Passive weak target detection method based on attention mechanism network

Abstract

The application discloses a passive weak target detection method based on an attention mechanism network, comprising the following steps: step 1, constructing a simulation weak target detection initial sample set and performing pretreatment to obtain a plurality of three-dimensional array element domain frequency domain data and corresponding weak target enhancement labels, and forming a simulation weak target detection sample set; step 2, constructing an attention mechanism network model; step 3, constructing a weak target weighted enhancement loss function for training of the attention mechanism network model; step 4, training the attention mechanism network model based on the simulation weak target detection sample set; step 5, taking unknown three-dimensional array element domain frequency domain data as input of the trained attention mechanism network model to obtain a predicted weak target inference result; and step 6, decoding the predicted weak target inference result to obtain a spatial energy spectrum.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of feature extraction of underwater targets and artificial intelligence, and particularly relates to a passive weak target detection method based on an attention mechanism network. BACKGROUND

[0002] Weak target detection is one of the focuses of current sonar signal processing research. With the application of various modern signal processing theories and technologies, the weak target detection performance of current broadband passive sonar systems has been improved to a certain extent, but the traditional detection system has not broken through, and the weak target detection capability under the influence of strong interference and the like is still difficult to fully meet the actual use requirements.

[0003] In recent years, deep learning technology has developed rapidly and has become a hot research method in the field of information processing. The application draws lessons from the latest trends and achievements of deep learning methods, proposes an adaptive attention mechanism network modeling and training method for weak target detection, takes array element domain frequency domain data as input, directly outputs spatial energy spectrum features, and improves the weak target detection capability by deeply mining useful information from bottom to top. SUMMARY

[0004] One of the purposes of the application is to provide a passive weak target detection method based on an attention mechanism network to solve the problem of poor weak target detection capability under the influence of strong interference and the like in the background art.

[0005] To achieve the above-mentioned purposes, the application provides the following technical solutions.

[0006] A passive weak target detection method based on an attention mechanism network, the method comprising the following steps:

[0007] Step 1, constructing an initial sample set of simulated weak target detection and performing pretreatment to obtain a plurality of three-dimensional array element domain frequency domain data and corresponding weak target enhanced labels, forming a simulated weak target detection sample set;

[0008] Step 2, constructing an attention mechanism network model;

[0009] Step 3, constructing a weak target weighted enhancement loss function for attention mechanism network model training;

[0010] Step 4, training the attention mechanism network model based on the simulated weak target detection sample set;

[0011] Step 5, taking unknown three-dimensional array element domain frequency domain data as the input of the trained attention mechanism network model to obtain a predicted weak target inference result;

[0012] Step 6, decoding the predicted weak target inference result to obtain a spatial energy spectrum.

[0013] Preferably, the step 1 comprises the following steps:

[0014] Step 1.1, constructing a passive broadband warning detection model, the input parameters of which include sonar array parameters, target parameters and background noise parameters, the target parameters including target quantity, target signal-to-noise ratio and target direction, and the output being two-dimensional array element domain frequency domain data and spatial energy spectrum;

[0015] Step 1.2, setting sonar array parameters according to actual parameter requirements of the sonar array, setting target quantity and signal-to-noise ratio range and background noise parameters according to use requirements;

[0016] Step 1.3, generating a single sample, randomly setting target quantity and corresponding signal-to-noise ratio according to the set target quantity and signal-to-noise ratio range, and randomly setting target corresponding direction, and generating two-dimensional array element domain frequency domain data of random target quantity, direction and signal-to-noise ratio and corresponding spatial energy spectrum through the passive broadband warning model;

[0017] Step 1.4, performing dimension expansion splicing operation on the two-dimensional array element domain frequency domain data obtained in step 1.3 to obtain three-dimensional array element domain frequency domain data;

[0018] Step 1.5, performing amplitude enhancement operation on the spatial energy spectrum obtained in step 1.3 to obtain a weak target enhancement label.

[0019] Preferably, the step 1.4 comprises the following steps:

[0020] Step 1.4.1, obtaining three two-dimensional matrices x f , x f，real , x f，imag from the two-dimensional array element domain frequency domain data x f，mod by taking real part, imaginary part and modulus respectively;

[0021] Step 1.4.2, expanding each two-dimensional matrix by 1 dimension in the same dimension to obtain three three-dimensional matrices x f，real_M , x f，imag_M , x f，mod_M , and performing splicing operation in the expanded dimension to obtain three-dimensional array element domain frequency domain data x Merge ;

[0022] Step 1.4.3, performing normalization processing on the three-dimensional array element domain frequency domain data x Merge to obtain normalized three-dimensional array element domain frequency domain data x Sample .

[0023] Preferably, the step 1.5 comprises the following steps:

[0024] Step 1.5.1, performing normalization processing on the spatial energy spectrum y to obtain normalized spatial energy spectrum y1;

[0025] Step 1.5.2, reduce the normalized spatial energy spectrum y1 by 50% to obtain the reduced spatial energy spectrum y2;

[0026] Step 1.5.3, according to the set target number and target direction, calculate the beam number n corresponding to each target beam , and perform amplitude enhancement y2(n beam ) = y2(n beam ) + N enhance , N enhance is the target enhancement amplitude, and obtain the weak target enhancement label y Lable = y2.

[0027] Preferably, the step 2 comprises the following steps:

[0028] Step 2.1: constructing a basic attention mechanism module;

[0029] Step 2.2, sample segmentation and linear dimension transformation are performed on the input three-dimensional element domain frequency domain data to obtain a slice feature sequence;

[0030] Step 2.3, constructing a slice position sequence and a category information sequence, and combining the slice feature sequence for multi-sequence weighted combination;

[0031] Step 2.4, the multi-sequence weighted combination result is subjected to feature extraction through a plurality of basic attention mechanism modules connected in series;

[0032] Step 2.5, merging the extracted features;

[0033] Step 2.6, dimension reduction is performed on the slice merging result through a fully connected layer to obtain a weak target reasoning result.

[0034] Preferably, the basic attention mechanism module comprises a basic module 1 and a basic module 2 connected in series;

[0035] The construction of the basic module 1 comprises the following steps:

[0036] Step 2.1.1, adding three fully connected layers connected in series to realize scaling of the three-dimensional matrix in the third dimension;

[0037] Step 2.1.2, sequentially adding a matrix dimension transformation layer and a matrix splitting processing layer connected in series to split the scaled three-dimensional matrix into three three-dimensional matrices M1, M2, M3;

[0038] Step 2.1.3, sequentially adding three parallel matrix dimension transformation layers for dimension transformation of the three-dimensional matrices M1, M2, M3, respectively, to obtain matrices M'1, M'2, M'3;

[0039] Step 2.1.4, a Dropout layer is added to perform Dropout processing on the matrix M'1, a matrix multiplication operation processing layer is added to multiply the matrices M'1 and M'2, obtaining a matrix N1; a coefficient weighting layer is added to multiply the matrix N1 by a preset coefficient, a Softmax layer and a Dropout layer are added, and the matrix N1 multiplied by the preset coefficient is sequentially processed;

[0040] Step 2.1.5, a matrix multiplication operation processing layer is added to implement the multiplication operation of the matrix N1 and the matrix M'3, obtaining a matrix N2; a matrix dimension transformation layer is added to perform dimension transformation on the matrix N2, obtaining a matrix N3, a fully connected layer and a Dropout layer are added, and N3 is sequentially processed to obtain the output result of the module;

[0041] The construction of the basic module 2 includes the following steps:

[0042] Step 2.2.1, sequentially add a matrix dimension transformation layer to convert the size (a1, a2, a3) of the input three-dimensional matrix in the spatial dimension to obtain a four-dimensional matrix (a1, a3, a2, 1);

[0043] Step 2.2.2, construct a series of convolution layers in series for feature extraction of the four-dimensional matrix, and perform Dropout processing on the adjacent two convolution layers through a Dropout layer, wherein the convolution layer and the Dropout layer are connected through a GELU activation function;

[0044] Step 2.2.3, add a Dropout layer and a matrix dimension transformation layer to convert the extracted features to a matrix with the same size (a1, a2, a3) as the original three-dimensional matrix after Dropout processing.

[0045] Preferably, the step 3 includes the following steps:

[0046] Step 3.1, divide the weak target enhancement label into a no-target sequence L Non and a target sequence L Target ;

[0047] Step 3.2, set a target discrimination threshold T, and divide the target sequence L Target into a strong target sequence L Target，1 and a weak target sequence L Target，2 ;

[0048] Step 3.3, based on the positions of the target sequences in the weak target enhancement label, obtain three corresponding sub-sequences B Non , B Target，1 , and B Target，2 in the predicted weak target inference result;

[0049] Step 3.4, constructing the weak target weighted enhancement loss function expression as

[0050]

[0051] wherein N1, N2, N3 are the sequence lengths of L Non , L Target，1 , L Target，2 , and α1, α2, α3 are the corresponding weak target weighting coefficients.

[0052] Preferably, the step 4 comprises the following steps:

[0053] Step 4.1, based on the SGD optimizer, the model is trained for fast fitting, and during the training, the α1, α2, α3 of the weak target weighted enhancement loss function are set to 1, 1.5, 2 respectively;

[0054] Step 4.2, based on the Adam optimizer, the model obtained in step 4.1 is trained for accurate fitting, and during the training, the α1, α2, α3 of the weak target weighted enhancement loss function are set to 1, 2, 3 respectively;

[0055] Step 4.3, based on the Adam optimizer, the model obtained in step 4.2 is trained for gradual fitting, and during the training, the α1, α2, α3 of the weak target weighted enhancement loss function are set to 1, 2, 5 respectively.

[0056] Preferably, the step 6 comprises the following steps:

[0057] Step 6.1, setting the amplitude reduction double threshold as N decode1 , N decode2 , and N enhance >N decode1 >N decode2 ;

[0058] Step 6.2, traversing the amplitude R(i) of the predicted weak target inference result, i = 1, 2,..., N Beam , judging R(i) > N decode2 , if yes, R(i) = R(i)-N decode2 , otherwise unchanged;

[0059] Step 6.3, traversing the amplitude R(i) of the weak target enhancement label obtained in step 6.2, i = 1, 2,..., N Beam , and judging R(i) > N decode1 , if yes, R(i) = R(i)-N decode1 , otherwise unchanged;

[0060] Step 6.4, the amplitude of the weak target enhancement label obtained in step 6.3 is increased by 100% as a whole, and the corresponding spatial energy spectrum is obtained.

[0061] Compared with the prior art, the present application has the following advantages:

[0062] The present application introduces the attention mechanism network method into passive weak target detection, and on the basis of constructing a deep network model adapting to the characteristics of array element domain frequency domain data, also proposes elements such as training sample label enhancement coding, weak target weighted loss function and multi-stage model training strategy, which can effectively enhance the learning ability of the attention mechanism network model to the characteristics of weak targets. Compared with traditional signal processing methods, the intelligent network model proposed by the present application has stronger non-linear mapping relationship learning ability, and can also improve the processing performance in different scenes by continuously learning new data, and has better generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The flowchart of the present application.

[0064] Figure 2 The weak target enhancement label construction schematic diagram.

[0065] Figure 3 The attention mechanism network modeling schematic diagram.

[0066] Figure 4 The weak target weighted enhancement loss function construction method schematic diagram.

[0067] Figure 5 The model training stage framework structure schematic diagram.

[0068] Figure 6 The simulation weak target detection result schematic diagram. DETAILED DESCRIPTION

[0069] 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 part 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 skilled in the art without creative labor fall within the protection scope of the present application.

[0070] Referring to Figure 1 As shown in the figure, a passive weak target detection method based on an attention mechanism network includes five stages of training sample set construction, attention mechanism network model construction, model loss function construction, attention mechanism network model training and weak target detection application, and specifically includes the following 6 steps.

[0071] Step 1, construct the initial sample set of simulation weak target detection and perform pretreatment to obtain a plurality of three-dimensional array element domain frequency domain data and corresponding weak target enhancement labels, forming a simulation weak target detection sample set.

[0072] In step 1 of the present application, the initial sample set for simulating weak target detection includes a plurality of samples, each sample including two-dimensional element domain frequency domain data and a corresponding spatial energy spectrum. The sample set for simulating weak target detection is obtained by preprocessing the samples. The specific process of step 1 is shown below.

[0073] In step 1.1, a passive wideband warning detection model is constructed. The input parameters of the model mainly include sonar array parameters, target parameters and background noise parameters, and additionally include fft points, etc. The output is two-dimensional element domain frequency domain data and a spatial energy spectrum. The sonar array parameters mainly include the number of array elements, the sampling rate, the processing frequency, the array position vector, etc. The target parameters include the number of targets, the target signal-to-noise ratio and the target bearing.

[0074] In step 1.1 of the present application, the passive wideband warning detection model is a commonly used array data processing model in the field. The model can generate two-dimensional element domain frequency domain data and a corresponding spatial energy spectrum. The construction of the model includes the following three sub-steps: step 1.1.1, constructing a passive array data simulation model. The input includes sonar array parameters, target parameters and background noise parameters. The output is two-dimensional element domain frequency domain data. Step 1.1.2, generating an array manifold matrix. The input mainly includes the observation space vector, the array position vector and the processing frequency. Step 1.1.3, processing the two-dimensional element domain frequency domain data x f based on a conventional beamforming algorithm model to output the corresponding spatial energy spectrum y, which is a one-dimensional sequence. The construction method of the passive wideband warning detection model is a conventional technical means in the field, and will not be described here.

[0075] In step 1.2, the sonar array parameters are set according to the actual parameter requirements of the sonar array. 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.

[0076] In step 1.3, a single sample is generated. The number of targets and the corresponding signal-to-noise ratio are randomly set according to the set number of targets and the signal-to-noise ratio range. The corresponding bearing of the target is also randomly set. The two-dimensional element domain frequency domain data x f and the corresponding spatial energy spectrum y of the random number of targets, bearings and signal-to-noise ratios are generated by the passive array data simulation model.

[0077] In step 1.4, the two-dimensional element domain frequency domain data x f obtained in step 1.3 is subjected to a dimension expansion splicing operation to obtain three-dimensional element domain frequency domain data x Sample .

[0078] The dimension expansion splicing operation specifically includes the following steps:

[0079] Step 1.4.1, the two-dimensional element domain frequency domain data xf Real part, imaginary part and modulus, respectively, three two-dimensional matrix x f，real , x f，imag , x f，mod ;

[0080] Step 1.4.2, each two-dimensional matrix is respectively extended by 1 dimension to obtain three three-dimensional matrices x f，real_M , x f，imag_M , x f，mod_M , so that the value of dimension 1 in the three-dimensional matrix is 1, the value of dimension 2 is consistent with the value of dimension 1 in the two-dimensional matrix, and the value of dimension 3 is consistent with the value of dimension 2 in the two-dimensional matrix, and the three three-dimensional matrices x f，real_M , x f，imag_M , x f，mod_M are spliced in dimension 1 to obtain three-dimensional cell domain frequency domain data x Merge ;

[0081] Step 1.4.3, the three-dimensional cell domain frequency domain data x Merge is normalized according to the following formula to obtain normalized three-dimensional cell domain frequency domain data x Sample .

[0082] In step 1.4.2 of the application, taking the two-dimensional matrix x f，real as an example, it is extended by 1 dimension to obtain the three-dimensional matrix x f，real_M , whose three dimensions are 1, x f，real dimension 1, x f，real dimension 2, and the value of x f，real_M in dimension 2 is consistent with the value of x f，real in dimension 1.

[0083] Step 1.5, the amplitude enhancement operation is performed on the spatial energy spectrum y obtained in step 1.3 to obtain a weak target enhancement label y lable .

[0084] In step 1.5 of the application, the amplitude of the weak target is strengthened by weighting the normalized target and background amplitude, which provides a basis for improving the training effect of the attention mechanism network model.

[0085] Referring to FIG. Figure 2 , the amplitude enhancement operation specifically includes the following steps:

[0086] Step 1.5.1, the spatial energy spectrum y is normalized according to to obtain the normalized spatial energy spectrum y1;

[0087] Step 1.5.2, the amplitude of the normalized spatial energy spectrum y1 is reduced by 50% as a whole to obtain the amplitude-reduced spatial energy spectrum

[0088] Step 1.5.3, according to the target number and target direction set, calculate the beam number n corresponding to each target beam , and the amplitude of each target is enhanced y2(n beam ) = y2(n beam ) + N enhanca , N enhance is the target enhancement amplitude, generally set interval is [0.4, 0.7], get weak target enhancement label y Lable = y2.

[0089] Step 2, construct attention mechanism network model, refer to Figure 3 In the attention mechanism network model, the sample segmentation and linear dimension transformation are carried out on the input array domain frequency domain data to obtain the slice feature sequence about the input sample, at the same time, the slice position sequence and the category information sequence are constructed, the above three are combined to form a unified numerical representation by feature sequence weighting, and then sent to the attention mechanism coding unit composed of multiple attention mechanism modules to realize feature deep abstraction and nonlinear mapping, which is the core element to realize weak target detection. Finally, the output features of the attention mechanism coding unit are reduced in dimension by slice merging and full connection layer to obtain the weak target detection result.

[0090] As a specific embodiment of constructing the attention mechanism network model, the main process is as follows:

[0091] Step 2.1, construct basic attention mechanism module, including basic module 1 and basic module 2 in series.

[0092] (1) The construction of basic module 1 includes the following process:

[0093] Add three full connection layers connected in turn to realize the scaling of three-dimensional matrix (a1, a2, a3) in the third dimension, the parameters are (a3, 3072), (3072, 12888) and (12288, 6144), a1, a2, a3 represent the size of each dimension of the matrix respectively, the same below, (a3, 3072) is taken as an example to illustrate, the two numbers represent the size before and after expansion respectively; through the above processing, the matrix dimension scaling process can be represented as (a1, a2, a3) -> (a1, a2, 3072) -> (a1, a2, 12288) -> (a1, a2, 6144), where the numbers in the brackets represent the size of each dimension of the output matrix at this stage, the same below;

[0094] Add matrix dimension transformation layer (a1, a2, 6144) -> (3, a1, a2, 2048), where (a1, a2, 6144) and (3, a1, a2, 2048) represent the matrix dimensions before and after transformation respectively;

[0095] Add a matrix splitting processing layer to split the matrix of dimension (3, a1, a2, 2048) to generate three sub-matrices of size (a1, a2, 2048), denoted as M1, M2 and M3;

[0096] Add three parallel matrix dimension transformation layers to realize dimension transformation of (a1, a2, 2048) -> (a1, 32, a2, 64), (a1, a2, 2048) -> (a1, 32, 64, a2) and (a1, a2, 2048) -> (a1, 32, a2, 64), and respectively used for dimension transformation of M1, M2 and M3, to obtain matrices M1', M2' and M3';

[0097] Add a Dropout layer to perform Dropout processing on M1';

[0098] Add a matrix multiplication operation processing layer to realize multiplication operation of M1' and M2' to obtain matrix N1;

[0099] Add a coefficient weighting layer to multiply N1 by a coefficient 0.125;

[0100] Add a Softmax layer and a Dropout layer to process N1 multiplied by the coefficient in turn;

[0101] Add a matrix multiplication operation processing layer to realize multiplication operation of N1 and M3' to obtain N2;

[0102] Add a matrix dimension transformation layer to realize dimension transformation of (a1, 32, a2, 64) -> (a1, a2, 2048) and used for processing N2 to obtain N3;

[0103] Add a fully connected layer (2048, a3) and a Dropout layer to process N3 in turn to output the result of this module.

[0104] (2) The construction of the basic module 2 includes the following process:

[0105] Add a matrix dimension transformation layer (a1, a2, a3) -> (a1, a3, a2, 1), a convolution layer (1x1, 2048, 1), a Dropout layer, a convolution layer (1x1, 3072, 1), a GELU activation function, a Dropout layer, a convolution layer (1x1, 2048, 1), a Dropout layer, a convolution layer (1x1, a3, 1) and a Dropout layer, a matrix dimension transformation layer (a1, a3, a2, 1) -> (a1, a2, a3) in turn, wherein the numbers in the brackets of the convolution layer represent the convolution kernel size, the number of convolution output channels and the convolution step in turn.

[0106] Step 2.2, constructing a training sample segmentation and linear dimension transformation module, including a matrix dimension transformation unit, a convolution unit and a matrix dimension transformation layer connected in sequence, outputting x Stage1 , wherein the matrix dimension transformation unit includes a plurality of matrix dimension transformation layers connected in series, and the convolution unit includes a plurality of convolution layers connected in series; the main process is as follows:

[0107] Set the matrix size of the model input as (N batch , 3, x1, x2), wherein N batch is the Batch size of the training sample, and 3, x1 and x2 are respectively the matrix dimensions of a single sample;

[0108] Set the size of the segmented sub-matrix as (N batch , 3, y1, y2);

[0109] Add a matrix dimension transformation layer to convert the matrix size of the model input sample from (N batch , 3, x1, x2) to (N batch , 3, N patch , y1, y2);

[0110] Add a matrix dimension transformation layer to convert the matrix size from (N batch , 3, N patch , y1, y2) to (N batch , 3·y1·y2, N patch , 1);

[0111] Add two convolution layers connected in series, and the parameters of the convolution layers are respectively (1×1, 2048, 1) and (1×1, 1024, 1);

[0112] Add a matrix dimension transformation layer to convert the matrix size from (N batch , 1024, N patch , 1) to (N batch , N patch , 1024), and finally output x Stage1 .

[0113] Step 2.3, constructing a category position information embedding module, including an embedding matrix generation unit, a matrix merging and expanding processing layer and a matrix addition processing layer; the embedding matrix generation unit is used to generate a category information embedding matrix with a size of (N batch , 1, 1024) and a position information embedding matrix with a size of (N batch , N patch +1, 1024), and the matrices are randomly initialized; the matrix merging and expanding processing layer is used to merge the category information embedding matrix with x Stage1 in dimension 1 to obtain x′ Stage1with size (N batch , N patch +1, 1024); a matrix addition processing layer is used to add the position information to the matrix x Stage1 , and output x Stage3 .

[0114] Step 2.4, a dropout module is constructed to perform dropout processing on x Stage2 , and output x Stage3 .

[0115] Step 2.5, a plurality of basic attention mechanism modules are repeatedly stacked to process x Stage3 , and output x Stage4 .

[0116] Step 2.6, an average processing layer is added to calculate the data average of x Stage4 in the matrix dimension 2, and obtain x Stage5 .

[0117] Step 2.7, a LayerNorm standardization layer, a full connection layer (1024, 512), and a full connection layer (512, N bf , N bf representing the spatial energy spectrum dimension, are added to process x Stage5 in turn, and output the result of the model.

[0118] Step 3, a weak target weighted enhancement loss function for attention mechanism network model training is constructed.

[0119] Referring to Figure 4 , the step 3 includes the following steps:

[0120] Step 3.1, the normalized weak target enhancement label is gradually zero-valued, and is divided into a no-target sequence L Non and a target sequence L Target , the no-target sequence L Non is a zero-value sequence, and the target sequence L Target is a non-zero-value sequence;

[0121] Step 3.2, a weak target discrimination threshold T is set, and the target sequence with an amplitude higher than T and lower than T is divided into a strong target sequence L Target，1 and a weak target sequence L Target，2 ;

[0122] Step 3.3, based on the position of each target sequence in the weak target enhancement label, three corresponding sub-sequences B Non , B Target，1 , and B Target，2 are obtained in the predicted weak target inference result.

[0123] Step 3.4, according to the amplitude division result, the weak target weighted enhancement loss function expression is constructed as

[0124]

[0125] Wherein, N1, N2, N3 are the sequence lengths of L Non , L Target，1 , L Target，2 , and α1, α2, α3 are the corresponding weak target weighting coefficients, which can be set according to the actual model training effect.

[0126] In step 3.3 of the application, according to the division result of y Lable , it can be known that L Non , L Target，1 , L Target，2 each value is extracted from the original position (i.e. beam number) before y Lable , according to the above three original position sequences, three corresponding sequences are extracted from the predicted weak target reasoning result B Model , that is, B Non , B Target，1 , B Target，2 .

[0127] In the application, for the weak target enhancement label corresponding to the array domain frequency domain data, it is statistically and divided into no target part, strong target part and weak target part according to the energy strength point by point, the above three parts of labels are weighted and combined with the attention mechanism model output result calculation and minimum loss value, and the weak target enhancement learning is realized.

[0128] Step 4, training the attention mechanism network model based on the simulation weak target detection sample set.

[0129] Referring to Figure 5 , the training of the attention mechanism network model includes the following steps:

[0130] Step 4.1, based on the SGD optimizer, the model is rapidly fitted and trained, and the specific process is as follows:

[0131] (4.1.1) set the SGD optimizer parameters, 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; set the model to learn 1 batch sample per round, and the batch_size is set to 64; α1, α2, α3 are set to 1.0, 1.5 and 2.0 respectively.

[0132] (4.1.2) randomly sample from the simulation weak target detection sample set with replacement to obtain the input sample S batch and the corresponding label Lbatch , S batch is processed based on the attention mechanism network model, and R batch is output. batch , L batch and R batch are calculated according to the loss function in step 3, and are optimized based on the SGD optimizer.

[0133] (4.1.3) Repeat step (4.1.2) to carry out multiple rounds of model training. Set every 50 rounds of training as an epoch, and the same applies below. Calculate the average training loss value of each epoch, and when the loss value reduction rate of every 100 epochs is less than 5%, stop the model training.

[0134] Step 4.2, based on the Adam optimizer, carry out precise fitting training on the model obtained in step 4.1, and the specific process is as follows:

[0135] (4.2.1) Set the learning rate (lr) of the Adam optimizer to 0.0001; set the model to learn 1 batch sample per round, and set natch_size to 64; set a1, a2, and a3 to 1.0, 2.0, and 3.0, respectively.

[0136] (4.2.2) Randomly sample from the simulated weak target detection sample set with replacement to obtain input sample S batch and corresponding label L batch , S batch is processed based on the attention mechanism network model, and R batch is output. batch , L batch and R batch are calculated according to the loss function in step 3, and are optimized based on the Adam optimizer.

[0137] (4.2.3) Repeat step (4.2.2) to carry out multiple rounds of model training. Calculate the average training loss value of each epoch, and when the loss value reduction rate of every 100 epochs is less than 1%, stop the model training.

[0138] Step 4.3, based on the Adam optimizer, carry out gradual fitting training on the model obtained in step 4.2, and the specific process is as follows:

[0139] (4.3.1) Set the learning rate (lr) of the Adam optimizer to 0.00001; set the model to learn 1 batch sample per round, and set batch_size to 64; set a1, a2, and a3 to 1.0, 2.0, and 5.0, respectively.

[0140] (4.3.2) Randomly sample from the simulated weak target detection sample set with replacement to obtain input sample S batch and corresponding label Lbatch S batch is processed based on the attention mechanism network model, and R batch is output. batch The loss value is calculated according to the loss function in step 3, and the optimization is performed based on the Adam optimizer. batch

[0141] (4.3.3) Repeat step (4.3.2) to carry out multiple rounds of model training. While calculating the average training loss value of each epoch and calculating the loss value reduction rate, test the model based on the validation sample set every 50 rounds (i.e. 1 epoch), calculate the loss value, and when the training loss value reduction rate is less than 1% every 100 epochs or the validation loss value no longer decreases every 100 epochs, stop the model training.

[0142] Step 5: Use the unknown three-dimensional element domain frequency domain data as the input of the trained attention mechanism network model to obtain the output result, i.e. the predicted weak target reasoning result R.

[0143] In step 5 of the present application, when the unknown two-dimensional element domain frequency domain data is obtained, it is processed according to step 1.5 to generate standardized samples, i.e. three-dimensional element domain frequency domain data, which is used as the input of the trained attention mechanism model to obtain the output result.

[0144] In step 5 of the present application, the model training process is divided into three stages: fast fitting, accurate fitting and gradual fitting. In each stage, the optimizer parameters, loss function weak target enhancement weighting coefficients, etc. are dynamically set according to real-time evaluation results, so that the model training is more suitable for the data.

[0145] Step 6: Decode the predicted weak target reasoning result R to obtain the spatial energy spectrum.

[0146] The step 6 includes the following steps:

[0147] Step 6.1: Set the amplitude reduction double threshold to N decode1 , N decode2 , and N enhance >N decode1 >N decode2 ;

[0148] Step 6.2: Traverse the amplitude R(i) of the predicted weak target reasoning result, i=1, 2,..., N Beam and compare it with N decode2 , if R(i) > N decode2 , then R(i) = R(i)-N decode2 , otherwise it remains unchanged;

[0149] ​Step 6.3. Traverse the amplitude R(i) of the weak target enhancement label obtained in step 6.2, i = 1, 2,..., N Beam And N decode1 Compare, if R(i) > N decode1 Then R(i) = R(i) - N decode2 , otherwise unchanged;

[0150] Step 6.4. The amplitude of the weak target enhancement label obtained in step 6.3 is increased by 100% as a whole, and a spatial energy spectrum is obtained, which is a one-dimensional sequence.

[0151] Figure 6 The simulation weak target detection result is shown, and it can be seen that, compared with the conventional beamforming result, the weak target detection ability of the result based on the attention mechanism network is significantly improved, verifying the effectiveness of the method and model proposed in the application.

Claims

1. A passive weak target detection method based on attention mechanism networks, characterized in that, The method includes the following steps: Step 1: Construct an initial sample set for simulated weak target detection and perform preprocessing to obtain several three-dimensional array element domain frequency domain data and corresponding weak target enhancement labels, thus forming a simulated weak target detection sample set; Step 2: Construct an attention mechanism network model; Step 3: Construct a weak objective weighted augmentation loss function for training attention mechanism network models; Step 4: Train the attention mechanism network model based on the simulated weak target detection sample set; Step 5: Use the unknown three-dimensional array element domain frequency domain data as input to the trained attention mechanism network model to obtain the predicted weak target inference result. Step 6: Decode the predicted weak target inference results to obtain the spatial energy spectrum; Step 2 includes the following steps: Step 2.1: Construct the basic attention mechanism module; Step 2.2: Perform sample segmentation and linear dimension transformation on the input three-dimensional array element domain frequency domain data to obtain the piecewise feature sequence; Step 2.3: Construct the segmentation location sequence and category information sequence, and combine them with the segmentation feature sequence to perform a multi-sequence weighted combination; Step 2.4: The weighted combination result of multiple sequences is used to extract features through several interconnected basic attention mechanism modules; Step 2.5: Segment and merge the extracted features; Step 2.6: The dimensionality of the merging results is reduced by a fully connected layer to obtain the weak target inference results; The basic attention mechanism module includes a basic module 1 and a basic module 2 connected in series; The construction of basic module 1 includes the following steps: Step 2.1.1: Add three cascaded fully connected layers to achieve scaling of the 3D matrix in the third dimension; Step 2.1.2: Add a cascaded matrix dimension transformation layer and a matrix splitting processing layer in sequence to split the scaled 3D matrix into three 3D matrices M1, M2, and M3. Step 2.1.3: Add three parallel matrix dimension transformation layers in sequence to perform dimension transformation on the three-dimensional matrices M1, M2, and M3 respectively, to obtain matrices M′1, M′2, and M′3; Step 2.1.4: Add a Dropout layer to perform Dropout processing on matrix M′1, add a matrix multiplication operation processing layer to multiply matrices M′1 and M′2 to obtain matrix N1; add a coefficient weighting layer to multiply matrix N1 by preset coefficients, add a Softmax layer and a Dropout layer, and process the matrix N1 after multiplying by preset coefficients in sequence. Step 2.1.5: Add a matrix multiplication operation processing layer to perform the multiplication operation of matrix N1 and matrix M′3 to obtain matrix N2; add a matrix dimension transformation layer to transform the dimension of matrix N2 to obtain matrix N3; add a fully connected layer and a Dropout layer to process N3 in sequence to obtain the output result of this module. The construction of basic module 2 includes the following steps: Step 2.2.1: Add matrix dimension transformation layers sequentially to transform the dimensions (a1, a2, a3) of the input three-dimensional matrix in spatial dimensions to obtain a four-dimensional matrix (a1, a3, a2, 1); Step 2.2.2: Construct several concatenated convolutional layers for feature extraction from the four-dimensional matrix. Adjacent convolutional layers are connected by a Dropout layer for Dropout processing. One convolutional layer and the Dropout layer are connected by the GELU activation function. Step 2.2.3: Add a Dropout layer and a matrix dimension transformation layer to convert the extracted features into a matrix with the same spatial dimensions as the original three-dimensional matrix (a1, a2, a3) after Dropout processing.

2. The passive weak target detection method based on an attention mechanism network as described in claim 1, characterized in that, Step 1 includes the following steps: 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 spatial energy spectrum. 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; Step 1.3: For single sample generation, the number of targets and the corresponding signal-to-noise ratio are randomly set according to the set number of targets and signal-to-noise ratio range. At the same time, the corresponding azimuth of the targets is randomly set. The two-dimensional array element domain frequency domain data and the corresponding spatial energy spectrum of random target number, azimuth and signal-to-noise ratio are generated by the passive broadband warning model. Step 1.4: Perform a dimensional expansion and stitching operation on the two-dimensional array element domain frequency domain data obtained in Step 1.3 to obtain three-dimensional array element domain frequency domain data; Step 1.5: Perform amplitude enhancement operation on the spatial energy spectrum obtained in Step 1.3 to obtain weak target enhancement labels.

3. The passive weak target detection method based on an attention mechanism network as described in claim 2, characterized in that, Step 1.4 includes the following steps: Step 1.4.1, process the two-dimensional array element 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 ; Step 1.4.2: Expand each two-dimensional matrix by one dimension in the same dimension to obtain a three-dimensional matrix x. f r eal_M x f，imag_M x f，mod_M And by performing a stitching operation on the extended dimension, we obtain the three-dimensional array element domain frequency domain data x. Merge ; Step 1.4.3, process the three-dimensional array element domain frequency domain data x Merge Normalization is performed to obtain the normalized three-dimensional array element domain frequency domain data x. Sample .

4. The passive weak target detection method based on an attention mechanism network as described in claim 2, characterized in that, Step 1.5 includes the following steps: Step 1.5.1: Normalize the spatial energy spectrum y to obtain the normalized spatial energy spectrum y1; Step 1.5.2: Reduce the amplitude of the normalized spatial energy spectrum y1 by 50% to obtain the reduced spatial energy spectrum y2; Step 1.5.3: Based on the set number of targets and their azimuths, calculate the beam number n corresponding to each target. beam And enhance the magnitude of each target y2(n) beam )=y2(n beam )+N enhance N enhance To determine the magnitude of target enhancement, we obtain the weak target enhancement label y. Lable =y2.

5. The passive weak target detection method based on an attention mechanism network as described in claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Divide the weak target enhancement labels into targetless sequences L. Non and target sequence L Target ; Step 3.2: Set the target discrimination threshold T, and set the target sequence L. Target Divided into strong target sequences L Target，1 and weak target sequence L Target，2 ; Step 3.3: Based on the position of each target sequence in the weak target enhancement label, obtain the corresponding three sub-sequences B from the predicted weak target inference results. Non B Target，1 B Target，2 ; Step 3.4, construct the expression for the weak target weighted augmentation loss function as follows: Where N1, N2, and N3 are respectively L Non L Target，1 L Target，2 The sequence length is α1, α2, and α3, which are the corresponding weak target weighting coefficients.

6. The passive weak target detection method based on an attention mechanism network as described in claim 5, characterized in that, Step 4 includes the following steps: Step 4.1: The model is trained quickly using the SGD optimizer. During training, the α1, α2, and α3 of the weak target weighted augmentation loss function are set to 1, 1.5, and 2, respectively. Step 4.2: Based on the Adam optimizer, perform exact fitting training on the model obtained in Step 4.

1. During training, the α1, α2, and α3 of the weak target weighted augmentation loss function are set to 1, 2, and 3, respectively. Step 4.3: Perform progressive fitting training on the model obtained in Step 4.2 based on the Adam optimizer. During training, the α1, α2, and α3 of the weak target weighted enhancement loss function are set to 1, 2, and 5, respectively.

7. The passive weak target detection method based on an attention mechanism network as described in claim 4, characterized in that, Step 6 includes the following steps: Step 6.1, set the amplitude reduction dual threshold to N. decode1 N decode2 And N enhance >N decode1 >N decode2 ; Step 6.2, iterate through the magnitudes R(i) of the predicted weak target inference results, i = 1, 2, ..., N Beam Determine if R(i) > N decode2 If so, then R(i) = R(i) - N decode2 Otherwise, it remains unchanged; Step 6.3: Iterate through the magnitude R(i) of the weak target enhancement labels obtained in Step 6.2, i = 1, 2, ..., N Beam And determine R(i) > N decode1 If so, then R(i) = R(i) - N decode1 Otherwise, it remains unchanged; Step 6.4: Increase the amplitude of the weak target enhancement label obtained in step 6.3 by 100% to obtain the corresponding spatial energy spectrum.

Images