A dual-path machine learning modulation pattern recognition method for the Internet of Things

Abstract

The present invention belongs to the technical field of communication signal modulation recognition, and specifically is a dual-path machine learning modulation pattern recognition method for the Internet of Things. The present invention includes: collecting image data sets; using prototype network small sample meta-learning to train models to cope with recognition capabilities from different types of images; the recognition stage is divided into fuzzy classification and precise classification dual classification stages: in the fuzzy classification stage, non-digital modulation signals and digital modulation signals are identified, and fuzzy classification results are provided for precise classification as a reference. In the precise classification stage, the distinguished digital modulation signals are restored by restoring the constellation diagram to obtain precise classification results; the two classification results are used to determine the final digital modulation signal results through a specific decision-making method. The dual-path dual recognition method based on the meta-learning neural network of the present invention makes the recognition signal type cover a wider range, with low complexity and higher accuracy, and improves the recognition accuracy by about 8% compared with a single method.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication signal modulation recognition, and in particular relates to a dual-path machine learning modulation pattern recognition method for the Internet of Things. Background Art

[0002] Modulation mode recognition of communication signals is a key technology in the field of signal processing and pattern recognition, and it is also a difficult technology. This technology is widely used in military and civilian fields, and has important application value and scientific significance. When transmitting information, in order to use the channel to quickly and effectively transmit real information, different signal modulation modes are often used in communication systems to fully utilize the channel capacity. Therefore, the communication signal modulation mode is one of the key features to distinguish different types of communication signals.

[0003] In response to the difficulty of modulation recognition, academic and industrial research can only perform intelligent recognition of some modulation modes. With the development of 5G communication technology, the modulation modes of communication signals are becoming more and more complex, and the types of modulation signals are increasing. Under the premise of ensuring recognition accuracy and low complexity, how to cope with the increase in the types of modulation signals and how to autonomously learn and recognize them has become a new development direction for modulation mode recognition technology in the future. Summary of the invention

[0004] The present invention provides a dual-path machine learning modulation pattern recognition method for the Internet of Things to solve the following problems in the prior art: the existing intelligent modulation pattern recognition can only recognize some specific modulation signals, which not only has a high training cost, but also cannot recognize linear modulation, nonlinear modulation and digital modulation.

[0005] The dual-path machine learning modulation pattern recognition method for the Internet of Things provided by the present invention includes:

[0006] Before training the network, an image data set is obtained, including obtaining a time-frequency feature image by extracting the time-frequency features of the modulated signal, and generating a constellation diagram by restoring the modulated signal;

[0007] In the network training phase, the prototype network small sample meta-learning method is used to train the model to cope with the recognition capabilities of different types of images;

[0008] In the recognition stage, it is divided into two stages: fuzzy classification and precise classification. In the fuzzy classification stage, the non-digital modulation signal and the digital modulation signal can be effectively identified by identifying the time-frequency feature image of the signal to be identified, and the fuzzy classification result is provided as a reference for precise classification. In the precise classification stage, the distinguished digital modulation signal is restored by restoring the constellation diagram to obtain the precise classification result. Through the decision classification algorithm, the two classification results are determined by a specific decision-making method to determine the final digital modulation signal result.

[0009] The specific steps are:

[0010] S1. Collecting image data sets: dividing the modulated signal into two paths: generating time-frequency feature images and restoring constellation diagrams, acquiring images and performing label processing respectively to form image data sets;

[0011] S2. performing image preprocessing on the labeled image dataset to obtain a preprocessed image dataset;

[0012] S3. Train and test the preprocessed data set through small sample meta-learning of prototype network to obtain the optimal model and build a meta-learning neural network;

[0013] S4. For the signal to be identified, fuzzy classification is first performed, the signal is normalized to generate a time-frequency feature image, and the fuzzy classification result is output through the trained meta-learning neural network. The non-digital modulation signal in the signal to be identified is removed to obtain the fuzzy classification digital modulation signal and the non-digital modulation signal result is output, and the fuzzy classification result of the digital modulation signal is passed to the decision classification module for storage;

[0014] S5. Accurately classify the fuzzy classified digital modulation signal, restore the constellation diagram and data enhancement to generate pre-processed images, obtain accurate classification results through the trained meta-learning neural network, and pass them to the decision classification module for storage;

[0015] S6. Use the decision classification algorithm to make a decision on the two classification results to obtain the type of digital modulation signal.

[0016] Furthermore, in the process of collecting the image data set in step S1, the modulation signals are divided into two mutually exclusive groups, and images are obtained through the above two paths respectively. After the images are obtained, the modulation type labels are annotated on the images. In order to ensure low complexity, the image data set is annotated as follows: the time-frequency feature images obtained through the time-frequency features are only annotated with fuzzy classification labels such as AM, FM, PSK, ASK, FSK, QAM, etc., and the constellation diagrams need to be annotated with precise classification labels such as modulation order in addition to the modulation type: {2~2 n}ASK,(n=2,3,4),{2~2 n}PSK,(n=2,3,4),{2~2 n}QAM, (n=2,3,…6)}.

[0017] Furthermore, the specific process of the preprocessing in step S2 is as follows:

[0018] S21. Perform size homogenization processing on the image data set to 30*30 pixels to reduce the computing pressure;

[0019] S22. Perform a 15° step rotation on the constellation map dataset to obtain 12 images with step size labels;

[0020] S23. Perform grayscale processing on the images processed by S21 and S22 to obtain a preprocessed image.

[0021] Furthermore, in step S3, during the training process of the meta-learning network, an internal and external double-loop algorithm is adopted, aiming to train the learning ability of the model itself, that is, to learn the distribution of similar tasks to find the optimal model parameters. At the beginning of each task, it starts from the optimal weights, and gradually increases the amount of tasks at each stage, simulating the human learning ability curve for learning, achieving convergence with fewer steps and obtaining better learning results.

[0022] Furthermore, in the constructed prototype network small sample meta-learning network, the prototype C of each category is first obtained k , mainly through the support set S train The average of all sample features in this category is obtained:

[0023]

[0024] Among them, |S k | is the number of all samples belonging to category k in the support set Strain, f θ (x i ) is the training sample x i The feature vector obtained after feature extraction; by comparing the Euclidean distance between each sample in the query set and the category prototype, the Euclidean distance d(.) is converted into probability through the normalized exponential function, and then the test sample is obtained For the probability distribution of category k:

[0025]

[0026] Finally, during the training process, in order to simulate the difference between the probability distribution of sample category prediction and its true probability distribution, a loss function is used:

[0027]

[0028] Furthermore, in the constructed meta-learning network, the prototype network is used as the basis to find the parameters θ of the model M for rapid adaptation. By training the model parameters θ, a specific model that can adapt to the task quickly is obtained, which is called the rapid model M. During the training process, the task is used as the minimum number of training steps. Before assigning the training task, the initial model parameters θ are randomly generated, and then the training data set D is used to obtain the model parameters θ. train The prior task p(T) sampled in is used to find the direction of meta-gradient descent to update the parameters θ. It applies the same forward propagation as the model M and uses a set of parameters shared by the model M. It optimizes the model parameters θ guided by the direction of the meta-gradient until it converges to obtain the optimal model for testing.

[0029] Furthermore, the staged meta-learning neural network algorithm consists in that given a classification model M initialized with random parameters θ = {W, B}, we consider a fast model denoted by M whose parameters are shared by the model M. In each stage batch of size 32 fixed for training or validation, the model parameters θ are first manually updated using data with support set D train Extract from the mth previous task p(T),m so that the parameters θ are updated quickly:

[0030]

[0031] Among them, γ represents the learning rate, is the gradient operator, and L p(T) (f θ ) is the loss function of the fast model M.

[0032] The fast parameters θ are shared with the fast model and the verification data Q train In the current task slice p(T), θ' is passed forward through the fast model M to obtain the cumulative loss of the meta stage; the parameters of the model M are updated through back propagation:

[0033]

[0034] Furthermore, the specific process of fuzzy classification of the signal to be identified in step S4 is as follows:

[0035] S41. Normalize the signal to be identified, compress the amplitude of the modulated signal between [0,1] by minimum and maximum standardization, and reduce the influence of channel fading on the signal amplitude:

[0036]

[0037] Where X represents the input vector, max(X) represents the maximum value of the input vector, and min(X) represents the minimum value of the input vector.

[0038] And generate time-frequency feature images, mainly using Choi-Williams distribution to suppress cross-terms to obtain higher time and frequency resolution for distinction:

[0039]

[0040] Among them, σ = 1 / α is the attenuation coefficient; is the fuzzy function of the signal, is a kernel function in exponential form;

[0041] S42. Obtain fuzzy classification results through a meta-learning neural network and directly output non-digital modulation signal results;

[0042] S43. Perform decision classification on the fuzzy classification of the digital modulation signal results, remove the non-digital modulation signals in the signal to be identified, and obtain the fuzzy classified digital modulation signal.

[0043] Furthermore, the fuzzy classification digital modulation signal is accurately classified in step S5, and the specific process is as follows:

[0044] S51. Restoring the constellation diagram of the fuzzy classified digital modulation signal;

[0045] S52. Normalize the size of the constellation diagram, perform grayscale processing, etc. to obtain a preprocessed image;

[0046] S53. Use the meta-learning neural network to obtain accurate classification results for the preprocessed images and further make classification decisions.

[0047] Furthermore, in step S6, the decision classification algorithm is used to make a decision on the two classification results to obtain the type of digital modulation signal. The specific process is as follows:

[0048] S61. Defining the modulation type by fuzzy classification results;

[0049] S62. Obtaining the modulation order by accurately classifying the results based on the known modulation type;

[0050] S63. Combine the two results as the digital modulation signal recognition result and output it.

[0051] The dual-path machine learning modulation pattern recognition method for the Internet of Things provided by the present invention adopts a prototype network small sample meta-learning neural training model to improve autonomy and robustness, and constructs a meta-learning neural network for two-stage signal recognition, fuzzy classification of unrecognized signals, and rapid distinction between digital modulation signals and linear modulation and nonlinear modulation signals according to time-frequency characteristics. The digital modulation signal is restored to a constellation diagram according to the characteristic quantity, and then the image is pre-processed in multiple aspects such as uniform size, color gamut, and offset through a data enhancement module, in order to resist the impact of harsh environments and pave the way for subsequent precise classification. The pre-processed image is then sent to the meta-learner neural network to obtain the result of accurate classification of the digital modulation signal, and the fuzzy classification result and the accurate classification result are compared and fused with each other in the decision classification module, and output as the digital modulation signal recognition result. Compared with the existing traditional training network modulation recognition method: its beneficial effects are:

[0052] The present invention uses a dual-path dual recognition method based on a meta-learning neural network, a prototype network small sample meta-learning neural network, and decision classification, so that it still has high recognition accuracy in a low signal-to-noise ratio environment, which is about 8% higher than the single recognition accuracy, and increases the range of recognition types. While creatively ensuring the recognition accuracy, on the basis of recognizing digital modulation signals, it can quickly and effectively recognize linear modulation signals and nonlinear modulation signals. In addition, the method has few training samples, low training cost, low complexity, and is easy to implement. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a dual-path modulation pattern recognition method based on prototype network small sample meta-learning provided by an embodiment of the present invention.

[0054] Figure 2 The embodiment of the present invention provides Figure 1 Flowchart for building a prototype network small sample meta-learning network.

[0055] Figure 3 The embodiment of the present invention provides Figure 2 Flowchart of the meta-learning network algorithm.

[0056] Figure 4 The embodiment of the present invention provides Figure 1 Examples of fuzzy classification, images using time-frequency features, and precise classification using constellation map restoration and data enhancement images.

[0057] Figure 5 The present invention provides Figure 1 Comparison chart of experimental simulation recognition accuracy. DETAILED DESCRIPTION

[0058] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided only for illustrative purposes and are not to be construed as limitations of the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute limitations on the scope of patent protection of the present invention, because many changes may be made to the present invention without departing from the scope of the spirit of the present invention.

[0059] (1) Method description

[0060] The dual-path modulation pattern recognition method based on prototype network small sample meta-learning provided by the present invention is as follows: Figure 1 As shown, the steps include:

[0061] S1. The modulated signal is divided into two paths: time-frequency feature image generation and constellation map restoration. Images are obtained and label processed to form an image dataset;

[0062] S2. performing image preprocessing on the labeled image dataset to obtain a preprocessed image dataset;

[0063] S3. Train and test the preprocessed data set through small sample meta-learning of prototype network to obtain the optimal model and build a meta-learning neural network;

[0064] S4. The signal to be identified is first fuzzy classified, the signal is normalized to generate a time-frequency feature image, and the fuzzy classification result is output through the trained meta-learning neural network, the non-digital modulation signal in the signal to be identified is removed to obtain the fuzzy classified digital modulation signal and the non-digital modulation signal result is output, and the fuzzy classification result of the digital modulation signal is passed to the decision classification module for storage;

[0065] S5. Accurately classify the fuzzy classified digital modulation signal, restore the constellation diagram and data enhancement to generate pre-processed images, obtain accurate classification results through the trained meta-learning neural network, and pass them to the decision classification module for storage;

[0066] S6. Use the decision classification module to make a decision on the two classification results to obtain the type of digital modulation signal.

[0067] Specifically, in step S1, during the image data set acquisition process, the modulated signals need to be divided into two mutually exclusive groups and the images are acquired through the above two paths respectively. After the images are acquired, the images are annotated with modulation type labels. The time-frequency feature images are only labeled with fuzzy classification labels such as AM, FM, PSK, ASK, and FSK. In addition to the type, the constellation diagrams also need to be labeled with precise classification labels such as modulation order.

[0068] Furthermore, the step S2 specifically includes the steps of:

[0069] S21. Perform size homogenization processing on the image data set to 30*30 pixels to reduce the computing pressure;

[0070] S22. Perform a 15° step rotation on the constellation map dataset to obtain 12 images with step size labels;

[0071] S23. Perform grayscale processing on the images processed by S21 and S22 to obtain a preprocessed image.

[0072] Furthermore, in step S3, the training of the meta-learning neural network is constructed as follows: Figure 2 As shown, the steps include:

[0073] S31. The labeled image dataset D obtained by the modulated signal is divided into a training set D according to the mutual exclusion principle. train and the test set D test ;

[0074] S32. The training set D train and the test set D test Classify the labels into support set and query set for training and testing respectively;

[0075] S33. For training set D train The support set S in train and the query set Q train Perform feature extraction and compare query set Q train The Euclidean distance between each sample and the category prototype is converted into probability, and the parameter convergence is judged by minimizing the loss function. If it does not converge, the feature extraction is repeated, and if it converges, the trained model is obtained.

[0076] S34. Set the test set D test The support set S in test and the query set Q test The trained model is used to perform similarity calculations, obtain test accuracy, and determine the generalization ability of the model. If the generalization ability is poor, the parameters are manually adjusted and retrained. If the generalization ability is strong, the optimal model is output and a meta-learning neural network is successfully built.

[0077] Furthermore, the step S4 specifically includes the steps of:

[0078] S41. Normalize the signal to be identified and generate a time-frequency feature image;

[0079] S42. Obtain fuzzy classification results through a meta-learning neural network and directly output non-digital modulation signal results;

[0080] S43. The fuzzy classification of the digital modulation signal result is passed to the decision classification module;

[0081] S44. Eliminate non-digital modulation signals in the signal to be identified to obtain a fuzzy classified digital modulation signal.

[0082] Furthermore, the step S5 specifically includes the steps of:

[0083] S51. Restoring the constellation diagram of the fuzzy classified digital modulation signal;

[0084] S52. Normalize the size of the constellation diagram, perform grayscale processing, etc. to obtain a preprocessed image;

[0085] S53. Obtain accurate classification results from the preprocessed images through a meta-learning neural network and pass them into the decision classification module.

[0086] Furthermore, the step S6 specifically includes the steps of:

[0087] S61. Defining the modulation type by fuzzy classification results;

[0088] S62. Obtaining the modulation order by accurately classifying the results based on the known modulation type;

[0089] S63. Combine the two results as the digital modulation signal recognition result and output it.

[0090] (2) Experimental simulation

[0091] In order to prove the recognition effect of the above method of this embodiment, this experimental example carried out a simulation comparison experiment, one group was a dual-path recognition scheme simulation experiment, and the other group was a constellation diagram single-path recognition scheme simulation experiment. In the selection of neural networks, due to the self-learning function of the meta-learning neural network, and in order to control variables, both were uniformly trained by the dual-path recognition scheme. The neural network. Since the constellation diagram only restores digital modulation signals, the basic modulation signals include {2ASK, 2PSK, 2FSK, 2QAM, 4ASK, 4FSK, 4PSK, 4QAM, 8ASK, 8FSK, 8PSK, 8QAM, 16ASK, 16FSK, 16PSK, 16QAM, 32QAM, 64QAM} (K = 18, i.e. 18 modulation signals), covering different modulation orders and different types of digital modulation methods, simulating complex modulation signal environments as much as possible. In the dual-path recognition scheme, in addition to the 18 basic modulation signals, AM and FM non-digital modulation signals are also included. A total of 20 modulation signals are used for simulation experiments. In the training phase of the meta-learning neural network, the signal-to-noise ratio of each modulation signal varied from 4dB to 10dB in 2dB intervals (4 signal-to-noise ratios), and 20 images of each type were collected at each signal-to-noise ratio as the training set (a total of 1600 images). In the testing phase, 40 images (800 images for dual paths and 720 images for constellation diagrams) were selected for each modulation signal as the test objects for testing.

[0092] like Figure 3 As shown in the figure, in the neural network algorithm, the model parameters are iteratively updated mainly through the internal and external double loop method. By training the model parameters, the model can have autonomous recognition ability. In the network structure design, a 5-layer convolution module design is adopted. Each convolution module is divided into a convolution layer 5*5*64+ReLU function, batch normalization+pooling layer, which improves the generalization ability of the model.

[0093] like Figure 5 As shown in the figure, since the constellation diagram single path recognition scheme cannot recognize non-digital modulation signals, only the recognition accuracy of digital modulation signals is compared. The figure shows the effect of the dual path recognition scheme and the constellation diagram single path recognition scheme on the recognition accuracy of digital modulation signals. It can be seen from the figure that the dual path recognition scheme can indeed improve the recognition accuracy under low signal-to-noise ratio conditions, and the maximum recognition accuracy improvement is close to 8%, which is feasible to a certain extent.

[0094] The above embodiments are preferred implementation modes of the method of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principles of the present invention shall be equivalent replacement methods and shall be included in the protection scope of the present invention.

Claims

1. A dual-path machine learning modulation pattern recognition method for the Internet of Things, characterized in that: include: Before training the network, an image data set is obtained, including obtaining a time-frequency feature image by extracting the time-frequency features of the modulated signal, and generating a constellation diagram by restoring the modulated signal; In the network training phase, the prototype network small sample meta-learning method is used to train the model to cope with the recognition capabilities of different types of images; In the modulation mode recognition stage, it is divided into two stages: fuzzy classification and precise classification. In the fuzzy classification stage, the non-digital modulation signal and the digital modulation signal can be effectively identified by identifying the time-frequency feature image of the signal to be identified, and the fuzzy classification result is provided as a reference for precise classification. In the accurate classification stage, the distinguished digital modulation signal is restored by restoring the constellation diagram to obtain the accurate classification result, and the two classification results are determined by the decision-making classification algorithm to determine the final digital modulation signal result; The specific steps are: S1. Collecting image data sets: dividing the modulated signal into two paths: generating time-frequency feature images and restoring constellation diagrams, acquiring images and performing label processing respectively to form image data sets; S2. performing image preprocessing on the labeled image dataset to obtain a preprocessed image dataset; S3. Train and test the preprocessed data set through small sample meta-learning of prototype network to obtain the optimal model and build a meta-learning neural network; S4. For the signal to be identified, fuzzy classification is first performed, the signal is normalized to generate a time-frequency feature image, and the fuzzy classification result is output through the trained meta-learning neural network, the non-digital modulation signal in the signal to be identified is removed to obtain the fuzzy classification digital modulation signal and the non-digital modulation signal result is output, and the fuzzy classification result of the digital modulation signal is saved; S5. Accurately classify the fuzzy classified digital modulation signal, restore the constellation diagram and data enhancement to generate pre-processed images, and obtain accurate classification results through the trained meta-learning neural network; S6. Use the decision classification algorithm to make a decision on the two classification results to obtain the type of digital modulation signal.

2. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 1 is characterized in that: In the process of collecting the image data set in step S1, the modulated signals are divided into two mutually exclusive groups, and images are obtained through the two paths of generating the time-frequency feature image and restoring the constellation diagram respectively. After the images are obtained, the images are annotated with modulation type labels: the time-frequency feature images obtained through the time-frequency features are only annotated with AM, FM, PSK, ASK, FSK, and QAM fuzzy classification labels, and the constellation diagram is annotated with modulation type and modulation order precise classification labels: 2ASK, 4ASK, 8ASK, 16ASK, 2PSK, 4PSK, 8PSK, 16PSK, 2QAM, 4QAM, 8QAM, 16QAM, 32QAM, 64QAM.

3. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 2 is characterized in that: The specific process of the pre-processing in step S2 is: S21. Perform size homogenization processing on the image dataset to 30*30 pixels; S22. Perform a 15° step rotation on the constellation map dataset to obtain 12 images with step size labels; S23. Perform grayscale processing on the images processed by S21 and S22 to obtain a preprocessed image.

4. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 3 is characterized in that: In the meta-learning neural network training process described in step S3, an internal and external double-loop algorithm is adopted to train the learning ability of the model itself, that is, to learn the distribution of similar tasks to find the optimal model parameters. At the beginning of each task, it starts from the optimal weights, and gradually increases the amount of tasks at each stage, simulating the human learning ability curve for learning, achieving convergence with fewer steps and obtaining better learning results.

5. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 4 is characterized in that: In the constructed prototype network small sample meta-learning network, we first obtain the prototype C of each category k : Among them, |S k | is the number of all samples belonging to category k in the support set Strain, f θ (x i ) is the training sample x i The feature vector obtained after feature extraction; by comparing the Euclidean distance between each sample in the query set and the category prototype, the Euclidean distance d(.) is converted into probability through the normalized exponential function, and then the test sample is obtained For the probability distribution of category k: Finally, during the training process, in order to simulate the difference between the probability distribution of sample category prediction and its true probability distribution, a loss function is used:

6. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 5 is characterized in that: In the constructed meta-learning network, the prototype network is used as the basis to find the parameters θ of the model M for rapid adaptation. By training the model parameters θ, a specific model that can adapt to the task quickly is obtained. This model is called the fast model M. During the training process, the task is used as the minimum number of training steps. Before assigning the training task, the initial model parameters θ are randomly generated, and then the training data set D is used to obtain the model parameters θ. train The prior task p(T) sampled in is used to find the direction of meta-gradient descent to update the parameters θ. It applies the same forward propagation as the model M and uses a set of parameters shared by the model M. It optimizes the model parameters θ guided by the direction of the meta-gradient until it converges to obtain the optimal model for testing.

7. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 6 is characterized in that: The staged meta-learning neural network algorithm consists of given a classification model M initialized with random parameters θ = {W, B}, considering a fast model denoted by M whose parameters are shared by the model M, first manually updating the model parameters θ using data in each stage batch of size 32 fixed for training or validation, with a support set D train Extract from the mth previous task p(T),m so that the parameters θ are updated quickly: Among them, γ represents the learning rate, is the gradient operator, and L p(T) (f θ ) is the loss function of the fast model M; The fast parameters θ are shared with the fast model and the verification data Q train In the current task slice p(T), θ′ is passed forward through the fast model M to obtain the cumulative loss of the meta stage; the parameters of the model M are updated through back propagation:

8. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 7 is characterized in that: The specific process of fuzzy classification of the signal to be identified in step S4 is as follows: S41. Normalize the signal to be identified, and compress the amplitude of the modulated signal between [0,1] by minimum and maximum standardization to reduce the impact of channel fading on the signal amplitude: Where X represents the input vector, max(X) represents the maximum value of the input vector, and min(X) represents the minimum value of the input vector. And generate time-frequency feature images, using Choi-Williams distribution to suppress cross-terms to obtain higher time and frequency resolution for distinction: Among them, σ = 1 / α is the attenuation coefficient; is the fuzzy function of the signal, is a kernel function in exponential form; S42. Obtain fuzzy classification results through a meta-learning neural network and directly output non-digital modulation signal results; S43. Perform decision classification on the fuzzy classification of the digital modulation signal results, remove the non-digital modulation signals in the signal to be identified, and obtain the fuzzy classified digital modulation signal.

9. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 8 is characterized in that: The specific process of accurately classifying the fuzzy classified digital modulation signal in step S5 is as follows: S51. Restoring the constellation diagram of the fuzzy classified digital modulation signal; S52. Normalize the size of the constellation diagram and perform grayscale processing to obtain a preprocessed image; S53. The preprocessed image is passed through a meta-learning neural network to obtain accurate classification results and further make classification decisions.

10. The dual-path machine learning modulation pattern recognition method for the Internet of Things according to claim 9 is characterized in that: In step S6, the decision classification algorithm is used to make a decision on the two classification results to obtain the type of digital modulation signal. The specific process is as follows: S61. Defining the modulation type by fuzzy classification results; S62. Obtaining the modulation order by accurately classifying the results based on the known modulation type; S63. Combine the two results as the digital modulation signal recognition result and output it.

Images