Modulation identification method and device based on multi-stream space-time network under low signal-to-noise ratio
By using EMD and EEMD algorithms to denoise signals in multi-stream spatiotemporal networks, and combining with RML2016.10a dataset training model, the problem of insufficient signal recognition accuracy under low signal-to-noise ratio is solved, and a higher recognition rate is achieved and precision.
Patent Information
- Application Number
- CN202411247332.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art signal recognition accuracy is not high enough under low signal-to-noise ratio conditions, especially in complex communication environments, and it is difficult for deep learning to effectively identify signal characteristics.
Multi-stream spatiotemporal networks are used to denoise the original signal by combining empirical modal decomposition (EMD) and ensemble empirical modal decomposition (EEMD) algorithms, and generate denoised signal data sets, and combine them with RML2016.10a data sets to divide them into training sets , test set and verification set, build a multi-stream spatiotemporal network model for training and verification, and is finally used for signal recognition.
Through improved denoising methods and multi-stream spatiotemporal networks, the signal recognition accuracy under low signal-to-noise ratio conditions is improved, and the robustness and generalization ability of the model are improved. The experimental results show that the recognition rate and accuracy have been significantly improved.
Smart Images

Figure CN119030839A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of signal modulation recognition, and in particular to a modulation recognition method and device based on a multi-stream space-time network under low signal-to-noise ratio. Background Art
[0002] In the field of wireless communications, signal analysis and processing technology mainly involves a series of complex processes such as signal detection, recognition and recovery under the condition that the channel and signal parameters are not completely known. The research on modulation recognition has roughly gone through three stages, namely modulation recognition research based on likelihood theory, feature extraction and deep learning. The modulation recognition research based on likelihood theory is divided into average likelihood ratio test method, generalized likelihood ratio test method and mixed likelihood ratio detection method. The modulation recognition method based on likelihood theory requires more prior knowledge and large amount of calculation, while the recognition method based on feature extraction has lower computational complexity, and many literatures start from this point for research. The modulation method based on feature extraction requires manual feature extraction, which depends largely on the feature extraction method of experts and is difficult to apply to the current complex communication environment. With the continuous deepening of artificial intelligence, modulation recognition based on deep learning has emerged.
[0003] Modulation recognition algorithms based on deep learning can be roughly divided into two categories. One is to use the deep learning network model to directly classify the original time domain signal; the other is to preprocess the signal and then input it into the deep learning network model. Deep learning can automatically extract and classify the features of the signal, replacing the step of relying on expert experience to identify features in the traditional recognition algorithm, and the recognition speed is fast and the accuracy is high.
[0004] The algorithms in the prior art have achieved certain results in the field of automatic modulation classification, but the average network recognition problem under low signal-to-noise ratio conditions still has the problem of insufficient recognition accuracy. Summary of the invention
[0005] In view of the above problems, an object of the present invention is to provide a modulation recognition method based on a multi-stream spatiotemporal network under low signal-to-noise ratio.
[0006] The second object of the present invention is to provide a modulation recognition device based on a multi-stream spatiotemporal network under low signal-to-noise ratio.
[0007] The first technical solution adopted by the present invention is: S1, denoising the acquired original signal using the EMD algorithm and the EEMD algorithm respectively to obtain two denoised signal data sets;
[0008] S2, combining the two denoised signal data sets with the RML2016.10a data set respectively to obtain two synthetic data sets RML2016.10a-01 and RML2016.10a-02, and dividing the multiple synthetic data sets into multiple training sets, test sets and validation sets;
[0009] S3, constructing a multi-stream spatiotemporal network model, and respectively using the multiple training sets and verification sets to train and verify the multi-stream spatiotemporal network model to obtain a trained model;
[0010] S4. Input the synthetic data set into the trained model to obtain a signal recognition result.
[0011] Optionally, the denoising by using the EMD algorithm in S1 includes:
[0012] S11A, determining the maximum and minimum points of the original signal, connecting them with a spline curve for data processing, and obtaining a maximum envelope and a minimum envelope;
[0013] S12A, calculating based on the maximum envelope and the minimum envelope to obtain a first average envelope;
[0014] S13A, using the first average envelope to calculate the residual signal function, update the original signal and return to S11A, until the updated residual signal satisfies the IMF signal, thereby obtaining the IMF signal;
[0015] S14A, calculating a residual signal value based on the IMF signal, and performing signal component analysis on the IMF signal to obtain an IMF component set;
[0016] S15A, calculating the mean square errors of all adjacent IMF components based on the IMF component set for comparison, and determining the value K of the number of IMF components according to the minimum mean square error;
[0017] S16A, reconstructing the EMD denoised signal based on the IMF component number value K.
[0018] Optionally, the denoising by using the EEMD algorithm in S1 includes:
[0019] S11 B, adding white noise to the original signal, and performing EMD decomposition on the signal after the white noise is included to obtain an IMF component;
[0020] S12B, averaging the obtained IMF components to obtain a component average;
[0021] S13B, calculating and updating a signal residual based on the component average value;
[0022] S14B, repeat S11 B-S13B, calculate the mean square errors of all adjacent IMF components for comparison, and determine the value K of the number of IMF components according to the minimum mean square error;
[0023] S15B, after several iterations, reconstruct the signal using the IMF component number value K to obtain the EEMD denoised signal.
[0024] Optionally, the step of training the multi-stream spatiotemporal network using the training set in S3 includes:
[0025] S31, inputting the signal in the training set into the constructed multi-stream spatiotemporal network to obtain a predicted category label;
[0026] S32, calculating the real category label and the predicted category label of the signal in the training set through a cross entropy loss function to obtain a loss value and update the network parameters, wherein the cross entropy loss function is as follows:
[0027] Loss=-Σt i log(p i )
[0028] In the formula, Loss is the value of the cross entropy loss function, t i and p i is the true label of the class and the predicted probability of the model for the class;
[0029] S33, repeat S31-S32, and complete the model training after several iterations.
[0030] Optionally, in S16A, the EMD denoising signal is constructed as follows:
[0031]
[0032] Where IMF[i] is the ith IMF signal, and k is the minimum mean square error.
[0033] Optionally, in S15B, the EEMD denoised signal is reconstructed according to the following formula:
[0034]
[0035] Where IMFK is the number of IMF components, rK is the signal residual after K iterations, and k is the minimum mean square error.
[0036] The second technical solution adopted by the present invention is: a denoising module, used to denoise the acquired original signal using the EMD algorithm and the EEMD algorithm respectively, to obtain two denoised signal data sets;
[0037] A data collection module, used for combining the two denoised signal data sets with the RML2016.10a data set respectively to obtain two synthetic data sets RML2016.10a-01 and RML2016.10a-02, and dividing the multiple synthetic data sets into multiple training sets, test sets and validation sets;
[0038] A model training module, used to construct a multi-stream spatiotemporal network model, and respectively train and verify the multi-stream spatiotemporal network model using the multiple training sets and verification sets to obtain a trained model;
[0039] The recognition module is used to input the synthetic data set into the trained model to obtain a signal recognition result.
[0040] Optionally, the denoising module uses an EMD algorithm to denoise the noise, including:
[0041] A data processing module is used to determine the maximum and minimum points of the original signal, and perform data processing using a spline curve connection to obtain a maximum envelope and a minimum envelope;
[0042] An envelope line calculation module, used for calculating based on the maximum envelope line and the minimum envelope line to obtain a first average envelope line;
[0043] A residual signal calculation module, used to calculate the residual signal function using the first average envelope, update the original signal and return to the data processing module until the updated residual signal satisfies the IMF signal to obtain the IMF signal;
[0044] A signal component module, used for calculating a residual signal value based on the IMF signal, and performing signal component analysis on the IMF signal to obtain an IMF component set;
[0045] A comparison module, used for calculating the mean square errors of all adjacent IMF components based on the IMF component set for comparison, and determining the value K of the number of IMF components according to the minimum mean square error;
[0046] The denoising signal construction module is used to reconstruct the EMD denoising signal based on the IMF component number value K.
[0047] Optionally, the denoising module uses the EEMD algorithm to denoise, including:
[0048] A signal decomposition module, used for adding white noise to the original signal, and performing EMD decomposition on the signal after adding the white noise to obtain an IMF component;
[0049] An average value calculation module, used for averaging the obtained IMF components to obtain a component average value;
[0050] A signal residual updating module, used for calculating and updating the signal residual based on the component average value;
[0051] An iteration module is used to repeatedly execute the signal decomposition module-signal residual update module, calculate the mean square error of all adjacent IMF components for comparison, and determine the value K of the number of IMF components according to the minimum mean square error;
[0052] The signal reconstruction module is used to reconstruct the signal using the IMF component number value K after several iterations to obtain the EEMD denoised signal.
[0053] Optionally, the model training module uses the training set to train the multi-stream spatiotemporal network, including:
[0054] A category prediction module, used for inputting the signal in the training set into the constructed multi-stream spatiotemporal network to obtain a predicted category label;
[0055] The cross entropy loss calculation module is used to calculate the real category label of the signal in the training set and the predicted category label through the cross entropy loss function to obtain the loss value and update the network parameters, wherein the cross entropy loss function is as follows:
[0056] Loss=-Σt i log(p i )
[0057] In the formula, Loss is the value of the cross entropy loss function, t i and p i is the true label of the class and the predicted probability of the model for the class;
[0058] The iterative training module is used to complete model training after several iterations.
[0059] Beneficial effects of the above technical solution:
[0060] (1) The improved signal denoising method under low signal-to-noise ratio conditions proposed in the present invention can improve the recognition accuracy of the network by denoising the low signal-to-noise ratio data.
[0061] (2) The use of complex convolution modules in the multi-stream spatiotemporal network used in the present invention can more comprehensively process IQ data and effectively capture the rich features between the two, deeply explore the intrinsic connection between IQ signals, thereby helping the model to more accurately understand the complex characteristics of the signal and improve the recognition accuracy of the signal under low signal-to-noise conditions.
[0062] (3) Based on the denoised signal datasets 1 and 2, data enhancement is performed on them combined with RML2016.10a to form RML2016.10a-01 and RML2016.10a-02, which reduces the sample imbalance ratio, improves the robustness of the model, and makes the model more generalizable.
[0063] (4) The experimental results show that the data after EMD / EEMD processing is fused with the original data set. In terms of the overall recognition rate, the RML2016.10a-01 data set formed by the fusion of data after EMD processing has a high overall recognition rate of 62.95%, of which the highest recognition accuracy is 93.90%. The recognition accuracy of the RML2016.10a-01 data set is improved to 92.69% between 0dB and 18dB, and the recognition accuracy of the RML2016.10a-02 data set is high between -20dB and -2dB, which has a promoting effect on the distinction of low signal-to-noise ratio. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] Figure 1 A flowchart of a modulation recognition method based on a multi-stream spatiotemporal network under low signal-to-noise ratio provided by the present invention;
[0065] Figure 2 A schematic diagram of the structure of a modulation recognition device based on a multi-stream spatiotemporal network under low signal-to-noise ratio provided by the present invention;
[0066] Figure 3 It is a schematic diagram of the signal denoising effect based on the improved empirical mode decomposition method;
[0067] Figure 4 It is a schematic diagram of the signal denoising effect based on the improved ensemble empirical mode decomposition method;
[0068] Figure 5 Schematic diagram of modulation recognition model based on multi-stream spatiotemporal network;
[0069] Figure 6 Schematic diagram of the recognition accuracy of each data set in the multi-stream spatiotemporal network;
[0070] Figure 7 Comparison chart of the training accuracy of each data set in the multi-stream spatiotemporal network;
[0071] Figure 8 Comparison chart of training loss of each data set in multi-stream spatiotemporal network;
[0072] Fig. 9 Comparison chart of verification accuracy of each data set in multi-stream spatiotemporal network;
[0073] Fig.10 Comparison chart of verification loss of each data set in multi-stream spatiotemporal network;
[0074] Fig.11 This is a comparison chart of RML2016.10a-01 when the signal-to-noise ratio is -6dB, 0dB, and 6dB;
[0075] Fig.12 This is a comparison chart of RML2016.10a-02 when the signal-to-noise ratio is -6dB, 0dB, and 6dB. DETAILED DESCRIPTION
[0076] The following detailed description of the embodiments of the present invention is further described in detail in conjunction with the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to exemplarily illustrate the principles of the present invention, but cannot be used to limit the scope of the present invention, that is, the present invention is not limited to the preferred embodiments described, and the scope of the present invention is defined by the claims.
[0077] In the description of the present invention, it should be noted that, unless otherwise specified, “plurality” means two or more than two; the terms “first”, “second”, etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance; for ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.
[0078] Embodiment 1
[0079] An embodiment of the present invention provides a modulation recognition method based on a multi-stream spatiotemporal network under low signal-to-noise ratio, such as Figure 1 As shown, a modulation recognition method based on a multi-stream spatiotemporal network under low signal-to-noise ratio is characterized by comprising:
[0080] S1, the acquired original signal is denoised using the EMD algorithm and the EEMD algorithm respectively to obtain two denoised signal data sets;
[0081] The denoised signal data set includes data set 1 and data set 2. An improved empirical mode decomposition (EMD) method is used to denoise some low signal-to-noise ratio signals in the data set to form a new data set 1. An improved ensemble empirical mode decomposition (EEMD) method is used to denoise some low signal-to-noise ratio signals in the data set to form a new data set 2.
[0082] It should be noted that the Empirical Mode Decomposition (EMD) algorithm is a heuristic technology for non-stationary and nonlinear signal processing. The decomposition idea is to use the extreme point information of the original signal from the time dimension to decompose the signal into several different components with high to low frequencies. This frequency component is called the intrinsic mode function (IMF). Noise filtering can be achieved by selecting the frequency band that you want to retain and reconstructing it in a certain way.
[0083] EMD algorithm denoising includes:
[0084] S11A, determining the maximum and minimum points of the original signal, connecting them with a spline curve for data processing, and obtaining a maximum envelope and a minimum envelope;
[0085] Find the maximum and minimum points of the original signal y(t) waveform, connect the extreme points with spline curves to form the maximum envelope curve e max (t) and the minimum envelope e min (t).
[0086] S12A, performing calculation based on the maximum envelope and the minimum envelope to obtain the first average envelope.
[0087] The first average envelope is m1(t) obtained by averaging the maximum / minimum envelope. The formula is:
[0088] m1(t)=(e max (t)+e min (t)) / 2
[0089] S13A, using the first average envelope to calculate the residual signal function, update the original signal and return to S11A, until the updated residual signal satisfies the IMF signal, thereby obtaining the IMF signal;
[0090] The new residual signal function h(t) is obtained as follows:
[0091] h(t)=y(t)-m1(t)
[0092] S14A, calculating a residual signal value based on the IMF signal, and performing signal component analysis on the IMF signal to obtain an IMF component set;
[0093] Let h(t) = y(t), repeat steps 1-3 above, and obtain the i-th h k,i (t)(i=2,3,4……) signal, if it satisfies the IMF signal, to calculate the new residual signal R k (t) value, the formula is:
[0094] R k (t) = h(t) - h k,i (t)
[0095] h k,i As long as the value of (t) (i=2, 3, 4, ...) does not satisfy the IMF condition, S14 is skipped and steps S11-S13 are directly repeated.
[0096] S15A, calculating the mean square errors of all adjacent IMF components based on the IMF component set for comparison, and determining the value K of the number of IMF components according to the minimum mean square error;
[0097] y(t) is decomposed into IMF signal components by EMD, and the value of variable K is determined when the continuous minimum mean square error (RMES) of two adjacent IMFs is taken. Variable K represents the value of the number of IMF components.
[0098] S16A, constructing an EMD denoising signal based on the number of IMF components.
[0099] When K! = 0, the K value is used to construct the denoised signal S_new. When K = 0, S_new = y(t)
[0100] In S16A, the EMD denoising signal is constructed as follows:
[0101]
[0102] Where IMF[i] is the ith IMF signal, and k is the minimum mean square error.
[0103] Ensemble empirical mode decomposition EEMD algorithm: Determine the addition of white noise to the original signal y(t), and add white noise to y(t) m Perform EMD decomposition to obtain the first IMF component. After completing the above steps, the IMF obtained each time is integrated and averaged to obtain the frequency component.
[0104] EEMD algorithm denoising includes:
[0105] S11B, adding white noise to the original signal, and performing EMD decomposition on the signal after the white noise is included to obtain an IMF component;
[0106] Determine the amplitude ε and the number m of white noise added to the original signal y(t); generate noise signals y(t)1, y(t)2...y(t) respectively. m Perform EMD decomposition on the above noise signals to obtain IMF components, corresponding to IMF1, IMF2...IMF m .
[0107] S12B, averaging the obtained IMF components to obtain a component average;
[0108] The average value of the IMF component is calculated as follows:
[0109]
[0110] S13B, calculating and updating a signal residual based on the component average value;
[0111] Update the signal residual r1, the corresponding formula is:
[0112] r1=y(t)-IMF1
[0113] S14B, repeat S11 B-S13B, calculate the mean square errors of all adjacent IMF components for comparison, and determine the value K of the number of IMF components according to the minimum mean square error;
[0114] When taking the continuous minimum mean square errors (RMES) of two adjacent IMFs, the value of the variable K is determined.
[0115] S15B, after several iterations, reconstruct the signal using the IMF component number value K to obtain the EEMD denoised signal.
[0116] The reconstructed EEMD denoised signal is as follows:
[0117]
[0118] Where IMFK is the number of IMF components, rK is the signal residual after K iterations, and k is the minimum mean square error.
[0119] S2, combining the two denoised signal data sets with the RML2016.10a data set respectively to obtain two synthetic data sets RML2016.10a-01 and RML2016.10a-02, and dividing the multiple synthetic data sets into multiple training sets, test sets and validation sets;
[0120] It should be noted that the RML2016.10a dataset is a wireless communication dataset used for modulation identification, which contains multiple modulation modes and multiple signal-to-noise ratio levels.
[0121] The newly constructed dataset 1 (EMD denoised signal) is combined with RML2016.10a for data enhancement to form the first synthetic dataset RML2016.10a-01, and the newly constructed dataset 2 (EEMD denoised signal) is combined with RML2016.10a for data enhancement to form the second synthetic dataset RML2016.10a-02. Then the first and second synthetic datasets are divided into a training set, a test set and a validation set, respectively, to obtain a first training set, a first test set and a first validation set, as well as a second training set, a second test set and a second validation set, for subsequent training, testing and validation of the model.
[0122] S3. Construct a multi-stream spatiotemporal network, and use the training set and the verification set to train and verify the multi-stream spatiotemporal network to obtain a trained model.
[0123] The multi-stream spatiotemporal network structure is as follows Figure 5 As shown, it is a complex-valued convolution fusion multi-stream spatiotemporal network, including a spatial feature module, LSTM and Bi-LSTM network structures and a temporal feature extraction module, wherein the spatial feature module includes a complex-valued convolution module, and the corresponding expression can be expressed as:
[0124] Y n =I n +jQ n
[0125] K m =K' m +jK” m
[0126] For the complex input data Y n With the complex convolution kernel K m Performing convolution operation, we get the following result:
[0127] Y n *K m =(I n +jQ n )*(K' m +jK” m )
[0128] Y n *K m =(I n *K' m -Q n *K” m )+j(I n *K m +Q n *K” m )
[0129] Y n Conduct Y n *K m When X is convolved as a real number DL The expression is:
[0130]
[0131] When X DL Do It is equivalent to X during linear transformation, realizing the process in which one-dimensional complex convolution is equivalent to two-dimensional real convolution after linear transformation.
[0132] Using the training set to train the multi-stream spatiotemporal network model includes:
[0133] S31, inputting the signal in the training set into the constructed multi-stream spatiotemporal network to obtain a predicted category label;
[0134] S32, calculating the real category label and the predicted category label of the signal in the training set through a cross entropy loss function to obtain a loss value and update the network parameters, wherein the cross entropy loss function is as follows:
[0135] Loss=-∑t i log(p i )
[0136] In the formula, Loss is the value of the cross entropy loss function, t i and p i is the true label of the class and the predicted probability of the model for the class;
[0137] S33, repeat S21-S22, and complete the model training after several iterations.
[0138] Then, the denoising method with the highest recognition accuracy in the verification sample is selected as the final model pre-training result.
[0139] Based on the above training and verification method, the multi-stream spatiotemporal network model is trained and verified using the first training set, the second training set, the first verification set, and the second verification set to obtain a trained model.
[0140] S4. Input the synthetic data set into the trained model to obtain a signal recognition result.
[0141] The first test set and the second test set are respectively input into the trained model to obtain predicted categories.
[0142] The present embodiment is a modulation recognition method based on a multi-stream space-time network under low signal-to-noise ratio. The original signal (-18dB~-10dB) is decomposed by using an improved empirical mode decomposition (EMD) / ensemble empirical mode decomposition (EEMD) algorithm to obtain IMF components that meet certain conditions. The optimal value is selected to achieve signal reconstruction based on the minimum root mean square error (RMSE) of adjacent roots. The signal samples reconstructed by using the empirical mode decomposition (EMD) form a denoised data set 1, and the signal samples reconstructed by using the ensemble empirical mode decomposition (EEMD) form a denoised data set 2. Then, the denoised data is fused with the original data set for data enhancement, and the data are respectively input into the multi-stream space-time network to extract signal features, aiming to improve the recognition accuracy under low signal-to-noise ratio conditions.
[0143] Embodiment 2
[0144] An embodiment of the present invention provides a modulation identification device 200 based on a multi-stream spatiotemporal network under low signal-to-noise ratio, such as Figure 2 As shown, including:
[0145] A denoising module 201 is used to denoise the acquired original signal using an EMD algorithm and an EEMD algorithm respectively to obtain two denoised signal data sets;
[0146] A data collection module 202 is used to combine the two denoised signal data sets with the RML2016.10a data set to obtain two synthetic data sets RML2016.10a-01 and RML2016.10a-02, and divide the multiple synthetic data sets into multiple training sets, test sets and validation sets;
[0147] A model training module 203 is used to construct a multi-stream spatiotemporal network model, and respectively train and verify the multi-stream spatiotemporal network model using the multiple training sets and verification sets to obtain a trained model;
[0148] The recognition module 204 is used to input the synthetic data set into the trained model to obtain a signal recognition result.
[0149] Furthermore, the denoising module 201 uses the EMD algorithm to denoise, including:
[0150] The data processing module 2011A is used to determine the maximum and minimum points of the original signal, and use spline curves to connect and process the data to obtain the maximum envelope and the minimum envelope;
[0151] An envelope line calculation module 2012A is used to calculate based on the maximum envelope line and the minimum envelope line to obtain a first average envelope line;
[0152] A residual signal calculation module 2013A is used to calculate the residual signal function using the first average envelope, update the original signal and return it to the data processing module until the updated residual signal satisfies the IMF signal to obtain the IMF signal;
[0153] A signal component module 2014A is used to calculate a residual signal value based on the IMF signal, and perform signal component analysis on the IMF signal to obtain an IMF component set;
[0154] A comparison module 2015A is used to calculate the mean square errors of all adjacent IMF components based on the IMF component set for comparison, and determine the IMF component number value K according to the minimum mean square error;
[0155] The denoising signal construction module 2016A is used to reconstruct the EMD denoising signal based on the IMF component number value K.
[0156] Furthermore, the denoising module 201 uses the EEMD algorithm to denoise, including:
[0157] The signal decomposition module 2011B is used to add white noise to the original signal, and perform EMD decomposition on the signal after the white noise is added to obtain IMF components;
[0158] An average value calculation module 2012B is used to average the obtained IMF components to obtain a component average value;
[0159] A signal residual updating module 2013B, configured to calculate and update the signal residual based on the component averages;
[0160] Iteration module 2014B is used to repeatedly execute the signal decomposition module-signal residual update module, calculate the mean square error of all adjacent IMF components for comparison, and determine the value K of the number of IMF components according to the minimum mean square error;
[0161] The signal reconstruction module 2015B is used to reconstruct the signal using the IMF component number value K after several iterations to obtain an EEMD denoised signal.
[0162] Furthermore, the model training module 203 uses the training set to train the multi-stream spatiotemporal network, including:
[0163] Category prediction module 2031, used for inputting the signal in the training set into the constructed multi-stream spatiotemporal network to obtain a predicted category label;
[0164] The cross entropy loss calculation module 2032 is used to calculate the real category label of the signal in the training set and the predicted category label through the cross entropy loss function to obtain the loss value and update the network parameters, wherein the cross entropy loss function is as follows:
[0165] Loss=-∑t i log(p i )
[0166] In the formula, Loss is the value of the cross entropy loss function, t i and p i is the true label of the class and the predicted probability of the model for the class;
[0167] The iterative training module 2033 is used to complete the model training after several iterations.
[0168] The present embodiment is a modulation recognition device based on a multi-stream spatiotemporal network under low signal-to-noise ratio, which is used to denoise the acquired original signal using the EMD algorithm and the EEMD algorithm respectively to obtain a denoised signal data set; the data collection module is used to combine the denoised signal data set with the RML2016.10a data set to obtain a synthetic data set and divide the synthetic data set into a training set, a test set and a verification set; the model training module is used to train and verify the constructed multi-stream spatiotemporal network using the training set and the verification set to obtain a trained model; the recognition module is used to input the test set into the trained model to obtain a signal recognition result.
[0169] The present invention utilizes an improved empirical mode decomposition (EMD) / ensemble empirical mode decomposition (EEMD) algorithm to decompose the original signal (-18dB to -10dB) to obtain IMF components that meet certain conditions, and selects the optimal value to achieve signal reconstruction based on the minimum adjacent root mean square error RMSE root, and forms a denoised data set 1 with signal samples reconstructed using empirical mode decomposition (EMD), and forms a denoised data set 2 with signal samples reconstructed using ensemble empirical mode decomposition (EEMD), then the denoised data is fused with the original data set for data enhancement, and input into a multi-stream spatiotemporal network to extract signal features, aiming to improve the recognition accuracy under low signal-to-noise ratio conditions.
[0170] Embodiment 3
[0171] (1) Use EMD / EEMD algorithm to denoise the original noisy signal. The effect is as follows Figure 3-Figure 4 shown.
[0172] (2) According to the decomposed IMF components, the optimal value K is selected based on the minimum adjacent root mean square error (RMSE).
[0173] (3) Signal reconstruction. When the K value is not 0, the EMD / EEMD algorithm is used to perform empirical mode decomposition on the input signal S to obtain a set of intrinsic mode functions IMFS. Then, an all-zero array S_new is created based on the length of the first element in the IMFS. Finally, the IMFS is traversed starting from index K_pre+1, and each IMF is added to S_new. In this way, S_new is obtained by performing empirical mode decomposition and accumulation on the input signal S; when the K value is 0, S_new = S.
[0174] (4) As shown in Table 1, the dataset generated by the denoised signal is fused with the original dataset to generate the corresponding RML2016.10a-01 and RML 2016.10a-02, which are respectively input into the multi-stream spatiotemporal network to realize automatic modulation recognition.
[0176] The experimental results are shown in Table 4 and Figure 6-Figure 12 shown.
[0177] Table 1 Dataset details
[0178]
[0179] The RML 2016.10a-1 dataset is a dataset in which the signal is synthesized with the original dataset after EMD denoising, and the RML 2016.10a-2 dataset is a dataset in which the signal is synthesized with the original dataset after EEMD denoising. It is divided into training, testing and validation sets in a ratio of 6:2:2.
[0180] Table 2 Hardware and software configuration
[0181]
[0182] Table 3 Model training hyperparameter settings
[0183]
[0184] Table 4 Comparison of recognition accuracy of each network on the RML2016.03 dataset
[0185]
[0186] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A modulation recognition method based on multi-stream spatiotemporal network under low signal-to-noise ratio, characterized in that: include: S1. De-noise the acquired original signal using the EMD algorithm and the EEMD algorithm respectively to obtain two de-noised signal data sets; S2, combining the two denoised signal data sets with the RML2016.10a data set respectively to obtain two synthetic data sets RML2016.10a-01 and RML2016.10a-02, and dividing the multiple synthetic data sets into multiple training sets, test sets and validation sets; S3, constructing a multi-stream spatiotemporal network model, and respectively using the multiple training sets and verification sets to train and verify the multi-stream spatiotemporal network model to obtain a trained model; S4. Input the synthetic data set into the trained model to obtain a signal recognition result.
2. The modulation recognition method based on multi-stream spatiotemporal network under low signal-to-noise ratio according to claim 1 is characterized in that: In S1, denoising by using the EMD algorithm includes: S11A, determining the maximum and minimum points of the original signal, connecting them with a spline curve for data processing, and obtaining a maximum envelope and a minimum envelope; S12A, calculating based on the maximum envelope and the minimum envelope to obtain a first average envelope; S13A, using the first average envelope to calculate the residual signal function, update the original signal and return to S11A, until the updated residual signal satisfies the IMF signal, thereby obtaining the IMF signal; S14A, calculating a residual signal value based on the IMF signal, and performing signal component analysis on the IMF signal to obtain an IMF component set; S15A, calculating the mean square errors of all adjacent IMF components based on the IMF component set for comparison, and determining the value K of the number of IMF components according to the minimum mean square error; S16A, reconstructing the EMD denoised signal based on the IMF component number value K.
3. The modulation recognition method based on multi-stream spatiotemporal network under low signal-to-noise ratio according to claim 1 is characterized in that: In S1, denoising by using the EEMD algorithm includes: S11 B, adding white noise to the original signal, and performing EEMD decomposition on the signal after the white noise is included to obtain an IMF component; S12B, averaging the obtained IMF components to obtain a component average; S13B, calculating and updating a signal residual based on the component average value; S14B, repeat S11B-S13B, calculate the mean square errors of all adjacent IMF components for comparison, and determine the value K of the number of IMF components according to the minimum mean square error; S15B, after several iterations, reconstruct the signal using the IMF component number value K to obtain the EEMD denoised signal.
4. The modulation recognition method based on multi-stream spatiotemporal network under low signal-to-noise ratio according to claim 1 is characterized in that: In S3, using the training set to train the multi-stream spatiotemporal network includes: S31, inputting the signal in the training set into the constructed multi-stream spatiotemporal network to obtain a predicted category label; S32, calculating the real category label and the predicted category label of the signal in the training set through a cross entropy loss function to obtain a loss value and update the network parameters, wherein the cross entropy loss function is as follows: Loss=-Σt i log(p i ) In the formula, Loss is the value of the cross entropy loss function, t i and p i is the true label of the i-th class and the predicted probability of the model for the i-th class; S33, repeat S31-S32, and complete the model training after several iterations.
5. The modulation recognition method based on multi-stream spatiotemporal network under low signal-to-noise ratio according to claim 2 is characterized in that: In S16A, the EMD denoising signal is constructed as follows: Where IMF[i] is the ith IMF signal, and k is the minimum mean square error.
6. The modulation recognition method based on multi-stream spatiotemporal network under low signal-to-noise ratio according to claim 3 is characterized in that: In S15B, the EEMD denoised signal is reconstructed according to the following formula: Where IMFK is the number of IMF components, rK is the signal residual after K iterations, and k is the minimum mean square error.
7. A modulation recognition device based on multi-stream spatiotemporal network under low signal-to-noise ratio, characterized in that: include: The denoising module is used to denoise the acquired original signal using the EMD algorithm and the EEMD algorithm respectively to obtain two denoised signal data sets; A data collection module, used for combining the two denoised signal data sets with the RML2016.110a data set respectively to obtain two synthetic data sets RML2016.10a-01 and RML2016.10a-02, and dividing the multiple synthetic data sets into multiple training sets, test sets and validation sets; A model training module, used to construct a multi-stream spatiotemporal network model, and respectively train and verify the multi-stream spatiotemporal network model using the multiple training sets and verification sets to obtain a trained model; The recognition module is used to input the synthetic data set into the trained model to obtain a signal recognition result.
8. The modulation identification device based on multi-stream spatiotemporal network under low signal-to-noise ratio according to claim 7, characterized in that: The denoising module comprises: A data processing module is used to determine the maximum and minimum points of the original signal, and perform data processing using a spline curve connection to obtain a maximum envelope and a minimum envelope; An envelope line calculation module, used for calculating based on the maximum envelope line and the minimum envelope line to obtain a first average envelope line; A residual signal calculation module, used to calculate the residual signal function using the first average envelope, update the original signal and return to the data processing module until the updated residual signal satisfies the IMF signal to obtain the IMF signal; A signal component module, used for calculating a residual signal value based on the IMF signal, and performing signal component analysis on the IMF signal to obtain an IMF component set; A comparison module, used for calculating the mean square errors of all adjacent IMF components based on the IMF component set for comparison, and determining the value K of the number of IMF components according to the minimum mean square error; The denoising signal construction module is used to reconstruct the EMD denoising signal based on the IMF component number value K.
9. The modulation identification device based on multi-stream spatiotemporal network under low signal-to-noise ratio according to claim 7, characterized in that: The denoising module also includes: A signal decomposition module, used for adding white noise to the original signal, and performing EMD decomposition on the signal after adding the white noise to obtain an IMF component; An average value calculation module, used for averaging the obtained IMF components to obtain a component average value; A signal residual updating module, used for calculating and updating the signal residual based on the component average value; An iteration module is used to repeatedly execute the signal decomposition module-signal residual update module, calculate the mean square error of all adjacent IMF components for comparison, and determine the value K of the number of IMF components according to the minimum mean square error; The signal reconstruction module is used to reconstruct the signal using the IMF component number value K after several iterations to obtain the EEMD denoised signal.
10. The modulation identification device based on multi-stream spatiotemporal network under low signal-to-noise ratio according to claim 7, characterized in that: The model training module includes: A category prediction module, used for inputting the signal in the training set into the constructed multi-stream spatiotemporal network to obtain a predicted category label; The cross entropy loss calculation module is used to calculate the real category label of the signal in the training set and the predicted category label through the cross entropy loss function to obtain the loss value and update the network parameters, wherein the cross entropy loss function is as follows: Loss=-Σt i log(p i ) In the formula, Loss is the value of the cross entropy loss function, t i and p i is the true label of the i-th class and the predicted probability of the model for the i-th class; The iterative training module is used to complete model training after several iterations.
Citation Information
Patent Citations
Aero-engine remaining life prediction method based on multi-stage information fusion
CN104166787A
Analysis method of gear train noise based on ensemble empirical mode decomposition and support vector machine
CN106778694A
Seismic signal detection algorithm based on GS transforming filtering and EMD denoising
CN109212603A
Space-time multi-channel deep learning system for automatic modulation identification
CN110598677A
Photovoltaic DC signal denoising method based on EMD combined with adaptive wavelet soft and hard thresholds
CN115577238A
Cited By
Clinical test data prediction method based on big data
CN120356693A
