Target depth attribute discrimination method based on lightweight residual neural network
Through the sound field elevation angle structure feature discrimination method based on lightweight residual neural network, the problem of traditional surface/underwater target classifiers' strong dependence on prior information is solved, and efficient and robust target depth attribute discrimination is achieved in complex underwater environments.
Patent Information
- Application Number
- CN202510924519.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-04
- Publication Date
- 2025-10-17
AI Technical Summary
Traditional surface/underwater target classifiers rely heavily on prior target information, have poor robustness, and are difficult to classify effectively in complex underwater environments.
A target depth attribute discrimination method based on a lightweight residual neural network is adopted. By generating a sound field elevation structure dataset and constructing a neural network consisting of multiple residual blocks for training, the target depth attribute is discriminated using the sound field elevation structure features.
The classification performance of surface and underwater targets is improved, the dependence on prior information is reduced, the robustness in the case of environmental mismatch is improved, and the computational complexity is reduced.
Smart Images
Figure CN120804684A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of underwater acoustic target depth classification, and particularly relates to a target depth attribute discrimination method based on a light residual neural network. BACKGROUND
[0002] In underwater counterwork, quickly and accurately discriminating the depth attribute of an underwater acoustic target, i.e. whether it is an underwater target or a surface target, is directly related to the timeliness of combat command and the effectiveness of safety defense. Therefore, target depth attribute discrimination is crucial to mastering the initiative on the battlefield. Two key technical links for realizing target depth attribute discrimination include target depth correlation feature extraction and underwater / surface target classifier design. In a shallow sea negative jump layer waveguide, the sound field elevation angle structure is a feature with target depth correlation, and therefore is often used for target depth attribute discrimination.
[0003] However, due to the complexity of the underwater environment, the diversity of targets and the variability of characteristics, the design of a surface / underwater target classifier has great difficulty. The commonly used classification methods are mostly based on the idea of a binary detector to construct a binary classifier, and discriminate by comparing the relationship between the feature quantity and the preset threshold value. However, such methods generally have a strong dependence on prior information of the target, and once the classification feature quantity construction or threshold selection does not match the real environment, the classification performance will decrease significantly, and the robustness is poor.
[0004] In recent years, with the development of deep learning technology, it has gradually become a research hotspot to apply its excellent feature learning ability to solve the surface / underwater target classification problem. Therefore, in order to solve the problem that the binary classifier has a strong dependence on prior information of the target and poor robustness, it is necessary to propose a surface / underwater target classification method based on deep learning technology. SUMMARY
[0005] The purpose of the present application is to solve the problem of strong dependence on prior information of the target and poor robustness of the traditional binary classifier, and a target depth attribute discrimination method based on a light residual neural network is proposed.
[0006] The technical scheme adopted by the present application to solve the above technical problems is: a target depth attribute discrimination method based on a light residual neural network, which specifically comprises the following steps:
[0007] Step one, under various hydrological conditions and various spatial positions of the target, generate sound pressure data received by the vertical array through numerical simulation;
[0008] Then, respectively process each group of sound pressure data received by the vertical array obtained by numerical simulation in the modal domain beam forming to obtain the sound field elevation angle structure data set;
[0009] Step two, constructing a lightweight residual neural network including a first residual block, a second residual block, a third residual block and a fourth residual block;
[0010] The first residual block includes a first convolutional layer, a first batch normalization layer, a first activation layer, a second convolutional layer, a second batch normalization layer and a second activation layer; and each residual block has the same structure;
[0011] The initial parameters in the lightweight residual neural network are obtained by random initialization or Kaiming initialization method;
[0012] Step three, dividing the sound field elevation structure data set obtained in step one into a training set and a validation set, training the lightweight residual neural network using the training set until the cross-entropy loss function of the lightweight residual neural network converges on the validation set, and stopping training to obtain the trained lightweight residual neural network;
[0013] Step four, receiving signals of a target to be classified by a vertical array, performing modal domain beamforming based on the received signals to obtain sound field elevation structure data corresponding to the target to be classified;
[0014] Step five, taking the sound field elevation structure data corresponding to the target to be classified as the input of the trained lightweight residual neural network, and outputting the depth attribute discrimination result of the target to be classified by the trained lightweight residual neural network.
[0015] Further, the hydrological condition includes a negative thermocline waveguide sound speed profile and a seabed bottom.
[0016] Further, the sound pressure data received by the vertical array is:
[0017]
[0018] wherein, p n (r) is the sound pressure data received by the nth element of the vertical array;
[0019] j is the imaginary unit;
[0020] z s is the target depth;
[0021] ρ(z s ) is the water density at the target depth z s ;
[0022] u m (·) is the eigenfunction of the mth mode;
[0023] k rm is the horizontal wave number of the mth mode;
[0024] αm the attenuation coefficient of the mth order mode;
[0025] z0 is the depth of the 0th element;
[0026] r is the horizontal distance between the target and the vertical array;
[0027] M represents the total order number of the normal mode.
[0028] Further, the vertical array received sound pressure data obtained by simulating each group of numerical values is subjected to modal domain beamforming processing to obtain a sound field elevation structure data set; the specific process is as follows:
[0029] Step one, for any one group of numerical simulation obtained vertical array received sound pressure data p n (r), n = 0, 1, …, N-1:
[0030] The 0th element at depth z0 is taken as the reference element, and the sound pressure data p n (r), n = 1, 2, …, N is subjected to modal domain beamforming processing to obtain the sound field elevation structure data B p (cosφ l ):
[0031]
[0032] Wherein, N is the number of elements in the vertical array;
[0033] Wave number k = 2πf / c, c is the sound speed, and f is the frequency;
[0034] φ l is the lth scanning angle of the array, l = 1, 2, …, L, and L is the total number of array scanning angles;
[0035] Step two, B p (cosφ l ) is recorded as x l , then the sound field elevation structure data corresponding to each scanning angle forms a vector x with a length of L:
[0036] x = [x1, x2, …, x l , …, x L ], l = 1, 2, …, L
[0037] The vector x is subjected to normalization processing to obtain the normalized sample x of the sound field elevation structure:
[0038]
[0039] Wherein, min(x) is the minimum element value in the vector x, and max(x) is the maximum element value in the vector x, for x l corresponding normalized value;
[0040] using l = 1, 2,..., L, the normalized sample is composed R is a real number, and as a sample in the sound field elevation structure data set;
[0041] Step III, using the method of step I and step II, the vertical array received sound pressure data obtained by traversing each group of numerical simulation is obtained, and the sound field elevation structure data set is obtained.
[0042] Further, the working process of the lightweight residual neural network is:
[0043] Step II, the sample is taken as the input of the lightweight residual neural network, and in the lightweight residual neural network, the sample passes through a third convolutional layer, a third batch normalization layer, a third activation layer and a maximum pooling layer in sequence;
[0044]
[0045] wherein, * represents a convolution operation; W0 represents a one-dimensional convolution kernel of the third convolutional layer; BN3(·) represents the third batch normalization layer; ReLU3(·) represents the third activation layer; and MP(·) represents the maximum pooling layer; represents the output of the maximum pooling layer;
[0046] Step II, the sample is taken as the input of the first residual block, and in the first residual block, passes through a first convolutional layer, a first batch normalization layer, a first activation layer, a second convolutional layer and a second batch normalization layer in sequence to obtain the output of the second batch normalization layer
[0047]
[0048] wherein, BN1(·) is the first batch normalization layer, ReLU1(·) is the first activation layer, BN2(·) is the second batch normalization layer, W1 is a one-dimensional convolution kernel of the first convolutional layer, and W2 is a one-dimensional convolution kernel of the second convolutional layer;
[0049] the output of the second batch normalization layer is element-wise added to , and the element-wise addition result is taken as the output of the first residual block:
[0050]
[0051] wherein z is an output of the first residual block;
[0052] Step two three, taking the output of the first residual block as an input of a second residual block, taking the output of the second residual block as an input of a third residual block, and taking the output of the third residual block as an input of a fourth residual block;
[0053] Taking the output of the fourth residual block as an input of a global average pooling layer, taking the output of the global average pooling layer as an input of a full connection layer, and taking the output of the full connection layer as an input of a Softmax layer, the probability P0 that the target is a water surface target and the probability P1 that the target is an underwater target are calculated through the Softmax layer;
[0054] The decision criterion is defined as follows:
[0055] H0: P0≥P1
[0056] H1: P0<P1
[0057] Wherein H0 represents a water surface target, and H1 represents an underwater target.
[0058] Further, the specific process of the step four is as follows:
[0059] Step four one, taking the received signal y(z n ,t) of the nth element in the vertical array as:
[0060] y(z n ,t)=s(t)*p(z n ,t)+n(t)
[0061] Wherein s(t) represents target radiation noise;
[0062] z n represents the depth of the nth element in the vertical array;
[0063] p(z n ,t) represents the channel impulse response at the element with the depth z n ;
[0064] n(t) is the received noise;
[0065] The Fourier transform of the received signal y(z n ,t) is performed to obtain the Fourier transform result y n (r,f) of y(z n ,t); after traversing the received signals of all elements in the vertical array, the Fourier transform results corresponding to each element are obtained respectively;
[0066] Step four two, modal domain beamforming is performed on the Fourier transform results corresponding to each array element to obtain the sound field elevation structure data corresponding to the target to be classified.
[0067] The beneficial effects of the present application are:
[0068] The present application designs a target depth attribute discrimination method based on a lightweight residual neural network, and the lightweight residual neural network can deeply mine the multi-level features of the sound field elevation structure data through the residual blocks stacked layer by layer, so as to effectively improve the classification performance of the water surface target and the underwater target, and the classification process does not need to rely on the prior information of the target, thereby solving the problem of high dependence on prior information of the traditional binary detector method, and the method of the present application can realize the target depth attribute discrimination based on the sound field elevation structure features under the condition of the vertical array, has robustness to environmental mismatch, and the lightweight residual neural network adopted has low computational complexity. BRIEF DESCRIPTION OF DRAWINGS
[0069] Figure 1 is a flowchart of a target depth attribute discrimination method based on a lightweight residual neural network of the present application;
[0070] Figure 2 is a structure diagram of the lightweight residual neural network based on the sound field elevation structure;
[0071] The left side of the dashed line is the entire lightweight residual neural network, and the right side of the dashed line is a structure diagram of the residual block;
[0072] Figure 3 is a sound velocity profile used when constructing a data set;
[0073] Figure 4 is a loss function in the training stage and the verification stage;
[0074] Figure 5 is the classification accuracy of the lightweight residual neural network and the comparative method for the underwater target;
[0075] Figure 6 is the classification accuracy of the lightweight residual neural network and the comparative method for the water surface target. DETAILED DESCRIPTION
[0076] Detailed implementation one: combined with Figure 1 The present application discloses a target depth attribute discrimination method based on a lightweight residual neural network, which specifically comprises the following steps:
[0077] Step one, under various hydrological conditions and various spatial positions of the target, sound pressure data received by the vertical array is generated through numerical simulation;
[0078] The vertical array received sound pressure data obtained by simulating each group of values is subjected to modal domain beamforming processing to obtain a sound field elevation structure data set;
[0079] Step two, constructing a lightweight residual neural network comprising a first residual block, a second residual block, a third residual block and a fourth residual block;
[0080] The first residual block comprises a first convolutional layer (Conv), a first batch normalization layer (BN), a first activation layer (ReLU), a second convolutional layer, a second batch normalization layer and a second activation layer, and the structure of each residual block is the same;
[0081] The initial parameters in the lightweight residual neural network are obtained by random initialization or Kaiming initialization method;
[0082] Step three, dividing the sound field elevation structure data set obtained in step one into a training set and a validation set, training the lightweight residual neural network using the training set until the cross-entropy loss function of the lightweight residual neural network converges on the validation set, and stopping the training to obtain the trained lightweight residual neural network (retaining the model parameters when the loss function converges on the validation set);
[0083] Step four, using the vertical array to receive the signal of the target to be classified, performing modal domain beamforming based on the received signal to obtain the sound field elevation structure data corresponding to the target to be classified;
[0084] Step five, taking the sound field elevation structure data corresponding to the target to be classified as the input of the trained lightweight residual neural network, outputting the depth attribute discrimination result of the target to be classified through the trained lightweight residual neural network, that is, giving the judgment result of whether the target belongs to a water surface target or an underwater target.
[0085] The lightweight residual neural network of the present application can effectively alleviate the gradient vanishing problem that easily occurs in the training process of a deep neural network by introducing a residual connection mechanism, and can deeply mine the potential deep feature differences in the sound field elevation structure, thereby improving the adaptability and classification accuracy of the model to environmental changes.
[0086] Specific implementation method two: different from the specific implementation method one, the hydrological condition comprises a negative thermocline waveguide sound speed profile and a seabed bottom.
[0087] The other steps and parameters are the same as those in the specific implementation method one.
[0088] The present application obtains a plurality of groups of numerical simulation conditions by combining different kinds of negative jump layer waveguide sound speed profiles and different kinds of seabed substrates, and combining different target depths and distances, each group of numerical simulation conditions including a negative jump layer waveguide sound speed profile, a seabed substrate, a target depth and a depth distance, and at least one factor being different in any two groups of numerical simulation conditions. Numerical simulation is respectively performed according to each group of numerical simulation conditions, and sound pressure data received by a vertical array under each group of numerical simulation conditions is respectively obtained, the vertical array being a uniform linear array perpendicular to the water surface.
[0089] Specific implementation mode three: the sound pressure data received by the vertical array in the present implementation mode is different from that in specific implementation mode one or two, and is:
[0090]
[0091] wherein, p n (r) is the sound pressure data received by the nth array element in the vertical array;
[0092] j is an imaginary unit;
[0093] z s is the target depth;
[0094] ρ(z s ) is the water density at the target depth z s ;
[0095] u m (·) is the eigenfunction of the mth mode;
[0096] k rm is the horizontal wave number of the mth mode;
[0097] α m is the attenuation coefficient of the mth mode;
[0098] z0 is the depth of the 0th array element (i.e. the depth of the array element with the shallowest depth);
[0099] r is the horizontal distance between the target and the vertical array;
[0100] M represents the total order number of the normal modes.
[0101] The other steps and parameters are the same as those in specific implementation mode one or two.
[0102] The sound pressure data in the present implementation mode is calculated by the acoustic field toolbox.
[0103] Specific implementation mode four: the present implementation mode is different from any one of specific implementation modes one to three in that the sound pressure data received by the vertical array obtained by each group of numerical simulation is subjected to a modal domain beamforming processing to obtain a sound field elevation angle structure data set, and the specific process is as follows:
[0104] Step 1: For any set of vertical array received sound pressure data p obtained by numerical simulation n (r),n=0,1,…,N-1:
[0105] Taking the array element 0 at the depth z0 as the reference array element, the sound pressure data p n (r),n=1,2,…,N perform modal domain beamforming processing to obtain the sound field elevation structure data B p (cosφ l ):
[0106]
[0107] Where N is the number of elements in the vertical array;
[0108] Wave number k = 2πf / c, c is the speed of sound, f is the frequency;
[0109] φ l is the l-th scanning angle of the array, l=1,2,…,L, L is the total number of array scanning angles;
[0110] Step 1 and 2: B p (cosφ l ) is denoted as x l , then the sound field elevation angle structure data corresponding to each scanning angle forms a vector x with a length of L:
[0111] x=[x1,x2,...,x l ,...,x L ],l=1,2,...,L
[0112] Then normalize the vector x to obtain the normalized sample of the sound field elevation structure
[0113]
[0114] Among them, min(x) is the minimum element value in vector x, max(x) is the maximum element value in vector x, is x l The corresponding normalized value;
[0115] use l=1,2,...,L constitutes the normalized sample R is a real number, As a sample in the sound field elevation structure dataset;
[0116] Step III, traverse to each group of numerical simulation to obtain vertical array receiving sound pressure data, and obtain sound field elevation angle structure data set by using the method of step I-I and step I-II.
[0117] Other steps and parameters are the same as one of the first to third embodiments.
[0118] Embodiment V: in combination Figure 2 This embodiment is described. The working process of the lightweight residual neural network is different from one of the first to fourth embodiments, which is:
[0119] Step II-I, the sample As the input of the lightweight residual neural network, the sample passes through a third convolutional layer, a third batch normalization layer, a third activation layer and a maximum pooling layer (MP) in sequence;
[0120]
[0121] Wherein, * represents convolution operation; W0 represents one-dimensional convolution kernel of the third convolutional layer; BN3(·) represents the third batch normalization layer; ReLU3(·) represents the third activation layer; MP(·) represents the maximum pooling layer; represents the output of the maximum pooling layer;
[0122] It should be noted that the weights, biases and scaling factors of the batch normalization layer and other parameters of the convolutional layer are initialized by random initialization or Kaiming initialization, and the parameters are optimized by using Adam optimizer and cross-entropy loss function for back propagation; In order to promote effective learning, dynamic learning rate adjustment is implemented in the whole training process, and early stopping strategy is adopted to avoid overfitting.
[0123] Step II-II, the sample As the input of the first residual block, the sample passes through a first convolutional layer, a first batch normalization layer, a first activation layer, a second convolutional layer and a second batch normalization layer in sequence to obtain the output of the second batch normalization layer
[0124]
[0125] Wherein, BN1(·) is the first batch normalization layer, ReLU1(·) is the first activation layer, BN2(·) is the second batch normalization layer, W1 is one-dimensional convolution kernel of the first convolutional layer, and W2 is one-dimensional convolution kernel of the second convolutional layer;
[0126] The output of the second batch normalization layer is combined with The element-wise addition result is taken as the output of the first residual block:
[0127]
[0128] wherein z is the output of the first residual block;
[0129] Step two three, the output of the first residual block is taken as the input of the second residual block, the output of the second residual block is taken as the input of the third residual block, and the output of the third residual block is taken as the input of the fourth residual block;
[0130] It should be noted that the working processes of the second residual block, the third residual block and the fourth residual block are the same as that of the first residual block, for example, for the second residual block, the output of the second BN layer in the second residual block is element-wise added with the input of the second residual block, and the element-wise addition result is taken as the output of the second residual block;
[0131] The output of the fourth residual block is taken as the input of the global average pooling layer, the output of the global average pooling layer is taken as the input of the full connection layer, and the output of the full connection layer is taken as the input of the Softmax layer, and the probability P0 that the target is a water surface target and the probability P1 that the target is an underwater target are calculated through the Softmax layer;
[0132] And the hypothesis testing method is used to define the decision criterion as follows:
[0133] H0: P0≥P1
[0134] H1: P0<P1
[0135] Wherein H0 represents the decision result of the water surface target, and H1 represents the decision result of the underwater target.
[0136] The other steps and parameters are the same as those in one of the first four embodiments.
[0137] The input part of step two one adopts the third convolutional layer to preliminarily extract features from the input sample, and then the preliminarily extracted features further extract key information in the feature map through each residual block, and then the key information in the extracted feature map is processed by the classification decision part including the global average pooling layer and the full connection layer, that is, the output of the last residual block is compressed into a feature vector through the global average pooling layer, and the compressed feature vector contains the target depth correlation features learned by the network, so the compressed feature vector is mapped to the class space through the full connection layer, and the result of the full connection layer output is probabilized through the Softmax layer, that is, the probabilities that the target is a water surface target and an underwater target are obtained respectively, if P0≥P1, the decision result is H0, and if P0<P1, the decision result is H1.
[0138] The related parameters of the lightweight residual neural network are shown in Table 1:
[0139] Table 1 Related parameters of the lightweight residual neural network
[0140]
[0141]
[0142] Specific implementation six: Different from one of the specific implementations one to five, the specific process of the step four is:
[0143] Step four one, the received signal y(z n ,t) of the nth element in the vertical array is expressed as:
[0144] y(z n ,t) = s(t) * p(z n ,t) + n(t)
[0145] Wherein, s(t) represents the target radiation noise;
[0146] z n represents the depth of the nth element in the vertical array;
[0147] p(z n ,t) represents the channel impulse response at the element with depth z n ;
[0148] n(t) is the received noise;
[0149] Then, the Fourier transform of the received signal y(z n ,t) is performed to obtain the Fourier transform result y n (r, f) of y(z n ,t); After traversing the received signals of all elements in the vertical array, the Fourier transform results corresponding to each element are obtained respectively;
[0150] Step four two, modal domain beamforming is performed on the Fourier transform results corresponding to each element to obtain the sound field elevation angle structure data corresponding to the target to be classified.
[0151] The other steps and parameters are the same as one of the specific implementations one to five.
[0152] The specific process of step four two is: selecting a frequency point to be analyzed, performing modal domain beamforming on the Fourier transform results corresponding to each element at the frequency point to be analyzed to obtain the sound field elevation angle structure:
[0153]
[0154] Wherein, Y n (r) is yn (r,f) the Fourier transform result of the frequency point f0 to be analyzed; φ l is the lth array scanning angle; B(cosφ l ) is the lth sound field elevation structure; k' is the wave number, k'=2πf0 / c; c is the average sound speed within the vertical array coverage depth; Δz is the distance between adjacent elements; z0 is the depth of the 0th element;
[0155] Traverse L array scanning angles, and then normalize the sound field elevation structures corresponding to the L array scanning angles. A sound field elevation structure normalized sample is formed by using the normalized sound field elevation structures, and the normalized sample is the sound field elevation structure data corresponding to the target to be classified.
[0156] The effectiveness and feasibility of the method are verified by experiments as follows:
[0157] Step 1: Generate sound pressure data of various hydrological conditions and target spatial positions through a simulation experiment based on a vertical array, and perform modal domain beamforming processing on the sound pressure data to obtain a sound field elevation structure data set for training a model. The related parameters are as follows:
[0158] The number of elements of the vertical array anchored to the water bottom is 6, the interval between adjacent elements is 5 m, the corresponding coverage depth is 187 m to 212 m, and the sea depth accounts for about 12%. The array scanning angle is set to a value range of (0, π). The target distance is 5 km.
[0159] Seven kinds of sound speed profiles are set as shown in Table 1, SSP#0 to SSP#6. Figure 3 Two kinds of seabed conditions are set as shown in Table 2, seabed#1 is a three-layer seabed, and seabed#2 is a half-space seabed.
[0160] Table 2 Related parameters of seabed conditions used when constructing the data set
[0161]
[0162] The target to be classified is a water surface target and an underwater target, and the depths of the two types of targets are randomly selected in the ranges of 1 m to 15 m and 16 m to 216 m, respectively.
[0163] Based on the above environmental parameters and sound source parameters, a training set and a validation set are generated, and the detailed parameters of the training set and the validation set are shown in Table 3:
[0164] Table 3 Related parameters for constructing the training set and the validation set
[0165]
[0166]
[0167]
[0168] Note: The representation such as "11:2:19" in the table is generally used to describe the value range and step of the parameter, which means that the signal-to-noise ratio is between 11 dB and 19 dB, with a step of 2 dB.
[0169] Step 2, construct a lightweight residual neural network based on the sound field elevation structure, and the structural diagram is as shown in Figure 2 The related parameters are shown in Table 1.
[0170] Step 3, based on the sound field elevation structure data sample in step 1, train and verify the lightweight residual neural network using the training set and the verification set, and the loss function in the training phase and the verification phase is as shown in Figure 4 The loss function in the training phase and the verification phase is as shown in
[0171] It can be seen that the loss functions in the training phase and the verification phase gradually converge with the increase of the training rounds, and the verification loss and the training loss are close, which indicates that under the current training sample amount, the lightweight residual neural network does not appear obvious overfitting phenomenon in the training process.
[0172] A test set composed of sound field elevation structure data samples is constructed, and the training set and the test set are conditionally mismatched. The trained lightweight residual neural network in step 3 is used to classify the sound field elevation structure data samples in the test set obtained under the environmental mismatch condition, and the judgment result of the water surface target or the underwater target is given, the classification performance of the trained lightweight residual neural network under the environmental mismatch condition is evaluated, and the TFMR method is compared and analyzed, wherein the TFMR method is based on the sound field elevation structure feature, and uses a binary classifier to classify the water surface / underwater target. Figure 5 And Figure 6 The classification accuracy of the lightweight residual neural network and the TFMR method under the environmental mismatch condition is given.
[0173] It can be seen that when the training set and the test set are conditionally mismatched, the lightweight residual neural network still exhibits robust classification ability, and the classification accuracy of the underwater target and the water surface target is 0.87 and 0.83 respectively, which is significantly higher than the classification accuracy of the TFMR method. This is mainly due to the training of the lightweight residual neural network through the diversified data set, which can excavate more discriminative features, so as to still maintain strong robustness when the environmental parameters change, proving the effectiveness of the method.
[0174] Step 4, use the vertical array to receive the signal of the target to be classified, perform mode domain beamforming based on the received signal, and obtain the sound field elevation structure data corresponding to the target to be classified;
[0175] Step 5, taking the sound field elevation structure data corresponding to the target to be classified as the input of the trained lightweight residual neural network, outputting the deep attribute discrimination result of the target to be classified through the trained lightweight residual neural network, that is, giving the judgment result of whether the target belongs to a water surface target or an underwater target.
[0176] The above examples of the present application are only used to illustrate the calculation model and calculation process of the present application, and are not intended to limit the embodiments of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art, and it is impossible to enumerate all the embodiments here. Any obvious changes or variations derived from the technical solutions of the present application still fall within the protection scope of the present application.
Claims
1. A method for distinguishing target depth attributes based on a lightweight residual neural network, characterized in that: The method specifically comprises the following steps: Step 1: Generate sound pressure data received by the vertical array through numerical simulation under various hydrological conditions and various spatial positions of the target; Then, modal domain beamforming processing is performed on each set of vertical array received sound pressure data obtained by numerical simulation to obtain the sound field elevation angle structure data set; Step 2: construct a lightweight residual neural network including a first residual block, a second residual block, a third residual block and a fourth residual block; The first residual block includes a first convolutional layer, a first batch normalization layer, a first activation layer, a second convolutional layer, a second batch normalization layer, and a second activation layer; and the structure of each residual block is the same; The initial parameters in the lightweight residual neural network are obtained by random initialization or Kaiming initialization; Step 3: Divide the sound field elevation structure dataset obtained in step 1 into a training set and a validation set, and use the training set to train the lightweight residual neural network until the cross entropy loss function of the lightweight residual neural network on the validation set converges, thereby obtaining a trained lightweight residual neural network. Step 4: Receive the signal of the target to be classified using a vertical array, perform modal domain beamforming based on the received signal, and obtain the sound field elevation structure data corresponding to the target to be classified; Step 5: The sound field elevation structure data corresponding to the target to be classified is used as the input of the trained lightweight residual neural network, and the trained lightweight residual neural network outputs the depth attribute discrimination result of the target to be classified.
2. The method for distinguishing target depth attributes based on a lightweight residual neural network according to claim 1, characterized in that: The hydrological conditions include the negative jump layer waveguide sound velocity profile and the seabed bottom.
3. The method for distinguishing target depth attributes based on a lightweight residual neural network according to claim 2, characterized in that: The sound pressure data received by the vertical array is: Among them, p n (r) is the sound pressure data received by the nth array element in the vertical array; j is the imaginary unit; z s is the target depth; ρ(z s ) is the target depth z s Water density at u m (·) is the eigenfunction of the mth mode; k rm is the horizontal wave number of the mth mode; α m is the attenuation coefficient of the m-th mode; z0 is the depth of array element 0; r is the horizontal distance between the target and the vertical array; M represents the total order of the simple normal wave.
4. The method for distinguishing target depth attributes based on a lightweight residual neural network according to claim 3, characterized in that: The modal domain beamforming process is performed on each set of vertical array received sound pressure data obtained by numerical simulation to obtain a sound field elevation angle structure data set. The specific process is as follows: Step 1: For any set of vertical array received sound pressure data p obtained by numerical simulation n (r),n=0,1,…,N-1: Taking the array element 0 at the depth z0 as the reference array element, the sound pressure data p n (r),n=1,2,…,N perform modal domain beamforming processing to obtain the sound field elevation structure data B p (cosφ l ): Where N is the number of elements in the vertical array; Wave number k = 2πf / c, c is the speed of sound, f is the frequency; φ l is the l-th scanning angle of the array, l=1,2,…,L, L is the total number of array scanning angles; Step 1 and 2: B p (cosφ l ) is denoted as x l , then the sound field elevation angle structure data corresponding to each scanning angle forms a vector x with a length of L: x=[x1,x2,…,x l ,…,x L ],l=1,2,…,L Then normalize the vector x to obtain the normalized sample of the sound field elevation structure Among them, min(x) is the minimum element value in vector x, max(x) is the maximum element value in vector x, is x l The corresponding normalized value; use l=1,2,…,L constitutes normalized samples R is a real number, As a sample in the sound field elevation structure dataset; Step 13: Use the method of step 11 and step 12 to traverse each group of vertical array received sound pressure data obtained by numerical simulation to obtain a sound field elevation angle structure data set.
5. The method for distinguishing target depth attributes based on a lightweight residual neural network according to claim 4, characterized in that: The working process of the lightweight residual neural network is as follows: Step 21: Sample As the input of the lightweight residual neural network, in the lightweight residual neural network, the sample It passes through the third convolution layer, the third batch normalization layer, the third activation layer and the maximum pooling layer in sequence; Where * represents the convolution operation; W0 represents the one-dimensional convolution kernel of the third convolutional layer; BN3(·) represents the third batch normalization layer; ReLU3(·) represents the third activation layer; MP(·) represents the maximum pooling layer; represents the output of the maximum pooling layer; Step 22: As the input of the first residual block, within the first residual block, After the first convolutional layer, the first batch normalization layer, the first activation layer, the second convolutional layer and the second batch normalization layer, the output of the second batch normalization layer is obtained. Among them, BN1(·) is the first batch normalization layer, ReLU1(·) is the first activation layer, BN2(·) is the second batch normalization layer, W1 is the one-dimensional convolution kernel of the first convolution layer, and W2 is the one-dimensional convolution kernel of the second convolution layer; Will and Perform element-by-element addition and use the result of element-by-element addition as the output of the first residual block: Where z is the output of the first residual block; Step 23: use the output of the first residual block as the input of the second residual block, then use the output of the second residual block as the input of the third residual block, and then use the output of the third residual block as the input of the fourth residual block; The output of the fourth residual block is used as the input of the global average pooling layer, the output of the global average pooling layer is used as the input of the fully connected layer, and the output of the fully connected layer is used as the input of the Softmax layer. The Softmax layer calculates the probability P0 that the target is a surface target and the probability P1 that the target is an underwater target; The decision criteria are defined as follows: H0: P0 ≥ P1 H1: P0<P1 Among them, H0 means the judgment result is a surface target, and H1 means the judgment result is an underwater target.
6. The method for distinguishing target depth attributes based on a lightweight residual neural network according to claim 5, characterized in that: The specific process of step 4 is as follows: Step 4.
1. Transform the received signal y(z n ,t) is expressed as: y(z n ,t)=s(t)*p(z n ,t)+n(t) Where s(t) represents the target radiation noise; z n Indicates the depth of the nth element in the vertical array; p(z n ,t) indicates the depth is z n The channel impulse response at the array element; n(t) is the received noise; Then the received signal y(z n ,t) is Fourier transformed to obtain y(z n ,t)’s Fourier transform result y n (r,f); after traversing the received signals of each array element in the vertical array, the Fourier transform results corresponding to each array element are obtained respectively; Step 42: Perform modal domain beamforming on the Fourier transform results corresponding to each array element to obtain the sound field elevation structure data corresponding to the target to be classified.