A Smart Recognition Method for Communication Signal Modulation Types for the Industrial Internet

Abstract

This invention discloses an intelligent method for identifying the modulation type of communication signals for the Industrial Internet, relating to the field of radio, to address the problems of poor robustness, incompleteness, and practicality in existing technologies. This invention utilizes a MobileNet-based signal identification network for initial classification of received signals; simultaneously, it uses a KAN-LSTM-based signal identification network for amplitude modulation and frequency modulation classification; based on the classification results of the first two steps, it outputs the signal modulation type identification result, or continues to use a KAN-LSTM-based signal identification network, a MobileNet-based signal identification network, or a multimodal feature fusion network for progressive subdivision to obtain the final signal modulation type identification result. This invention improves the accuracy and robustness of the signal modulation type identification method, can accurately identify high-order CPFSK signals, and solves the problem of poor practicality in existing algorithms.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radio, in particular to a communication signal modulation type intelligent identification method for industrial internet. BACKGROUND

[0002] With the increase of wireless terminal devices in the industrial internet, various wireless signals are scattered in different frequency bands, which poses a challenge to how to efficiently manage and control large-scale wireless signals. General radio management and control methods only consider the frequency, bandwidth and power of wireless signals, which is difficult to find abnormal signals and has great limitations.

[0003] With the development of automatic modulation identification technology, a new radio management and control method is proposed. Specifically, when managing and controlling wireless signals, the modulation type, a deeper feature of the signal, is extracted. Generally, the communication between industrial devices will follow a specific frequency band and modulation type. If a modulation type different from the expected mode suddenly appears, it may indicate abnormal activities in the system, such as malicious attacks or unauthorized access.

[0004] Chinese patent document with publication number CN117354106A proposes a communication signal modulation identification method based on a re-parameterized causal convolutional network. This method uses neural networks to automatically learn features from I / Q data, avoiding the dependence on manual identification features. However, this method is easily affected by noise and channel. When the signal-to-noise ratio and channel factors change slightly, the I / Q sequence of the signal will change dramatically, and the recognition performance of the algorithm will decrease sharply, and the robustness of the algorithm is poor.

[0005] Chinese patent document with publication number CN105763499A proposes a CPFSK modulation identification method based on signal instantaneous frequency sliding average. This method extracts the instantaneous frequency of the signal and identifies 2CPFSK, 4CPFSK and 8CPFSK according to the variance of the extracted instantaneous frequency. However, it does not consider higher-order 16CPFSK, and the completeness of the algorithm needs to be improved.

[0006] Chinese patent document with publication number CN117857270A proposes a communication signal modulation identification method based on a double-channel VisionTransformer neural network. This method extracts the real and imaginary parts of the time-frequency features of the signal, aggregates the real and imaginary parts of the time-frequency features according to the self-attention coding feature map, and identifies FM, AM, CPFSK and other signals, and achieves good recognition results. However, this method uses MATLAB simulation data and does not consider real FM, AM broadcast signals and signals transmitted by different signal sources. The algorithm has performance problems when identifying real signals and different signal sources, and the recognition algorithm cannot be truly put into use. SUMMARY

[0007] The application aims at providing a communication signal modulation type intelligent identification method for industrial internet to solve at least one of the problems of poor robustness, poor completeness and poor practicability in the prior art.

[0008] The technical scheme adopted by the application is as follows:

[0009] The communication signal modulation type intelligent identification method for industrial internet comprises the following steps: S1, performing normalization processing on the power of a received signal S(n) to obtain a unit power signal S a (n); S2, extracting the frequency spectrum feature of the signal S a (n), and using a signal identification network based on MobileNet to obtain a preliminary classification result of the signal S a (n) in AM, FM, {AM broadcast, ASK}, {CPFSK, FM broadcast, PAM}; S3, using a signal identification network based on KAN-LSTM to distinguish the signal S a (n) as {AM, AM broadcast, ASK, PAM} or {FM, FM broadcast, CPFSK} according to the I / Q sequence of the signal S a (n); S4, if the signal S a (n) belongs to the AM modulation type, outputting the identification result of the signal modulation type as AM according to the classification results of steps S2 and S3; S5, if the signal S a (n) belongs to the FM modulation type, outputting the identification result of the signal modulation type as FM according to the classification results of steps S2 and S3; S6, if the signal S a (n) belongs to the {AM broadcast, ASK} modulation type, extracting the baseband signal feature and constellation feature of the signal S a (n), and combining the signal identification network based on KAN-LSTM and the signal identification network based on MobileNet to obtain the identification result of the modulation type; S7, if the signal S a (n) belongs to the PAM modulation type, extracting the multi-time constellation feature of the signal S a (n), and combining the multi-modal feature fusion network to obtain the identification result of the signal modulation type; S8, if the signal S a (n) belongs to the {CPFSK, FM broadcast} modulation type, extracting the baseband signal feature and constellation feature of the signal S aThe frequency information, multi-time constellation and spectrum characteristics of (n) are combined with a KAN-LSTM-based signal recognition network, a multi-modal feature fusion network and a MobileNet-based signal recognition network to obtain a recognition result of the signal modulation type.

[0010] In summary, due to the adoption of the technical solutions described above, the present application has the following beneficial effects:

[0011] 1. The present application fully considers the influence of noise, channel and other factors on signals, analyzes signals from multiple dimensions, extracts high-robustness features of signals, and further digs the essential features of signals in a learning-driven manner in combination with a deep learning algorithm, thereby improving the accuracy and robustness of the signal modulation type recognition method.

[0012] 2. In view of the problems of few existing CPFSK modulation recognition algorithms and incomplete recognition types, the present application extracts the frequency information of CPFSK, and extracts the multi-time constellation feature of signals through "pseudo-demodulation", which can accurately recognize 2CPFSK, 4CPFSK, 8CPFSK and 16CPFSK signals.

[0013] 3. In view of the problems that real FM broadcast is easily confused with CPFSK signals and AM broadcast is easily confused with ASK signals, the present application extracts the baseband signal features of these signals, relies on the characteristics that the baseband signal frequency of FM and AM broadcast always changes, improves the above problems, and improves the practicability of the recognition algorithm; in addition, in view of the problem that most existing algorithms use simulation data for algorithm research, and have insufficient performance when recognizing real signals, the present application uses two types of signal sources to transmit signals, collects signals through a spectrum monitoring receiver, and uses the data of the two types as a training set and a test set respectively, thereby fully verifying the practicability of the algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0014] The present application will be described by way of example and with reference to the accompanying drawings, in which:

[0015] Figure 1 is a flowchart of the communication signal modulation type intelligent recognition method of the present application in one embodiment;

[0016] Figure 2 is a signal recognition network structure diagram based on MobileNet in the present application;

[0017] Figure 3 is a signal recognition network structure diagram based on KAN-LSTM in the present application;

[0018] Figure 4 is a multi-modal feature fusion network structure diagram in the present application;

[0019] Figure 5is a common communication signal spectrum diagram listed by the embodiment of the application;

[0020] Figure 6 is a real part waveform diagram of the signal after removing the carrier residual of the AM broadcast in the embodiment of the application;

[0021] Figure 7 is a constellation diagram of the ASK modulation type in the embodiment of the application;

[0022] Figure 8 is a multi-time constellation diagram of the PAM modulation type in the embodiment of the application;

[0023] Figure 9 is a waveform diagram of the frequency information of the FM broadcast, CPFSK in the embodiment of the application;

[0024] Figure 10 is a multi-time constellation diagram of the CPFSK modulation in the embodiment of the application;

[0025] Figure 11 is a spectrum diagram of the signal 2 times after MSK, 2CPFSK in the embodiment of the application;

[0026] Figure 12 is a recognition accuracy curve diagram of 13 common communication signal modulation types in the embodiment of the application;

[0027] Figure 13 is a recognition confusion matrix diagram of 13 common communication signal modulation types in the embodiment of the application. DETAILED DESCRIPTION

[0028] All features disclosed in this specification, and / or the steps of any method or process specified in this specification, can be combined in any combination, except combinations where at least some of the features and / or steps are mutually exclusive.

[0029] Any feature disclosed in this specification, unless stated otherwise, can be replaced by any equivalent or other technically equivalent feature. That is, unless stated otherwise, each feature is one example only of a number of equivalent or similar features.

[0030] Embodiment 1

[0031] As shown in Figure 1 , the communication signal modulation type intelligent identification method for the industrial internet includes the following steps:

[0032] S1, the power of the received signal S(n) is normalized to obtain a unit power signal S a (n).

[0033] S11, the received signal S(n) is a signal in a real non-simulation scene, and the signal must have a carrier residual. The received signal is represented as:

[0034] S(n) = Y(n)e j2πnf ,

[0035] where Y(n) is an ideal received signal without carrier residual, and f is the carrier residual of the signal S(n).

[0036] S12, the power of the received signal S(n) is normalized:

[0037]

[0038] where N is the number of sampling points of the signal S(n), i.e. the signal length.

[0039] S2, extract the spectral features of the signal S a (n), and use the MobileNet-based signal recognition network to obtain the preliminary classification result of the signal S a (n) in AM, FM, {AM broadcast, ASK}, {CPFSK, FM broadcast, PAM} four modulation types according to the extracted spectral features of the signal.

[0040] Referring to Figure 2 , this step S2 performs fast Fourier transform and modulus value processing on the unit power signal of the historical received signal, extracts the spectral features, and uses the extracted spectral features to supervise the training of the MobileNet-based signal recognition network, thereby constructing a signal preliminary classification model based on the spectral features, with the classification results being AM, FM, {AM broadcast, ASK}, {CPFSK, FM broadcast, PAM}; the first spectral feature F(k) obtained by the fast Fourier transform and modulus value processing of the signal S a (n) is preliminarily classified by using the signal preliminary classification model, and the preliminary classification result is one of AM, FM, {AM broadcast, ASK}, {CPFSK, FM broadcast, PAM}.

[0041] The method of fast Fourier transform and modulus value processing is:

[0042]

[0043] The MobileNet-based signal recognition network is composed of a cascaded convolution module, a depth separable convolution module a, a depth separable convolution module b, a depth separable convolution module c, and a first classification module:

[0044] The convolution module includes a two-dimensional convolution layer, a two-dimensional batch normalization layer, and an activation function ReLU in sequence;

[0045] The deep separable convolution module a comprises, in sequence, a two-dimensional convolution layer a, a two-dimensional batch normalization layer a, an activation function ReLU, a two-dimensional convolution layer b, a two-dimensional batch normalization layer b, and an activation function ReLU;

[0046] The deep separable convolution module b comprises, in sequence, a two-dimensional convolution layer a, a two-dimensional batch normalization layer a, an activation function ReLU, a two-dimensional convolution layer b, a two-dimensional batch normalization layer b, and an activation function ReLU;

[0047] The deep separable convolution module c comprises, in sequence, a two-dimensional convolution layer a, a two-dimensional batch normalization layer a, an activation function ReLU, a two-dimensional convolution layer b, a two-dimensional batch normalization layer b, and an activation function ReLU;

[0048] The first classification module comprises, in sequence, a two-dimensional average pooling layer, a feature unfolding function, a linear layer, and a Softmax function.

[0049] S3, according to the signal S a (n) I / Q sequence, using a KAN-LSTM-based signal recognition network, distinguishing the signal S a (n) is {AM, AM broadcast, ASK, PAM} or {FM, FM broadcast, CPFSK}.

[0050] Reference Figure 3 : This step sends the I / Q sequence of the signal S a (n) into a KAN-LSTM-based signal recognition network, and divides the signal into amplitude modulation and frequency modulation. The identification result of the signal modulation type is {AM, AM broadcast, ASK, PAM} and {FM, FM broadcast, CPFSK}. The KAN-LSTM-based signal recognition network here has undergone sample training when used. The training samples select the I / Q sequence of the signal, and the labels are {AM, AM broadcast, ASK, PAM} and {FM, FM broadcast, CPFSK}.

[0051] The KAN-LSTM-based signal recognition network is composed of a cascaded spatial feature extraction module a, a spatial feature extraction module b, a time sequence feature extraction module a, a time sequence feature extraction module b, and a second classification module:

[0052] The spatial feature extraction module a comprises, in sequence, a one-dimensional convolution layer, a one-dimensional average pooling layer, a KAN network layer, a one-dimensional batch normalization layer, and an activation function ReLU;

[0053] The spatial feature extraction module b comprises, in sequence, a one-dimensional convolution layer, a one-dimensional average pooling layer, a KAN network layer, a one-dimensional batch normalization layer, and an activation function ReLU;

[0054] The time sequence feature extraction module a sequentially comprises an LSTM, an activation function ReLU, a one-dimensional batch normalization layer, a KAN network layer, and an activation function GELU.

[0055] The time sequence feature extraction module b sequentially comprises an LSTM, an activation function ReLU, a one-dimensional batch normalization layer, a KAN network layer, and an activation function GELU.

[0056] The second classification module sequentially comprises a KAN network layer a, a feature unfolding function, a one-dimensional average pooling layer, a KAN network layer b, and an activation function Softmax.

[0057] The KAN network is an application of the Kolmogorov-Arnold representation theorem to a neural network. The theorem states that any continuous function f(x1, x2,.., x d ) on the closed interval [0, 1] d can be represented as a superposition and composition of some univariate functions, specifically as follows:

[0058]

[0059] where Φ q and φ pq are continuous functions on the interval [0, 1]. Based on this theorem, the KAN network is responsible for replacing the weight parameters with learnable univariate functions.

[0060] S4, according to the classification results of steps S2 and S3, if the signal S a (n) belongs to the AM modulation type, the recognition result of the signal modulation type is AM.

[0061] That is, in step S2, it is identified that S a (n) belongs to the AM modulation type, and in step S3, it is identified that S a (n) belongs to the {AM, AM broadcast, ASK, PAM} modulation type, and the intersection is taken (the same reasoning applies), then the recognition result of the signal modulation type is AM.

[0062] S5, according to the classification results of steps S2 and S3, if the signal S a (n) belongs to the FM modulation type, the recognition result of the signal modulation type is FM.

[0063] S6, according to the classification results of steps S2 and S3, if the signal S a (n) belongs to the {AM broadcast, ASK} modulation type, the baseband signal features and constellation features of the signal S a (n) are extracted, and the modulation type recognition result is obtained by combining the KAN-LSTM-based signal recognition network and the MobileNet-based signal recognition network.

[0064] Step S2 in identifying S a (n) belongs to {AM broadcast, ASK}, Step S3 in identifying S a (n) belongs to {AM, AM broadcast, ASK, PAM}, and the intersection of the two identification results is taken, then the signal S a (n) belongs to {AM broadcast, ASK} modulation type. At this time, Step S6 includes:

[0065] S61, carrier residual estimation is performed on the signal S a (n) to obtain a carrier residual estimation value f a , and according to the carrier residual estimation value f a , the carrier residual of the signal S a (n) is removed to obtain the signal S fre-no (n).

[0066] The carrier residual estimation process includes:

[0067] S611, a fast Fourier transform and a modulo value processing are performed on the 4th power of the signal S a (n) to obtain a frequency feature F a (k), and the specific calculation process is as follows:

[0068]

[0069] S612, according to the position of the peak value of the F a (k) spectrum line, the carrier residual value f a is estimated, and the specific calculation process is as follows:

[0070] index = argmax F a (k),

[0071]

[0072] wherein f s is the sampling rate of the signal S a (n).

[0073] According to the carrier residual estimation value f a , the carrier residual of the signal S a (n) is removed to obtain the signal S fre-no (n) includes:

[0074] S62, according to the I / Q sequence of the signal S fre-no (n), a signal recognition network based on KAN-LSTM is used to obtain a preliminary identification type in {AM broadcast, ASK}, and the preliminary identification type is AM broadcast or {2ASK, 4ASK, 8ASK}.

[0075] The network structure of the signal recognition network based on KAN-LSTM here is the same as that in the above step S3, except that the dimension of the output is changed to 2, and the corresponding label change is AM broadcast and {2ASK, 4ASK, 8ASK}. In addition, corresponding to the change of the predicted label, the training sample also changes to the I / Q sequence of the signal and the two types of labels of AM broadcast, {2ASK, 4ASK, 8ASK}.

[0076] S63, if the signal S fre-no (n) belongs to the AM broadcast modulation type, the output signal modulation type recognition result is AM.

[0077] S64, if the signal S fre-no (n) belongs to {2ASK, 4ASK, 8ASK} modulation type, the symbol rate estimation is performed on the signal S fre-no (n), and the symbol rate estimation value R a of the signal is obtained; according to the symbol rate estimation value R a , the sampling point number spsi corresponding to each symbol is calculated; according to the sampling point number spsi of the symbol, the timing error of the signal S fre-no (n) is calculated by using the O&M square estimation algorithm. Further, the constellation diagram of the signal S fre-no (n) is calculated.

[0078] The symbol rate estimation process is as follows:

[0079] S641, the rate signal S a (n) of the signal S r (n) is calculated, and the specific calculation process is as follows:

[0080] S r (n) = |S fre-no (n+1) - S fre-no (n) | * f s , n = 0, 1, 2,..., N-2,

[0081] S642, the rate signal S r (n) is subjected to fast Fourier transform and the modulus value is taken, and the specific calculation process is as follows:

[0082]

[0083] S643, according to the position where the F r (k) spectrum line peak value is located, the symbol rate estimation value R a is calculated, and the specific calculation process is as follows:

[0084]

[0085] sps1 and timing error The calculation process includes:

[0086] S644, the number of sampling points sps1 of the symbol is calculated by:

[0087]

[0088] represents the rounding up operation.

[0089] Calculate the signal S fre-no (n) timing error The flow includes:

[0090]

[0091] Wherein, arg{·} represents the phase operation, is the estimated timing error.

[0092] Calculate the constellation of signal S a (n)

[0093] S645, according to the number of sampling points N of signal S(n), the number of sampling points sps1 of the symbol and the timing error Calculate the down-sampling position vector:

[0094]

[0095] Wherein, x(m) represents the mth down-sampling position, and all down-sampling positions are arranged in sequence to form a down-sampling position vector, represents the rounding down operation;

[0096] According to the down-sampling position vector, the signal S fre-no (n) is down-sampled to obtain the down-sampling vector y1;

[0097] According to the down-sampling vector y1, the constellation of signal S a (n) is drawn.

[0098] S65, according to the constellation of signal S fre-no (n), using the signal recognition network based on MobileNet, the recognition result of modulation type in 2ASK, 4ASK, 8ASK is obtained.

[0099] The structure of the signal recognition network based on MobileNet is the same as that in step S2 described above, except that the input features and the predicted labels are changed. The input features become the constellation diagram of the signal, and the predicted labels become 2ASK, 4ASK, and 8ASK in three dimensions. Correspondingly, the training sample pair also changes.

[0100] S7, according to the classification results of steps S2 and S3, if the signal S a (n) belongs to the PAM modulation type, the multi-time constellation features of the signal S a (n) are extracted, and a multi-modal feature fusion network is combined to obtain the recognition result of the signal modulation type.

[0101] Reference Figure 4 , the present step includes, in some embodiments:

[0102] S71, carrier residual estimation is performed on the signal S a (n) to obtain a carrier residual estimation value f a , and according to the carrier residual estimation value f a , the carrier residual of the signal S a (n) is removed to obtain the signal S fre-no (n).

[0103] The carrier residual estimation process and the carrier residual removal process are the same as those described in step S61 above, and will not be described here again.

[0104] S72, symbol rate estimation is performed on the signal S fre-no (n) to obtain a symbol rate estimation value R fre-no (n) of the signal S a ; according to the symbol rate estimation value R a , the number of sampling points spsi corresponding to each symbol is calculated; and according to the number of sampling points spsi, the signal S fre-no (n) is down-sampled multiple times to extract the multi-time constellation of the signal S a (n).

[0105] The symbol rate estimation and the calculation process of the symbol sampling points are the same as those described in step S64 above, and will not be described here again. Here, the method for extracting the multi-time constellation of the signal S a (n) is mainly described, which includes:

[0106] According to the number of sampling points N of the signal S(n) and the number of sampling points spsi of the symbol, the down-sampling position matrix is calculated:

[0107] x(n, m) = n + spsi·m; n = 0, 1, 2, …, spsi-1;

[0108] x(n,m) represents the m-th downsampling position in the n-th sample group;

[0109] Based on the downsampling position matrix, the signal S fre-no (n) performs downsampling, that is, downsampling the signal S fre-no (n) Perform sps1 group downsampling to obtain sps1 group downsampling values. Each group has m downsampling points, and finally obtain the downsampling matrix y2.

[0110] Plot the signal S based on the downsampling matrix y2. a (n) is a multi-time constellation chart, i.e., sps1 constellation chart.

[0111] S73, according to signal S a The multi-temporal constellation diagram (n) is used to obtain the modulation type recognition results in 2PAM, 4PAM and 8APM by using a multimodal feature fusion network.

[0112] The multimodal feature fusion network consists of a cascaded feature extraction module, a feature fusion module, and a third classification module.

[0113] The feature extraction module includes, in sequence, a convolution module, a depthwise separable convolution module a, a depthwise separable convolution module b, a depthwise separable convolution module c, and a two-dimensional average pooling layer feature unrolling function;

[0114] The feature fusion module includes, in sequence, a feature concatenation function, a KAN network layer a, an activation function GELU, a KAN network layer b, an activation function Sigmoid, and a feature multiplication function;

[0115] The third classification module includes, in sequence, a feature expansion function, a KAN network layer, and an activation function Softmax.

[0116] Since the designed multimodal feature fusion network supports a maximum of 15 constellation diagrams as input, the number of sampling points of the extracted symbols in step S72, sps1 ≤ 15. In addition, if the number of constellation diagram inputs is less than 15, the insufficient part of the input is 0 by default.

[0117] S8. Based on the classification results of steps S2 and S3, if signal S a If (n) belongs to the {CPFSK, FM broadcast} modulation type, then extract the signal S. a The frequency information, multi-time constellation diagram and spectral features of (n) are combined with a signal recognition network based on KAN-LSTM, a multimodal feature fusion network and a signal recognition network based on MobileNet to obtain the signal modulation type recognition result.

[0118] In some embodiments, step S8 includes:

[0119] S81, for signal Sa Phase of (n) Perform differential calculations to extract signal S a The frequency information f(n) of (n).

[0120] Signal S a The process of extracting the frequency information f(n) of (n) includes:

[0121] Extract signal S a Phase of (n)

[0122]

[0123] phase Perform differential calculations to obtain signal S a Frequency information f(n) of (n):

[0124]

[0125] S82, according to signal S a The frequency information f(n) is used to obtain the initial identification type of modulation type in {CPFSK, FM broadcast} using a signal recognition network based on KAN-LSTM. The initial identification type is either FM broadcast or two categories: {MSK, 2CPFSK, 4CPFSK, 8CPFSK, 16CPFSK}.

[0126] The signal recognition network based on KAN-LSTM here has the same structure as in step S3 above. The difference is that the predicted label is changed to either FM broadcast or two categories: {MSK, 2CPFSK, 4CPFSK, 8CPFSK, 16CPFSK}. Correspondingly, the sample labels of the training samples are also changed to either FM broadcast or two categories: {MSK, 2CPFSK, 4CPFSK, 8CPFSK, 16CPFSK}.

[0127] S83. If the frequency information f(n) is identified as FM broadcast modulation type, then the output signal modulation type identification result is FM.

[0128] S84. If the frequency information f(n) is identified as belonging to the modulation type {MSK, 2CPFSK, 4CPFSK, 8CPFSK, 16CPFSK}, then the symbol rate of the frequency information f(n) is estimated to obtain the symbol rate estimate R of f(n). b Based on the symbol rate estimate R b Calculate the number of sampling points sps2 corresponding to each symbol; based on the number of sampling points sps2 of the symbol, downsample the frequency information f(n) multiple times to extract the signal S. a (n) is a multi-time constellation diagram.

[0129] The method of symbol rate estimation and the method of symbol sample point number calculation are the same as the aforementioned step S64, and the only difference is that the estimation object is changed from the signal S fre-no (n) to the frequency information f(n), so the specific calculation process is not described in detail here.

[0130] The extraction method of the multi-time constellation diagram is the same as the aforementioned step S72, and the only difference is that the extraction object is changed from the signal S fre-no (n) to the frequency information f(n).

[0131] S85, according to the multi-time constellation diagram of the frequency information f(n), the modulation type is identified in {MSK, 2CPFSK}, 4CPFSK, 8CPFSK or 16CPFSK using a multi-modal feature fusion network.

[0132] The network structure of the multi-modal feature fusion network here is the same as that of step S73, and the only difference is that the predicted label is changed from 2PAM, 4PAM, 8APM to {MSK, 2CPFSK}, 4CPFSK, 8CPFSK, 16CPFSK. The label of the corresponding training sample is also changed to {MSK, 2CPFSK}, 4CPFSK, 8CPFSK, 16CPFSK.

[0133] S86, if the frequency information f(n) is identified as belonging to the {MSK, 2CPFSK} modulation type, the square of the signal S a (n) is subjected to Fourier transform and the modulus value is taken to obtain the signal S a (n) The second spectral feature F b (k) is calculated.

[0134] The second spectral feature F b (k) is calculated as follows:

[0135]

[0136] S87, according to the second spectral feature F b (k), a binary classification result of the modulation type is identified using a MobileNet-based signal recognition network.

[0137] The network structure of the MobileNet-based signal recognition network here is the same as that in step S2, and the only difference is that the predicted label is changed from AM, FM, {AM broadcast, ASK}, {CPFSK, FM broadcast, PAM} to MSK, 2CPFSK. Correspondingly, the label of the training sample is also changed to MSK, 2CPFSK.

[0138] Embodiment 2

[0139] This embodiment focuses on 13 common communication signal modulation types, such as amplitude modulation (AM), frequency modulation (FM), 2-ary amplitude shift keying (2ASK), 2-ary pulse amplitude modulation (2PAM), 2-ary continuous phase frequency shift keying (2CPFSK), etc., to experimentally verify the performance of the method of embodiment 1.

[0140] All received signals in this embodiment are collected by the ThinkSpectrum 3943B spectrum monitoring receiver with an antenna. The signals are generated by Agilent E8267D and R&S SMW200A vector signal generators and transmitted through the antenna. In addition, in order to fully consider the real communication scenario, the ThinkSpectrum 3943B spectrum monitoring receiver collects AM and FM broadcast signals at different frequencies. When collecting signals, the ThinkSpectrum 3943B spectrum monitoring receiver needs to select an appropriate demodulation bandwidth according to the signal bandwidth. The demodulation bandwidth has a certain ratio relationship with the sampling rate, as shown in Table 1.

[0141] Table 1 Corresponding relationship between demodulation bandwidth and sampling rate

[0142] Demodulation bandwidth Sampling rate 300 KHz 1.4 MHz 500 KHz 2 MHz 800 KHz 3.5 MHz 1000 KHz 4 MHz 1250 KHz 5.6 MHz 1500 KHz 7 MHz 2000 KHz 7 MHz 5000 KHz 14 MHz 8000 KHz 28 MHz 10000 KHz 28 MHz 12500 KHZ 28 MHz 15000 KHZ 28 MHz 20000 KHZ 28 MHz

[0143] When constructing the training data set, considering the actual scenario, the receiving end only needs to ensure that the signal is within the demodulation bandwidth to complete signal reception, and the carrier residual of the received signal fluctuates in a large range. Therefore, when generating the training data set, the carrier frequencies of the transmitting end and the receiving end are aligned, and are affected by the local oscillator error. At this time, the collected signal has a very small carrier residual. Further, by digital signal processing, a carrier residual is added to the received signal. The numerical range of the carrier residual is determined by the signal bandwidth and the demodulation bandwidth of the receiving end, and the signal can be ensured to be within the demodulation bandwidth. The parameters of the signal transmitted by the signal source are shown in Table 2. In addition, the AM broadcast is collected at 15195KHz, 13919KHz, 13655KHz, and 13610KHz, and the FM broadcast is collected at 107.8MHz, 106.6MHz, 106.1MHz, 105.5MHz, 104.9MHz, 104.3MHz, 103.5MHz, 101.8MHz, and 101.1MHz. Among them, due to the collection of broadcast signals, there is a certain error in the carrier frequency.

[0144] When constructing the test data set, considering the actual scenario, as long as the signal is within the demodulation bandwidth at the receiving end, the carrier residual fluctuates in a large range.

[0145] Table 2 Parameters of communication signal modulation types

[0146]

[0147] According to a specific embodiment of the present invention, the present invention only requires knowledge of the receiving end sampling rate during identification, without any other prior knowledge, and can identify 13 communication signal modulation types. Among them, AM and FM broadcasts are identified as AM and FM modulation respectively, and are not treated as separate modulation types. The number of signal sampling points N = 8192 used in the present invention to identify the signal modulation type.

[0148] In the experimental verification of the signal recognition network based on MobileNet, signal S was extracted. a (n) When dealing with the spectrum, the number of signal sampling points used is N = 8192, and the image dimension of the spectrum graph input to the MobileNet-based signal recognition network is 64 × 64 × 3. The network parameters of the MobileNet-based signal recognition network are shown in Table 3, where the output dimension L = 4 of the linear layer and the Softmax activation function represents the probability that the recognition result may be AM, FM, {AM broadcast, ASK} and {CPFSK, FM broadcast, PAM}.

[0149] Table 3. Signal recognition network parameters based on MobileNet

[0150]

[0151]

[0152] In the experimental verification of the signal recognition network based on KAN-LSTM, the signal S with N=2048 sampling points was used. a The I / Q sequences of (n) are concatenated according to the channel dimension and input into the network. The I / Q sequence dimension of the input signal recognition based on KAN-LSTM is 2×2048. The network parameters of the signal recognition network based on KAN-LSTM are shown in Table 4, where the output dimension L=2 of the KAN network layer b and the activation function Softmax represents the probability that the recognition result may be {AM, AM broadcast, ASK, PAM} and {FM, FM broadcast, CPFSK}.

[0153] Table 4. Parameters of the signal recognition network based on KAN-LSTM

[0154]

[0155]

[0156] Considering the modulation type identification results of the two networks above, the comprehensive identification results are AM, FM, {AM broadcast, ASK}, PAM and {CPFSK, FM broadcast}.

[0157] If the signal S a If (n) belongs to AM modulation, then the output signal modulation type identification result is AM.

[0158] If the signal S a (n) belongs to FM modulation, the identification result of the modulation type of the output signal is FM.

[0159] If the signal S a (n) belongs to {AM broadcast, ASK} modulation, carrier residual estimation and removal are performed on the signal S a (n) to obtain a signal S fre-no (n), and the number of sampling points of the signal used is N = 8192; further, the I / Q sequence of the signal S fre-no (n) is sent to a KAN-LSTM-based signal identification network, and the dimension of the I / Q sequence input into the network is 2x2048. The network parameters of the KAN-LSTM-based signal identification network used are shown in Table 4, wherein the output dimension L of the linear layer and the activation function Softmax is 2, representing the probability that the identification result may be AM broadcast, {2ASK, 4ASK, 8ASK}.

[0160] If the signal S fre-no (n) belongs to {2ASK, 4ASK, 8ASK} modulation, symbol rate estimation and timing error estimation are performed on the signal S fre-no (n), and the number of sampling points of the signal used is N = 8192; further, the constellation diagram is sent to a MobileNet-based signal identification network, and the image dimension of the constellation diagram input into the MobileNet is 64x64x3. The MobileNet network parameters used are similar to those in Table 3, except that the output dimension L of the linear layer and the activation function Softmax is 3, representing the probability that the identification result may be 2ASK, 4ASK, 8ASK.

[0161] If the signal S fre-no (n) belongs to AM broadcast modulation, the identification result of the modulation type of the output signal is AM.

[0162] If the signal S a (n) belongs to PAM modulation, carrier residual estimation and symbol rate estimation are performed on the signal S a (n), and the number of sampling points of the signal used is N = 8192; further, the multi-time constellation features of the signal are extracted, and the multi-time constellation is sent to a multi-modal feature fusion network. The multi-time constellation is 15 constellation diagrams, and the image dimension of each constellation diagram is 64x64x3. If sps<15, i.e., the number of constellation diagrams is less than 15, the default is that the insufficient constellation diagrams are all 0. The MobileNet-Sttention network parameters used are shown in Table 5, wherein the output dimension L of the KAN network layer and the activation function Softmax is 3, representing the probability that the identification result may be 2PAM, 4PAM, 8APM.

[0163] Table 5 Modulation type recognition network parameters based on MobileNet-Attention

[0164]

[0165]

[0166] If the signal S a (n) belongs to {CPFSK, FM broadcast} modulation, the frequency information feature of the signal is extracted, the number of sampling points of the signal used is N = 8192, and the number of sampling points of the frequency information f(n) is N-1; further, f(n) is sent into the signal recognition network based on KAN-LSTM, the dimension of input f(n) is 1x2048. The KAN-LSTM network parameters used are similar to Table 44, wherein the input channel number of the one-dimensional convolution layer in the spatial feature extraction module is 1, in addition, the output dimension L = 2 of the KAN network layer b and the activation function Softmax represents the probability that the recognition result may be FM broadcast, {MSK, 2CPFSK, 4CPFSK, 8CPFSK, 16CPFSK}.

[0167] If f(n) belongs to FM broadcast modulation, the output signal modulation type recognition result is FM.

[0168] If f(n) belongs to {MSK, 2CPFSK, 4CPFSK, 8CPFSK, 16CPFSK} modulation, symbol rate estimation is performed on f(n), and the number of sampling points used is N-1 = 8191; further, the multi-time constellation features of the signal are extracted, and the multi-time constellation is sent into the multi-modal feature fusion network. Among them, the multi-time constellation is 15 constellation diagrams, and the image dimension of each constellation diagram is 64x64x3, if sps<15, i.e. the number of constellation diagrams is less than 15, then the insufficient constellation diagrams are all 0 by default. The network parameters of the multi-modal feature fusion network used are similar to Table 5, wherein the output dimension L = 4 of the KAN network layer and the activation function Softmax represents the probability that the recognition result may be {MSK, 2CPFSK}, 4CPFSK, 8CPFSK, 16CPFSK.

[0169] If f(n) belongs to {MSK, 2CPFSK} modulation, the 2nd power of the signal S a (n) is subjected to Fourier transform, and the number of sampling points of the signal used is N = 8192; further, the frequency spectrum F b(k) into the MobileNet-based signal recognition network, the image dimension of the spectrum graph input into the MobileNet-based signal recognition network is 64x64x3. The network parameters of the MobileNet-based signal recognition network used are similar to those in Table 3, except that the output dimension L=3 of the linear layer and the activation function Softmax represents the probability that the recognition result can be MSK, 2CPFSK.

[0170] Experimental content and results

[0171] Experiment 1: from Figure 5 It can be seen that the spectrum graphs of 15 signals in turn are AM, FM, etc. It is not difficult to find that the spectrum of AM has three peak lines, the spectrum of AM broadcast and ASK modulation has one peak line, and the spectrum of FM, FM broadcast and CPFSK is a very wide band, among which the spectrum of FM presents low in the middle and high on both sides within the signal bandwidth.

[0172] Experiment 2: from Figure 6 It can be seen that the real part of S fre-no (n) of AM broadcast, ASK modulation and carrier removal signal in turn, it is not difficult to find that the signal frequency corresponding to AM broadcast has changed, while the signal frequency corresponding to ASK modulation remains unchanged.

[0173] Experiment 3: from Figure 7 It can be seen that the constellation graphs of 2ASK, 4ASK and 8ASK in turn, it is not difficult to find that the constellation graphs of 2ASK, 4ASK and 8ASK are 2, 4 and 8 points respectively.

[0174] Experiment 4: from Figure 8 It can be seen that the multi-time constellation graphs of 2PAM, 4PAM and 8PAM in turn, it is not difficult to find that at a certain time, the characteristics of the constellation graph are very obvious corresponding to the best sampling time, in addition, the constellation graphs of 2PAM, 4PAM and 8PAM are very different.

[0175] Experiment 5: from Figure 9 It can be seen that the frequency information of FM broadcast and CPFSK in turn, it is not difficult to find that the signal frequency corresponding to FM broadcast has changed, while the signal frequency corresponding to CPFSK modulation remains unchanged.

[0176] Experiment 5: from Figure 10 It can be seen that the multi-time constellation graphs of MSK, 2CPFSK, 4CPFSK, 8CPFSK and 16CPFSK in turn, it is not difficult to find that at a certain time, the characteristics of the constellation graph are very obvious corresponding to the best sampling time, in addition, the constellation graphs of {MSK, 2CPFSK}, 4CPFSK, 8CPFSK and 16CPFSK are very different.

[0177] Experiment 6: from Figure 11It can be seen that the frequency spectrum diagram of the square of the signal of MSK and 2CPFSK in turn, it is not difficult to see that the frequency spectrum corresponding to MSK has two obvious spectral peaks.

[0178] Experiment 7: from Figure 12 It can be seen that when the signal-to-noise ratio is 20dB, the overall accuracy of signal recognition is more than 98%. To verify the effectiveness of the algorithm, two comparison algorithms are selected. Algorithm 1 is a modulation recognition method based on convolutional neural network, which only extracts the time domain features of the signal. It can be seen that the accuracy of signal recognition is poor. Algorithm 2 is a modulation recognition method based on convolutional neural network and long short-term memory network, which further captures the time sequence features of the signal. The accuracy of signal recognition is greatly improved, but the accuracy of signal recognition still needs to be improved.

[0179] Experiment 8: from Figure 13 As shown in the confusion matrix diagram of the recognition of 13 common signal modulation types in the embodiment, it is not difficult to find that when the signal-to-noise ratio is 20dB, the recognition accuracy of AM, FM, 2ASK, 2PAM and 2CPFSK can reach more than 99%; the recognition accuracy of QPSK and 128QAM can reach more than 97%; the recognition accuracy of 4ASK, 4PAM and MSK can reach more than 98%; the recognition accuracy of 8ASK, 8PAM, 4CPFSK and 8CPFSK can reach more than 97%; and the recognition accuracy of 16CPFSK can reach more than 96%.

[0180] The present application is not limited to the foregoing specific embodiments. The present application extends to any novel one, or any new combination, of the characteristics disclosed in this specification, as well as to any novel method or process disclosed in any of the steps or any new combination.

Claims

1.A method for intelligent identification of communication signal modulation type for industrial internet, characterized in that, The application relates to a signal classification method and device. S1, normalizing power of the received signal to obtain a unit power signal ; S2, extract signal First spectral feature The first spectral features are analyzed using a preliminary signal classification model trained by a MobileNet-based signal recognition network. Perform preliminary classification and obtain signals Preliminary classification results in AM, FM, {AM broadcast, ASK} and {CPFSK, FM broadcast, PAM}; The construction method of the signal preliminary classification model is as follows: unit power signals of historical received signals are subjected to fast Fourier transform and modulus value processing to extract frequency spectrum characteristics, a MobileNet-based signal recognition network is subjected to supervised training by using the extracted frequency spectrum characteristics, and a signal preliminary classification model based on frequency spectrum characteristics and having classification results of AM, FM, {AM broadcast, ASK} and {CPFSK, FM broadcast, PAM} is constructed; The MobileNet-based signal recognition network is composed of cascaded convolution modules, a depth separable convolution module a, a depth separable convolution module b, a depth separable convolution module c and a first classification module. The convolution module sequentially comprises a two-dimensional convolution layer, a two-dimensional batch normalization layer and an activation function ReLU. The depth separable convolution module a sequentially comprises a two-dimensional convolution layer a, a two-dimensional batch normalization layer a, an activation function ReLU, a two-dimensional convolution layer b, a two-dimensional batch normalization layer b and an activation function ReLU. The depth separable convolution module b sequentially comprises a two-dimensional convolution layer a, a two-dimensional batch normalization layer a, an activation function ReLU, a two-dimensional convolution layer b, a two-dimensional batch normalization layer b and an activation function ReLU. The depth separable convolution module c sequentially comprises a two-dimensional convolution layer a, a two-dimensional batch normalization layer a, an activation function ReLU, a two-dimensional convolution layer b, a two-dimensional batch normalization layer b and an activation function ReLU. The first classification module sequentially comprises a two-dimensional average pooling layer, a feature unfolding function, a linear layer and a Softmax function. S3, distinguishing the signal from the I / Q sequence of the signal using a trained KAN-LSTM based signal recognition network is {AM, AM broadcast, ASK, PAM} or {FM, FM broadcast, CPFSK}; The KAN-LSTM-based signal recognition network is trained by using training samples, the training samples are I / Q sequences of historical received signals, and labels are two categories of {AM, AM broadcast, ASK, PAM} or {FM, FM broadcast, CPFSK}. The KAN-LSTM-based signal recognition network is composed of cascaded spatial feature extraction module a, spatial feature extraction module b, time sequence feature extraction module a, time sequence feature extraction module b and a second classification module. The spatial feature extraction module a sequentially comprises a one-dimensional convolution layer, a one-dimensional average pooling layer, a KAN network layer, a one-dimensional batch normalization layer and an activation function ReLU. The spatial feature extraction module b sequentially comprises a one-dimensional convolution layer, a one-dimensional average pooling layer, a KAN network layer, a one-dimensional batch normalization layer and an activation function ReLU. The time sequence feature extraction module a sequentially comprises an LSTM, an activation function ReLU, a one-dimensional batch normalization layer, a KAN network layer and an activation function GELU. The time sequence feature extraction module b sequentially comprises an LSTM, an activation function ReLU, a one-dimensional batch normalization layer, a KAN network layer and an activation function GELU. The second classification module sequentially comprises a KAN network layer a, a feature unfolding function, a one-dimensional average pooling layer, a KAN network layer b and an activation function Softmax. S4. Determining a signal of the intersection of the classification results according to steps S2 and S3 The modulation type of the signal; if the signal belongs to the AM modulation type, output the identification result of the signal modulation type as AM; S5、if the signal belongs to the FM modulation type, the identification result of the modulation type of the output signal is FM; S6. If the signal belongs to the {AM broadcast, ASK} modulation type, then perform: S61, to the signal a carrier residual estimation is performed to obtain a carrier residual estimation value , and according to the carrier residual estimation value , the carrier residual of the signal is removed to obtain a signal ; S62、According to the I / Q sequence of the signal , using the KAN-LSTM-based signal recognition network, the preliminary identification type of the adjustment type in {AM broadcast, ASK} is obtained, and the preliminary identification type is AM broadcast or {2ASK, 4ASK, 8ASK}. S63, if the signal belongs to the AM broadcast modulation type, the identification result of the signal modulation type is AM; S64, if the signal If it belongs to the {2ASK, 4ASK, 8ASK} modulation type, then for the signal Perform symbol rate estimation to obtain the symbol rate estimate of the signal. Based on symbol rate estimation Calculate the number of sampling points corresponding to each symbol. Based on the number of sampling points of the code element The signal is calculated using the O&M squared estimation algorithm. Timing error Further calculation of the signal A constellation chart; S65, according to the constellation diagram of the signal the recognition result of the modulation type in 2ASK, 4ASK, 8ASK is obtained by using a signal recognition network based on MobileNet. S7. If the signal belongs to the PAM modulation type, then perform: S71, to the signal performing carrier residual estimation to obtain a carrier residual estimation value , and removing the carrier residual of the signal according to the carrier residual estimation value to obtain a signal ; S72, the signal S72, the signal S72, the signal S72, the signal S72, the signal S72, the signal S72, the signal S72, the signal S72, the signal S73, according to the signal of the multi-time constellation, using a multi-modal feature fusion network, obtaining the recognition result of the modulation type in 2PAM, 4PAM, 8APM; The multi-modal feature fusion network is composed of cascaded feature extraction modules, a feature fusion module and a third classification module. The feature extraction module sequentially comprises a convolution module, a depth separable convolution module a, a depth separable convolution module b, a depth separable convolution module c and a two-dimensional average pooling layer feature unfolding function; The feature fusion module sequentially comprises a feature splicing function, a KAN network layer a, an activation function GELU, a KAN network layer b, an activation function Sigmoid and a feature multiplication function; The third classification module sequentially comprises a feature unfolding function, a KAN network layer and an activation function Softmax; S8. If the signal belongs to the modulation type {CPFSK, FM broadcast}, then perform: S81, phase of the signal is calculated difference, extracting frequency information of the signal ;​ S82, according to the signal frequency information , using a KAN-LSTM-based signal recognition network, obtaining a preliminary recognition type of the modulation type in {CPFSK, FM broadcast}, the preliminary recognition type being FM broadcast or {MSK, 2CPFSK, 4CPFSK, 8CPFSK, 16CPFSK}. S83、if the identified frequency information belongs to the FM broadcast modulation type, then the identification result of the signal modulation type is FM; S84, If frequency information is identified If it belongs to the {MSK, 2CPFSK, 4CPFSK, 8CPFSK, 16CPFSK} modulation type, then the frequency information... Perform symbol rate estimation to obtain symbol rate estimate Based on symbol rate estimation Calculate the number of sampling points corresponding to each symbol. Based on the number of sampling points of the code element Frequency information Perform multiple downsampling to extract the signal Multi-time constellation chart; S85、According to the frequency information The multi-mode feature fusion network identifies the modulation type from among {MSK, 2CPFSK}, 4CPFSK, 8CPFSK, or 16CPFSK. S86、if the identification frequency information belongs to {MSK, 2CPFSK} modulation type, then the Fourier transform of the square of the signal is taken and the modulus value is obtained, to get the second spectral feature of the signal ; S87、According to the second spectral feature A binary classification result of the modulation type is identified by using a MobileNet-based signal recognition network. 2.The industrial internet of things oriented communication signal modulation type intelligent identification method according to claim 1, wherein, Computing a constellation of signals The method of computing a constellation of signals comprises: According to the signal Number of sampling points Number of sampling points of the code element and timing error Calculate the downsampling position vector; down-sampling a signal according to a down-sampling position vector to obtain a down-sampled vector ; According to the down-sampling vector Plot the constellation diagram of the signal . 3.The industrial internet of things oriented communication signal modulation type intelligent identification method of claim 1, wherein, The number of sampling points of the symbol calculated in step S72 . 4.The industrial internet of things oriented communication signal modulation type intelligent identification method according to claim 1 or 3, characterized in that, In step S72, the signal The method of multi-time constellation diagram includes: According to the signal Number of sampling points Number of sampling points of the symbol Calculate the downsampling position matrix; down-sampling the signal according to the down-sampling position matrix to obtain a down-sampled matrix ; According to the down-sampling matrix Plot the signal of the multi-time constellation. 5.The industrial internet of things oriented communication signal modulation type intelligent identification method according to claim 1, wherein, The step S81 comprises: extracting the signal the phase of : ； on the phase :​​ 。

Images