Construction method of signal direction of arrival and modulation pattern combined intelligent estimation model

Abstract

The present application relates to a kind of signal wave direction and modulation pattern combined intelligent estimation model construction method, belong to signal analysis processing technical field, solve the existing technology in the wave direction estimation and modulation pattern identification, there is the problems of algorithm complexity, high hardware cost etc..The present application is by constructing multi-task deep neural network structure, and collecting multiple synchronous sampling IQ data stream, forms training sample and test sample, trains neural network structure, and increases modulation pattern and wave direction estimation extraction module, obtains wave direction and modulation pattern combined intelligent estimation model, then estimates test sample, obtains wave direction estimation result and modulation pattern identification result, completes the construction of model.Realize with less performance loss while realizing higher precision wave direction estimation and higher accuracy modulation pattern identification, it is favorable to reduce the workload required for the correlation of two kinds of results, while the demand for processing equipment algorithm can be reduced.

Description

Technical Field

[0001] This invention relates to the field of signal analysis and processing technology, and in particular to a method for constructing a joint intelligent estimation model of signal direction of arrival and modulation pattern. Background Technology

[0002] Radio direction finding and signal modulation pattern recognition are widely used in the field of wireless communication. On the one hand, radio direction finding refers to the determination of the angle between the ray from the direction finding station to the radiation source and a specified reference direction by measuring and calculating electromagnetic field parameters. On the other hand, signal modulation pattern recognition refers to the process of receiving communication signals, processing and analyzing the signals, and determining the modulation pattern of the communication signals.

[0003] For radio direction finding, considering the characteristics of electromagnetic waves, the direction of arrival information from the radiation source is usually reflected in the amplitude or phase response of the antenna's received signal. Therefore, based on the different information responses used to obtain the direction, commonly used radio direction finding techniques are mainly amplitude response and phase response type direction finding. More modern subspace-based direction finding utilizes the array response information of the signal. Due to the Doppler effect, when there is relative motion between the radiation source and the direction finder, the signal received by the direction finder exhibits a Doppler frequency shift. A Doppler direction finder obtains the target direction by measuring this frequency shift using a rotating antenna, which is a frequency response type of direction finding. Additionally, the arrival time of the multiple antennas from the radiation source signal at the direction finder varies; using these time differences to obtain the target direction is a time response type of direction finding. Traditional radio direction finding systems have gradually revealed some shortcomings in certain scenarios during practical applications: First, direction finding performance deteriorates significantly under complex conditions, including complex receiving systems (antenna element mutual coupling, antenna element position errors, inconsistent receiving channels, etc.), complex signals (broadband signals, etc.), and complex channels (reflection channels, etc.); second, they are increasingly unable to meet the requirements for further improvement in direction finding accuracy and resolution; and third, multi-frequency direction finding correction is extremely labor-intensive or cannot achieve full coverage correction. In these scenarios, intelligent direction finding methods demonstrate significant advantages due to the powerful modeling capabilities of machine learning models (especially deep neural network models), achieving significant improvements in direction finding accuracy in complex scenarios, improved resolution in multi-signal direction finding, and improved direction finding accuracy at frequencies not found in the correction table or in the direction of arrival. On the other hand, signal modulation pattern recognition is a crucial step in non-cooperative communication processes, a prerequisite for demodulation and information acquisition, and is commonly used in electronic interference and spectrum monitoring. In recent years, deep learning methods have also attracted researchers' attention in the field of modulation pattern recognition due to their excellent performance. Their advantage lies in achieving high-precision modulation pattern recognition at low signal-to-noise ratios without the need for manually designed features. In signal monitoring and spectrum regulation, there are often scenarios where both direction of arrival (ROA) estimation and modulation pattern identification (MCI) are required for a given signal. Currently, intelligent models for ROA estimation and MCI identification often operate independently. Designing separate intelligent models for ROA estimation and MCI identification offers advantages such as strong targeting and optimal performance, but it also has some drawbacks: First, correlating the ROA results with the MCI identification results requires additional work. Specifically, to confirm that the ROA estimation and MCI identification results originate from the same signal, additional information (such as timestamps) and a correlation algorithm based on this information are needed. Second, it places high demands on the computing power of the processing equipment, requiring the equipment to have the computing power to deploy and run both intelligent models simultaneously.

[0004] In summary, existing technologies for estimating the direction of arrival and identifying the modulation pattern require separate algorithm design and model construction, which suffers from drawbacks such as algorithm complexity and high hardware costs. Summary of the Invention

[0005] Based on the above analysis, the present invention aims to provide a method for constructing a joint intelligent estimation model of signal direction of arrival and modulation pattern, in order to solve the problems of algorithm complexity and high hardware cost in the prior art, which require separate algorithm design and model construction for signal direction of arrival estimation and modulation pattern identification.

[0006] The objective of this invention is mainly achieved through the following technical solutions:

[0007] This invention provides a method for constructing a joint intelligent estimation model of signal direction of arrival and modulation pattern, comprising the following steps:

[0008] Construct a multi-task deep neural network structure and collect multiple synchronously sampled IQ data streams to form training samples and test samples;

[0009] The multi-task deep neural network structure is trained using training samples to obtain the trained multi-task deep neural network structure.

[0010] Based on the trained multi-task deep neural network structure, a modulation pattern and direction of arrival estimation extraction module is added to obtain a joint intelligent estimation model of direction of arrival and modulation pattern. The model is used to perform joint intelligent estimation on the test sample to obtain the direction of arrival estimation result and modulation pattern recognition result of the test sample, thus completing the construction of the joint intelligent estimation model of signal direction of arrival and modulation pattern.

[0011] Based on a further improvement of the above method, the multi-task deep neural network structure includes an input layer, four sequentially connected convolutional blocks, and an incoming direction estimation branch and a modulation pattern recognition branch; wherein,

[0012] Each of the four convolutional blocks includes a convolutional layer, a batch normalization layer, and a linear rectifier unit;

[0013] After the fourth convolutional block, it is divided into the direction of arrival estimation branch and the modulation pattern recognition branch; wherein...

[0014] The incoming wave direction estimation branch includes, in sequence, a first fully connected layer, a fifth linear rectifier unit, and a second fully connected layer;

[0015] The modulation pattern recognition branch includes, in sequence, a third fully connected layer, a sixth linear rectifier unit, a fourth fully connected layer, and a softmax layer.

[0016] Based on a further improvement of the above method, the input to the input layer is a multi-channel synchronously sampled IQ data stream transmitted from the signal receiving array antenna, including:

[0017] In an M-element antenna array, the IQ data of a single sample of element j is represented as:

[0018] s Ij =[s Ij1 ,s Ij2 ,...,s IjN ]

[0019] s Qj =[s Qj1 ,s Qj2 ,...,s QjN ]

[0020] Among them, s Ij For I-channel data, s Qj For Q-channel data, s Ijk For the k-th I-channel sampled data of unit j, s Ijp Let N be the p-th Q-channel sampled data of unit j, where N is the number of sampling points, 1≤k, p≤N;

[0021] The network input for unit j is:

[0022] s j =[s Ij1 ,s Qj1 ,s Ij2 ,s Qj2 ,...s IjN ,s QjN ]

[0023] The input to the multi-task deep neural network structure, i.e., the multi-channel synchronously sampled IQ data stream, is expressed as:

[0024]

[0025] Based on a further improvement of the above method, the incoming wave direction estimation branch is used to output the sine and cosine estimates of the radio incoming wave incident azimuth angle, denoted as O. DoA , is represented as:

[0026] O DoA =[sinα,cosα] T

[0027] Where α is the incident azimuth angle of the radio wave.

[0028] Based on a further improvement to the above method, the modulation pattern recognition branch is used to output the probability value of the radio incoming wave being identified as a certain signal modulation type, denoted as O. MTR ,include:

[0029] Let a be the input to the softmax layer.

[0030] For the softmax layer, the category labels are constructed as follows:

[0031] i∈{1,2,...,N} MTR}

[0032] Where, N MTR Number of categories identified for modulation patterns;

[0033] The conditional probability that the incoming modulation pattern predicted by the softamax layer belongs to category i is:

[0034]

[0035] Where, ω i Let be the weight vector for the i-th class;

[0036] Then O MTR =p(i∣I).

[0037] Based on further improvements to the above method, the multi-task deep neural network structure is trained, including:

[0038] Based on the outputs of the direction of arrival estimation task and the modulation pattern recognition task using a multi-task deep neural network structure, loss functions for the direction of arrival estimation task and the modulation pattern recognition task are constructed.

[0039] Based on the loss function of the direction of arrival estimation task and the loss function of the modulation pattern recognition task, a loss function for a multi-task neural network is constructed.

[0040] Based on the loss function of the multi-task neural network structure, the stochastic gradient method is used to train and update the model parameters to obtain the trained multi-task deep neural network structure.

[0041] Based on further improvements to the above method, a loss function for the direction of arrival estimation task is constructed based on the output of the multi-task deep neural network structure, including:

[0042] The loss function for the incoming wave direction estimation task is constructed as the root mean square error function. Therefore, the loss for estimating the incoming wave direction on the training samples is:

[0043]

[0044] Where, N t The number of training samples, The output is the estimated direction of incoming waves for the j'-th training sample. This is the estimated value of the incoming wave direction for the j'th training sample.

[0045] Based on further improvements to the above method, a loss function for the modulation pattern recognition task is constructed based on the output of a multi-task deep neural network structure, including:

[0046] If the loss function for the modulation pattern recognition task is constructed as the average cross-entropy function, then the modulation pattern recognition loss for the training samples is:

[0047]

[0048] in, For the i'th element in the modulation pattern recognition result of the j'th training sample, N MTR Number of categories to identify for modulation patterns.

[0049] Based on the further improvement of the above method, a loss function for a multi-task neural network is constructed based on the loss function of the direction of arrival estimation task and the loss function of the modulation pattern recognition task, which is expressed as:

[0050] Loss = Loss DoA +η·Loss MTR

[0051] Where η is the weight adjustment parameter, and η > 0.

[0052] Based on a further improvement of the above method, the modulation pattern and direction of arrival estimation extraction module is represented as follows:

[0053]

[0054]

[0055] Where tj represents the test sample. Estimate the final output of the branch for the incoming wave direction of tj. The final output of the modulation pattern recognition branch for tj; These are the first and second elements of the final output of the incoming wave direction estimation branch of tj, namely the sine estimate and the cosine estimate, respectively. N represents the estimated incident azimuth angle of tj; MTR represents the number of modulation style categories; i is the modulation style classification number of tj.

[0056] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0057] 1. This invention discloses a method for constructing a joint intelligent estimation model of signal arrival direction and modulation pattern. It can obtain the incident angle estimate and modulation pattern of radio wave using only one model, without the need to build separate models and configure high computing power equipment, thus effectively saving hardware costs.

[0058] 2. The algorithm designed in this invention for constructing a joint intelligent estimation model of signal arrival direction and modulation pattern is simple and effective. It does not require correlation between the signal arrival direction result and the modulation pattern identification result, that is, it does not require designing a correlation algorithm, thus saving the computing power of hardware devices.

[0059] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from the description and drawings, which are particularly pointed out. Attached Figure Description

[0060] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0061] Figure 1 This is a diagram of the architecture of a joint intelligent estimation model for signal arrival direction and modulation pattern.

[0062] Figure 2 A schematic diagram of a uniform five-element circular receiving array;

[0063] Figure 3 This is the loss function for a deep neural network based on training samples over 200 training epochs. Detailed Implementation

[0064] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0065] Example 1

[0066] A specific embodiment of the present invention discloses a method for constructing a joint intelligent estimation model of signal direction of arrival and modulation pattern, such as... Figure 1 As shown, it includes the following steps:

[0067] S1. Construct a multi-task deep neural network structure and collect multiple synchronously sampled IQ data streams to form training samples and test samples.

[0068] S2. Using training samples, train the multi-task deep neural network structure to obtain the trained multi-task deep neural network structure.

[0069] S3. Based on the trained multi-task deep neural network structure, add a modulation pattern and direction of arrival estimation extraction module to obtain a joint intelligent estimation model of direction of arrival and modulation pattern; use the model to perform joint intelligent estimation on the test sample to obtain the direction of arrival estimation result and modulation pattern recognition result of the test sample, and complete the construction of the joint intelligent estimation model of signal direction of arrival and modulation pattern.

[0070] Example 2

[0071] Based on Example 1, step S1 can be further refined into the following steps:

[0072] S11. Construct a multi-task deep neural network structure that includes multiple convolutional blocks.

[0073] For example, the number of convolutional blocks is 4, and the entire multi-task deep neural network structure has a total of 20 layers, such as... Figure 1 As shown, it includes, in sequence:

[0074] The input layer, first convolutional block, second convolutional block, third convolutional block, fourth convolutional block, and arrival direction estimation branch and modulation pattern recognition branch; among which,

[0075] The first convolutional block includes, in sequence, the first convolutional layer, the first batch of normalization layers, and the first linear rectifier unit;

[0076] The second convolutional block consists of a second convolutional layer, a second batch normalization layer, and a second linear rectifier unit.

[0077] The third convolutional block consists of the third convolutional layer, the third batch normalization layer, and the third linear rectifier unit in sequence.

[0078] The fourth convolutional block consists of the fourth convolutional layer, the fourth batch normalization layer, and the fourth linear rectifier unit.

[0079] The fourth convolutional block is then divided into an incoming direction estimation branch and a modulation pattern recognition branch; wherein...

[0080] The incoming wave direction estimation branch includes, in sequence, a first fully connected layer, a fifth linear rectifier unit, and a second fully connected layer;

[0081] The modulation pattern recognition branch includes, in sequence, a third fully connected layer, a sixth linear rectifier unit, a fourth fully connected layer, and a softmax layer.

[0082] S12. Using the data stream as input to the multi-task deep neural network structure, obtain the output of the multi-task deep neural network structure.

[0083] Specifically, the input to the input layer is a multi-channel synchronous sampling IQ data stream sent by the signal receiving array antenna.

[0084] The multiple convolutional blocks, based on the input original multi-channel synchronous sampling IQ data stream, are used to extract high-dimensional features of the data stream layer by layer. When the high-dimensional features have sufficient representational power, the data is output to the direction of arrival estimation branch and the modulation pattern recognition branch.

[0085] Specifically, for any convolutional layer in a convolutional block, it contains multiple convolutional kernels. Each element of the convolutional kernel corresponds to a weight coefficient and a bias, similar to a neuron in a feedforward neural network. When the convolutional kernel is working, it will regularly scan the input features, perform matrix element multiplication and summation on the input features within the receptive field, and add the bias, thereby extracting the high-dimensional features of the input data.

[0086] In the first convolutional block, the output feature dimension is 1×128; in the second convolutional block, the output feature dimension is 1×512.

[0087] For any convolutional layer in a convolutional block, the operation steps are as follows: first, perform a convolution operation with the input of the convolutional layer using the learnable weight vector to obtain the convolution operation result; then, sum the convolution operation result with the learnable bias vector to obtain the calculation result of the convolutional layer, and output it to the corresponding next layer.

[0088] For any batch normalization layer in a convolutional block, its function is to uniformly perform characteristic statistics on the input of the subsequent neural network, accelerate the network learning speed, and effectively prevent gradient explosion and gradient vanishing.

[0089] One of the functions of the linear rectified unit is to detect whether features are relevant to the task, retain only useful features and remove irrelevant features, thus ensuring the sparsity of the network. The derivative of the linear rectified unit function is only 0 or 1, which can effectively avoid the gradient vanishing problem during backpropagation. In addition, the linear rectified unit function is also simple and fast to compute.

[0090] The incoming wave direction estimation branch integrates various high-dimensional features for incoming wave direction estimation through two fully connected layers and a linear rectifier unit. Since using alpha will lead to error penalty, in order to avoid this problem, this application constructs the incoming wave direction estimation as a regression problem. The incoming wave direction estimation branch is used to output the sine and cosine estimates of the incident azimuth angle of the radio incoming wave. The first fully connected layer of the incoming wave direction estimation branch outputs a feature dimension of 1×128, and the second fully connected layer outputs a feature dimension of 1×2.

[0091] The modulation pattern recognition branch integrates the high-dimensional features through two fully connected layers and a linear rectifier unit, and normalizes the membership degree of the modulation pattern through a softmax layer. In this application, the modulation pattern recognition task is constructed as a classification task, and the output of the modulation pattern recognition task is a one-bit valid code of the modulation pattern classification number. The modulation pattern recognition branch is used to output the probability value of the radio wave being identified as a certain signal modulation type. The first fully connected layer of the modulation pattern recognition branch outputs a feature dimension of 1×128, and the second fully connected layer outputs a feature dimension of 1×4.

[0092] For example, the following is a specific computational process of a multi-task deep neural network structure:

[0093] In an M-element antenna array, the IQ data of element j in a single sampling is:

[0094] s Ij =[s Ij1 ,s Ij2 ,...,s IjN ]

[0095] s Qj =[s Qj1 ,s Qj2 ,...,s QjN ]

[0096] Among them, s Ij For I-channel data, s Qj For Q-channel data, s Ijk For the k-th I-channel sampled data of unit j, s Ijp Let N be the p-th Q-channel sampled data of unit j, where N is the number of sampling points, 1≤k, p≤N.

[0097] The network input for unit j is:

[0098] s j =[s Ij1 ,s Qj1 ,s Ij2 ,s Qj2 ,...s IjN ,s QjN ]

[0099] The input to the entire multi-task deep neural network structure is:

[0100]

[0101] The output of the incoming wave direction estimation task is:

[0102] O DoA =[sinα,cosα] T

[0103] Where α is the incident azimuth angle of the radio wave;

[0104] A multi-task deep neural network structure can be represented as a nonlinear mapping from input to output, where the input I is equal to the output O. DoA The mapping is represented as:

[0105]

[0106] The output of the modulation pattern recognition task is O MTR Then from input to output O MTR The mapping is represented as:

[0107]

[0108] in, Let be the i-th function of the multi-task deep neural network structure, where For the input layer, These are the first, second, third, and fourth convolutional layers, respectively. These are the first batch of normalized layers, the second batch of normalized layers, the third batch of normalized layers, and the fourth batch of normalized layers. These are the first linear rectifier unit, the second linear rectifier unit, the third linear rectifier unit, the fourth linear rectifier unit, the fifth linear rectifier unit, and the sixth linear rectifier unit, respectively. These are the first fully connected layer, the second fully connected layer, the third fully connected layer, and the fourth fully connected layer, respectively. It is a softmax layer.

[0109] Specifically, the input to the softmax layer is denoted as a;

[0110] For the softmax layer, the category labels are constructed as follows:

[0111] i∈{1,2,...,N} MTR}

[0112] Where, N MTR Number of categories identified for modulation patterns;

[0113] The conditional probability that the incoming modulation pattern predicted by the softamax layer belongs to category i is:

[0114]

[0115] Where, ω i Let be the weight vector for the i-th class;

[0116] Then O MTR =p(i∣I).

[0117] Preferably, step S2 can be further refined into the following steps:

[0118] S21. Based on the multi-task deep neural network structure, further construct the network loss function and parameter update method, and then train and update the network parameters of the multi-task deep neural network structure.

[0119] Specifically, based on the outputs of the direction of arrival estimation task and the modulation pattern recognition task using a multi-task deep neural network structure, corresponding loss functions are constructed; among them,

[0120] The loss function for the incoming wave direction estimation task is constructed as the root mean square error function. Therefore, the loss for estimating the incoming wave direction on the training samples is:

[0121]

[0122] Where, N t The number of training samples, The output is the estimated direction of incoming waves for the j'-th training sample. This is the estimated value of the incoming wave direction for the j'th training sample.

[0123] If the loss function for the modulation pattern recognition task is constructed as the average cross-entropy function, then the modulation pattern recognition loss for the training samples is:

[0124]

[0125] in, For the i'th element in the modulation pattern recognition result of the j'th training sample, N MTR Number of categories to identify for modulation patterns.

[0126] S22. Based on the loss function of the direction of arrival estimation task and the loss function of the modulation pattern recognition task, the loss function for constructing the multi-task neural network structure is:

[0127] Loss = Loss DoA +η·Loss MTR

[0128] Where η is the weight adjustment parameter, and η > 0; when η is small, the network focuses more on the accuracy of the direction of arrival estimation, and when η is large, the network focuses more on the accuracy of modulation pattern recognition. In practical applications, the parameter η can be adjusted as needed.

[0129] For example, based on engineering experience, we take η = 1.

[0130] Based on the loss function of the multi-task neural network structure, the stochastic gradient method is used to train and update the model parameters to obtain the trained multi-task deep neural network structure.

[0131] Preferably, step S3 can be further refined into the following steps:

[0132] S31. Based on the trained multi-task deep neural network structure, at its output end, namely the direction of arrival estimation branch and the modulation pattern recognition branch, a modulation pattern and direction of arrival estimation extraction module is added to obtain a joint intelligent estimation model of direction of arrival and modulation pattern.

[0133] The model is used to perform joint intelligent estimation on the test samples to obtain the estimated direction of arrival and modulation pattern recognition results of the test samples.

[0134] Specifically, the inputs to the modulation pattern and direction of arrival estimation extraction module include:

[0135] In the trained multi-task deep neural network structure, the output of the incoming wave direction estimation branch is the sine and cosine estimates of the incident azimuth angle of the radio incoming wave.

[0136] The output of the modulation pattern recognition branch is the probability value of the incoming radio wave being identified as a certain signal modulation type.

[0137] Specifically, the output of the modulation pattern and direction of arrival estimation extraction module includes:

[0138] Estimated results of the incident azimuth angle of the radio wave;

[0139] Radio wave modulation style classification number.

[0140] For example, for test sample tj, the output of its incoming direction estimation branch is: And the output of its modulation pattern recognition branch is The modulation pattern and direction of arrival estimation extraction module is represented as follows:

[0141]

[0142]

[0143] in, These are the first and second elements of the final output of the incoming wave direction estimation branch of tj, namely the sine estimate and the cosine estimate, respectively. N represents the estimated incident azimuth angle of tj; MTR represents the number of modulation style categories; i is the modulation style classification number of tj.

[0144] Thus, the direction of arrival estimation and modulation pattern identification of test sample j have been completed.

[0145] S32. Test the joint intelligent estimation model of direction of arrival and modulation pattern, and complete the construction of the joint intelligent estimation model of direction of arrival and modulation pattern.

[0146] For example, the application scenario and effect of the intelligent estimation model for the direction of arrival and modulation pattern constructed by the present invention are tested through a typical application case.

[0147] Specifically, a uniform five-element circular array is used to receive the signal, with the incoming wave direction angle ranging from 0° to 360°, such as... Figure 2 As shown. The types of incoming radio waves are BPSK, QPSK, 8PSK and MSK modulation styles.

[0148] For each modulation pattern, 50 sets of samples ranging from 1° to 360° were generated in 1° increments, totaling 72,000 samples. 90% of these samples were randomly selected as training samples for the multi-task deep neural network structure, with the remainder serving as test samples. For each sample, the signal oversampling rate was 8, and each set of samples contained 128 points, meaning the input size of the multi-task deep neural network structure was 5×256, with a signal-to-noise ratio of 20dB.

[0149] The minimum batch size for model training is set to 1024, with 200 training cycles. Additionally, η is set to 1. Figure 3 The loss function of the deep neural network based on training samples over 200 training cycles is presented. Ultimately, based on the test samples, the root mean square error (RMSE) of the direction of arrival estimation is 2.9584°, and the modulation pattern recognition accuracy is 96.25%. For comparison, based on the same training and test samples, a deep neural network with the same structure that only estimates the direction of arrival (i.e., η = 0) has an RMSE of 2.5829°, while a deep neural network that only recognizes the modulation pattern (i.e., η = ∞) achieves a modulation pattern recognition accuracy of 96.76%.

[0150] The loss function of the deep neural network based on the training samples over 200 training epochs, such as Figure 3 As shown.

[0151] For each modulation pattern, under a signal-to-noise ratio of 10dB, 10 sets of samples ranging from 1° to 360° were generated in 1° increments, totaling 14,400 test samples. The multi-task deep neural network trained under a signal-to-noise ratio of 20dB was used for prediction. The root mean square error of the signal direction of arrival was 3.7512°, and the modulation pattern recognition accuracy was 95.54%, indicating that the model has good generalization performance.

[0152] As can be seen from the above results, the intelligent estimation model for the direction of arrival and modulation pattern disclosed in this invention can achieve high-precision estimation of the direction of arrival and high-accuracy identification of the modulation pattern with less performance loss. This helps to reduce the workload required to correlate the two types of results and reduces the computing power required for processing equipment.

[0153] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0154] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a joint intelligent estimation model of signal direction of arrival and modulation pattern, characterized in that, Includes the following steps: Construct a multi-task deep neural network structure and collect multiple synchronously sampled IQ data streams to form training samples and test samples; The multi-task deep neural network structure is trained using training samples to obtain the trained multi-task deep neural network structure. Based on the trained multi-task deep neural network structure, a modulation pattern and direction of arrival estimation extraction module is added to obtain a joint intelligent estimation model of direction of arrival and modulation pattern; the model is used to perform joint intelligent estimation on test samples to obtain the direction of arrival estimation result and modulation pattern recognition result of the test samples, thus completing the construction of the joint intelligent estimation model of signal direction of arrival and modulation pattern. The modulation pattern and direction of arrival estimation extraction module is represented as follows: in, The input to the softmax layer is tj, which represents the test sample. Estimate the final output of the branch for the incoming wave direction of tj. The final output of the modulation pattern recognition branch for tj; , These are the first and second elements of the final output of the incoming wave direction estimation branch of tj, namely the sine estimate and the cosine estimate, respectively; This is the estimated result of the incident azimuth angle of tj; represents the number of modulation style categories; i is the modulation style classification number of tj.

2. The method for constructing a joint intelligent estimation model of signal arrival direction and modulation pattern according to claim 1, characterized in that, The multi-task deep neural network structure includes an input layer, four sequentially connected convolutional blocks, and a direction-of-arrival estimation branch and a modulation pattern recognition branch; wherein, Each of the four convolutional blocks includes a convolutional layer, a batch normalization layer, and a linear rectifier unit; After the fourth convolutional block, it is divided into the direction of arrival estimation branch and the modulation pattern recognition branch; wherein... The incoming wave direction estimation branch includes, in sequence, a first fully connected layer, a fifth linear rectifier unit, and a second fully connected layer; The modulation pattern recognition branch includes, in sequence, a third fully connected layer, a sixth linear rectifier unit, a fourth fully connected layer, and a softmax layer.

3. The method for constructing a joint intelligent estimation model of signal arrival direction and modulation pattern according to claim 2, characterized in that, The input to the input layer is a multi-channel synchronously sampled IQ data stream transmitted from the signal receiving array antenna, including: one M In a single-element antenna array, the element j The IQ data from a single sample is represented as: in, For I-channel data, For Q-channel data, For unit j The One I-channel sampling data, For unit j The Q-channel sampling data, The number of sampling points. ; Then unit j The network input is: The input to the multi-task deep neural network structure, i.e., the multi-channel synchronously sampled IQ data stream, is expressed as: 。 4. The method for constructing a joint intelligent estimation model of signal arrival direction and modulation pattern according to claim 3, characterized in that, The incoming wave direction estimation branch is used to output the sine and cosine estimates of the radio incoming wave incident azimuth angle, denoted as . , is represented as: in, The azimuth angle of the incoming radio wave.

5. The method for constructing a joint intelligent estimation model of signal direction of arrival and modulation pattern according to claim 4, characterized in that, The modulation pattern recognition branch is used to output the probability value of the radio wave being identified as a certain signal modulation type, denoted as . ,include: For the softmax layer, the category labels are constructed as follows: in, Number of categories identified for modulation patterns; The arrival modulation pattern predicted by the softamax layer belongs to category [category missing]. i The conditional probability is: in, For the first i The weight vector of the class; but = p ( i |I).

6. The method for constructing a joint intelligent estimation model of signal arrival direction and modulation pattern according to claim 5, characterized in that, Training the multi-task deep neural network structure includes: Based on the outputs of the direction of arrival estimation task and the modulation pattern recognition task using a multi-task deep neural network structure, loss functions for the direction of arrival estimation task and the modulation pattern recognition task are constructed. Based on the loss function of the direction of arrival estimation task and the loss function of the modulation pattern recognition task, a loss function for a multi-task neural network is constructed. Based on the loss function of the multi-task neural network structure, the stochastic gradient method is used to train and update the model parameters to obtain the trained multi-task deep neural network structure.

7. The method for constructing a joint intelligent estimation model of signal direction of arrival and modulation pattern according to claim 6, characterized in that, Based on the output of the direction of arrival estimation task using a multi-task deep neural network structure, a loss function for the direction of arrival estimation task is constructed, including: The loss function for the incoming wave direction estimation task is constructed as the root mean square error function. Therefore, the loss for estimating the incoming wave direction on the training samples is: in, The number of training samples, For the first j’ The output of the estimated direction of arrival for each training sample. For the first j’ The estimated value of the incoming wave direction for each training sample.

8. The method for constructing a joint intelligent estimation model of signal direction of arrival and modulation pattern according to claim 6, characterized in that, Based on the output of a modulation pattern recognition task using a multi-task deep neural network structure, a loss function for the modulation pattern recognition task is constructed, including: If the loss function for the modulation pattern recognition task is constructed as the average cross-entropy function, then the modulation pattern recognition loss for the training samples is: in, The number of training samples, For the first j’ The modulation pattern recognition result of the training sample is the first one. i’ One element, Number of categories to identify for modulation patterns.

9. The method for constructing a joint intelligent estimation model of signal arrival direction and modulation pattern according to claim 6, characterized in that, Based on the loss functions of the direction of arrival estimation task and the modulation pattern recognition task, the loss function of the multi-task neural network is constructed as follows: in, The loss function for estimating the direction of incoming waves. The loss function for the modulation pattern recognition task is... For weight adjustment parameters, and .

Images