Metamultitask network-based small sample communication modulation identification method, system and device, and medium

Abstract

The invention provides a small sample communication modulation identification method, system and device based on a meta-multi-task network, and a medium, belongs to the field of communication, and solves the problem of low modulation identification accuracy. Comprising the steps that a basic network, a full connection layer and a Softmax layer form a element module; sequentially connecting the basic networks to obtain a basic module; a first basic network in each group of extraction units in the basic module extracts features of a certain easily-confused modulation mode signal in the support set to obtain a first network parameter; the sequentially connected second basic network extracts the characteristics of all modulation mode signals in the support set; obtaining the second network parameter and sharing the first network parameter; the features of all the easily-confused modulation mode signals in the support set are extracted; the element module inputs a network parameter set into a full connection layer, and calculates a loss value through a Softmax layer output value; the loss value converges the output weight file to obtain a meta-multi-task network; obtaining a classification identification result through the test set; according to the invention, the modulation identification accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a small sample communication modulation recognition method, system, device and medium based on a meta-multitask network. Background Art

[0002] The recognition of communication signal modulation types plays a key role in cognitive radio, threat assessment, and spectrum monitoring, and has promoted researchers' research on this technology. However, in non-cooperative scenarios, due to the lack of understanding of relevant information such as unknown signal modulation methods, it is difficult to intercept a large number of high-quality and highly reliable communication modulation signal samples, which increases the cost of sample labeling; communication signal modulation recognition often faces the problem of small sample numbers, such as non-cooperative satellite communication signal modulation recognition and communication reconnaissance system modulation recognition. When the small number of samples makes it impossible for small samples to fully represent signal characteristics, the communication modulation recognition algorithm often exhibits weak generalization ability and low recognition rate.

[0003] At present, the small sample modulation recognition method based on the data level is divided into two stages: data enhancement and modulation recognition. While the data enhancement method is used to solve the problem of small sample size, the modulation recognition result is still affected by the link relationship between the two stages. At the same time, data enhancement is not unlimited. When the amount of data is very small, there is an upper limit even if the data is enhanced.

[0004] Most existing research on communication modulation recognition uses tens of thousands of data for network training. However, in actual non-cooperative scenarios, the amount of intercepted signal data is small and the cost of obtaining high-quality labeled samples is high, resulting in insufficient labeled sample data for the acquired communication modulation signals, which affects the recognition accuracy of the end-to-end network based on deep learning (DL). Since the small sample data set has a small number of samples and contains limited features, the features learned by the network cannot classify and recognize the communication modulation signals well. At the same time, due to the limited small sample features learned by the network, signals with similar modulation modes, such as QAM16 and QAM64, are easily confused during the recognition process, which further reduces the modulation recognition accuracy. Therefore, the accuracy of small sample modulation recognition is also affected by the problem of easy confusion.

[0005] In order to improve the recognition rate of communication signal modulation recognition under small sample conditions, it is necessary to carry out research on communication signal modulation recognition under small sample conditions. In order to ensure the use of modulation recognition technology in the military and civilian security fields, a communication signal modulation recognition method with high recognition rate is studied to deal with the problem of low recognition rate of small samples and easy confusion of similar modulation modes. It has theoretical value for space electromagnetic signal processing methods and has positive reference significance for improving the engineering method of communication modulation signal recognition. Summary of the invention

[0006] In view of the small amount of intercepted signal data in non-cooperative scenarios, the acquired communication modulation signal has insufficient label sample data, which leads to overfitting of network training, resulting in low recognition accuracy based on DL end-to-end network; and due to the limited small sample features learned by the network, signals with similar modulation modes are easily confused during the recognition process, resulting in a further decrease in modulation recognition accuracy. The present invention proposes a small sample communication modulation recognition method, system, device and medium based on a meta-multi-task network, which adopts the idea of ​​meta-learning to deeply mine the feature information of small sample data to improve the recognition rate of small sample communication modulation signals. Combined with the idea of ​​multi-task learning, the feature mining of easily confused modulation signals is used as an auxiliary task, and the recognition of all modulation signals is used as the main task. The auxiliary tasks optimize the main tasks to ultimately achieve the purpose of improving the modulation recognition accuracy of small sample communication signals. The technical solution disclosed in the present invention realizes end-to-end modulation recognition, reduces the degree of confusion of similar modulation modes under small sample conditions, improves the modulation recognition accuracy, and lays a solid foundation for subsequent signal detection estimation and recognition analysis.

[0007] The purpose of the present invention is specifically achieved through the following technical solutions:

[0008] The present invention discloses a small sample communication modulation recognition method based on a meta-multitask network, comprising:

[0009] Step 1: Use stratified sampling method to extract signals in the modulation recognition task data set according to a preset ratio and divide them into a training set and a test set;

[0010] Step 2: A neural network capable of extracting features from modulated signals is used as the basic network, and a base model is composed of the basic network, a Dense fully connected layer, and a Softmax layer.

[0011] Step 3: Determine the number of modulation modes that are easily confused in the modulation recognition task, and sequentially connect two times the number of basic networks to obtain a basic module Base model; in the basic module Base model, divide all the sequentially connected basic networks into two groups to obtain multiple groups of extraction units;

[0012] Step 4: Divide the training set into a support set and a target set according to a preset ratio. The first basic network in each group of extraction units is used to extract the features of a certain easily confused modulation signal in the support set to obtain the corresponding first network parameters. The second basic networks connected in sequence are used to extract the features of all modulation signals in the support set. The corresponding second network parameters are obtained, and the first network parameters are shared at the same time. The features of all easily confused modulation signals in the support set are extracted by the extraction units in the base module Base model. All the obtained network parameters are combined into a network parameter set.

[0013] Step 5, the meta-module Meta model obtains the network parameter set updated on the support set, inputs it into the Dense fully connected layer, and calculates the total loss value on the target set through the Softmax layer output value; if the total loss value does not meet the convergence conditions, iterate steps 4 and 5 until the total loss value converges, output the weight file of the meta-multi-task network MMTN, and obtain the meta-multi-task network MMTN;

[0014] Step 6: Input the test set into the meta-multi-task network MMTN to obtain the classification and recognition results. By comparing the classification and recognition results with the actual categories, relevant evaluation indicators are obtained to realize the performance analysis of the meta-multi-task network MMTN.

[0015] In step 1, the numpy.random.choice function in the python language is used to divide the training set and the test set.

[0016] In step 2, the neural network capable of extracting features of the modulated signal includes but is not limited to a CNN, LSTM or CLDNN network.

[0017] In step 3, the number of groups of extracted units is the same as the number of modulation modes that are easily confused in the modulation recognition task.

[0018] In step 5, the calculation method of the total loss value includes:

[0019]

[0020] Among them, θ0 is the initial value of the meta-parameter θ, is the total loss value, which is used to measure the quality of the initial value θ0. The initial value θ0 is optimized until minimize; B is the total batch of tasks, b is the index of a specific task in a batch of tasks; T b represents the target set of task b, Represents the cross entropy loss function categorical_crossentropy calculated in T b The loss value on ; is the support set S in task b b The updated network weight value is obtained by performing N gradient updates on Indicates that the initial network weight value is θ0 and the updated network weight value is The network model, express In T b The loss value on .

[0021] Among them, update the network weight value The calculation method is:

[0022]

[0023] Among them, S b is the support set for task b, For S b The network weight value after N-1 gradient updates, α represents the learning rate of a set of basic modules Base model and meta module Meta model, represents the gradient of the meta-parameter θ, express The gradient of the meta-parameter θ is, express In S b The loss value after N-1 updates, Represents the cross entropy loss function categorical_crossentropy calculated in S b The loss value on Indicates that the network weight value is network model.

[0024] In step 5, the calculation method of the output meta-multi-task network MMTN weight file includes:

[0025]

[0026] Among them, θ0' is the weight file, β is the learning rate of the meta-multi-task network MMTN, express The gradient of the meta-parameter θ is, for In task b, the target set T b The total loss value on .

[0027] The present invention also provides a small sample communication modulation recognition system based on a meta-multitask network, comprising:

[0028] A division unit is used to extract signals in the modulation recognition task data set according to a preset ratio by using a stratified sampling method, and divide the signals into a training set and a test set;

[0029] The meta-module building unit is used to use a neural network capable of extracting features of a modulated signal as a basic network, and the meta-module Base model is composed of a basic network, a Dense fully connected layer, and a Softmax layer;

[0030] A basic module building unit is used to determine the number of modulation modes that are easily confused in the modulation recognition task, and a basic module Base model is obtained by sequentially connecting two times the number of basic networks; in the basic module Base model, all the sequentially connected basic networks are divided into two groups to obtain multiple groups of extraction units;

[0031] The network parameter set construction module is used to divide the training set into a support set and a target set according to a preset ratio. The first basic network in each group of extraction units is used to extract the characteristics of a certain easily confused modulation signal in the support set to obtain the corresponding first network parameters; the second basic network connected in sequence is used to extract the characteristics of all modulation signals in the support set to obtain the corresponding second network parameters, and the first network parameters are shared at the same time; until the characteristics of all easily confused modulation signals in the support set are extracted by the extraction units in the base module Base model; all the obtained network parameters are combined into a network parameter set;

[0032] A meta-multi-task network construction unit is used to obtain the network parameter set updated on the support set through the meta-module Meta model, input it into the Dense fully connected layer, and calculate the loss value on the target set through the Softmax layer output value; if the loss value does not meet the convergence condition, iteratively execute the network parameter set construction module and the meta-multi-task network construction unit until the loss value converges, output the weight file of the meta-multi-task network MMTN, and obtain the meta-multi-task network MMTN;

[0033] The performance analysis unit is used to input the test set into the meta-multi-task network MMTN to obtain the classification and recognition results, and to obtain relevant evaluation indicators by comparing the classification and recognition results with the actual categories to realize the performance analysis of the meta-multi-task network MMTN.

[0034] The present invention also provides a small sample communication modulation identification device based on a meta-multitasking network, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of a small sample communication modulation identification method based on a meta-multitasking network.

[0035] The present invention also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the steps of a small sample communication modulation identification method based on a meta-multitask network are implemented.

[0036] Beneficial effects:

[0037] The present invention proposes a small sample communication modulation recognition method, system, device and medium based on meta-multi-task network, which solves the small sample problem faced by modulation mode recognition due to low communication signal interception rate and high cost of obtaining high-quality and high-reliability samples in non-cooperative scenarios; through the "learn to The idea of ​​"learn" is to deeply mine the feature information of small sample data to improve the recognition rate of small sample communication modulation signals, and solve the problem of overfitting of network training caused by insufficient sample number; and combined with the idea of ​​multi-task learning, multiple tasks are connected in series to realize the transmission and sharing of network parameters, and the feature mining of easily confused modulation signals is used as an auxiliary task, and the recognition of all modulation signals is used as the main task. The main task obtained by the second basic network shares the feature knowledge learned by the auxiliary task obtained by the first basic network, and the learnable network parameters are transferred to the downstream tasks after feature extraction through the basic network. The auxiliary tasks are carried out alternately with the main tasks, and the auxiliary tasks optimize the main tasks, strengthen the attention to the easily confused modulation methods, realize end-to-end modulation recognition, and reduce the degree of confusion of the modulation recognition methods of similar signals under small sample conditions, so as to finally achieve the purpose of improving the modulation recognition accuracy of small sample communication signals; wherein, the parallel multi-task needs to set the importance percentage according to the importance of each task, and the setting of the importance percentage is artificially stipulated, and the deviation of different settings is large, and the corresponding setting specifications cannot be accurately found. The network weight parameters of the serial multi-task disclosed in the present invention are shared and do not need to set the importance percentage. They belong to automatic optimization parameters, and the error will be smaller than that of parallel multi-tasks, and the effect will be better.

[0038] The present invention has the advantage of low signal-to-noise ratio signal recognition while improving the accuracy of communication signal modulation recognition under small sample conditions, plays a key role in achieving strong robustness and high recognition rate of the modulation algorithm, and lays a solid foundation for subsequent signal detection, estimation, and recognition and analysis; the technical solution disclosed in the present invention has theoretical value for electromagnetic signal processing methods under small sample conditions, and has positive reference significance for improving the accuracy of communication modulation signal recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention is further described in detail below based on the accompanying drawings and embodiments.

[0040] Figure 1 It is a schematic diagram of the meta-multitask network MMTN provided by the present invention;

[0041] Figure 2 The present invention provides a small sample communication modulation recognition training flow chart based on a meta-multi-task network, taking three easily confused modulation modes and six tasks as examples;

[0042] Figure 3It is a schematic diagram of a comparison curve of task strategy structure recognition accuracy provided by the present invention;

[0043] FIG4( a ) is a schematic diagram showing a comparison of the confusion matrix of structure 1 at -4 dB for the task strategy structure provided by the present invention;

[0044] FIG4( b ) is a schematic diagram showing the comparison of the confusion matrix of structure2 under -4 dB of the task strategy structure provided by the present invention;

[0045] FIG4( c ) is a schematic diagram showing the comparison of the confusion matrix of structure 3 at -4 dB for the task strategy structure provided by the present invention;

[0046] FIG4( d ) is a schematic diagram showing the comparison of the confusion matrix of structure 4 at -4 dB for the task strategy structure provided by the present invention;

[0047] Figure 5 It is a schematic diagram of the comparison curve of the accuracy of the recognition method under small samples provided by the present invention;

[0048] FIG6( a ) is a schematic diagram showing the comparison of the confusion matrix of SVM at 0 dB for the recognition method provided by the present invention;

[0049] FIG6( b ) is a schematic diagram showing the comparison of the confusion matrix of KNN at 0 dB for the recognition method provided by the present invention;

[0050] FIG6( c ) is a schematic diagram showing the comparison of the confusion matrix of the decision tree at 0 dB of the recognition method provided by the present invention;

[0051] FIG6( d ) is a schematic diagram showing the comparison of the confusion matrix of the random forest at 0 dB for the recognition method provided by the present invention;

[0052] FIG6( e ) is a schematic diagram showing the comparison of the confusion matrix of CNN at 0 dB for the recognition method provided by the present invention;

[0053] FIG6( f ) is a schematic diagram showing the comparison of the confusion matrix of LSTM at 0 dB for the recognition method provided by the present invention;

[0054] FIG6( g ) is a schematic diagram showing the comparison of the confusion matrix of CLDNN at 0 dB for the recognition method provided by the present invention;

[0055] FIG6( h ) is a schematic diagram showing the comparison of the confusion matrix of CCC-MMTN at 0 dB of the recognition method provided by the present invention. DETAILED DESCRIPTION

[0056] like Figure 1 As shown, the first embodiment of the present invention provides a small sample communication modulation recognition method based on a meta-multitask network, including:

[0057] Step 1: Use stratified sampling method to extract signals in the modulation recognition task data set according to a preset ratio and divide them into a training set and a test set;

[0058] Step 2: A neural network capable of extracting features from modulated signals is used as the basic network, and a base model is composed of the basic network, a Dense fully connected layer, and a Softmax layer.

[0059] Step 3: Determine the number of modulation modes that are easily confused in the modulation recognition task, and sequentially connect two times the number of basic networks to obtain a basic module Base model; in the basic module Base model, divide all the sequentially connected basic networks into two groups to obtain multiple groups of extraction units;

[0060] Step 4: Divide the training set into a support set and a target set according to a preset ratio. The first basic network in each group of extraction units is used to extract the features of a certain easily confused modulation signal in the support set to obtain the corresponding first network parameters. The second basic networks connected in sequence are used to extract the features of all modulation signals in the support set. The corresponding second network parameters are obtained, and the first network parameters are shared at the same time. The features of all easily confused modulation signals in the support set are extracted by the extraction units in the base module Base model. All the obtained network parameters are combined into a network parameter set.

[0061] In this step, the network learning network extracts features from small sample data through the "learn to learn" idea of ​​meta-learning. The main task obtained by the second basic network shares the feature knowledge learned by the auxiliary task obtained by the first basic network. The auxiliary task is performed alternately with the main task to strengthen attention to easily confused modulation methods, improve the confusion problem, and thus improve the recognition accuracy of easily confused modulation methods.

[0062] Step 5, the meta-module Meta model obtains the network parameter set updated on the support set, inputs it into the Dense fully connected layer, and calculates the total loss value on the target set through the Softmax layer output value; if the total loss value does not meet the convergence conditions, iterate steps 4 and 5 until the total loss value converges, output the weight file of the meta-multi-task network MMTN, and obtain the meta-multi-task network MMTN;

[0063] Step 6: Input the test set into the meta-multi-task network MMTN to obtain the classification and recognition results. By comparing the classification and recognition results with the actual categories, relevant evaluation indicators are obtained to realize the performance analysis of the meta-multi-task network MMTN.

[0064] In step 1, the numpy.random.choice function in the python language is used to divide the training set and the test set.

[0065] In step 2, the neural network capable of extracting features of the modulated signal includes but is not limited to a CNN, LSTM or CLDNN network.

[0066] In step 3, the number of groups of extracted units is the same as the number of modulation modes that are easily confused in the modulation recognition task.

[0067] In step 5, the calculation method of the total loss value includes:

[0068]

[0069] Among them, θ0 is the initial value of the meta-parameter θ, is the total loss value, which is used to measure the quality of the initial value θ0. The initial value θ0 is optimized until minimize; B is the total batch of tasks, b is the index of a specific task in a batch of tasks; T b represents the target set of task b, Represents the cross entropy loss function categorical_crossentropy calculated in T b The loss value on ; is the support set S in task b b The updated network weight value is obtained by performing N gradient updates on Indicates that the initial network weight value is θ0 and the updated network weight value is The network model, express In T b The loss value on .

[0070] Among them, update the network weight value The calculation method is:

[0071]

[0072] Among them, S b is the support set for task b, For S b The network weight value after N-1 gradient updates, α represents the learning rate of a set of basic modules Base model and meta module Meta model, represents the gradient of the meta-parameter θ, express The gradient of the meta-parameter θ is, express In S b The loss value after N-1 updates, Represents the cross entropy loss function categorical_crossentropy calculated in Sb The loss value on Indicates that the network weight value is network model.

[0073] In step 5, the calculation method of the output meta-multi-task network MMTN weight file includes:

[0074]

[0075] Among them, θ0' is the weight file, β is the learning rate of the meta-multi-task network MMTN, express The gradient of the meta-parameter θ is, for In task b, the target set T b The total loss value on .

[0076] Embodiment 2 of the present invention provides a small sample communication modulation recognition system based on a meta-multitask network, including:

[0077] A division unit is used to extract signals in the modulation recognition task data set according to a preset ratio by using a stratified sampling method, and divide the signals into a training set and a test set;

[0078] The meta-module building unit is used to use a neural network capable of extracting features of a modulated signal as a basic network, and the meta-module Base model is composed of a basic network, a Dense fully connected layer, and a Softmax layer;

[0079] A basic module building unit is used to determine the number of modulation modes that are easily confused in the modulation recognition task, and a basic module Base model is obtained by sequentially connecting two times the number of basic networks; in the basic module Base model, all the sequentially connected basic networks are divided into two groups to obtain multiple groups of extraction units;

[0080] The network parameter set construction module is used to divide the training set into a support set and a target set according to a preset ratio. The first basic network in each group of extraction units is used to extract the characteristics of a certain easily confused modulation signal in the support set to obtain the corresponding first network parameters; the second basic network connected in sequence is used to extract the characteristics of all modulation signals in the support set to obtain the corresponding second network parameters, and the first network parameters are shared at the same time; until the characteristics of all easily confused modulation signals in the support set are extracted by the extraction units in the base module Base model; all the obtained network parameters are combined into a network parameter set;

[0081] A meta-multi-task network construction unit is used to obtain the network parameter set updated on the support set through the meta-module Meta model, input it into the Dense fully connected layer, and calculate the loss value on the target set through the Softmax layer output value; if the loss value does not meet the convergence condition, iteratively execute the network parameter set construction module and the meta-multi-task network construction unit until the loss value converges, output the weight file of the meta-multi-task network MMTN, and obtain the meta-multi-task network MMTN;

[0082] The performance analysis unit is used to input the test set into the meta-multi-task network MMTN to obtain the classification and recognition results, and to obtain relevant evaluation indicators by comparing the classification and recognition results with the actual categories to realize the performance analysis of the meta-multi-task network MMTN.

[0083] Embodiment 3 of the present invention provides a small sample communication modulation identification device based on a meta-multi-task network, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of a small sample communication modulation identification method based on a meta-multi-task network.

[0084] Embodiment 4 of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the steps of a small sample communication modulation identification method based on a meta-multitask network are implemented.

[0085] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and media described in the above embodiments can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0086] In order to facilitate those skilled in the art to understand the present invention, specific examples are now provided to illustrate the technical solutions of the present invention in detail.

[0087] Take the small sample modulation recognition basic data set shown in Table 1 as an example:

[0088] Table 1

[0089]

[0090] After the dataset in Table 1 is denoised by F-CNRN, modulation recognition is performed on the dataset in Table 2:

[0091] Table 2

[0092]

[0093] The complex cycle convolutional network (CCCN) is used as the basic network for experiments.

[0094] like Figure 2As shown, taking the small sample modulation recognition basic data set shown in Table 1 as an example, the three modulation modes of 8PSK, QAM16, and QAM64 are easy to be confused;

[0095] During the MMTN training process, tasks 1, 3, and 5 use the basic network to learn the features of three easily confused signals, 8PSK, QAM16, and QAM64 (it doesn't matter which one is in which order). Tasks 2, 4, and 6 use the basic network to learn the features of the eight modulation signals in the data set, that is, the features of all modulation signals. The learning of single signal features is carried out alternately with the learning of multi-signal features, and the learned network parameters are transmitted and shared in the six tasks. Then, the Dense layer and the Softmax layer are used to flatten and classify, and the data feature learning and classification process is repeated until the network converges. Finally, the converged network weight file is saved to complete the network training. During the test process, the data set containing eight communication modulation signals is input into the trained MMTN-based network model to obtain the classification results. Finally, the classification recognition results are compared with the actual categories to obtain relevant evaluation indicators to realize the performance analysis of the modulation recognition model.

[0096] Example 1 Small Sample Communication Modulation Recognition Method Based on Meta Multi-task Network

[0097] The effectiveness of the task structure designed according to the multi-task idea in the meta-multi-task network framework is verified. Complex Circular Convolutional Nets (CCCN) is used as the basic network. The structural parameters of the basic network CCCN in the modulation recognition algorithm are kept unchanged, and the modulation recognition accuracy of the four task strategies is calculated. The first strategy is that the Base model uses one CCCN to directly extract features of 8 modulated signals; the second strategy is that the Base model uses two CCCNs to complete two tasks respectively. The first task is used as an auxiliary task to use CCCN to extract features of 8PSK, QAM16 and QAM64 signals, and the second task is used as the main task to directly extract features of 8 modulated signals using CCCN; the third strategy is that the Base model uses six CCCNs to complete six tasks respectively, among which the first, third and fifth tasks are used as auxiliary tasks to extract features of the three groups of signals "8PSK, QAM16", "8PSK, QAM64" and "QAM16, QAM64", and the second, fourth and sixth tasks are used to extract features of 8 modulated signals; the fourth strategy is also the CCC-MMTN proposed in the present invention. The Base model uses six CCCNs to complete six tasks respectively, among which the first, third and fifth tasks are used as auxiliary tasks to extract features of 8PSK, QAM16 and QAM64 signals respectively. One task only learns one signal feature, and the second, fourth and sixth tasks are used to extract features of 8 modulated signals. By comparing the recognition accuracy under four task strategies, it can be shown that the multi-task strategy designed in the small sample modulation recognition method CCC-MMTN proposed in this invention has a positive effect on improving the accuracy. Figure 3 And as shown in Table 3.

[0098] Table 3

[0099]

[0100] from Figure 3It can be seen from the recognition rate curve that when the SNR is -20 to -2dB, the recognition accuracy of CCC-MMTN represented by the red line is significantly higher than that of the recognition network under the other three task strategies. Table 3 shows that the recognition accuracy of structure4 at -20 to -2dB and 0 to 18dB is 71.83% and 82.33% respectively, and the difference is the smallest among the four structures, 11.05%, indicating that the structure has strong robustness in dealing with different noise intensities. At 0 to 18dB, the recognition accuracy of structure4 is 82.88%, which is 0.59% higher than structure1. This is because the auxiliary task of CCC-MMTN performs separate feature learning on the three easily confused modulation signals, which is equivalent to adding an attention mechanism to these three signals, so that the recognition effect is improved compared with structure1. The recognition rate of structure4 is 0.75% and 0.33% lower than that of structure2 and structure3 at 0-18dB, but the Acc is 3.4%, 2.2% and 1.34%, 0.93% higher at -20--2dB and -20-18dB, respectively. This shows that the multi-task strategy in CCC-MMTN has the advantage of low signal-to-noise ratio signal recognition while improving the learning effect of small sample features, which plays a key role in achieving strong robustness and high recognition rate of modulation algorithm.

[0101] In order to verify that the multi-task strategy proposed in the present invention has the characteristic of reducing the confusion of similar modulation modes, the modulation recognition confusion matrices under the four task strategies when the signal-to-noise ratio is -4dB are listed. Figure 4 (a) is the confusion matrix of structure1. It can be seen that most of the QAM16 signals are mistakenly identified as QAM64. Figure 4 (b) is the confusion matrix of structure2. It can be seen that the correct recognition rate of 8PSK has been improved compared with structure1. Although the proportion of QAM16 mistakenly identified as QAM64 has decreased, the proportion of mistakenly identified as 8PSK has increased. At the same time, the proportion of QAM64 mistakenly identified as 8PSK has also increased. Figure 4 (c) is the confusion matrix of structure3. It can be seen that the proportion of 8PSK and QPSK mistakenly identified as CPFSK has decreased significantly, the proportion of QAM16 mistakenly identified as QAM64 has decreased, but the proportion of mistakenly identified as 8PSK has increased. The construction of auxiliary tasks of structure2 and structure3 causes new confusion in the main task. This is because the auxiliary task of simultaneously learning features of two or three modulated signals for small sample data cannot well learn the difference features of easily confused signals, but instead causes the confusion to worsen. Therefore, the present invention proposes a multi-task strategy represented by structure4. It can be seen from Figure 4(a) and Figure 4(d) that although QAM16 will still be mistakenly identified as QAM64, the error rate has dropped significantly, and the correct recognition rate of 8PSK has also increased significantly. This shows that the feature extraction and learning of a single signal by the auxiliary task is similar to adding an attention mechanism, which can effectively improve the problem of easy confusion in the recognition of similar signals under small sample conditions.

[0102] Example 2: Small Sample Communication Modulation Recognition Method Based on Meta-Multi-Task Network

[0103] In order to further verify the effectiveness of the small sample communication modulation recognition method proposed in this paper, the modulation recognition effect of CCC-MMTN is compared with seven other recognition methods, including model-driven methods: SVM, KNN, decision tree and random forest, and data-driven methods: CNN, LSTM and CLDNN. Figure 5 And as shown in Table 4.

[0104] Table 4

[0105]

[0106] pass Figure 5It can be seen from the recognition rate curve that the CCC-MMTN modulation recognition method shown by the red line has a significantly higher accuracy rate than the other 7 recognition methods under small sample conditions. As shown in Table 4, KNN, as a model-driven modulation recognition method for dealing with small sample problems, has the highest recognition accuracy among model-driven methods, with an overall recognition rate of 67.68%. The overall Acc of the modulation recognition algorithm proposed in the present invention is 77.35% at -20 to 18dB, which is 9.67% higher than that of KNN, verifying the effectiveness of CCC-MMTN in solving small sample modulation recognition. In Table 4, the effect of modulation recognition using CNN, LSTM, and CLDNN is the worst, which further illustrates that the effect of the data-driven recognition method is affected by the number of samples, and indirectly verifies the advantages of the data-driven end-to-end communication signal modulation recognition method CCC-MMTN proposed in the present invention in dealing with small sample data sets. The experimental results show that CCC-MMTN can effectively solve the classification and recognition problems of small sample communication modulation signals and improve the recognition accuracy.

[0107] In order to further verify the effectiveness of the small sample modulation recognition method proposed in the present invention, the confusion matrices of the eight recognition methods at 0 dB in Table 4 are visually compared. Figure 6(a) to Figure 6(h) As shown. Figure 6(a) to Figure 6(h) It can be clearly seen that although CCC-MMTN still has confusion problems, the diagonal color is the darkest, indicating that the recognition effect is the best. The KNN confusion matrix shown in Figure 6(b) has the clearest diagonal color except for Figure 6(h), indicating that KNN has the potential to solve the small sample problem, but compared with the CCC-MMTN modulation recognition method proposed in the present invention, new confusion situations have appeared in 8PSK and QPSK. The color distribution of the confusion matrix in Figure 6(a)(c)(d)(e)(f)(g) is messy, indicating that the recognition effect is very poor and a large number of misrecognitions occur, indicating that these recognition methods are seriously disturbed by the small sample bottleneck.

[0108] Beneficial effects:

[0109] The present invention proposes a small sample communication modulation recognition method, system, device and medium based on meta-multi-task network, which solves the small sample problem faced by modulation mode recognition due to low communication signal interception rate and high cost of obtaining high-quality and high-reliability samples in non-cooperative scenarios; through the "learn to The idea of ​​"learn" is to deeply mine the feature information of small sample data to improve the recognition rate of small sample communication modulation signals, and solve the problem of overfitting of network training caused by insufficient sample number; and combined with the idea of ​​multi-task learning, multiple tasks are connected in series to realize the transmission and sharing of network parameters, and the feature mining of easily confused modulation signals is used as an auxiliary task, and the recognition of all modulation signals is used as the main task. The main task obtained by the second basic network shares the feature knowledge learned by the auxiliary task obtained by the first basic network, and the learnable network parameters are transferred to the downstream tasks after feature extraction through the basic network. The auxiliary tasks are carried out alternately with the main tasks, and the auxiliary tasks optimize the main tasks, strengthen the attention to the easily confused modulation methods, realize end-to-end modulation recognition, and reduce the degree of confusion of the modulation recognition methods of similar signals under small sample conditions, so as to finally achieve the purpose of improving the modulation recognition accuracy of small sample communication signals; wherein, the parallel multi-task needs to set the importance percentage according to the importance of each task, and the setting of the importance percentage is artificially stipulated, and the deviation of different settings is large, and the corresponding setting specifications cannot be accurately found. The network weight parameters of the serial multi-task disclosed in the present invention are shared and do not need to set the importance percentage. They belong to automatic optimization parameters, and the error will be smaller than that of parallel multi-tasks, and the effect will be better.

[0110] The present invention has the advantage of low signal-to-noise ratio signal recognition while improving the accuracy of communication signal modulation recognition under small sample conditions, plays a key role in achieving strong robustness and high recognition rate of the modulation algorithm, and lays a solid foundation for subsequent signal detection, estimation, and recognition and analysis; the technical solution disclosed in the present invention has theoretical value for electromagnetic signal processing methods under small sample conditions, and has positive reference significance for improving the accuracy of communication modulation signal recognition.

[0111] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A small sample communication modulation recognition method based on meta-multitask network, characterized in that: include: Step 1: Use stratified sampling method to extract signals in the modulation recognition task data set according to a preset ratio and divide them into a training set and a test set; Step 2: A neural network capable of extracting features from modulated signals is used as the basic network, and a base model is composed of the basic network, a Dense fully connected layer, and a Softmax layer. Step 3: Determine the number of modulation modes that are easily confused in the modulation recognition task, and sequentially connect two times the number of basic networks to obtain a basic module Base model; in the basic module Base model, divide all the sequentially connected basic networks into two groups to obtain multiple groups of extraction units; Step 4: Divide the training set into a support set and a target set according to a preset ratio. The first basic network in each group of extraction units is used to extract the features of a certain easily confused modulation signal in the support set to obtain the corresponding first network parameters. The second basic networks connected in sequence are used to extract the features of all modulation signals in the support set. The corresponding second network parameters are obtained, and the first network parameters are shared at the same time. The features of all easily confused modulation signals in the support set are extracted by the extraction units in the base module Base model. All the obtained network parameters are combined into a network parameter set. Step 5: The meta-module Meta model obtains the network parameter set updated on the support set, inputs it into the Dense fully connected layer, and calculates the total loss value on the target set through the Softmax layer output value; If the total loss value does not meet the convergence condition, iterate steps 4 and 5 until the total loss value converges, output the weight file of the meta-multi-task network MMTN, and obtain the meta-multi-task network MMTN; Step 6: Input the test set into the meta-multi-task network MMTN to obtain the classification and recognition results. By comparing the classification and recognition results with the actual categories, relevant evaluation indicators are obtained to realize the performance analysis of the meta-multi-task network MMTN.

2. A small sample communication modulation recognition method based on meta-multitask network as claimed in claim 1, characterized in that: In step 1, the numpy.random.choice function in the python language is used to divide the training set and the test set.

3. The small sample communication modulation recognition method based on meta-multitask network as claimed in claim 1, characterized in that: In step 2, the neural network capable of extracting features of the modulated signal includes but is not limited to a CNN, LSTM or CLDNN network.

4. The small sample communication modulation recognition method based on meta-multitask network as claimed in claim 1, characterized in that: In step 3, the number of groups of extracted units is the same as the number of modulation modes that are easily confused in the modulation recognition task.

5. The small sample communication modulation recognition method based on meta-multitask network as claimed in claim 1, characterized in that: In step 5, the calculation method of the total loss value includes: Among them, θ0 is the initial value of the meta-parameter θ, is the total loss value, which is used to measure the quality of the initial value θ0. The initial value θ0 is optimized until minimize; B is the total batch of tasks, b is the index of a specific task in a batch of tasks; T b represents the target set of task b, Represents the cross entropy loss function categorical_crossentropy calculated in T b The loss value on ; is the support set S in task b b The updated network weight value is obtained by performing N gradient updates on Indicates that the initial network weight value is θ0 and the updated network weight value is The network model, express In T b The loss value on .

6. A small sample communication modulation recognition method based on meta-multitask network as claimed in claim 5, characterized in that: Update network weights The calculation method is: Among them, S b is the support set for task b, For S b The network weight value after N-1 gradient updates, α represents the learning rate of a set of basic modules Base model and meta module Meta model, represents the gradient of the meta-parameter θ, express The gradient of the meta-parameter θ is, express In S b The loss value after N-1 updates, Represents the cross entropy loss function categorical_crossentropy calculated in S b The loss value on Indicates that the network weight value is network model.

7. The small sample communication modulation recognition method based on meta-multitask network as claimed in claim 6, characterized in that: In step 5, the calculation method of the output meta-multi-task network MMTN weight file includes: Among them, θ0' is the weight file, β is the learning rate of the meta-multi-task network MMTN, express The gradient of the meta-parameter θ is, for In task b, the target set T b The total loss value on .

8. A small sample communication modulation recognition system based on meta-multitask network, characterized in that: include: A division unit is used to extract signals in the modulation recognition task data set according to a preset ratio by using a stratified sampling method, and divide the signals into a training set and a test set; The meta-module building unit is used to use a neural network capable of extracting features of a modulated signal as a basic network, and the meta-module Base model is composed of a basic network, a Dense fully connected layer, and a Softmax layer; A basic module building unit is used to determine the number of modulation modes that are easily confused in the modulation recognition task, and a basic module Base model is obtained by sequentially connecting two times the number of basic networks; in the basic module Base model, all the sequentially connected basic networks are divided into two groups to obtain multiple groups of extraction units; The network parameter set construction module is used to divide the training set into a support set and a target set according to a preset ratio. The first basic network in each group of extraction units is used to extract the characteristics of a certain easily confused modulation signal in the support set to obtain the corresponding first network parameters; the second basic network connected in sequence is used to extract the characteristics of all modulation signals in the support set to obtain the corresponding second network parameters, and the first network parameters are shared at the same time; until the characteristics of all easily confused modulation signals in the support set are extracted by the extraction units in the base module Base model; all the obtained network parameters are combined into a network parameter set; A meta-multi-task network construction unit is used to obtain the network parameter set updated on the support set through the meta-module Meta model, input it into the Dense fully connected layer, and calculate the loss value on the target set through the Softmax layer output value; if the loss value does not meet the convergence condition, iteratively execute the network parameter set construction module and the meta-multi-task network construction unit until the loss value converges, output the weight file of the meta-multi-task network MMTN, and obtain the meta-multi-task network MMTN; The performance analysis unit is used to input the test set into the meta-multi-task network MMTN to obtain the classification and recognition results, and to obtain relevant evaluation indicators by comparing the classification and recognition results with the actual categories to realize the performance analysis of the meta-multi-task network MMTN.

9. A small sample communication modulation recognition device based on a meta-multitask network, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images