A method and system for open set recognition of deep neural network modulated signals

Abstract

The present invention provides a method and system for open-set recognition of modulation signals using a deep neural network, wherein the method comprises: establishing a data set X of known modulation types and a data set V of unknown modulation types in a first environment, initializing the parameters of the generator G and the discriminator D; obtaining a predetermined parameter update times threshold; updating the parameters of the discriminator D and the generator G according to the predetermined parameter update times threshold, and performing signal enhancement on the data set X and the data set V; modeling the modulation signal through a one-dimensional convolutional residual network to obtain a 1D-ResNet network model; inputting a test set S into the 1D-ResNet network model to obtain a first recognition result. The method solves the technical problems in the prior art that the modulation recognition algorithm is sensitive to noise, does not consider the open-set recognition in a real environment, cannot distinguish unknown modulation type signals, and has high requirements for signal quality.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a method and system for open-set recognition of deep neural network modulated signals. Background Art

[0002] In recent years, with the rise of neural network (NN) technology, neural networks have begun to be applied to modulation recognition tasks due to their excellent feature extraction and data mapping capabilities. In a complex real electromagnetic environment, it is difficult to collect all types of debugging data for neural network training in the training set. Therefore, there will be some modulation types that have not appeared in the training set in the test set. When performing recognition, it is necessary not only to recognize signals of known modulation types, but also to recognize signals of unknown modulation types. Such recognition tasks are called open set recognition. Since wireless signals are easily affected by various interferences and noises when propagating in real environments, the signal quality of the signal received by the receiving end is often poor, which will lead to the failure of modulation recognition.

[0003] However, the inventors of this application found that the above technology has at least the following technical problems in the process of implementing the technical solution of the invention in the embodiment of this application:

[0004] There are technical problems in the existing technology that the modulation recognition algorithm is sensitive to noise, does not consider the problem of open set recognition in real environment, cannot distinguish unknown modulation type signals, has high requirements on signal quality, and has poor algorithm robustness and generalization. Summary of the invention

[0005] The embodiment of the present application provides a method and system for open-set recognition of modulation signals using a deep neural network, which solves the technical problems in the prior art that the modulation recognition algorithm is sensitive to noise, does not consider the problem of open-set recognition in a real environment, cannot distinguish unknown modulation type signals, and has high requirements for signal quality. The problem of being sensitive to noise, unable to distinguish unknown modulation type signals, and having high requirements for signal quality is solved through signal enhancement. By adding unknown modulation type signals to the training data and adding a loss function, open-set recognition is achieved, the recognition accuracy is improved, and the robustness and generalization ability are enhanced.

[0006] In view of the above problems, an embodiment of the present application provides a method and system for open-set recognition of deep neural network modulated signals.

[0007] In a first aspect, an embodiment of the present application provides a method for open-set recognition of deep neural network modulation signals, wherein the method comprises: establishing a data set X of known modulation types and a data set V of unknown modulation types in a first environment, wherein X={x1, x2, ..., x n},n=1,2,...,N, V={v1,v2,...,v m},m=1,2,...,M; initialize the parameters of the generator G and the discriminator D; obtain a predetermined parameter update number threshold; according to the predetermined parameter update number threshold, update the parameters of the discriminator D and the generator G, and perform signal enhancement on the data set X and the data set V; model the modulated signal through a one-dimensional convolutional residual network based on the data set X and the data set V after signal enhancement, and obtain a 1D-ResNet network model; input the test set S into the 1D-ResNet network model to obtain a first recognition result.

[0008] On the other hand, an embodiment of the present application provides a deep neural network modulation signal open set recognition system, wherein the system comprises: a first establishment unit, the first establishment unit is used to establish a data set X of known modulation type and a data set V of unknown modulation type in a first environment, wherein X = {x1, x2, ..., x n},n=1,2,...,N, V={v1,v2,...,v m},m=1,2,...,M;a first execution unit, the first execution unit is used to initialize the parameters of the generator G and the discriminator D;a first acquisition unit, the first acquisition unit is used to obtain a predetermined parameter update times threshold;a second execution unit, the second execution unit is used to update the parameters of the discriminator D and the generator G according to the predetermined parameter update times threshold, and perform signal enhancement on the data set X and the data set V;a second acquisition unit, the second acquisition unit is used to model the modulated signal through a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement, and obtain a 1D-ResNet network model;a third acquisition unit, the third acquisition unit is used to input the test set S into the 1D-ResNet network model to obtain a first recognition result.

[0009] In a third aspect, an embodiment of the present application provides a deep neural network modulated signal open set identification system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods described in the first aspect when executing the program.

[0010] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0011] Since a dataset X of known modulation types and a dataset V of unknown modulation types are established in the first environment, where X = {x1, x2, ..., x n},n=1,2,...,N, V={v1,v2,...,v m},m=1,2,...,M;initialize the parameters of the generator G and the discriminator D;obtain a predetermined parameter update times threshold;update the parameters of the discriminator D and the generator G according to the predetermined parameter update times threshold,and perform signal enhancement on the data set X and the data set V;model the modulated signal through a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement to obtain a 1D-ResNet network model;input the test set S into the 1D-ResNet network model to obtain a technical solution for the first recognition result. The embodiment of the present application provides a method and system for open set recognition of modulated signals using a deep neural network, which solves the problems of being sensitive to noise, being unable to distinguish unknown modulation type signals, and having high requirements on signal quality through signal enhancement. By adding unknown modulation type signals and a loss function to the training data, open set recognition is achieved, and the recognition accuracy is improved, and the robustness and generalization ability are enhanced.

[0012] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A schematic diagram of a flow chart of an open set recognition method for a deep neural network modulated signal according to an embodiment of the present application;

[0014] Figure 2 A schematic diagram of a process of establishing a data set for an open set recognition method of a deep neural network modulated signal according to an embodiment of the present application;

[0015] Figure 3 A schematic diagram of a single parameter update process of an open set recognition method for a deep neural network modulated signal according to an embodiment of the present application;

[0016] Figure 4 A schematic diagram of a flow chart of a training discriminator of an open set recognition method for a deep neural network modulated signal according to an embodiment of the present application;

[0017] Figure 5 A schematic diagram of a flow chart of a training generator for an open set recognition method of a deep neural network modulated signal according to an embodiment of the present application;

[0018] Figure 6 A schematic diagram of a process of generating a denoised modulation data set and an unknown modulation signal data set in a method for open set recognition of a modulation signal using a deep neural network according to an embodiment of the present application;

[0019] Figure 7 A schematic diagram of a flow chart of modeling a modulated signal in a method for open-set recognition of modulated signals using a deep neural network according to an embodiment of the present application;

[0020] Figure 8 A schematic diagram of a flow chart of obtaining recognition accuracy of an open set recognition method for a deep neural network modulated signal according to an embodiment of the present application;

[0021] Fig. 9 This is a schematic diagram of the structure of an open-set recognition system for a deep neural network modulated signal according to an embodiment of the present application;

[0022] Fig.10 It is a schematic diagram of the structure of an exemplary electronic device according to an embodiment of the present application.

[0023] Explanation of the reference numerals: first establishing unit 11 , first executing unit 12 , first obtaining unit 13 , second executing unit 14 , second obtaining unit 15 , third obtaining unit 16 , electronic device 300 , memory 301 , processor 302 , communication interface 303 , bus architecture 304 . DETAILED DESCRIPTION

[0024] The embodiment of the present application provides a method and system for open-set recognition of modulation signals using a deep neural network, which solves the technical problems in the prior art that the modulation recognition algorithm is sensitive to noise, does not consider the problem of open-set recognition in a real environment, cannot distinguish unknown modulation type signals, and has high requirements for signal quality. The problem of being sensitive to noise, unable to distinguish unknown modulation type signals, and having high requirements for signal quality is solved through signal enhancement. By adding unknown modulation type signals to the training data and adding a loss function, open-set recognition is achieved, the recognition accuracy is improved, and the robustness and generalization ability are enhanced.

[0025] Application Overview

[0026] In recent years, with the rise of neural network (NN) technology, neural networks have begun to be applied to modulation recognition tasks due to their excellent feature extraction and data mapping capabilities. In a complex real electromagnetic environment, it is difficult for the training set to collect all types of debugging data for neural network training. Therefore, there will be some modulation types in the test set that have not appeared in the training set. When performing recognition, it is necessary not only to recognize signals of known modulation types, but also to recognize signals of unknown modulation types. Such recognition tasks are called open set recognition. Since wireless signals are easily affected by various interferences and noises when propagating in a real environment, the signal quality of the signal received by the receiving end is often poor, which will lead to the failure of modulation recognition. The technical problems that the modulation recognition algorithm in the prior art is sensitive to noise, does not consider the problem of open set recognition in a real environment, cannot distinguish unknown modulation type signals, and has high requirements for signal quality are solved.

[0027] In response to the above technical problems, the overall idea of ​​the technical solution provided by this application is as follows:

[0028] The embodiment of the present application provides a method for open set recognition of deep neural network modulation signals, wherein the method comprises: establishing a data set X of known modulation types and a data set V of unknown modulation types in a first environment, wherein X={x1, x2, ..., x n},n=1,2,...,N, V={v1,v2,...,v m},m=1,2,...,M; initialize the parameters of the generator G and the discriminator D; obtain a predetermined parameter update number threshold; according to the predetermined parameter update number threshold, update the parameters of the discriminator D and the generator G, and perform signal enhancement on the data set X and the data set V; model the modulated signal through a one-dimensional convolutional residual network based on the data set X and the data set V after signal enhancement, and obtain a 1D-ResNet network model; input the test set S into the 1D-ResNet network model to obtain a first recognition result.

[0029] After introducing the basic principles of the present application, various non-limiting implementation methods of the present application will be specifically described below in conjunction with the drawings in the specification.

[0030] Embodiment 1

[0031] like Figure 1 As shown, an embodiment of the present application provides a method for identifying an open set of deep neural network modulated signals, wherein the method includes:

[0032] S100: Establishing a dataset X of known modulation types and a dataset V of unknown modulation types in a first environment, wherein X={x1, x2, ..., xn},n=1,2,...,N, V={v1,v2,...,v m},m=1,2,...,M;

[0033] Specifically, the first environment refers to any complex environment including Gaussian white noise environment with different signal-to-noise ratios and real environmental noise channel environment, etc., and the known modulation types include amplitude modulation (AM), frequency modulation (FM), frequency shift keying (FSK), phase shift keying (PSK), binary phase shift keying (BPSK), quadrature phase shift keying (QPSK), etc. The noisy signal data set obtained by adding Gaussian white noise with different signal-to-noise ratios and real environmental noise to the pure signal data set of the known modulation type is the data set X of the known modulation type, where the data set X can be expressed as X={x1,x2,...,x n},n=1,2,...,N, where N is the number of signal samples. Further, two noisy modulated signal data sets V={v1,v2,...,v m}, m = 1, 2, ..., M, where M is the number of signal samples. This can lay a solid foundation for subsequent signal enhancement and modulated signal open set recognition operations.

[0034] S200: Initialize the parameters of the generator G and the discriminator D;

[0035] Specifically, the modulated signals under various complex channels are enhanced through the conditional generative adversarial network (ConditionalGAN, CGAN). CGAN transforms the noisy modulated signal into a denoised signal of the same modulation type by learning a nonlinear mapping function F. CGAN consists of two deep neural networks, namely the generator (Generator, G) and the discriminator (Discriminator, D). The generator G is used to act as the mapping function F, and the input is the noisy modulated signal x n After that, the denoised modulation signal is generated The discriminator D measures the denoised modulated signal x generated by the generator G n With pure signal y nThe initialization parameters refer to the process of initializing the weights and biases of each node before network training. The initialization of parameters is related to whether the network can be trained to produce good results or how fast it converges to obtain good training results.

[0036] S300: Obtaining a predetermined parameter update times threshold;

[0037] S400: updating the parameters of the discriminator D and the generator G according to the predetermined parameter update times threshold, and performing signal enhancement on the data set X and the data set V;

[0038] Specifically, the predetermined parameter update number threshold is used to set the number of times the generator G and the discriminator D are updated. If it is set to 50, the parameters of the generator G and the discriminator D are updated 50 times. The data set X and the data set V are signal enhanced by updating the parameters, and the generator after the last parameter update is used to construct the data set required for the subsequent open set recognition task of the modulated signal, that is, the signals in the noisy modulated signal data set X are sequentially input into the generator G for enhancement to obtain the noise-reduced modulated data set, and the output signals after the signals in the data set V are sequentially input into the generator G constitute the unknown modulated signal data set. Thereby, the effect of using CGAN to enhance the signal of the data set is achieved. By generating adversarial networks for signal enhancement, it is possible to perform recognition in a low signal-to-noise ratio environment with good robustness.

[0039] S500: Modeling the modulated signal through a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement to obtain a 1D-ResNet network model;

[0040] S600: Input the test set S into the 1D-ResNet network model to obtain a first recognition result.

[0041] Specifically, since there are both known modulation type signals and unknown modulation type signals in the real environment, we build a one-dimensional convolutional residual network (1D-ResNet) to model the modulated signals and construct an open set recognition dataset for modulated signals. That is, a dataset of known modulation types and a dataset of unknown modulation signals Merge to get the open set recognition data set U = {u1,u2,...,u n}, n = 1, 2, ... N, N is the number of signal samples. Then, the samples in the data set U are shuffled and divided into a training set T, a validation set E and a test set S in a ratio of 8: 1: 1. Then, a 1D-ResNet network structure is constructed, which consists of two residual modules, and the 512-dimensional modulated signal in the data set U is input. The samples in the training set T are input into the 1D-ResNet, and the 1D-ResNet is trained. After each round of training, the loss value of the model on the training set and the validation set is calculated once. When the preset number of training times is reached or the loss value of the network has not decreased for 5 consecutive times, the training process of the neural network is completed, and a trained 1D-ResNet is obtained. The test set S is input into the trained 1D-ResNet, and the recognition result obtained is the result of the first recognition. The one-dimensional convolutional residual network is used to automatically extract the features of the input modulated signal for recognition, avoiding the cumbersome steps of manually extracting signal features in the traditional method of the prior art. At the same time, by adding signals of unknown modulation types to the training data and adding the central error loss function to the fully connected layer of the network, the network is able to perform open set recognition of modulated signals, thereby improving the generalization ability of the model.

[0042] Further, such as Figure 2 As shown, in the first environment, a data set X of known modulation type and a data set V of unknown modulation type are established, and step S100 further includes:

[0043] S110: Generate a clean signal data set Y={y1,y2,...,y n}, n = 1, 2, ..., N, N is the number of signal samples;

[0044] S120: Adding Gaussian white noise with different signal-to-noise ratios and real environmental noise to the clean signal in the clean signal data set Y to obtain a noisy signal data set as the known modulation type data set X = {x1, x2, ..., x n},n=1,2,...,N;

[0045] S130: Pair the clean signal data set Y with the data set X to form a training data pair set T = {(x1, y1), (x2, y2), ..., (x n ,y n )},n=1,2,...,N;

[0046] S140: Generate two noisy modulation signal data sets of Gaussian white noise with different signal-to-noise ratios, amplitude shift keying and minimum frequency shift keying, and a real environmental noise environment as the unknown modulation type data set V = {v1, v2, ..., v m}, m = 1, 2, ..., M, M is the number of signal samples.

[0047] Specifically, the software radio software GNU-Radio is used to generate the modulated signal. GNU-Radio is a Python-based architecture platform, based on which users can write a variety of wireless applications in a software programming manner. The GNU-Radio platform is generally combined with the radio peripheral device USPR, and the transmission and reception of the radio are defined by software, thus forming a complete hardware and software communication system. The noisy modulated signal is generated by using GNU-Radio to generate 100 pure signals of seven modulation types: amplitude modulation, frequency modulation, frequency shift keying, two-phase frequency shift keying, phase frequency shift keying, amplitude shift keying, and minimum frequency shift keying. Each signal has a length of 512 points and a modulation frequency of 1kHz. Gaussian white noise of different energies is added to each pure modulated signal to expand it into a noisy signal with a signal-to-noise ratio range of -10dB to 20dB, with 1000 signals of each type. At the same time, USRP hardware devices are deployed in a real environment, with a distance of 50m between the transceiver and the transmitter. The above seven pure modulated signals are sent at the transmitter, and the modulated signals affected by environmental noise will be received at the receiver.

[0048] After adding Gaussian white noise with different signal-to-noise ratios and real environmental noise to the clean signal in the clean signal data set Y, a noisy signal data set X is obtained, and Y and X are paired to form a training data pair set T, which is used by the neural network for fitting and learning to perform network training. The data set X is used as a data set of known modulation type, and the data set V is used as a data set of unknown modulation type for open set recognition, taking into account the problem of open set recognition in a real environment, thereby improving the recognition accuracy.

[0049] Further, such as Figure 3 As shown, according to the predetermined parameter update times threshold, the parameters of the discriminator D and the generator G are updated, and the signal of the data set X and the data set V is enhanced, wherein the discriminator D and the generator G are updated with a single parameter, and step S400 further includes:

[0050] S410: Fix the parameters of the generator G and train the discriminator D;

[0051] S420: Fix the parameters of the discriminator D and train the generator G.

[0052] Specifically, the generator G consists of an encoder module and a decoder module. The encoder consists of 6 one-dimensional convolutional layers (the convolution kernel size is 6 and the step size is 2). The PReLU activation function is connected after each layer. The convolution depth gradually increases (16, 32, 64, 128, 256, 512), and the input signal length gradually decreases (512, 256, 128, 64, 32, 16). The decoder input is the sum of the output of the last layer of the encoder and a normally distributed noise vector Z. The decoder is also composed of 6 one-dimensional convolutional layers, and its convolution depth gradually decreases (256, 128, 64, 32, 16, 1). The PReLU activation function is connected after each layer, and the output signal length gradually increases (16, 32, 64, 128, 256, 512), and each layer splices the output of the corresponding layer of the encoder at the input, that is, skip connection. The structure of the discriminator is similar to that of the encoder, containing 6 one-dimensional convolutional layers with the same convolution parameters as the encoder. The generator G and the discriminator D are trained for 50 epochs using RMSprop as the optimizer (learning rate is 0.0001). The updated generator G lays the foundation for the subsequent construction of the dataset.

[0053] Further, such as Figure 4 As shown, the parameters of the generator G are fixed, and the discriminator D is trained. Step S410 further includes:

[0054] S411: Randomly select N groups of signal pairs (x n ,y n ), the noisy modulated signal x n Input the generator G to get the output G(x n );

[0055] S412: G(x n ) and the x n Spliced ​​into a two-dimensional signal (G(x n ),x n ) Input the discriminator D so that the target output of the discriminator is '0';

[0056] S413: The y n With the x n Spliced ​​into a two-dimensional signal (x n ,y n ) is input into the discriminator D so that the target output of the discriminator is '1'.

[0057] Specifically, the above training process of the discriminator D is to minimize the loss function L CGAN (D) is used to update the parameters of the discriminator D through reverse error propagation:

[0058]

[0059] where y n is the nth pure modulated signal, x n is the nth noise modulated signal, Indicates the accumulation of errors from the first modulation signal to the Nth modulation signal. G(x n ) indicates that the generator G generates a denoised modulation signal based on the n-th noisy modulation signal. D(·) indicates that the input debugging signal is judged to be a pure modulation signal or a modulation signal generated by the generator G. It means updating the parameters of the discriminator D by minimizing the loss function. The discriminator D is trained so that it can continuously improve its "discrimination ability" to distinguish between modulated signals and pure signals.

[0060] Further, such as Figure 5 As shown, the parameters of the discriminator D are fixed and the generator G is trained. Step S420 includes:

[0061] S421: The two-dimensional signal (G(x n ),x n ) Input the discriminator D, and train the parameters of the generator G by making the output of the discriminator '1';

[0062] S422: By making the output G(x n ) approximates the y n To train the parameters of the generator G.

[0063] Specifically, the two-dimensional signal (G(x n ),x n ) is given by G(x n ) and x n The above steps are equivalent to minimizing the loss function L CGAN (G) Perform error back propagation to update the parameters of the generator G:

[0064]

[0065] Where yn is the nth pure modulation signal, x n is the nth noise modulated signal, Indicates the accumulation of errors from the first modulation signal to the Nth modulation signal. G(x n ) indicates that the generator G generates a denoised modulation signal based on the n-th noisy modulation signal. D(·) indicates the output result of the input debugging signal discriminator D, λ is a weight between 0 and 1, and ||·||1 is the sum of the absolute values ​​of each element in the vector. It means updating the parameters of the generator G by minimizing the loss function. The training of the generator G enables the generator G to continuously improve its "generation ability" to achieve the purpose of confusing the discriminator D.

[0066] Further, such as Figure 6 As shown, the parameters of the discriminator D and the generator G are updated according to the predetermined parameter update times threshold, and step S400 further includes:

[0067] S430: Input the signals in the data set X into the generator G in sequence for enhancement to obtain a noise reduction modulation data set

[0068] S440: The output signals of the signals in the data set V are sequentially input into the generator G to form an unknown modulation signal data set

[0069] Specifically, the generator G is fed with a noisy modulation signal x n , capable of generating a denoised modulated signal When the generator G continuously improves its "generation ability" through training and reaches the level of confusing the discriminator D, the generator G generates the discriminator D. It can be regarded as a pure signal y n The noise reduction modulation data set can be obtained by sequentially inputting the signal in the data set X and the signal in the data set V into the trained generator G. and the unknown modulation signal data set It can achieve the effect of signal enhancement and overcome Gaussian white noise or environmental noise interference.

[0070] Further, such as Figure 7 As shown, the modulated signal is modeled by a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement to obtain a 1D-ResNet network model, and step S500 also includes:

[0071] S510: constructing a signal open set identification data set U, wherein the data set U includes signals in the data set X and the data set V after signal enhancement;

[0072] S520: Scrambling the samples in the data set U and dividing them into a training set T, a validation set E and the test set S according to a predetermined ratio;

[0073] S530: Construct 1D-ResNet network structure;

[0074] S540: Inputting the training set T into the 1D-ResNet network structure for training;

[0075] S550: Obtain a 1D-ResNet loss function by weighted summing a multi-classification cross entropy loss function and a center error loss function;

[0076] S560: Calculate the loss value of the model on the training set T and the validation set E according to the 1D-ResNet loss function;

[0077] S570: When the loss value reaches a preset number of training times or the loss value of the network does not decrease for a predetermined number of consecutive times, the 1D-ResNet network model is obtained.

[0078] Specifically, a signal open set recognition data set U is constructed, and the data set U contains both signals of known modulation types and signals of unknown modulation types, that is, it contains the signals in the data set X and the data set V after signal enhancement. The samples in the data set U are shuffled and divided into a training set T, a validation set E and a test set S in a ratio of 8:1:1. Further, the 1D-ResNet network structure is constructed, and the 1D-ResNet network structure consists of two residual modules. The convolution layer of the residual module is composed of a one-dimensional convolution layer combined with a batch normalization layer operation and a Relu activation function, and a skip connection operation is added to the corresponding convolution layer, the convolution kernel size is 5, and the step size is 3. The output of each residual module is followed by a maximum pooling operation, and the pooling window size is 2. The parameters of the fully connected layer are set to 64 fully connected neurons in the first fully connected layer, 3 neurons in the second fully connected layer, and the output fully connected layer is set to 6 output neural units, and the activation function is the softmax function. A dropout operation is also added to the fully connected layer to prevent overfitting.

[0079] The input of 1D-ResNet is the 512-dimensional modulation signal in the data set U, that is, the training set T. Then, the parameters of the 7 one-dimensional convolutional layers and 3 fully connected layers in 1D-ResNet are set, and the weighted sum of the multi-classification cross entropy loss function and the center error loss function is used as the loss function of 1D-ResNet.

[0080] The multi-classification cross entropy loss function is as follows:

[0081]

[0082] in To indicate whether the nth modulated signal belongs to the kth category, the value is 0 or 1. The probability that the nth modulated signal belongs to the kth category predicted by the network Indicates the accumulation of errors for the Kth category. (K is equal to 6, consisting of 5 known debugging categories: AM, FM, FSK, BPSK, QPSK, and the unknown category to which ASK and MSK belong: 'Unknown'), N is the number of signal samples, It means that the errors of all samples are accumulated.

[0083] The center error loss function is as follows:

[0084]

[0085] in Represents the input signal g of 1D-ResNet n , the output value of the forward propagation to the last fully connected layer, It is the category center corresponding to each category learned through error back propagation. is the Euclidean distance between two vectors, It means that the error between each sample and its category center is accumulated. By adding the center error loss function, the signals of known modulation types can be mapped to their respective category centers in the high-dimensional space, and the signals of unknown modulation types can be mapped to the category center of the 'Unknown' class in the high-dimensional space, thereby achieving the clustering effect.

[0086] Train 1D-ResNet, input the samples in the training set T into 1D-ResNet, and select the model optimization method as the adaptive optimization algorithm Adam. Set the number of network training times to minimize the loss function J 1D-ResNet To train the network parameters by reverse error propagation, as shown in the following formula.

[0087] J 1D-ResNet =J1+β·J2

[0088] Where J1 represents the classification cross entropy loss, J2 represents the center error loss, β is a weight value between 0 and 1, and min(·) represents training network parameters by minimizing the loss function. After each round of training, the loss value of the model on the training set and the validation set is calculated once. When the preset number of training times is reached or the loss value of the network has not decreased for 5 consecutive times, the training process of the neural network is completed and the trained 1D-ResNet is obtained. A one-dimensional convolutional residual network is used to automatically extract the features of the input modulated signal for recognition, avoiding the tedious steps of manually extracting signal features in traditional methods. At the same time, by adding signals of unknown modulation types to the training data and adding the center error loss function to the fully connected layer of the network, the network is able to perform open set recognition of modulated signals.

[0089] Further, such as Figure 8As shown, after inputting the test set S into the 1D-ResNet network model and obtaining the first recognition result, step S600 further includes:

[0090] S610: Obtaining real category data;

[0091] S620: Compare the first recognition result with the real category data to obtain a recognition accuracy rate.

[0092] Specifically, the real category data is mastered, and the test set S is input into the trained 1D-ResNet to obtain the recognition result. The recognition result is compared with the real category, and the statistical recognition accuracy is the modulation signal open set recognition accuracy, and the recognition accuracy is the percentage of correctly recognized test samples to test samples. The modulation signal open set recognition accuracy can be observed to evaluate the accuracy of the recognition method.

[0093] In summary, the method and system for open-set recognition of deep neural network modulated signals provided by the embodiments of the present application have the following technical effects:

[0094] 1. Since a dataset X of known modulation type and a dataset V of unknown modulation type are established in the first environment, where X = {x1, x2, ..., x n},n=1,2,...,N, V={v1,v2,...,v m},m=1,2,...,M;initialize the parameters of the generator G and the discriminator D;obtain a predetermined parameter update times threshold;update the parameters of the discriminator D and the generator G according to the predetermined parameter update times threshold,and perform signal enhancement on the data set X and the data set V;model the modulated signal through a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement to obtain a 1D-ResNet network model;input the test set S into the 1D-ResNet network model to obtain a technical solution for the first recognition result. The embodiment of the present application provides a method and system for open set recognition of modulated signals using a deep neural network, which solves the problems of being sensitive to noise, being unable to distinguish unknown modulation type signals, and having high requirements on signal quality through signal enhancement. By adding unknown modulation type signals and a loss function to the training data, open set recognition is achieved, and the recognition accuracy is improved, and the robustness and generalization ability are enhanced.

[0095] 2. The automatic extraction of input modulated signal features for identification avoids the tedious steps of manually extracting signal features in traditional methods. At the same time, by adding signals of unknown modulation types to the training data and adding a central error loss function to the fully connected layer of the network, the network is able to perform open-set identification of modulated signals.

[0096] Embodiment 2

[0097] Based on the same inventive concept as the method for identifying an open set of deep neural network modulated signals in the aforementioned embodiment, Fig. 9 As shown, an embodiment of the present application provides a deep neural network modulated signal open set recognition system, wherein the system includes:

[0098] The first establishing unit 11 is used to establish a data set X of known modulation types and a data set V of unknown modulation types in a first environment, wherein X={x1, x2, ..., x n},n=1,2,...,N, V={v1,v2,...,v m},m=1,2,...,M;

[0099] A first execution unit 12, wherein the first execution unit 12 is used to initialize parameters of the generator G and the discriminator D;

[0100] A first obtaining unit 13, wherein the first obtaining unit 13 is used to obtain a predetermined parameter update times threshold;

[0101] A second execution unit 14, the second execution unit 14 is used to update the parameters of the discriminator D and the generator G according to the predetermined parameter update times threshold, and perform signal enhancement on the data set X and the data set V;

[0102] A second obtaining unit 15, wherein the second obtaining unit 15 is used to model the modulated signal through a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement, so as to obtain a 1D-ResNet network model;

[0103] The third obtaining unit 16 is used to input the test set S into the 1D-ResNet network model to obtain a first recognition result.

[0104] Furthermore, the system comprises:

[0105] The first generating unit is used to generate a clean signal data set Y={y1, y2, ..., y n}, n = 1, 2, ..., N, N is the number of signal samples;

[0106] The second execution unit is used to add Gaussian white noise with different signal-to-noise ratios and real environmental noise to the clean signal in the clean signal data set Y to obtain a noisy signal data set as the data set X of the known modulation type = {x1, x2, ..., x n},n=1,2,...,N;

[0107] The third execution unit is used to pair the clean signal data set Y with the data set X to form a training data pair set T = {(x1, y1), (x2, y2), ..., (x n ,y n )},n=1,2,...,N;

[0108] The fourth execution unit is used to generate two noisy modulation signal data sets of Gaussian white noise with different signal-to-noise ratios, namely amplitude shift keying and minimum frequency shift keying, and a real environmental noise environment, as the data set V of the unknown modulation type = {v1, v2, ..., v m}, m = 1, 2, ..., M, M is the number of signal samples.

[0109] Furthermore, the system comprises:

[0110] A first training unit, wherein the first training unit is used to fix the parameters of the generator G and train the discriminator D;

[0111] The second training unit is used to fix the parameters of the discriminator D and train the generator G.

[0112] Furthermore, the system comprises:

[0113] A fifth execution unit, the fifth execution unit is used to randomly select N groups of signal pairs (x n ,y n ), the noisy modulated signal x n Input the generator G to get the output G(x n );

[0114] A sixth execution unit, the sixth execution unit is used to convert the G(x n ) and the x n Spliced ​​into a two-dimensional signal (G(x n ),x n ) Input the discriminator D so that the target output of the discriminator is '0';

[0115] A seventh execution unit, the seventh execution unit is used to convert the y n With the x n Spliced ​​into a two-dimensional signal (x n ,y n ) is input into the discriminator D so that the target output of the discriminator is '1'.

[0116] Furthermore, the system comprises:

[0117] An eighth execution unit, the eighth execution unit is used to convert the two-dimensional signal (G(x n ),x n ) Input the discriminator D, and train the parameters of the generator G by making the output of the discriminator '1';

[0118] A ninth execution unit, the ninth execution unit is used to make the output G(x n ) approximates the y n To train the parameters of the generator G.

[0119] Furthermore, the system comprises:

[0120] A fourth obtaining unit, the fourth obtaining unit is used to sequentially input the signals in the data set X into the generator G for enhancement to obtain a noise reduction modulation data set

[0121] A tenth execution unit, the tenth execution unit is used to form an unknown modulated signal data set by sequentially inputting the output signals of the signals in the data set V into the generator G.

[0122] Furthermore, the system comprises:

[0123] A first construction unit, the first construction unit is used to construct a signal open set identification data set U, wherein the data set U includes signals in the data set X and the data set V after signal enhancement;

[0124] An eleventh execution unit, the eleventh execution unit is used to shuffle the samples in the data set U and divide them into a training set T, a validation set E and the test set S according to a predetermined ratio;

[0125] A second construction unit, wherein the second construction unit is used to construct a 1D-ResNet network structure;

[0126] A twelfth execution unit, the twelfth execution unit is used to input the training set T into the 1D-ResNet network structure for training;

[0127] A fifth obtaining unit, wherein the fifth obtaining unit is used to obtain a 1D-ResNet loss function by weighted summing a multi-classification cross entropy loss function and a center error loss function;

[0128] A thirteenth execution unit, the thirteenth execution unit is used to calculate the loss value of the model on the training set T and the verification set E according to the 1D-ResNet loss function;

[0129] The sixth obtaining unit is used to obtain the 1D-ResNet network model when the loss value reaches a preset number of training times or the loss value of the network does not decrease for a predetermined number of consecutive times.

[0130] Furthermore, the system comprises:

[0131] a seventh obtaining unit, the seventh obtaining unit being used to obtain real category data;

[0132] An eighth obtaining unit is used to compare the first recognition result with the real category data to obtain a recognition accuracy rate.

[0133] Exemplary Electronic Devices

[0134] Reference below Fig.10 To describe the electronic device of the embodiment of the present application,

[0135] Based on the same inventive concept as the method for identifying an open set of deep neural network modulated signals in the aforementioned embodiment, the embodiment of the present application further provides a system for identifying an open set of deep neural network modulated signals, including: a processor, the processor is coupled to a memory, the memory is used to store a program, when the program is executed by the processor, the system executes the method described in any one of the first aspects

[0136] The electronic device 300 includes: a processor 302, a communication interface 303, and a memory 301. Optionally, the electronic device 300 may also include a bus architecture 304. The communication interface 303, the processor 302, and the memory 301 may be interconnected via the bus architecture 304; the bus architecture 304 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus architecture 304 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Fig.10 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0137] The processor 302 may be a CPU, a microprocessor, an ASIC, or one or more integrated circuits for controlling the execution of the program of the present application.

[0138] The communication interface 303 uses any transceiver-like system for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), wired access networks, etc.

[0139] The memory 301 may be a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, an optical disc storage (including a compressed optical disc, a laser disc, an optical disc, a digital versatile disc, a Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory may exist independently and be connected to the processor via a bus architecture 304. The memory may also be integrated with the processor.

[0140] The memory 301 is used to store computer-executable instructions for executing the solution of the present application, and the execution is controlled by the processor 302. The processor 302 is used to execute the computer-executable instructions stored in the memory 301, thereby realizing a deep neural network modulated signal open set recognition method provided in the above embodiment of the present application.

[0141] Optionally, the computer-executable instructions in the embodiments of the present application may also be referred to as application code, which is not specifically limited in the embodiments of the present application.

[0142] The embodiment of the present application provides a method for open set recognition of deep neural network modulation signals, wherein the method comprises: establishing a data set X of known modulation types and a data set V of unknown modulation types in a first environment, wherein X={x1, x2, ..., x n},n=1,2,...,N, V={v1,v2,...,v m},m=1,2,...,M; initialize the parameters of the generator G and the discriminator D; obtain a predetermined parameter update number threshold; according to the predetermined parameter update number threshold, update the parameters of the discriminator D and the generator G, and perform signal enhancement on the data set X and the data set V; model the modulated signal through a one-dimensional convolutional residual network based on the data set X and the data set V after signal enhancement, and obtain a 1D-ResNet network model; input the test set S into the 1D-ResNet network model to obtain a first recognition result.

[0143] Those of ordinary skill in the art will appreciate that the various digital numbers such as the first and second involved in the present application are only for the convenience of description, and are not used to limit the scope of the embodiments of the present application, nor do they represent the order of precedence. "And / or" describes the association relationship of the associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship. "At least one" refers to one or more. At least two refers to two or more. "At least one", "any one" or similar expressions refer to any combination of these items, including any combination of single items (individuals) or plural items (individuals). For example, at least one item (individual, kind) of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0144] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server, data center, etc. that includes one or more available media integrated. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

[0145] The various illustrative logic units and circuits described in the embodiments of the present application can be implemented or operated by a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic system, a discrete gate or transistor logic, a discrete hardware component, or the design of any combination of the above functions. The general-purpose processor can be a microprocessor, and optionally, the general-purpose processor can also be any traditional processor, controller, microcontroller or state machine. The processor can also be implemented by a combination of a computing system, such as a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a digital signal processor core, or any other similar configuration to implement.

[0146] The steps of the method or algorithm described in the embodiments of the present application can be directly embedded in hardware, a software unit executed by a processor, or a combination of the two. The software unit can be stored in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium in the art. Exemplarily, the storage medium can be connected to the processor so that the processor can read information from the storage medium and can write information to the storage medium. Optionally, the storage medium can also be integrated into the processor. The processor and the storage medium can be arranged in an ASIC, and the ASIC can be arranged in a terminal. Optionally, the processor and the storage medium can also be arranged in different components in the terminal. These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are performed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0147] Although the present application has been described in conjunction with specific features and embodiments thereof, it is obvious that various modifications and combinations may be made thereto without departing from the spirit and scope of the present application. Accordingly, this specification and the drawings are merely exemplary illustrations of the present application as defined by the appended claims, and are deemed to have covered any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is intended to include these modifications and variations.

Claims

1. A method for open set recognition of deep neural network modulated signals, wherein: The method comprises: In the first environment, a dataset X of known modulation type and a dataset V of unknown modulation type are established, wherein: , ; Initialize the parameters of the generator G and the discriminator D; Obtaining a predetermined parameter update times threshold; According to the predetermined parameter update times threshold, update the parameters of the discriminator D and the generator G, and perform signal enhancement on the data set X and the data set V; Modeling the modulated signal through a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement to obtain a 1D-ResNet network model, wherein a 1D-ResNet loss function of the 1D-ResNet network model is obtained by weighted summation of a multi-classification cross entropy loss function and a center error loss function; The test set S is input into the 1D-ResNet network model to obtain the first recognition result.

2. The method of claim 1, wherein: The establishing of a data set X of known modulation type and a data set V of unknown modulation type under the first environment includes: Generate clean signal data sets of various modulation types including AM, FM, FSK, BFSK, QFSK , N is the number of signal samples; For the clean signal dataset The pure signal in the above example is added with Gaussian white noise of different signal-to-noise ratios and real environmental noise to obtain a noisy signal data set as the data set of the known modulation type. ; The clean signal dataset Paired with the data set X to form a training data pair set ; Generate two noisy modulation signal data sets of amplitude shift keying and minimum frequency shift keying with different signal-to-noise ratios of Gaussian white noise and real environmental noise environment as the data set of the unknown modulation type , M is the number of signal samples.

3. The method of claim 2, wherein: The method of updating the parameters of the discriminator D and the generator G according to the predetermined parameter update times threshold and performing signal enhancement on the data set X and the data set V, wherein the single parameter update of the discriminator D and the generator G includes: Fix the parameters of the generator G and train the discriminator D; The parameters of the discriminator D are fixed and the generator G is trained.

4. The method of claim 3, wherein: The fixing of the parameters of the generator G and training of the discriminator D comprises: Randomly select N groups of signal pairs from the training data pair set T ( , ), the noisy modulated signal Input the generator G to get the output ; The With the Splice into two-dimensional signal Input the discriminator D so that the target output of the discriminator is '0'; The With the Splice into two-dimensional signal Input the discriminator D so that the target output of the discriminator is '1'.

5. The method of claim 4, wherein: The step of fixing the parameters of the discriminator D and training the generator G comprises: The two-dimensional signal Input the discriminator D, and train the parameters of the generator G by making the output of the discriminator '1'; By making the output of the generator G Approximation To train the parameters of the generator G.

6. The method of claim 1, wherein: The updating of parameters of the discriminator D and the generator G according to the predetermined parameter update times threshold comprises: The signals in the data set X are sequentially input into the generator G for enhancement to obtain a noise reduction modulation data set. ; The output signals of the signals in the data set V after being sequentially input into the generator G constitute an unknown modulated signal data set .

7. The method of claim 1, wherein: The method of modeling the modulated signal through a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement to obtain a 1D-ResNet network model includes: Constructing a signal open set identification data set U, wherein the data set U includes signals in the data set X and the data set V after signal enhancement; The samples in the data set U are shuffled and divided into a training set T, a validation set E and the test set S according to a predetermined ratio; Construct 1D-ResNet network structure; Input the training set T into the 1D-ResNet network structure for training; Calculate the loss value of the model on the training set T and the validation set E according to the 1D-ResNet loss function; When the loss value reaches a preset number of training times or the loss value of the network does not decrease for a predetermined number of consecutive times, the 1D-ResNet network model is obtained.

8. The method of claim 1, wherein: After inputting the test set S into the 1D-ResNet network model and obtaining the first recognition result, the method includes: Get real category data; The first recognition result is compared with the real category data to obtain a recognition accuracy rate.

9. A deep neural network modulated signal open set recognition system, wherein: The system comprises: A first establishing unit, the first establishing unit is used to establish a data set X of known modulation type and a data set V of unknown modulation type in a first environment, wherein: , ; A first execution unit, wherein the first execution unit is used to initialize parameters of the generator G and the discriminator D; A first obtaining unit, the first obtaining unit is used to obtain a predetermined parameter update times threshold; A second execution unit, the second execution unit is used to update the parameters of the discriminator D and the generator G according to the predetermined parameter update number threshold, and perform signal enhancement on the data set X and the data set V; A second obtaining unit, wherein the second obtaining unit is used to model the modulated signal through a one-dimensional convolutional residual network according to the data set X and the data set V after signal enhancement, to obtain a 1D-ResNet network model, wherein a 1D-ResNet loss function of the 1D-ResNet network model is obtained by weighted summation of a multi-classification cross entropy loss function and a center error loss function; The third obtaining unit is used to input the test set S into the 1D-ResNet network model to obtain a first recognition result.

10. A deep neural network modulated signal open set recognition system, comprising: A processor, wherein the processor is coupled to a memory, wherein the memory is used to store a program, and when the program is executed by the processor, the system executes the method according to any one of claims 1 to 8.

Images