A vertical array pulse signal ranging and sounding method based on a deep learning network

Abstract

A vertical array pulse signal ranging and depth measurement method based on deep learning networks belongs to the field of underwater information fusion technology. The method utilizes marine environmental parameters to simulate and calculate the arrival delay and amplitude of direct waves and sea surface reflected waves at different grids, generating received signals and adding noise to obtain a frequency-elevation map. Matched filtering and sound path transformation are applied to the frequency-elevation map to obtain a sound path-elevation map, which is then cropped based on the location of bright spots. A network is constructed consisting of a decoupled strip attention feature extraction module, a two-layer MLP physical coordinate mapping module, a homoscedastic uncertainty multi-task regression module, and an unsupervised domain adversarial module. The image and coordinates are input into the network, and the distance and depth are output. This method aligns the feature distribution of simulated and measured data through domain adversarial training, uses strip attention to separate the distance and depth coupling features, and dynamically balances multi-task learning with weights to enhance multipath features under low signal-to-noise ratio conditions, thereby improving the accuracy and generalization ability of positioning.

Description

Technical Field

[0001] This invention belongs to the field of underwater information fusion technology, specifically relating to a method for ranging and depth measurement using vertical array pulse signals based on deep learning networks. Background Technology

[0002] Matched field processing is a common method for underwater acoustic positioning. However, due to the time-varying and spatially varying nature of marine environmental parameters, the replicated field in matched field processing often mismatches with the actual sound field, leading to decreased positioning accuracy. In recent years, with the development of artificial intelligence, data-driven models, through powerful feature extraction capabilities, can more accurately learn the relationship between data and the sound field, thereby matching the sound field and improving positioning accuracy. Active sonar often uses deterministic pulse signals for target detection. Vertical arrays can obtain target range and elevation angle information. This range and elevation angle information can be input into a network for learning to estimate horizontal range and depth.

[0003] Existing ranging and depth sounding techniques based on matching multipath arrival features can be broadly categorized into two types depending on the receiver configuration:

[0004] 1. In a single hydrophone environment, multipath delay and three-dimensional positioning in complex terrain are used to achieve sound source localization using only a single hydrophone by matching the measured arrival pattern with a model copy. This method is easy to deploy, but the attenuation of seabed reflected waves is large. When the signal-to-noise ratio of the received signal is low, the seabed reflected waves may be submerged by noise and the information cannot be extracted.

[0005] In the deep-sea shadow region, direct waves cannot reach the sound field, which is mainly formed by interference from multiple paths, such as a single seabed reflection. The received sound intensity exhibits two different interference periods with frequency. By extracting these two interference periods with a single hydrophone and matching them with the model, passive ranging and depth determination can be achieved without a broadband guiding source and a large copy field calculation. This method extends passive positioning with a single hydrophone from the direct sound region to the deep-sea shadow region. The disadvantage is that it requires a broadband signal with a high signal-to-noise ratio to ensure that observable interference fringes and periods can be formed.

[0006] 2. In a vertical array environment, the Sparse Bayesian Learning (SBL) algorithm is used to extract the angle of arrival (AOA) of direct waves and sea surface reflected waves with high resolution. The measured AOA is then matched with a pre-calculated AOA model dictionary to achieve distance and depth localization of the sound source. This method has the ability to extract AOA with ultra-high resolution, but the computational complexity is very high, and there is a computational bottleneck for systems with a large number of array elements or requiring high real-time performance. Alternatively, an adaptive iterative frequency domain deconvolution algorithm can be used to estimate the AOA with high resolution. Then, by matching the feature template of the AOA, the distance and depth of the target can be determined. This method has a significantly improved computational efficiency, but its robustness is reduced in low signal-to-noise ratio environments. Excessive iteration can easily produce false angle peaks (artifacts), and it is also sensitive to the noise covariance matrix.

[0007] Depth-distance estimation can also be achieved by combining artificial intelligence methods. Using multi-task learning convolutional neural networks trained on simulated data, an end-to-end joint estimation of the distance and depth of deep-sea broadband sound sources can be realized, improving resistance to environmental mismatch. However, a relatively high signal-to-noise ratio is required to output reliable positioning results. At extremely low signal-to-noise ratios, its maximum working distance will be shortened. Summary of the Invention

[0008] The purpose of this invention is to address the problems of data mismatch between simulation training sets and sea trial test sets, horizontal distance and depth coupling, and model performance degradation under low signal-to-noise ratio environments in existing methods. Therefore, this invention proposes a vertical array pulse signal ranging and depth measurement method based on deep learning networks, comprising the following steps:

[0009] S1. Obtain marine environmental parameters, vertical array parameters, and waveform parameters of pulse signals. Simulate and calculate the arrival delay and amplitude of direct waves and sea surface reflected waves in the preset horizontal distance-depth grid. Generate the simulated received signal and add noise. Perform a Fourier transform on the signal to obtain the frequency domain received signal. Calculate the steering vector based on the array parameters and construct a frequency-elevation diagram.

[0010] S2. According to the principle of matched filtering, the local signal is obtained by time-domain inversion of the transmitted signal. Based on the frequency-elevation diagram and the local signal, the time delay-elevation diagram is obtained by convolution. Then, the time delay axis in the time delay-elevation diagram is converted into the sound path axis by using the speed of sound to obtain the sound path-elevation diagram.

[0011] S3. Based on the coordinates of the direct wave bright spot and the coordinates of the sea surface reflected wave bright spot, the sound path-elevation angle diagram is cropped to obtain the cropped rectangular window image and its midpoint coordinates.

[0012] S4. Input the cropped rectangular window image and its midpoint coordinates into the deep learning model to obtain the horizontal distance and depth estimates of the sound source.

[0013] Furthermore, the generation of simulated received signal ,in , For the arrival time delay of direct waves and sea-emitted waves, , For the corresponding amplitude, To transmit a signal.

[0014] Furthermore, the cropping of the sound path-elevation diagram includes:

[0015] The path length extension of the cut rectangular window is at least the product of the maximum time delay difference between the direct wave and the sea surface reflected wave under the preset grid and the sampling rate, and the azimuth extension length satisfies the coverage of the bright spot and its azimuth sidelobe.

[0016] Furthermore, the deep learning model includes a decoupled strip attention feature extraction module, a two-layer MLP physical coordinate mapping module, and a homoscedastic uncertainty multi-task regression module. The decoupled strip attention feature extraction module receives the cropped rectangular window image and performs feature extraction, outputting an image feature vector. The two-layer MLP physical coordinate mapping module receives the coordinates of the midpoint of the cropped rectangular window, performs feature mapping, and outputs a coordinate feature vector. The image feature vector and the coordinate feature vector are concatenated to obtain a multimodal fusion feature vector, which is then input into the homoscedastic uncertainty multi-task regression module for distance prediction and depth prediction.

[0017] Furthermore, the decoupled strip attention feature extraction module sequentially includes an asymmetric convolution primary unit and a downsampling unit, an intermediate convolution and downsampling unit, a decoupled strip attention mechanism unit, and a deep feature mapping and convergence unit;

[0018] The decoupled strip attention mechanism unit performs adaptive average pooling and adaptive max pooling in both the height and width directions, respectively. It interacts with information through a non-linear activation function of a convolutional kernel, and then decouples the obtained features into row feature vectors and column feature vectors to obtain row attention weights and column attention weights. The row attention weights and column attention weights are then multiplied with the input feature map of the decoupled strip attention mechanism unit by a Hadamard product to obtain the enhanced output feature map.

[0019] The deep feature mapping and convergence unit obtains a one-dimensional image feature vector, i.e., the output image feature vector, based on the output of the decoupled strip attention mechanism unit.

[0020] Furthermore, the dual-layer MLP physical coordinate mapping module is a multilayer perceptron, including two sequentially connected fully connected layers, each followed by a normalization layer and a PReLU activation layer; the dual-layer MLP physical coordinate mapping module maps the center coordinates of the clipping window to high-dimensional coordinates.

[0021] Furthermore, the homoscedastic uncertainty multi-task regression module includes two parallel and independent sub-networks, which are used to perform distance prediction and depth prediction, respectively. Each sub-network contains multiple sequentially connected fully connected layers, layer normalization layers, PReLU activation function layers, and random deactivation layers. The last fully connected layer outputs the distance or depth prediction value.

[0022] Furthermore, during deep learning model training, the homoscedastic uncertainty multi-task regression module introduces two learnable noise parameters, combining the mean squared error loss of the distance task and the depth task to construct a joint multi-task loss function under maximum likelihood estimation:

[0023] ,

[0024] in and To define the mean squared error loss for distance and depth tasks, and Two learnable noise parameters are introduced.

[0025] Furthermore, the deep learning model also includes an unsupervised domain adversarial module. During model training, the multimodal fused feature vector is input to the unsupervised domain adversarial discriminant module. The unsupervised domain adversarial module includes a gradient reversal layer and a domain discriminator branch. The gradient reversal layer is concatenated between the multimodal fused feature vector and the domain discriminator branch. During forward propagation, the gradient reversal layer is propagated identically, and during backward propagation, the domain discrimination error gradient is multiplied by a negative coefficient for reversal. The domain discriminator branch is a fully connected subnetwork, with a Sigmoid activation function layer connected to its end, outputting the probability that the input feature belongs to the target domain; thus, the domain classification loss is obtained. ,based on and Adversarial training is performed using gradient inversion layers and domain discriminators to achieve feature alignment between the source and target domains.

[0026] Furthermore, the adversarial training process of the deep learning network model is a minimax game process. By multiplying the gradient of the domain discrimination error through the gradient reversal layer by a negative coefficient, the feature extractor parameters are forced to be updated in the direction of increasing the domain classification error, so as to align the probability distribution of the source domain simulation data and the target domain real data in the feature space.

[0027] The beneficial effects of this invention are:

[0028] 1. By introducing a domain adversarial training mechanism, the distribution of simulated acoustic data and real ocean measurement data is aligned in the feature space, enabling the network to extract domain-invariant features, thereby reducing positioning errors caused by environmental mismatch and improving the model's generalization ability under different sea conditions.

[0029] 2. By using decoupled strip attention to enhance the multipath structure representation along the acoustic path dimension and pitch angle dimension respectively, and combining homoscedastic uncertainty multi-task regression to dynamically balance the weights of distance and depth estimation, the coupling interference between the two is reduced and the estimation accuracy of distance and depth is improved.

[0030] 3. Feature enhancement is performed in multipath bright areas through an attention mechanism, and the path-elevation angle map formed by matched filtering is used as network input. This can suppress sidelobes and background noise under low signal-to-noise ratio, maintain the ability to extract effective multipath features, and make the localization results more stable. Attached Figure Description

[0031] Figure 1 This is a parameter table for the training and validation sets involved in the simulation experiments of this invention;

[0032] Figure 2 This refers to the sound velocity profile involved in the simulation experiment of this invention;

[0033] Figure 3 This is the sound path-pitch diagram involved in the simulation experiment of this invention;

[0034] Figure 4 This is the cut sound path-pitch angle diagram involved in the simulation experiment of this invention;

[0035] Figure 5 This is a structural diagram of the deep learning model involved in the simulation experiment of this invention;

[0036] Figure 6 This is a simulation test prediction effect diagram related to the simulation experiment of this invention. Detailed Implementation

[0037] The present invention will now be described in further detail with reference to specific embodiments.

[0038] The main steps of the method of the present invention are described below:

[0039] Step 1: In the marine environment where the application is needed, obtain the sound velocity profile, marine environment parameters, vertical array parameters, and pulse signal waveform parameters. Assume the incident signal frequency is... ,wavelength The number of frequency points is The elements of the vertical array are equally spaced, and the element spacing is... The number of array elements is The angle between the direction of arrival of the wave and the horizontal plane is defined as... The number of scanning angles is .

[0040] Acoustic software was used to simulate and calculate the arrival delay of direct waves and sea surface emitted waves for each horizontal range-depth grid under this marine environment and array parameters. , and the corresponding amplitude , , where the subscript z represents the direct wave, the subscript s represents the sea surface reflected wave, and the subscript i represents the i-th horizontal distance-depth grid.

[0041] Assume the transmitted signal is The simulated received signal can then be expressed as:

[0042] ,

[0043] Calculate the signal power of each array element, and add noise within a certain signal-to-noise ratio range according to the power signal-to-noise ratio definition. The simulated received signal can then be expressed as:

[0044] ,

[0045] in It is Gaussian white noise that is uncorrelated with the target signal.

[0046] Performing a Fourier transform on the received signal yields the frequency domain representation of the simulated received signal: The steering vector is calculated based on the array parameters as follows:

[0047] ,

[0048] Where c is the speed of sound; This represents the Fourier transform.

[0049] The interference matrix can be obtained by multiplying the conjugate matrix of the steering vector by the simulated received signal:

[0050] ,

[0051] That is, the frequency-elevation diagram, the expanded interference matrix is matrix.

[0052] Step 2: Since the original signal form of the known transmitted signal is... Based on the principle of matched filtering, the local signal is obtained by inverting the time domain of the original signal. Convolving it with the interference matrix completes the matched filtering to form the time delay-elevation diagram; given the speed of sound c, the sound path... ,in The time delay axis parameter in the time delay-elevation plot is transformed into a sound path-elevation plot after sound path conversion. The sound path-elevation plot can be represented as:

[0053]

[0054] IFFT stands for Inverse Fourier Transform.

[0055] Based on the defined horizontal distance-depth grid, calculate the sound path-pitch angle matrix for all grids, and then proceed to the next step.

[0056] Step 3: The path length of the path-pitch diagram depends on the application range. In cases of long distance and high depth, the size of the path-pitch diagram will be too large to be applied, and the bright spots and their side lobes in the image will occupy a very small proportion of the image. Therefore, the image needs to be appropriately cropped according to the position of the bright spots in the image.

[0057] Since two sound rays are considered—the direct wave and the sea-reflected wave—two bright spots will be observed in the sound path-elevation diagram. Let the coordinates of the direct wave bright spot be... The coordinates of the bright spot of the sea surface reflected wave are The coordinates of the cropped window are ,in The path length extension is The azimuth axis extension length is To ensure that the rectangular clipping window can cover all bright spots and their slant-distance side lobes, the following must be satisfied: ,in It is the maximum time delay difference between the direct wave and the sea surface reflected wave in the defined horizontal distance-depth grid. It refers to the sampling rate; the azimuth axis extension length. It must meet the requirements of covering the bright spot and its azimuth sidelobe.

[0058] The final size of the cropped rectangular window is: In practical applications, the rectangular window can be slightly enlarged based on this formula to allow for some margin. The cropped rectangular window is one of the features input into the network, and the coordinates of the midpoint of the rectangular window... It is also one of the features input into the network, while the network labels are horizontal distance and depth.

[0059] Step 4: Construct an unsupervised domain adaptation-based vertical array underwater sound source ranging and depth sounding model (ARA-Net) in Python. The core of the model consists of four modules: a decoupled strip attention feature extraction module, a two-layer MLP physical coordinate mapping module, a homoscedastic uncertainty multi-task regression module, and an unsupervised domain adversarial module.

[0060] The system's data flow is as follows: the cropped sound path-elevation angle map is input to the decoupled strip attention feature extraction module, which outputs an image feature vector; the midpoint of the cropped rectangular window is input to the two-layer MLP physical coordinate mapping module, which outputs a coordinate feature vector; the two are concatenated to obtain a multimodal fusion feature vector. This fusion feature vector is independently input to the homoscedastic uncertainty multi-task regression module, and also input to the unsupervised adversarial discrimination module through a gradient inversion layer.

[0061] 4.1 Decoupling the strip attention feature extraction module:

[0062] The decoupled strip attention feature extraction module includes a deep asymmetric convolutional neural network (CNN) for receiving the path-pitch map (RD map) obtained from step 3, with dimensions of height (H) × width (W), and its internal structure is sequentially connected to the following units:

[0063] A1. Asymmetric Convolutional Primary Unit and Downsampling Unit: Contains a unit with a size of... A two-dimensional convolutional layer (without bias) is followed by an instance normalization layer (InstanceNorm2d), a PReLU activation function layer, and a pooling kernel with a size of [size missing]. Step size is A two-dimensional max-pooling layer.

[0064] A2. Intermediate Convolutional and Downsampling Units: Each unit contains one... The system consists of a 2D convolutional layer, a 2D batch normalization layer (BatchNorm2d), a PReLU activation function layer, and a pooling kernel. Step size is A two-dimensional max-pooling layer.

[0065] A3. Decoupling Strip Attention Mechanism Unit: Let the input feature map tensor be... Where C is the number of channels; adaptive average pooling and adaptive max pooling are performed in the height and width directions respectively to extract extreme energy features and background smoothing features:

[0066] ,

[0067] Will Dimension transpose and with splicing along spatial dimensions, through convolution Interacting with non-linear activation functions (PReLu):

[0068] ,

[0069] Where r is the channel scaling factor.

[0070] The fused feature y is decoupled and cut into row feature vectors of length H. and column eigenvectors of length W (Transposition required). Separately via independent... Convolution restores the number of channels to C, and the sigmoid activation function is used to generate a decoupled attention weight matrix:

[0071]

[0072] in This is the Sigmoid activation function.

[0073] Finally, the row and column attention weights are multiplied by the input feature map X using the Hadamard product to obtain the enhanced output feature map: .

[0074] A4. Deep Feature Mapping and Convergence Unit: Contains a Two-dimensional convolutional layer, batch normalization layer, PReLU activation layer, one A two-dimensional max pooling layer, and finally an adaptive average pooling layer (with a fixed output dimension). The output is flattened into a one-dimensional image feature vector.

[0075] 4.2 Two-layer MLP physical coordinate mapping module:

[0076] The module is a multilayer perceptron (MLP) and includes:

[0077] B1. First fully connected layer: A linear fully connected layer with an input dimension of 2 and an output dimension of 32, followed by a normalization layer (LayerNorm) and a PReLU activation function layer;

[0078] B2. Second fully connected layer: A linear fully connected layer with an input dimension of 32 and an output dimension of 64, followed by a normalization layer and a PReLU activation function layer.

[0079] Map the center coordinates of the cropping window to a high-dimensional coordinate feature vector:

[0080] ,

[0081] in , For the first Learnable weights and biases of layers Layer normalization (LayerNorm) The PReLU activation function is used, where P is the coordinate of the midpoint of the clipped rectangular window. This is the dimension of the output coordinate features.

[0082] Let the flattened image feature vector extracted by the backbone network be... The flattened image feature vector is concatenated with the coordinate feature vector output by this module along the feature dimension to obtain the final multimodal fusion feature vector:

[0083] .

[0084] 4.3 Homoscedasticity Uncertainty Multi-Task Regression Module:

[0085] This module consists of two parallel and independent sub-networks used to simultaneously perform sound source distance prediction (Range, typically in kilometers) and depth prediction (Depth, typically in hundreds of meters). Each sub-network sequentially contains the following neural layers: a first fully connected linear layer (output 256 dimensions) receiving the fused feature vector, a layer normalization layer (LayerNorm), a PReLU activation function layer, a dropout layer (dropout rate 0.3), a second fully connected linear layer (output 64 dimensions), another PReLU activation function layer, and a third fully connected linear layer (output 1 dimension).

[0086] The third linear fully connected layer directly outputs a scalar prediction value, and is not connected to any nonlinear activation function layer thereafter.

[0087] During deep learning model training, the mean squared error loss for distance tasks and depth tasks are defined as follows: and :

[0088] ,

[0089] in , For network predictions, Labeled as physically real.

[0090] Two learnable noise parameters are introduced. and Construct a joint multi-task loss function under maximum likelihood estimation. :

[0091] ,

[0092] During the backpropagation optimization process, when the error of a certain task (such as...) When the value is large, the network tends to increase. To reduce its weight ratio and use the latter term for regularization penalty, dynamic and stable alignment of gradients across multiple tasks can be achieved.

[0093] 4.4 Unsupervised Domain Adversarial Module:

[0094] To eliminate environmental noise and parameter mismatch between purely simulated acoustic data (source domain S) and real sea trial acoustic data (target domain T), this model introduces an unsupervised domain adversarial mechanism. This unsupervised domain adversarial module operates only during training and consists of two parts:

[0095] C1. Gradient Reversal Layer (GRL): Serially connected between the multimodal fusion feature vector and the domain discriminator, the gradient reversal layer is identity-transferred during forward propagation. In the backward propagation computation graph, this layer multiplies the gradient tensor by a negative coefficient. The reverse function.

[0096] C2. Domain Discriminator Branch: Receives the fused features passed through the GRL. This branch, also a fully connected sub-network, sequentially includes: a first linear layer (256-dimensional output), a layer normalization layer, a PReLU layer, a Dropout layer (0.3 dropout rate), a second linear layer (64-dimensional output), another PReLU layer, and a third linear layer (1-dimensional output). Unlike the regression module, this branch forcibly connects a Sigmoid activation function layer after the third linear layer.

[0097] The decoupled strip attention feature extraction module and the two-layer MLP physical coordinate mapping module are collectively referred to as the feature extractor (including the fusion layer), and the corresponding mapping function is... Output multimodal fusion feature vector For short, fusion features .

[0098] Let the mapping function of the homoscedastic uncertainty multi-task regression module be: Calculate the aforementioned .

[0099] Let the mapping function of the domain discriminator be: Its output characteristics belong to the real sea trial domain (target domain) The probability of ) The domain classification loss uses binary cross-entropy loss:

[0100] ,

[0101] This enables the domain discriminator to accurately determine the probability that a feature belongs to the source or target domain based on the input multimodal fusion feature.

[0102] The goal of adversarial training is a Minimax game process:

[0103] ,

[0104] in, The dynamic domain fitness coefficients are used. To solve this minimax problem in one go using the backpropagation algorithm within a standard deep learning framework, this invention... and A gradient inversion layer (GRL) is inserted between the layers. During backpropagation of the computation graph, the GRL discriminates the error gradient of the domains it flows through. Multiply by a negative coefficient :

[0105] ,

[0106] Therefore, the feature extractor parameters The update gradient automatically becomes:

[0107] ,

[0108] This mechanism forces The update direction continuously increases the domain classification error, ultimately achieving a match between the source and target domains in the feature space. Alignment of probability distributions in (i.e.) ).

[0109] Example:

[0110] Assume the original transmitted signal waveform is a set of hyperbolic FM signals, consisting of one rising hyperbolic FM and one falling hyperbolic FM, with a signal bandwidth of 500Hz, a center frequency of 750Hz, a total pulse width of 2s, a sampling rate of 3000, 128 array elements, a cropped rectangular window size of (1201*51), and a simulated signal-to-noise ratio of -20dB for each array element. Test points with a distance of 3.9km and a depth of 5:2:201m are selected for testing.

[0111] like Figure 1 and Figure 2 The provided parameter descriptions for the training and validation sets, as well as the sound velocity profile, were obtained in this invention using an acoustic toolbox and numerical calculation software. Figure 3 A sound path-pitch diagram at a certain horizontal distance-depth. Figure 4 The red border in the image represents the cropped path-elevation plot, one of the features input into the deep learning network. This plot is constructed in Python. Figure 5 Deep learning networks ultimately yield... Figure 6 The prediction results, from Figure 6 The prediction results show that both the depth estimation MAE and the distance estimation MAE are within acceptable ranges even at extremely low signal-to-noise ratios. These simulation results verify the feasibility of this invention.

[0112] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A vertical array pulse signal ranging and sounding method based on a deep learning network, characterized in that, Includes the following steps: S1. Obtain marine environmental parameters, vertical array parameters, and waveform parameters of pulse signals. Simulate and calculate the arrival delay and amplitude of direct waves and sea surface reflected waves in the preset horizontal distance-depth grid. Generate the simulated received signal and add noise. Perform a Fourier transform on the signal to obtain the frequency domain received signal. Calculate the steering vector based on the array parameters and construct a frequency-elevation diagram. S2. According to the principle of matched filtering, the local signal is obtained by time-domain inversion of the transmitted signal. Based on the frequency-elevation diagram and the local signal, the time delay-elevation diagram is obtained by convolution. Then, the time delay axis in the time delay-elevation diagram is converted into the sound path axis by using the speed of sound to obtain the sound path-elevation diagram. S3. Based on the coordinates of the direct wave bright spot and the coordinates of the sea surface reflected wave bright spot, the sound path-elevation angle diagram is cropped to obtain the cropped rectangular window image and its midpoint coordinates. S4. Input the cropped rectangular window image and its midpoint coordinates into the deep learning model to obtain the horizontal distance and depth estimates of the sound source.

2. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 1, characterized in that, the generated simulation reception signal wherein , is the time delay of the direct wave and the sea surface emitted wave, , is the corresponding amplitude, is the transmitted signal.

3. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 1, characterized in that, The cropping of the sound path-elevation diagram includes: The path length extension of the cut rectangular window is at least the product of the maximum time delay difference between the direct wave and the sea surface reflected wave under the preset grid and the sampling rate, and the azimuth extension length satisfies the coverage of the bright spot and its azimuth sidelobe.

4. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 1, characterized in that, The deep learning model includes a decoupled strip attention feature extraction module, a two-layer MLP physical coordinate mapping module, and a homoscedastic uncertainty multi-task regression module. The decoupled strip attention feature extraction module receives a cropped rectangular window image and extracts features, outputting an image feature vector. The two-layer MLP physical coordinate mapping module receives the coordinates of the midpoint of the cropped rectangular window, performs feature mapping, and outputs a coordinate feature vector. The image feature vector and the coordinate feature vector are concatenated to obtain a multimodal fusion feature vector, which is then input into the homoscedastic uncertainty multi-task regression module for distance prediction and depth prediction.

5. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 4, characterized in that, The decoupled strip attention feature extraction module sequentially includes an asymmetric convolution primary unit and a downsampling unit, an intermediate convolution and downsampling unit, a decoupled strip attention mechanism unit, and a deep feature mapping and convergence unit. The decoupled strip attention mechanism unit performs adaptive average pooling and adaptive max pooling in both the height and width directions, respectively. It interacts with information through a non-linear activation function of a convolutional kernel, and then decouples the obtained features into row feature vectors and column feature vectors to obtain row attention weights and column attention weights. The row attention weights and column attention weights are then multiplied with the input feature map of the decoupled strip attention mechanism unit by a Hadamard product to obtain the enhanced output feature map. The deep feature mapping and convergence unit obtains a one-dimensional image feature vector, i.e., the output image feature vector, based on the output of the decoupled strip attention mechanism unit.

6. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 4, characterized in that, The dual-layer MLP physical coordinate mapping module is a multilayer perceptron, including two sequentially connected fully connected layers, each followed by a normalization layer and a PReLU activation layer; the dual-layer MLP physical coordinate mapping module maps the center coordinates of the clipping window to high-dimensional coordinates.

7. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 4, characterized in that, The homoscedastic uncertainty multi-task regression module includes two parallel and independent sub-networks, which are used to perform distance prediction and depth prediction, respectively. Each sub-network contains multiple sequentially connected fully connected layers, layer normalization layers, PReLU activation function layers and random deactivation layers. The last fully connected layer outputs the distance or depth prediction value.

8. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 4, characterized in that, During deep learning model training, the homoscedastic uncertainty multi-task regression module introduces two learnable noise parameters, and combines the mean squared error loss of the distance task and the depth task to construct a joint multi-task loss function under maximum likelihood estimation: ， in and To define the mean squared error loss for distance and depth tasks, and Two learnable noise parameters are introduced.

9. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 8, characterized in that, The deep learning model also includes an unsupervised domain adversarial module. During model training, the multimodal fused feature vector is input to the unsupervised domain adversarial discriminant module. The unsupervised domain adversarial module includes a gradient reversal layer and a domain discriminator branch. The gradient reversal layer is concatenated between the multimodal fused feature vector and the domain discriminator branch. During forward propagation, the gradient reversal layer is propagated identically, and during backward propagation, the domain discrimination error gradient is multiplied by a negative coefficient for reversal. The domain discriminator branch is a fully connected subnetwork, with a Sigmoid activation function layer at its end, outputting the probability that the input feature belongs to the target domain; thus, the domain classification loss is obtained. ,based on and Adversarial training is performed using gradient inversion layers and domain discriminators to achieve feature alignment between the source and target domains.

10. The method for ranging and depth measurement using vertical array pulse signals based on deep learning networks according to claim 9, characterized in that, The adversarial training process of the deep learning network model is a minimax game process. By multiplying the gradient of the domain discrimination error through the gradient reversal layer by a negative coefficient, the feature extractor parameters are forced to be updated in the direction of increasing the domain classification error, so as to align the probability distribution of the source domain simulation data with the target domain real data in the feature space.

Images