signal acquisition and preprocessing steps, a signal quality assessment module is added;

[0162] This module evaluates the quality of the collected signal in real time by calculating the signal-to-noise ratio, energy distribution and other indicators;

[0163] If the signal quality is below a preset threshold, the system can automatically adjust sampling parameters or take additional signal enhancement measures to ensure the accuracy of subsequent processing;

[0164] Assume the signal is x(t) and the noise is n(t), then the signal-to-noise ratio is

[0165]

[0166] Where T is the signal length, which represents the number of discrete points in time;

[0167] x(t) represents the signal value at time t, n(t) represents the noise value at time t, SNR represents the signal-to-noise ratio, which reflects the relative strength of the signal and the noise;

[0168] The energy distribution is obtained by calculating the energy spectral density of the signal. Assuming the Fourier transform of the signal is X(f), the energy spectral density is

[0169] S(f)=|X(f)| 2

[0170] Here f represents the frequency variable, X(f) represents the Fourier transform result of the signal, and S(f) represents the energy spectrum density, which reflects the energy distribution of the signal at different frequencies.

[0171] Preferably, in the time domain feature extraction step, the time domain dynamic characteristics of the signal are further explored;

[0172] For example, the instantaneous frequency and instantaneous amplitude of the signal are calculated to change with time. These characteristics can better reflect the rapid changes and non-stationary characteristics of the signal in the time dimension.

[0173] Let the signal be x(t), and its Hilbert transform is

[0174]

[0175] The instantaneous amplitude of the signal is

[0176]

[0177] The instantaneous frequency is

[0178]

[0179] Here t represents the time variable, x(t) represents the original signal, y(t) represents the signal after Hilbert transform, A(t) represents the instantaneous amplitude, which reflects the amplitude of the signal at time t, and ω(t) represents the instantaneous frequency, which reflects the frequency change of the signal at time t.

[0180] The present invention designs a deep neural network model that can simultaneously receive time domain features and frequency domain features as input, fully utilizing the information of the signal in different dimensions. The model structure is flexible and diverse, including convolutional neural networks (CNNs), recurrent neural networks (RNNs) or their combinations, and can introduce structures such as attention mechanisms and residual networks to improve the model's processing capabilities and recognition accuracy for complex features. Multimodal fusion technology is used to better integrate time domain features and frequency domain features to improve model performance.

[0181] Preferably, in the frequency domain feature extraction step, high-order statistics of the frequency domain features are introduced;

[0182] For example, by calculating the skewness and kurtosis statistics of frequency domain signals, these statistics can provide more information about the frequency distribution of the signal and help improve the model's ability to distinguish different types of signals.

[0183] Preferably, in the step of building a deep learning model, different network architectures and hyperparameter combinations are tried;

[0184] For example, adjusting the convolution kernel size, number of layers, activation function and other parameters of the convolutional neural network, or changing the unit type and hidden layer size of the recurrent neural network;

[0185] By comparing the performance of different combinations through experiments, the optimal model architecture and hyperparameter settings are selected.

[0186] Preferably, the method further includes a step of continuously optimizing the deep learning model;

[0187] When new communication signal types emerge or the communication environment changes, the system can automatically collect new data for incremental learning and gradually adjust the model to improve its adaptability and accuracy;

[0188] Let the new data be D new , the old model parameters are θ old , then the new model parameters can be obtained by updating the old model parameters, that is,

[0189]

[0190] It is the gradient of the loss function on the new data to the model parameters, which is used to guide the update direction of the parameters;

[0191] Here D new represents the newly collected data, θ old represents the old model parameters, θ new Represents the updated model parameters, α represents the learning rate, which controls the step size of parameter update.

[0192] Preferably, the method is applicable to a wider range of communication scenarios, including but not limited to intelligent transportation communication, industrial automation communication, aerospace communication, etc., and has a wider application prospect and practical value. At the same time, combined with actual application needs, corresponding visualization tools and interfaces are developed to facilitate users to view and analyze signal recognition results.

[0193] Preferably, during the model training process, a data enhancement method can be used to increase the diversity of training data and improve the generalization ability of the model;

[0194] Data enhancement methods can include random cropping, random flipping, random rotation, adding noise, etc. For example, for image signals, random cropping and random flipping can be performed to increase the diversity of the image; for audio signals, different types of noise can be added to simulate different environmental noise conditions.

[0195] Preferably, in order to improve the real-time performance of the method, a parallel computing method can be used to accelerate processes such as signal acquisition, preprocessing, feature extraction, and model training;

[0196] Parallel computing uses multi-core processors, graphics processing units (GPUs), and distributed computing technology;

[0197] Use GPU to accelerate the training of deep learning models and increase the training speed; use distributed computing framework to distribute signal acquisition and preprocessing tasks to multiple nodes for parallel processing to improve processing efficiency.

[0198] Compared with the prior art, the present invention has the following beneficial effects:

[0199] 1. The present invention designs a deep neural network model that can simultaneously receive time domain features and frequency domain features as inputs, fully utilizes the information of the signal in different dimensions, and has a flexible and diverse model structure, including convolutional neural networks (CNNs), recurrent neural networks (RNNs) or a combination thereof, and can introduce structures such as attention mechanisms and residual networks to improve the model's processing capabilities and recognition accuracy for complex features. It uses multimodal fusion technology to better integrate time domain features and frequency domain features, improve model performance, and solve the problem that traditional deep learning algorithms usually use simple neural network structures, the model does not pay enough attention to important features, and is prone to problems such as gradient disappearance, which affects the performance and recognition accuracy of the model.

[0200] 2. The present invention also improves the recognition accuracy and generalization ability of the model by using a large amount of labeled dual-mode communication signal data for training, adjusting model parameters, selecting a suitable optimizer and designing a reasonable loss function, and adopts hyperparameter adjustment methods, such as grid search and random search, to select the optimal hyperparameter settings.

[0201] 3. The present invention also combines a variety of frequency domain transformation methods, such as Fourier transform, discrete cosine transform (DCT), and wavelet transform, to perform multi-level frequency domain analysis on the signal to obtain more comprehensive frequency domain feature information. At the same time, it introduces high-order statistics of frequency domain features, such as skewness and kurtosis, to provide more information about the signal frequency distribution and improve the model's ability to distinguish different types of signals. DETAILED DESCRIPTION

[0202] In order to facilitate those skilled in the art to understand the technical solution of the present invention, the technical solution of the present invention is further explained in conjunction with the specification.

[0203] A dual-mode communication signal recognition method based on deep learning, comprising the following steps:

[0204] S1, signal acquisition and preprocessing;

[0205] S1.1, signal reception;

[0206] receiving a dual-mode communication signal from an intelligent transportation communication system;

[0207] In the intelligent transportation communication scenario, communication signals come from vehicle-to-vehicle communication, communication between vehicles and road infrastructure, and communication between the traffic management center and various traffic nodes.

[0208] S1.2, adaptive sampling;

[0209] Use advanced adaptive sampling technology for sampling;

[0210] This technology dynamically adjusts the sampling frequency and number of sampling points according to the characteristics of the signal;

[0211] Let the original signal be x(t), where t represents the time variable and the sampling frequency is f s , the number of sampling points is N, then the sampled signal can be expressed as:

[0212] x(n)=x(nT s )

[0213] in is the sampling period, n represents the index of the discrete time point;

[0214] S1.3, noise removal;

[0215] Adopt adaptive noise removal algorithm to remove noise;

[0216] Among them, the adaptive noise removal algorithm can dynamically adjust the noise removal parameters according to the characteristics of the signal;

[0217] For mean filtering, let the signal length be N and the filter window size be M. Then the filtered signal y(n) is

[0218]

[0219] in Here i is the index used to traverse the signal points in the filter window, and y(n) represents the signal value at time point n after mean filtering;

[0220] S1.4, signal enhancement;

[0221] Signal enhancement is performed by signal amplification, filtering, etc. Signal amplification can be performed by linear amplification or nonlinear amplification. If the amplification factor is k, the amplified signal is y(n)=kx(n);

[0222] Here k represents the amplification factor, which determines the degree of signal enhancement;

[0223] The filter adopts high-pass filter, and its transfer function is

[0224]

[0225] where f c is the cut-off frequency, f represents the frequency variable;

[0226] S1.5, normalization processing;

[0227] Perform normalization processing to make the signal data in a uniform numerical range to obtain high-quality signal data;

[0228] The normalization method selects maximum and minimum normalization. Let the original signal be x(n) and the maximum value be x max , the minimum value is x min , then the normalized signal is

[0229]

[0230] Here x max and x min They represent the maximum and minimum values ​​in the original signal, respectively, and y(n) is the signal value at time point n after normalization;

[0231] S2, time domain feature extraction;

[0232] S2.1 Basic time domain feature extraction;

[0233] Perform time domain analysis on the preprocessed signal data to extract the time domain characteristics of the signal's waveform, amplitude, period, and phase, reflecting the signal's changing characteristics in the time dimension;

[0234] Waveform features can include detailed features of peaks, troughs, rising edges, and falling edges. By analyzing these features, we can understand the morphological changes of the signal.

[0235] The amplitude feature can calculate the maximum, minimum, mean, and variance statistics of the signal. Suppose the signal is x(n) and its length is N, then the mean is

[0236]

[0237] The variance is

[0238]

[0239] Here μ represents the mean of the signal, reflecting the average level of the signal;

[0240] σ 2 It represents the variance, which reflects the discreteness of the signal;

[0241] The periodic characteristics are determined by the autocorrelation function method. The autocorrelation function is

[0242]

[0243] When the autocorrelation function reaches a peak, the corresponding time delay τ is the period of the signal;

[0244] Here τ represents the time delay, R(τ) represents the autocorrelation function value, which reflects the similarity of the signal under different time delays; the phase feature is extracted by the Hilbert transform method. Let the signal be x(t), and its Hilbert transform is

[0245]

[0246] Then the instantaneous phase of the signal is

[0247]

[0248] Here y(t) represents the signal after Hilbert transform, Indicates the instantaneous phase, reflecting the phase information of the signal at different time points;

[0249] S2.2, extended time domain analysis;

[0250] Short-time Fourier transform (STFT) analysis method can also be used to extract richer time domain features. Short-time Fourier transform can divide the signal into short time periods and perform Fourier transform on each time period. Suppose the signal is x(t) and the window function is g(t), then the short-time Fourier transform is

[0251]

[0252] Here t represents the time variable, f represents the frequency variable, g(t) represents the window function, and STFT(t, f) represents the result after short-time Fourier transform, which reflects the local characteristics of the signal at different times and frequencies;

[0253] The autocorrelation function can be used to detect the periodicity and similarity of signals, and the cross-correlation function can be used to compare the similarity and time delay between two signals. Suppose the two signals are x(t) and y(t), then the cross-correlation function is

[0254]

[0255] Here τ represents the time delay, Rxy (τ) represents the cross-correlation function value, which reflects the similarity of two signals at different time delays;

[0256] S2.3, deep learning time domain feature extraction;

[0257] A time domain feature extraction method based on deep learning is introduced, and the autoencoder is used to reconstruct the signal and extract more representative time domain features;

[0258] The autoencoder automatically learns the potential feature representation of the signal through unsupervised learning;

[0259] Assume that the input signal is x, the encoder function of the autoencoder is f(x), and the decoder function is g(f(x)). The goal of the autoencoder is to minimize the reconstruction error L(x, g(f(x))), using the mean square error loss function

[0260]

[0261] Here x represents the input signal, f(x) represents the encoder function, g(f(x)) represents the decoder function, and L(x, g(f(x))) represents the reconstruction error, which reflects the autoencoder's ability to reconstruct the input signal;

[0262] S3, frequency domain feature extraction;

[0263] S3.1, basic frequency domain transform;

[0264] Use Fourier transform to convert time domain signal into frequency domain signal;

[0265] Fourier transform is the most commonly used frequency domain transformation method, which can convert the signal from the time domain to the frequency domain. Let the signal be x(t), and its Fourier transform is

[0266]

[0267] in

[0268]

[0269] Here n represents the index of the discrete time point, k represents the index of the discrete frequency point, and X(k) represents the result after discrete cosine transform, which reflects the energy distribution of the signal at different frequencies;

[0270] Wavelet transform can perform multi-resolution analysis on signals and is suitable for processing non-stationary signals. Suppose the signal is x(t) and the wavelet function is ψ(t), then the wavelet transform is:

[0271]

[0272] Here t represents the time variable, a is the scale parameter, which determines the width of the wavelet function, b is the translation parameter, which determines the position of the wavelet function, and WT(a, b) represents the result after wavelet transformation, which reflects the local characteristics of the signal at different scales and positions;

[0273] S3.2, multi-level frequency domain analysis;

[0274] First, use Fourier transform to get the basic frequency distribution of the signal, and then use wavelet transform to perform more detailed analysis on specific frequency intervals;

[0275] Different wavelet basis functions can be selected according to the characteristics of the signal, such as Haar wavelet and Daubechies wavelet, to adapt to different signal characteristics;

[0276] S3.3, high-order frequency domain statistics;

[0277] Introducing high-order statistics of frequency domain features, such as calculating the skewness and kurtosis statistics of frequency domain signals, which can provide more information about the frequency distribution of the signal;

[0278] Assume the frequency domain signal is X(f), the mean is μx, and the standard deviation is σx, then the skewness is

[0279]

[0280] The kurtosis is

[0281]

[0282] Here μx represents the mean of the frequency domain signal, σx represents the standard deviation, S represents the skewness, which reflects the distribution symmetry of the frequency domain signal, and K represents the kurtosis, which reflects the peak degree of the distribution of the frequency domain signal;

[0283] S4. Build a deep learning model;

[0284] S4.1, Model structure selection;

[0285] Design a deep neural network model that can simultaneously receive time domain features and frequency domain features as input and output the classification results of communication signals;

[0286] The model structure can include convolutional neural network (CNN), recurrent neural network (RNN), or a combination thereof;

[0287] CNN is good at processing local features of images and signals, while RNN is good at processing the temporal dependencies of sequence data;

[0288] The appropriate model structure can be selected according to the characteristics of the signal and the task requirements;

[0289] For example, for signals with spatially local features, CNN can be selected;

[0290] For signals with time series characteristics, RNN or CNN-RNN combination can be selected;

[0291] Let the input signal be x, and the convolutional layer output of CNN be

[0292] y conv =W*X+b

[0293] Here W is the convolution kernel, which determines the weight of the convolution operation, X represents the input feature map, b is the bias term, and y conv Represents the output result of the convolutional layer;

[0294] The hidden state update formula of RNN is

[0295] h t =tanh(W xh x t +W hh h t-1 +b h )

[0296] Where W xh is the weight matrix input to the hidden layer, W hh is the weight matrix from hidden layer to hidden layer, b h is the bias term;

[0297] Here x t represents the input signal at time point t, h t represents the hidden state at time point t;

[0298] S4.2, introduce special structures;

[0299] Introducing the attention mechanism and residual network structure to improve the model's ability to process complex features and recognition accuracy;

[0300] The attention mechanism allows the model to focus on the important parts of the signal and ignore the unimportant parts;

[0301] Residual networks can solve the gradient vanishing problem in deep neural networks and improve the training effect of the model;

[0302] Different types of attention mechanisms can be used, such as spatial attention, channel attention, etc., and selected according to the characteristics of the signal;

[0303] The residual network can adopt different residual block structures, such as basic residual block, bottleneck residual block, etc., to adapt to different task requirements;

[0304] Let the input feature be X, and the output of the attention mechanism be

[0305] Y=Att(X)

[0306] Where Att is the attention function;

[0307] The output of the residual network is

[0308] Y=X+F(X)

[0309] Where F is the residual function;

[0310] Here X represents the input feature, Y represents the output result of the attention mechanism or residual network, Att(X) represents the processing result of the attention function on the input feature, and F(X) represents the processing result of the residual function on the input feature;

[0311] S4.3, multimodal fusion;

[0312] Adopt multimodal fusion technology to better integrate time domain features and frequency domain features;

[0313] Use feature splicing and attention weighting methods;

[0314] Feature splicing is to directly splice the time domain features and frequency domain features together as the input of the model;

[0315] Attention weighting is to perform weighted fusion of time domain features and frequency domain features by learning an attention weight matrix, highlighting important features and suppressing unimportant features;

[0316] Let the time domain feature be T, the frequency domain feature be F, and the feature after feature concatenation be Y = [T; F];

[0317] The feature after attention weighting is

[0318] Y=αT+(1-α)F

[0319] Where α is the attention weight;

[0320] Here T represents the time domain features, F represents the frequency domain features, Y represents the fused features, [T; F] represents the splicing of the time domain features and the frequency domain features together, and α represents the attention weight, which determines the relative importance of the time domain features and the frequency domain features.

[0321] S5, model training and optimization;

[0322] S5.1. Data preparation;

[0323] Using a large amount of labeled bimodal communication signal data, including its time domain and frequency domain features;

[0324] Divide the data into training set, validation set and test set;

[0325] The training set is used to train the model, the validation set is used to adjust the model's hyperparameters, and the test set is used to evaluate the model's performance;

[0326] S5.2, parameter adjustment;

[0327] The deep learning model is trained by adjusting model parameters, selecting appropriate optimizers, and designing reasonable loss functions to improve the recognition accuracy and generalization ability of the model;

[0328] Model parameters include convolution kernel size, number of layers, and number of hidden layer nodes;

[0329] The optimizer selects an optimization algorithm with adaptive learning rate, such as AdamW optimizer, which can automatically adjust the learning rate according to the gradient of each parameter to speed up the convergence of the model;

[0330] The loss function can be a combination of cross entropy loss and mean square error loss. Let the true label be y and the predicted label be The loss function is

[0331] L=αL ce +βL mse

[0332] in is the cross entropy loss; is the mean square error loss;

[0333] Here y represents the true label, represents the predicted label, L represents the loss function, which reflects the difference between the model prediction result and the true label, L ce represents the cross entropy loss, which measures the difference in probability distribution in the classification task, L mse represents the mean square error loss, which measures the numerical difference in the regression task. α and β represent weight coefficients, which are used to adjust the relative importance of the cross entropy loss and the mean square error loss in the total loss function.

[0334] S5.3. Hyperparameter adjustment;

[0335] Compare the performance of different hyperparameter combinations through experiments and select the optimal hyperparameter settings;

[0336] Hyperparameters include learning rate, batch size, and number of training rounds;

[0337] Grid search and random search methods can be used to adjust hyperparameters;

[0338] S6, signal recognition and classification;

[0339] S6.1, signal preprocessing and feature extraction;

[0340] After preprocessing, time domain feature extraction and frequency domain feature extraction of the dual-mode communication signal to be identified, a high-quality feature representation of the signal is obtained;

[0341] S6.2, Model input and output;

[0342] Input into the trained deep learning model;

[0343] The model outputs the classification result of the communication signal according to the input feature vector, realizing the accurate recognition of the dual-mode communication signal;

[0344] S6.3, Result evaluation and feedback;

[0345] To evaluate the recognition results, the accuracy, recall rate, and F1 value indicators can be used for evaluation;

[0346] Assume that the true positive example is TP, the true negative example is TN, the false positive example is FP, and the false negative example is FN, then the accuracy is

[0347]

[0348] The recall rate is

[0349]

[0350] The F1 value is

[0351]

[0352] in

[0353]

[0354] Here TP represents true positive examples, that is, the number of samples correctly classified as positive;

[0355] TN represents true negative examples, that is, the number of samples correctly classified as negative;

[0356] FP stands for false positive, that is, the number of samples that are incorrectly classified as positive;

[0357] FN represents false negative examples, that is, the number of samples that are misclassified as negative;

[0358] The accuracy reflects the overall accuracy of the classification results, the recall reflects the ability to identify positive samples, and the F1 value comprehensively considers the accuracy and recall, and is a balanced evaluation indicator.

[0359] In an embodiment of the present invention, a signal quality assessment module is added in the signal acquisition and preprocessing steps;

[0360] This module evaluates the quality of the collected signal in real time by calculating the signal-to-noise ratio, energy distribution and other indicators;

[0361] If the signal quality is below a preset threshold, the system can automatically adjust sampling parameters or take additional signal enhancement measures to ensure the accuracy of subsequent processing;

[0362] Let the signal be x(t) and the noise be n(t), then the signal-to-noise ratio is

[0363]

[0364] Where T is the signal length, which represents the number of discrete points in time;

[0365] x(t) represents the signal value at time t, n(t) represents the noise value at time t, SNR represents the signal-to-noise ratio, which reflects the relative strength of the signal and the noise;

[0366] The energy distribution is obtained by calculating the energy spectral density of the signal. Assuming the Fourier transform of the signal is X(f), the energy spectral density is

[0367] S(f)=|X(f)| 2

[0368] Here f represents the frequency variable, X(f) represents the Fourier transform result of the signal, and S(f) represents the energy spectrum density, which reflects the energy distribution of the signal at different frequencies.

[0369] In an embodiment of the present invention, in the time domain feature extraction step, the time domain dynamic characteristics of the signal are further explored;

[0370] For example, the instantaneous frequency and instantaneous amplitude of the signal are calculated to change with time. These characteristics can better reflect the rapid changes and non-stationary characteristics of the signal in the time dimension.

[0371] Let the signal be x(t), and its Hilbert transform is

[0372]

[0373] The instantaneous amplitude of the signal is

[0374]

[0375] The instantaneous frequency is

[0376]

[0377] Here t represents the time variable, x(t) represents the original signal, y(t) represents the signal after Hilbert transform, A(t) represents the instantaneous amplitude, which reflects the amplitude of the signal at time t, and ω(t) represents the instantaneous frequency, which reflects the frequency change of the signal at time t.

[0378] In an embodiment of the present invention, in the frequency domain feature extraction step, high-order statistics of the frequency domain features are introduced;

[0379] For example, the skewness and kurtosis statistics of the frequency domain signal are calculated. These statistics can provide more information about the frequency distribution of the signal and help improve the model's ability to distinguish different types of signals.

[0380] In the step of building a deep learning model, try different network architectures and hyperparameter combinations;

[0381] For example, adjusting the convolution kernel size, number of layers, activation function and other parameters of the convolutional neural network, or changing the unit type and hidden layer size of the recurrent neural network;

[0382] Compare the performance of different combinations through experiments and select the optimal model architecture and hyperparameter settings;

[0383] It also includes steps for continuous optimization of deep learning models;

[0384] When new communication signal types emerge or the communication environment changes, the system can automatically collect new data for incremental learning and gradually adjust the model to improve its adaptability and accuracy;

[0385] Let the new data be D new , the old model parameters are θ old , then the new model parameters can be obtained by updating the old model parameters, that is,

[0386]

[0387] It is the gradient of the loss function on the new data to the model parameters, which is used to guide the update direction of the parameters;

[0388] Here D new represents the newly collected data, θ old represents the old model parameters, θ new Represents the updated model parameters, α represents the learning rate, which controls the step size of parameter update.

[0389] In the embodiments of the present invention, the method is applicable to a wider range of communication scenarios, including but not limited to intelligent transportation communication, industrial automation communication, aerospace communication, etc., and has a wider application prospect and practical value. At the same time, in combination with actual application needs, corresponding visualization tools and interfaces are developed to facilitate users to view and analyze signal recognition results;

[0390] During the model training process, data enhancement methods can be used to increase the diversity of training data and improve the generalization ability of the model;

[0391] Data enhancement methods can include random cropping, random flipping, random rotation, adding noise, etc. For example, for image signals, random cropping and random flipping can be performed to increase the diversity of the image; for audio signals, different types of noise can be added to simulate different environmental noise conditions;

[0392] In order to improve the real-time performance of the method, parallel computing methods can be used to accelerate the processes of signal acquisition, preprocessing, feature extraction, and model training;

[0393] Parallel computing uses multi-core processors, graphics processing units (GPUs), and distributed computing technology;

[0394] Use GPU to accelerate the training of deep learning models and increase the training speed; use distributed computing framework to distribute signal acquisition and preprocessing tasks to multiple nodes for parallel processing to improve processing efficiency.

[0395] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A dual-mode communication signal recognition method based on deep learning, characterized in that: The following steps are involved: S1, signal acquisition and preprocessing; S1.1, signal reception; receiving a dual-mode communication signal from an intelligent transportation communication system; In the intelligent transportation communication scenario, communication signals come from vehicle-to-vehicle communication, communication between vehicles and road infrastructure, and communication between the traffic management center and various traffic nodes. S1.2, adaptive sampling; Use advanced adaptive sampling technology for sampling; This technology dynamically adjusts the sampling frequency and number of sampling points according to the characteristics of the signal; Let the original signal be x(t), where t represents the time variable and the sampling frequency is f s , the number of sampling points is N, then the sampled signal can be expressed as: x(n)=x(nT s ) in is the sampling period, n represents the index of the discrete time point; S1.3, noise removal; Adopt adaptive noise removal algorithm to remove noise; Among them, the adaptive noise removal algorithm can dynamically adjust the noise removal parameters according to the characteristics of the signal; For mean filtering, let the signal length be N and the filter window size be M. Then the filtered signal y(n) is in Here i is the index used to traverse the signal points in the filter window, and y(n) represents the signal value at time point n after mean filtering; S1.4, signal enhancement; Signal enhancement is performed by signal amplification, filtering, etc. Signal amplification can be performed by linear amplification or nonlinear amplification. If the amplification factor is k, the amplified signal is y(n)=kx(n); Here k represents the amplification factor, which determines the degree of signal enhancement; The filter adopts high-pass filter, and its transfer function is where f c is the cut-off frequency, f represents the frequency variable; S1.5, normalization processing; Perform normalization processing to make the signal data in a uniform numerical range to obtain high-quality signal data; The normalization method selects maximum and minimum normalization. Let the original signal be x(n) and the maximum value be x max , the minimum value is x min , then the normalized signal is Here x max and x min They represent the maximum and minimum values ​​in the original signal, respectively, and y(n) is the signal value at time point n after normalization; S2, time domain feature extraction; S2.1 Basic time domain feature extraction; Perform time domain analysis on the preprocessed signal data to extract the time domain characteristics of the signal's waveform, amplitude, period, and phase, reflecting the signal's changing characteristics in the time dimension; Waveform features can include detailed features of peaks, troughs, rising edges, and falling edges. By analyzing these features, we can understand the morphological changes of the signal. The amplitude feature can calculate the maximum, minimum, mean, and variance statistics of the signal. Suppose the signal is x(n) and its length is N, then the mean is The variance is Here μ represents the mean of the signal, reflecting the average level of the signal; σ 2 It represents the variance, which reflects the discreteness of the signal; The periodic characteristics are determined by the autocorrelation function method. The autocorrelation function is When the autocorrelation function reaches a peak, the corresponding time delay τ is the period of the signal; Here τ represents the time delay, and R(τ) represents the autocorrelation function value, which reflects the similarity of the signals under different time delays; The phase feature is extracted by Hilbert transform method. Let the signal be x(t), and its Hilbert transform is Then the instantaneous phase of the signal is Here y(t) represents the signal after Hilbert transform, Indicates the instantaneous phase, reflecting the phase information of the signal at different time points; S2.2, extended time domain analysis; Short-time Fourier transform (STFT) analysis method can also be used to extract richer time domain features. Short-time Fourier transform can divide the signal into short time periods and perform Fourier transform on each time period. Suppose the signal is x(t) and the window function is g(t), then the short-time Fourier transform is Here t represents the time variable, f represents the frequency variable, g(t) represents the window function, and STFT(t, f) represents the result after short-time Fourier transform, which reflects the local characteristics of the signal at different times and frequencies; The autocorrelation function can be used to detect the periodicity and similarity of signals, and the cross-correlation function can be used to compare the similarity and time delay between two signals. Suppose the two signals are x(t) and y(t), then the cross-correlation function is Here τ represents the time delay, R xy (τ) represents the cross-correlation function value, which reflects the similarity of two signals at different time delays; S2.3, deep learning time domain feature extraction; A time domain feature extraction method based on deep learning is introduced, and the autoencoder is used to reconstruct the signal and extract more representative time domain features; The autoencoder automatically learns the potential feature representation of the signal through unsupervised learning; Assume that the input signal is x, the encoder function of the autoencoder is f(x), and the decoder function is g(f(x)). The goal of the autoencoder is to minimize the reconstruction error L(x, g(f(x))), using the mean square error loss function Here x represents the input signal, f(x) represents the encoder function, g(f(x)) represents the decoder function, and L(x, g(f(x))) represents the reconstruction error, which reflects the autoencoder's ability to reconstruct the input signal; S3, frequency domain feature extraction; S3.1, basic frequency domain transform; Use Fourier transform to convert time domain signal into frequency domain signal; Fourier transform is the most commonly used frequency domain transformation method, which can convert the signal from the time domain to the frequency domain. Let the signal be x(t), and its Fourier transform is in Here n represents the index of the discrete time point, k represents the index of the discrete frequency point, and X(k) represents the result after discrete cosine transform, which reflects the energy distribution of the signal at different frequencies; Wavelet transform can perform multi-resolution analysis on signals and is suitable for processing non-stationary signals. Suppose the signal is x(t) and the wavelet function is ψ(t), then the wavelet transform is: Here t represents the time variable, a is the scale parameter, which determines the width of the wavelet function, b is the translation parameter, which determines the position of the wavelet function, and WT(a, b) represents the result after wavelet transformation, which reflects the local characteristics of the signal at different scales and positions; S3.2, multi-level frequency domain analysis; First, use Fourier transform to get the basic frequency distribution of the signal, and then use wavelet transform to perform more detailed analysis on specific frequency intervals; Different wavelet basis functions can be selected according to the characteristics of the signal, such as Haar wavelet and Daubechies wavelet, to adapt to different signal characteristics; S3.3, high-order frequency domain statistics; Introducing high-order statistics of frequency domain features, such as calculating the skewness and kurtosis statistics of frequency domain signals, which can provide more information about the frequency distribution of the signal; Assume the frequency domain signal is X(f), the mean is μx, and the standard deviation is σx, then the skewness is The kurtosis is Here μx represents the mean of the frequency domain signal, σx represents the standard deviation, S represents the skewness, which reflects the distribution symmetry of the frequency domain signal, and K represents the kurtosis, which reflects the peak degree of the distribution of the frequency domain signal; S4, build a deep learning model; S4.1, Model structure selection; Design a deep neural network model that can simultaneously receive time domain features and frequency domain features as input and output the classification results of communication signals; The model structure can include convolutional neural network (CNN), recurrent neural network (RNN), or a combination thereof; CNN is good at processing local features of images and signals, while RNN is good at processing the temporal dependencies of sequence data; The appropriate model structure can be selected according to the characteristics of the signal and the task requirements; For example, for signals with spatially local features, CNN can be selected; For signals with time series characteristics, RNN or CNN-RNN combination can be selected; Let the input signal be x, and the convolutional layer output of CNN be y conv =W*X+b Here W is the convolution kernel, which determines the weight of the convolution operation, X represents the input feature map, b is the bias term, and y conv Represents the output result of the convolutional layer; The hidden state update formula of RNN is h t =tanh(W xh x t +W hh h t-1 +b h ) Where W xh is the weight matrix input to the hidden layer, W hh is the weight matrix from hidden layer to hidden layer, b h is the bias term; Here x t represents the input signal at time point t, h t represents the hidden state at time point t; S4.2, introduce special structures; Introducing the attention mechanism and residual network structure to improve the model's ability to process complex features and recognition accuracy; The attention mechanism allows the model to focus on the important parts of the signal and ignore the unimportant parts; Residual networks can solve the gradient vanishing problem in deep neural networks and improve the training effect of the model; Different types of attention mechanisms can be used, such as spatial attention, channel attention, etc., and selected according to the characteristics of the signal; The residual network can adopt different residual block structures, such as basic residual block, bottleneck residual block, etc., to adapt to different task requirements; Let the input feature be X, and the output of the attention mechanism be Y=Att(X) Where Att is the attention function; The output of the residual network is Y=X+F(X) Where F is the residual function; Here X represents the input feature, Y represents the output result of the attention mechanism or residual network, Att(X) represents the processing result of the attention function on the input feature, and F(X) represents the processing result of the residual function on the input feature; S4.3, multimodal fusion; Use multimodal fusion technology to better integrate time domain features and frequency domain features; Use feature splicing and attention weighting methods; Feature splicing is to directly splice the time domain features and frequency domain features together as the input of the model; Attention weighting is to perform weighted fusion of time domain features and frequency domain features by learning an attention weight matrix, highlighting important features and suppressing unimportant features; Let the time domain feature be T, the frequency domain feature be F, and the feature after feature concatenation be Y = [T; F]; The feature after attention weighting is Y=αT+(1-α)F Where α is the attention weight; Here T represents the time domain features, F represents the frequency domain features, Y represents the fused features, [T; F] represents the splicing of the time domain features and the frequency domain features together, and α represents the attention weight, which determines the relative importance of the time domain features and the frequency domain features. S5, model training and optimization; S5.

1. Data preparation; Using a large amount of labeled bimodal communication signal data, including its time domain and frequency domain features; Divide the data into training set, validation set and test set; The training set is used to train the model, the validation set is used to adjust the model's hyperparameters, and the test set is used to evaluate the model's performance; S5.2, parameter adjustment; The deep learning model is trained by adjusting model parameters, selecting appropriate optimizers, and designing reasonable loss functions to improve the recognition accuracy and generalization ability of the model; Model parameters include convolution kernel size, number of layers, and number of hidden layer nodes; The optimizer selects an optimization algorithm with adaptive learning rate, such as AdamW optimizer, which can automatically adjust the learning rate according to the gradient of each parameter to speed up the convergence of the model; The loss function can be a combination of cross entropy loss and mean square error loss. Let the true label be y and the predicted label be The loss function is L=αL ce +βL mse in is the cross entropy loss; is the mean square error loss; Here y represents the true label, represents the predicted label, L represents the loss function, which reflects the difference between the model prediction result and the true label, L ce represents the cross entropy loss, which measures the difference in probability distribution in the classification task, L mse represents the mean square error loss, which measures the numerical difference in the regression task. α and β represent weight coefficients, which are used to adjust the relative importance of the cross entropy loss and the mean square error loss in the total loss function. S5.

3. Hyperparameter adjustment; Compare the performance of different hyperparameter combinations through experiments and select the optimal hyperparameter settings; Hyperparameters include learning rate, batch size, and number of training rounds; Grid search and random search methods can be used to adjust hyperparameters; S6, signal recognition and classification; S6.1, signal preprocessing and feature extraction; After preprocessing, time domain feature extraction and frequency domain feature extraction of the dual-mode communication signal to be identified, a high-quality feature representation of the signal is obtained; S6.2, model input and output; Input into the trained deep learning model; The model outputs the classification result of the communication signal according to the input feature vector, realizing the accurate recognition of the dual-mode communication signal; S6.3, Result evaluation and feedback; To evaluate the recognition results, the accuracy, recall rate, and F1 value indicators can be used for evaluation; Assume that the true positive example is TP, the true negative example is TN, the false positive example is FP, and the false negative example is FN, then the accuracy is The recall rate is The F1 value is in Here TP represents true positive examples, that is, the number of samples correctly classified as positive; TN represents true negative examples, that is, the number of samples correctly classified as negative; FP stands for false positive, that is, the number of samples that are incorrectly classified as positive; FN represents false negative examples, that is, the number of samples that are misclassified as negative; The accuracy reflects the overall accuracy of the classification results, the recall reflects the ability to identify positive samples, and the F1 value comprehensively considers the accuracy and recall, and is a balanced evaluation indicator.

2. According to a dual-mode communication signal recognition method based on deep learning according to claim 1, it is characterized in that: In the signal acquisition and preprocessing steps, a signal quality assessment module is added; This module evaluates the quality of the collected signal in real time by calculating the signal-to-noise ratio, energy distribution and other indicators; If the signal quality is below a preset threshold, the system can automatically adjust sampling parameters or take additional signal enhancement measures to ensure the accuracy of subsequent processing; Let the signal be x(t) and the noise be n(t), then the signal-to-noise ratio is Where T is the signal length, which represents the number of discrete points in time; x(t) represents the signal value at time t, n(t) represents the noise value at time t, SNR represents the signal-to-noise ratio, which reflects the relative strength of the signal and the noise; The energy distribution is obtained by calculating the energy spectral density of the signal. Assuming the Fourier transform of the signal is X(f), the energy spectral density is S(f)=|X(f)| 2 Here f represents the frequency variable, X(f) represents the Fourier transform result of the signal, and S(f) represents the energy spectrum density, which reflects the energy distribution of the signal at different frequencies.

3. According to a dual-mode communication signal recognition method based on deep learning according to claim 2, it is characterized in that: In the time domain feature extraction step, the time domain dynamic characteristics of the signal are further explored; For example, the instantaneous frequency and instantaneous amplitude of the signal are calculated to change with time. These characteristics can better reflect the rapid changes and non-stationary characteristics of the signal in the time dimension. Let the signal be x(t), and its Hilbert transform is The instantaneous amplitude of the signal is The instantaneous frequency is Here t represents the time variable, x(t) represents the original signal, y(t) represents the signal after Hilbert transform, A(t) represents the instantaneous amplitude, which reflects the amplitude of the signal at time t, and ω(t) represents the instantaneous frequency, which reflects the frequency change of the signal at time t.

4. The dual-mode communication signal recognition method based on deep learning according to claim 3 is characterized in that: In the frequency domain feature extraction step, high-order statistics of frequency domain features are introduced; For example, by calculating the skewness and kurtosis statistics of frequency domain signals, these statistics can provide more information about the frequency distribution of the signal and help improve the model's ability to distinguish different types of signals.

5. A dual-mode communication signal recognition method based on deep learning according to claim 4, characterized in that: In the step of building a deep learning model, try different network architectures and hyperparameter combinations; For example, adjusting the convolution kernel size, number of layers, activation function and other parameters of the convolutional neural network, or changing the unit type and hidden layer size of the recurrent neural network; By comparing the performance of different combinations through experiments, the optimal model architecture and hyperparameter settings are selected.

6. A dual-mode communication signal recognition method based on deep learning according to claim 5, characterized in that: It also includes steps for continuous optimization of deep learning models; When new communication signal types emerge or the communication environment changes, the system can automatically collect new data for incremental learning and gradually adjust the model to improve its adaptability and accuracy; Let the new data be D new , the old model parameters are θ old , then the new model parameters can be obtained by updating the old model parameters, that is, It is the gradient of the loss function on the new data to the model parameters, which is used to guide the update direction of the parameters; Here D new represents the newly collected data, θ old represents the old model parameters, θ new Represents the updated model parameters, α represents the learning rate, which controls the step size of parameter update.

7. A dual-mode communication signal recognition method based on deep learning according to claim 6, characterized in that: The method is applicable to a wider range of communication scenarios, including but not limited to intelligent transportation communication, industrial automation communication, aerospace communication, etc., and has broader application prospects and practical value. At the same time, combined with actual application needs, corresponding visualization tools and interfaces are developed to facilitate users to view and analyze signal recognition results.

8. A dual-mode communication signal recognition method based on deep learning according to claim 7, characterized in that: During the model training process, data enhancement methods can be used to increase the diversity of training data and improve the generalization ability of the model; Data enhancement methods can include random cropping, random flipping, random rotation, adding noise, etc. For example, for image signals, random cropping and random flipping can be performed to increase the diversity of the image; for audio signals, different types of noise can be added to simulate different environmental noise conditions.

9. The dual-mode communication signal recognition method based on deep learning according to claim 8, characterized in that: In order to improve the real-time performance of the method, parallel computing methods can be used to accelerate the processes of signal acquisition, preprocessing, feature extraction, and model training; Parallel computing uses multi-core processors, graphics processing units (GPUs), and distributed computing technology; Use GPU to accelerate the training of deep learning models and increase the training speed; use distributed computing framework to distribute signal acquisition and preprocessing tasks to multiple nodes for parallel processing to improve processing efficiency.

Images