A conventional beamforming learning method based on deep convolutional neural regression network

Abstract

The application discloses a conventional beam forming learning method based on a deep convolutional neural regression network, comprising the following steps: step 1, obtaining a data set, comprising a plurality of data pairs, any data pair comprising array element domain frequency domain data and corresponding spatial energy spectrum; step 2, performing feature pre-extraction on the array element domain frequency domain data to obtain array element domain frequency domain features and construct a sample set; step 3, constructing a deep convolutional neural network model with the array element domain frequency domain features as input and the corresponding spatial energy spectrum as target output; step 4, training the deep convolutional neural network model based on the sample set; and step 5, after performing feature pre-extraction on unknown array element domain frequency domain data, inputting the unknown array element domain frequency domain data into the trained deep convolutional neural network model to output a predicted spatial energy spectrum. The method and model disclosed by the application are a conventional beam forming framework independent of current array signal processing methods, and have the characteristics of autonomous learning evolution, efficient parallel computing and the like.

Description

Technical Field

[0001] This invention belongs to the fields of underwater acoustic signal detection technology and artificial intelligence technology, specifically relating to a conventional beamforming learning method based on a deep convolutional neural regression network. Background Technology

[0002] Beamforming is a major component of sonar array signal processing. Conventional beamformers (CBFs) have advantages such as good tolerance and low computational cost, and are currently the most widely used beamforming technology in sonar signal detection.

[0003] In recent years, artificial intelligence and neural network technologies, represented by deep learning, have developed rapidly. By superimposing and applying several hidden computation layers, deep learning can achieve hierarchical feature extraction and fitting of complex nonlinear relationships. Therefore, there is an urgent need for a method for sonar array signal processing using deep learning. Summary of the Invention

[0004] The purpose of this invention is to provide a conventional beamforming learning method based on a deep convolutional neural regression network. By learning the mapping relationship between array element domain frequency domain data and conventional beamforming results, the network model is endowed with conventional beamforming capabilities.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A conventional beamforming learning method based on a deep convolutional neural regression network includes the following steps:

[0007] Step 1: Obtain the dataset, which includes several data pairs. Each data pair includes array element domain frequency domain data and the corresponding spatial energy spectrum.

[0008] Step 2: Perform feature pre-extraction on the array element domain frequency domain data to obtain array element domain frequency domain features and construct a sample set;

[0009] Step 3: Construct a deep convolutional neural network model with the frequency domain features of the array element domain as input and the corresponding spatial energy spectrum as the target output;

[0010] Step 4: Train the deep convolutional neural network model based on the sample set;

[0011] Step 5: After pre-extracting features from the frequency domain data of the unknown array element domain, input the data into the trained deep convolutional neural network model to output the predicted spatial energy spectrum.

[0012] Preferably, step 1 includes the following steps:

[0013] Step 1.1: Construct a passive broadband early warning and detection model. Its input parameters include sonar array parameters, target parameters, and background noise parameters. The target parameters include the number of targets, the target signal-to-noise ratio, and the target azimuth. The output is two-dimensional array element domain frequency domain data and one-dimensional spatial energy spectrum.

[0014] Step 1.2: Set the sonar array parameters according to the actual parameter requirements of the sonar array, and set the target quantity, signal-to-noise ratio range and background noise parameters according to the usage requirements;

[0015] Step 1.3: Randomly set the number of targets and the corresponding signal-to-noise ratio according to the set number of targets and signal-to-noise ratio range, and at the same time randomly set the corresponding azimuth of the targets. Generate the array element domain frequency domain data and the corresponding spatial energy spectrum of random target number, azimuth and signal-to-noise ratio through the passive broadband warning model.

[0016] Step 1.4, repeat step 1.3, to obtain the dataset.

[0017] Preferably, step 2 includes the following steps:

[0018] Step 2.1, pairwise element-domain frequency domain data x f Finding the real part, imaginary part, and modulus yields three two-dimensional matrices x. f，real x f，imag x f，mod ;

[0019] Step 2.2: Expand each two-dimensional matrix by one dimension in the same dimension to obtain a three-dimensional matrix x. f，real_M x f，imag_M x f，mod_M And by performing splicing operations in the extended dimensions, three-dimensional array element domain frequency domain data is obtained;

[0020] Step 2.3: Normalize the three-dimensional array element domain frequency domain data to obtain the three-dimensional array element domain frequency domain features x. Sample ;

[0021] Step 2.4: Associate the frequency domain features of each array element with the corresponding spatial energy spectrum to form several samples and obtain a sample set.

[0022] Preferably, the deep convolutional neural network model includes several feature extraction and compression modules, a global average pooling layer, and a fully connected layer connected in series. Each feature extraction and compression module includes a size compression unit and several feature extraction units connected in sequence.

[0023] Preferably, the feature extraction unit includes two parallel branches. Branch 1 includes a convolutional layer, a LayerNorm layer, a convolutional layer, a GELU activation function, a convolutional layer, and a coefficient weighting layer connected in sequence. Branch 2 is a directly connected layer. The convolutional features of branch 1 and branch 2 are connected through a convolutional feature summation layer.

[0024] Preferably, the size compression unit includes two parallel branches. Branch 1 includes a convolutional layer, a LayerNorm layer, a convolutional layer, a GELU activation function, a convolutional layer, and a coefficient weighting layer connected in sequence. Branch 2 includes a convolutional layer. The convolutional features of branch 1 and branch 2 are connected through a convolutional feature summation layer.

[0025] Preferably, step 4 includes the following steps:

[0026] Step 4.1: Divide the sample set into a training sample set and a validation sample set according to a preset ratio;

[0027] Step 4.2: Set the training parameters, set the optimizer to Adam, and set the loss function to the minimum mean square error function;

[0028] Step 4.3: Based on the training sample set and the validation sample set, the deep convolutional neural network model is trained using cross-validation.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] This invention learns the frequency domain data of array elements in the sample through a deep convolutional neural network model, thereby obtaining a target output consistent with the corresponding spatial energy spectrum. This deep learning model is a conventional beamforming framework independent of current array signal processing methods, and has the characteristics of autonomous learning evolution and efficient parallel computing. This conventional beamforming method based on deep convolutional neural regression network has been applied to simulation data processing and has achieved good results. Attached Figure Description

[0031] Figure 1 This is a flowchart of the process framework of the present invention.

[0032] Figure 2 This is a structural diagram of the feature extraction unit and the size compression unit.

[0033] Figure 3 This is a structural diagram of a deep convolutional neural regression network model.

[0034] Figure 4 The simulation results are for the conventional beamforming method based on deep convolutional neural regression networks.

[0035] Figure 5 A modeling scheme for deep convolutional neural regression networks. Detailed Implementation

[0036] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0037] Reference Figure 1 A conventional beamforming learning method based on deep convolutional neural regression networks includes four stages: construction of training sample set, construction of deep convolutional neural regression network model, training of deep convolutional neural regression network model, and conventional beamforming application, which is implemented by the following five steps.

[0038] Step 1: Obtain the dataset, which includes several data pairs. Each data pair includes array element domain frequency domain data and the corresponding spatial energy spectrum.

[0039] Step 2: Perform feature pre-extraction on the array element domain frequency domain data to obtain array element domain frequency domain features and construct a sample set;

[0040] Step 3: Construct a deep convolutional neural network model that takes the frequency domain features of the array element domain as input and the corresponding spatial energy spectrum as the target output.

[0041] Step 4: Train the deep convolutional neural network model based on the sample set.

[0042] Step 5: After pre-extracting features from the frequency domain data of the unknown array element domain, input the data into the trained deep convolutional neural network model to output the predicted spatial energy spectrum.

[0043] In step 1 of this invention, the dataset consists of a large number of data pairs, with a one-to-one correspondence between the array element domain frequency domain data and the spatial energy spectrum. Obtaining this dataset includes the following steps:

[0044] Step 1.1: Construct a passive broadband early warning and detection model. Its input parameters mainly include sonar array parameters, target parameters, and background noise parameters, as well as the number of FFT points. The target parameters include the number of targets, the target signal-to-noise ratio, and the target azimuth. The output is two-dimensional array element domain frequency domain data and one-dimensional spatial energy spectrum.

[0045] This passive broadband early warning and detection model is a commonly used array data processing model in this field. It can generate two-dimensional array element domain frequency domain data and corresponding spatial energy spectra. The model's construction can be summarized as follows: Step 1.1.1 Constructing a passive array data simulation model, whose inputs include sonar array parameters, target parameters, and background noise parameters, and whose output is two-dimensional array element domain frequency domain data; Step 1.1.2 Generating the array manifold matrix, whose inputs mainly include the observation space vector, array position vector, and processing frequency; Step 1.1.3 Processing the array element domain frequency domain data based on a conventional beamforming algorithm model, outputting a one-dimensional spatial energy spectrum. The above construction method is common knowledge in this field and will not be elaborated upon here.

[0046] Step 1.2: Set the sonar array parameters according to the actual parameter requirements of the sonar array, and set the target quantity, signal-to-noise ratio range, and background noise parameters according to the usage requirements.

[0047] Step 1.3: Randomly set the number of targets and the corresponding signal-to-noise ratio according to the set number of targets and signal-to-noise ratio range, and at the same time randomly set the corresponding azimuth of the targets. Generate the array element domain frequency domain data and the corresponding spatial energy spectrum of random target number, azimuth and signal-to-noise ratio through the passive broadband warning model.

[0048] Step 1.4, repeat step 1.3, to obtain a large number of data pairs, thus obtaining the dataset.

[0049] In step 2 of this invention, the sample set includes a large number of samples, the total number of which is determined by the total number of data pairs in the dataset. Each sample in the sample set includes array element domain frequency domain features and a corresponding spatial energy spectrum. The constructed deep convolutional neural network model is trained using these samples to optimize the model parameters, thereby improving the model's fitting performance. Step 2 specifically includes the following steps:

[0050] Step 2.1, pairwise element-domain frequency domain data x f Finding the real part, imaginary part, and modulus yields three two-dimensional matrices x. f，real x f，imag x f，mod ;

[0051] Step 2.2: Expand each two-dimensional matrix by one dimension in the same dimension to obtain a three-dimensional matrix. Make the following three-dimensional matrices such that the value of dimension 1 is 1, the value of dimension 2 is the same as the value of dimension 1 in the two-dimensional matrix, and the value of dimension 3 is the same as the value of dimension 2 in the two-dimensional matrix. Perform a stitching operation on dimension 1 to obtain two-dimensional array element frequency domain data;

[0052] Step 2.3: Normalize the three-dimensional array element domain frequency domain data to obtain the three-dimensional array element domain frequency domain features x. Sample ;

[0053] Step 2.4: Associate the frequency domain features of each array element with the corresponding spatial energy spectrum to form several samples and obtain a sample set.

[0054] In step 2.2 of this invention, a two-dimensional matrix x is used. f，real For example, by extending it by one dimension, we obtain a three-dimensional matrix x. f，real_M Its three dimensions are 1, x f，real Dimension 1, x f，real Dimension 2, x f，real_M The value in dimension 2 and x f，real The values ​​are consistent in dimension 1.

[0055] Reference Figure 2 In step 3 of this invention, the deep convolutional neural network model uses two basic convolutional module frameworks alternately, combined with global average pooling layers and fully connected layers, to obtain useful information from the frequency domain features of the array element domain and output conventional beamforming results. The specific process of step 3 is as follows.

[0056] Step 3.1: Construct a feature extraction unit, with its output channel number set to an adjustable parameter x; add two parallel branches: branch 1 includes a convolutional layer (7×7, x, 1), a LayerNorm layer, a convolutional layer (1×1, 4x, 1), a GELU activation function, a convolutional layer (1×1, x, 1), and a coefficient weighting layer; branch 2 is a direct connection layer; add the convolutional features output from the two branches to obtain the final output result.

[0057] Step 3.2: Construct a size compression unit with its output channel number set to an adjustable parameter x; add two parallel branches. Branch 1 consists of a convolutional layer (3×3, x, 1), a LayerNorm layer, a convolutional layer (1×1, 4x, 1), a GELU activation function, a convolutional layer (1×1, x, 2), and a coefficient weighting layer; Branch 2 consists of a convolutional layer (1×1, x, 2); add the convolutional features output from the two branches to obtain the final output result.

[0058] In this invention, the three parameters of the convolutional layer represent the kernel size, the number of convolutional channels, and the stride, respectively. Taking (7×7, x, 1) as an example, 7×7 is the kernel size, x is the number of convolutional channels, and 1 is the stride. For the coefficient weighting layer, it implements the dot product of the convolutional features in the channel dimension into a weighted vector, and the length of this weighted vector is the same as the number of convolutional feature channels.

[0059] Step 3.3: Construct the entire deep convolutional neural regression network model, which includes several feature extraction and compression modules, a global average pooling layer, and a fully connected layer connected in series. Each feature extraction and compression module includes a size compression unit and several feature extraction units connected in sequence.

[0060] Figure 5 This paper presents a deep convolutional neural regression network modeling scheme. For conventional beamforming under different parameters, a model with an appropriate scale needs to be constructed for the corresponding array element domain frequency domain data. The input size of the array element domain frequency domain data sample is set to a×b×3, where a>16, b>16, and int() represents rounding down. R out The model represents the length of the spatial energy spectrum and includes multiple convolutional layers, pooling layers, size compression units, and feature extraction units.

[0061] As one specific embodiment of step 3.3 of the present invention, the process mainly includes:

[0062] (3.3.1) Add one size compression unit and set the number of output channels to 64.

[0063] (3.3.2) Add 3 cascaded feature extraction units and set the number of output channels to 64.

[0064] (3.3.3) Add one size compression unit and set the number of output channels to 128.

[0065] (3.3.4) Add 3 cascaded feature extraction units and set the number of output channels to 128.

[0066] (3.3.5) Add one size compression unit and set the number of output channels to 256.

[0067] (3.3.6) Add 16 cascaded feature extraction units and set the number of output channels to 256.

[0068] (3.3.7) Add one size compression unit and set the number of output channels to 512.

[0069] (3.3.8) Add 3 cascaded feature extraction units and set the number of output channels to 512.

[0070] (2.2.9) Add a global average pooling layer.

[0071] (2.2.10) Add a fully connected layer (512, N) spatial_spec ), N spatial_spec This indicates the length of the output spatial energy spectrum.

[0072] In step 4 of this invention, the spatial energy spectrum corresponding to the frequency domain features of the array element domain input to the network model is used as the target output. The goal is to reduce the loss value between the network model's output and the target output. The network model is then trained so that the output of the trained network model can infinitely approach the corresponding spatial energy spectrum. The specific process of step 4 is as follows.

[0073] Step 4.1: Divide the sample set into an independent training sample set and a validation sample set according to a preset ratio, specifically, the preset ratio is 100:1.

[0074] Step 4.2: Set the training parameters. Set the optimizer to Adam, the loss function to the minimum mean square error function, and the batch size for each training session to 64.

[0075] Step 4.3: Based on the training sample set and the validation sample set, the deep convolutional neural network model is trained using cross-validation. The specific process is as follows.

[0076] (4.3.1) For a single model training, 64 samples are selected from the training sample set based on the random replacement sampling method, and the beamforming results are output through the deep convolutional neural network model.

[0077] (4.3.2) Minimize the error between the deep convolutional neural network model result and the corresponding spatial energy spectrum based on the Adam optimizer.

[0078] (4.3.3) Repeat steps (4.3.1) and (4.3.2) to conduct more rounds of model training. After every 50 rounds of model training, perform performance testing on the model based on the validation sample set, calculate the error between the output result of the validation sample set and the label, and stop model training when the error decrease rate of every 100 performance tests does not exceed 1%.

[0079] Figure 4 The simulation verification results of the method of the present invention are shown. The model has been trained and improved using a large-scale simulation sample set, combined with... Figure 4 It can be seen that the output of the deep convolutional neural regression network model is very close to the results of conventional beamforming, which verifies the effectiveness of using this network model for conventional beamforming learning.

Claims

1. A conventional beamforming learning method based on deep convolutional neural regression network, characterized in that, The method comprises the following steps: Step 1, obtaining a data set comprising a plurality of data pairs, any data pair comprising array element domain frequency domain data and corresponding spatial energy spectrum; Step 2, performing feature pre-extraction on the array element domain frequency domain data to obtain array element domain frequency domain features and constructing a sample set; Step 3, constructing a deep convolutional neural network model taking the array element domain frequency domain features as input and the corresponding spatial energy spectrum as target output; The deep convolutional neural network model comprises a plurality of feature extraction compression modules, a global average pooling layer and a fully connected layer connected in series, any feature extraction compression module comprising a size compression unit and a plurality of feature extraction units connected in sequence; The feature extraction unit comprises two parallel branches, branch 1 comprising a convolutional layer, a LayerNorm layer, a convolutional layer, a GELU activation function, a convolutional layer and a coefficient weighting layer connected in sequence, and branch 2 being a direct connection layer, the convolutional features of branch 1 and branch 2 being connected through a convolutional feature addition layer; The size compression unit comprises two parallel branches, branch 1 comprising a convolutional layer, a LayerNorm layer, a convolutional layer, a GELU activation function, a convolutional layer and a coefficient weighting layer connected in sequence, and branch 2 comprising a convolutional layer, the convolutional features of branch 1 and branch 2 being connected through a convolutional feature addition layer; Step 4, training the deep convolutional neural network model based on the sample set; Step 5, inputting the unknown array element domain frequency domain data after feature pre-extraction into the trained deep convolutional neural network model to output a predicted spatial energy spectrum. 2.The conventional beamforming learning method based on deep convolutional neural regression network according to claim 1, wherein, The step 1 comprises the following steps: Step 1.1, constructing a passive wideband warning detection model, the input parameters of which comprise sonar array parameters, target parameters and background noise parameters, the target parameters comprising target quantity, target signal-to-noise ratio and target direction, the output being two-dimensional array element domain frequency domain data and one-dimensional spatial energy spectrum; Step 1.2, setting the sonar array parameters according to the actual parameter requirements of the sonar array, setting the target quantity and signal-to-noise ratio range and the background noise parameters according to the use requirements; Step 1.3, randomly setting the target quantity and the corresponding signal-to-noise ratio according to the set target quantity and signal-to-noise ratio range, and randomly setting the corresponding target direction, generating array element domain frequency domain data and corresponding spatial energy spectrum of random target quantity, direction and signal-to-noise ratio through the passive wideband warning model; Step 1.4, repeating step 1.3 to obtain the data set. 3.The conventional beamforming learning method based on deep convolutional neural regression network according to claim 1, wherein, The step 2 comprises the following steps: Step 2.1, take the real part, the imaginary part and the modulus of the array domain frequency domain data xf, respectively to obtain three two-dimensional matrices x f，real , x f，imag , x f，mod ; Step 2.2, each two-dimensional matrix is respectively extended by 1 dimension in the same dimension to obtain a three-dimensional matrix x f，real_M , x f，imag_M , x f，mod_M , and a splicing operation is performed on the extended dimension to obtain three-dimensional array element domain frequency domain data; Step 2.3, normalizing the three-dimensional array element domain frequency domain data to obtain three-dimensional array element domain frequency domain features x Sample ; Step 2.4, associating each array element domain frequency domain feature with the corresponding spatial energy spectrum to form a plurality of samples and obtain a sample set. 4.The conventional beamforming learning method based on deep convolutional neural regression network according to claim 1, wherein, The step 4 comprises the following steps: Step 4.1, dividing the sample set into a training sample set and a verification sample set according to a predetermined proportion; Step 4.2, setting the training parameters, setting the optimizer as Adam and setting the loss function as the least mean square error function; Step 4.3, training the deep convolutional neural network model through cross-validation based on the training sample set and the verification sample set.

Images