Self-adaptive beam forming method and system based on long and short term memory network

Abstract

The invention discloses a self-adaptive beam forming method and system based on a long and short term memory network, and aims to avoid complex matrix calculation processes of a covariance matrix or an array manifold matrix and the like when the frequency and the direction of arrival of an unknown incident signal change in real time. The signal after beam forming is directly synthesized based on the input signal data, real-time beam forming of array receiving signals is ensured, the signal-to-noise ratio of the array receiving signals is stably improved, the weak signal detection capability of a receiver is enhanced, a system can capture target signals more accurately and stably, and the purpose of electronic reconnaissance is achieved. According to the invention, when prior information based on reference signals, such as the frequency and the direction of arrival of the incident signals, is difficult to obtain in a low signal-to-noise ratio environment, adaptive beam forming can still be accurately carried out on the array incident signals, and the signals after beam forming are directly output, so that the signal-to-noise ratio of the incident signals is greatly improved, and the signal-to-noise ratio of the incident signals is improved. And the sensitivity of the signal receiver is improved.

Description

Technical Field

[0001] The present invention belongs to the fields of array signal processing and electronic reconnaissance, and particularly relates to an adaptive beamforming method and system based on a long short-term memory network. Background Art

[0002] Adaptive beamforming technology is a signal processing technology that uses an antenna array to perform spatial filtering and enhancement on received signals. By precisely controlling the phases and amplitudes of each element in the antenna array, the antenna array can enhance or suppress signals in a specific spatial direction, and can significantly improve the signal-to-noise ratio of the received signal by using the equivalent antenna aperture brought by the antenna array, thereby improving the sensitivity of the signal receiver. In addition, beamforming also enables the electronic reconnaissance system to respond efficiently and flexibly in a multi-target environment, enabling the system to cover a wider area and intercept information from multiple signal sources in real time by quickly scanning and adjusting the beam direction. This ability enables the electronic reconnaissance system to acquire and process a large amount of signals in real time. In this context, the application of adaptive beamforming technology is gradually becoming a key element in the field of electronic reconnaissance.

[0003] Traditional adaptive beamforming methods are roughly divided into two categories. One is the method based on a reference signal, which requires the transmitting end to send a time-domain reference signal. The other is the blind algorithm, which does not require the transmitting end to provide too much information.

[0004] Currently, most traditional beamforming technologies and adaptive beamforming methods are methods based on reference signals, which require prior knowledge of the frequency and direction of arrival of the incident signals on the array as prior information for beamforming. However, in a low signal-to-noise ratio environment, it is extremely difficult to obtain the frequency and direction of arrival of the incident signals on the array, which directly leads to a significant decline in the performance of the beamforming method based on the reference signal. In order to perform more effective adaptive beamforming, in the field of array signal processing, many researchers have begun to combine adaptive beamforming technology with neural networks and apply them to directions such as electronic reconnaissance and wireless communication.

[0005] Existing solution one: Saeed Mohammadzadeh et al. proposed an adaptive beamforming technology based on a complex-valued convolutional neural network, which can accurately calculate an almost optimal beamformer using the sample covariance matrix without the direction of arrival of the target signal.

[0006] Existing solution two: Ioannis Mallioras et al. used a recurrent neural network based on the gated recurrent unit architecture as an adaptive beamformer. This architecture is an optimal solution to the adaptive beamforming problem, but this architecture uses a large number of gated units, making the network structure too complex and redundant.

[0007] Existing Solution 3: Bai Moyu, Liu Hao, etc. proposed an adaptive beamforming method based on a deep neural network. This method can quickly perform beamforming under large echo data, improving the calculation speed of the adaptive beamforming algorithm to a certain extent.

[0008] Although the research on these existing blind beamforming methods based on deep learning can provide better effects than digital beamforming, most of them focus on ideal signal-to-noise ratio conditions. When in a low signal-to-noise ratio environment, information such as the inherent frequency and direction of arrival of the incident signal is not obvious and is difficult to be obtained by some existing neural networks. This makes it impossible to perform matching in dimensions such as frequency and angle, and unable to concentrate the signal energy on the frequency point or direction of arrival of a certain signal, thus unable to perform effective beamforming and meet some demanding beamforming tasks.

[0009] To solve the problem that it is difficult to obtain prior information such as the frequency and direction of arrival of the incident signal based on the reference signal in a low signal-to-noise ratio environment, resulting in the inability to accurately perform adaptive beamforming on the incident signal, the present invention proposes an adaptive beamforming method and system based on a long short-term memory network. Summary of the Invention

[0010] The purpose of the present invention is to provide an adaptive beamforming method and system based on a long short-term memory network. The method uses the established uniform linear array model to receive spatial domain signals with unknown directions of arrival and frequencies, then constructs a long short-term memory network, and trains the neural network using the spatial domain signals received by the array model and the signal source signals obtained in advance.

[0011] The purpose of the present invention is achieved through the following technical solutions:

[0012] An adaptive beamforming method based on a long short-term memory network, the specific steps are as follows:

[0013] Step 1: Establish a uniform linear array model including N independent antenna elements, with N input channels and one output channel;

[0014] Step 2: Construct a long short-term memory network with two LSTM layers and one fully connected layer;

[0015] Step 3: Obtain a data set according to Step 1 as the input of the network in Step 2;

[0016] Step 4: Set the feature extraction structure of the long short-term memory network in Step 2;

[0017] Step 5: Define a loss function to obtain a loss value;

[0018] Step 6: Initialize hyperparameters;

[0019] Step 7: According to the loss value calculated in Step 5, combined with the gradient descent method and the least mean square error criterion, use the backpropagation algorithm to update the hyperparameters initialized in Step 6;

[0020] Step 8: Use the Adam neural network optimizer to optimize the loss function of the neural network;

[0021] Step 9: Beamforming; Send the array received signal into the trained long short-term memory network to form a beam pointing to the incident angle of the signal, and output the original signal after beamforming.

[0022] Further, in Step 1, the incident angle of the spatial domain signal received by the uniform linear array is θ, and the frequency range of the incident signals that the array can receive is [f 1 , f 2 , then the distance between two adjacent antenna elements can be set to λ / 2, where λ = f 2 / c, and c is the speed of light.

[0023] Further, in the overall structure of the neural network model in Step 2, two LSTM layers appear first, followed by a fully connected layer; the LSTM layer captures and models the long-term dependencies in the array received signal, and effectively extracts and transmits information in the multi-step sequence input to the neural network; the fully connected layer generates the finally predicted signal after beamforming according to the task type; in the neural network model, the LSTM layer uses the tanh function as the activation function to make the network perform non-linear mapping, enabling the LSTM to model complex time series relationships, and the fully connected layer uses the Leaky ReLU function as the activation function.

[0024] Further, in Step 3, select the leftmost antenna element in the array as the reference element of the uniform linear array. The reference element in the array reception model receives a continuous spatial domain signal x(t) of a total of M points. Then the signals received by each element from left to right are expressed as:

[0025]

[0026] Among them, is the phase difference between the signal received by the element that is one element interval d away from the reference element and the signal received by the reference element, and n is the number of element intervals between this element and the reference element; then the signal received by the array reception model is N channels, and each channel has a signal X(t) of a total of M points. Then the signal received by the array has a total of L points (L = N·M), and this signal will be used as the input of the neural network, and the signal is expressed as follows:

[0027] X(t) = [x 1 (t), x 2(t),…,x n (t)], n = 1, 2, …, N

[0028] The labels of the training dataset are replaced by the source signals, obtaining a single-channel signal with a total of M points, that is, the noise-free version y(t) of the signal x(t). This signal will be used as the training label of the long short-term memory network and corresponds one-to-one with the signal received by the array reception model.

[0029] Furthermore, for the signal X(t) with a total of L points input in step four, it is reconstructed into s time steps, denoted as x k , where k = 1, 2, …, s, and each time step has input data with g features (L = s·g). At this time, the feature structure is 1×s×g. Subsequently, it is fed into the first LSTM layer. After being activated by the tanh function, the processing results of all time steps of the LSTM layer are output, that is, h 1 , h 2 , …, h k , …, h s , and the feature structure of the neural network is transformed into 1×s×a. Then, h k , k = 1, 2, …, s is used as the input data of the second LSTM layer. After being activated by the tanh function, the processing result of the last time step of this LSTM layer is taken as the output of the neural network, that is, p s , and the feature structure of the neural network is transformed into 1×b. Finally, the output result p s of the last time step of the second-layer LSTM is input into the fully connected layer. After being processed by the fully connected layer, the final reconstruction result y pre (t) of the neural network is obtained, and the feature structure is 1×M. This signal is the signal after adaptive beamforming.

[0030] Furthermore, in step five, the time-domain minimum mean square error criterion is used as the loss function to constrain the neural network. The specific formula is as follows:

[0031] loss = MMSE(pre(input i ), label i )

[0032] where MMSE(y pre , y) represents calculating the minimum mean square error of the data y pre , y, pre(X) represents the reconstruction output of the neural network for the data X, input i represents an input data of the neural network, and label i is the label data corresponding to input i .

[0033] Further, in Step 6, the initial value of the learning rate is defined before training, and a learning rate decay rate is specified for it. The learning rate decay rate in this part refers to the magnification factor by which the learning rate decays exponentially after the neural network has been trained μ times. The decay process is as follows:

[0034] μ = steps / size

[0035] lr a = lr b ·η Sμ

[0036] where steps is the total number of samples in the training dataset, size is the size of a single training batch, η is the learning rate decay rate, lr b is the learning rate before decay, lr a is the learning rate after decay, and S is the total number of steps that the current neural network has been trained; then, the local variables and global variables used in the neural network training process are initialized with parameters.

[0037] Further, in Step 8, when the loss value reaches a minimum within a set range and oscillates near the minimum value, the training is stopped, indicating that the training is completed, and then the performance of the long short - term memory network is tested.

[0038] A computer device / system, including a memory, a processor, and a computer program stored on the memory, where the processor executes the computer program to implement the steps of an adaptive beamforming method based on a long short - term memory network.

[0039] A computer - readable storage medium, on which a computer program / instructions are stored, and when the computer program / instructions are executed by a processor, the steps of an adaptive beamforming method based on a long short - term memory network are implemented.

[0040] The beneficial effects of the present invention are as follows:

[0041] When it is difficult to obtain prior information based on reference signals such as the frequency and direction of arrival of incident signals in a low - signal - to - noise ratio environment, the present invention can still accurately perform adaptive beamforming on the incident signals of the array and directly output the signals after beamforming, greatly improving the signal - to - noise ratio of the incident signals and thus improving the sensitivity of the signal receiver. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is the overall flowchart of the long short - term memory network;

[0043] Figure 2 is the uniform linear array reception model;

[0044] Figure 3It is the characteristic structure diagram of the long short-term memory network;

[0045] Figure 4 It is the comparison diagram of the effects after processing by the LSTM-ABF method and the CBF method. Specific implementation manner

[0046] The present invention will be further described below with reference to the accompanying drawings.

[0047] An adaptive beamforming method and system based on a long short-term memory network according to the present invention aims at the problem that it is difficult to obtain prior information based on a reference signal such as the frequency and direction of arrival of an incident signal in a low signal-to-noise ratio environment, resulting in the inability to accurately perform adaptive beamforming on the incident signal. An adaptive beamforming method based on a long short-term memory network applicable to the field of electronic reconnaissance is proposed. The core of this method is to simulate the signal received by the array with low signal-to-noise ratio spatial signals with unknown direction of arrival and frequency based on a large amount of spatial statistical information, and use the long short-term memory network to perform adaptive beamforming on the array received signal.

[0048] According to Figure 1 shown, the specific implementation steps of the present invention are as follows:

[0049] Step 1: Establish a uniform linear array model. The array reception model is as Figure 2 shown. This model is planned to be composed of N independent antenna elements, so this model will have N input channels and one output channel. Assume that the incident angle of the spatial signal received by the uniform linear array is θ, and the frequency range of the incident signal that the array can receive is [f 1 , f 2 . Then the distance between two adjacent antenna elements can be set to λ / 2, where λ = f 2 / c, and c is the speed of light.

[0050] Step 2: Construct a long short-term memory network. A long short-term memory network with two LSTM layers and one fully connected layer is established. In the overall structure of the neural network model, the two LSTM layers appear first, and then are followed by one fully connected layer. The role of the LSTM layer is to capture and model the long-term dependencies in the array received signal and effectively extract and transmit information in the multi-step sequence input to the neural network. The role of the fully connected layer is to generate the final prediction result (the signal after beamforming) according to the task type, which enables the model to better adapt to the needs of specific tasks and ensures sufficient generalization. In the neural network model, the LSTM layer uses the tanh function as the activation function to enable the network to perform non-linear mapping, so that the LSTM can model complex time series relationships. The fully connected layer uses the Leaky ReLU function as the activation function, which helps to reduce the complexity of the model, reduce the risk of overfitting, and improve the sensitivity of the network to input features.

[0051] Step 3: Obtain the dataset. Select the antenna element at the leftmost end of the array as the reference element of the uniform linear array. Assume that the reference element in the array reception model receives a continuous spatial domain signal x(t) with a total of M points. Then, the signals received by each element from left to right can be expressed as:

[0052]

[0053] where is the phase difference between the signal received by the element separated from the reference element by an element spacing d and the signal received by the reference element, and n is the number of element spacings between this element and the reference element. Then, the signal received by the array reception model is N channels, and each channel has a signal X(t) with M points. Therefore, the signal received by the array has a total of L points (L = N·M). This signal will be used as the input of the neural network, and the signal is expressed as follows:

[0054] X(t) = [x 1 (t), x 2 (t), …, x n (t)], n = 1, 2, …, N

[0055] The label of the training dataset is replaced by the source signal, obtaining a single-channel signal with M points, that is, the noise-free version y(t) of the signal x(t). This signal will be used as the training label of the long short-term memory network and corresponds one-to-one with the signal received by the array reception model.

[0056] Step 4: Set the feature extraction structure of the long short-term memory network. The overall feature structure of each layer of this part of the neural network is as Figure 3 shown. For an input signal X(t) with a total of L points, it is reconstructed into s time steps, denoted as x k , where k = 1, 2, …, s, and each time step has input data with g features (L = s·g). At this time, the feature structure is 1×s×g. Then, it is sent into the first LSTM layer. After being activated by the tanh function, the processing results of all time steps of the LSTM layer are output, that is, h 1 , h 2 , …, h k , …, h s . The feature structure of the neural network is transformed into 1×s×a. Then, h k , k = 1, 2, …, s are used as the input data of the second LSTM layer. After being activated by the tanh function, the processing result of the last time step of this LSTM layer is taken as the output of the neural network, that is, p s, the feature structure of the neural network is transformed into 1×b (it should be noted that both a and b here are hidden layer parameters of the neural network and do not have fixed values). Finally, the output result p of the last time step of the second-layer LSTM is s input into the fully connected layer, and the final reconstruction result y of the neural network is obtained after passing through the fully connected layer. pre (t), the feature structure is 1×M, and this signal is the signal after adaptive beamforming.

[0057] Step Five: Define the loss function. In the present invention, the time-domain minimum mean square error criterion is used as the loss function to constrain the neural network. The specific formula is as follows:

[0058] loss = MMSE(pre(input i ), label i )

[0059] where MMSE(y pre , y) represents calculating the minimum mean square error of data y pre , y, pre(X) represents the reconstruction output of the neural network for data X, input i represents a certain input data of the neural network, and label i is the corresponding label data of input i .

[0060] Step Six: Hyperparameter initialization. Define the initial value of the learning rate before training and specify a learning rate decay rate for it. The learning rate decay rate in this part refers to the exponential decay multiple of the learning rate after the neural network is trained μ times. The decay process is as follows:

[0061] μ = steps / size

[0062] lr a = lr b ·η Sμ

[0063] where steps is the total number of samples in the training dataset, size is the size of a single training batch, η is the learning rate decay rate, lr b is the learning rate before decay, lr a is the learning rate after decay, and S is the total number of steps that the current neural network has been trained.

[0064] Subsequently, initialize the parameters of the local variables and global variables that may be used during the training process of the neural network.

[0065] Step 7: Update hyperparameters. According to the loss value calculated in Step 5 above, combined with the gradient descent method and the least mean square error criterion, use the backpropagation algorithm to update the hyperparameters (including weights and biases) in the network.

[0066] Step 8: Use a neural network optimizer. In the present invention, Adaptive Moment Estimation (Adam) is adopted as the parameter optimization algorithm for the neural network. The Adam optimizer combines the advantages of the Momentum algorithm and the RMSprop algorithm, and can achieve fast and stable optimization in problems with a large amount of data and parameters.

[0067] Step 9: Beamforming. Feed the array received signal into the trained long short-term memory network to form a beam pointing to the incident angle of the signal, and output the original signal after beamforming.

[0068] When the present invention faces the situation where the frequency and the incoming wave direction of the unknown incident signal change in real time, it no longer performs complex matrix calculation processes such as covariance matrix or array manifold matrix, but directly synthesizes the signal after beamforming based on the input signal data, ensuring real-time beamforming of the array received signal, stably improving the signal-to-noise ratio of the array received signal, enhancing the detection ability of the receiver for weak signals, enabling the system to capture the target signal more accurately and stably, and achieving the purpose of electronic reconnaissance.

[0069] Example 1: To verify the effectiveness of the present invention, the following steps are carried out for testing:

[0070] Step 1: Establish a uniform linear array model. In the test session, use eight independent omnidirectional antennas to form a uniform linear array, so this model will have N input channels and one output channel. Assume that the wavelength of the received signal is λ, and the distance between two adjacent antenna elements is λ / 2, where λ is the wavelength corresponding to the 2.3 GHz signal. In addition, assume that the incident angle of the spatial signal received by the uniform linear array model is set as θ, the value range of θ is [-30°, 30°], and the frequency of the spatial signal received is in the range of [1.8 GHz, 2.3 GHz].

[0071] Step 2: Construct a long short-term memory network. Build a long short-term memory network with two LSTM layers and one fully connected layer. In the overall structure of the neural network model, the two LSTM layers appear first, followed by one fully connected layer. The role of the LSTM layer is to capture and model the long-term dependencies in the array received signals and effectively extract and transmit information in the multi-step sequence input to the neural network. The role of the fully connected layer is to generate the final prediction result (the signal after beamforming) according to the task type, which enables the model to better adapt to the requirements of specific tasks and ensures sufficient generalization. In the neural network model, the LSTM layer uses the tanh function as the activation function to enable the network to perform non-linear mapping, enabling the LSTM to model complex temporal relationships. The fully connected layer uses the Leaky ReLU function as the activation function, which helps to reduce the complexity of the model, reduce the risk of overfitting, and improve the sensitivity of the network to input features.

[0072] Step 3: Obtain the dataset. Select the antenna element at the leftmost end of the array as the reference element of the uniform linear array. During the training process of the long short-term memory network, select 30,000 randomly generated single-frequency signals to form the training set. The range of the incident angle of the signals in the training set is [-30°, 30°], and the range of the frequency is [1.8 GHz, 2.3 GHz]. Each signal has 400 points in total, and eight elements correspond to eight channels. The signals received by each element are spliced to obtain an input data of 3,200 points in total. The label corresponding to the input data is based on the signal source signal in the same time period, and the length of the label is 400 points for a single channel.

[0073] Step 4: Set the feature extraction structure of the long short-term memory network. The overall feature structure of each layer of this part of the neural network is as Figure 3 shown. For an array received signal with a total of 3,200 points in input, reconstruct it into 64 time steps, and each time step has an input data with 50 features. At this time, the feature structure is 1×64×50. Then send it into the first LSTM layer. After being activated by the tanh function, output the processing results of all time steps of the LSTM layer. The feature structure of the neural network is transformed into 1×64×256. Then send it into the second LSTM layer. After being activated by the tanh function, output the processing result of the last time step of this LSTM layer. The feature structure of the neural network is transformed into 1×500 (it should be noted that both 256 and 500 here are the hidden layer parameters of the neural network and do not have fixed values). Finally, input the output result of the second LSTM layer into the fully connected layer. After being processed by the fully connected layer, obtain the final reconstruction result of the neural network, and the feature structure is 1×400. Thus, the reconstructed signal y output by the long short-term memory network with a length of 1×400 is obtained. pre(t), this signal is the signal after adaptive beamforming.

[0074] Step Five: Define the loss function. In the present invention, the time-domain minimum mean square error criterion is adopted as the loss function to constrain the neural network. The specific formula is as follows:

[0075] loss = MMSE(pre(input i ), label i )

[0076] where MMSE(y pre , y) represents calculating the minimum mean square error of data y pre , y, pre(X) represents the reconstructed output of the neural network for data X, input i represents a certain input data of the neural network, and label i is the label data corresponding to input i .

[0077] Step Six: Initialize hyperparameters. Before training, define the initial learning rate as 0.001, the learning rate decay rate as 0.9, and define that the learning rate decays once every 10,000 training steps. Then initialize the parameters of the local variables and global variables that may be used during the training process of the neural network.

[0078] Step Seven: Update hyperparameters. According to the loss value calculated in Step Five above, combined with the gradient descent method and the minimum mean square error criterion, use the backpropagation algorithm to update the hyperparameters (including weights and biases) in the network.

[0079] Step Eight: Use the neural network optimizer. In the present invention, Adaptive Moment Estimation (Adam) is adopted as the parameter optimization algorithm of the neural network to optimize the loss function of the neural network. When the loss value reaches a minimum value within a certain range and oscillates near this minimum value, stop training, indicating that the training is completed. Then test the performance of this long short-term memory network.

[0080] Step Nine: Beamforming. After the parameters in the long short-term memory network are trained, input the signal data in the test set into the neural network model, and a directional beam that is expected to be synthesized can be obtained.

[0081] In the present invention, there are 3,000 signals in the test set, and the value ranges of the signal frequency and the incident direction are consistent with those of the training set data. After inputting the training set data into the long short-term memory network, the reconstruction result of the test set by the long short-term memory network is obtained. After testing the signal-to-noise ratio of the reconstruction result of the test set and taking the average value, the average signal-to-noise ratio gain brought by the long short-term memory network for the test set signals is about 19.6 dB, which is much higher than the 9 dB signal-to-noise ratio gain that can be brought by the conventional beamforming. Among them, when obtaining this gain by the two methods, the adaptive beamforming method based on the long short-term memory network is obtained when the incident direction and frequency of the incident signal are unknown, while the conventional beamforming method is obtained when the incident direction and frequency of the signal are known. It can be clearly seen from this that the effect of the adaptive beamforming method based on the long short-term memory network is better than that of the conventional beamforming method.

[0082] Figure 4 Fig. shows the comparison diagram of the output signal spectra after beamforming of a single-carrier signal with a signal-to-noise ratio of 0 dB and an incident angle of 3° incident on the array by the conventional beamforming (CBF) and the adaptive beamforming method based on the long short-term memory network respectively. Among them, the main lobe of the beam pattern of the CBF method is set to 0°. The angle deviation set in this part is to simulate the situation of beamforming for incident signals with unknown incident directions in a low signal-to-noise ratio environment. It can be clearly seen from the figure that when the incident direction of the unknown incident signal is unknown in a low signal-to-noise ratio environment, the spectral noise floor of the signal processed by the LSTM-ABF method proposed in the present invention is significantly lower than the result processed by the CBF method. Moreover, since the main lobe of the beam pattern of the CBF method does not align with the incident direction of the incident signal, the power of the main peak of the signal spectrum processed by it is lower than the power of the main peak of the signal spectrum processed by the LSTM-ABF method.

[0083] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An adaptive beamforming method based on long short-term memory network, characterized in that: The specific steps are as follows: Step 1: Establish a uniform linear array model including N independent antenna elements, and having N input channels and one output channel; Step 2: Build a long short-term memory network with two LSTM layers and one fully connected layer; Step 3: Obtain the data set according to step 1 as the input of the network in step 2; Step 4: Set the feature extraction structure of the long short-term memory network in step 2; Step 5: Define the loss function and get the loss value; Step 6: Initialize hyperparameters; Step 7: Based on the loss value calculated in step 5, the back propagation algorithm is used to update the hyperparameters initialized in step 6, combining the gradient descent method and the minimum mean square error criterion; Step 8: Use the Adam neural network optimizer to optimize the loss function of the neural network; Step 9: Beamforming: Send the array received signal into the trained long short-term memory network to form a beam pointing to the signal incident angle, and output the original signal after beamforming.

2. The adaptive beamforming method based on long short-term memory network according to claim 1, characterized in that: In step 1, the incident angle of the spatial signal received by the uniform linear array is θ, and the frequency range of the incident signal that the array can receive is [f1, f2]. Then the distance between two adjacent antenna elements can be set to λ / 2, where λ=f2 / c, and c is the speed of light.

3. The adaptive beamforming method based on long short-term memory network according to claim 1, characterized in that: In the overall structure of the neural network model, in step 2, two LSTM layers appear first, followed by a fully connected layer; the LSTM layer captures and models the long-term dependencies in the array received signals, and effectively extracts and transmits information in the multi-step sequence of the input neural network; the fully connected layer generates the final predicted beamformed signal according to the task type; In the neural network model, the LSTM layer uses the tanh function as the activation function to enable the network to perform nonlinear mapping, so that the LSTM can model complex temporal relationships. The fully connected layer uses the Leaky ReLU function as the activation function.

4. The adaptive beamforming method based on long short-term memory network according to claim 1, characterized in that: In step 3, the leftmost antenna element in the array is selected as the reference element of the uniform linear array. The reference element in the array receiving model receives a continuous spatial signal x(t) with a total of M points. The signals received by each element from left to right are expressed as: in, It is the phase difference between the signal received by the reference element and the array element separated by an array element interval d, and n is the number of array element intervals between the array element and the reference element; therefore, the signal received by the array receiving model is N channels, and each channel has a total of M points of signal X(t), so the array receives a total of L points (L = N·M), and the signal will be used as the input of the neural network. The signal is expressed as follows: X(t)=[x1(t),x2(t),…,x n (t)],n=1,2,…,N The label of the training data set is replaced by the signal source signal, and a single-channel signal with M points is obtained, that is, the noise-free version y(t) of the signal x(t). This signal will be used as the training label of the long short-term memory network, corresponding one-to-one to the signal received by the array receiving model.

5. The adaptive beamforming method based on long short-term memory network according to claim 1, characterized in that: The input signal X(t) with L points in step 4 is reconstructed into s time steps, set as x k , where k = 1, 2, …, s, and each time step has g features (L = s g) of input data. At this time, the feature structure is 1×s×g, which is then sent to the first LSTM layer. After being activated by the tanh function, the processing results of all time steps of the LSTM layer are output, that is, h1, h2, …, h k ,…,h s , the feature structure of the neural network is transformed into 1×s×a, and then h k , k=1,2,…,s is used as the input data of the second LSTM layer. After being activated by the tanh function, the processing result of the last time step of the LSTM layer is taken as the output of the neural network, that is, p s , the feature structure of the neural network is transformed into 1×b, and finally the output result p of the last time step of the second layer LSTM is s Input to the fully connected layer, and after processing by the fully connected layer, the final reconstruction result y of the neural network is obtained pre (t), the characteristic structure is 1×M, and this signal is the signal after adaptive beamforming.

6. The adaptive beamforming method based on long short-term memory network according to claim 1, characterized in that: The step 5 uses the time domain minimum mean square error criterion as the loss function to constrain the neural network. The specific formula is as follows: loss=MMSE(pre(input i ),label i ) Among them, MMSE(y pre ,y) represents the calculated data y pre , the minimum mean square error of y, pre(X) represents the reconstructed output of the neural network for data X, input i Represents an input data of the neural network, label i is the same as input i The corresponding label data.

7. The adaptive beamforming method based on long short-term memory network according to claim 1, characterized in that: The step 6 defines the initial value of the learning rate before training and specifies a learning rate decay rate for it. The learning rate decay rate in this part refers to the multiple of the exponential decay of the learning rate after each μ training of the neural network. The decay process is as follows: μ=steps / size lr a =lr b ·η Sμ Among them, steps is the total number of samples in the training data set, size is the size of a single training batch, η is the learning rate decay rate, lr b is the learning rate before decay, lr a is the attenuated learning rate, S is the total number of steps that the current neural network has been trained; then the parameters of the local variables and global variables used in the neural network training process are initialized.

8. The adaptive beamforming method based on long short-term memory network according to claim 1, characterized in that: In step eight, when the loss value reaches a minimum value within the set range and oscillates near the minimum value, the training is stopped, indicating that the training has been completed, and the performance of the long short-term memory network is tested again.

9. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images