Multi-layer convolutional neural network DOA estimation method based on spatial feature feedback

Abstract

The present invention discloses a multi-layer convolutional neural network DOA estimation method based on spatial feature feedback, firstly calculating the covariance matrix of the received signal, then decomposing the covariance matrix into two channels of real part and imaginary part and inputting them into the multi-layer DOA estimation model to obtain the final arrival angle estimation value; the multi-layer DOA estimation model includes L classification layers arranged in sequence, wherein the first classification layer is a coarse classification module, and each of the other classification layers has a fine classification module; the number of classification layers containing the fine classification modules is determined by the estimation resolution. The present invention improves the resolution and meets the estimation requirements by refining the DOA estimation range layer by layer, avoiding the high complexity problem caused by over-fine division, reducing the calculation cost, and also has the advantages of strong approximation ability and fast convergence speed.

Description

Technical Field

[0001] The invention belongs to the technical field of array signal processing, and in particular relates to a DOA estimation method. Background Art

[0002] Direction of Arrival (DOA) estimation is one of the core issues in the field of signal processing and is widely used in many fields such as radar, sonar, communication and medical imaging. Its main purpose is to determine the direction of arrival of the signal source by receiving the output data of the array sensor. Traditional DOA estimation methods mainly rely on spatial signal processing techniques, such as minimum variance distortion-free response and MUSIC algorithm. These classic methods perform estimation by analyzing the spatial correlation of signals. However, in complex environments, it is often necessary to divide the signal space in detail to obtain high-precision estimates. Excessive division leads to a significant increase in the computational complexity of the model, which greatly affects the efficiency of real-time processing and the overall performance of the system.

[0003] With the continuous advancement of array processing technology, DOA estimation methods have undergone continuous development from classical beamforming methods to high-resolution subspace methods, sparse representation methods, and deep learning methods. Although deep learning, especially convolutional neural networks (CNNs), has shown superiority in feature extraction and can automatically learn the spatial features of input signals, thereby improving the accuracy of DOA estimation, existing CNN models often use fixed classification strategies in applications and fail to effectively solve the over-division problem caused by traditional methods. This makes the model still face the challenges of increased complexity and computational cost under high precision requirements. Summary of the invention

[0004] The present invention proposes a multi-layer convolutional neural network DOA estimation method based on spatial feature feedback, the purpose of which is to solve the complexity and computational cost problems caused by over-fine division of traditional models and to improve the accuracy and efficiency of DOA estimation.

[0005] The technical solution of the present invention is as follows:

[0006] A multi-layer convolutional neural network DOA estimation method based on spatial feature feedback first calculates the covariance matrix of the received signal, then decomposes the covariance matrix into two channels, real part and imaginary part, and inputs them into the multi-layer DOA estimation model to obtain the final arrival angle estimation value;

[0007] The multi-layer DOA estimation model includes sequentially setting classification layers, wherein the first classification layer is a coarse classification module, and each of the other classification layers has a fine classification module; the number of classification layers containing the fine classification modules is determined by the estimation resolution; each classification layer outputs an arrival angle estimation value, and the first The arrival angle estimate output by the classification layer is recorded as , the arrival angle estimate output by the last classification layer That is the output value of the multi-layer DOA estimation model;

[0008] The input of the coarse classification module contains two channels, namely the real part of the covariance matrix and the imaginary part of the covariance matrix, and the output is the arrival angle estimation value output by the classification layer ;

[0009] For classification layers, The input of the fine classification module contains 4 channels, namely: the real part of the covariance matrix, the imaginary part of the covariance matrix, the arrival angle estimate output by the previous classification layer and the feedback term calculated based on the arrival angle estimate of the previous classification layer through spatial features , the output of the fine classification module is the accuracy increment The classification layer outputs the estimated arrival angle as .

[0010] As a further improvement of the multi-layer convolutional neural network DOA estimation method based on spatial feature feedback: each additional layer of fine classification module will increase the resolution of the precision increment by 10 times relative to the output of the previous classification layer; As the number of classification layers containing the refined classification module, Indicates the rounding operation. is the resolution to be achieved.

[0011] As a further improvement of the multi-layer convolutional neural network DOA estimation method based on spatial feature feedback: Assume that the received signal is ,in For the signal vectors of time snapshots, each signal vector Contains the signals received by all array elements at this time snapshot, and the calculation method of the signal covariance matrix is:

[0012] ;

[0013] in, for The conjugate transpose of .

[0014] As a further improvement of the multi-layer convolutional neural network DOA estimation method based on spatial feature feedback: the coarse classification module and the fine classification module each include a feature extraction submodule and a classification submodule.

[0015] As a further improvement of the multi-level convolutional neural network DOA estimation method based on spatial feature feedback: in the coarse classification module, the real part of the covariance matrix and the imaginary part of the covariance matrix are input into the feature extraction submodule to obtain a feature map, and then the feature map is input into the classification submodule to obtain a probability distribution estimate. The category corresponding to the maximum probability in the probability distribution estimate is the arrival angle estimate output by the coarse classification module. ;

[0016] In the fine classification module, the real part of the covariance matrix, the imaginary part of the covariance matrix, the arrival angle estimate output by the previous classification layer, and the feedback item obtained based on the arrival angle estimate of the previous classification layer are input into the feature extraction submodule to obtain a feature map, and then the feature map is input into the classification submodule to obtain an accuracy increment.

[0017] As a further improvement of the multi-layer convolutional neural network DOA estimation method based on spatial feature feedback: the feature extraction submodule includes feature extraction blocks arranged in sequence, and each feature extraction block is arranged in sequence with a convolution layer, an activation function and a pooling layer;

[0018] Set up The feature map input in the feature extraction block is obtained after the convolution layer and activation function. ,but:

[0019] ;

[0020] in, represents the convolution operation, and are the convolution kernel weights and bias terms of the convolution layer in the feature extraction block respectively; Extract the nonlinear activation function in the block for this feature; For the The feature map output by the feature extraction block is After The feature map output after the pooling layer in the feature extraction block.

[0021] As a further improvement of the multi-level convolutional neural network DOA estimation method based on spatial feature feedback: in the classification submodule, the feature map from the feature extraction submodule is first flattened into a one-dimensional vector , the one-dimensional vector Include elements, is the total number of categories of the classification submodule, and the categories are obtained by dividing the range of angles output by the classification layer according to the resolution to be output by the current classification layer; then the one-dimensional vector Input to the fully connected layer for feature fusion to obtain the fusion result ,in and are the weight matrix and bias vector of the fully connected layer, respectively. Include elements; then merge the results Input into the softmax classifier to obtain the probability distribution estimate. The Elements , the probability value obtained by the softmax classifier is , That is the The probability value of each category; the category corresponding to the maximum probability in the probability distribution estimate is the output result of the classification submodule.

[0022] As a further improvement of the multi-layer convolutional neural network DOA estimation method based on spatial feature feedback: classification layers, , and obtain the estimated angle of arrival of this layer After that, the corresponding feedback term is calculated by the following method :

[0023] First, calculate the steering vector :

[0024] ;

[0025] in, , is the array element spacing of the received signal, is the signal wavelength, is the number of array elements, is the sequence index of the array antenna;

[0026] Then, calculate the arrival angle estimate for this layer With the covariance matrix Deviation in space :

[0027] ;

[0028] in, Represents the conjugate transpose of a matrix;

[0029] Finally, the deviation Perform normalization processing to obtain the feedback item :

[0030] ;

[0031] in, is the Frobenius norm.

[0032] As a further improvement of the multi-layer convolutional neural network DOA estimation method based on spatial feature feedback: the multi-layer DOA estimation model updates model parameters by training from coarse classification module to fine classification module layer by layer, first training the coarse classification module, and then training each fine classification module in turn; when the previous classification model is trained, the training of the next classification module is started;

[0033] For each classification module, training is completed by iteratively updating the network parameters of the classification module. In each iteration, the loss value is first calculated based on the arrival angle estimate and the true angle value output by the classification module, and then the back propagation algorithm is used to calculate the gradient of the loss value with respect to the network parameters. Then, the gradient descent algorithm is used to multiply the gradient by the step size to obtain the amplitude of the parameter update, and the network parameters of the current classification module are updated accordingly. When the loss value converges or reaches the preset number of iterations, the training is stopped.

[0034] As a further improvement of the multi-level convolutional neural network DOA estimation method based on spatial feature feedback: for each classification module, before the current classification module starts iterative updating, the step size is set to the initial value; The estimated angle of arrival after iterations is , the true angle value is , then The deviation after iterations is , No. The step size used in the next iteration will be updated to ,in is the step size adjustment coefficient, which is used to control the amplitude of the step size adjustment. is a sign function.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The multi-layer DOA estimation model of the present invention adopts a multi-layer classification architecture, which improves the resolution and meets the estimation requirements by increasing the number of layers and refining the DOA estimation range layer by layer. Each layer of the classification module is only accurate to the specified resolution, and the next layer further determines the accuracy increment of the higher resolution on this basis to obtain a more accurate DOA estimation, thereby avoiding the high complexity problem caused by over-division and reducing the computational cost.

[0037] 2. A dynamic feedback mechanism of spatial features is added to the fine classification module. This mechanism adjusts the feedback items by analyzing the difference between the current output and the true DOA spatial features, guides the estimation of the next layer, and gradually approaches the true value, so that each layer can be adaptively corrected according to the current estimation results, thereby improving the ability to approximate the true angle.

[0038] 3. The adaptive step size adjustment mechanism in the present invention can dynamically adjust the update amplitude of the parameters according to the estimated error, accelerate the convergence when the error is large, and improve the accuracy when the error is small, thereby improving the convergence speed. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is described in detail below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0040] A multi-layer convolutional neural network DOA estimation method based on spatial feature feedback first calculates the covariance matrix of the received signal, then decomposes the covariance matrix into two channels, real part and imaginary part, which are input into the multi-layer DOA estimation model to obtain the final arrival angle estimation value.

[0041] The multi-layer DOA estimation model includes sequentially setting There are three classification layers, the first of which is a coarse classification module, and the other classification layers each have a fine classification module. The number of classification layers containing the fine classification modules is determined by the estimation resolution. Each classification layer outputs an arrival angle estimate. The arrival angle estimate output by the classification layer is recorded as , the arrival angle estimate output by the last classification layer That is the output value of the multi-layer DOA estimation model.

[0042] The input of the coarse classification module contains two channels, namely the real part of the covariance matrix and the imaginary part of the covariance matrix, and the output is the arrival angle estimation value output by the classification layer .

[0043] For classification layers, The input of the fine classification module contains 4 channels, namely: the real part of the covariance matrix, the imaginary part of the covariance matrix, the arrival angle estimate output by the previous classification layer and the feedback term calculated based on the arrival angle estimate of the previous classification layer through spatial features , the output of the fine classification module is the accuracy increment The classification layer outputs the estimated arrival angle as .

[0044] In this embodiment, the resolution of the arrival angle estimation value output by the coarse classification module is 1°. For example, the output is 30°, which means that the range of the actual arrival angle predicted by the coarse classification module is greater than or equal to 30° and less than 31°. With each additional layer of fine classification module, the resolution of the obtained precision increment will be increased by 10 times relative to the output of the previous classification layer. For example, the resolution of the precision increment output by the first fine classification module is 0.1°, and the range of the precision increment is greater than or equal to 0° and less than or equal to 0.9°. If the precision increment finally output by the fine classification module of this layer is 0.2°, then the arrival angle estimation value output by this classification layer is 30°+0.2°=30.2°. Similarly, the resolution of the precision increment output by the second fine classification module is 0.01°, and the precision increment is greater than or equal to 0° and less than or equal to 0.09°. If the precision increment finally output by the fine classification module of this layer is 0.03°, then the arrival angle estimation value output by this classification layer is 30.2°+0.03°=30.23°. Therefore, As the number of classification layers containing the refined classification module, Indicates the rounding operation. For the resolution to be achieved, we have:

[0045] .

[0046] As the number of fine classification modules increases, the estimation interval will gradually narrow and eventually reach the required resolution.

[0047] Furthermore, let the received signal be ,in For the signal vectors of time snapshots, each signal vector Contains the signals received by all array elements at this time snapshot, and the calculation method of the signal covariance matrix is:

[0048] ;

[0049] in, for The conjugate transpose of .

[0050] In this method, in order to adapt to the input of CNN, the covariance matrix is ​​decomposed into real and imaginary parts, which are used as multi-channel inputs respectively.

[0051] Furthermore, the coarse classification module and the fine classification module each include a feature extraction submodule and a classification submodule.

[0052] In the coarse classification module, the real part of the covariance matrix and the imaginary part of the covariance matrix are input into the feature extraction submodule to obtain a feature map, and then the feature map is input into the classification submodule to obtain a probability distribution estimate. The category corresponding to the maximum probability in the probability distribution estimate is the arrival angle estimate output by the coarse classification module. .

[0053] Specifically, the feature extraction submodule in the coarse classification module contains feature extraction blocks arranged in sequence, and each feature extraction block is arranged in sequence with a convolution layer, an activation function, and a pooling layer. The convolution layer is used to perform convolution operations on the input feature map to extract local features, thereby generating a new feature map. The pooling layer is after the convolution layer and is used to reduce the size of the feature map, reduce the amount of calculation, and extract the most important features. This alternating convolution-pooling structure helps to control the computational complexity while extracting the high-order spatial features of the signal layer by layer.

[0054] Assume The feature map input in the feature extraction block is obtained after the convolution layer and activation function. ,but:

[0055] ;

[0056] in, represents the convolution operation, and are the convolution kernel weights and bias terms of the convolution layer in the feature extraction block respectively; Extract the nonlinear activation function in the block for this feature; For the The feature map output by the feature extraction block is After The feature map output after the pooling layer in the feature extraction block.

[0057] By stacking multiple layers of convolution, CNN can gradually extract the high-order spatial features of the signal and reveal the DOA pattern of the signal.

[0058] In the classification submodule of the coarse classification module, the feature map from the feature extraction submodule is first flattened into a one-dimensional vector , the one-dimensional vector Include elements, is the total number of categories of the classification submodules, and the categories are obtained by dividing the range of angles output by the current classification layer according to the resolution to be output by the classification layer. Assuming that the estimated angle range in this embodiment is from -90° to 90°, and the resolution of the coarse classification module is 1° (i.e., the classification interval is 1°), then , that is, each 1° is a classification. Then the one-dimensional vector Input to the fully connected layer for feature fusion to obtain the fusion result ,in and are the weight matrix and bias vector of the fully connected layer, respectively. Include elements; then merge the results Input into the softmax classifier to obtain the probability distribution estimate. The Elements , the probability value obtained by the softmax classifier is , That is the The category corresponding to the maximum probability in the probability distribution estimate is the arrival angle estimate output by the coarse classification module. .

[0059] Furthermore, in the fine classification module, the real part of the covariance matrix, the imaginary part of the covariance matrix, the arrival angle estimate output by the previous classification layer, and the feedback item obtained based on the arrival angle estimate of the previous classification layer are input into the feature extraction submodule to obtain a feature map, and then the feature map is input into the classification submodule to obtain an accuracy increment.

[0060] Specifically, the feature extraction submodule in the fine classification module has the same structure as the feature extraction submodule in the coarse classification module, and only the number of channels processed is different, which will not be described here.

[0061] The processing of the classification submodule in the fine classification module is the same as that of the coarse classification module, with the only difference being that the category corresponding to the maximum probability in the probability distribution estimate is the precision increment output by the fine classification module. , rather than an arrival angle estimate.

[0062] Furthermore, for the classification layers, , and obtain the estimated angle of arrival of this layer After that, the corresponding feedback term is calculated by the following method :

[0063] First, calculate the steering vector :

[0064] ;

[0065] in, , is the array element spacing of the received signal, is the signal wavelength, is the number of array elements, is the sequence index of the array antenna.

[0066] Then, calculate the arrival angle estimate for this layer With the covariance matrix Deviation in space :

[0067] ;

[0068] in, Represents the conjugate transpose of a matrix.

[0069] Finally, the deviation Perform normalization processing to obtain the feedback item :

[0070] ;

[0071] in, is the Frobenius norm.

[0072] The purpose of normalization is to avoid the extreme influence of excessive eigenvalues.

[0073] The dynamic feedback mechanism of spatial features based on feedback terms is the core of the multi-layer DOA estimation model. It dynamically adjusts the feedback terms by analyzing the deviation between the model output and the spatial features. This mechanism is not only used in the training process, but also plays a key role in the inference process. The core of this mechanism lies in the feedback analysis of spatial features, that is, refining the DOA estimate by the feature difference of the steering vector. The dynamic feedback mechanism is the basis of the precision increment of the fine layer, which enables each layer of the network to make adaptive corrections based on the current estimation results. The role of the feedback term is not to simply repeat the covariance matrix or the DOA estimate of the previous layer, but to provide an error perception of the current DOA estimate in the spatial features. In this way, the input of each layer is integrated with the DOA estimate and feedback terms of the previous layer, so that the model can learn and utilize this dynamic information layer by layer, thereby improving the approximation ability to the true DOA.

[0074] Furthermore, the multi-layer DOA estimation model updates model parameters by training from the coarse classification module to the fine classification module layer by layer, first training the coarse classification module, and then training each fine classification module in turn. When the previous classification model is trained, the training of the next classification module begins.

[0075] For each classification module, training is completed by updating the network parameters of the classification module in multiple iterations. In each iteration, the loss value (such as the cross entropy loss function) is first calculated based on the estimated arrival angle value and the true angle value output by the classification module, and then the gradient of the loss value with respect to the network parameters is calculated using the back propagation algorithm. The gradient is then multiplied by the step size using the gradient descent algorithm to obtain the amplitude of the parameter update, and the network parameters of the current classification module are updated accordingly. When the loss value converges or reaches the preset number of iterations, the training stops.

[0076] Furthermore, in the training process of each classification module, the convergence speed can be accelerated by an adaptive step size adjustment mechanism: before the current classification module starts iterative updating, the step size is set to the initial value; The estimated angle of arrival after iterations is , the true angle value is , then The deviation after iterations is , No. The step size used in the next iteration will be updated to ,in is the step size adjustment coefficient, which is used to control the amplitude of the step size adjustment. is a sign function.

[0077] Through the adaptive step size adjustment mechanism, when the error is large, the step size is increased to improve the convergence speed. When the error is small, the step size is reduced to increase the accuracy. In this way, the model can quickly approach the true DOA value in the initial stage, and gradually improve the accuracy of DOA estimation in the subsequent layers of fine classification.

[0078] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. A multi-layer convolutional neural network DOA estimation method based on spatial feature feedback, characterized in that: First, the covariance matrix of the received signal is calculated, and then the covariance matrix is ​​decomposed into two channels, the real part and the imaginary part, which are input into the multi-layer DOA estimation model to obtain the final arrival angle estimation value; The multi-layer DOA estimation model includes sequentially setting classification layers, wherein the first classification layer is a coarse classification module, and each of the other classification layers has a fine classification module; the number of classification layers containing the fine classification modules is determined by the estimation resolution; each classification layer outputs an arrival angle estimation value, and the first The arrival angle estimate output by the classification layer is recorded as , the arrival angle estimate output by the last classification layer That is the output value of the multi-layer DOA estimation model; The input of the coarse classification module contains two channels, namely the real part of the covariance matrix and the imaginary part of the covariance matrix, and the output is the arrival angle estimation value output by the classification layer ; For classification layers, The input of the fine classification module contains 4 channels, namely: the real part of the covariance matrix, the imaginary part of the covariance matrix, the arrival angle estimate output by the previous classification layer and the feedback term calculated based on the arrival angle estimate of the previous classification layer through spatial features , the output of the fine classification module is the accuracy increment The classification layer outputs the estimated arrival angle as ; For classification layers, , and obtain the estimated angle of arrival of this layer After that, the corresponding feedback term is calculated by the following method : First, calculate the steering vector : ； in, , is the array element spacing of the received signal, is the signal wavelength, is the number of array elements, is the sequence index of the array antenna; Then, calculate the arrival angle estimate for this layer With the covariance matrix Deviation in space : ； in, Represents the conjugate transpose of a matrix; Finally, the deviation Perform normalization processing to obtain the feedback item : ； in, is the Frobenius norm.

2. The multi-layer convolutional neural network DOA estimation method based on spatial feature feedback as claimed in claim 1, characterized in that: Each additional layer of fine classification module will increase the resolution of the precision increment by 10 times compared with the output of the previous classification layer. As the number of classification layers containing the refined classification module, Indicates the rounding operation. is the resolution to be achieved.

3. The multi-layer convolutional neural network DOA estimation method based on spatial feature feedback as claimed in claim 1, characterized in that: Assume the received signal is ,in For the signal vectors of time snapshots, each signal vector Contains the signals received by all array elements at this time snapshot, and the calculation method of the signal covariance matrix is: ； in, for The conjugate transpose of .

4. The multi-layer convolutional neural network DOA estimation method based on spatial feature feedback as claimed in claim 1, characterized in that: The coarse classification module and the fine classification module each include a feature extraction submodule and a classification submodule.

5. The multi-layer convolutional neural network DOA estimation method based on spatial feature feedback as claimed in claim 4, characterized in that: In the coarse classification module, the real part of the covariance matrix and the imaginary part of the covariance matrix are input into the feature extraction submodule to obtain a feature map, and then the feature map is input into the classification submodule to obtain a probability distribution estimate. The category corresponding to the maximum probability in the probability distribution estimate is the arrival angle estimate output by the coarse classification module. ; In the fine classification module, the real part of the covariance matrix, the imaginary part of the covariance matrix, the arrival angle estimate output by the previous classification layer, and the feedback item obtained based on the arrival angle estimate of the previous classification layer are input into the feature extraction submodule to obtain a feature map, and then the feature map is input into the classification submodule to obtain an accuracy increment.

6. The multi-layer convolutional neural network DOA estimation method based on spatial feature feedback according to claim 4 or 5, characterized in that: The feature extraction submodule includes feature extraction blocks arranged in sequence, and each feature extraction block is arranged in sequence with a convolution layer, an activation function and a pooling layer; Set up The feature map input in the feature extraction block is obtained after the convolution layer and activation function. ,but: ； in, represents the convolution operation, and are the convolution kernel weights and bias terms of the convolution layer in the feature extraction block respectively; Extract the nonlinear activation function in the block for this feature; For the The feature map output by the feature extraction block is After The feature map output after the pooling layer in the feature extraction block.

7. The multi-layer convolutional neural network DOA estimation method based on spatial feature feedback as claimed in claim 5, characterized in that: In the classification submodule, the feature map from the feature extraction submodule is first flattened into a one-dimensional vector , the one-dimensional vector Include elements, is the total number of categories of the classification submodule, and the categories are obtained by dividing the range of angles output by the classification layer according to the resolution to be output by the current classification layer; then the one-dimensional vector Input to the fully connected layer for feature fusion to obtain the fusion result ,in and are the weight matrix and bias vector of the fully connected layer, respectively. Include elements; then merge the results Input into the softmax classifier to obtain the probability distribution estimate. The Elements , the probability value obtained by the softmax classifier is , That is the The probability value of each category; The category corresponding to the maximum probability in the probability distribution estimate is the output result of the classification submodule.

8. The multi-layer convolutional neural network DOA estimation method based on spatial feature feedback as claimed in claim 1, characterized in that: The multi-layer DOA estimation model updates model parameters by training from the coarse classification module to the fine classification module layer by layer, first training the coarse classification module, and then training each fine classification module in turn; When the previous classification model is trained, the training of the next classification module will start; For each classification module, training is completed by iteratively updating the network parameters of the classification module. In each iteration, the loss value is first calculated based on the arrival angle estimate and the true angle value output by the classification module, and then the back propagation algorithm is used to calculate the gradient of the loss value with respect to the network parameters. Then, the gradient descent algorithm is used to multiply the gradient by the step size to obtain the amplitude of the parameter update, and the network parameters of the current classification module are updated accordingly. When the loss value converges or reaches the preset number of iterations, the training is stopped.

9. The multi-layer convolutional neural network DOA estimation method based on spatial feature feedback as claimed in claim 8, characterized in that: For each classification module, before the current classification module starts iterative updating, the step size is set to the initial value; The estimated arrival angle after iterations is , the true angle value is , then The deviation after iterations is , No. The step size used in the next iteration will be updated to ,in is the step size adjustment coefficient, which is used to control the amplitude of the step size adjustment. is a sign function.

Images