A TCSE-ResNet50 numerical-analog hybrid signal recognition method combining frequency spectrum and fourth-order spectrum

Abstract

The application discloses a TCSE-ResNet50 digital-analog hybrid signal recognition method combined with frequency spectrum and fourth power spectrum. The application forms an image with rich features by combining signal frequency spectrum and fourth power spectrum, solves the problem that common constellation features cannot recognize analog signals and the problem that single spectrum features are only applicable to recognizing similar signals. And on the basis of the traditional ResNet50, a self-attention mechanism is added, and cross-entropy and triplet two types of loss functions are combined to optimize the network. By combining cross-entropy and triplet two types of loss functions, the cross-entropy is used to enlarge the feature distance between different types of signals with similar frequency domain forms, and the triplet is used to shorten the feature distance between similar signals caused by random baseband symbols or random additive noise. The application pays more attention to useful features, increases the discrimination degree of different types of samples, improves the robustness of the model, and can effectively recognize a digital-analog hybrid signal set of {FM, AM, 2ASK, BPSK, 2FSK, 16QAM, 16APSK}.

Description

Technical Field

[0001] This invention specifically relates to a TCSE-ResNet50 mixed-signal identification method based on the fourth power spectrum of the joint spectrum. Background Technology

[0002] Currently, the application of deep learning algorithms to modulation type recognition scenarios mostly focuses on the recognition of single digital modulation types, with less attention paid to the recognition of mixed digital and analog modulation types. Because current recognition networks use limited signal features, and analog signals lack commonly used recognition features such as cyclic spectra and constellation diagrams, existing mapping methods are unsuitable for recognizing mixed digital-analog signal sets.

[0003] Li Junchao's paper, "Identification of Time-Frequency Aliased Modulation Signals Based on Lightweight Networks," inputs a signal constellation diagram into the network for identification. Since analog signals lack constellation diagram features, this approach is unsuitable for identifying mixed-signal sets. Zhou Yuhang et al.'s paper, "Spectrum Signal Identification Based on Frequency Domain Superposition and Deep Learning," proposes a signal modulation identification algorithm based on frequency domain superposition and the YOLOv5-Radio network. It uses Fourier transform to convert the I / Q channels of the signal to be identified into spectrum data, and then uses frequency domain superposition to plot the spectrum data as a dataset. This approach is ineffective among similar signals with similar spectra. The above identification algorithms, due to their limited feature sets, are unsuitable for diverse mixed-signal sets, and these features are highly susceptible to noise, resulting in poor identification performance.

[0004] The paper "Modulation Intelligent Recognition Based on Phase Fusion Features and Residual Networks" published by Luo Qunping et al. uses signal constellation diagrams and vector trajectory diagrams as signal feature extraction data, and then inputs the data into a residual network for recognition. This mapping method is unsuitable for recognizing mixed-signal signal sets because analog signals lack constellation diagrams and vector trajectory features. Although the recognition algorithm uses joint features, these joint features are not applicable to mixed-signal signal sets, and there is still room for improvement in recognition performance.

[0005] The modulation scheme identification algorithms described above struggle to effectively identify signal sets that combine digital and analog modulation types. Furthermore, since both digital and analog modulation types exist in actual maritime satellite communication channels, common single-type digital or analog modulation identification algorithms face significant challenges in identifying signal sets with mixed digital and analog modulation. Therefore, developing effective classification and identification methods for mixed digital-analog signal sets is crucial. Summary of the Invention

[0006] The purpose of this invention is to improve the signal recognition capability of signal sets containing {FM (frequency modulation), AM (amplitude modulation), 2ASK (binary amplitude keying), BPSK (binary phase shift keying), 2FSK (binary frequency shift keying), 16QAM (hexadecimal quadrature amplitude modulation), 16APSK (hexadecimal amplitude phase shift keying)}, and to propose a TCSE-ResNet50 mixed-signal recognition method based on the fourth power spectrum of the combined spectrum.

[0007] This invention addresses the problem that common constellation diagram features cannot identify analog signals and that single spectral features are only suitable for identifying similar signals by combining signal spectra and fourth-order spectra to form feature-rich images. Furthermore, it adds a self-attention mechanism to the traditional ResNet50 and combines cross-entropy and triplet loss functions for network optimization. Because the proposed method of combining fourth-order spectra results in more features on discrete spectral lines, and the area of ​​these discrete spectral lines is narrower than the entire feature map, the attention mechanism module allows the model to focus more on these discrete spectral lines, minimizing interference from other background areas or random noise on the feature map. Simultaneously, the combined cross-entropy and triplet loss functions increase the feature distance between different types of signals with similar frequency domain representations using cross-entropy, and shorten the feature distance between similar signals caused by random baseband symbols or additive noise using triplet loss functions. Compared to the traditional ResNet50 network, the TCSE-ResNet50 proposed in this invention focuses more on useful features, while increasing the discriminative power of different sample types and improving the model's robustness. It can effectively identify mixed analog-digital signal sets {FM, AM, 2ASK, BPSK, 2FSK, 16QAM, 16APSK}. The algorithm proposed in this invention achieves an average signal recognition rate of over 93% at a signal-to-noise ratio of -2dB, outperforming single-feature input and traditional convolutional network recognition models.

[0008] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0009] Step 1: Generate experimental signals. The signal parameters involved in this invention are set as follows: sampling frequency of 4MHz, symbol rate of 500kHz, signal frequency deviation of 500Hz, and number of sampling points of 16384. The {2ASK, BPSK, 16QAM, 16APSK} signals are processed by a root-raised cosine pulse shaping filter with a shaping coefficient of 0.2. Additive white Gaussian noise is added using Matlab 2020a, with a signal-to-noise ratio range of -8 to 2dB.

[0010] Step 2: Constructing Feature Images. For the various experimental signals generated in Step 1, their spectral and fourth-order spectral data are obtained through Fourier transform. Then, the spectral and fourth-order spectral data for each type of experimental signal are obtained through data conversion. Data conversion can lead to information loss. To minimize conversion loss and more intuitively reflect the correlation between signal symbols and the integrity of signal features, this invention uniformly normalizes the image resolution to a 224×224 three-channel grayscale image. Finally, the spectral and fourth-order spectral data corresponding to each type of experimental signal are stitched together to obtain a new feature map.

[0011] Step 3: Build the TCSE-ResNet50 network model.

[0012] The TCSE-ResNet50 network uses the ResNet50 residual network and SENet (Squeeze-and-Excitation Networks) as its backbone network. By fusing residual and attention mechanisms, the SE module from SENet is embedded as an attention gating unit into the ResNet50 residual network, resulting in the SE-ResNet50 backbone network. SENet is divided into Squeeze and Excitation parts. Cross-entropy is used to widen the feature distance between different classes of experimental signals with similar frequency domain representations, while triples are used to narrow the feature distance between similar experimental signals caused by baseband symbol randomness or additive noise. Combining triples and cross-entropy loss improves the model's robustness, enabling faster convergence during training.

[0013] Step 4: Design the loss function.

[0014] In network training, triplet loss and cross-entropy loss are used to optimize the network to improve its representational ability. During signal recognition, due to the randomness of baseband symbols and additive noise, the feature distance between images within the same signal class can vary significantly. Cross-entropy loss only constrains the feature distance between classes, and its effect is not ideal for signals with similar discrete spectral representations. Therefore, this invention introduces triplet loss to expand the feature distance between different classes of signals with similar discrete spectral representations, and to reduce the feature distance between the same signals under random baseband symbols and random additive noise conditions.

[0015] Step 1 is implemented as follows:

[0016] The FM and AM signal models are as follows:

[0017]

[0018] Where A is the constant amplitude of the carrier wave, ω c K is the carrier angular frequency. fLet n be the modulation sensitivity, t be the current transmission time, and n0(t) be the expression that at time t follows the N(0,σ) modulation sensitivity. 2 The noise is Gaussian white noise, and τ is the integral variable. In this invention, m(t) is the voice signal at time t, and the FM signal is wideband frequency modulation (WBFM).

[0019] The 2FSK signal model is as follows:

[0020] s 2FSK (t)=a(n)cosω1t+b(n)cosω2t+n0(t) (2)

[0021] Where a(n) and b(n) are both unipolar pulse sequences, ω1 is the carrier angular frequency corresponding to a(n), and ω2 is the carrier angular frequency corresponding to b(n).

[0022] The amplitude and phase signal set includes {2ASK, BPSK, 16QAM, 16APSK}, a total of 4 modulation types, and its general signal model is:

[0023]

[0024] Where j represents an imaginary number, a(n) represents the amplitude of the nth symbol, T is the symbol period, representing the duration of a symbol, and g T (t) represents the impulse response of the root-mean-square pulse shaping filter, f c Let φ(t) be the center frequency of the signal, φ(t) be the modulation phase of the signal, which remains unchanged during the duration of one symbol, and φ0 represent the initial phase of the signal.

[0025] Step 2 is implemented as follows:

[0026] Existing deep learning models are primarily designed to solve image classification problems. The preprocessed signals are sequences, and to meet the requirements of deep learning models, specific data samples with recognizable features need to be converted into images. Since real-world maritime satellite communication channels contain both digital and analog modulation types, common recognition algorithms typically use signal constellation diagrams as feature images to classify digital signals. However, because analog signals lack constellation diagram features, recognition algorithms are unsuitable for the mixed digital and analog modulation signal sets found in real-world maritime satellite communications.

[0027] Since the signal spectrum does not have the limitations of constellation diagram features, it can be used for signal sets with mixed digital and analog modulation types. However, due to the problem of similar spectral features between different types of signals, this invention compensates for the shortcomings of single spectral feature recognition by combining the signal spectrum and the fourth power spectrum to form an image with rich recognition features. The spectral and fourth power spectrum feature expressions of the signal to be identified are as follows:

[0028]

[0029]

[0030] Both the signal's spectrum and its fourth power spectrum contain useful identification information. However, a single signal spectrum or its fourth power spectrum cannot effectively identify a set of seven mixed digital-analog signals.

[0031] Expanding the time-domain expression of the FM signal using trigonometric formulas, and combining it with the properties of the Bessel function, the frequency-domain expression obtained through Fourier transform shows that the spectrum of the FM signal is determined by the carrier component ω. c Composed of numerous side frequencies, where ω c =2πf c As shown in equations (1), (2), and (3), the AM, 2FSK, and 2ASK signals exhibit discrete impulse lines in their spectra. However, BPSK, 16QAM, and 16APSK signals, due to their zero-mean characteristic, exhibit a suppressed carrier effect in the frequency domain, causing their spectra to disperse within a certain envelope. Among these, the {FM, BPSK, 16QAM, and 16APSK} signals have low spectral discrimination, making identification difficult using a single spectral feature. Furthermore, reducing the signal-to-noise ratio blurs the spectral characteristics of other signals, affecting the recognition rate.

[0032] The time-domain expression of the signal can be simplified as shown in equation (6):

[0033] s(t)=m(t)·exp(jω c t)=m(t)·(cos(ω c t)+j·(sin(ω c t)) (6)

[0034] From equation (6), the time-domain expression for the fourth power of the signal is:

[0035] s 4 (t)=m 4 (t)·exp(j4ω c t)=m 4 (t)·(cos(4ω c t)+j·sin(4ω c t)) (7)

[0036] Since the spectral characteristics of a signal are actually a manifestation of the periodic and stationary statistical properties inherent in the signal itself, and because 16APSK signals have a constrained relationship between amplitude and phase, the statistical expectation mean of its fourth-power spectrum is approximately zero, therefore its fourth-power spectrum does not exhibit discrete impulse spectral characteristics. BPSK and 16QAM signals, on the other hand, have statistically independent amplitude and phase, i.e., m... 4(t) is a function with a period of T, so the fourth power spectrum of BPSK and 16QAM signals will have impact lines at the frequency points {0, ±1 / T, ±2 / T}.

[0037] Data conversion can lead to information loss. To minimize conversion loss and more intuitively reflect the correlation between signal symbols and the integrity of signal features, this invention uniformly normalizes the image resolution to a 224×224 three-channel grayscale image. The feature images of the seven signals to be identified at a 10dB signal-to-noise ratio are shown below. Figure 1 As shown.

[0038] As can be seen from the feature images, a single spectral feature or a single fourth-order spectral feature is insufficient for effectively identifying mixed-signal sets. Combining single spectral features and single fourth-order spectral features can effectively compensate for the shortcomings of each feature, improving the ability to identify mixed-signal sets. For example, FM, BPSK, 16QAM, and 16APSK signals cannot be identified using a single spectral feature, but can be identified by superimposing a fourth-order spectral feature; similarly, 2ASK, BPSK, and 16QAM signals are difficult to identify using a single fourth-order spectral feature, but can be distinguished by superimposing a spectral feature. Therefore, a new feature image can be formed by stitching the spectral and fourth-order spectral images together for identifying mixed-signal sets {FM, AM, 2ASK, BPSK, 2FSK, 16QAM, 16APSK}.

[0039] Step 3 is implemented as follows:

[0040] The TCSE-ResNet50 network uses ResNet50 (a residual network) and SENet (Squeeze-and-Excitation Networks) as its backbone. By fusing residual and attention mechanisms, the SE module is embedded as an attention gating unit into the ResNet50 network, resulting in the SE-ResNet50 backbone. SENet consists of two parts: Squeeze and Excitation. Due to the proposed joint fourth-order spectrum mapping method, features are primarily concentrated on discrete spectral lines. Since the region containing discrete spectral lines is narrower than the entire feature map, the SE module focuses attention more on these discrete spectral lines, minimizing interference from background regions or random noise on the feature map. Cross-entropy is used to widen the feature distance between different classes of signals with similar frequency domain representations, while triples are used to narrow the feature distance between similar signals caused by baseband symbol randomness or additive noise. Combining triples and cross-entropy loss improves the model's robustness, enabling faster convergence during training. The TCSE-ResNet50 network model is as follows: Figure 2 As shown.

[0041] SENet employs a "feature recalibration" method for feature channel fusion. It learns and assigns weights to feature channels, increasing the weights of useful feature channels while decreasing the weights of less relevant ones. The SE module, for example... Figure 3 As shown. H and W represent the height and width of the feature map, respectively, and C represents the number of channels. During the Squeeze phase of the SE module, the feature map... Global average pooling is performed, and its formula is shown in Equation (8), resulting in a 1×1×C feature map. In the excitation stage, a two-layer multilayer perceptron (MLP) is used to model the correlation between channels. The number of neurons in the first layer is C / r (r is the compression ratio), which realizes feature dimensionality reduction. The activation function is ReLU (Rectified Linear Units). The number of neurons in the second layer is C, which increases the feature dimensionality back to the original dimension. The purpose of this is to increase the nonlinearity of the features. After activation by the Sigmoid function, a one-dimensional vector with the same number of channels C is obtained as the weight of each channel. This weight is then weighted onto the feature of each channel through a scaling operation, that is, the calculated weight M(X) is multiplied by the feature X of the corresponding channel to obtain the recalibrated feature Y, as shown in Equations (9) and (10). This process makes the model pay more attention to the features of channels with high information content, while ignoring the features of unimportant channels as much as possible.

[0042]

[0043] M(X)=σ(W1·ReLU(W0X avg (9)

[0044]

[0045] Among them, X avg σ represents the average global feature information, σ is the Sigmoid normalization function, and W0 and W1 represent the weights of the fully connected operation.

[0046] Step 4 is implemented as follows:

[0047] In network training, to improve the representational ability of the network model, triplet loss and cross-entropy loss are used to optimize the network. The calculation formulas are as follows:

[0048] Loss = Loss Triplet +λLoss cross-entropy (11)

[0049] In the formula, Loss is the joint loss. cross-entropyIt's cross-entropy loss, Loss Triplet λ is the triplet loss, and λ is a hyperparameter representing the weight relationship between the cross-entropy loss and the triplet loss. In this invention, it is set to 2.

[0050] 4-1. The formula for cross-entropy loss is as follows:

[0051]

[0052] In the formula, N represents the total number of images in the training batch of the network model. b The number of images in a batch is represented by D, and the number of signal types is represented by f. i v represents the feature vector of the feature image of the i-th signal. i f i The corresponding real category, U j Let b represent the j-th column vector of matrix U. j This indicates the bias term.

[0053] In signal recognition, due to the randomness of baseband symbols and additive noise, the feature distance between images within the same signal class can vary significantly. Cross-entropy loss only constrains the feature distance between classes, and its effect is not ideal for signals with similar discrete spectral patterns. Therefore, this invention introduces triplet loss to expand the feature distance between different classes of signals with similar discrete spectral patterns, and to reduce the feature distance between the same signals under random baseband symbols and random additive noise conditions.

[0054] 4-2. The formula for triplet loss is as follows:

[0055] Loss Triplet =∑[||f(X) a )-f(X p )||-||f(X a )-f(X n )||β] (13)

[0056] In the formula, X a X is a feature image obtained by combining the fourth power spectrum of the current signal. p Is with X a Difficult sample images of the same category and batch; X n Is with X a Difficult sample images from the same batch of different categories; f(X) a f(X) p f(X) n) represent the feature vectors output by the model from the feature images of the corresponding signals; ||·|| is the square of the 2-norm; β is the minimum interval between the feature distances between feature images of different signals and between feature images of the same type of signal, which is set to 0.4 in this invention. In the triplet loss, feature distances between feature images of different signals less than β are considered hard samples, i.e., feature images of different types of signals with similar spectral representations; feature distances between feature images of the same type of signal greater than β are considered hard samples, i.e., feature images of the same type of signal with different spectral representations due to baseband symbol randomness and additive noise randomness. The triplet loss focuses more on training hard samples. Through continuous network training, the triplet loss continuously decreases, making the feature distances between feature images of different types of signals with similar spectral representations and between feature images of the same type of signal with different spectral representations due to baseband symbol randomness and additive noise randomness closer to β.

[0057] The beneficial effects of this invention are as follows:

[0058] 1. This invention targets a signal set of seven types {FM, AM, 2ASK, BPSK, 2FSK, 16QAM, 16APSK}, and employs a TCSE-ResNet50 mixed-signal recognition algorithm based on the fourth power spectrum of the joint spectrum. By integrating deep learning technology into the field of modulation mode recognition, it is more innovative than traditional modulation mode recognition methods and improves classification and recognition capabilities.

[0059] 2. Compared with other deep learning-based modulation scheme recognition algorithms, this invention does not suffer from the disadvantage of insufficient single feature image input features, nor the disadvantage of joint feature input being unsuitable for analog signals. It can more effectively identify digital-analog mixed signal sets.

[0060] 3. The TCSE-ResNet50 network proposed in this invention has the advantage of being more suitable for mixed-signal datasets compared to the ResNet network. It introduces an attention mechanism module, allowing the model to focus more on discrete spectral lines and minimize interference from other background regions or random noise on the feature map for signal recognition. Simultaneously, it combines two loss functions: cross-entropy and triplet. Cross-entropy widens the feature distance between different classes of signals with similar frequency domain representations, while triplet narrows the feature distance between similar signals caused by baseband symbol randomness or additive noise.

[0061] 4. The signal-to-noise ratio is higher than -2dB, and the recognition model can achieve a recognition rate of over 93% for FM, AM, 2ASK, BPSK, 2FSK, 16QAM, and 16APSK.

[0062] In summary, the algorithm of this invention features low complexity, high recognition rate, and strong robustness, and maintains a high recognition rate even under conditions of low signal-to-noise ratio and frequency offset. Attached Figure Description

[0063] Figure 1 Three-channel images of seven modulation types at a signal-to-noise ratio of 10dB;

[0064] Figure 2 Diagram of the TCSE-ResNet50 network model;

[0065] Figure 3 This is a diagram of the SE module;

[0066] Figure 4 Feature visualization heatmaps for the two network models;

[0067] Figure 5 The classification and recognition performance of different networks under different signal-to-noise ratios. Detailed Implementation

[0068] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0069] Step 1: Generate the signals {FM, AM, 2ASK, BPSK, 2FSK, 16QAM, 16APSK} using Matlab 2020a. The signal parameters are set as follows: sampling frequency 4MHz, symbol rate 500kHz, signal frequency deviation 500Hz, number of sampling points 16384. The {2ASK, BPSK, 16QAM, 16APSK} signals are filtered by a root-raised cosine pulse shaping filter with a shaping coefficient of 0.2. The signal-to-noise ratio range is -8 to 2dB.

[0070] Step 2: Since the signal features contained in the signal spectrum are not limited by constellation diagram features, they can be used for signal sets with mixed digital and analog modulation types. However, due to the problem of similar spectral features between different types of signals, this invention compensates for the shortcomings of single spectral feature recognition by combining the signal spectrum and the fourth power spectrum to form an image with rich recognition features. For example, BPSK, 16QAM, and 16APSK signals have zero mean characteristics, which manifest as carrier suppression in the frequency domain, causing the signal spectrum to be dispersed within a certain envelope range. Among them, the spectral differentiation of {FM, BPSK, 16QAM, 16APSK} signals is low, making it difficult to complete the identification through a single spectral feature; 2ASK, BPSK, and 16QAM signals are difficult to complete the identification through a single fourth power spectrum feature, but after superimposing the spectral features, 2ASK, BPSK, and 16QAM can be distinguished. Therefore, the spectrum diagram and the fourth power spectrum diagram are stitched together to form a new feature map for the identification of mixed digital and analog signal sets {FM, AM, 2ASK, BPSK, 2FSK, 16QAM, 16APSK}.

[0071] Step 3: The TCSE-ResNet50 network uses the residual network ResNet50 and SENet as the backbone network. By fusing residual and attention mechanisms, the SE module is embedded into the ResNet50 network as an attention gating unit to obtain the backbone network SE-ResNet50

[10] . Among them, SENet is divided into two parts: Squeeze and Excitation.

[0072] SENet employs a "feature recalibration" method for fusing feature channels. It learns and assigns weights to feature channels, increasing the weights of useful channels while decreasing the weights of less relevant ones. Here, H and W represent the height and width of the feature map, respectively, and C represents the number of channels. During the Squeeze phase of the SE module, the feature map... Global average pooling is performed to obtain a 1×1×C feature map. In the excitation stage, a two-layer multilayer perceptron (MLP) is used to model the correlation between channels. The first layer has C / r neurons (r is the compression ratio) to achieve feature dimensionality reduction, using ReLU (Rectified Linear Units) as the activation function. The second layer has C neurons, which increases the feature dimensionality back to the original dimension. This is done to increase the non-linearity of the features. After activation by the sigmoid function, a one-dimensional vector with the same number of channels C is obtained as the weight of each channel. This weight is then applied to the features of each channel through a scaling operation, i.e., multiplying the calculated weight by the corresponding channel's feature. This process allows the model to pay more attention to the features of channels with high information content, while ignoring features of less important channels as much as possible.

[0073] Step 4: During signal recognition, due to the randomness of baseband symbols and additive noise, the feature distance between images within the same signal class may vary significantly. Cross-entropy loss only constrains the feature distance between classes, which is not ideal for signals with similar discrete spectral representations. Therefore, this invention introduces triplet loss to expand the feature distance between different classes of signals with similar discrete spectral representations and reduce the feature distance between the same signals under random baseband symbols and additive noise. In triplet loss, feature distances less than β between feature images of different signals are considered hard samples, i.e., feature images of different classes of signals with similar spectral representations; feature distances greater than β between feature images of the same class of signals are considered hard samples, i.e., feature images of the same class of signals with different spectral representations due to random baseband symbols and additive noise. Triplet loss focuses more on training hard samples. Through continuous network training, triplet loss continuously decreases, making the feature distances between feature images of different classes of signals with similar spectral representations and between feature images of the same class of signals with different spectral representations due to random baseband symbols and additive noise closer to β.

[0074] Example:

[0075] 1. Experimental Dataset

[0076] Simulations of seven mixed-signal signals (FM, AM, 2ASK, BPSK, 2FSK, 16QAM, 16APSK) were implemented using MATLAB software. A joint feature map of the spectrum and fourth power spectrum was constructed, and based on this, training and test set image samples were obtained for this invention, with no overlap between the two image samples. The signal parameters were set as follows: sampling frequency of 4MHz, symbol rate of 500kHz, signal frequency deviation of 500Hz, and number of sampling points of 16384. The {2ASK, BPSK, 16QAM, 16APSK} signals were filtered by a root-raised cosine pulse shaping filter with a shaping coefficient of 0.2. Additive white Gaussian noise was added using Matlab 2020a, with a signal-to-noise ratio (SNR) ranging from -8 to 2 dB, resulting in image data for 7 different signal types and 6 different SNRs: -8 dB, -6 dB, -4 dB, -2 dB, 0 dB, and 2 dB. The training set contained 4000 images for each signal type and each SNR, while the validation set contained 1000 images for each signal type and each SNR.

[0077] 2. Experimental Environment and Network Configuration

[0078] The experimental server GPU was an NVIDIA TITAN RTX 3090, the deep learning framework used was PyTorch, the input image was a 3-channel image with a resolution of 224×224, the model training used a batch size of 64, the number of training iterations was 50, the initial learning rate was 3e-4, and the learning rate decay factor for each training was 0.7.

[0079] 3. Feature Visualization

[0080] Under a 2dB signal-to-noise ratio environment, the joint fourth-order spectrum image of seven digital-analog mixed signals is visualized in the feature heatmap of the TCSE-ResNet50 network and the ResNet50 network proposed in this invention, as shown below. Figure 4 As shown, different colors in the heatmap represent the degree of attention paid to regional features, with the red areas in the heatmap representing the feature regions that the network model focuses on during classification. Figure 4 It is known that the ResNet50 network has problems such as focusing too much on the noise base region and the surrounding background region. The TCSE-ResNet50 network proposed in this invention can effectively avoid the network model focusing too much on the noise base region and background interference, so that the network model focuses on the region with obvious spectral feature distribution, thereby improving the classification and recognition performance.

[0081] 4. Recognition Performance Analysis

[0082] The algorithm proposed in this invention, the recognition algorithm in "Spectrum Signal Recognition Based on Frequency Domain Superposition and Deep Learning" (Reference 1), and the recognition method for "Mixed-Analog Signal Recognition Method Based on Multi-Level Dynamic Blind Digital Receiver" (Reference 2) were respectively used in MATLAB 2018a software for signal recognition simulation tests. The recognition rates of the three algorithms are as follows: Figure 5 As shown.

[0083] from Figure 5 It can be seen that when the signal-to-noise ratio is higher than 0dB, the recognition rate of the joint spectrum quartic spectrum TCSE-ResNet50 hybrid signal recognition algorithm proposed in this invention can reach more than 99%. In addition, under the traditional ResNet50 network, when the signal-to-noise ratio is higher than 0dB, the new synthesis method proposed in this invention can achieve an average recognition rate of more than 95%. Compared with the traditional algorithm, the classification and recognition method based on deep learning has a significant improvement under different signal-to-noise ratios. Under low signal-to-noise ratios, the deep learning method has a significant improvement over the traditional method. In particular, when the signal-to-noise ratio is lower than -6dB, the deep learning method is significantly better than the traditional method. Compared with the single spectrum feature recognition algorithm mentioned in reference [1], since FM, BPSK, 16QAM and 16APSK signals have similar frequency domain representations, it is difficult for a single spectrum feature to identify all signals. When the signal-to-noise ratio is higher than 2dB, the average recognition rate of the signal can only reach more than 70%. When the signal-to-noise ratio is lower than -6dB, since noise is suppressed by spectrum superposition, the performance of the single spectrum feature recognition algorithm is not significantly different from the recognition algorithm proposed by us. However, as the signal-to-noise ratio increases, the effectiveness of noise suppression decreases, making the limitations of single-spectrum feature recognition algorithms more apparent.

[0084] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A TCSE-ResNet50 mixed-signal recognition method based on the fourth power spectrum of the combined spectrum, which combines the signal spectrum and the fourth power spectrum to form a feature-rich image, solving the problems that common constellation diagram features cannot recognize analog signals and that single spectral features are only applicable to recognizing similar signals; characterized in that, The method of constructing the fourth power spectrum of the joint spectrum, based on ResNet50, adds a self-attention mechanism and combines two types of loss functions, cross-entropy and triplet, for network optimization. The specific steps include the following: Step 1: Generate experimental signals; The sampling frequency is 4MHz, the symbol rate is 500kHz, the signal frequency deviation is 500Hz, and the number of sampling points is 16384. The {2ASK, BPSK, 16QAM, 16APSK} signals are passed through a root-raised cosine pulse shaping filter with a shaping coefficient of 0.

2. Additive white Gaussian noise is added using Matlab2020a, and the signal-to-noise ratio ranges from -8 to 2dB. Step 2: Construct feature images; For each type of experimental signal generated, its spectrum and fourth power spectrum data are obtained by Fourier transform. Then, the spectrum and fourth power spectrum of each type of experimental signal are obtained by data conversion. Then, the resolution of the spectrum and fourth power spectrum is uniformly normalized to a three-channel grayscale image of size 224×224. Finally, the spectrum and fourth power spectrum of each type of experimental signal are stitched together to obtain a new feature image. Step 3: Build the TCSE-ResNet50 network model; The TCSE-ResNet50 network uses ResNet50 and SENet as its backbone networks. By fusing residual and attention mechanisms, the SE module in SENet is embedded as an attention gating unit into the ResNet50 residual network to obtain the backbone network SE-ResNet50. SENet is divided into two parts: Squeeze and Excitation. Furthermore, cross-entropy is used to widen the feature distance between different types of experimental signals with similar frequency domain representations, and triples are used to narrow the feature distance between similar experimental signals caused by random baseband symbols or random additive noise. By combining triples and cross-entropy loss, the robustness of the model is improved, enabling the model to converge faster during training. Step 4: Loss function design; The network is optimized using triplet loss and cross-entropy loss. Cross-entropy is used to increase the feature distance between different classes of experimental signals with similar frequency domain representations, and triplet loss is used to increase the feature distance between different classes of signals with similar discrete spectral representations. The feature distance between the same signals under random baseband symbols and random additive noise is also reduced.

2. The TCSE-ResNet50 mixed-signal recognition method based on the fourth power spectrum of the joint spectrum according to claim 1, characterized in that... Step 1 is implemented as follows: The FM and AM signal models are as follows: in, For the constant amplitude of the carrier wave, The carrier angular frequency, To modulate sensitivity, At the current transmission time, for Always obey Gaussian white noise, For integration variables; in this invention for The voice signal at any given time is a wideband frequency modulation (WBFM) signal. The 2FSK signal model is as follows: in, and All are unipolar pulse sequences. for The corresponding carrier angular frequency, for The corresponding carrier angular frequency; The amplitude and phase signal set includes {2ASK, BPSK, 16QAM, 16APSK}, a total of 4 modulation types, and its general signal model is: in, represents an imaginary number, This represents the amplitude of the nth symbol. The symbol period represents the duration of a symbol. This represents the impulse response of the root-mean-square pulse shaping filter. The center frequency of the signal. The modulation phase of the signal remains constant throughout the duration of one symbol. This indicates the initial phase of the signal.

3. The TCSE-ResNet50 mixed-signal recognition method based on the fourth power spectrum of the joint spectrum according to claim 2, characterized in that... Step 2 is implemented as follows: The limitations of single-spectral feature recognition are compensated by combining the signal spectrum and the fourth-order spectrum to form an image with rich recognition features. The spectral and fourth-order spectrum feature expressions of the signal to be recognized are shown below: Both the spectrum and the fourth power spectrum of a signal contain useful identification information, but a single signal spectrum or the fourth power spectrum cannot effectively identify a set of seven mixed digital and analog signals. Expanding the time-domain expression of the FM signal using trigonometric formulas, and combining it with the properties of the Bessel function, the frequency-domain expression obtained through Fourier transform shows that the spectrum of the FM signal is composed of the carrier component. Composed of numerous side frequencies, among which As can be seen from equations (1), (2), and (3), the AM, 2FSK, and 2ASK signals exhibit discrete impact lines in their spectra. Meanwhile, the BPSK, 16QAM, and 16APSK signals, due to their zero-mean characteristics, exhibit a suppressed carrier effect in the frequency domain, causing their spectra to disperse within a certain envelope. Among these, the {FM, BPSK, 16QAM, and 16APSK} signals have low spectral differentiation, making identification difficult through a single spectral feature. Furthermore, reducing the signal-to-noise ratio can blur the spectral features of other signals, affecting the recognition rate. The time-domain expression of the signal can be simplified as shown in equation (6): From equation (6), the time-domain expression for the fourth power of the signal is: Since the spectral characteristics of a signal are actually a manifestation of the periodic and stable statistical properties inherent in the signal itself, and because 16APSK signals have a constrained relationship between amplitude and phase, the statistical expectation mean of its fourth-power spectrum is approximately zero, therefore its fourth-power spectrum does not exhibit discrete impulse spectral characteristics; BPSK and 16QAM signals, on the other hand, have statistically independent amplitude and phase, i.e. It is a function with a period of T, so the fourth power spectrum of BPSK and 16QAM signals will be... Impact spectral lines appear at frequency points; The spectrogram and the fourth power spectrogram are spliced ​​together to form a new feature map for the identification of mixed digital-analog signal sets {FM, AM, 2ASK, BPSK, 2FSK, 16QAM, 16APSK}.

4. The TCSE-ResNet50 mixed-signal recognition method based on the fourth power spectrum of the joint spectrum according to claim 2 or 3, characterized in that... Step 3 is implemented as follows: SENet employs a "feature recalibration" method for feature channel fusion. It learns and assigns weights to feature channels, increasing the weights of useful feature channels while decreasing the weights of less correlated ones. During the Squeeze phase of the SE module, the feature maps are... Global average pooling is performed, and its formula is shown in equation (8), yielding... Feature map To achieve feature dimensionality reduction, The compression ratio is [value]; the activation function is ReLU; and the number of neurons in the second layer is [value]. The feature map is then upscaled back to its original dimension, and then activated by the Sigmoid function to obtain a sum of channel numbers. The same one-dimensional vector is used as the weight for each channel, and this weight is then applied to the features of each channel through a scale operation, resulting in the calculated weights. Features of the corresponding channel Multiply to obtain the recalibrated features. As shown in equations (9) and (10): in, Represents average global feature information. The Sigmoid normalization function, , This represents the weight of the fully connected operation. These represent the height and width of the feature map, respectively. Indicates the number of channels.

5. The TCSE-ResNet50 mixed-signal recognition method based on the fourth power spectrum of the joint spectrum according to claim 4, characterized in that... Step 4 is implemented as follows: The network is optimized using triplet loss and cross-entropy loss, calculated as follows: In the formula, It is a joint loss. It is cross-entropy loss. It is a triplet loss. It is a hyperparameter that represents the weighted relationship between cross-entropy loss and triplet loss; 4-1. The formula for cross-entropy loss is as follows: In the formula, N represents the total number of images in the training batch of the network model. This indicates the number of images in a batch. Indicates the number of signal types. Indicates the first Feature vector of a feature image of a signal express The corresponding real category, Representation matrix The Column vectors Indicates the bias term; 4-2. The formula for triplet loss is as follows: In the formula, It is a feature image obtained by combining the fourth power spectrum of the current signal. Is with Images of difficult samples from the same category and batch; Is with Images of difficult samples from the same batch but from different categories; , , These are the feature vectors output by the model from the feature images of the corresponding signals; ||·|| is the square of the 2-norm; It is the minimum interval between the feature distances between feature images of different signals and between feature distances between feature images of the same type of signal, set to 0.4.

Images