Intelligent anti-interference beam forming method based on dynamic prediction from interference direction

Abstract

The invention provides an intelligent anti-interference beam forming method based on dynamic prediction of an interference coming direction, and belongs to the crossing field of array signal processing and artificial intelligence. The method comprises the following steps: firstly, constructing an interference coming direction prediction model based on an autonomous attention full-connection network, and training the prediction model by using interference coming direction historical observation data; then, the trained prediction model is used for obtaining the interference coming direction of the next moment, an objective function is constructed in combination with the interference pointing requirement, an improved particle swarm method is used for forming the phase of each array element in the array, and an anti-interference wave beam is formed; and finally, along with the dynamic change of the coming direction of the interference, the coming direction of the interference is predicted again, an anti-interference beam is formed, and the dynamic intelligent anti-interference capability is realized. According to the method, the interference coming direction rule is extracted by using the interference coming direction historical data and prediction is realized, and intelligent interference beam forming in combination with the prediction result and the improved particle swarm method is realized.

Description

Technical Field

[0001] The present invention belongs to the intersection field of array signal processing and artificial intelligence, and in particular to an intelligent anti-interference beamforming method based on dynamic prediction of interference. Background Art

[0002] With the increasing number of frequency-using equipment in the physical world, the electromagnetic spectrum space is becoming increasingly complex, and interference phenomena such as co-frequency, co-source, and intermodulation are becoming increasingly frequent. Using antenna arrays to form an antenna radiation pattern with "low gain in the direction of the interference signal and high gain in the direction the antenna is pointing" has become a key means of anti-interference.

[0003] In the patent application of the 54th Research Institute of China Electronics Technology Group Corporation, with the publication number CN111030748B, entitled "A digital beamforming terminal device for satellite communication", a detailed implementation scheme of the zeroing anti-interference beamforming terminal device is given from the perspective of hardware implementation. However, the patent does not disclose the beamforming algorithm.

[0004] In the patent "A hybrid coherent interference suppression beamforming method in an impulsive noise background" applied for by Sun Yat-sen University with application number CN202410646145.2, nonlinear compression processing and spatial decomposition coherence technology are combined to improve the optimal beamforming method based on the minimum variance distortion-free response criterion, solving the problem of poor robustness of the beamforming method when resisting coherent interference.

[0005] In the article "Study on Adaptive Beamforming Method for Wide Null Interference Suppression Based on the Fastest Method" published by Gao Zhiqi et al. in Signal Processing, the spatial smoothing weight preprocessing mechanism is first used to solve the coherence and extract features, and the initial weights are obtained based on the constructed projection matrix. Finally, the covariance matrix reconstruction and the fastest iteration are combined to achieve interference suppression capability. However, the above method process is relatively complicated, and the technical implementation requires a high level of theoretical knowledge.

[0006] In the patent application of Shandong University of Science and Technology, application number CN115296709A, "Adaptive Anti-interference Beamforming Method Based on Preprocessing Deep Reinforcement Learning", the reinforcement learning method is used to control the array phase, so as to achieve the premise of keeping the main lobe of the beam pointed under control, while the side lobe is nulled to prevent interference signals. Although this method uses the reinforcement learning method to form an anti-interference beamforming paradigm in a robust environment, it requires a delay to estimate the interference source in a non-steady complex environment. Therefore, when the interference source is changing dynamically, it is necessary to perform spatial anti-interference processing for future interference sources. Summary of the invention

[0007] In view of this, the present invention provides an intelligent anti-interference beamforming method based on dynamic prediction of interference direction. The present invention can cope with dynamic changes in interference direction and realize stable anti-interference beamforming.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] The intelligent anti-interference beamforming method based on dynamic prediction of interference direction includes the following steps:

[0010] Step 1: Construct interference source data collection, build an interference source prediction model based on the autonomous attention fully connected network, and train the interference source prediction model;

[0011] Step 2: The interference direction is predicted by the interference direction prediction model, and the anti-interference beamforming objective function is constructed. The phase of each antenna array element is calculated based on the particle swarm algorithm to achieve anti-interference beamforming;

[0012] Step 3: Repeat step 2 to implement dynamic beamforming for dynamic changes in the interference direction.

[0013] Furthermore, the specific method of step 1 is:

[0014] Step 1-1: Take the array antenna as the origin and continuously collect the interference observation data at T+1 time. The interference observation data at each time Including the direction and position of the interference source , Distance to interference source and interference source ; Among them, the interference observation data of the first T moments is recorded as sample data D, D = , the interference source in the observed data at the T+1th moment is As label L; repeat the above process K times to construct a sample set and label sets , where the sample set = , label set = ; The sample set and label set are proportional Split to get the training sample set , test sample set , training label set and the test label set , the lengths are , , and , Indicates rounding down;

[0015] Step 1-2: Build an interference prediction model, including five linear fully connected layers. The input and output dimensions of the first four linear fully connected layers are both 3T, and the input dimension of the fifth linear fully connected layer is 3T, and the output dimension is 1;

[0016] The first three linear fully connected layers output the query vector query, the key vector key and the value vector value respectively, from which the attention scores scores are calculated:

[0017]

[0018] in, is the transpose of the key vector;

[0019] Apply the softmax function to the last dimension of scores, the sequence dimension, to generate the attention weight:

[0020] attention=softmax(scores, dim=−1)

[0021] Among them, dim=−1 represents the last dimension;

[0022] Perform matrix multiplication of the attention weight and the value vector value to obtain the weighted value Value:

[0023] Value = attention·value

[0024] The fourth linear fully connected layer linearly transforms the weighted value Value to obtain the output out;

[0025] Input out into the fifth linear fully connected layer, and the output is the predicted value;

[0026] Step 1-3: Set the loss function to mean square error, the optimizer to Adam, the learning rate to lr, and the optimal loss value to infinity;

[0027] Set the interference source prediction model to training mode and select the training sample set and training label set Take out B pieces of data, clear the optimizer gradient to 0, input interference to the prediction model, and get B prediction values; use the loss function to compare the B prediction values ​​with the training label set The loss value of the extracted B data is calculated, the gradient of the loss value to the model parameters is calculated, and the model parameters are updated according to the rules of the optimizer algorithm until the training sample set is and training label set The data will be retrieved.

[0028] Set the interference source prediction model to test mode and set the test sample set Input interference to the prediction model to obtain the prediction results, and use the loss function to compare the prediction results with the test label set. Calculate the loss value of the data in; if the loss value is less than the optimal loss value, save the interference to the prediction model as the optimal model and assign the loss value to the optimal loss value; otherwise, do not save the model and do not assign a value;

[0029] Repeat the training and testing. When the number of repetitions corresponding to the optimal loss value is greater than the threshold number E, and the number of repetitions is greater than twice the number of repetitions corresponding to the optimal loss value, stop the training.

[0030] Furthermore, the specific method of step 2 is:

[0031] Step 2-1: Input the interference observation data collected continuously for T moments into the interference prediction model to obtain the interference prediction value;

[0032] Step 2-2: Construct an anti-interference objective function with the input of the function being the preset pointing direction , interference prediction direction and the antenna element phase vector , antenna element phase vector The phase of each antenna element Composition, n=1,2,...,N, N is the number of array elements; the output of the function is the target value:

[0033]

[0034] Where d is the antenna element spacing, To round down, To round up;

[0035] Step 2-3: Construct a set of candidate phase vectors , contains a total of M phase vectors, each phase vector The dimension is N, and the value range is , m=1,2,...,M;

[0036] Construct a velocity set, which contains M velocity vectors. Each velocity vector The dimension is N;

[0037] Step 2-4: Set the optimal value of each phase vector ,all Composition Collection , using an alternative phase vector set The value in is Each of Initialize, then Substitute them into the anti-interference objective function respectively, and get M target values ​​in total. The individual optimal fitness value of M individuals is combined into an optimal fitness value set , The largest individual optimal fitness value is set as the global optimal fitness value , the corresponding phase vector is the global optimal vector ;

[0038] Step 2-5: Set inertia weight w, cognitive parameters and social parameters , and initialize the counter stable_counter=0 used to check the stability of the global optimal value;

[0039] Step 2-6: Update each phase vector in the candidate phase vector set The value of:

[0040]

[0041] in, and is a random number between [0, 1];

[0042] The updated phase vector Limit the values ​​in the range Then, the phase vector Substitute the anti-interference objective function for calculation. If the calculated target value is less than the corresponding individual optimal fitness value, then use the target value to update The corresponding individual optimal fitness value in renew The optimal value of the corresponding phase vector in ;

[0043] Step 2-7: If the optimal fitness value set There is a fitness value less than the global optimal The fitness value of is assigned to the global optimal fitness value. , and its corresponding phase vector Assign the global optimal vector , and set the counter stable_counter to 0, otherwise add 1 to the counter stable_counter;

[0044] Step 2-8: Repeat steps 2-6 to 2-7 until the stable_counter value exceeds the preset upper limit. The vector in is the phase of each antenna element.

[0045] The beneficial effects of the present invention are:

[0046] 1. The present invention has the ability to predict the direction of interference. The present invention constructs a self-attention fully connected deep neural network to extract the law of dynamic changes in interference sources, realize the prediction of the direction of interference, increase the redundancy of system response time, and help improve the anti-interference effect and robustness.

[0047] 2. The present invention constructs an anti-interference beamforming paradigm. The objective function is constructed based on the main lobe pointing and interference, and the array phase is solved using the improved particle swarm method to form an anti-interference antenna pattern, thus realizing the intelligent airspace anti-interference capability.

[0048] 3. The present invention has the ability to cope with dynamic interference environments. Through the iteration of interference prediction and intelligent anti-interference beamforming, the airspace anti-interference capability against dynamic changes in interference sources is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The present invention is a flow chart of an intelligent anti-interference beamforming method based on dynamic prediction of interference.

[0050] Figure 2 Schematic diagram of the relationship between loss value and training times.

[0051] Figure 3 This is the anti-interference antenna pattern formed for the first time.

[0052] Figure 4 This is the anti-interference antenna pattern formed for the second time.

[0053] Figure 5 This is the anti-interference antenna pattern formed for the third time.

[0054] Figure 6 This is the anti-interference antenna pattern formed for the fourth time. DETAILED DESCRIPTION

[0055] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0056] The intelligent anti-interference beamforming method based on dynamic prediction of interference direction first constructs an interference direction prediction model based on self-attention fully connected deep neural network to predict future interference directions. Then, the objective function is constructed based on the predicted interference direction and main lobe direction, and the generation of anti-interference antenna pattern is realized through the improved particle swarm algorithm. Finally, as the interference direction changes dynamically, the direction is continuously predicted and the anti-interference antenna pattern is generated, realizing the intelligent anti-interference beamforming capability to cope with dynamic changes of interference.

[0057] The overall process is as follows Figure 1 The specific steps include:

[0058] Step 1: Build interference source data collection, build an interference source prediction model based on the autonomous attention fully connected network, and realize interference source prediction based on historical observation data by training and acquiring the interference source model; the specific process is as follows:

[0059] Step 1-1: Interference with data collection;

[0060] Taking the array antenna as the origin, continuously collect the interference observation data at T+1 time, where T is 5, and the observation data at each time Including the direction and position of the interference source , Distance to interference source and interference source The interference observation data of the first T moments is recorded as sample data D, D = ; The interference source in the sample data at the T+1th moment comes from As label L. Repeat the above process K times, K is 895021, and construct the sample set and label sets , where the sample set = and label sets = . The number of sample sets and label sets is calculated according to The ratio is divided into Take 0.9 to get the training sample set , test sample set , training label set and the test label set , the lengths are , , and .

[0061] Step 1-2: Build a prediction model based on the interference of the autonomous attention fully connected network;

[0062] Construct five linear fully connected layers. The input and output dimensions of the first four are both 3T, that is, 15, and the input dimension of the fifth is 3T and the output dimension is 1.

[0063] Input the input data into the first three fully connected layers to obtain the query vector query, key vector key and value vector value respectively;

[0064] Calculate the attention scores:

[0065]

[0066] in The transpose of the key vector.

[0067] Apply softmax to the last dimension of scores, the sequence dimension, to generate attention weights:

[0068] attention=softmax(scores, dim=−1)

[0069] Perform matrix multiplication of the attention weight and the value vector value to obtain the weighted value Value:

[0070] Value = attention·value

[0071] The weighted value Value is input into the fourth fully connected layer for further linear transformation to obtain the output out.

[0072] Input out into the fifth fully connected layer to get the predicted value.

[0073] Step 1-3: Train the interference prediction model and obtain the optimal interference prediction model;

[0074] The loss function is defined as mean square error, the optimizer is Adam, the learning rate is lr=0.001, and the optimal loss value is infinity.

[0075] Set the prediction model to training mode. and training label set Take out B data, B takes 32 values, clear the optimizer gradient to 0, input the prediction model to get B prediction values, and compare them with the training label set Take out B data and use the loss function to find the loss value, calculate the gradient of the loss value to the model parameters, and update the prediction model parameters according to the rules of the optimizer algorithm until the training sample set is and training label set The data will be retrieved.

[0076] Set the prediction model to test mode. Input the prediction model to get the prediction results and test label set The loss value is calculated through the loss function. If the loss value is less than the optimal loss value, the prediction model is saved as the optimal model and the loss value is assigned to the optimal loss value. If the loss value is greater than or equal to the optimal loss value, the model is not saved and no value is assigned.

[0077] Repeat the above steps. When the number of repetitions corresponding to the optimal loss value is greater than the threshold number E=100, and the number of repetitions is greater than twice the number of repetitions corresponding to the optimal loss value, stop the above process. The relationship between the loss value and the number of training times is as follows: Figure 2 shown.

[0078] Step 2: According to the predicted interference direction, an anti-interference beamforming objective function is constructed, and the phase of each antenna is calculated based on the improved particle swarm algorithm to achieve anti-interference beamforming; the specific process is as follows:

[0079] Step 2-1: Input the interference direction observation data collected continuously for T moments into the interference direction prediction model to obtain the interference direction prediction value. The detected interference direction data is [[11823.3086, 13286.2754, 48.3345,11683.4307, 13294.4619, 48.6904, 11543.5537, 13302.6475, 49.0497, 11403.6758,13310.8330, 49.4126, 11263.7979, 13319.0195, 49.7790], and the prediction value is 50.1101, which is used as the interference prediction direction. .

[0080] Step 2-2: Construct the anti-interference objective function. The input is the preset pointing direction , interference prediction direction and the antenna element phase vector , where the phase of each antenna element is ,in , N is the number of array elements, which is 32. The output target value is:

[0081]

[0082] Where d is the antenna element spacing, which is 0.5 meters. To round down, To round up, preset the pointing direction It is -20°.

[0083] Step 2-3: Construct a set of candidate phase vectors , contains M phase vectors, M is 50, each phase vector The dimension is N, and the value range is , corresponding to the number of antenna array elements, then The dimension of is M×N. The dimension of the velocity set is the same as that of the candidate phase set. The velocity vector corresponding to each phase vector is .

[0084] Step 2-4: Set the optimal value of each phase vector ,all Composition Collection , using an alternative phase vector set The value in is Each of Initialize, then Substitute them into the anti-interference objective function respectively, and get M target values ​​in total. The individual optimal fitness value of M individuals is combined into an optimal fitness value set , The largest individual optimal fitness value is set as the global optimal fitness value , the corresponding phase vector is the global optimal vector .

[0085] Step 2-5: Set inertia weight w, cognitive parameters and social parameters , take 0.7, 1.4 and 1.4 respectively, and initialize the counter stable_counter=0 used to check the stability of the global optimal value.

[0086] Step 2-6: Update each phase vector in the candidate phase vector set The value of:

[0087]

[0088] in and is a random number between [0, 1].

[0089] The updated phase vector Limit the values ​​in The phase vector Used for objective function calculation. If the calculated fitness value is less than the individual optimal fitness value, it will be updated to the optimal fitness value set of the corresponding vector. And update the value of the phase vector to the corresponding optimal value middle.

[0090] Step 2-7: If the optimal fitness value set There is a fitness value less than the global optimal The fitness value of is set as the global optimal fitness value. , and its corresponding phase vector Assigned as the global optimal vector , and set the counter stable_counter to 0. If it does not exist, the counter is incremented by 1.

[0091] Step 2-8: Repeat the above process from step 2-6 to step 2-7 until the value of stable_counter exceeds the preset threshold stable_threshold, where stable_threshold is 40. The global optimal vector obtained is The antenna pattern formed is as follows Figure 3 shown.

[0092] Step 3: Repeat step 2, and update the latest T interference direction observation data in step 2-1 to achieve dynamic beamforming for dynamic changes in interference direction.

[0093] The interference direction data detected in the second detection is [11683.4307, 13294.4619, 48.6904,11543.5537, 13302.6475, 49.0497, 11403.6758, 13310.8330, 49.4126, 11263.7979,13319.0195, 49.7790, 11123.9199, 13327.2051, 50.1490], and the predicted value is 50.4842, which is used as the interference prediction direction. Repeat step 2 to obtain the global optimal vector The antenna pattern formed is as follows Figure 4 shown.

[0094] The interference direction data detected in the third detection is [11543.5537, 13302.6475, 49.0497,11403.6758, 13310.8330, 49.4126, 11263.7979, 13319.0195, 49.7790, 11123.9199,13327.2051, 50.1490, 10984.0420, 13335.3906, 50.5225], and the predicted value is 50.8618, which is used as the interference prediction direction. Repeat step 2 to obtain the global optimal vector The antenna pattern formed is as follows Figure 5 shown.

[0095] The interference direction data detected in the third detection is [11403.6758, 13310.8330, 49.4126,11263.7979, 13319.0195, 49.7790, 11123.9199, 13327.2051, 50.1490, 10984.0420,13335.3906, 50.5225, 10844.1641, 13343.5771, 50.8997], and the predicted value is 51.2430, which is used as the interference prediction direction. Repeat step 2 to obtain the global optimal vector The antenna pattern formed is as follows Figure 6 shown.

[0096] It can be seen that this method can intelligently realize dynamic beamforming for the dynamic change of interference direction.

[0097] In summary, the present invention utilizes interference historical data to extract interference patterns and realizes prediction, thereby realizing intelligent interference beamforming that combines prediction results with an improved particle swarm method.

Claims

1. An intelligent anti-interference beamforming method based on dynamic prediction of interference, characterized in that: The following steps are involved: Step 1: Construct interference source data collection, build an interference source prediction model based on the autonomous attention fully connected network, and train the interference source prediction model; Step 2: The interference direction is predicted by the interference direction prediction model, and the anti-interference beamforming objective function is constructed. The phase of each antenna array element is calculated based on the particle swarm algorithm to achieve anti-interference beamforming; Step 3: Repeat step 2 to implement dynamic beamforming for dynamic changes in the interference direction.

2. The intelligent anti-interference beamforming method based on dynamic prediction of interference sources according to claim 1 is characterized in that: The specific method of step 1 is: Step 1-1: Take the array antenna as the origin and continuously collect the interference observation data at T+1 time. The interference observation data at each time Including the direction and position of the interference source , Distance to interference source and interference source ; Among them, the interference observation data of the first T moments is recorded as sample data D, D = , the interference source in the observed data at the T+1th moment is As label L; repeat the above process K times to construct a sample set and label sets , where the sample set = , label set = ; The sample set and label set are proportional Split to get the training sample set , test sample set , training label set and the test label set , the lengths are , , and , Indicates rounding down; Step 1-2: Build an interference prediction model, including five linear fully connected layers. The input and output dimensions of the first four linear fully connected layers are both 3T, and the input dimension of the fifth linear fully connected layer is 3T, and the output dimension is 1; The first three linear fully connected layers output the query vector query, the key vector key and the value vector value respectively, from which the attention scores scores are calculated: in, is the transpose of the key vector; Apply the softmax function to the last dimension of scores, the sequence dimension, to generate the attention weight: attention=softmax(scores, dim=−1) Among them, dim=−1 represents the last dimension; Perform matrix multiplication of the attention weight and the value vector value to obtain the weighted value Value: Value = attention·value The fourth linear fully connected layer linearly transforms the weighted value Value to obtain the output out; Input out into the fifth linear fully connected layer, and the output is the predicted value; Step 1-3: Set the loss function to mean square error, the optimizer to Adam, the learning rate to lr, and the optimal loss value to infinity; Set the interference source prediction model to training mode and select the training sample set and training label set Take out B pieces of data, clear the optimizer gradient to 0, input interference to the prediction model, and get B prediction values; use the loss function to compare the B prediction values ​​with the training label set The loss value of the extracted B data is calculated, the gradient of the loss value to the model parameters is calculated, and the model parameters are updated according to the rules of the optimizer algorithm until the training sample set is and training label set The data will be retrieved. Set the interference source prediction model to test mode and set the test sample set Input interference to the prediction model to obtain the prediction results, and use the loss function to compare the prediction results with the test label set. Calculate the loss value of the data in; if the loss value is less than the optimal loss value, save the interference to the prediction model as the optimal model and assign the loss value to the optimal loss value; otherwise, do not save the model and do not assign a value; Repeat the training and testing. When the number of repetitions corresponding to the optimal loss value is greater than the threshold number E, and the number of repetitions is greater than twice the number of repetitions corresponding to the optimal loss value, stop the training.

3. The intelligent anti-interference beamforming method based on dynamic prediction of interference sources according to claim 2 is characterized in that: The specific method of step 2 is: Step 2-1: Input the interference observation data collected continuously for T moments into the interference prediction model to obtain the interference prediction value; Step 2-2: Construct an anti-interference objective function with the input of the function being the preset pointing direction , interference prediction direction and the antenna element phase vector , antenna element phase vector The phase of each antenna element Composition, n=1,2,...,N, N is the number of array elements; the output of the function is the target value: Where d is the antenna element spacing, To round down, To round up; Step 2-3: Construct a set of candidate phase vectors , contains a total of M phase vectors, each phase vector The dimension is N, and the value range is , m=1,2,...,M; Construct a velocity set, which contains M velocity vectors. Each velocity vector The dimension is N; Step 2-4: Set the optimal value of each phase vector ,all Composition Collection , using an alternative phase vector set The value in is Each of Initialize, then Substitute them into the anti-interference objective function respectively, and get M target values ​​in total, which are used as The individual optimal fitness value of M individuals is combined into an optimal fitness value set , The largest individual optimal fitness value is set as the global optimal fitness value , the corresponding phase vector is the global optimal vector ; Step 2-5: Set inertia weight w, cognitive parameters and social parameters , and initialize the counter stable_counter=0 used to check the stability of the global optimal value; Step 2-6: Update each phase vector in the candidate phase vector set The value of: in, and is a random number between [0, 1]; The updated phase vector Limit the values ​​in the range Then, the phase vector Substitute the anti-interference objective function for calculation. If the calculated target value is less than the corresponding individual optimal fitness value, then use the target value to update The corresponding individual optimal fitness value in renew The optimal value of the corresponding phase vector in ; Step 2-7: If the optimal fitness value set There is a fitness value less than the global optimal The fitness value of is assigned to the global optimal fitness value. , and its corresponding phase vector Assign the global optimal vector , and set the counter stable_counter to 0, otherwise add 1 to the counter stable_counter; Step 2-8: Repeat steps 2-6 to 2-7 until the stable_counter value exceeds the preset upper limit. The vector in is the phase of each antenna element.

Images