Underwater target positioning method based on data-model dual drive

Abstract

The invention discloses an underwater target positioning method based on data-model dual drive. The method specifically comprises the following steps: data acquisition: acquiring initial data by using a receiver; performing sight distance reasoning: constructing and training a sight distance information reasoning neural network based on data driving, preprocessing the collected initial data, inputting the preprocessed data into the neural network, and outputting arrival time estimation of a sight distance path; joint estimation: adopting a model driving method, calculating the arrival time difference of sound signals received by a receiver by using the obtained sight distance arrival time, designing underwater target position and signal propagation speed joint estimation based on the arrival time difference, and calculating initial estimation values of the target position and the propagation speed by using a weighted least square method; and robust iteration: based on a maximum likelihood estimation function, performing robust iteration optimization on the obtained preliminary estimation value to obtain a stable and convergent final joint estimation result. The target positioning method is convenient to popularize and apply in different underwater environments, and has relatively high adaptability and expansibility.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of underwater target monitoring, and particularly relates to an underwater target positioning method based on data-model double driving. BACKGROUND

[0002] At present, underwater target positioning has important applications in the fields of marine resource development, marine environment monitoring and underwater robot navigation. However, the underwater environment is complex and changeable, and the sound wave propagation often produces serious multipath effect due to factors such as seabed topography, water organisms and water surface fluctuations, which leads to that the traditional positioning method based on the direct path is easily disturbed by non-line-of-sight propagation in actual application, resulting in measurement error and insufficient positioning accuracy. The existing methods either adopt constraint optimization or rely on traditional least square solution, and it is difficult to balance real-time performance and high accuracy in a severe multipath environment, and part of the technology cannot be popularized and applied due to the difficulty of data collection and high computational complexity. SUMMARY

[0003] The technical problem to be solved by the present application is to provide an underwater target positioning method based on data-model double driving, which combines the advantages of data driving and model driving, and solves the problems of difficulty in accurately extracting line-of-sight information, large target positioning error and high computational complexity of traditional methods in a complex multipath environment.

[0004] In order to solve the above technical problems, the present application is realized by the following way: An underwater target positioning method based on data-model double driving, specifically comprising the following steps: S1, data collection: collecting initial data by using a receiver; S2, line-of-sight reasoning: constructing and training a line-of-sight information reasoning neural network based on data driving, inputting the initial data collected in step S1 into the neural network after preprocessing, and outputting the time of arrival estimation of the line-of-sight path; S3, joint estimation: using the line-of-sight time of arrival obtained in step S2, calculating the time difference of arrival of the sound signal received by the receiver by using a model driven method, and designing a joint estimation of the underwater target position and the signal propagation speed based on the time difference of arrival, and using a weighted least square method to solve the preliminary estimation value of the target position and the propagation speed; S4, robust iteration: based on the maximum likelihood estimation function, the preliminary estimation value obtained in step S3 is robustly iterated and optimized to obtain the final joint estimation result of stable convergence.

[0005] Further, the specific method of step S1 is as follows: M (M≥6) receivers are arranged in space, initialization: determining a space rectangular coordinate system, determining the position coordinates s1, s2,..., s MDetermine the estimated propagation speed of underwater acoustic signals. c (That is, to determine the approximate propagation speed of underwater acoustic signals based on prior knowledge in this environment).

[0006] Multiple receivers are used to receive the sound wave signals emitted by the sound source, forming a set of observations including arrival time, propagation loss, and angle of arrival.

[0007] Furthermore, the specific method for step S2 is as follows: S21. Dataset Construction: Construct a high-fidelity, multi-scenario neural network training dataset, use the Bellhop ray tracing model to generate multipath channel characteristic data, and preprocess the dataset. S22, Neural Network Model Design; S23. Training and inference of the neural network model.

[0008] Furthermore, step S21 specifically includes the following steps: S211, Dynamic Configuration of Environmental Parameters Set the vertical layered sound velocity profile parameters and seabed topographic undulation characteristic parameters of the simulation model to cover complex topographic scenarios in both deep and shallow seas. S212, Multipath Channel Simulation Execution Running the Bellhop model to perform ray tracing calculations will generate the following output files: Multipath information file: Records the arrival time series, amplitude, and phase information of multipath paths. Path trajectory file: Stores the geometric parameters and propagation time of the sound ray propagation trajectory; S213. Selection of Multipath Characteristic Parameters Parse the Bellhop output file and select the following physical characteristic parameters: Propagation delay τ : Read the path arrival time from the path trajectory file as a multipath delay feature; transmission loss TL The specific expression is as follows: in, Indicates the absolute value of the vocal timbre amplitude; Angle of arrival θ : Read the path arrival angle from the path trajectory file; Line-of-sight path delay label generation: based on the geometric distance between the sound source and the receiver. d and current sound speed estimate c est Calculate the theoretical line-of-sight delay as a supervision label for neural network training. τ los The specific expression is as follows: τ los =d / c est S214. For the nth specific simulation scenario (where n ranges from 1 to n), N After performing steps S211-S213, the feature parameters of P multipath paths in the scene can be extracted. These parameters are then combined to form the feature matrix X of the single scene. (n) Its expression is as follows: ; S215, Changing the location of the sound source and receiver, seabed topography N Repeat steps S211-S214 each time to obtain the training dataset, as shown in the following expression: The corresponding supervision label vector expression is as follows: ; S216, Data Preprocessing From the feature matrix of the dataset In the middle, extract the valid multipath path as follows: Path sorting: For each scene's P sound lines, sort them by propagation delay. τ Sort in ascending order to ensure that the earliest arriving line-of-sight path is placed first. Path filtering: Extract the first 4 paths to form a new matrix. , Mask generation: Constructing a binary mask matrix This identifies the valid path location, expressed as follows: ; For the truncated matrix X cut Data standardization is performed by applying Z-score normalization to each feature and label of the matrix to achieve zero mean and unit variance, resulting in the normalized matrix, as shown in the following expression: ,in, Let represent the normalized multipath delay, propagation loss, angle of arrival matrix, and the normalized supervision label vector, respectively. y norm .

[0009] Furthermore, step S22 includes the following steps: S221. Neural Network Model Architecture Construction Input features: Normalized feature matrix Multipath delay including normalization Dissemination loss Angle of arrival From X norm Extract the delay of the first path. As a residual direct connection signal; Low-dimensional parameter processing: Input features are encoded into the latent space through linear mapping, and a trainable weight matrix W is introduced. e With bias b e The expression is as follows: H in = X norm W e + b e , Self-attention core module: First, perform H on the input features. in = LayerNorm(H in ) Parametric projection and split calculation: Feature extraction is achieved through a parallel 4-head attention mechanism, with each head internally implemented through parameter matrix decomposition and query matrix. Key matrix Value matrix ,in, They represent the first i The query, key, and value weight matrix for each attention head; Head fusion and residual connectivity: The outputs from each head are spliced ​​together and then linearly projected using MultiHead = Concat( head 1 ,..., head 4 W0, residual connections preserve the original features H mid = H in + MultiHead(Q,K,V), where W0 represents the output weight matrix of the multi-head attention mechanism; Gated nonlinear transformation: For H mid First, layer normalization is performed. Then, the first-level expansion expands the dimension by a factor of 4 using the first feedforward network weight matrix W1, and the GELU gated activation function achieves non-linear interaction. The second-level compression restores the original dimension using the second feedforward network weight matrix W2, as shown in the following expression: ; Delay prediction output module: Path dimension compression, through W r Multi-path features are aggregated into a single path, as shown in the following expression: H red = H out W r Among them, W r This represents the path compression weight matrix, whose function is to compress the previously obtained feature H, which contains information about multiple paths. out A linear transformation is performed to aggregate or compress this multi-path information, resulting in a representative single-path feature representation H. red .

[0010] Through gated nonlinear transformation, the expression is as follows: Delay Residual Fusion: The gated nonlinear transformation result is coupled with the first path delay, as expressed below: H out = GELU(H red W1) W2 Delay Residual Fusion: The gated nonlinear transformation result is coupled with the first path delay, as expressed below: Finally, the estimated line-of-sight path arrival time is obtained. As input for subsequent algorithms.

[0011] Furthermore, step S23 includes the following steps: S231, Neural Network Model Training The predicted line-of-sight path delay value output by the neural network Compared with the actual delay label The mean squared error loss is calculated to drive model parameter optimization. The expression for the mean squared error loss is as follows: ,in, N Indicates the number of samples in the batch. Indicates the first i The predicted latency value for each scenario Indicates the first i The real-time delay labels for each scenario are used to calculate the gradient of the loss function with respect to the network parameters using the gradient descent algorithm. The neural network parameters are then updated, and training is completed after multiple iterations. S232, Neural Network Model Inference The set of multipath observations collected in step S1, including arrival time, propagation loss, and angle of arrival, is preprocessed and input into the trained neural network model to output the estimated arrival time of the line-of-sight path from the sound source to the receiver for use in step S3.

[0012] Furthermore, the specific method for step S3 is as follows: The estimated line-of-sight arrival time of sound waves from the sound source to each receiver is obtained from the output of the line-of-sight inference neural network. t i Based on the time of arrival at the first receiver, calculate the... i The arrival time difference between the first receiver and the second receiver And based on prior knowledge, the propagation speed of underwater sound waves is set to... c ; Based on the time difference of arrival (TDOA) measurement model, the estimated target position u and propagation speed are constructed.v = c + x The pseudolinear equation is as follows: ,in, , , n i Indicates the first i Measurement noise of each receiver, , Indicates the first i Position error of each receiver; To eliminate the pathological nature of the original problem, G1 is divided into G... a = G1(:, 1: N ) and G b = G1(:, N+ 1: N+ 4) Construct a structure that satisfies The projection matrix V is used to project the equation onto a full-rank subspace using null space projection, transforming it into a full-rank equation expression as follows: Where, h2 = V T h1, G2 = V T G a A2 = V T A1, B2 = V T B1; The initial estimate of the target location is obtained by weighted least squares, and the specific expression is as follows: Among them, the weight matrix Q t Q is the covariance matrix of the measurement noise n. s Position error The covariance matrix; use Replace α1 r 1. Construct an equation that depends only on the propagation speed variable, as shown in the following expression: ,in, ; Then, a preliminary estimate of the target location is obtained through weighted least squares, as shown in the following expression: Among them, the weight matrix ; Propagation speed variable x initial estimate x' = α2(1), the preliminary estimate of the propagation speed is v' = c + x' .

[0013] Furthermore, the specific method for step S4 is as follows: The maximum likelihood objective function is constructed based on the relationship between the time difference of arrival, the target position, and the propagation speed, as shown in the following expression: ,in, ，q i This represents the i-th diagonal element of Q. , The arrival time difference with the initial estimated result z' as the parameter. Q t Q represents the covariance matrix of the measurement noise. s The covariance matrix represents the position error; Then, construct a system containing auxiliary variables. τ i The factor graph model is constructed, and two factor nodes are defined: The sum-product algorithm (SPA) in the factor graph is used, and the nonlinear function is applied. Perform Taylor expansion and linearization to optimize the posterior marginal distribution of the target z to a Gaussian distribution. The optimized objective estimation result is expressed as follows: in, , ; To achieve robust estimation of the target variable, a forgetting factor is introduced. λ > 0 is used to control the proportion of historical information retained, and is combined with the scaling factor. ρ j Numerical normalization is achieved, thereby constructing an iterative estimation and update mechanism with a maximum number of iterations. J =100000, No. j In this iteration, the estimation results are expressed as follows: in, , ; Finally, the convergence condition is met. Or reach the maximum number of iterations J Then, the iteration is terminated, and the final estimation result is output as follows: .

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes a classical underwater acoustic signal propagation model to generate channel parameters, providing physical priors for the data-driven module. This effectively reduces the data-driven algorithm's dependence on large amounts of measured data, solving the challenge of underwater data acquisition. A deep neural network based on the Transformer architecture is employed to intelligently extract line-of-sight (LoS) features from dense multipath signals, eliminating or correcting errors caused by non-line-of-sight paths. A joint estimation model of target position and sound velocity is established based on Time Difference of Arrival (TDOA), and solved using weighted least squares to obtain preliminary joint estimation results. Based on the preliminary joint estimation, a robust iterative convergence joint parameter optimization method is used to optimize the joint estimation results, further improving its accuracy and robustness. This enhances both the accuracy of the positioning algorithm and ensures real-time performance. The system's modular design facilitates application in various underwater environments and possesses high adaptability and scalability. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the target localization method of the present invention; Figure 2 This is a schematic diagram of the line-of-sight inference module of the Transformer of the present invention; Figure 3 This is a schematic diagram of the factor graph model of the present invention; Figure 4 This is a schematic diagram of a specific implementation scenario of the present invention. Detailed Implementation

[0016] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0017] like Figures 1-4 As shown, an underwater target localization method based on a data-model dual-drive approach includes the following steps: S1. Data Acquisition: Initial data is acquired using a receiver, as follows: Will M ( M ≥6 receivers are arranged in space. Initialization: determine the spatial rectangular coordinate system, and determine the position coordinates s1, s2, ..., s of each receiver. M Determine the estimated propagation speed of underwater acoustic signals. c (That is, to determine the approximate propagation speed of underwater acoustic signals based on prior knowledge in this environment).

[0018] Multiple receivers are used to receive the sound wave signals emitted by the sound source, forming a set of observations including arrival time, propagation loss, and angle of arrival.

[0019] S2. Line-of-sight inference: Construct and train a data-driven line-of-sight information inference neural network. The initial data collected in step S1 is preprocessed and input into the neural network, outputting an estimate of the arrival time of the line-of-sight path. The specific method is as follows: S21. Dataset Construction: Construct a high-fidelity, multi-scenario neural network training dataset. Use the Bellhop ray tracing model to generate multipath channel characteristic data, and preprocess the dataset, specifically including the following steps: S211, Dynamic Configuration of Environmental Parameters Set the vertical layered sound velocity profile parameters and seabed topographic undulation characteristic parameters of the simulation model to cover complex terrain scenarios in both deep and shallow seas.

[0020] S212, Multipath Channel Simulation Execution Running the Bellhop model to perform ray tracing calculations will generate the following output files: Multipath information file: Records the arrival time series, amplitude, and phase information of multipath paths. Path trajectory file: Stores the geometric parameters and propagation time of the sound ray propagation trajectory; S213. Selection of Multipath Characteristic Parameters Parse the Bellhop output file and select the following physical characteristic parameters: Propagation delay τ : Read the path arrival time from the path trajectory file as a multipath delay feature; transmission loss TL The specific expression is as follows: ,in, Indicates the absolute value of the vocal timbre amplitude; Angle of arrival θ : Read the path arrival angle from the path trajectory file; Line-of-sight path delay label generation: based on the geometric distance between the sound source and the receiver. d and current sound speed estimate c est Calculate the theoretical line-of-sight delay as a supervision label for neural network training. τ los The specific expression is as follows: τ los = d / c est S214. For the nth specific simulation scenario (where n ranges from 1 to n), N After performing steps S211-S213, the feature parameters of P multipath paths in the scene can be extracted. These parameters are then combined to form the feature matrix X of the single scene. (n) Its expression is as follows: ; S215, Changing the location of the sound source and receiver, seabed topography N Repeat steps S211-S214 each time to obtain the training dataset, as shown in the following expression: The corresponding supervision label vector expression is as follows: ; S216, Data Preprocessing From the feature matrix of the dataset In the middle, extract the valid multipath path as follows: Path sorting: For each scene's P sound lines, sort them by propagation delay. τ Sort in ascending order to ensure that the earliest arriving line-of-sight path is placed first. Path filtering: Extract the first 4 paths to form a new matrix. , Mask generation: Constructing a binary mask matrix This identifies the valid path location, expressed as follows: ; For the truncated matrix X cut Data standardization is performed by applying Z-score normalization to each feature and label of the matrix to achieve zero mean and unit variance, resulting in the normalized matrix, as shown in the following expression: ,in, Let represent the normalized multipath delay, propagation loss, angle of arrival matrix, and the normalized supervision label vector, respectively. y norm .

[0021] S22. Neural network model design, including the following steps: S221. Neural Network Model Architecture Construction Input features: Normalized feature matrix Multipath delay including normalization Dissemination loss Angle of arrival From X norm Extract the delay of the first path. As a residual direct connection signal; Low-dimensional parameter processing: Input features are encoded into the latent space through linear mapping, and a trainable weight matrix W is introduced. e With bias b e The expression is as follows: H in = X norm W e + b e Self-attention core module: First, perform H on the input features. in = LayerNorm(H in ) Parametric projection and split calculation: Feature extraction is achieved through a parallel 4-head attention mechanism, with each head internally implemented through parameter matrix decomposition and query matrix. Key matrix Value matrix ,in, They represent the first i The query, key, and value weight matrix for each attention head; Head fusion and residual connectivity: The outputs from each head are spliced ​​together and then linearly projected using MultiHead = Concat( head 1 ,..., head 4 W0, residual connections preserve the original features H mid = H in + MultiHead(Q,K,V), where W0 represents the output weight matrix of the multi-head attention mechanism; Gated nonlinear transformation: For H mid First, layer normalization is performed. Then, the first-level expansion expands the dimension by a factor of 4 using the first feedforward network weight matrix W1, and the GELU gated activation function achieves non-linear interaction. The second-level compression restores the original dimension using the second feedforward network weight matrix W2, as shown in the following expression: ; Delay prediction output module: Path dimension compression, through W r Multi-path features are aggregated into a single path, as shown in the following expression: H red = H out W r Among them, W r This represents the path compression weight matrix; its function is to compress the previously obtained feature H, which contains information about multiple paths. out The purpose of performing a linear transformation is to aggregate or compress this multi-path information to obtain a representative single-path feature representation H. red .

[0022] Through gated nonlinear transformation, the expression is as follows: H out = GELU(H redW1) W2, Delay Residual Fusion: The gated nonlinear transformation result is coupled with the first path delay, as shown in the following expression: Finally, the estimated line-of-sight path arrival time is obtained. As input for subsequent algorithms.

[0023] S23. Training and inference of the neural network model, including the following steps: S231, Neural Network Model Training The predicted line-of-sight path delay value output by the neural network Compared with the actual delay label The mean squared error loss is calculated to drive model parameter optimization. The expression for the mean squared error loss is as follows: ,in, N Indicates the number of samples in the batch. Indicates the first i The predicted latency value for each scenario Indicates the first i The real-time delay labels for each scenario are used to calculate the gradient of the loss function with respect to the network parameters using the gradient descent algorithm. The neural network parameters are then updated, and training is completed after multiple iterations. S232, Neural Network Model Inference The set of multipath observations collected in step S1, including arrival time, propagation loss, and angle of arrival, is preprocessed and input into the trained neural network model to output the estimated arrival time of the line-of-sight path from the sound source to the receiver for use in step S3.

[0024] S3. Joint Estimation: Using a model-driven approach, the time difference of arrival (TDOA) of the sound signal received by the receiver is calculated using the line-of-sight time of arrival (LOA) obtained in step S2. A joint estimation of the underwater target position and signal propagation speed based on the TOA is designed. The preliminary estimates of the target position and propagation speed are obtained using the weighted least squares method. The specific method is as follows: The estimated line-of-sight arrival time of sound waves from the sound source to each receiver is obtained from the output of the line-of-sight inference neural network. t i Based on the time of arrival at the first receiver, calculate the... i The arrival time difference between the first receiver and the second receiver And based on prior knowledge, the propagation speed of underwater sound waves is set to... c ( c (Take a speed of 1400-1600 m / s). Based on the time difference of arrival (TDOA) measurement model, the estimated target position u and propagation speed are constructed. v = c + x The pseudolinear equation is as follows: ,in, , n i Indicates the first i Measurement noise of each receiver, , Indicates the first i Position error of each receiver; To eliminate the pathological nature of the original problem, G1 is divided into G... a = G1(:, 1: N ) and G b = G1(:, N+ 1: N+ 4) Construct a structure that satisfies The projection matrix V is used to project the equation onto a full-rank subspace using null space projection, transforming it into a full-rank equation expression as follows: Where, h2 = V T h1, G2 = V T G a A2 = V T A1, B2 = V T B1; The initial estimate of the target location is obtained through weighted least squares (WLS), and the specific expression is as follows: Among them, the weight matrix ; use Replace α1 r 1. Construct an equation that depends only on the propagation speed variable, as shown in the following expression: ,in, ; Then, a preliminary estimate of the target location is obtained through weighted least squares (WLS), as shown in the following expression: Among them, the weight matrix ; Propagation speed variable x initial estimate x' = α2(1), the preliminary estimate of the propagation speed is v' = c + x' .

[0025] S4. Robust Iteration: Based on the maximum likelihood estimation function, the preliminary estimate obtained in step S3 is robustly iterated and optimized to obtain a stable and convergent final joint estimation result. The specific method is as follows: The maximum likelihood objective function is constructed based on the relationship between the time difference of arrival, the target position, and the propagation speed, as shown in the following expression: in, , q i This represents the i-th diagonal element of Q. , The arrival time difference with the initial estimated result z' as the parameter. Q t Q represents the covariance matrix of the measurement noise. s The covariance matrix represents the position error; Then, construct a system containing auxiliary variables. τ i The factor graph model is constructed, and two factor nodes are defined: The sum-product algorithm (SPA) in the factor graph is employed, and the nonlinear function is... Perform Taylor expansion and linearization to optimize the posterior marginal distribution of the target z to a Gaussian distribution. The optimized objective estimation result is expressed as follows: in, , ; To achieve robust estimation of the target variable, a forgetting factor is introduced. λ > 0 is used to control the proportion of historical information retained, and is combined with the scaling factor. ρ j Numerical normalization is achieved, thereby constructing an iterative estimation and update mechanism with a maximum number of iterations. J =100000, No. j In this iteration, the estimation results are expressed as follows: in, , ; Finally, the convergence condition is met. Or reach the maximum number of iterations J Then, the iteration is terminated, and the final estimation result is output as follows: .

[0026] Example: This application sets up 10 receivers in the water, with the coordinates of each observation station being (24.84, 26.83, 0), (93.09, 26.71, 0), (32.38, 112.10, 10), (226.15, 4.34, 10), (-11.71, 63.75, 20), (188.29, 121.89, 20), (78.96, 149.36, 30), (288.37, 215.71, 30), (-23.47, 204.60, 40), and (327.13, 179.38, 40). The true three-dimensional coordinates of the target sound source location are (0, 900, 50) (unit: meters), the actual speed of underwater sound wave propagation is 1500 m / s, and its specific scenario is as follows: Figure 4 As shown.

[0027] A dual-drive data model-based underwater target localization method includes the following steps: S1. Activate all 10 receivers mentioned above to synchronously monitor and collect the acoustic signals emitted by the target sound source, and obtain channel data including the underwater target sound source and the dense multipath effect caused by the complex marine environment.

[0028] S2. For the channel data received by each receiver, preprocess the data to adapt it to the input requirements of the neural network; each receiver generates a set of standardized multipath feature parameters, which serve as the input to the deep neural network. Using the pre-trained deep neural network model, the estimated line-of-sight path arrival times (TDOAs) for the 10 receivers are obtained through neural network inference: 0.584360, 0.586764, 0.528369, 0.618687, 0.558614, 0.534552, 0.504126, 0.495835, 0.464533, and 0.528556. These estimates will be used for TDOA calculation and initial target localization in the subsequent S3 step.

[0029] S3. Select the first receiver as a reference and construct a measurement model based on TDOA, with the following values: 0.002404, -0.055991, 0.034327, -0.025746, -0.049809, -0.080234, -0.088525, -0.119827 and -0.055804. Then, based on the joint estimation of the underwater target position and signal propagation speed using the time difference of arrival, the preliminary estimates of the target position and propagation speed are obtained using the weighted least squares method, which are (1.3, 891.2, 66.2) and 1483.8 m / s, respectively.

[0030] S4. Perform robust iterative optimization on the initial estimate and perform joint optimization on the initial estimate to obtain the final solution. The sound source location is (-0.821, 905.145, 42.541), and the signal propagation speed is 1493.160m / s.

[0031] The above description is merely an embodiment of the present invention. It should be reiterated that those skilled in the art can make several improvements to the present invention without departing from the principle of the present invention, and these improvements are also included within the scope of protection of the claims of the present invention.

Claims

1. A data-model dual-driven underwater target localization method, characterized in that: Specifically, the following steps are included: S1. Data Acquisition: Initial data is acquired using a receiver; S2, Line-of-sight reasoning: Construct and train a data-driven line-of-sight information reasoning neural network. The initial data collected in step S1 is preprocessed and then input into the neural network to output the estimated arrival time of the line-of-sight path. S3. Joint estimation: Using a model-driven approach, the time difference of arrival of the sound signal received by the receiver is calculated using the line-of-sight time of arrival obtained in step S2. A joint estimation of the underwater target position and signal propagation speed based on the time difference of arrival is designed, and the preliminary estimates of the target position and propagation speed are solved using the weighted least squares method. S4. Robust Iteration: Based on the maximum likelihood estimation function, the preliminary estimate obtained in step S3 is robustly iterated and optimized to obtain a stable and convergent final joint estimation result.

2. The underwater target localization method based on data-model dual-drive as described in claim 1, characterized in that: The specific method for step S2 is as follows: S21. Dataset Construction: Construct a high-fidelity, multi-scenario neural network training dataset, use the Bellhop ray tracing model to generate multipath channel characteristic data, and preprocess the dataset. S22, Neural Network Model Design; S23. Training and inference of the neural network model.

3. The underwater target localization method based on data-model dual-drive as described in claim 2, characterized in that: Step S21 specifically includes the following steps: S211, Dynamic Configuration of Environmental Parameters Set the vertical layered sound velocity profile parameters and seabed topographic undulation characteristic parameters of the simulation model to cover complex terrain scenarios in both deep and shallow seas. S212, Multipath Channel Simulation Execution Running the Bellhop model to perform ray tracing calculations will generate the following output files: Multipath information file: Records the arrival time series, amplitude, and phase information of multipath paths. Path trajectory file: Stores the geometric parameters and propagation time of the sound ray propagation trajectory; S213. Selection of Multipath Characteristic Parameters Parse the Bellhop output file and select the following physical characteristic parameters: Propagation delay τ : Read the path arrival time from the path trajectory file as a multipath delay feature; transmission loss TL The specific expression is as follows: ,in, Indicates the absolute value of the vocal timbre amplitude; Angle of arrival θ : Read the path arrival angle from the path trajectory file; Line-of-sight path delay label generation: based on the geometric distance between the sound source and the receiver. d and current sound speed estimate c est Calculate the theoretical line-of-sight delay as a supervision label for neural network training. τ los The specific expression is as follows: τ los = d / c est S214. For the nth specific simulation scenario, after executing steps S211~S213, extract the feature parameters of P multipaths in that scenario, and combine these parameters to form the feature matrix X of that single scenario. (n) Its expression is as follows: ； S215, Changing the location of the sound source and receiver, seabed topography N Repeat steps S211-S214 each time to obtain the training dataset, as shown in the following expression: The corresponding supervision label vector expression is as follows: ; S216, Data Preprocessing From the feature matrix of the dataset In the middle, extract the valid multipath path as follows: Path sorting: For each scene's P sound lines, sort them by propagation delay. τ Sort in ascending order to ensure that the earliest arriving line-of-sight path is placed first. Path filtering: Extract the first 4 paths to form a new matrix. , Mask generation: Constructing a binary mask matrix This identifies the valid path location, expressed as follows: ； For the truncated matrix X cut Data standardization is performed by applying Z-score normalization to each feature and label of the matrix to achieve zero mean and unit variance, resulting in the normalized matrix, as shown in the following expression: , in, Let represent the normalized multipath delay, propagation loss, angle of arrival matrix, and the normalized supervision label vector, respectively. y norm .

4. The underwater target localization method based on data-model dual-drive according to claim 2, characterized in that: Step S22 specifically includes the following steps: S221. Neural Network Model Architecture Construction Input features: Normalized feature matrix Multipath delay including normalization Dissemination loss Angle of arrival From X norm Extract the delay of the first path. As a residual direct connection signal; Low-dimensional parameter processing: Input features are encoded into the latent space through linear mapping, and a trainable weight matrix W is introduced. e With bias b e The expression is as follows: H in = X norm W e +b e Self-attention core module: First, perform H on the input features. in = LayerNorm(H in ) Parametric projection and split calculation: Feature extraction is achieved through a parallel 4-head attention mechanism, with each head internally implemented through parameter matrix decomposition and query matrix. Key matrix Value matrix ,in, They represent the first i The query, key, and value weight matrix for each attention head; Head fusion and residual connectivity: The outputs from each head are spliced ​​together and then linearly projected using MultiHead = Concat( head 1 ,..., head 4 W0, residual connections preserve the original features H mid = H in + MultiHead(Q,K,V), where W0 represents the output weight matrix of the multi-head attention mechanism; Gated nonlinear transformation: For H mid First, layer normalization is performed. Then, the first-level expansion expands the dimension by a factor of 4 using the first feedforward network weight matrix W1, and the GELU gated activation function achieves non-linear interaction. The second-level compression restores the original dimension using the second feedforward network weight matrix W2, as shown in the following expression: ; Delay prediction output module: Path dimension compression, through W r Multi-path features are aggregated into a single path, as shown in the following expression: H red = H out W r Among them, W r Represents the path compression weight matrix; Through gated nonlinear transformation, the expression is as follows: H out = GELU(H red W1) W2, Delay Residual Fusion: The gated nonlinear transformation result is coupled with the first path delay, as expressed below: Finally, the estimated line-of-sight path arrival time is obtained. As input for subsequent algorithms.

5. The underwater target localization method based on data-model dual-drive according to claim 2, characterized in that: The specific method for step S23 is as follows: S231, Neural Network Model Training The predicted line-of-sight path delay value output by the neural network Compared with the actual delay label The mean squared error loss is calculated to drive model parameter optimization. The expression for the mean squared error loss is as follows: ,in, N Indicates the number of samples in the batch. Indicates the first i The predicted latency value for each scenario Indicates the first i The real-time delay labels for each scenario are used to calculate the gradient of the loss function with respect to the network parameters using the gradient descent algorithm. The neural network parameters are then updated, and training is completed after multiple iterations. S232, Neural Network Model Inference The set of multipath observations collected in step S1, including arrival time, propagation loss, and angle of arrival, is preprocessed and input into the trained neural network model to output the estimated arrival time of the line-of-sight path from the sound source to the receiver for use in step S3.

6. The underwater target localization method based on data-model dual-drive as described in claim 1, characterized in that: The specific method for step S3 is as follows: The estimated line-of-sight arrival time of sound waves from the sound source to each receiver is obtained from the output of the line-of-sight inference neural network. t i Based on the time of arrival at the first receiver, calculate the... i The arrival time difference between the first receiver and the second receiver And based on prior knowledge, the propagation speed of underwater sound waves is set to... c ; Based on the time difference of arrival (TDOA) measurement model, the estimated target position u and propagation speed are constructed. v = c + x The pseudolinear equation is as follows: ,in, , n i Indicates the first i Measurement noise of each receiver, , Indicates the first i Position error of each receiver; To eliminate the pathological nature of the original problem, G1 is divided into G... a =G1(:, 1: N ) and G b =G1(:, N+ 1: N+ 4) Construct a structure that satisfies The projection matrix V is used to project the equation onto a full-rank subspace using null space projection, transforming it into a full-rank equation expression as follows: Where h2 = V T h1, G2 = V T G a A2 = V T A1, B2 = V T B1; The initial estimate of the target location is obtained by weighted least squares, and the specific expression is as follows: Among them, the weight matrix ; use Replace α1 r 1. Construct an equation that depends only on the propagation speed variable, as shown in the following expression: in, ; Then, a preliminary estimate of the target location is obtained through weighted least squares, as shown in the following expression: Among them, the weight matrix ; Propagation speed variable x initial estimate x' = α2(1), the preliminary estimate of the propagation speed is v' = c + x' .

7. The underwater target localization method based on data-model dual-drive as described in claim 1, characterized in that: The specific method for step S4 is as follows: The maximum likelihood objective function is constructed based on the relationship between the time difference of arrival, the target position, and the propagation speed, as shown in the following expression: in, , q i This represents the i-th diagonal element of Q. , The arrival time difference with the initial estimated result z' as the parameter. Q t Q represents the covariance matrix of the measurement noise. s The covariance matrix represents the position error; Then, construct a system containing auxiliary variables. τ i The factor graph model is constructed, and two factor nodes are defined: ； The summation-product algorithm in the factor graph is adopted, and the nonlinear function is applied. Perform Taylor expansion and linearization to optimize the posterior marginal distribution of the target z to a Gaussian distribution. The optimized objective estimation result is expressed as follows: in, , ; Introducing the forgetting factor λ > 0 is used to control the proportion of historical information retained, and is combined with the scaling factor. ρ j Implement numerical normalization, construct an iterative estimation and update mechanism, and determine the maximum number of iterations. J = 100000, no. j In this iteration, the estimation results are expressed as follows: in, , ; Finally, the convergence condition is met. Or reach the maximum number of iterations J Then, the iteration is terminated, and the final estimation result is output as follows: 。

Images