Method and apparatus for unsupervised recognition of radio signal modulation

Abstract

The application discloses a kind of based on the identification method and device of radio signal modulation mode of unsupervised learning, the method includes: obtaining the data of radio signal in source domain;The source domain data of radio signal is carried out data transformation and obtains the data of radio signal in transformation domain;Radio signal source domain data and transformation domain data are input to the modulation mode automatic identification model based on the training of unsupervised learning, obtain the modulation mode of radio signal, and the trained identification model is trained according to the data of training signal in source domain and in transformation domain.The method provided in the application, by inputting radio signal to the modulation mode automatic identification model based on the training of unsupervised learning, obtains the modulation mode of signal, since model is trained according to the data of signal in source domain and transformation domain, can maximize the mutual information between the learning signal in different characterization domain, and does not need additional data labeling, so as to reduce the identification cost, enhance the identification effect.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a method and apparatus for identifying unsupervised radio signal modulation patterns. Background Technology

[0002] Automatic Modulation Recognition (AMR) is a communication reconnaissance technology widely used in military, aviation, and wireless communication fields. In current popular wireless communication systems, the purpose of AMR is to identify the modulation scheme of radio signals to determine their type and nature. AMR is a crucial processing step in signal detection and demodulation in non-cooperative scenarios such as electronic reconnaissance and intelligence gathering, playing a vital role in subsequent signal demodulation and the acquisition of decoded information. AMR also has significant implications in cognitive radio, spectrum sensing, blind zone signal processing, and interference identification.

[0003] The powerful feature extraction capability of multi-layer stacked artificial neurons promotes the development of a large number of deep learning (DL) based radio signal detection and recognition. Therefore, the current modulation mode recognition method is basically based on neural network and deep learning method. The current deep learning based modulation mode recognition method is mostly based on supervised learning method. According to different modulation signal types and different signal features, different networks are selected and built, a large number of labeled data are input into the network to obtain a trained model, for new downstream tasks, the learned parameters are transferred and fine-tuned on new labeled tasks to obtain a network that can adapt to new tasks. Then a large number of labeled samples are used to train the designed deep neural network, the mapping function containing a large number of parameters is optimized by using the training samples to extract high-dimensional features from the input signal to distinguish different modulation types, thereby completing the modulation recognition of the signal. The supervised learning method is roughly divided into two categories, one is the AMR based on maximum likelihood theory (Likelihood Based-Automatic Modulation Recognition, LB-AMR), and the other is the AMR based on features (Feature Based-Automatic Modulation Recognition, FB-AMR). LB-AMR is to regard the modulation mode recognition problem as a multivariate hypothesis testing problem, usually does not need a large number of training samples, but obtains the best recognition accuracy from the perspective of Bayesian estimation, but its computational complexity is relatively high. The FB-AMR algorithm is to extract representative features with discrimination from the signal samples, such as instantaneous features, statistical features, transform domain features, etc., and classify the signals through a classifier to obtain the modulation mode of the signal. The common classifier is based on support vector machine, decision tree, neural network, clustering algorithm, etc.

[0004] Although the supervised learning based recognition method has shown superior performance in some public data sets; and compared with the traditional modulation recognition method, it can significantly improve the recognition accuracy; but when facing unknown modulation modes, transmission and reception parameters in complex environments, this method cannot effectively classify and recognize them; and the supervised learning method also faces the problem of high data labeling cost.

[0005] Timothy J. O'Shea proposed in the literature that low-level time series features learned using convolutional neural networks on raw sampled radio time series data can be used to effectively cluster many radio signal modulation types without the need for explicitly labeled training data, verifying the feasibility of modulation recognition methods based on unsupervised learning. On this basis, the modulation recognition method based on unsupervised learning currently proposed mainly uses automatic encoders, sparse automatic encoders, and restricted Boltzmann machines for unsupervised training, thereby realizing modulation recognition for unlabeled signals. However, the existing unsupervised AMR method's recognition model only uses the signal's original sequence or the signal's features in a certain representation domain for unsupervised learning, which can only recognize a limited number of modulation methods, resulting in poor recognition results.

[0006] Therefore, in the current recognition method, the method based on supervised learning needs to label the data, which is costly; the method based on unsupervised learning only learns the signal's features in one representation domain, which can only recognize a limited number of modulation methods, resulting in poor recognition results. SUMMARY

[0007] The embodiments of the present application provide a radio signal modulation method and device based on unsupervised learning, which can solve the problems of high cost and poor effect of the current signal modulation method.

[0008] In a first aspect, the embodiments of the present application provide a radio signal modulation method based on unsupervised learning, the method comprising:

[0009] obtaining source domain data of a radio signal in a source domain;

[0010] performing data transformation processing on the source domain data of the radio signal to obtain transformation domain data of the radio signal in a transformation domain;

[0011] inputting the source domain data and the transformation domain data of the radio signal into a trained modulation method automatic recognition model based on unsupervised learning to obtain the modulation method of the radio signal, the trained modulation method automatic recognition model being trained according to sample data of a radio training signal in the source domain and transformation domain data in the transformation domain, the sample data and the transformation domain data of the radio training signal both comprising data of multiple radio training signals.

[0012] In a possible implementation manner of the first aspect, the modulation method automatic recognition model is trained by a back propagation algorithm under the constraint of a double-cycle contrast loss function.

[0013] For example, the double-cycle contrast loss function is determined according to an inter-domain loss function and an intra-domain loss function.

[0014] Exemplarily, the intra-domain loss function is determined according to the sample data of the radio training signal in the source domain, and the inter-domain loss function is determined according to the sample data in the source domain and the transformed domain data in the transformed domain.

[0015] In a possible implementation of the first aspect, the intra-domain loss function is determined according to the source domain enhanced data in the first positive sample pair and the first negative sample pair of the radio training signal. The inter-domain loss function is determined according to the sample data in the source domain and the transformed domain data in the transformed domain in the second positive sample pair and the second negative sample pair.

[0016] Exemplarily, the first positive sample pair and the first negative sample pair each include first type source domain enhanced data and second type source domain enhanced data, the data enhancement method of the first type source domain enhanced data and the second type source domain enhanced data is different, the first type source domain enhanced data and the second type source domain enhanced data in the first positive sample pair are obtained according to the same radio training signal, and the first type source domain enhanced data and the second type source domain enhanced data in the first negative sample pair are obtained according to different radio training signals.

[0017] Exemplarily, the second positive sample pair and the second negative sample pair each include a group of sample data and a group of transformed domain data, the sample data and the transformed domain data in the second positive sample pair are obtained according to the same radio training signal, and the sample data and the transformed domain data in the second negative sample pair are obtained according to different radio training signals.

[0018] In a possible implementation of the first aspect, the modulation mode automatic identification model can include a source domain representation learning module, a target domain contrast learning module, and a classification module. The modulation mode automatic identification model can be trained by the following method:

[0019] The source domain representation learning module determines the first positive sample pair and the first negative sample pair according to the source domain enhanced data;

[0020] The source domain representation learning module determines the representation feature vector of the radio training signal according to the source domain enhanced data in the first positive sample pair and the first negative sample pair, respectively;

[0021] The target domain contrast learning module determines the second positive sample pair and the second negative sample pair according to the sample data of the radio training signal and the transformed domain data;

[0022] The target domain contrast learning module determines the original feature vector of the radio training signal according to the sample data in the second positive sample pair and the second negative sample pair, respectively;

[0023] The target domain contrast learning module determines the modulation feature vector of the radio training signal according to the transformed domain data in the second positive sample pair and the second negative sample pair, respectively;

[0024] The target domain contrast learning module constructs a source domain feature dictionary according to the original feature vector, where the source domain feature dictionary is used to store the original feature vector.

[0025] The target domain contrast learning module constructs a target domain feature dictionary according to the modulation feature vector, where the target domain feature dictionary is used to store the modulation feature vector.

[0026] The intra-domain loss function is determined according to the representation feature vector of the radio training signal.

[0027] The inter-domain loss function is determined according to the original feature vector and the modulation feature vector of the radio training signal.

[0028] The double-cycle contrast loss function is determined according to the intra-domain loss function and the inter-domain loss function.

[0029] The source domain representation learning module and the target domain contrast learning module are trained under the constraint of the double-cycle contrast loss function to obtain the trained source domain representation learning module and the target domain contrast learning module.

[0030] The classification module is trained according to the representation feature vector and the modulation feature vector of the radio training signal to obtain the trained classification module.

[0031] The trained modulation mode automatic identification model is obtained according to the trained source domain representation learning module, the target domain contrast learning module and the classification module.

[0032] In a possible implementation manner of the first aspect, the source domain representation learning module can include a first encoder, a second encoder, a first projection head and a second projection head. The first encoder can extract features of the radio training signal according to the first type of source domain enhanced data in the first positive sample pair and the first negative sample pair to obtain a query feature vector of the first type of source domain enhanced data. The first projection head can extract a projection of the query feature vector of the first type of source domain enhanced data to obtain a representation feature vector of the first type of source domain enhanced data. The second encoder can extract features of the radio training signal according to the second type of source domain enhanced data in the first positive sample pair and the first negative sample pair to obtain a query feature vector of the second type of source domain enhanced data. The second projection head can extract a projection of the query feature vector of the second type of source domain enhanced data to obtain a representation feature vector of the second type of source domain enhanced data.

[0033] Illustratively, the parameters of the first encoder are updated iteratively according to a back propagation method under the constraint of the double-cycle contrast loss function.

[0034] Illustratively, the parameters of the second encoder are updated gradually according to the parameters of the first encoder in the training process.

[0035] In a possible implementation manner of the first aspect, the target domain contrast learning module can include a third encoder, a fourth encoder, a third projection head, and a fourth projection head. The third encoder can extract features of the radio training signal according to the sample data in the second positive sample pair and the second negative sample pair, to obtain a query feature vector of the sample data. The third projection head can extract a projection of the query feature vector of the sample data, to obtain an original feature vector. The fourth encoder can extract features of the radio training signal according to the transformed domain data in the second positive sample pair and the second negative sample pair, to obtain a query feature vector of the transformed domain data. The fourth projection head can extract a projection of the query feature vector of the transformed domain data, to obtain a modulation feature vector.

[0036] In a possible implementation manner of the first aspect, the intra-domain loss function is determined according to similarities between the representation feature vectors in the first positive sample pair, and similarities between the representation feature vectors in the first negative sample pair.

[0037] In a possible implementation manner of the first aspect, the inter-domain loss function is determined according to the source domain-target domain loss function and the target domain-source domain loss function.

[0038] For example, the source domain-target domain loss function and the target domain-source domain loss function are determined according to similarities between the original feature vector and the modulation feature vector in the second positive sample pair, and similarities between the original feature vector and the modulation feature vector in the second negative sample pair.

[0039] According to the method provided in the present application, the modulation mode of the signal is obtained by inputting the radio signal into the automatic modulation mode recognition model based on unsupervised learning. Since the model is trained according to the data of the radio training signal in the source domain and at least one transformed domain, the mutual information between signals in different representation domains, i.e., the similarities and differences of different forms of information in different representation domains, can be maximized, and no additional data labeling is required, thereby reducing the recognition cost and enhancing the recognition effect.

[0040] In the second aspect, the embodiments of the present application provide a radio signal modulation mode recognition device based on unsupervised learning. The recognition device can include an acquisition unit and a processing unit. The processing unit can include an automatic radio signal modulation mode recognition model based on unsupervised learning.

[0041] The acquisition unit is configured to acquire a radio signal.

[0042] The processing unit is configured to perform data transformation processing on the radio signal, to obtain transformed domain data of the radio signal in a transformed domain.

[0043] The processing unit is further configured to input the radio signal and the transform domain data of the radio signal into the trained modulation mode automatic identification model based on unsupervised learning to obtain the modulation mode of the radio signal, the trained modulation mode automatic identification model being trained according to sample data of radio training signals in the source domain and the transform domain data in the transform domain, and the sample data including a plurality of radio training signals.

[0044] In a possible implementation of the second aspect, the modulation mode automatic identification model is trained by a back propagation algorithm under the constraint of a dual-cycle contrast loss function.

[0045] For example, the dual-cycle contrast loss function is determined according to an inter-domain loss function and an intra-domain loss function.

[0046] For example, the intra-domain loss function is determined according to sample data of radio training signals in the source domain, and the inter-domain loss function is determined according to the sample data in the source domain and the transform domain data in the transform domain.

[0047] In a possible implementation of the second aspect, the intra-domain loss function is determined according to source domain enhanced data in a first positive sample pair and a first negative sample pair of the radio training signals. The inter-domain loss function is determined according to sample data in the source domain and the transform domain data in the transform domain in a second positive sample pair and a second negative sample pair.

[0048] For example, the first positive sample pair and the first negative sample pair each include first source domain enhanced data and second source domain enhanced data, the data enhancement method for obtaining the first source domain enhanced data and the second source domain enhanced data is different, the first source domain enhanced data and the second source domain enhanced data in the first positive sample pair are obtained from the same radio training signal, and the first source domain enhanced data and the second source domain enhanced data in the first negative sample pair are obtained from different radio training signals.

[0049] For example, the second positive sample pair and the second negative sample pair each include a group of sample data and a group of transform domain data, the sample data and the transform domain data in the second positive sample pair are obtained from the same radio training signal, and the sample data and the transform domain data in the second negative sample pair are obtained from different radio training signals.

[0050] In a possible implementation of the second aspect, the modulation mode automatic identification model can include a source domain representation learning module, a target domain contrast learning module, and a classification module. The modulation mode automatic identification model can be trained by the following method:

[0051] The source domain representation learning module determines the first positive sample pair and the first negative sample pair according to the source domain enhanced data.

[0052] The source domain representation learning module determines a representation feature vector of the radio training signal according to source domain enhanced data in the first positive sample pair and the first negative sample pair respectively;

[0053] The target domain contrast learning module determines a second positive sample pair and a second negative sample pair according to sample data and transformed domain data of the radio training signal;

[0054] The target domain contrast learning module determines an original feature vector of the radio training signal according to sample data in the second positive sample pair and the second negative sample pair respectively;

[0055] The target domain contrast learning module determines a modulation feature vector of the radio training signal according to transformed domain data in the second positive sample pair and the second negative sample pair respectively;

[0056] The target domain contrast learning module constructs a source domain feature dictionary according to the original feature vector, wherein the source domain feature dictionary is used to store the original feature vector;

[0057] The target domain contrast learning module constructs a target domain feature dictionary according to the modulation feature vector, wherein the target domain feature dictionary is used to store the modulation feature vector;

[0058] A domain-intra loss function is determined according to the representation feature vector of the radio training signal;

[0059] A domain-inter loss function is determined according to the original feature vector and the modulation feature vector of the radio training signal;

[0060] A double-cycle contrast loss function is determined according to the domain-intra loss function and the domain-inter loss function;

[0061] Under the constraint of the double-cycle contrast loss function, the source domain representation learning module and the target domain contrast learning module are trained to obtain trained source domain representation learning module and target domain contrast learning module;

[0062] A classification module is trained according to the representation feature vector and the modulation feature vector of the radio training signal to obtain a trained classification module;

[0063] A trained modulation mode automatic identification model is obtained according to the trained source domain representation learning module, the target domain contrast learning module and the classification module.

[0064] In a possible implementation manner of the second aspect, the source domain representation learning module can include a first encoder, a second encoder, a first projection head, and a second projection head. The first encoder can be configured to extract features of the radio training signal to obtain query feature vectors of the first type of source domain enhanced data according to the first type of source domain enhanced data in the first positive sample pair and the first negative sample pair. The first projection head can be configured to extract projections of the query feature vectors of the first type of source domain enhanced data to obtain representation feature vectors of the first type of source domain enhanced data. The second encoder can be configured to extract features of the radio training signal to obtain query feature vectors of the second type of source domain enhanced data according to the second type of source domain enhanced data in the first positive sample pair and the first negative sample pair. The second projection head can extract projections of the query feature vectors of the second type of source domain enhanced data to obtain representation feature vectors of the second type of source domain enhanced data.

[0065] For example, the parameters of the first encoder are iteratively updated according to a back propagation method under the constraint of the double-loop contrastive loss function.

[0066] For example, the parameters of the second encoder are gradually updated according to the parameters of the first encoder in the training process.

[0067] In a possible implementation manner of the second aspect, the target domain contrastive learning module can include a third encoder, a fourth encoder, a third projection head, and a fourth projection head. The third encoder can be configured to extract features of the radio training signal to obtain query feature vectors of the sample data according to the sample data in the second positive sample pair and the second negative sample pair. The third projection head can be configured to extract projections of the query feature vectors of the sample data to obtain original feature vectors. The fourth encoder can be configured to extract features of the radio training signal to obtain query feature vectors of the transformed domain data according to the transformed domain data in the second positive sample pair and the second negative sample pair. The fourth projection head can be configured to extract projections of the query feature vectors of the transformed domain data to obtain modulation feature vectors.

[0068] In a possible implementation manner of the second aspect, the intra-domain loss function is determined according to similarities between the representation feature vectors in the first positive sample pair and similarities between the representation feature vectors in the first negative sample pair.

[0069] In a possible implementation manner of the second aspect, the inter-domain loss function is determined according to a source domain-target domain loss function and a target domain-source domain loss function.

[0070] For example, the source domain-target domain loss function and the target domain-source domain loss function are determined according to similarities between the original feature vectors and the modulation feature vectors in the second positive sample pair and similarities between the original feature vectors and the modulation feature vectors in the second negative sample pair.

[0071] It can be understood that the beneficial effects of the second aspect described above can be seen in the relevant description of the first aspect described above, and will not be repeated here.

[0072] The beneficial effects of the embodiment of the present application compared with the prior art are: according to the method provided by the present application, the modulation mode of the signal is obtained by inputting the radio signal into the modulation mode automatic identification model based on unsupervised learning, since the model is trained according to the data of the radio training signal in the source domain and at least one transformed domain, the mutual information between the signals in different representation domains can be maximized, that is, the similarities and differences of different forms of information of the signals in different representation domains, and no additional data labeling is required, thereby reducing the recognition cost and enhancing the recognition effect. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 A flowchart of a training method of a modulation mode automatic identification model based on unsupervised learning provided by the embodiment of the present application;

[0074] Figure 2 An effect diagram of data enhancement provided by the embodiment of the present application;

[0075] Figure 3 A flowchart of a recognition method of a radio signal modulation mode based on unsupervised learning provided by the embodiment of the present application;

[0076] Figure 4 A structural diagram of a recognition device of a radio signal modulation mode based on unsupervised learning provided by the embodiment of the present application;

[0077] Figure 5 A comparison diagram of recognition effect accuracy rate provided by the embodiment of the present application;

[0078] Figure 6 Another comparison diagram of recognition effect accuracy rate provided by the embodiment of the present application;

[0079] Figure 7 A visualization effect diagram of unsupervised features provided by the embodiment of the present application. DETAILED DESCRIPTION

[0080] In the following description, specific details such as specific system structures, techniques, etc. are presented in order to thoroughly understand the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits and methods are omitted to avoid unnecessary details that hinder the description of the present application.

[0081] It should be understood that the word “comprise” or variations such as “comprises” or “comprising”, when used in this specification and in the accompanying claims, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0082] It should also be understood that the term “and / or” when used in this specification and in the following claims is intended to mean one or the other or both of the associated listed items and that no combinations of one or more items included in the processes, methods, systems or computer readable media are intended to be excluded.

[0083] As used in this specification and in the appended claims, the term “if’ can be interpreted as meaning “when” or “once” or “in response to a determination” or “in response to a detection” depending on the context. Similarly, the phrase “if it is determined” or “if [a described condition or event] is detected” can be interpreted as meaning “once it is determined” or “in response to a determination” or “once [the described condition or event] is detected” or “in response to a detection [of the described condition or event]” depending on the context.

[0084] In addition, the terms “first”, “second”, “third”, etc. are used in the description of the specification and the appended claims only for distinguishing between similar elements and do not necessarily have an implied relative importance.

[0085] Reference in the specification to “one embodiment” or “some embodiments” or “an embodiment” or “some implementations” means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. The appearances of the phrases “in one embodiment”, “in some embodiments”, “in other embodiments”, “in additional embodiments” or the like in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily referring to some, but not all embodiments, unless otherwise indicated. The terms “comprise”, “comprises”, “comprising”, “include”, “includes”, “including” and the like are synonymous with “containing” or “comprising” and are used in the sense of “including but not limited to”, unless otherwise indicated.

[0086] Figure 1 Shown is a training method of a modulation mode automatic identification model based on self-supervised learning provided by an embodiment of the application. The modulation mode automatic identification model can include a source domain representation learning module, a target domain contrast learning module, and a classification module. As an example but not limitation, the method 100 can include steps S101-S113, which are described below.

[0087] S101, the source domain representation learning module determines a first positive sample pair and a first negative sample pair according to an enhanced result of sample data in the source domain.

[0088] Exemplarily, the enhancement result of the sample data of the source domain is source domain enhanced data. The source domain enhanced data is obtained by performing data enhancement processing on the sample data of the radio training signal in the source domain by using multiple different data enhancement methods.

[0089] In some embodiments, before the source domain enhanced data is input into the source domain representation learning module, the sample data of the radio training signal in the source domain can be subjected to data enhancement processing to obtain the source domain enhanced data.

[0090] Exemplarily, the radio training signal can be a signal sample in the RML series dataset.

[0091] The RML series dataset is data generated by O'Shea et al. using GNU radio. There are 220,000 radio training signal samples in the RML2016.10a dataset, and the modulation schemes of the signals include 8PSK, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, PAW4, QAM16, QAM64, QPSK, and WBFM, a total of 11 modulation types, each modulation type contains 20 levels of signal-to-noise ratio (SNR), and the signal-to-noise ratio ranges from -20 dB to 20 dB with an interval of 2 dB. Each sample has a length of 128 and a channel number of 2, and is divided into I and Q two paths. Since the RML dataset considers the channel effects commonly seen in most radio systems during generation, including frequency offset, phase offset, sampling rate offset, multipath effect and fading, and additive white noise, etc., therefore, using it as sample data to train the modulation mode automatic identification model can better simulate the propagation characteristics of radio signals in harsh environments.

[0092] In a possible implementation, the signal samples in the RML2016.10a dataset can be divided into a training set and a test set according to a certain proportion, such as 7:3. Then the data in the training set is used as the sample data of the radio training signal in the source domain.

[0093] Specifically, for each modulation type and signal-to-noise ratio, 1000 signals are randomly divided into 700 signals for training and 300 signals for testing. Each normalized signal sample is placed in a matrix and used as input.

[0094] Exemplarily, the sample data of the radio training signal in the source domain can be the original time domain IQ data set of the radio training signal.

[0095] In an example, the sample data can be subjected to data enhancement processing by using enhancement methods such as flipping, rotation, additive Gaussian noise, small area occlusion, etc. to obtain the source domain enhanced data.

[0096] ​Exemplarily, the source domain enhanced data input into the source domain enhancement module in each round of training includes first type source domain enhanced data and second type source domain enhanced data. The first type source domain enhanced data and the second type source domain enhanced data are obtained by using different data enhancement methods.

[0097] For example, for sample data, see (a) in Figure 2 , after the above-mentioned four data enhancement processes are performed respectively, four types of source domain enhanced data are obtained, see (b), (c), (e) in Figure 2 , wherein Figure 2 (b) is obtained by flipping the sample data, Figure 2 (c) is obtained by rotating the sample data, Figure 2 (d) is obtained by adding Gaussian noise to the sample data, Figure 2 (e) is obtained by region occlusion and the like. However, only two types of source domain enhanced data are selected for training each time, and two types of source domain enhanced data are selected for training after the completion of the current round of training.

[0098] In some embodiments, before training the source domain representation learning module and the target domain contrast learning module, the training parameters can be initialized.

[0099] Exemplarily, the minimum batch size of data, the initial learning rate and the maximum number of training iterations can be set each time.

[0100] For example, the maximum number of iterations can be set as: , the minimum batch size is set to 256, the initial learning rate , and after 120 rounds of training, the learning rate is decayed by 0.1 every 40 rounds.

[0101] In some embodiments, after the source domain enhanced data is input into the source domain representation learning module, the source domain representation learning module can determine the first positive sample pair and the first negative sample pair according to the source domain enhanced data.

[0102] In one example, the first positive sample pair and the first negative sample pair both include the first type source domain enhanced data and the second type source domain enhanced data.

[0103] Exemplarily, the source domain enhanced data in the first positive sample pair is obtained according to the same radio training signal.

[0104] Exemplarily, the source domain enhanced data in the first negative sample pair is obtained according to different radio training signals.

[0105] For example, the first type source domain enhanced data of the i radio training signal can be denoted as , the second type of source domain enhanced data of the i-th radio training signal is denoted as ; then the positive sample pair can satisfy: , wherein, is the positive sample pair. If the in the positive sample pair is taken as the query sample, the negative sample pair can satisfy: , wherein, is the negative sample pair, is the second type of source domain enhanced data of the i-th radio training signal; if the in the positive sample pair is taken as the query sample, the negative sample pair can satisfy: .

[0106] In S102, the source domain representation learning module determines the representation feature vector of the radio training signal according to the source domain enhanced data in the first positive sample pair and the first negative sample pair, respectively.

[0107] In a possible implementation, the source domain representation learning module can include a first encoder, a second encoder, a first projection head, and a second projection head. The first encoder and the second encoder can perform source domain representation learning on the radio signal.

[0108] In an example, the first encoder can extract features of the first type of source domain enhanced data in the first positive sample pair and the first negative sample pair to obtain query feature vectors of the first type of source domain enhanced data. The first projection head can project and focus the query feature vectors of the first type of source domain enhanced data to obtain representation feature vectors of the first type of source domain enhanced data.

[0109] For example, the representation feature vectors of the first type of enhanced data can satisfy the following formula:

[0110]

[0111] , wherein, is the representation feature vector of the i-th radio training signal, is the first type of source domain enhanced data of the i-th radio training signal, is the first encoder, is the first projection head. In an example, the second encoder can extract features of the second type of source domain enhanced data in the first positive sample pair and the first negative sample pair to obtain query feature vectors of the second type of source domain enhanced data. The second projection head can project and focus the query feature vectors of the second type of source domain enhanced data to obtain representation feature vectors of the second type of source domain enhanced data.

[0112] , wherein,

[0113] ​​Exemplarily, the representation vector of the second type of enhanced data can satisfy the following formula:

[0114]

[0115] wherein, is a representation feature vector of the i-th radio training signal, is the second type of source domain enhanced data of the i-th radio training signal, is a second encoder, is a second projection head. S103, the target domain contrast learning module determines a second positive sample pair and a second negative sample pair according to the sample data and the transformed domain data of the radio training signal.

[0116] In some embodiments, before the sample data and the transformed domain data of the radio training signal are input into the target domain contrast learning module, the sample data of the radio training signal in the source domain can be processed by a data transformation method to obtain the transformed domain data of the radio training signal in the transformed domain.

[0117] In some embodiments, before the sample data and the transformed domain data of the radio training signal are input into the target domain contrast learning module, the sample data of the radio training signal in the source domain can be processed by a data transformation method to obtain the transformed domain data of the radio training signal in the transformed domain.

[0118] In a possible implementation, the IQ sequence of the radio signal can be processed to obtain the transformed domain data of the radio training signal in the amplitude-phase domain, the wavelet domain, and the frequency domain.

[0119] In an example, if the transformed domain data of the radio training signal in multiple transformed domains is obtained, similarly, only the sample data in the source domain and the transformed domain data in one transformed domain of the radio training signal are input into the target domain contrast learning module in each round of training.

[0120] In some embodiments, after the transformed domain data of the radio training signal is obtained, the sample data and the transformed domain data of the radio training signal can be input into the target domain contrast learning module.

[0121] Exemplarily, the second positive sample pair and the second negative sample pair each include a set of sample data and a set of transformed domain data.

[0122] Exemplarily, the sample data and the transformed domain data in the second positive sample pair are obtained according to the same radio training signal.

[0123] Exemplarily, the sample data and the transformed domain data in the second negative sample pair are obtained according to different radio training signals.

[0124] For example, the sample data of the radio training signal can be denoted as , and the transformed domain data can be denoted as . Then, the second positive sample pair is: ​.

[0125] The target domain contrast learning module determines the original feature vector of the radio training signal according to the sample data in the second positive sample pair and the second negative sample pair, respectively.

[0126] In a possible implementation, the target domain contrast learning module can include a third encoder and a third projection head.

[0127] In an example, the third encoder can extract features of the radio training signal according to the sample data in the first positive sample pair and the second negative sample pair, to obtain a query feature vector of the sample data. The third projection head can project and focus on the query feature vector of the sample data, to obtain the original feature vector of the radio training signal.

[0128] Exemplarily, the original feature vector of the radio training signal can satisfy the following formula:

[0129]

[0130] wherein, is the original feature vector of the i-th radio training signal, i is the third encoder, is the third projection head.

[0131] The target domain contrast learning module determines the modulation feature vector of the radio training signal according to the transformed domain data in the second positive sample pair and the second negative sample pair, respectively.

[0132] In a possible implementation, the target domain contrast learning module can include a fourth encoder and a fourth projection head.

[0133] In an example, the parameters of the first encoder to the fourth encoder can also be initialized when the training parameters are initialized.

[0134] In an example, the fourth encoder can extract features of the radio training signal according to the transformed domain data in the second positive sample pair and the second negative sample pair, to obtain a query feature vector of the transformed domain data. The fourth projection head can project and focus on the query feature vector of the transformed domain data, to obtain the modulation feature vector of the radio training signal.

[0135] Exemplarily, the modulation feature vector of the radio training signal can satisfy the following formula:

[0136]

[0137] wherein, is the modulation feature vector of the i-th radio training signal, i is the fourth encoder, and​ is a fourth encoder, is a fourth projection head.

[0138] For example, the features of the second positive sample pair may be extracted by the third encoder and the fourth encoder respectively, to obtain the query feature vector of the sample data in the second positive sample pair , and the query feature vector of the transformed domain in the second positive sample pair . By projecting and focusing the features through the third projection head, the original feature vector of the radio training signal may be obtained. By projecting and focusing the features through the fourth projection head, the modulation feature vector of the radio training signal may be obtained.

[0139] S106, the target domain contrast learning module constructs a source domain feature dictionary according to the original feature vector.

[0140] Illustratively, the source domain feature dictionary is used to store the original feature vector.

[0141] Illustratively, the source domain feature dictionary can satisfy the following formula:

[0142]

[0143] wherein, is the source domain feature dictionary, is the original feature vector of the th radio training signal, is a positive integer.

[0144] S107, the target domain contrast learning module constructs a target domain feature dictionary according to the modulation feature vector.

[0145] Illustratively, the target domain feature dictionary is used to store the modulation feature vector.

[0146] Illustratively, the target domain feature dictionary can satisfy the following formula:

[0147]

[0148] wherein, is the target domain feature dictionary, is the original feature vector of the th radio training signal, is the maximum value of the feature vector stored in the feature dictionary, is a positive integer.

[0149] In one example, the size of the feature dictionary can also be set when initializing the training parameters, such as setting the size of the source domain feature dictionary and the size of the target domain feature dictionary are both set to 16384.

[0150] S108, determining the intra-domain loss function according to the representation feature vectors of the radio training signals.

[0151] In some embodiments, since the representation learning of the radio training signals in the source domain representation learning module pays more attention to the characteristics of the signals in the time domain, it is expected that the features of the results enhanced according to the two different ways are closer in the positive sample pair. Therefore, the intra-domain loss function can be determined according to the similarity between the representation feature vectors in the first positive sample pair and the similarity between the representation feature vectors in the first negative sample pair.

[0152] In one possible implementation, the intra-domain loss function of single batch learning can satisfy the following formula:

[0153]

[0154] wherein, is the intra-domain loss function, is the similarity between the representation feature vectors in the first positive sample pair, is the similarity between the representation feature vectors in the first negative sample pair.

[0155] Exemplarily, the similarity between the representation feature vectors in the first positive sample pair can satisfy the following formula:

[0156]

[0157] wherein, is the similarity between the representation feature vectors in the first positive sample pair, is the representation feature vector of the first type of source domain enhanced data, is the representation feature vector of the second type of source domain enhanced data, is a temperature coefficient.

[0158] Similarly, the similarity between the representation feature vectors in the first negative sample pair can satisfy the following formula:

[0159]

[0160] wherein, is the similarity between the representation feature vectors in the first negative sample pair.

[0161] In one example, the value of the temperature coefficient can be set when initializing the training parameters of the automatic identification model before training the automatic identification model.

[0162] For example, the temperature coefficient can be 0.007.

[0163] S109, according to the original feature vector of the radio training signal and the modulation feature vector, determine the inter-domain loss function.

[0164] In some embodiments, from the perspective of the high-dimensional representation space, any one-dimensional representation domain is only a representation form of the signal. Therefore, not only the loss when traversing the target domain with the source domain as the query set, that is, the loss of the transformation domain, can be considered, but also the loss from the target domain to the source domain after exchanging the query relationship can be considered. Therefore, the inter-domain loss function can be determined according to the source domain-target domain loss function and the target domain-source domain loss function.

[0165] For example, the inter-domain loss function can satisfy the following formula:

[0166]

[0167] Wherein, is the inter-domain loss function, is the source domain-target domain loss function, is the target domain-source domain loss function.

[0168] In a possible implementation, the source domain-target domain loss function and the target domain-source domain loss function can be determined according to the similarity between the original feature vectors and the modulation feature vectors in the second positive sample pair, and the similarity between the original feature vectors and the modulation feature vectors in the negative positive sample pair.

[0169] For example, the original feature vectors and the modulation feature vectors of the second positive sample pair and the second negative sample pair can be respectively obtained from the source domain feature dictionary and the transformation domain feature dictionary.

[0170] For example, the source domain-target domain loss function can satisfy the following formula:

[0171]

[0172] Wherein, is the source domain-target domain loss function, is the similarity between the original feature vectors and the modulation feature vectors in the second positive sample pair, is the similarity between the original feature vectors and the modulation feature vectors in the negative positive sample pair.

[0173] For example, the target domain-source domain loss function can satisfy the following formula:

[0174]

[0175] Wherein, is a target domain-source domain loss function.

[0176] In one example, the similarity between the original feature vectors and the modulated feature vectors between the second positive sample pairs can satisfy the following formula:

[0177]

[0178] wherein, is the similarity between the original feature vectors and the modulated feature vectors between the second positive sample pairs, is the original feature vector, is the modulated feature vector, is a temperature coefficient, the temperature coefficient can adjust the dynamic score of the similarity, the greater the value of the temperature coefficient, the more the loss function focuses on the negative sample pairs with smaller similarity differences.

[0179] S110, determining a double-cycle contrastive loss function according to the intra-domain loss function and the inter-domain loss function.

[0180] Exemplarily, the double-cycle contrastive loss function satisfies the following formula:

[0181]

[0182] wherein, is the double-cycle contrastive loss function, is the intra-domain loss function, is the inter-domain loss function, is the source domain, is the transformation domain, , are respectively the proportion coefficients of the source domain and the transformation domain.

[0183] If the model is trained only according to the data of the source domain and one transformation domain, the double-cycle contrastive loss function can be determined according to one intra-domain loss function and one inter-domain loss function; if the model is trained according to the data of the source domain and multiple transformation domains, the double-cycle contrastive loss function can be determined according to multiple intra-domain loss functions and multiple inter-domain loss functions.

[0184] S111, the source domain representation learning module and the target domain contrastive learning module are trained under the constraint of the double-cycle contrastive loss function, to obtain the trained source domain representation learning module and the target domain contrastive learning module.

[0185] In some embodiments, the source domain representation learning module and the target domain contrastive learning module can be trained under the constraint of the double-cycle contrastive loss function by a back propagation algorithm, to obtain the trained modulation mode automatic identification model.

[0186] In a possible implementation, the first encoder of the source domain representation learning module can be updated by a back propagation algorithm under the constraint of the double-loop contrastive loss function. The parameters of the second encoder can be updated by a momentum update method based on the parameters of the first encoder.

[0187] For example, the parameters of the second encoder can satisfy the following formula:

[0188]

[0189] wherein, is the parameters of the second encoder in the n+1th training round, is a proportionality control hyperparameter, , is the parameters of the first encoder in the nth training round, is the parameters of the second encoder in the nth training round.

[0190] Optionally, when initializing the training parameters, the momentum update parameter, i.e., the value of the proportionality control hyperparameter, can be set. For example, the value is set to 0.9.

[0191] In a possible implementation, the third encoder and the fourth encoder of the target domain contrastive learning module can be updated by a back propagation algorithm under the constraint of the double-loop contrastive loss function.

[0192] For example, after the source domain representation learning module and the target domain contrastive learning module are trained, the parameters of the first encoder to the fourth encoder can be fixed and output.

[0193] In a possible implementation, the first projection head to the fourth projection head can each be composed of a fully connected layer and a ReLU activation function.

[0194] In S112, the classification module is trained according to the representation feature vector and the modulation feature vector of the radio training signal, to obtain a trained classification module.

[0195] For example, the source domain enhanced data, the transform domain data and the sample data of the radio training signal in the test set can be input to the trained source domain representation learning module and the target domain contrastive learning module, to obtain the representation feature vector and the modulation feature vector of the radio training signal. Then, the representation feature vector and the modulation feature vector of the radio training signal are spliced to obtain a spliced final feature vector. The spliced final feature vector is input to the classification module, the modulation mode of the radio training signal is identified by the classification module, and the classification module is trained to obtain a trained classification module.

[0196] Exemplarily, before training the classification module, the classification module can be set to train a total of 80 rounds, the batch size is set to 256, the initial learning rate is 0.01, and after the first 40 rounds of training, the learning rate is decayed by 0.2 every 10 rounds.

[0197] Exemplarily, the classification module can be a two-layer linear classifier, the first layer of the classifier is a fully connected layer with 128 neurons and ReLU activation. The second layer is a Softmax layer with 11 neurons. And through dropout to alleviate overfitting.

[0198] S113, according to the trained source domain representation learning module, target domain contrast learning module and learning module, a trained modulation mode automatic identification model is obtained.

[0199] Exemplarily, the trained source domain representation learning module, target domain contrast learning module and learning module are obtained, that is, the trained modulation mode automatic identification model is obtained.

[0200] According to the training method of the modulation mode automatic identification model provided by the application, the model is trained by the data of the radio training signal in the source domain and the transform domain, which can maximize the mutual information between the signals in different representation domains, that is, the similarities and differences of different forms of information in different representation domains, and the model based on self-supervised learning does not need additional data labeling, thereby reducing the recognition cost and enhancing the recognition effect.

[0201] The radio signal modulation mode recognition method based on unsupervised provided by the embodiment of the application can be applied to a supercomputer, a notebook computer and other processing devices, and the embodiment of the application does not limit the specific type of the processing device.

[0202] Figure 3 A schematic flowchart of a radio signal modulation mode recognition method based on unsupervised provided by the embodiment of the application is shown. As an example but not limitation, the method can be applied to the above processing devices. The method 300 can include steps S301-S303, which will be described below.

[0203] S301, obtaining source domain data of a radio signal in a source domain.

[0204] Exemplarily, the source domain data of the radio signal in the source domain can be the original time domain IQ data set of the obtained radio signal.

[0205] S302, performing data transform processing on the source domain data of the radio signal to obtain transform domain data of the radio signal in a transform domain.

[0206] In one example, the source domain data of the radio signal can be data-transformed based on the IQ sequence to obtain the transform domain data of the radio signal.

[0207] Exemplarily, the transform domain data can be data of the radio signal in the amplitude-phase domain, the wavelet domain, the frequency domain, etc.

[0208] Exemplarily, the transform domain of the transform domain data of the radio signal should be the same as the transform domain of the transform domain data of the radio training signal used by the modulation mode automatic identification model during training.

[0209] For example, the transform domain data of the radio training signal in the wavelet domain and the transform domain data in the frequency domain are used during training, and then the radio signal should be transformed to obtain the transform domain data of the radio signal in the frequency domain and / or the wavelet domain when identifying the radio signal.

[0210] S303, input the source domain data and the transform domain data of the radio signal into the trained modulation mode automatic identification model based on unsupervised learning to obtain the modulation mode of the radio signal.

[0211] Exemplarily, the trained modulation mode automatic identification model is obtained by training the sample data of the radio training signal in the source domain and the transform domain data in the transform domain according to the training method 100.

[0212] Exemplarily, the sample data and the transform domain data of the radio training signal both include data of multiple radio training signals

[0213] In one possible implementation, the automatic identification model can include a source domain feature learning module, a target domain contrast learning module, and a classification module.

[0214] The source domain feature learning module can be used to determine the feature vector of the radio signal according to the source domain data of the radio signal.

[0215] In one example, the source domain feature learning module can include a first encoder and a first projection head, and the parameters of the first encoder and the first projection head can be determined by the training method in method 100.

[0216] The first encoder is used to extract the features of the radio signal to obtain the query feature vector of the radio signal.

[0217] The first projection head is used to project and focus the query feature vector of the radio signal to obtain the feature vector of the radio signal.

[0218] The target domain contrast learning module can be used to determine the modulation feature vector of the radio signal according to the transform domain data of the radio signal.

[0219] In one example, the target domain contrastive learning module can include a fourth encoder and a fourth projection head. The parameters of the fourth encoder and the fourth projection head are determined by the training parameters in the method 100.

[0220] The fourth encoder is used to extract the features of the radio signal in the transform domain, to obtain the query feature vector of the radio signal transform domain data.

[0221] The fourth projection head is used to focus and project the transform domain data of the radio signal, to obtain the modulation feature vector of the radio signal.

[0222] The learning module can be used to determine the modulation mode of the radio signal according to the final feature vector of the radio signal.

[0223] In an example, the final feature vector of the radio signal is obtained by splicing the representation feature vector and the modulation feature vector of the radio signal.

[0224] According to the method provided by the application, the modulation mode of the signal is obtained by inputting the radio signal into the modulation mode automatic identification model based on unsupervised learning. Since the model is trained according to the data of the radio training signal in the source domain and the transform domain, the mutual information between the signals in different representation domains can be maximized, that is, the similarities and differences of different forms of information in different representation domains, and no additional data labeling is required, thereby reducing the recognition cost and enhancing the recognition effect.

[0225] Figure 4 An apparatus for identifying the modulation mode of a radio signal based on unsupervised learning is shown. As an example but not limitation, the apparatus 400 can include an acquisition unit 410 and a processing unit 420, and the processing unit 420 can include a modulation mode automatic identification model 421.

[0226] The acquisition unit 410 is used to acquire a radio signal.

[0227] The processing unit 420 is used to perform data transform processing on the radio signal to obtain transform domain data of the radio signal in the transform domain.

[0228] The processing unit 420 is further used to input the radio signal and the transform domain data of the radio signal into the trained modulation mode automatic identification model based on unsupervised learning, to obtain the modulation mode of the radio signal.

[0229] In an example, the trained modulation mode automatic identification model is trained according to the sample data of the radio training signal in the source domain and the transform domain data in the transform domain.

[0230] Exemplarily, the sample data includes a plurality of radio training signals.

[0231] In some embodiments, the modulation mode automatic identification model 421 can include a source domain representation learning module 421a, a target domain contrast learning module 421b, and a classification module 421c.

[0232] The source domain representation learning module 421a is configured to determine a representation feature vector of the radio signal according to source domain data of the radio signal.

[0233] The target domain contrast learning module 421b is configured to determine a modulation feature vector of the radio signal according to the transformed domain data of the radio signal.

[0234] The classification module 421c is configured to determine the modulation mode of the radio signal according to the final feature vector of the radio signal.

[0235] Exemplarily, the final feature vector of the radio signal is obtained by splicing the representation feature vector and the modulation feature vector of the radio signal.

[0236] In a possible implementation, the modulation mode automatic identification model is trained by a back propagation algorithm under the constraint of a double-loop contrast loss function.

[0237] Exemplarily, the double-loop contrast loss function is determined according to an inter-domain loss function and an intra-domain loss function.

[0238] Exemplarily, the intra-domain loss function is determined according to sample data of the radio training signal in the source domain, and the inter-domain loss function is determined according to sample data in the source domain and transformed domain data in the transformed domain.

[0239] In a possible implementation, the intra-domain loss function is determined according to source domain enhanced data in a first positive sample pair and a first negative sample pair of the radio training signal. The inter-domain loss function is determined according to sample data in the source domain and transformed domain data in the transformed domain in a second positive sample pair and a second negative sample pair.

[0240] Exemplarily, the first positive sample pair and the first negative sample pair each include first type source domain enhanced data and second type source domain enhanced data, the data enhancement method of the first type source domain enhanced data and the second type source domain enhanced data is different, the first type source domain enhanced data and the second type source domain enhanced data in the first positive sample pair are obtained according to the same radio training signal, and the first type source domain enhanced data and the second type source domain enhanced data in the first negative sample pair are obtained according to different radio training signals.

[0241] Exemplarily, the second positive sample pair and the second negative sample pair each include a set of sample data and a set of transform domain data, the sample data and the transform domain data in the second positive sample pair are obtained according to a same radio training signal, and the sample data and the transform domain data in the second negative sample pair are obtained according to different radio training signals.

[0242] In a possible implementation, the modulation mode automatic identification model can be trained by the following method:

[0243] The source domain representation learning module determines the first positive sample pair and the first negative sample pair according to the source domain enhanced data.

[0244] The source domain representation learning module respectively determines the representation feature vector of the radio training signal according to the source domain enhanced data in the first positive sample pair and the first negative sample pair.

[0245] The target domain contrast learning module determines the second positive sample pair and the second negative sample pair according to the sample data and the transform domain data of the radio training signal.

[0246] The target domain contrast learning module respectively determines the original feature vector of the radio training signal according to the sample data in the second positive sample pair and the second negative sample pair.

[0247] The target domain contrast learning module respectively determines the modulation feature vector of the radio training signal according to the transform domain data in the second positive sample pair and the second negative sample pair.

[0248] The target domain contrast learning module constructs a source domain feature dictionary according to the original feature vector, where the source domain feature dictionary is used to store the original feature vector.

[0249] The target domain contrast learning module constructs a target domain feature dictionary according to the modulation feature vector, where the target domain feature dictionary is used to store the modulation feature vector.

[0250] The intra-domain loss function is determined according to the representation feature vector of the radio training signal.

[0251] The inter-domain loss function is determined according to the original feature vector and the modulation feature vector of the radio training signal.

[0252] The double-cycle contrast loss function is determined according to the intra-domain loss function and the inter-domain loss function.

[0253] The source domain representation learning module and the target domain contrast learning module are trained under the constraint of the double-cycle contrast loss function to obtain the trained source domain representation learning module and the target domain contrast learning module.

[0254] The classification module is trained according to the representation feature vector and the modulation feature vector of the radio training signal to obtain the trained classification module.

[0255] According to the trained source domain representation learning module, the target domain contrast learning module, and the classification module, a trained modulation mode automatic recognition model is obtained.

[0256] In a possible implementation, the source domain representation learning module can include a first encoder, a second encoder, a first projection head, and a second projection head.

[0257] The first encoder can be configured to extract features of the radio training signal according to the first-class source domain enhanced data in the first positive sample pair and the first negative sample pair to obtain a query feature vector of the first-class source domain enhanced data. The first projection head can be configured to extract a projection of the query feature vector of the first-class source domain enhanced data to obtain a representation feature vector of the first-class source domain enhanced data. The second encoder can be configured to extract features of the radio training signal according to the second-class source domain enhanced data in the first positive sample pair and the first negative sample pair to obtain a query feature vector of the second-class source domain enhanced data. The second projection head can be configured to extract a projection of the query feature vector of the second-class source domain enhanced data to obtain a representation feature vector of the second-class source domain enhanced data.

[0258] For example, the parameters of the first encoder are updated iteratively according to a back propagation method under the constraint of a double-loop contrast loss function.

[0259] For example, the parameters of the second encoder are updated gradually according to the parameters of the first encoder in the training process.

[0260] For example, the first projection head and the second projection head can be composed of a fully connected layer and a ReLU activation function.

[0261] In a possible implementation, the target domain contrast learning module can include a third encoder, a fourth encoder, a third projection head, and a fourth projection head.

[0262] The third encoder can be configured to extract features of the radio training signal according to the sample data in the second positive sample pair and the second negative sample pair to obtain a query feature vector of the sample data. The third projection head can be configured to extract a projection of the query feature vector of the sample data to obtain an original feature vector. The fourth encoder can be configured to extract features of the radio training signal according to the transform domain data in the second positive sample pair and the second negative sample pair to obtain a query feature vector of the transform domain data. The fourth projection head can be configured to extract a projection of the query feature vector of the transform domain data to obtain a modulation feature vector.

[0263] For example, the third projection head and the fourth projection head can be composed of a fully connected layer and a ReLU activation function.

[0264] Exemplarily, the parameters of the third encoder and the fourth encoder can be updated by a back propagation algorithm under the constraint of a double-loop contrastive loss function.

[0265] In a possible implementation, the intra-domain loss function is determined according to similarities between the representation feature vectors in the first positive sample pair and similarities between the representation feature vectors in the first negative sample pair.

[0266] In a possible implementation, the inter-domain loss function is determined according to a source domain-target domain loss function and a target domain-source domain loss function.

[0267] Exemplarily, the source domain-target domain loss function and the target domain-source domain loss function are determined according to similarities between the original feature vectors and the modulation feature vectors in the second positive sample pair and similarities between the original feature vectors and the modulation feature vectors in the second negative sample pair.

[0268] In order to better illustrate the beneficial effects of the unsupervised radio signal modulation mode recognition method provided by the present application, the following simulation experiment is designed:

[0269] Exemplarily, the radio training signal in the test set data can be recognized by the method provided by the present application, and the results are analyzed.

[0270] Figure 5 A comparison diagram of recognition accuracy provided by an embodiment of the present application is provided.

[0271] Exemplarily, the recognition results of the method provided by the present application can be compared with the methods based on SAE, MoCoV1, MoCoV2 and SimSiam and other unsupervised learning frameworks.

[0272] Exemplarily, by the method provided by the present application, the amplitude-phase domain, the wavelet domain and the frequency domain are respectively taken as the target domain of inter-domain contrastive learning, and the recognition results are simply referred to as MRC-D1-MRC-D3.

[0273] Referring to Figure 5 The method based on the MRC-D1 model, i.e. the method provided by the present application, is superior to a simple generative unsupervised sparse autoencoder SAE in that the method is based on unsupervised learning to automatically recognize the modulation mode of the features of the signals in multiple representation domains. The model framework proposed by the present application currently shows the most promising classification performance in unsupervised recognition, and is superior to the results of the currently popular unsupervised algorithm in the field of computer vision in the AMR field.

[0274] Figure 6 Another comparison diagram of recognition accuracy provided by an embodiment of the present application is provided.

[0275] Exemplarily, the recognition result of the method provided by the application can be compared with a method based on a supervised learning framework such as SP-CNN.

[0276] Referring to Figure 6 , the results of different domains and source domains combined under the MRC framework are better than the corresponding supervised models in terms of test accuracy, proving the effectiveness of the MRC unsupervised framework, and the inter-domain contrast learning can comprehensively use the dual-domain information of the source domain and the target domain.

[0277] Figure 7 The unsupervised feature visualization results of the method provided by the application in different training rounds are provided for the embodiments of the application.

[0278] Exemplarily, a high-dimensional feature vector is mapped to a two-dimensional space for visualization using (T-Distributed Stochastic Neighbor Embedding, T-SNE), and t-SNE first calculates the similarity between high-dimensional data points, uses a Gaussian distribution or a t-distribution to measure the similarity between points, and then uses the similarity information to construct a probability distribution in a two-dimensional space, so that similar high-dimensional data points have a higher probability of being adjacent in the low-dimensional space. When visualizing, the clustering results are pasted with their original categories to observe the effect of unsupervised representation learning.

[0279] Referring to Figure 7 , Figure 7 (a), (b), (c), (d), (e), (f) in (a), (b), (c), (d), (e), (f) are the visualization results of unsupervised features in the amplitude-phase domain when the signal SNR is 6db after being modulated by AM-SSB, PAM4, CPFSK, WBEMB, GFSK, QAM645, BPSK, QPSK and the like.

[0280] As can be seen from the perspective of feature visualization, the MRC proposed in the application can perform unsupervised learning of signal representation under the condition of no labeled information according to the data enhancement and domain transformation signals of the modulation signal and for classification.

[0281] According to the method provided by the application, the modulation mode of the signal is obtained by inputting the radio signal into the modulation mode automatic recognition model based on unsupervised learning. Since the model is trained according to the data of the radio training signal in the source domain and at least one transformation domain, the mutual information between the signals in different representation domains can be maximized, that is, the similarities and differences of different forms of information in different representation domains, and no additional data labeling is required, thereby reducing the recognition cost and enhancing the recognition effect.

[0282] In the above embodiments, the description of each embodiment has its own focus, and the parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0283] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

Claims

1. A method for unsupervised recognition of radio signal modulation type, characterized in that, The method comprises: obtaining source domain data of a radio signal in a source domain; performing data transformation processing on the source domain data of the radio signal to obtain transformation domain data of the radio signal in a transformation domain; inputting the source domain data and the transformation domain data of the radio signal into a trained modulation mode automatic identification model based on unsupervised learning to obtain a modulation mode of the radio signal, wherein the trained modulation mode automatic identification model is trained according to sample data of a radio training signal in a source domain and transformation domain data in a transformation domain, and the sample data of the radio training signal and the transformation domain data both comprise data of multiple radio training signals; wherein the modulation mode automatic identification model is trained by a back propagation algorithm under the constraint of a double-cycle contrast loss function; the double-cycle contrast loss function is determined according to an inter-domain loss function and an intra-domain loss function; wherein the intra-domain loss function is determined according to sample data of the radio training signal in the source domain, and the inter-domain loss function is determined according to sample data of the radio training signal in the source domain and transformation domain data in the transformation domain; the intra-domain loss function is determined according to source domain enhanced data in a first positive sample pair and a first negative sample pair of the radio training signal; the source domain enhanced data is obtained by performing data enhancement processing on the sample data in the source domain by multiple different data enhancement methods; wherein the first positive sample pair and the first negative sample pair both comprise first type source domain enhanced data and second type source domain enhanced data, the data enhancement methods for obtaining the first type source domain enhanced data and the second type source domain enhanced data are different, the first type source domain enhanced data and the second type source domain enhanced data in the first positive sample pair are obtained from the same radio training signal, and the first type source domain enhanced data and the second type source domain enhanced data in the first negative sample pair are obtained from different radio training signals; the inter-domain loss function is determined according to the sample data in the source domain and the transformation domain data in the transformation domain in a second positive sample pair and a second negative sample pair; wherein the second positive sample pair and the second negative sample pair both comprise a group of sample data and a group of transformation domain data, the sample data and the transformation domain data in the second positive sample pair are obtained from the same radio training signal, and the sample data and the transformation domain data in the second negative sample pair are obtained from different radio training signals.

2. The method of claim 1, wherein, The double-cycle contrast loss function satisfies the following formula: wherein, is the dual-cycle contrast loss function, is the intra-domain loss function, is the inter-domain loss function, is the source domain, is the transform domain, , are scaling factors for the source domain and the transform domain, respectively.

3. The method of claim 1, wherein, The modulation mode automatic identification model comprises a source domain feature learning module, a target domain contrast learning module and a classification module, and the training method of the modulation mode automatic identification model comprises: the source domain feature learning module determines the first positive sample pair and the first negative sample pair according to the source domain enhanced data; The source domain representation learning module respectively determines a representation feature vector of the radio training signal according to the source domain enhanced data in the first positive sample pair and the first negative sample pair; The target domain contrast learning module determines the second positive sample pair and the second negative sample pair according to the sample data and the transformed domain data of the radio training signal; The target domain contrast learning module respectively determines an original feature vector of the radio training signal according to the sample data in the second positive sample pair and the second negative sample pair; The target domain contrast learning module respectively determines a modulation feature vector of the radio training signal according to the transformed domain data in the second positive sample pair and the second negative sample pair; The target domain contrast learning module constructs a source domain feature dictionary according to the original feature vector, wherein the source domain feature dictionary is used to store the original feature vector; The target domain contrast learning module constructs a target domain feature dictionary according to the modulation feature vector, wherein the target domain feature dictionary is used to store the modulation feature vector; The intra-domain loss function is determined according to the representation feature vector of the radio training signal; The inter-domain loss function is determined according to the original feature vector and the modulation feature vector of the radio training signal; The double-cycle contrast loss function is determined according to the intra-domain loss function and the inter-domain loss function; The source domain representation learning module and the target domain contrast learning module are trained under the constraint of the double-cycle contrast loss function to obtain the trained source domain representation learning module and the target domain contrast learning module; The classification module is trained according to the representation feature vector and the modulation feature vector of the radio training signal to obtain the trained classification module; A trained modulation mode automatic identification model is obtained according to the trained source domain representation learning module, the target domain contrast learning module and the classification module.

4. The method of claim 3, the source domain representation learning module comprises a first encoder, a second encoder, a first projection head and a second projection head; wherein The source domain representation learning module determines the representation feature vector of the radio training signal according to the source domain enhanced data in the first positive sample pair and the first negative sample pair, comprising: The first encoder extracts features of the radio training signal according to the first type of source domain enhanced data in the first positive sample pair and the first negative sample pair to obtain query feature vectors of the first type of source domain enhanced data; The first projection head extracts projections of the query feature vectors of the first type of source domain enhanced data to obtain representation feature vectors of the first type of source domain enhanced data; The second encoder extracts features of the radio training signal according to the second type of source domain enhanced data in the first positive sample pair and the first negative sample pair to obtain query feature vectors of the second type of source domain enhanced data; The second projection head extracts projections of the query feature vectors of the second type of source domain enhanced data to obtain representation feature vectors of the second type of source domain enhanced data; The parameters of the first encoder are updated iteratively according to a back propagation method under the constraint of the dual-cycle contrast loss function, and the parameters of the second encoder are gradually updated according to the parameters of the first encoder during training.

5. The method of claim 3, wherein, The target domain contrast learning module comprises a third encoder, a fourth encoder, a third projection head, and a fourth projection head. The target domain contrast learning module determines the original feature vector of the radio training signal according to the sample data in the second positive sample pair and the second negative sample pair, respectively, including: The third encoder extracts the features of the radio training signal according to the sample data in the second positive sample pair and the second negative sample pair, to obtain the query feature vector of the sample data. The third projection head extracts the projection of the query feature vector of the sample data, to obtain the original feature vector. The target domain contrast learning module determines the modulation feature vector of the radio training signal according to the transform domain data in the second positive sample pair and the second negative sample pair, respectively, including: The fourth encoder extracts the features of the radio training signal according to the transform domain data in the second positive sample pair and the second negative sample pair, to obtain the query feature vector of the transform domain data. The fourth projection head extracts the projection of the query feature vector of the transform domain data, to obtain the modulation feature vector. The parameters of the third encoder and the fourth encoder are updated iteratively by a back propagation method under the constraint of the dual-cycle contrast loss function.

6. The method according to any one of claims 3-5, characterized in that, The intra-domain loss function is determined according to the similarity between the representation feature vectors in the first positive sample pair and the similarity between the representation feature vectors in the first negative sample pair.

7. The method of claim 6, wherein, The inter-domain loss function is determined according to a source domain-target domain loss function and a target domain-source domain loss function. The source domain-target domain loss function and the target domain-source domain loss function are determined according to the similarity between the original feature vector and the modulation feature vector in the second positive sample pair, and the similarity between the original feature vector and the modulation feature vector in the second negative sample pair.

8. An identification device based on an unsupervised radio signal modulation method, characterized in that, The device comprises a processing unit and an acquisition unit, the processing unit comprises an unsupervised learning-based radio signal modulation mode automatic identification model; The acquisition unit is configured to acquire a radio signal; The processing unit is configured to perform data transform processing on the radio signal to obtain transform domain data of the radio signal in a transform domain; The processing unit is further configured to input the radio signal and the transform domain data of the radio signal into a trained unsupervised learning-based modulation mode automatic identification model to obtain the modulation mode of the radio signal, the trained modulation mode automatic identification model being trained according to sample data of radio training signals in a source domain and transform domain data in a transform domain, the sample data comprising a plurality of radio training signals; The modulation mode automatic identification model is trained by a back propagation algorithm under the constraint of a double-cycle contrast loss function. The double-cycle contrast loss function is determined according to an inter-domain loss function and an intra-domain loss function. The intra-domain loss function is determined according to sample data of the radio training signal in the source domain, and the inter-domain loss function is determined according to sample data of the radio training signal in the source domain and transformed domain data in the transformed domain. The intra-domain loss function is determined according to source domain enhanced data in a first positive sample pair and a first negative sample pair of the radio training signal. The source domain enhanced data is obtained by performing data enhancement processing on the sample data in the source domain by using a plurality of different data enhancement methods. The first positive sample pair and the first negative sample pair each include first type source domain enhanced data and second type source domain enhanced data, the data enhancement methods for obtaining the first type source domain enhanced data and the second type source domain enhanced data are different, the first type source domain enhanced data and the second type source domain enhanced data in the first positive sample pair are obtained from the same radio training signal, and the first type source domain enhanced data and the second type source domain enhanced data in the first negative sample pair are obtained from different radio training signals. The inter-domain loss function is determined according to the sample data in the source domain and the transformed domain data in the transformed domain in a second positive sample pair and a second negative sample pair. The second positive sample pair and the second negative sample pair each include a group of the sample data and a group of the transformed domain data, the sample data and the transformed domain data in the second positive sample pair are obtained from the same radio training signal, and the sample data and the transformed domain data in the second negative sample pair are obtained from different radio training signals.

Images