A blind identification method for communication signal modulation mode for intelligent instruments

Abstract

The present invention provides a blind recognition method for communication signal modulation mode for intelligent instruments, which aims to empower the instrument, free it from dependence on manual experience, and realize automated and intelligent electromagnetic signal measurement. The received signal is subjected to rough frequency estimation, frequency correction and signal detection. If the signal detection result is a communication signal, the pre-recognition type of the signal is determined by a signal recognition network based on convolutional attention (CNN-Attention). According to the pre-recognition type, the communication signal modulation mode is further classified, and the amplitude keying (ASK) type, phase shift keying (PSK) type, and quadrature amplitude modulation (QAM) type are identified respectively. The method proposed in the present invention gets rid of the dependence on manual experience, effectively alleviates the influence of noise, time-frequency asynchrony and radio frequency channels, solves the problems of low modulation recognition accuracy and poor algorithm robustness brought about by the existing methods, and greatly improves the accuracy of modulation recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent processing of communication signals, and in particular to a blind identification method for communication signal modulation modes for intelligent instruments. The method can be deployed in a new generation of intelligent electronic measuring instruments for electromagnetic signal monitoring scenarios and radio management scenarios. Background Art

[0002] Currently, the intelligence level of electronic measuring instruments is relatively low. When measuring electromagnetic signals, such as communications and radar, measurement parameters must be configured based on manual experience. This not only complicates the measurement process but also relies heavily on manual experience, making standardized, accurate measurements difficult to achieve. Communication signal modulation identification is one of the key measurement parameters in electronic measuring instruments and a primary task in radio monitoring. Regarding communication signal modulation identification, on the one hand, due to the influence of noise, time-frequency asynchrony, and the RF channel, the feature differentiation between different modulation modes is poor, making traditional modulation identification methods that rely solely on expert knowledge difficult to achieve accurate identification. On the other hand, existing deep learning-based modulation identification methods mostly use the original signal waveform or simple transform domain features to directly train deep learning networks for modulation identification. These methods fail to consider the generalizability and deployability of the technology and are therefore unsuitable for real-world scenarios. Therefore, high-precision automated modulation identification technology is crucial to improving the intelligence level of instruments. As a key parameter of communication signals, modulation is a key technical module in the new generation of intelligent electronic measuring instruments. Its goal is to empower instruments, freeing them from reliance on manual experience and enabling automated and intelligent electromagnetic signal measurement.

[0003] Patent document CN112769722A (application number: CN202110008169.1) discloses a method for automatically identifying the modulation type of a communication signal. This method extracts features through methods such as spectrum peaks and time domain envelope standard deviations, and recognizes the modulation mode of the communication signal based on the extracted multidimensional features. This method is extremely susceptible to noise, time-frequency asynchrony, and radio frequency channels, and the robustness of the recognition algorithm needs to be improved.

[0004] Patent document CN116319210A (application number: CN202310291803.6) discloses a lightweight automatic modulation recognition method for signals based on deep learning. This method uses a lightweight convolutional neural network to identify the modulation mode of communication signals, breaking away from the traditional method's reliance on manual experience, and the model complexity is low. However, it only considers a single feature of the communication signal modulation mode, and the generalization of the recognition algorithm needs to be improved.

[0005] Patent document CN114611550A (Application Number: CN202210210961.X) discloses a multi-feature automatic modulation recognition method based on a complex convolution module. This method integrates features from multiple signal dimensions, fully extracting the complex domain characteristics of the signal, and achieving excellent recognition results. However, this method does not consider technical modules such as signal detection and coarse signal estimation, making it unsuitable for actual wireless communication scenarios. It only recognizes the modulation mode of communication signals after preprocessing by the receiver. Summary of the Invention

[0006] In response to the problems existing in the prior art, a blind identification method for communication signal modulation mode for intelligent instruments is provided. This method gets rid of the reliance on manual experience, can effectively alleviate the influence of noise, time-frequency asynchrony and radio frequency channels, solves the problems of low modulation recognition accuracy and poor algorithm robustness brought by traditional methods, and greatly improves the accuracy of modulation recognition; and the deep learning networks used in this technology are all lightweight deep neural networks, which have the advantages of small algorithm model and low algorithm computational complexity, and can be deployed in various types of measuring instruments; in addition, in actual wireless communication scenarios, the frequency, code rate and other information of the signal are unknown, and the receiver needs to obtain a rough estimate of the frequency and complete the preliminary correction of the frequency. Furthermore, the signal detection technology is used to determine whether the received signal is a communication signal. The present invention fully considers the needs of actual application scenarios and designs receiver preprocessing modules such as signal detection and signal rough estimation, which can be applicable to actual wireless communication scenarios.

[0007] The technical solution adopted by the present invention is as follows: a method for blindly identifying the modulation mode of a communication signal, oriented to intelligent electronic measuring instruments, comprising:

[0008] Step 1: Perform power normalization processing on the received signal to obtain a normalized received signal;

[0009] Step 2: performing coarse frequency estimation and coarse frequency correction on the normalized received signal to obtain a corrected signal;

[0010] Step 3: Use the energy-based signal frequency domain detection method to determine whether the corrected signal is a pure noise signal or a communication signal. If it is a pure noise signal, proceed to step 6. If it is a communication signal, record it as a signal to be identified and proceed to step 4.

[0011] Step 4: Input the signal to be identified into the trained CNN-Attention-based signal recognition network and output the signal class pre-identification type. The signal class pre-identification type includes three modulation modes: ASK, PSK, and QAM.

[0012] Step 5. If the signal to be identified belongs to the ASK modulation mode, the signal modulation mode is identified by combining the multi-time constellation diagram and deep learning. If the signal to be identified belongs to the PSK modulation mode, the signal to be identified is subjected to multiple square spectrum transformations, and then input into the signal recognition network based on the residual network, which integrates the multi-dimensional domain features of the signal and outputs the signal modulation mode identification result. If the signal to be identified belongs to the QAM modulation mode, the signal modulation mode is identified by combining the fourth-power spectrum transformation, the multi-time constellation diagram and deep learning.

[0013] Step 6: Output the modulation mode identification result of the signal modulation.

[0014] Furthermore, the specific sub-steps of step 2 include:

[0015] Step 2.1, perform power spectrum estimation on the received signal to obtain the signal power spectrum;

[0016] Step 2.2, perform amplitude spectrum estimation on the received signal to obtain the signal amplitude spectrum;

[0017] Step 2.3: Calculate a rough estimate of the frequency offset of the received signal based on the signal power spectrum and the signal amplitude spectrum;

[0018] Step 2.4: Perform coarse frequency correction on the received signal based on the coarse frequency offset estimate.

[0019] Furthermore, the specific sub-steps of step 3 include:

[0020] Step 3.1, perform a fast Fourier transform on the corrected signal, and perform a zero operation on the amplitude outside the demodulation bandwidth of the fast Fourier transform result to obtain a spectrum;

[0021] Step 3.2, set the calculation threshold factor;

[0022] Step 3.3, sort the frequency points in the spectrum in ascending order according to the amplitude, and mark the frequency points at the front of the sort as a frequency point set with no signal;

[0023] Step 3.4, calculating the threshold value of the initial state according to the frequency point set with no signal;

[0024] Step 3.5: Compare the threshold value with the amplitude of the next frequency point outside the set of no-signal frequencies. If the threshold value is larger, the frequency point is included in the set of no-signal frequencies and the process proceeds to step 3.4 until all frequencies are traversed. At this point, the signal is pure noise. If the threshold value is smaller, communication signals exist at the current frequency point and subsequent frequency points.

[0025] Furthermore, in step 4, the CNN-Attention network is supervisedly trained using the time domain signal waveform to establish a signal class pre-recognition model. Based on the signal class pre-recognition model, the signal to be identified is analyzed and the modulation mode identification result of the signal is output. The identification result is one of the ASK, PSK or QAM modulation modes.

[0026] Furthermore, in step 5, if the signal to be identified belongs to the ASK modulation mode, the specific identification method includes:

[0027] Step 5.1.1. Perform symbol rate estimation and frequency precision estimation on the signal to be identified to obtain a precise frequency offset estimate and symbol rate estimate of the signal;

[0028] Step 5.1.2: Perform precise frequency correction on the signal based on the precise frequency offset estimate, and calculate the number of sampling points for each symbol based on the symbol rate estimate.

[0029] Step 5.1.3, calculate the multi-time constellation diagram of the signal based on the number of sampling points of each symbol and the corrected signal;

[0030] Step 5.1.4: Input the multi-time constellation diagram into the VGG-based signal recognition network in sequence. If the constellation diagram analysis result at a certain moment is one of the 2ASK, 4ASK or 8ASK modulation modes, then output the modulation mode recognition result corresponding to the signal.

[0031] Furthermore, in step 5, if the signal to be identified belongs to a PSK modulation mode, the specific identification method includes:

[0032] Step 5.2.1, perform multiple power spectrum transformations on the signal to be identified to obtain the signal's second power spectrum, fourth power spectrum, and eighth power spectrum;

[0033] Step 5.2.2: Process the 2nd, 4th, and 8th power spectra into corresponding 3D tensors respectively, and concatenate them into a 3D tensor of the signal according to the channel dimension.

[0034] Step 5.2.3: Input the three-dimensional tensor of the signal into the trained residual network-based signal recognition network to complete the recognition of PSK modulation mode.

[0035] Furthermore, in step 5, if the signal to be identified belongs to a QAM modulation mode, the specific identification method includes:

[0036] Step 5.3.1, perform a fourth-power spectrum transformation on the signal to be identified to obtain the fourth-power spectrum of the signal;

[0037] Step 5.3.2: Input the 4th power spectrum into the MLP-based signal recognition network and output the pre-identification type of the QAM modulation mode. If the pre-identification type of the signal to be identified is 16QAM, 64QAM, or 256QAM, proceed to step 5.3.3. If the pre-identification type of the signal to be identified is 32QAM or 128QAM, proceed to step 5.3.4.

[0038] Step 5.3.3: Perform precise frequency estimation and symbol rate estimation on the signal to be identified, obtaining precise frequency deviation and symbol rate estimates. Then, perform precise frequency correction on the signal based on the precise frequency deviation estimate, and calculate the number of sampling points for each symbol based on the symbol rate estimate. Then, calculate the multi-time constellation of the signal based on the number of sampling points for each symbol. Finally, input the multi-time constellation into the VGG-based signal recognition network, and output the specific modulation mode identification result of the signal.

[0039] Step 5.3.4: Estimate the symbol rate of the signal to be identified and obtain the symbol rate estimate of the signal; then calculate the number of sampling points for each symbol based on the symbol rate estimate; then calculate the multi-time constellation diagram of the signal based on the number of sampling points for each symbol; finally, input the multi-time constellation diagram into the VGG-based signal recognition network in sequence, and output the specific modulation method recognition result of the signal.

[0040] Furthermore, the frequency precise estimation method is as follows: performing a fourth power spectrum transformation on the signal to be identified to obtain the fourth power spectrum of the signal, and obtaining a precise frequency deviation estimation value based on the peak spectrum line of the fourth power spectrum.

[0041] Furthermore, the CNN-Attention-based signal recognition network includes a feature extraction module, a compression excitation module, a multi-head self-attention mechanism module and a classifier module;

[0042] The feature extraction module is composed of a convolution module a, a convolution module b and a convolution module c in sequence, wherein the convolution module a includes a connected convolution layer, batch normalization, and an activation function ReLu, the convolution module b includes a connected convolution layer, batch normalization, an average pooling layer, and an activation function ReLu, and the convolution module c includes a connected convolution layer, batch normalization, an activation function ReLu, and a dropout layer;

[0043] The compression excitation module is composed of a convolution module c, a compression module, an excitation module and a feature reweighting module in sequence. The convolution module c includes a connected convolution layer a, an activation function ReLu, and a convolution layer b. The compression module is a global average pooling layer. The excitation module includes a connected fully connected layer a, an activation function ReLu, a fully connected layer b, and an activation function Sigmoid.

[0044] The multi-head self-attention mechanism module is sequentially composed of a multi-head self-attention mechanism and a feedforward network submodule, wherein the multi-head self-attention mechanism module includes a connected causal convolution layer, a multi-head self-attention layer, and a residual unit, and the feedforward network submodule includes a connected layer normalization a, a fully connected layer a, a layer normalization b, a fully connected layer b, and a residual unit;

[0045] The classifier is composed of a flattening module, an average pooling layer, a fully connected layer, and a Softmax layer in sequence. The flattening module is used to convert multi-dimensional input features into one dimension.

[0046] The feature reweighting module is used to reweight different feature channels.

[0047] Furthermore, the VGG-based signal recognition network is composed of a convolution module a, a convolution module b, and a classifier module in sequence;

[0048] The convolution module a includes a connected two-dimensional convolution layer a, a two-dimensional convolution layer b, and a pooling layer;

[0049] The convolution module b includes a connected two-dimensional convolution layer a, a two-dimensional convolution layer b, a two-dimensional convolution layer c, and a pooling layer;

[0050] The classifier module includes a connected flattening module, a fully connected layer a, an activation function ReLu, a fully connected layer b, a fully connected layer c, and a softmax layer.

[0051] Furthermore, the signal recognition network based on the residual network is composed of a two-dimensional convolutional layer, a residual module a, a residual module b, a residual module c, a residual module d, a fully connected layer a, a fully connected layer b, and a Softmax layer in sequence;

[0052] The two-dimensional convolutional layer adopts a three-dimensional input mode of H*W*C;

[0053] The residual module a, residual module b, residual module c, and residual module d are respectively composed of 3 basic modules, 4 basic modules, 6 basic modules, and 3 basic modules in sequence, wherein the basic module is composed of a two-dimensional convolution layer a, batch normalization, an activation function ReLu, a two-dimensional convolution layer b, batch normalization, a residual unit, and an activation function ReLu in sequence.

[0054] Furthermore, the MLP-based signal recognition network is composed of a convolution module and a classifier module in sequence;

[0055] The convolution module includes a two-dimensional convolution layer, a pooling layer, an activation function ReLu, and batch normalization;

[0056] The classifier module includes a flattening module fully connected layer a, a fully connected layer b, an activation function ReLu, a fully connected layer c, and a Softmax layer.

[0057] Compared with the existing technology, the beneficial effects of adopting the above technical solution are:

[0058] (1) The present invention fully considers the needs of practical application scenarios. Since the frequency, code rate and other information of the signal in actual wireless communication scenarios are unknown, the receiver needs to obtain a rough estimate of the signal frequency and perform preliminary correction of the signal frequency. Furthermore, it uses signal detection technology to determine whether the received signal is a communication signal, and then completes the subsequent communication signal modulation mode identification function. Therefore, the present invention designs receiver preprocessing modules such as signal detection and rough signal estimation, which can be applied to actual wireless communication scenarios.

[0059] (2) The present invention proposes a method for blind recognition of communication signal modulation modes, which integrates signal processing methods and deep learning methods. On the one hand, the method uses signal processing methods to extract multi-dimensional domain features of signals, fully explores the different characteristics of different communication signal modulation modes, and effectively improves the generalization of modulation recognition; on the other hand, the method uses deep learning methods to automatically explore the potential characteristics of multi-dimensional domain features, gets rid of the dependence on manual experience, and can effectively alleviate the influence of noise, time-frequency asynchrony and radio frequency channels, solving the problems of low modulation recognition accuracy and poor algorithm robustness brought by traditional methods, and greatly improving the accuracy of modulation recognition.

[0060] (3) The present invention fully considers the technical deployment issues in actual application scenarios and can be deployed in various measuring instruments. On the one hand, each submodule designed in the present invention takes into account the trade-off between computational complexity and algorithm performance. The deep learning networks used are all lightweight networks, and the length of the signal sequences used is as short as possible. On the other hand, the data sets used for training and testing the deep learning networks of the present invention are all real signal data, which are transmitted by a signal generator, passed through a wireless channel, and received by a signal and spectrum analyzer. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is an implementation flow chart of the present invention.

[0062] Figure 2 This is the signal recognition network structure diagram based on CNN-Attention in the present invention.

[0063] Figure 3 This is a structural diagram of the signal recognition network based on VGG in the present invention.

[0064] Figure 4 This is a structural diagram of the signal recognition network based on residual in the present invention.

[0065] Figure 5 This is a structural diagram of the signal recognition network based on MLP in the present invention.

[0066] Figure 6 (a)~ Figure 6 (c) is a multi-time constellation diagram of the ASK modulation method in the present invention.

[0067] Figure 7 (a)~ Figure 7 (c) is the quadratic spectrum, the fourth power spectrum, and the eighth power spectrum of the PSK modulation method in the present invention.

[0068] Figure 8 It is the fourth power spectrum of the QAM modulation method in the present invention.

[0069] Figure 9 (a)~ Figure 9 (e) is a multi-time constellation diagram of the QAM type modulation method in the present invention.

[0070] Figure 10 This is a curve diagram of recognition accuracy of 14 typical communication signal modulation modes in the present invention.

[0071] Figure 11 This is a confusion matrix diagram for identifying 14 typical communication signal modulation modes in the present invention. DETAILED DESCRIPTION

[0072] The embodiments of the present application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar modules or modules with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application and are not to be construed as limiting the present application. On the contrary, the embodiments of the present application include all changes, modifications, and equivalents that fall within the spirit and scope of the appended claims.

[0073] In view of the problems that the existing solutions are difficult to accurately identify by simply relying on expert knowledge, and the modulation recognition methods based on deep learning mostly use the original waveform of the signal or simple transform domain features to directly train the deep learning network to achieve modulation recognition, without considering the generalization and deployability of the technology, and cannot adapt to the needs of actual scenarios, the embodiment of the present invention proposes a blind recognition method for communication signal modulation mode for intelligent instruments, aiming to empower the instrument, free it from dependence on manual experience, and realize automated and intelligent electromagnetic signal measurement. Specifically, according to the needs of actual application scenarios, the received signal is preprocessed, including rough frequency estimation, frequency correction and signal detection. If the signal detection result is a communication signal, the pre-recognition type of the signal is determined by a signal recognition network based on convolutional attention (CNN-Attention). Furthermore, according to the pre-identification type, the communication signal modulation mode is further classified, among which, the recognition of the amplitude keying (ASK) class adopts the multi-time constellation diagram and VGG network to classify the signal; the recognition of the phase shift keying (PSK) class adopts the quadratic spectrum, the fourth power spectrum, and the eighth power spectrum method to extract the characteristics of the signal respectively, and fuses the generated composite features, and then adopts the residual network to classify the signal; the recognition of the quadrature amplitude modulation (QAM) class first adopts the fourth power spectrum method to extract the signal characteristics, and then adopts the multilayer perceptron (MLP) network to pre-classify the signal, and then combines the multi-time constellation diagram and the VGG network to realize the signal classification of the submodule. The method proposed in the present invention gets rid of the dependence on manual experience, effectively alleviates the influence of noise, time-frequency asynchrony and radio frequency channel, solves the problems of low modulation recognition accuracy and poor algorithm robustness brought by the existing methods, and greatly improves the accuracy of modulation recognition. At the same time, the actual wireless communication scenarios are fully considered. Not only receiver preprocessing modules such as signal detection and rough signal estimation are designed, but also a lightweight deep neural network suitable for signal processing is designed, which can be deployed in various measuring instruments.

[0074] Please refer to Figure 1 The specific plan is as follows:

[0075] Step 1: Perform power normalization processing on the received signal to obtain a normalized received signal.

[0076] In this embodiment, the received signal is denoted as R(n), which is a communication signal collected in an actual communication scenario. The signal has a frequency offset, and the expression is as follows:

[0077] R(n)=S(n)e j2πnf

[0078] Where S(n) is the frequency-offset-free signal in an ideal scenario, and f is the actual frequency offset value of the signal.

[0079] The normalization process is as follows:

[0080]

[0081]

[0082] Where, P f Represents the power value of the received signal R(n), N represents the length of the signal, n∈[0,N], R a (n) is the normalized received signal.

[0083] Step 2: Perform coarse frequency estimation and coarse frequency correction on the normalized received signal to obtain a corrected signal.

[0084] Before signal recognition, it is necessary to normalize the received signal R a (n) Perform a rough frequency estimation, extract the frequency information of the signal, obtain a rough frequency estimate, and then correct the signal based on the rough frequency estimate. The specific steps are as follows:

[0085] Step 2.1: For signal R a (n) Perform power spectrum estimation to obtain the signal power spectrum P(k), which is calculated as follows:

[0086]

[0087] Step 2.2: For signal R a (n) Perform amplitude spectrum estimation to obtain the signal amplitude spectrum F(k), which is calculated as follows:

[0088]

[0089] Step 2.3: Calculate the frequency deviation estimate f of the signal based on the signal power spectrum P(k) and the signal amplitude spectrum F(k). a , the calculation formula is as follows:

[0090]

[0091] Step 2.4: Perform coarse frequency correction on the received signal based on the coarse frequency offset estimate. The calculation formula is as follows:

[0092]

[0093] Among them, S a (n) is the correction signal.

[0094] Step 3: Determine whether the correction signal is a pure noise signal or a communication signal.

[0095] In this embodiment, the signal S a (n) Perform fast Fourier transform and preprocessing to obtain the spectrum F a(k), then use the energy-based signal frequency domain detection method to detect the signal and determine the signal S a (n) is a pure noise signal or a communication signal. If it is a communication signal, it is recorded as the signal to be identified S b (n). The specific steps are as follows:

[0096] Step 3.1: Correct the signal S a (n) Perform fast Fourier transform and then zero the amplitude outside the demodulation bandwidth to obtain the spectrum F a (k), calculated as follows:

[0097] F a (k)=S(k)w(k)

[0098]

[0099]

[0100] Where S(k) is the spectrum without preprocessing, dBW is the demodulation bandwidth of the signal, and f s is the sampling frequency of the signal.

[0101] Step 3.2, set an appropriate calculation threshold factor μ;

[0102] Step 3.3: Spectrum F a The frequency points of (k) are sorted in ascending order by amplitude (excluding points with zero amplitude), and the first few frequency points after sorting are marked as frequency point set A with no signal. n , the interior point of the set is denoted as a i ,1≤i≤n, where n represents the size of the set.

[0103] Step 3.4: Calculate the threshold value of the initial state

[0104] Step 3.5: Compare the threshold value th with the frequency set A n The next frequency point a n+1 Amplitude, if the threshold is larger, then a n+1 Classified into set A n , update the set size n, and return to 3.4, update the threshold until all frequency points are traversed, at which point the signal is considered to be a pure noise signal; otherwise, it is considered a n+1 There are communication signals at the frequencies above and after it.

[0105] Step 4: Pre-identification between signal classes.

[0106] In this embodiment, the signal S b(n) Input the trained CNN-Attention based signal recognition network and output the signal class pre-recognition type. For details, please refer to Figure 2 The signal recognition network based on CNN-Attention includes a feature extraction module, a compression excitation module, a multi-head self-attention mechanism module, and a classifier module. The CNN-Attention network is supervised and trained using the same time domain signal waveform to establish a signal inter-class pre-recognition model. Based on the signal inter-class pre-recognition model, the signal to be identified S is analyzed. b (n) and outputs the modulation mode identification result of the signal, which is one of ASK, PSK or QAM modulation modes.

[0107] Among them, the feature extraction module is used to extract the signal to be identified S b The time-domain waveform features of (n) are sequentially constructed by convolutional modules a, b, and c. Convolutional module a includes a convolutional layer, batch normalization, and a ReLU activation function. Convolutional module b includes a convolutional layer, batch normalization, an average pooling layer, and a ReLU activation function. Convolutional module c includes a convolutional layer, batch normalization, a ReLU activation function, and a dropout layer.

[0108] The compression-excitation module is used to enhance important features and suppress irrelevant ones. It consists of a convolution module, a compression module, an excitation module, and a feature reweighting module. The convolution module includes convolution layer a, ReLu activation function, and convolution layer b. The compression module is a global average pooling layer, which performs average pooling on the features of each channel. The excitation module includes a fully connected layer a, ReLu activation function, a fully connected layer b, and a Sigmoid activation function. The feature reweighting module reweights different feature channels to increase the network's focus on important features.

[0109] The multi-head self-attention module consists of a multi-head self-attention mechanism and a feedforward network submodule. The multi-head self-attention module includes a causal convolutional layer, a multi-head self-attention layer, and a residual unit. The feedforward network submodule includes layer normalization a, fully connected layer a, layer normalization b, fully connected layer b, and a residual unit.

[0110] The classifier module is composed of a flattening module, an average pooling layer, a fully connected layer, and a softmax layer. The flattening module can transform multi-dimensional input features into one dimension. The fully connected layer fuses and generates signal composite features, and outputs a vector x L×1 , the calculation formula is as follows:

[0111] x L×1 =[x1,..,x l ,...,x L ] T ,l∈[1,L]

[0112] Where x l The weight representing the recognition result of the signal modulation type is the lth type, and L represents the number of signal modulation types. The function of the Softmax layer is to calculate the recognition probability q of the signal modulation type, q∈[0,1], and the calculation formula is as follows:

[0113]

[0114] Where y l The probability that the identification result of the signal modulation type is the lth type is calculated as follows:

[0115]

[0116] Step 5. For the pre-identification type between signal classes, different specific modulation mode identification methods are used for different types of modulation modes in this embodiment. If the signal to be identified belongs to the ASK modulation mode, the signal modulation mode is identified by combining the multi-time constellation diagram and deep learning; if the signal to be identified belongs to the PSK modulation mode, the signal to be identified is subjected to multiple square spectrum transformations, and then input into the signal recognition network based on the residual network, the multi-dimensional domain features of the signal are integrated, and the modulation mode identification result of the signal is output; if the signal to be identified belongs to the QAM modulation mode, the signal modulation mode is identified by combining the fourth-power spectrum transformation, the multi-time constellation diagram and deep learning. Specifically:

[0117] Step 5.1: If the signal to be identified S b (n) belongs to the ASK modulation mode. The specific identification method is as follows:

[0118] Step 5.1.1: Treat the identification signal S b (n) Perform symbol rate estimation and frequency precision estimation to obtain the signal frequency deviation precision estimation value f b and symbol rate estimate R s The precise frequency estimation is used to obtain the precise frequency information of the signal, which is different from the coarse frequency estimation.

[0119] The frequency estimation process is as follows:

[0120] Treat the identification signal S b (n) Perform a 4th power spectrum transformation to obtain the 4th power spectrum P of the signal b (k), calculated as follows:

[0121]

[0122] According to P b The frequency deviation estimate f is obtained by using the peak spectrum line of (k) b , the calculation formula is as follows:

[0123]

[0124] Where, f s represents the sampling frequency of the signal, f represents the frequency of the signal, and N represents the length of the signal.

[0125] The symbol rate estimation process is as follows:

[0126] Calculate signal S b (n) rate signal:

[0127] S r (n)=|S c (n+1)-S c (n)|*f s

[0128] For rate signal S r (n) Perform fast Fourier transform, and the calculation formula is as follows:

[0129]

[0130] According to P r (k) Solve the frequency point f corresponding to the peak of the spectrum line r , the frequency point f r The value of is the estimated symbol rate R s , the calculation formula is as follows:

[0131]

[0132] Step 5.1.2: Based on the frequency deviation estimate f b The signal is frequency-corrected according to the symbol rate estimate R s Calculate the number of samples per symbol sps.

[0133] The frequency correction expression is as follows:

[0134]

[0135] Where S c (n) represents the signal after frequency precision correction.

[0136] The calculation formula for the number of sampling points per symbol sps is as follows:

[0137]

[0138] Where, f s represents the sampling rate of the signal, and [·] represents the rounding operation.

[0139] Step 5.1.3: According to the number of sampling points per symbol sps and the corrected signal S c(n) Calculate the multi-time constellation diagram of the signal.

[0140] The calculation formula of the multi-time constellation diagram is as follows:

[0141] C k =m(r k )

[0142] r k =r(K*(n-1)+1+k),k∈[0,1,...,K-1]

[0143] K=sps,r=S c (n)

[0144] Where m(·) represents the constellation mapping function of the signal on the I / Q plane, r k represents the nth signal sequence of the extracted signal with a frequency shift of k, C k Represents signal sequence r k Constellation diagram, K is the oversampling factor, r is the signal S after frequency correction c (n).

[0145] Step 5.1.4: Input the multi-time constellation diagrams sequentially into the VGG-based signal recognition network. If the constellation diagram analysis result at a certain moment indicates one of the modulation modes: 2ASK, 4ASK, or 8ASK, the corresponding modulation mode identification result is output. The VGG network is a network model obtained through supervised training based on the constellation diagram at the optimal sampling time.

[0146] In this embodiment, please refer to Figure 3 The VGG-based signal recognition network consists of a convolutional module a, a convolutional module b, and a classifier module. Convolutional module a includes two-dimensional convolutional layers a and b, and a pooling layer. Convolutional module b includes two-dimensional convolutional layers a, b, c, and a pooling layer. The classifier module includes a flattening module, a fully connected layer a, a Reluctant Lu activation function, fully connected layers b and c, and a softmax layer.

[0147] Step 5.2: If the signal to be identified S b (n) belongs to the PSK modulation mode. The specific identification method is as follows:

[0148] Step 5.2.1: Treat the identification signal S b (n) Perform multiple power spectrum transformations to obtain the signal's second power spectrum P2(k), fourth power spectrum P4(k), and eighth power spectrum P8(k).

[0149] The calculation formulas for the square spectrum P2(k), the fourth spectrum P4(k), and the eighth spectrum P8(k) are as follows:

[0150]

[0151]

[0152]

[0153] Step 5.2.2: Process the quadratic spectrum, the fourth power spectrum, and the eighth power spectrum into corresponding three-dimensional tensors respectively, and concatenate them into a three-dimensional tensor of the signal according to the channel dimension.

[0154] That is, the second power spectrum P2(k), fourth power spectrum P4(k), and eighth power spectrum P8(k) are processed as Then splice it into a three-dimensional tensor of H*W*C according to the channel dimension.

[0155] Step 5.2.3: Input the three-dimensional tensor of the signal into the trained residual network-based signal recognition network to complete the recognition of PSK modulation mode.

[0156] In this embodiment, please refer to Figure 4 ,The signal recognition network based on residual network is composed of two-dimensional convolution layer, residual module a, residual module b, residual module c, residual module d, fully connected layer a, fully connected layer b, and Softmax layer in sequence.

[0157] Among them, the two-dimensional convolutional layer adopts the three-dimensional input mode of H*W*C, where H, W, and C represent the height, width, and channel dimensions respectively:

[0158]

[0159]

[0160]

[0161] Where, Represent the height of the 2nd power spectrum, the 4th power spectrum, and the 8th power spectrum respectively. Represent the width of the second power spectrum, the fourth power spectrum, and the eighth power spectrum respectively. They represent the channel dimensions of the quadratic spectrum, the fourth power spectrum, and the eighth power spectrum respectively.

[0162] Residual module a, residual module b, residual module c, and residual module d are composed of 3 basic modules, 4 basic modules, 6 basic modules, and 3 basic modules, respectively, and are used to extract multi-dimensional domain features of the signal. Among them, the basic module is composed of a 2D convolution layer a, batch normalization, activation function ReLu, a 2D convolution layer b, batch normalization, residual unit, and activation function ReLu, in that order.

[0163] Step 5.3: If the signal to be identified S b (n) Belongs to QAM modulation mode. Specific identification methods include:

[0164] Step 5.3.1: Treat the identification signal S b (n) Perform a 4th power spectrum transformation to obtain the 4th power spectrum P of the signal b (k).

[0165] Among them, the fourth power spectrum P b The calculation formula for (k) is as follows:

[0166]

[0167] Step 5.3.2, the fourth power spectrum P b (k) Input the MLP-based signal recognition network and output the pre-recognition type of the QAM modulation mode. If the pre-recognition type of the signal to be recognized is 16QAM, 64QAM, or 256QAM, proceed to step 5.3.3. If the pre-recognition type of the signal to be recognized is 32QAM or 128QAM, proceed to step 5.3.4. The MLP network is based on the fourth power spectrum P b (k) The network model obtained after supervised training.

[0168] In this embodiment, please refer to Figure 5 The MLP-based signal recognition network consists of a convolutional module and a classifier module. The convolutional module includes a two-dimensional convolutional layer, a pooling layer, a ReLu activation function, and batch normalization. The classifier module includes a flattening module, a fully connected layer a, a fully connected layer b, a ReLu activation function, a fully connected layer c, and a softmax layer.

[0169] Step 5.3.3: Treat the identification signal S b (n) Perform frequency estimation and symbol rate estimation to obtain the signal frequency deviation estimation value f b and the symbol rate estimate R s ; Then according to the frequency deviation accurate estimate f b Correct the signal frequency according to the symbol rate estimate R s Calculate the number of sampling points for each symbol; then calculate the multi-time constellation diagram of the signal based on the number of sampling points sps for each symbol; finally, input the multi-time constellation diagram into the VGG-based signal recognition network in sequence, and output the specific modulation method recognition result of the signal.

[0170] Among them, the frequency offset precise estimation, symbol rate estimation method, signal frequency correction method, symbol sampling point number calculation method, and multi-time constellation diagram calculation method are the same as steps 5.1.1 to 5.1.3 and will not be repeated here.

[0171] The VGG-based signal recognition network is the same as step 5.1.4.

[0172] Step 5.3.4: Treat the identification signal S b (n) Perform symbol rate estimation to obtain the signal symbol rate estimate R s ; Then according to the symbol rate estimate R s Calculate the number of sampling points sps for each symbol; then calculate the multi-time constellation diagram of the signal based on the number of sampling points sps for each symbol; finally, input the multi-time constellation diagram into the VGG-based signal recognition network in sequence, and output the specific modulation mode recognition result of the signal.

[0173] The symbol rate estimation method and the symbol sampling point calculation method are the same as those in steps 5.1.1 to 5.1.2, and the VGG-based signal recognition network is the same as that in step 5.1.4, so they will not be described here. The multi-time constellation calculation formula is the same as that in step 5.1.3, except that the frequency-corrected signal S c (n) is replaced by signal S b (n).

[0174] The recognition method of the present invention is further verified by simulation experiments below.

[0175] 1. Simulation parameter settings

[0176] In a simulation experiment for a blind recognition method for communication signal modulation for intelligent electronic measuring instruments, the algorithm performance was verified using 14 classic communication signal modulation methods, including binary amplitude-shift keying (2ASK), binary phase-shift keying (BPSK), offset quadrature phase-shift keying (OQPSK), π / 4 quadrature differential phase-shift keying (π / 4DQPSK), and hexadecimal quadrature amplitude modulation (16QAM). This method fully considers the signal recognition requirements in actual communication scenarios. The communication signal modulation method used is generated and transmitted by a Rohde & Schwarz SMW200A vector signal generator, passes through a wireless channel, and is then received and collected by a Rohde & Schwarz FSW signal and spectrum analyzer. The signal and spectrum analyzer can select an appropriate demodulation bandwidth based on the bandwidth of the acquired signal. The corresponding relationship between demodulation bandwidth and sampling rate is shown in Table 1.

[0177] Table 1 Correspondence between demodulation bandwidth and sampling rate

[0178] Demodulation bandwidth Sampling rate 300KHz 1.4MHz 500KHz 2MHz 800KHz 3.5MHz 1000KHz 4MHz 1250KHz 5.6MHz 1500KHz 7MHz 2000KHz 7MHz 5000KHz 14MHz 8000KHz 28MHz 10000KHz 28MHz

[0179] For the signal training dataset, the signal source and spectrum analyzer were set to the same center frequency. This resulted in only a small frequency offset due to differences in the local oscillators between the instruments. However, in actual communication scenarios, the receiver cannot pre-determine the signal's carrier frequency and can only constrain the received signal within a suitable demodulation bandwidth. This results in a wide range of frequency shifts. Considering the massive data required for deep learning model training and the demands of actual communication scenarios, data augmentation is performed on the received signal. This involves artificially adding a frequency offset. The value of this frequency offset is determined by the demodulation bandwidth of the signal, and the frequency offset is randomly added within the range that can be contained by the demodulation bandwidth. Table 2 shows the parameters collected for the communication signal modulation scheme.

[0180] For the signal test data set, the signal source and spectrum analyzer are no longer set to the same center frequency. Instead, the collected signals take into account actual communication scenarios, including both small and large frequency offsets. Furthermore, the signal recognition method of the present invention is completely blind, eliminating the need for obtaining prior information about the signal, such as symbol rate, frequency offset, filter type, and filter coefficients.

[0181] Table 2 Parameters of communication signal modulation method

[0182]

[0183]

[0184] According to the specific implementation scheme of the present invention, there are 14 identifiable communication signal modulation modes; the signal length N used for the signal power normalization processing operation is 8192; the signal length N used for the rough estimation of the signal frequency is 4096, and the signal length N used for the signal frequency correction operation is 8192; the signal length N used for signal detection is 8192, and the threshold factor μ is 3.

[0185] In the signal recognition simulation based on CNN-Attention, the signal to be recognized S with length N = 2500 is b The real and imaginary parts of (n) are concatenated according to the channel dimension and then input into the network. The network parameters of the CNN-Attention are shown in Table 3. The output dimension of the fully connected layer b in the classifier module is L = 3, representing the three pre-recognition types: ASK, PSK, and QAM. The output dimension of the softmax layer is L = 3, and the dimension with the highest probability value represents the modulation recognition result of the signal.

[0186] Table 3. Parameters of signal recognition network based on CNN-Attention

[0187]

[0188]

[0189] If the signal to be identified S b (n) belongs to ASK modulation mode, for S b (n) Symbol rate estimation and frequency precision estimation are performed, using a signal length of N = 8192. Furthermore, the multi-time constellation diagram of the signal is calculated and sequentially input into the VGG-based signal recognition network. In the multi-time constellation method, all signal lengths are N = 8192 / sps, and the number of sampling points per symbol, sps, is calculated from the sampling rate of the input signal and the symbol rate estimate. The image size of the multi-time constellation diagram input to the VGG network is 64 × 64 × 3. The VGG network parameters are shown in Table 4. The output dimension of the fully connected layer c is L = 3, representing the three recognizable ASK modulation types: 2ASK, 4ASK, and 8ASK.

[0190] Table 4. Parameters of signal recognition network based on VGG

[0191]

[0192]

[0193] If the signal to be identified S b (n) belongs to the PSK modulation mode, for S b (n) performs multiple power spectrum transformations to obtain the signal's quadratic spectrum P2(k), quadratic spectrum P4(k), and octadecane spectrum P8(k). The signal length used is N = 8192, and the image size of the signal spectrum line diagram is 64 × 64 × 3. Furthermore, the quadratic spectrum P2(k), quadratic spectrum P4(k), and octadecane spectrum P8(k) are preprocessed into 64 × 64 × 3 three-dimensional tensors, which are then concatenated into 64 × 64 × 9 tensors based on the channel dimension and input into the residual-based signal recognition network. The residual network parameters are shown in Table 5. The output dimension of the fully connected layer b is L = 6, representing the six recognized PSK modulation types: BPSK, QPSK, 8PSK, 16PSK, OQPSK, and π / 4DQPSK.

[0194] Table 5. Parameters of residual-based signal recognition network

[0195]

[0196]

[0197] If the signal to be identified S b (n) belongs to QAM modulation mode, for S b (n) Perform a 4th power spectrum transformation to obtain the 4th power spectrum P b(k), the signal length used is N = 8192, and the image size of the signal spectrum is 64 × 64 × 3. Furthermore, the fourth-order spectrum is input into the MLP-based signal recognition network. The parameters of the MLP-based signal recognition network are shown in Table 6. The output dimension of the fully connected layer c is L = 2, representing the pre-identified types of the two QAM modulation modes: 16QAM / 64QAM / 256QAM and 32QAM / 128QAM.

[0198] Table 6 Parameters of the signal recognition network based on MLP

[0199]

[0200] When the signal S b (n) When the modulation mode is 16QAM / 64QAM / 256QAM, the S b (n) Perform frequency and symbol rate estimation, using a signal length of N = 8192. Furthermore, the multi-time constellation diagram of the signal is calculated and sequentially input into the VGG-based signal recognition network. In the multi-time constellation method, all signal lengths are N = 8192 / sps, and the number of sampling points per symbol, sps, is calculated from the sampling rate of the input signal and the symbol rate estimate. The image size of the multi-time constellation diagram input to the VGG network is 64 × 64 × 3. The VGG network parameters used are shown in Table 3. The fully connected layer c has an output dimension of L = 3, representing the three recognizable QAM modulation types: 16QAM, 64QAM, and 256QAM.

[0201] When the signal S b (n) When the modulation mode is 32QAM / 128QAM, the S b (n) Symbol rate estimation is performed, using a signal length of N = 8192. Furthermore, the signal's multi-time constellation is calculated and sequentially fed into a VGG-based signal recognition network. In the multi-time constellation method, all signals have a length of N = 8192 / sps, and the number of sampling points per symbol, sps, is calculated from the input signal's sampling rate and the symbol rate estimate. The image size of the multi-time constellation fed into the VGG network is 64 × 64 × 3. The VGG network parameters used are similar to those shown in Table 3, with the only difference being the output dimension of the fully connected layer c, which has a dimension of L = 2, representing the two recognized QAM modulation types: 32QAM and 128QAM.

[0202] 2. Simulation Experiment Content and Results

[0203] Simulation 1: By Figure 6 (a)~ Figure 6(c) It can be seen that the multi-time constellation diagrams of 2ASK, 4ASK and 8ASK respectively, show that there is a certain moment when the characteristics of the constellation diagram are particularly significant. Among them, 2ASK presents two constellation points, 4ASK presents four constellation points, and 8ASK presents eight constellation points. The constellation diagrams of 2ASK, 4ASK and 8ASK are quite different and can be easily distinguished by the deep learning network.

[0204] Simulation 2: By Figure 7 (a)~ Figure 7 (c) It can be seen that these are the quadratic spectrum, the fourth spectrum, and the eighth spectrum of the PSK modulation method, respectively.

[0205] For the quadratic spectrum, the quadratic spectrum of BPSK presents a single-frequency peak shape; the quadratic spectrum of OQPSK presents a dual-frequency peak plus a triangular wide-band shape; and the quadratic spectrum of QPSK, π / 4DQPSK, 8PSK, and 16PSK presents a triangular wide-band shape, which are highly similar and difficult to distinguish.

[0206] For the fourth-power spectrum, the fourth-power spectrum of BPSK and OQPSK presents a single-frequency peak shape, with high similarity and difficult to distinguish; the fourth-power spectrum of QPSK presents a triple-frequency peak shape; the fourth-power spectrum of π / 4DQPSK presents a dual-frequency peak shape; 8PSK and 16PSK present a triangular wide-band shape, with high similarity and difficult to distinguish.

[0207] For the 8th power spectrum, the 8th power spectrum of BPSK, QPSK, and OQPSK presents a single-frequency peak shape, with high similarity and difficult to distinguish; the 8th power spectrum of π / 4DQPSK and 8PSK presents a triple-frequency peak shape, with high similarity and difficult to distinguish; and the 8th power spectrum of 16PSK presents a triangular wide-band shape.

[0208] By combining the characteristics of the square spectrum, fourth power spectrum, and eighth power spectrum of PSK modulation methods, it can be concluded that the single characteristics of some signals are similar, and PSK modulation methods cannot be accurately identified by relying on a single feature. However, by combining the square spectrum, fourth power spectrum, and eighth power spectrum of PSK modulation methods, it is theoretically possible to accurately identify the communication signal modulation type from the multi-dimensional domain characteristics of the signal.

[0209] Simulation 3: By Figure 8It can be seen that these are the fourth-power spectra of QAM modulation methods. Among them, the fourth-power spectra of 16QAM, 64QAM and 256QAM present the shape of a single frequency peak plus a triangular wide-band, while the fourth-power spectra of 32QAM and 128QAM present the shape of a triangular wide-band. The fourth-power spectra of 32QAM / 128QAM and 16QAM / 64QAM / 256QAM are quite different, and can be easily distinguished by deep learning networks.

[0210] Simulation 4: By Figure 9 (a)~ Figure 9 (e) It can be seen that the multi-time constellation diagrams of 16QAM, 32QAM, 64QAM, 128QAM and 256QAM, respectively, show that at a certain moment, the characteristics of the constellation diagram are particularly significant. Among them, the constellation diagrams of 16QAM, 64QAM and 256QAM are quite different. The constellation diagrams of 32QAM and 128QAM are also quite different, which can be distinguished by the deep learning network.

[0211] Simulation 5: By Figure 10 As shown in the figure, the recognition accuracy curves for 14 typical communication signal modulation methods show that at a signal-to-noise ratio of 14 dB, the overall accuracy of signal recognition reaches over 98%. To verify the effectiveness of the algorithm, two comparison algorithms were selected. Algorithm 1 is a modulation recognition method based on a convolutional neural network and a long short-term memory network. This method only extracts the time domain features of the signal, resulting in poor signal recognition accuracy. Algorithm 2 adds a coarse signal estimation and signal detection module to Algorithm 1. Compared to Algorithm 1, the signal recognition accuracy is significantly improved. However, due to the influence of noise, time-frequency asynchrony, and the RF channel, the signal recognition accuracy still needs to be improved.

[0212] Simulation 6: By Figure 11 As shown in the figure, the recognition confusion matrix diagram of 14 typical communication signal modulation modes in the present invention is shown. It can be seen that when the signal-to-noise ratio is 14dB, the recognition accuracy of 2ASK, 4ASK, 8ASK, BPSK, OQPSK, π / 4DQPSK, 16PSK, and 16QAM is relatively high, reaching more than 98%; the recognition accuracy of QPSK and 128QAM is more than 97%; the recognition accuracy of 8PSK, 32QAM, 64QAM and 256QAM is more than 95%.

[0213] It should be noted that, in the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "setting" and "connection" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection, or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances; the drawings in the embodiments are used to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations.

[0214] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A blind recognition method for communication signal modulation mode of intelligent instruments, characterized in that: include: Step 1: Perform power normalization processing on the received signal to obtain a normalized received signal; Step 2: performing coarse frequency estimation and coarse frequency correction on the normalized received signal to obtain a corrected signal; Step 3: Use the energy-based signal frequency domain detection method to determine whether the corrected signal is a pure noise signal or a communication signal. If it is a pure noise signal, proceed to step 6. If it is a communication signal, record it as a signal to be identified and proceed to step 4. Step 4: Input the signal to be identified into the trained CNN-Attention-based signal recognition network and output the signal class pre-identification type. The signal class pre-identification type includes three modulation modes: ASK, PSK, and QAM. Step 5. If the signal to be identified belongs to the ASK modulation mode, the signal modulation mode is identified by combining the multi-time constellation diagram and deep learning. If the signal to be identified belongs to the PSK modulation mode, the signal to be identified is subjected to multiple square spectrum transformations, and then input into the signal recognition network based on the residual network, which integrates the multi-dimensional domain features of the signal and outputs the signal modulation mode identification result. If the signal to be identified belongs to the QAM modulation mode, the signal modulation mode is identified by combining the fourth-power spectrum transformation, the multi-time constellation diagram and deep learning. Step 6: Output the modulation mode identification result of the signal modulation; In step 4, supervised training of the CNN-Attention network is performed using the time domain signal waveform to establish a signal inter-class pre-recognition model. Based on the signal inter-class pre-recognition model, the signal to be identified is analyzed and a modulation mode recognition result of the signal is output, where the recognition result is one of the ASK, PSK or QAM modulation modes. In step 5, if the signal to be identified belongs to the ASK modulation mode, the specific identification method includes: Step 5.1.

1. Perform symbol rate estimation and frequency precision estimation on the signal to be identified to obtain a precise frequency offset estimate and symbol rate estimate of the signal; Step 5.1.2: Perform precise frequency correction on the signal based on the precise frequency offset estimate, and calculate the number of sampling points for each symbol based on the symbol rate estimate. Step 5.1.3, calculate the multi-time constellation diagram of the signal based on the number of sampling points of each symbol and the corrected signal; Step 5.1.4: Input the multi-time constellation diagrams into the VGG-based signal recognition network in sequence. If the constellation diagram analysis result at a certain moment is one of the modulation modes of 2ASK, 4ASK, or 8ASK, then output the modulation mode identification result corresponding to the signal; In step 5, if the signal to be identified belongs to the PSK modulation mode, the specific identification method includes: Step 5.2.1, perform multiple power spectrum transformations on the signal to be identified to obtain the signal's second power spectrum, fourth power spectrum, and eighth power spectrum; Step 5.2.2: Process the 2nd, 4th, and 8th power spectra into corresponding 3D tensors respectively, and concatenate them into a 3D tensor of the signal according to the channel dimension. Step 5.2.3: Input the three-dimensional tensor of the signal into the trained residual network-based signal recognition network to complete the recognition of PSK-type modulation mode; In step 5, if the signal to be identified belongs to the QAM modulation mode, the specific identification method includes: Step 5.3.1, perform a fourth-power spectrum transformation on the signal to be identified to obtain the fourth-power spectrum of the signal; Step 5.3.2: Input the 4th power spectrum into the MLP-based signal recognition network and output the pre-identification type of the QAM modulation mode. If the pre-identification type of the signal to be identified is 16QAM, 64QAM, or 256QAM, proceed to step 5.3.

3. If the pre-identification type of the signal to be identified is 32QAM or 128QAM, proceed to step 5.3.

4. Step 5.3.3: Perform precise frequency estimation and symbol rate estimation on the signal to be identified, obtaining precise frequency deviation and symbol rate estimates. Then, perform precise frequency correction on the signal based on the precise frequency deviation estimate, and calculate the number of sampling points for each symbol based on the symbol rate estimate. Then, calculate the multi-time constellation of the signal based on the number of sampling points for each symbol. Finally, input the multi-time constellation into the VGG-based signal recognition network, and output the specific modulation mode identification result of the signal. Step 5.3.4: Estimate the symbol rate of the signal to be identified to obtain the symbol rate estimate of the signal; then calculate the number of sampling points for each symbol based on the symbol rate estimate; then calculate the multi-time constellation of the signal based on the number of sampling points for each symbol; finally, input the multi-time constellation into the VGG-based signal recognition network in sequence, and output the specific modulation mode recognition result of the signal; The frequency precision estimation method comprises the following steps: performing a fourth power spectrum transformation on the signal to be identified to obtain the fourth power spectrum of the signal, and obtaining a precise frequency deviation estimation value based on the peak spectrum line of the fourth power spectrum; The signal recognition network based on CNN-Attention includes a feature extraction module, a compression excitation module, a multi-head self-attention mechanism module and a classifier module; The feature extraction module is composed of a convolution module a, a convolution module b and a convolution module c in sequence, wherein the convolution module a includes a connected convolution layer, batch normalization, and an activation function ReLu, the convolution module b includes a connected convolution layer, batch normalization, an average pooling layer, and an activation function ReLu, and the convolution module c includes a connected convolution layer, batch normalization, an activation function ReLu, and a dropout layer; The compression excitation module is composed of a convolution module c, a compression module, an excitation module and a feature reweighting module in sequence. The convolution module c includes a connected convolution layer a, an activation function ReLu, and a convolution layer b. The compression module is a global average pooling layer. The excitation module includes a connected fully connected layer a, an activation function ReLu, a fully connected layer b, and an activation function Sigmoid. The multi-head self-attention mechanism module is sequentially composed of a multi-head self-attention mechanism and a feedforward network submodule, wherein the multi-head self-attention mechanism module includes a connected causal convolution layer, a multi-head self-attention layer, and a residual unit, and the feedforward network submodule includes a connected layer normalization a, a fully connected layer a, a layer normalization b, a fully connected layer b, and a residual unit; The classifier is composed of a flattening module, an average pooling layer, a fully connected layer, and a Softmax layer in sequence. The flattening module is used to convert multi-dimensional input features into one dimension. The feature reweighting module is used to reweight different feature channels; The VGG-based signal recognition network is composed of a convolution module a, a convolution module b, and a classifier module in sequence; The convolution module a includes a connected two-dimensional convolution layer a, a two-dimensional convolution layer b, and a pooling layer; The convolution module b includes a connected two-dimensional convolution layer a, a two-dimensional convolution layer b, a two-dimensional convolution layer c, and a pooling layer; The classifier module includes a connected flattening module, a fully connected layer a, an activation function ReLu, a fully connected layer b, a fully connected layer c, and a softmax layer; The signal recognition network based on the residual network is composed of a two-dimensional convolutional layer, a residual module a, a residual module b, a residual module c, a residual module d, a fully connected layer a, a fully connected layer b, and a Softmax layer in sequence; The two-dimensional convolutional layer adopts Three-dimensional input mode; where H, W, and C represent the height, width, and channel dimensions respectively; The residual module a, residual module b, residual module c, and residual module d are respectively composed of 3 basic modules, 4 basic modules, 6 basic modules, and 3 basic modules in sequence, wherein the basic module is composed of a two-dimensional convolution layer a, batch normalization, an activation function ReLu, a two-dimensional convolution layer b, batch normalization, a residual unit, and an activation function ReLu in sequence; The MLP-based signal recognition network is composed of a convolution module and a classifier module in sequence; The convolution module includes a two-dimensional convolution layer, a pooling layer, an activation function ReLu, and batch normalization; The classifier module includes a flattening module fully connected layer a, a fully connected layer b, an activation function ReLu, a fully connected layer c, and a Softmax layer.

2. The method for blindly identifying communication signal modulation modes for intelligent instruments according to claim 1, characterized in that: The specific sub-steps of step 2 include: Step 2.1, perform power spectrum estimation on the received signal to obtain the signal power spectrum; Step 2.2, perform amplitude spectrum estimation on the received signal to obtain the signal amplitude spectrum; Step 2.3: Calculate a rough estimate of the frequency offset of the received signal based on the signal power spectrum and the signal amplitude spectrum; Step 2.4: Perform coarse frequency correction on the received signal based on the coarse frequency offset estimate.

3. The method for blindly identifying communication signal modulation modes for intelligent instruments according to claim 1, characterized in that: The specific sub-steps of step 3 include: Step 3.1, perform a fast Fourier transform on the corrected signal, and perform a zero operation on the amplitude outside the demodulation bandwidth of the fast Fourier transform result to obtain a spectrum; Step 3.2, set the calculation threshold factor; Step 3.3, sort the frequency points in the spectrum in ascending order according to the amplitude, and mark the frequency points at the front of the sort as a frequency point set with no signal; Step 3.4, calculating the threshold value of the initial state according to the frequency point set with no signal; Step 3.5: Compare the threshold value with the amplitude of the next frequency point outside the set of no-signal frequencies. If the threshold value is larger, the frequency point is included in the set of no-signal frequencies and the process proceeds to step 3.4 until all frequencies are traversed. At this point, the signal is pure noise. If the threshold value is smaller, communication signals exist at the current frequency point and subsequent frequency points.

Images