Source DOA estimation method based on fitting strategy

Abstract

The invention provides a method for estimating a direction of arrival angle based on a fitting strategy. The method comprises the following steps: S1, a receiving end receives a multi-target signal through an array antenna; s2, performing data compression processing on the data and performing feature extraction at the same time; s3, randomly dividing the data into a training set, a verification set and a test set; s4, training the network by using the training set data, updating parameters of the network, and constructing an FCNN neural network model; and S5, performing DOA estimation on the test set data by using the designed FCNN neural network model. The DOA estimation method can infinitely approach the real angle of arrival theoretically, and compared with a mainstream method for classifying by using a neural network, the DOA estimation real-time performance can be ensured, and the estimation precision can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of array signal processing, and particularly relates to a method for estimating the direction of arrival (DOA) of a signal source based on a fitting strategy. Background Art

[0002] Array signal processing realizes the acquisition and processing of signals in space by arranging multiple sensing devices at different positions in space in a specific manner. Among them, the direction of arrival estimation is a key technology in array signal processing and has wide applications in many military and civilian fields such as radar, wireless communication, smart antennas, navigation, and guidance.

[0003] With the development of technology, the direction of arrival estimation is constantly progressing. The current development trend is to combine with deep learning technology more to improve the adaptability and robustness of the algorithm. Most use the strategy of multi-classification by neural networks, but if the number of classification categories is small, the estimation accuracy is low; if the number of classification categories is large, the training time is too long and the classification accuracy rate decreases.

[0004] In 1989, K Hornik confirmed a key property of neural network algorithms through research: if the hidden layer has enough neurons, this algorithm can achieve any required accuracy to simulate various complex continuous functions. This basic architecture based on neural networks is constructed by non-linear elements and has strong non-linear mapping and adaptability. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides a method for estimating the DOA of a signal source based on a fitting strategy, which specifically includes the following steps:

[0006] S1. The receiving end receives multi-target signals through an array antenna;

[0007] The array antenna is a uniform linear array ULA, which receives multi-target signals X with a size of M×N, where M is the number of array elements and N is the number of snapshots;

[0008] S2. Perform data compression processing on the multi-target signals X and extract features;

[0009] The signal received at time t is

[0010] X(t) = [x1(t), x2(t), … x M (t)] T

[0011] where x1(t), x2(t), … x M (t) respectively represent the signals received by the 1st, 2nd, …, Mth array elements of the array, and T represents matrix transpose;

[0012] Extract the real and imaginary parts of X(t) and combine them into a new vector s(t).

[0013] s(t) = [Re(X(t)); Im(X(t))]

[0014] Calculate the covariance matrix R

[0015]

[0016] where H represents the conjugate transpose of the matrix, and the matrix size of R is 2M×2M;

[0017] Select the lower triangular matrix of the covariance matrix R, remove the zero elements in the lower triangular part, and reshape it into a column vector z;

[0018] Perform normalization processing on z

[0019] S = (z - mean(z) / std(z))'

[0020] where S represents the data after normalization processing of Z, mean(·) represents calculating the arithmetic mean of the data, std(·) represents calculating the standard deviation of the data, and "'" represents the conjugate transpose of the matrix;

[0021] S3. Randomly divide the processed S data in step 2 into a training set, a validation set, and a test set;

[0022] S4. Use the training set data to train the network, update the network parameters, and construct an FCNN neural network model;

[0023] Use the BP algorithm to update the multi-layer neural network, perform forward propagation of the input data through the network to obtain the output result, and according to Calculate the loss function MSE, where L represents the total number of training samples, θ(l) represents the actual true angle of the l-th training sample, is the estimated value of the actual true angle of the l-th training sample, and evaluate the difference between the output and the true value; if the error continuously decreases, then use the gradient descent method to adjust the weights, and repeat the processes of forward propagation, calculating the loss, backpropagation, and updating the weights; if the loss function converges or reaches the predetermined number of iterations, then use the test samples for testing and apply the designed network model;

[0024] S5. Apply the designed FCNN neural network model to perform DOA estimation on the test set data;

[0025] The designed FCNN neural network model is such that each neuron is connected to all neurons in the previous layer but not to other neurons in the same layer; the fully connected neural network structure is connected through "weighted summation"; in the fully connected neural network structure, each arrow has a weight w, x is the value of the neuron, and the value of the next layer neuron is y = ∑w i x i + b, where x1, x2, x3, …, x i are input nodes, w1, w2, w3, … w i are weights, b is the bias, and then a non - linear mapping is performed, that is, the output output = f(∑w i x i + b), where f(·) represents the activation function.

[0026] In step S3 of a specific embodiment of the present invention, the S data is divided into a training set, a validation set, and a test set according to the ratio of 6:2:2.

[0027] In step S5 of another specific embodiment of the present invention, the activation function is the ReLU function.

[0028] The present invention proposes a method for estimating the DOA of a signal source based on a fitting strategy, which uses a fully connected neural network (FullConnect Neural Network, FCNN) to estimate the DOA of the signal. The purpose is to ensure the real - time performance of the DOA estimation and improve the estimation accuracy at the same time.

[0029] The present invention uses a fitting method and uses the fitting strategy of a fully connected neural network (Full Connect Neural Network, FCNN) to estimate the DOA of the signal. From a theoretical perspective, this can make it as close as possible to the real angle. By fitting, a model is obtained, and using this model for DOA estimation can ensure the real - time performance and robustness of the algorithm, and improve the accuracy of direction finding at the same time.

[0030] The advantages of the present invention are as follows:

[0031] 1. The specific method for data compression processing and feature extraction of the multi - target signal X in step S2 is proposed for the first time in the present invention;

[0032] 2. Most of the existing technologies for direction estimation use a convolutional neural network based on classification. The present invention proposes for the first time to use a fitting method and use the fitting strategy of a fully connected neural network to estimate the DOA of the signal, making it as close as possible to the real angle. By fitting, an FCNN model is obtained, and using this model for DOA estimation can ensure the real - time performance and accuracy of direction finding. Brief Description of the Drawings

[0033] Figure 1 It is a schematic flowchart of a method for estimating the DOA of a signal source based on a fitting strategy according to the present invention;

[0034] Figure 2 It is the uniform linear array structure of the present invention;

[0035] Figure 3 It is the neural network structure of the present invention;

[0036] Figure 4 It is the structure of a single node in the neural network of the present invention;

[0037] Figure 5 It is the process of the MSE changing with the number of Epochs;

[0038] Figure 6 It is the error histogram;

[0039] Figure 7 It is the R coefficient of the new data after training. Detailed implementation manners

[0040] To achieve the above object, the present invention provides a method for estimating the DOA of a signal source based on a fitting strategy, as Figure 1 shown, which specifically includes the following steps:

[0041] S1. The receiving end receives multi-target signals through an array antenna;

[0042] The array antenna is a uniform linear array (ULA). The received multi-target signal X has a size of M×N, where M is the number of array elements and N is the number of snapshots.

[0043] S2. Perform data compression processing on the multi-target signal X and extract features;

[0044] The signal received at time t is

[0045] X(t) = [x1(t), x2(t), … x M (t)] T

[0046] where x1(t), x2(t), … x M (t) respectively represent the signals received by the 1st, 2nd, … Mth array elements of the array, and T represents the matrix transpose.

[0047] Extract the real part and the imaginary part of X(t) respectively, and combine them into a new vector s(t)

[0048] s(t) = [Re(X(t)); Im(X(t))]

[0049] Find the covariance matrix R

[0050]

[0051] where H represents the conjugate transpose of the matrix, and the matrix size of R is 2M×2M.

[0052] Select the lower triangular matrix of the covariance matrix R, remove the zero elements in the lower triangular part, and reshape it into a column vector z. Reshaping into a column vector is the default element arrangement order in Matlab, using column-major order.

[0053] Perform normalization processing on z

[0054] S = (z - mean(z) / std(z))'

[0055] where S represents the data after normalizing Z, mean(·) represents calculating the arithmetic mean of the data, std(·) represents calculating the standard deviation of the data, and "'" represents the conjugate transpose of the matrix.

[0056] S3. Randomly divide the processed S data in step 2 into a training set, a validation set, and a test set;

[0057] Divide the S data, for example, in a ratio of 6:2:2 into a training set, a validation set, and a test set.

[0058] S4. Use the training set data to train the network, update the network parameters, and construct an FCNN neural network model;

[0059] Use the BP algorithm to update the multi-layer neural network. First, propagate the input data forward through the network to obtain the output result. According to Calculate the loss function MSE, where L represents the total number of training samples, θ(l) represents the actual true angle of the l-th training sample, is the estimated value of the actual true angle of the l-th training sample, and then evaluate the difference between the output and the true value. If the error continuously decreases, then use the gradient descent method to adjust the weights, repeat the forward propagation, calculate the loss, backpropagation, and update the weights process; if the loss function converges or reaches the predetermined number of iterations, then use the test samples for testing and apply the designed network model.

[0060] S5. Apply the designed FCNN neural network model to perform DOA estimation on the test set data;

[0061] For the designed FCNN neural network model, each neuron is connected to all neurons in the previous layer but not to other neurons in the same layer. As Figure 4 shown, the fully connected neural network structure is connected through "weighted summation". Figure 4Among them, each arrow is associated with a weight w, x is the value of the neuron, and the value of the next-layer neuron is y = ∑w i x i + b, where x1, x2, x3, …, x i are the input nodes, w1, w2, w3, …, w i are the weights, b is the bias, and then a non-linear mapping is performed, that is, the output output = f(∑w i x i + b), where f(·) represents the activation function. The activation function is, for example, the ReLU function.

[0062] The present invention will be clearly and completely described below in conjunction with the accompanying drawings and MATLAB simulation examples.

[0063] As Figure 2 is a uniform linear array, M array elements are evenly arranged on a straight line, and there are K far-field narrowband signals incident, with the signal wavelength being λ. Assume M = 6 and the array spacing is d = 0.4. The signal sources are set to two signal sources with equal power, K = 2, the signal-to-noise ratio SNR = 5 dB, and the number of snapshots is set to 64. The arrival angle range is [-90°, 90°]. The data input to the neural network is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. Since there are 6 arrays, the covariance matrix is a 6×6 matrix. Take its lower triangular matrix and reshape the non-zero elements into a row. The length of this row of data is M×(2×M + 1) = 78. 80% of the training data in the dataset, and the label is the true arrival direction angle. The neural network structure is as Figure 3 shown. The network hidden layer is set with three layers, each layer is set with 25 nodes, and the structure of a single node is as Figure 4 shown.

[0064] Epoch represents the process of completing a full forward propagation and backward propagation for the entire training dataset. The mean square error (MSE) is defined as:

[0065]

[0066] In the formula, L represents the total number of training samples, θ(l) represents the actual true angle of the l-th training sample, is the estimated value of the actual true angle of the l-th training sample.

[0067] The performance of the MSE metric in each Epoch during the training, validation, and testing processes is as Figure 5As shown, the smaller the value of MSE, the better the prediction, and 0 indicates no error. The circle corresponding to the 32nd Epoch in the figure represents the best mean square error value of the validation set corresponding to the iteration times at this time, which also shows that the training effect is the most ideal when the network is trained to the 32nd Epoch. It can be seen from this figure that the performance of the training set is very good. After training for 5 Epochs, the MSE drops to 10 -2 After training for 12 Epochs, the MSE drops to 10 -4 Below, the validation set reaches the optimal performance at the 32nd Epoch.

[0068] The error histogram shows the error between the predicted output and the target output. As Figure 6 shown, the smaller the Errors, the better the training effect. The abscissa has positive and negative values, indicating the error between the predicted output and the target output; the ordinate represents the number of examples that conform to the corresponding error. It can be seen from the figure that the abscissa value is very small, indicating that the training error is very small. In particular, the error of most data is within 0.00224, proving that the performance of this network is good.

[0069] The fitting performance of the network training set, validation set and test set is visually displayed. As Figure 7 shown, where R represents the fitting result. If the R value approaches 1, it means that the correlation between the prediction and the actual output is relatively high and the prediction effect is good; on the contrary, if the R value is close to 0, it means that the connection between the two is relatively random and the prediction performance is poor. The figure shows the correlation of the trained data on the training set, validation set, test set and the overall result. The ordinate represents the fitting degree between the prediction and the target, and the abscissa represents the output of the target. The R coefficient in the training set is 0.99999, the R coefficient in the validation set is 0.99996, the R coefficient in the test set is 0.99996, and the overall R coefficient is 0.99998. From these fitting coefficients, the fitting effect of this neural network is very good and the direction finding accuracy is guaranteed.

[0070] Therefore, the DOA algorithm based on the fitting strategy proposed by the present invention can ensure the real-time performance of DOA estimation and effectively improve the estimation accuracy.

Claims

1. A method for estimating the DOA of a signal source based on a fitting strategy, characterized in that, Specifically, it includes the following steps: S1. The receiving end receives multi-target signals through an array antenna; The array antenna is a uniform linear array ULA, which receives multi-target signals X with a size of M×N, where M is the number of array elements and N is the number of snapshots; S2. Perform data compression processing on the multi-target signals X and extract features; The signal received at time t is X(t) = [x1(t), x2(t), … x M (t)] T Among them, x1(t), x2(t), … x M (t) respectively represent the signals received by the 1st, 2nd, … Mth array elements of the array, and T represents matrix transpose; Extract the real part and the imaginary part of X(t) respectively and combine them into a new vector s(t) s(t) = [Re(X(t)); Im(X(t))] Calculate the covariance matrix R where H represents the conjugate transpose of the matrix, and the size of the matrix R is 2M×2M; Select the lower triangular matrix of the covariance matrix R, remove the zero elements in the lower triangular part, and reshape it into a column vector z; Perform normalization processing on z S = (z - mean(z) / std(z))' where S represents the data after normalizing Z, mean(·) represents calculating the arithmetic mean of the data, std(·) represents calculating the standard deviation of the data, and " ' " represents the conjugate transpose of the matrix; S3. Randomly divide the processed S data in step 2 into a training set, a validation set, and a test set; S4. Use the training set data to train the network, update the parameters of the network, and construct a FCNN neural network model; Update the multi-layer neural network using the BP algorithm, perform forward propagation of the input data through the network to obtain the output result, and according to Calculate the loss function MSE, where L represents the total number of training samples, and θ(l) represents the actual true angle of the l-th training sample. is the estimated value of the actual true angle of the l-th training sample, and evaluate the difference between the output and the true value; if the error continues to decrease, then use the gradient descent method to adjust the weights, and repeat the processes of forward propagation, loss calculation, backpropagation, and weight update; if the loss function converges or reaches the predetermined number of iterations, then use the test samples for testing and apply the designed network model; S5. Apply the designed FCNN neural network model to perform DOA estimation on the test set data; The designed FCNN neural network model has each neuron connected to all neurons in the previous layer but not to other neurons in the same layer; the fully connected neural network structure is connected through "weighted summation"; in the fully connected neural network structure, each arrow has a weight w, x is the value of the neuron, and the value of the next layer neuron is y = ∑w i x i + b, where, x1, x2, x3, …, x i are input nodes, w1, w2, w3, … w i are weights, b is the bias, and then a non - linear mapping is performed, that is, the output output = f(∑w i x i + b), where f(·) represents the activation function.

2. The method for estimating the DOA of a signal source based on a fitting strategy according to claim 1, wherein, In step S3, the S data is divided into a training set, a validation set, and a test set in a ratio of 6:2:

2.

3. The method for estimating the DOA of a signal source based on a fitting strategy according to claim 1, characterized in that, In step S5, the activation function is the ReLU function.

Images