A passive high-resolution processing method based on a multi-scale deep convolutional neural regression network

Abstract

The application discloses a passive high-resolution processing method based on a multi-scale deep convolution neural regression network, which comprises the following steps: step 1, constructing a training sample set, wherein any training sample in the training sample set comprises three-dimensional array element domain frequency domain data x sample and a corresponding real high-resolution spatial energy spectrum label y lable ; step 2, constructing a multi-scale deep convolution neural regression network model based on the training sample; step 3, constructing a weighted enhancement loss function about the multi-scale deep convolution neural regression network model; step 4, training the multi-scale deep convolution neural regression network model based on the training sample set and the weighted enhancement loss function; and step 5, taking new three-dimensional array element domain frequency domain data as input, and obtaining a predicted high-resolution spatial energy spectrum through the trained multi-scale deep convolution neural regression network model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of underwater acoustic signal detection and artificial intelligence, and particularly relates to a passive high-resolution processing method based on a multi-scale deep convolutional neural regression network. BACKGROUND

[0002] Signal detection is one of the important functions of sonar signal processing, and improving spatial resolution is an important goal of signal detection technology development. Although increasing the array aperture can improve the spatial resolution performance, in most cases, it is not realistic to increase the array aperture, so many high-resolution spatial spectrum estimation algorithms have been developed and emerged. With the application of various modern signal processing theories and technologies, the passive detection resolution has gradually improved, but the traditional detection system has not broken through and does not have the ability of autonomous learning and evolution. The performance in complex environments such as multi-target and strong interference still has a lot of room for improvement.

[0003] In recent years, deep learning technology has developed rapidly and has been widely used in computer vision, speech signal processing and other fields. The present application applies deep learning to passive high-resolution processing. On the basis of constructing standardized high-resolution reinforcement learning samples, a multi-scale deep convolutional neural regression network modeling, training and application method is proposed. The complex nonlinear relationship fitting capability of the deep network is used to realize array domain and frequency domain data processing, which helps to improve the spatial resolution of passive target detection. SUMMARY

[0004] The present application aims to provide a passive high-resolution processing method based on a multi-scale deep convolutional neural regression network to solve the problem of low detection performance of the existing passive detection spatial resolution method in complex environments such as multi-target and strong interference.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] A passive high-resolution processing method based on a multi-scale deep convolutional neural regression network, the method comprising the following steps:

[0007] Step 1, constructing a training sample set, any training sample in the training sample set comprising three-dimensional array domain and frequency domain data x sample and the corresponding real high-resolution spatial energy spectrum label y lable ;

[0008] Step 2, constructing a multi-scale deep convolutional neural regression network model based on the training sample;

[0009] Step 3, constructing a weighted reinforcement loss function for the multi-scale deep convolutional neural regression network model;

[0010] Step 4, training the multi-scale deep convolutional neural regression network model based on the training sample set and the weighted enhanced loss function;

[0011] Step 5, taking the new three-dimensional array element domain frequency domain data as input, obtaining the predicted high-resolution spatial energy spectrum through the trained multi-scale deep convolutional neural regression network model.

[0012] Preferably, the construction of any of the training samples comprises the following steps:

[0013] Step 1.1, setting the range of input parameters of the passive wideband warning detection model, the input parameters including target parameters, the target parameters including target quantity, target signal-to-noise ratio and target direction;

[0014] Step 1.2, setting the parameter values of the input parameters, and generating two-dimensional array element domain frequency domain data x f and corresponding one-dimensional spatial energy spectrum y through the passive wideband warning detection model;

[0015] Step 1.3, performing three-dimensional extension splicing processing on the two-dimensional array element domain frequency domain data x f to obtain three-dimensional array element domain frequency domain data x sample ;

[0016] Step 1.4, performing high-resolution enhancement processing on the one-dimensional spatial energy spectrum y to obtain the true high-resolution spatial energy spectrum label y lable .

[0017] Preferably, the step 1.3 comprises the following steps:

[0018] Step 1.3.1, calculating the real part, the imaginary part and the modulus of the two-dimensional array element domain frequency domain data x f to obtain three two-dimensional matrices x f，real , x f，imag , x f，mod , respectively;

[0019] Step 1.3.2, respectively extending the three two-dimensional matrices to three-dimensional matrices x f，real_M , x f，imag_M , x f，mod_M , and performing splicing operation in the extension dimension to obtain the three-dimensional array element domain frequency domain data x sample .

[0020] Preferably, the step 1.4 comprises the following steps:

[0021] Step 1.4.1, sequentially taking out a sequence y sub of a sliding window length from the one-dimensional spatial energy spectrum y according to the beam number;

[0022] Step 1.4.2, based on the sequence y subThe minimum m values are calculated to obtain the background sequence y back ;

[0023] Step 1.4.3, generate a zero-value sequence y with the same length as the length of the one-dimensional spatial energy spectrum target , corresponding to the beam number sequence N of the target direction target The zero-value sequence is assigned to the corresponding spatial energy spectrum, denoted as y target (N target )=y(N target );

[0024] Step 1.4.4, calculate the real high-resolution spatial energy spectrum y lable The label is the sum of the assigned zero-value sequence and the background sequence y lable =y target +y back .

[0025] Preferably, the training sample set includes a first training sample set and a second training sample set, and in the generation process of the second training sample set, the set target number is not less than 2, and the interval between at least two target directions is not greater than 2 beams.

[0026] Preferably, the construction of the multi-scale convolutional neural regression network model includes the following steps:

[0027] Add a convolution layer and a LayerNorm layer;

[0028] Add several basic modules 1;

[0029] Add a LayerNorm layer and a convolution layer;

[0030] Add several basic modules 2;

[0031] Add a LayerNorm layer and a convolution layer;

[0032] Add several basic modules 3;

[0033] Add a LayerNorm layer and a convolution layer;

[0034] Add several basic modules 4;

[0035] Add a global average pooling layer and a fully connected layer.

[0036] Preferably, the basic module 1, the basic module 2, the basic module 3 and the basic module 4 each include 3 parallel branches;

[0037] The branch 1 and the branch 2 each comprise a convolution layer, a LayerNorm layer of convolution channel dimension normalization, a convolution layer, a GELU activation function, a full connection layer and a coefficient weighting layer connected in sequence, wherein the coefficient weighting layer is used for point multiplication of a weighting vector to the convolution feature at the channel dimension, and the weighting vector has the same length as the number of convolution feature channels; the branch 3 is a direct connection layer; the convolution features output by the three branches are added to obtain the output of the module.

[0038] Preferably, the step 3 comprises the following steps.

[0039] Step 3.1, performing gradual zero value statistics on the real high-resolution spatial energy spectrum y lable and dividing into non-zero subsequences and zero value subsequences

[0040] Wherein, L0 and L1 represent the results of no target and target in the real high-resolution spatial energy spectrum respectively, length(y Lable ) represents the sequence length of the real high-resolution spatial energy spectrum, and i0 and i1 are indexes of the high-resolution spatial energy spectrum.

[0041] Step 3.2, performing same position division on the predicted high-resolution enhanced spatial energy spectrum R Model according to the real non-zero subsequence and the real zero value subsequence

[0042] Step 3.3, constructing a weighted enhancement loss value Wherein, N B is the batch size during training, and alpha0 and alpha1 are weighting coefficients.

[0043] Preferably, the step 4 comprises the following steps.

[0044] Step 4.1, training the multi-scale deep convolution neural regression network model through a first training sample set, wherein alpha0 and alpha1 in the weighted enhancement loss function are both 1.0.

[0045] Step 4.2, performing secondary training on the multi-scale deep convolution neural regression network model through a second training sample set, wherein alpha0 = 1.0 and alpha1 = 2.0 in the weighted enhancement loss function.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] The present application uses a deep learning algorithm adapted to the characteristics of the array element domain frequency domain data for processing, and compared with the traditional array signal processing, the nonlinear fitting effect is better, and the model learning and evolution ability is better, and the performance can be improved with long-term training, and the tolerance is better. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 Flow chart of the present application.

[0049] Figure 2 Schematic diagram for construction of high-resolution spatial energy spectrum.

[0050] Figure 3 Schematic diagram of a multi-scale convolution basic module structure used in constructing a multi-scale deep convolution neural regression network model in the present application.

[0051] Figure 4 Structure diagram of a multi-scale convolution neural network model proposed in the present application.

[0052] Figure 5 Schematic diagram of a deep convolution neural regression network modeling scheme.

[0053] Figure 6 Multi-scale convolution neural network model reinforcement learning framework.

[0054] Figure 7 Simulation verification results of a passive high-resolution processing method based on a multi-scale convolution neural regression network model. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0056] Figure 1 A passive high-resolution processing method based on a multi-scale deep convolution neural regression network is shown in the present application, referring to Figure 1 The passive high-resolution processing method includes the following five steps.

[0057] Step 1, constructing a training sample set, any training sample in the training sample set includes three-dimensional array element domain frequency domain data x sample and corresponding real high-resolution spatial energy spectrum y lable .

[0058] In step 1 of the present application, the training sample set is composed of a large number of training samples, each training sample can be represented as (x sample , y lable ), that is, the three-dimensional array element domain frequency domain data and the real high-resolution spatial energy spectrum are in a corresponding relationship.

[0059] This step 1 specifically includes the following steps:

[0060] Step 1.1, set the range of input parameters of the passive broadband warning detection model, mainly including sonar array parameters, target parameters and background noise parameters; the sonar array parameters include the number of array elements, the sampling rate, the processing frequency, the array position vector, etc., and the parameter values of the sonar array parameters are set according to the requirements of the sonar array parameters; the target parameters include the number of targets, the target signal-to-noise ratio and the target direction, and the number of targets and the signal-to-noise ratio range are set according to the use requirements, and the background noise parameters are set;

[0061] Step 1.2, set the parameter values of the target parameters according to the set range of the target parameters, and generate two-dimensional array element domain frequency domain data x f and corresponding one-dimensional spatial energy spectrum y through the passive broadband warning detection model;

[0062] Step 1.3, the two-dimensional array element domain frequency domain data x f is subjected to three-dimensional extension splicing processing to obtain three-dimensional array element domain frequency domain data x sample ;

[0063] Step 1.4, the one-dimensional spatial energy spectrum y is subjected to high-resolution enhancement processing to obtain a real high-resolution spatial energy spectrum label y lable ;

[0064] Step 1.5, repeat steps 1.2-1.4 to obtain a first training sample set;

[0065] Step 1.6, repeat steps 1.2-1.4, wherein the number of input targets is not less than 2, and the interval between at least two target directions is not greater than 2 beams, to obtain a second training sample set.

[0066] In step 1.1 of the present application, the passive broadband warning detection model is a commonly used array data processing model in the field, and simply speaking, the construction of the passive broadband warning detection model includes the following steps: constructing a passive array data simulation model, the input of which includes sonar array parameters and target parameters, and the output is two-dimensional array element domain frequency domain data x f ; generating an array manifold matrix, the input of which mainly includes an observation space vector, an array position vector and a working frequency; the two-dimensional array element domain frequency domain data x f is processed through a conventional beam forming algorithm model to obtain a corresponding one-dimensional spatial energy spectrum y; the above construction method is a conventional technical means in the field, and will not be repeated here.

[0067] In step 1.2 of the application, for single sample generation, the target quantity and the corresponding signal-to-noise ratio are randomly set according to the target quantity and the signal-to-noise ratio range, and the target direction is set according to certain rules, and then the two-dimensional array element domain frequency domain data and the corresponding one-dimensional spatial energy spectrum are generated through the passive wideband warning detection model.

[0068] In the application, the construction of the training sample generation model is completed through steps 1.1-1.4, and then steps 1.1-1.4 are repeatedly executed according to the requirements of step 1.5 to obtain a large-scale sample set as a first training sample set; after the construction of the first training sample set is completed, steps 1.1-1.4 are repeatedly executed according to the requirements of step 1.6 to obtain a second large-scale sample set as a second training sample set.

[0069] In step 1.3 of the application, specifically, the three-dimensional extension splicing processing includes the following steps:

[0070] Step 1.3.1, the two-dimensional array element domain frequency domain data x f is extended to three dimensions to obtain three two-dimensional matrices x f，real , x f，imag and x xf，mod .

[0071] Step 1.3.2, each two-dimensional matrix is extended to one dimension to obtain three-dimensional matrices x f，real_M , x f，imag_M and x f，mod_M , and splicing operation is performed on the extension dimension to obtain three-dimensional array element domain frequency domain data x sample .

[0072] Taking x f，real_M as an example, the three dimensions after the extension of one dimension are 1, x f，real_M dimension 1 and x f，real_M dimension 2, the values on the dimension 2 and the dimension 3 correspond to x f，real_M , and the calculation processes of x f，imag_M , x f，mod_M are the same as x f，real_M , then x f，real_M , x f，imag_M and x f，mod_M are spliced on the dimension 1 to form a three-dimensional matrix x sample , and the dimensions are 3, x f dimension 1 and x f dimension 2 in turn.

[0073] Figure 2The schematic diagram of high-resolution enhanced label construction is shown. The spatial energy spectrum results obtained by conventional beamforming on the simulated array element domain frequency domain data are combined with the known number and orientation of targets in advance. The spectrum estimation results are constructed only for the orientation of the target, thereby significantly improving the orientation resolution capability of the training sample labels. Specifically, in step 1.4, the high-resolution enhancement processing includes the following steps:

[0074] Step 1.4.1: Take out the sequence y of sliding window length from the one-dimensional space energy spectrum y according to the beam number. sub ;

[0075] Step 1.4.2, based on sequence y sub The median of the smallest m values ​​is calculated and used as the background of the beam to obtain the background sequence y back ;

[0076] Step 1.4.3, generate a zero-value sequence y with the same length as the one-dimensional space energy spectrum length(y) target , according to the beam number sequence N corresponding to the target direction target Assign the zero-value sequence to the corresponding spatial energy spectrum, denoted as y target (N target )=y(N target );

[0077] Step 1.4.4, calculate the true high-resolution spatial energy spectrum y lable is the sum of the zero value sequence and the background sequence after assignment y lable =y target +y back .

[0078] Here, the beam number is the sequence number of the one-dimensional space energy spectrum y. That is, if the one-dimensional space energy spectrum y has 256 points, then the beam number 1 corresponds to the first point. For example, if the sliding window length is 2L, then in step 1.4.1, the sequence y of length 2L is taken out from the one-dimensional space energy spectrum y in sequence according to the beam number. sub , that is, for beam number n beam , take out the spatial energy spectrum of L length at both ends, if one end exceeds n beam The range of possible values, that is, n beam -L<0 or n beam +L>length(y), the value of this end starts from 0 or length(y), and the remaining Ln beam A point is taken from the other end of the one-dimensional energy spectrum y.

[0079] In step 1.4.3 of the present invention, the beam number sequence N corresponding to the target direction is targetThe zero value sequence is assigned to the corresponding spatial energy spectrum, for example, if the target number set in step 1.2 is 2 and the target direction is 1 and 3, the real high-resolution spatial energy spectrum label y obtained through step 1.4 is {1, 0.2, 1.5, 0.35, 0.25}, and then y target (N target )={1, 0, 1.5, 0, 0}; that is, the amplitude of the spatial energy spectrum without the target is set to 0, and the amplitude of the target is kept unchanged.

[0080] Step 2, constructing a multi-scale deep convolutional neural regression network model based on the training sample.

[0081] In step 2 of the present application, the neural network model is constructed based on the convolutional neural network framework, specifically including two parts of constructing four basic modules and constructing the entire deep convolutional network model.

[0082] The main process of constructing four basic modules is as follows:

[0083] Step 2.1.1, constructing basic module 1. Adding three parallel branches. Branch 1 includes, in sequence, a convolutional layer (7x7, 64, 1), a LayerNorm layer for convolutional channel dimension normalization, a convolutional layer (1x1, 256, 1), a GELU activation function, a fully connected layer (1x1, 64, 1), and a coefficient weighting layer; Branch 2 includes, in sequence, a convolutional layer (5x5, 64, 1), a LayerNorm layer for convolutional channel dimension normalization, a convolutional layer (1x1, 256, 1), a GELU activation function, a fully connected layer (1x1, 64, 1), and a coefficient weighting layer; Branch 3 is a straight connection layer; the convolutional features output by the three branches are added to obtain the final output result of the module.

[0084] Step 2.1.2, constructing basic module 2. Adding three parallel branches. Branch 1 includes, in sequence, a convolutional layer (7x7, 128, 1), a LayerNorm layer for convolutional channel dimension normalization, a convolutional layer (1x1, 512, 1), a GELU activation function, a fully connected layer (1x1, 128, 1), and a coefficient weighting layer; Branch 2 includes, in sequence, a convolutional layer (5x5, 128, 1), a LayerNorm layer for convolutional channel dimension normalization, a convolutional layer (1x1, 512, 1), a GELU activation function, a fully connected layer (1x1, 128, 1), and a coefficient weighting layer; Branch 3 is a straight connection layer; the convolutional features output by the three branches are added to obtain the final output result of the module.

[0085] Step 2.1.3, constructing basic module 3. Three parallel branches are added. Branch 1 comprises, in sequence, a convolutional layer (7x7, 256, 1), a LayerNorm layer with convolutional channel dimension normalization, a convolutional layer (1x1, 1024, 1), a GELU activation function, a fully connected layer (1x1, 256, 1), and a coefficient weighting layer; branch 2 comprises, in sequence, a convolutional layer (5x5, 256, 1), a LayerNorm layer with convolutional channel dimension normalization, a convolutional layer (1x1, 1024, 1), a GELU activation function, a fully connected layer (1x1, 256, 1), and a coefficient weighting layer; branch 3 is a straight connection layer; the convolutional features output by the three branches are added to obtain the final output result of the module.

[0086] Step 2.1.4, constructing basic module 4. Three parallel branches are added. Branch 1 comprises, in sequence, a convolutional layer (7x7, 512, 1), a LayerNorm layer with convolutional channel dimension normalization, a convolutional layer (1x1, 2048, 1), a GELU activation function, a fully connected layer (1x1, 512, 1), and a coefficient weighting layer; branch 2 comprises, in sequence, a convolutional layer (5x5, 512, 1), a LayerNorm layer with convolutional channel dimension normalization, a convolutional layer (1x1, 2048, 1), a GELU activation function, a fully connected layer (1x1, 512, 1), and a coefficient weighting layer; branch 3 is a straight connection layer; the convolutional features output by the three branches are added to obtain the final output result of the module.

[0087] In the present application, the coefficient weighting layer point multiplies a weighting vector with the convolutional features at the channel dimension, and the length of the weighting vector is the same as the number of channels of the convolutional features.

[0088] Figure 3 A multi-scale convolutional basic module structure diagram used in the construction of the multi-scale deep convolutional neural regression network model in the present application is shown, where x is the number of output channels of the module, the module is constructed based on the current mainstream deep convolutional neural network architecture, branch 1 and branch 2 in each basic module can enhance the adaptability to different scales by configuring different convolutional operation process parameters, combining LN layer standardization and GELU activation function, thereby improving the ability to discover data dynamics and the timing of capturing fine features, and branch 3 is a straight connection branch, which can enhance the trainability of the network in the case of large depth.

[0089] Figure 5 A multi-scale deep convolutional neural regression network modeling scheme diagram is shown, the input size of the three-dimensional array element domain frequency domain data sample is set to a x b x 3, a > 16, b > 16, int() represents rounding down, R out represents the length of the high-resolution spatial energy spectrum, and the scheme includes multiple convolutional layers, pooling layers, and basic modules.

[0090] Referring toFigure 4 and 5 The main process of constructing the multi-scale deep convolutional neural regression network model is as follows:

[0091] Step 2.2.1, add a convolution layer (4x4, 64, 2) and a LayerNorm layer;

[0092] Step 2.2.2, add three basic modules 1 in series;

[0093] Step 2.2.3, add a LayerNorm layer and a convolution layer (2x2, 128, 2);

[0094] Step 2.2.4, add three basic modules 2 in series;

[0095] Step 2.2.5, add a LayerNorm layer and a convolution layer (2x2, 256, 2);

[0096] Step 2.2.6, add nine basic modules 3 in series;

[0097] Step 2.2.7, add a LayerNorm layer and a convolution layer (2x2, 512, 2);

[0098] Step 2.2.8, add three basic modules 4 in series;

[0099] Step 2.2.9, add a global average pooling layer and a full connection layer (512, N spatial_spec , N spatial_spec represents the length of the output spatial energy spectrum.

[0100] In the present application, the features are compressed between adjacent different basic modules through the convolution layer, and the data are normalized through the LayerNorm layer. By comprehensively utilizing the multi-scale convolution basic modules under multiple different parameters, and combining the use of convolution layer, global average pooling layer and full connection layer, the array domain frequency domain data dimension reduction and feature abstraction are realized, and finally the high-resolution spatial energy spectrum is output.

[0101] Step 3, constructing a weighted enhanced loss function for the multi-scale deep convolutional neural regression network model.

[0102] In step 3 of the present application, the model loss function is constructed based on the least mean square error method, and according to whether there is a target in each orientation corresponding to the label of the training sample, a corresponding weighting coefficient is given to form a multi-target weighted loss function. On this basis, the model training process is divided into two stages, and the optimizer, learning rate, multi-target weighting coefficient and the like are reasonably set, and at the same time, the model training termination condition is set according to the model performance test trend based on the validation sample set.

[0103] The step 3 comprises the following steps:

[0104] Step 3.1, for the real high-resolution spatial energy spectrum label y lable Gradually zero-value statistics are performed and divided into real non-zero subsequences and real zero-value subsequences

[0105] Wherein, L0 and L1 represent the results of no target and target in the real high-resolution spatial energy spectrum label respectively, length(y Lable ) represents the sequence length of the real high-resolution spatial energy spectrum label, i0 and i1 are indexes of the high-resolution spatial energy spectrum;

[0106] Step 3.2, according to the division of the real non-zero subsequence and the real zero-value subsequence, the predicted high-resolution enhanced spatial energy spectrum R Model is divided in the same position

[0107] Step 3.3, constructing a weighted enhanced loss value Wherein, N B is the batch size during training, and a0 and a1 are weighting coefficients.

[0108] Step 4, training the multi-scale deep convolutional neural regression network model based on the training sample set and the weighted enhanced loss function.

[0109] In step 4 of the present application, the established multi-scale convolutional neural network model is trained, and the model training process is divided into two stages, including that in the first stage, the first training sample set is randomly divided into independent training sample set 1 and test sample set 1 according to 500:1, and the model is trained based on the first training sample set, and in the second stage, the second training sample set is randomly divided into independent training sample set 2 and test sample set 2 according to 500:1, and the model trained in the first stage is trained again. Referring to Figure 6 , the specific process is as follows.

[0110] Step 4.1, training stage 1, the main steps are as follows:

[0111] Step 4.1.1, the model is trained based on the SGD optimizer, wherein the learning rate (lr) is set to 0.001, the impulse (momentum) is set to 0.9, and the weight decay is set to 0.00001; the model is set to learn 1 batch sample per round, and the batch-size is set to 64; a0 and a1 are both set to 1.0, that is, degenerate into a non-weighted loss function.

[0112] Step 4.1.2, a batch of input samples Sbatch and the corresponding label L batch , S batch and L batch are normalized, S batch is taken as the input of the multi-scale convolutional neural network model, and R batch , L batch and R batch are calculated according to the loss function in step 3, and are optimized based on the Adam optimizer.

[0113] Step 4.1.3, repeat step 4.1.2, and carry out multiple rounds of model training. Set every 50 rounds of training as an epoch, and the same below. After each epoch training, test the model performance based on the test sample set 1, and calculate the target orientation recognition accuracy. When the target orientation calculation accuracy rises by less than 3% per 100 epochs, stop the model training.

[0114] Step 4.2, training phase 2, the main steps are as follows:

[0115] Step 4.2.1, based on the Adam optimizer, the model is trained, wherein the learning rate (lr) is set to 0.0001; set the model to learn 1 batch sample per round, and the batch-size is set to 64; and a0 and a1 are set to 1.0 and 2.0 respectively.

[0116] Step 4.2.2, randomly sample from the training sample set 2 with replacement to obtain a batch of input samples S batch and the corresponding label L batch , S batch and L batch are normalized, S batch is taken as the input of the multi-scale convolutional neural network model, and R batch , L batch and R batch are calculated according to the loss function in step 3, and are optimized based on the Adam optimizer.

[0117] Step 4.2.3, repeat step 4.2.2, and carry out multiple rounds of model training. Set every 50 rounds of training as an epoch, and the same below. After each epoch training, test the model performance based on the test sample set 2, and calculate the target orientation recognition accuracy. When the target orientation calculation accuracy rises by less than 1% per 100 epochs, stop the model training.

[0118] Step 5, take the new three-dimensional array element domain frequency domain data as the input, and obtain the predicted high-resolution spatial energy spectrum through the trained multi-scale deep convolutional neural regression network model.

[0119] In step 5 of the present application, when new unknown two-dimensional element domain frequency domain data appears, it can be processed according to steps 1.4 and 1.5 to obtain three-dimensional element domain frequency domain data, which is used as the input of the trained multi-scale deep convolution neural regression network model obtained in step 4, and the high-resolution spatial energy spectrum can be directly predicted.

[0120] Figure 7 The simulation verification results of the passive high-resolution processing method based on the multi-scale convolution neural regression network model are shown. The model has been trained and improved using a large-scale simulation sample set. It can be seen that, compared with the conventional beamforming result, the resolution of the result output by the multi-scale convolution neural network model is higher, and the effectiveness of the network model is verified.

[0121] Based on the construction of a high-resolution enhanced training sample set, a multi-scale convolution framework is designed, an adaptive deep regression network model is established, a multi-target weighted model loss function is proposed, a model optimization learning training strategy is designed, and the complex nonlinear relationship fitting between the three-dimensional element domain frequency domain data and the high-resolution spatial energy spectrum label is realized. Compared with the traditional method, the generalization ability of the method is stronger. The method is applied to the simulation data processing, and good results are obtained.

Claims

1. A passive high-resolution processing method based on a multi-scale deep convolutional neural regression network, characterized in that: The method comprises the following steps: Step 1: Construct a training sample set. Any training sample in the training sample set includes three-dimensional array element domain frequency domain data x sample and the corresponding true high-resolution spatial energy spectrum label y lable ; Step 2: Build a multi-scale deep convolutional neural regression network model based on the training samples; Step 3: Construct a weighted enhancement loss function for the multi-scale deep convolutional neural regression network model; Step 4: training a multi-scale deep convolutional neural regression network model based on the training sample set and the weighted enhancement loss function; Step 5: Using the new 3D array element domain frequency domain data as input, the trained multi-scale deep convolutional neural regression network model is used to obtain the predicted high-resolution spatial energy spectrum. The construction of the multi-scale convolutional neural regression network model includes the following steps: Add convolutional layers and LayerNorm layers; Add several basic modules 1; Add LayerNorm layer and convolution layer; Add several basic modules 2; Add LayerNorm layer and convolution layer; Add several basic modules 3; Add LayerNorm layer and convolution layer; Add several basic modules 4; Add global average pooling layer and fully connected layer; The step 3 comprises the following steps: Step 3.1, the real high-resolution spatial energy spectrum label y lable Perform gradual zero value statistics and divide into non-zero subsequences and zero value subsequences Among them, L0 and L1 represent the results without target and with target in the real high-resolution spatial energy spectrum, respectively. Lable ) represents the sequence length of the real high-resolution spatial energy spectrum label, i0 and i1 are the indexes of the high-resolution spatial energy spectrum; Step 3.2: According to the division of true non-zero subsequences and true zero-value subsequences, the predicted high-resolution enhanced spatial energy spectrum R Model Perform same-position division ; Step 3.3, construct weighted enhancement loss value , where N B is the batch size during training, and α0 and α1 are weighting coefficients.

2. A passive high-resolution processing method based on a multi-scale deep convolutional neural regression network according to claim 1, characterized in that: The construction of any of the training samples comprises the following steps: Step 1.1, setting the range of input parameters of the passive broadband warning detection model, wherein the input parameters include target parameters, which include target number, target signal-to-noise ratio, and target direction; Step 1.2: Set the parameter values ​​of the input parameters and generate the two-dimensional array element domain frequency domain data x through the passive broadband warning detection model. f And the corresponding one-dimensional space energy spectrum y; Step 1.3, for the two-dimensional array element domain frequency domain data x f Perform three-dimensional expansion and splicing processing to obtain three-dimensional array element domain frequency domain data x sample ; Step 1.4: Perform high-resolution enhancement processing on the one-dimensional space energy spectrum y to obtain the true high-resolution space energy spectrum label y lable .

3. The passive high-resolution processing method based on a multi-scale deep convolutional neural regression network according to claim 2, characterized in that: The step 1.3 includes the following steps: Step 1.3.1, for the two-dimensional array element domain frequency domain data x f Calculate the real part, imaginary part and modulus to obtain three two-dimensional matrices x f，real 、x f，imag 、x f，mod ; Step 1.3.2: Each two-dimensional matrix is ​​expanded by one dimension to obtain a three-dimensional matrix x f，real_M 、x f，imag_M 、x f，mod_M , and perform splicing operations on the extended dimension to obtain the three-dimensional array element domain frequency domain data x sample .

4. The passive high-resolution processing method based on a multi-scale deep convolutional neural regression network according to claim 2, characterized in that: The step 1.4 includes the following steps: Step 1.4.1: Take out the sequence y of sliding window length from the one-dimensional space energy spectrum y according to the beam number. sub ; Step 1.4.2, based on sequence y sub The median of the smallest m values ​​is calculated and used as the background of the beam to obtain the background sequence y back ; Step 1.4.3, generate a zero-value sequence y with the same length as the one-dimensional space energy spectrum length(y) target , according to the beam number sequence N corresponding to the target direction target Assign the zero-value sequence to the corresponding spatial energy spectrum, denoted as y target (N target )=y(N target ); Step 1.4.4, calculate the true high-resolution spatial energy spectrum label y lable is the sum of the zero value sequence and the background sequence after assignment y lable =y target +y back .

5. The passive high-resolution processing method based on a multi-scale deep convolutional neural regression network according to claim 2, characterized in that: The training sample set includes a first training sample set and a second training sample set. During the generation process of the second training sample set, the number of targets set is not less than 2, and there are at least 2 targets with an azimuth interval of not more than 2 beams.

6. The passive high-resolution processing method based on a multi-scale deep convolutional neural regression network according to claim 1, characterized in that: The basic module 1, basic module 2, basic module 3 and basic module 4 all include 3 parallel branches; branch 1 and branch 2 both include a convolution layer, a LayerNorm layer for normalizing the convolution channel dimension, a convolution layer, a GELU activation function, a fully connected layer, and a coefficient weighted layer connected in sequence, wherein the coefficient weighted layer multiplies the convolution feature by a weighted vector in the channel dimension, and the length of the weighted vector is the same as the number of convolution feature channels; branch 3 is a direct connection layer; the convolution features output by the three branches are added to obtain the output of the module.

7. The passive high-resolution processing method based on a multi-scale deep convolutional neural regression network according to claim 1, characterized in that: The step 4 comprises the following steps: Step 4.1: train the multi-scale deep convolutional neural regression network model using the first training sample set, where both α0 and α1 in the weighted enhancement loss function are 1.0; Step 4.2: Perform secondary training on the multi-scale deep convolutional neural regression network model using the second training sample set, where α0 = 1.0 and α1 = 2.0 in the weighted enhancement loss function.

Images