Modulation identification and direction of arrival estimation method based on multi-task learning network

Abstract

The invention discloses a modulation identification and direction of arrival estimation method based on a multi-task learning network, comprising the following steps: A, simulating a modulation signal, calculating an autocorrelation matrix of the modulation signal, and making a training set and a test set; b, designing a multi-task learning network, and training the multi-task learning network by using the samples and the labels in the training set, including a pre-training stage and a parameter fine tuning stage; and C, testing the designed multi-task learning network by using samples and labels in the test set, and calculating the modulation identification accuracy and the direction of arrival estimation root-mean-square error. According to the method, the fine mode difference of the modulation signal and the array signal space feature required by direction of arrival estimation can be accurately extracted, the feature utilization rate is effectively improved, the modulation recognition accuracy is improved, and the error of direction of arrival estimation is reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and particularly relates to a modulation recognition and direction of arrival estimation method based on a multi-task learning network. BACKGROUND

[0002] Under non-cooperative wireless communication conditions, modulation recognition and direction of arrival estimation can respectively realize the identification of the modulation type of unknown modulation signals and the estimation of the direction of arrival of signals, both of which are key technical links in the field of radio signal processing. In recent years, with the rapid development of wireless communication technology, modulation recognition and direction of arrival estimation have received increasingly widespread attention in civil and military fields such as cognitive radio, network security, electronic countermeasures, etc. Among them, modulation recognition is an important step between radio signal detection and demodulation, and direction of arrival estimation provides key support for signal source positioning, channel modeling and anti-jamming techniques in wireless communication.

[0003] The technical methods of modulation recognition and direction of arrival estimation can be mainly summarized as multiple different systems: modulation recognition methods mainly include maximum likelihood function-based and time sequence signal feature-based classification, and direction of arrival estimation methods cover subspace-based algorithms, maximum likelihood-based algorithms and traditional beamforming algorithms based on array signal processing. The breakthrough progress of deep learning theory and the leap-forward improvement of computer power have simultaneously promoted the research process of the two types of technologies: for modulation recognition, deep neural networks are widely used for feature extraction and classification of time sequence signals; for direction of arrival estimation, deep learning models are gradually applied to spatial feature analysis and direction parameter estimation of array signals. However, there are still common challenges in the existing technical systems: for modulation recognition, the deep neural network is still not sufficient for extracting the modulation mode features hidden in the time sequence signal, and the feature robustness in complex noise environment is insufficient, which makes it difficult to further improve the recognition accuracy; for direction of arrival estimation, traditional algorithms are sensitive to array errors and non-ideal channel conditions, and deep learning-based methods face problems such as incomplete spatial feature representation and weak generalization ability, which restrict the breakthrough of estimation accuracy. SUMMARY

[0004] The purpose of the present application is to provide a modulation recognition and direction of arrival estimation method based on a multi-task learning network, which can accurately extract the subtle pattern differences of modulation signals and the spatial features of array signals required for direction of arrival estimation, effectively improve the feature utilization rate, and further improve the modulation recognition accuracy and reduce the error of direction of arrival estimation.

[0005] To achieve the above purpose, the present application provides a modulation recognition and direction of arrival estimation method based on a multi-task learning network, comprising the following steps:

[0006] A: simulate the modulated signal, calculate the autocorrelation matrix of the modulated signal, and make a training set and a test set;

[0007] B: design a multi-task learning network, and train the multi-task learning network using the samples and labels in the training set, including two stages of pre-training and parameter fine-tuning;

[0008] C: test the designed multi-task learning network using the samples and labels in the test set, and calculate the modulation recognition accuracy and the root mean square error of the direction of arrival estimation.

[0009] As a further scheme of the application, step A specifically comprises the following steps:

[0010] A1: calculate the modulated signal vector r(k) received by the array antenna through simulation, wherein k represents a discrete time variable and satisfies k∈{1, …, K}, and K represents the total length of the discrete time signal;

[0011] A2: calculate the autocorrelation matrix R of the received modulated signal matrix r;

[0012] A3: repeat steps A1 and A2 for N times, and use the autocorrelation matrix R and the corresponding modulation type label y Mod and the direction of arrival estimation label y DOA to form a data set containing N samples and 2N labels Wherein, the subscript n represents the serial number of the simulation times; and the data set is divided according to the ratio of 8:2 to make a training set and a test set.

[0013] As a further scheme of the application, the modulated signal vector r(k) received by the array antenna satisfies formula (1):

[0014] r(k)=a(θ)·x(k)+v(k) (1)

[0015] In formula (1), a(θ) represents a steering vector, θ represents the direction of arrival, x(k) represents the transmitted modulated signal, and v(k) represents a Gaussian noise vector. The steering vector a(θ) satisfies formula (2):

[0016] a(θ)=[1 e j2πd(sinθ) / λ …e j2π(M-1)d(sinθ)λ ] T (2)

[0017] In formula (2), j represents an imaginary unit, M represents the total number of array antennas, d represents the distance between array antennas, λ represents the wavelength of the signal, and (·) T represents the transpose operator;

[0018] The modulated signal matrix r received by the array antenna satisfies formula (3):

[0019] r = [r(1) r(2) … r(K)] (3)

[0020] The autocorrelation matrix R satisfies formula (4):

[0021] R = E(r·r H ) (4)

[0022] In formula (4), E(·) represents a statistical expectation operator, (·) H represents a conjugate transpose operator.

[0023] As a further scheme of the present application: the multi-task learning network comprises a feature extraction module and a multi-task processing module, the feature extraction module mainly realizes extraction of common features, and the feature extraction module is sequentially composed of a size transformation layer, a convolution layer, a scaled exponential linear activation unit, a convolution layer, a scaled exponential linear activation unit, two transformers, a convolution layer and a transformer; the multi-task processing module comprises two branches, one is a modulation recognition classifier, and the other is a direction of arrival estimation classifier, both of which are composed of two fully connected layers.

[0024] As a further scheme of the present application: step B specifically comprises the following steps:

[0025] 1) designing a multi-task learning network;

[0026] 2) when pre-training, inputting samples into the feature extraction module of the designed multi-task learning network;

[0027] 3) inputting the output features of the feature extraction module into the multi-task processing module of the multi-task learning network;

[0028] 4) in the multi-feature processing module, respectively passing the input features to the modulation recognition classifier and the direction of arrival estimation classifier to respectively obtain preliminary results of modulation recognition and direction of arrival estimation;

[0029] 5) calculating the cross-entropy loss of the preliminary results of modulation recognition and modulation recognition labels;

[0030] 6) calculating the root mean square error loss of the preliminary results of direction of arrival estimation and direction of arrival estimation labels;

[0031] 7) combining the cross-entropy loss and the mean square error loss into a multi-task loss function in a weighted sum manner;

[0032] 8) according to the multi-task loss function, using AdamW as the optimizer of the network parameters, setting the learning rate to 0.001, and optimizing the parameters;

[0033] 9) repeatedly perform single pre-training until the multi-task loss function in 7) converges, and the pre-training ends;

[0034] 10) when performing parameter fine-tuning, according to the modulation recognition accuracy and the root mean square error of the direction of arrival estimation after pre-training, the weights of the cross-entropy loss and the mean square error loss in the multi-task loss function are adjusted;

[0035] 11) the values of the input batch and the learning rate are adjusted to achieve acceptable modulation recognition accuracy and root mean square error of the direction of arrival estimation, and the parameter fine-tuning ends.

[0036] As a further scheme of the application: step C specifically comprises the following steps:

[0037] C1: sequentially inputting the samples in the test set into the trained feature extraction module, then inputting the features output by the feature extraction module into the multi-task processing module, and taking the outputs of the modulation recognition classifier and the direction of arrival estimation classifier in the multi-task processing module as the modulation recognition result and the direction of arrival estimation result respectively, and calculating the modulation recognition result accuracy and the root mean square error of the direction of arrival estimation according to the modulation recognition labels and the direction of arrival estimation labels corresponding to the samples in the test set;

[0038] C2: calculating the modulation recognition accuracy P Acc and the root mean square error RMSE of the direction of arrival estimation; the modulation recognition accuracy P Acc and the root mean square error RMSE of the direction of arrival estimation satisfy formula (5) and formula (6):

[0039]

[0040] In formula (5) and formula (6), N Test represents the total number of samples in the test set, represents the total number of samples in the test set that are accurately recognized modulation types, and respectively represent the nth direction of arrival estimation label in the test set and the corresponding direction of arrival estimation result.

[0041] Compared with the prior art, the application has the following beneficial effects:

[0042] In the training phase, a multi-task learning network based on a convolutional neural network and a transformer module is designed. The convolutional neural network and the transformer module have efficient long sequence modeling capability, can fully mine the complex features in the time domain, frequency domain and space domain of the signal, can accurately extract the subtle pattern differences required for modulation recognition and the spatial features required for direction of arrival estimation, can effectively improve the utilization rate of features, and thus improve the modulation recognition accuracy and reduce the error of the direction of arrival estimation.

[0043] Through the multi-task learning mechanism, two tasks of modulation recognition and direction of arrival estimation are completed at the same time by using the same multi-task learning network, parameter sharing and feature joint optimization are realized, redundant calculation required by the traditional method for separately processing the two tasks is avoided, signal processing efficiency is greatly improved, operation cost is reduced, and the demand for fast processing in a complex wireless communication scene is met. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a flowchart of the modulation recognition and direction of arrival estimation method based on the multi-task learning network of the present application.

[0045] Figure 2 is a structure diagram of the multi-task learning framework in the modulation recognition and direction of arrival estimation method based on the multi-task learning network of the present application.

[0046] Figure 3 is a modulation recognition average accuracy curve diagram of the modulation recognition and direction of arrival estimation method based on the multi-task learning network of the present application.

[0047] Figure 4 is a direction of arrival estimation root mean square error curve diagram of the modulation recognition and direction of arrival estimation method based on the multi-task learning network of the present application. DETAILED DESCRIPTION

[0048] The present application will be further described below in combination with the drawings.

[0049] As Figure 1 shown, a modulation recognition and direction of arrival estimation method based on a multi-task learning network comprises the following steps:

[0050] A: simulate a modulation signal, calculate the autocorrelation matrix of the modulation signal, and make a training set and a test set.

[0051] Further, step A specifically comprises the following steps:

[0052] A1: calculate the modulation signal vector r(k) received by the array antenna through simulation, k represents a discrete time variable and satisfies k∈{1,…,K}, K represents the total length of the discrete time of the signal;

[0053] A2: calculate the autocorrelation matrix R of the received modulation signal matrix r received by the array antenna;

[0054] A3: repeat steps A1 and A2 for N times, and use the autocorrelation matrix R and the corresponding modulation type label y Mod and the direction of arrival (DOA) estimation label y DOAA data set composed of N samples and 2N labels Wherein, the subscript n represents the serial number of the simulation times; and the data set is divided according to the ratio of 8:2 to make a training set and a test set.

[0055] Further, the modulation signal vector r(k) received by the array antenna satisfies formula (1):

[0056] r(k)=a(θ)·x(k)+v(k) (1)

[0057] In formula (1), a(θ) represents a steering vector, θ represents a direction of arrival, x(k) represents a transmitted modulation signal, v(k) represents a Gaussian noise vector, and the steering vector a(θ) satisfies formula (2):

[0058] a(θ)=[1 e j2πd(sinθ) / λ …e j2π(M-1)d(sinθ) / λ ] T (2)

[0059] In formula (2), j represents an imaginary unit, M represents the total number of array antennas, d represents the spacing of array antennas, λ represents the wavelength of a signal, and (·) T represents a transpose operator;

[0060] The modulation signal matrix r received by the array antenna satisfies formula (3):

[0061] r=[r(1) r(2)…r(K)] (3)

[0062] The autocorrelation matrix R satisfies formula (4):

[0063] R=E(r·r H ) (4)

[0064] In formula (4), E(·) represents a statistical expectation operator, and (·) H represents a conjugate transpose operator

[0065] B: Design a multi-task learning network, and train the multi-task learning network using the samples and labels in the training set, including two stages of pre-training and parameter fine-tuning.

[0066] Further, step B specifically includes the following steps:

[0067] 1) Design a multi-task learning network, such as Figure 2As shown, the multi-task learning network includes a feature extraction module and a multi-task processing module. The feature extraction module mainly realizes extraction of common features. The feature extraction module sequentially comprises a size conversion layer, a convolutional layer, a Scaled Exponential Linear Unit (SELU), a convolutional layer, a Scaled Exponential Linear Unit, two Transformers, a convolutional layer, and a Transformer. The multi-task processing module includes two branches, one is a modulation recognition classifier, and the other is a direction of arrival estimation classifier, both of which are composed of two fully connected layers.

[0068] 2) During pre-training, the sample is input to the feature extraction module of the designed multi-task learning network;

[0069] 3) The output features of the feature extraction module are input to the multi-task processing module of the multi-task learning network;

[0070] 4) In the multi-feature processing module, the input features are respectively transmitted to the modulation recognition classifier and the direction of arrival estimation classifier to obtain preliminary results of modulation recognition and direction of arrival estimation, respectively;

[0071] 5) The cross-entropy loss of the preliminary result of modulation recognition and the modulation recognition label is calculated;

[0072] 6) The root mean square error loss of the preliminary result of direction of arrival estimation and the direction of arrival estimation label is calculated;

[0073] 7) The cross-entropy loss and the mean square error loss are combined into a multi-task loss function in a weighted sum manner;

[0074] 8) According to the multi-task loss function, AdamW is used as the optimizer of the network parameters, the learning rate is set to 0.001, and the parameters are optimized;

[0075] 9) Repeat the single pre-training until the multi-task loss function in 7) converges, and the pre-training ends;

[0076] 10) During parameter fine-tuning, the weights of the cross-entropy loss and the mean square error loss in the multi-task loss function are adjusted according to the modulation recognition accuracy and the root mean square error of direction of arrival estimation after pre-training;

[0077] 11) The values of the input batch and the learning rate are adjusted to achieve acceptable modulation recognition accuracy and root mean square error of direction of arrival estimation, and the parameter fine-tuning ends.

[0078] C: Test the designed multi-task learning network using samples and labels in the test set, and calculate the modulation recognition accuracy and the root mean square error of direction of arrival estimation.

[0079] Further, step C specifically comprises the following steps:

[0080] C1: inputting the samples in the test set into the trained feature extraction module in turn, inputting the features output by the feature extraction module into the multi-task processing module, taking the outputs of the modulation recognition classifier and the direction of arrival estimation classifier in the multi-task processing module as the modulation recognition result and the direction of arrival estimation result respectively, and calculating the modulation recognition result accuracy and the direction of arrival estimation root mean square error according to the modulation recognition labels and the direction of arrival estimation labels corresponding to the samples in the test set;

[0081] C2: calculating the modulation recognition accuracy P Acc and the direction of arrival estimation root mean square error RMSE; the modulation recognition accuracy P Acc and the direction of arrival estimation root mean square error RMSE satisfy formula (5) and formula (6):

[0082]

[0083] In formula (5) and formula (6), N Test represents the total number of samples in the test set, represents the total number of samples in the test set that are accurately recognized modulation types, and respectively represent the nth direction of arrival estimation label in the test set and the corresponding direction of arrival estimation result.

[0084] The experimental results of the present application are shown in Figure 3 and Figure 4 Figure 3 is a modulation recognition accuracy curve diagram, Figure 4 is a direction of arrival estimation root mean square error curve diagram. The experimental results prove that the method proposed in the present application can realize modulation recognition and direction of arrival estimation based on a multi-task learning network.​

Claims

1. A modulation recognition and direction of arrival estimation method based on a multi-task learning network, characterized in that, The method comprises the following steps: A: simulate a modulation signal, calculate the autocorrelation matrix of the modulation signal, and make a training set and a test set; B: design a multi-task learning network, and train the multi-task learning network using samples and labels in the training set, including two stages of pre-training and parameter fine-tuning; C: test the designed multi-task learning network using samples and labels in the test set, and calculate the modulation recognition accuracy and the root mean square error of the direction of arrival estimation.

2. The modulation recognition and direction of arrival estimation method based on multi-task learning network according to claim 1, characterized in that, Step A specifically comprises the following steps: A1: calculate the modulation signal vector r(k) received by the array antenna through simulation, wherein k represents a discrete time variable and satisfies k∈{1, …, K}, and K represents the total length of the discrete time signal; A2: calculate the autocorrelation matrix R of the received modulation signal matrix r; A3: Repeat step A1 and step A2 for N times, and use the autocorrelation matrix R and the corresponding modulation type label y Mod and the direction of arrival estimation label y DOA constitute a data set containing N samples and 2N labels wherein the subscript n represents the serial number of the simulation times; and then split the data set in a ratio of 8:2 to make a training set and a test set.

3. The modulation recognition and direction of arrival estimation method based on multi-task learning network according to claim 2, characterized in that, The modulation signal vector r(k) received by the array antenna satisfies formula (1): r(k)=a(θ)·x(k)+v(k) (1) In formula (1), a(θ) represents a steering vector, θ represents the direction of arrival, x(k) represents the transmitted modulation signal, and v(k) represents a Gaussian noise vector. The steering vector a(θ) satisfies formula (2): a(θ) = [1 e j2πd(sinθ) / λ … e j2π(M-1)d(sinθ) / λ ] T (2) In formula (2), j represents an imaginary unit, M represents the total number of array antennas, d represents the interval of array antennas, λ represents the wavelength of a signal, (·) T denotes a transposition operator; The modulation signal matrix r received by the array antenna satisfies formula (3): r＝[r(1) r(2) … r(K)] (3) The autocorrelation matrix R satisfies formula (4): R = E(r · r H ) (4) In Equation (4), E(·) represents a statistical expectation operator, (·) H denotes a conjugate transpose operator.

4. The modulation recognition and direction of arrival estimation method based on multi-task learning network according to claim 2 or 3, characterized in that, The multi-task learning network comprises a feature extraction module and a multi-task processing module. The feature extraction module mainly realizes extraction of common features, and is sequentially composed of a size transformation layer, a convolution layer, a scaled exponential linear activation unit, a convolution layer, a scaled exponential linear activation unit, two transformers, a convolution layer, and a transformer. The multi-task processing module comprises two branches, one is a modulation recognition classifier, and the other is a direction of arrival estimation classifier, both of which are composed of two fully connected layers.

5. The modulation recognition and direction of arrival estimation method based on multi-task learning network according to claim 4, characterized in that, Step B specifically comprises the following steps: 1) design a multi-task learning network; 2) during pre-training, input the sample into the feature extraction module of the designed multi-task learning network; 3) input the output features of the feature extraction module into the multi-task processing module of the multi-task learning network; 4) in the multi-feature processing module, the input features are respectively transmitted to the modulation recognition classifier and the direction of arrival estimation classifier to obtain preliminary results of modulation recognition and direction of arrival estimation, respectively; 5) calculate the cross-entropy loss of the preliminary results of modulation recognition and the modulation recognition label; 6) calculate the root mean square error loss of the preliminary results of the direction of arrival estimation and the direction of arrival estimation label; 7) combine the cross-entropy loss and the mean square error loss into a multi-task loss function in a weighted sum manner; 8) according to the multi-task loss function, use AdamW as the optimizer of the network parameters, set the learning rate to 0.001, and optimize the parameters; 9) repeatedly perform single pre-training until the multi-task loss function in 7) converges, and the pre-training is completed; 10) during parameter fine-tuning, adjust the weights of the cross-entropy loss and the mean square error loss in the multi-task loss function according to the modulation recognition accuracy and the root mean square error of the direction of arrival estimation after pre-training. 11) adjust the value of the input batch and the value of the learning rate to reach an acceptable modulation recognition accuracy and root mean square error of direction of arrival estimation, and the parameter fine-tuning ends.

6. The modulation recognition and direction of arrival estimation method based on multi-task learning network according to claim 4, characterized in that, Step C specifically comprises the following steps: C1: input the samples in the test set into the trained feature extraction module one by one, then input the features output by the feature extraction module into the multi-task processing module, and take the outputs of the modulation recognition classifier and the direction of arrival estimation classifier in the multi-task processing module as the modulation recognition result and the direction of arrival estimation result respectively, and calculate the modulation recognition result accuracy and the root mean square error of direction of arrival estimation according to the modulation recognition labels and the direction of arrival estimation labels corresponding to the samples in the test set respectively; C2: Calculate modulation identification accuracy P Acc and root mean square error of direction of arrival estimation RMSE; modulation identification accuracy P Acc and root mean square error of direction of arrival estimation RMSE satisfy formula (5) and formula (6): In formula (5) and formula (6), N Test denotes the total number of samples in the test set, denotes the total number of samples in the test set that are accurately identified modulation types, and denote the n th DOA estimation label in the test set and the corresponding DOA estimation result, respectively.

Images