Vertical array target distance estimation system and method based on twin neural network

Abstract

The invention provides a vertical array target distance estimation system and method based on a twin neural network, and belongs to the technical field of estimation methods, and the system comprises a data collection unit which comprises an L-element hydrophone array; the preprocessing unit is in communication connection with the data acquisition unit; the sample pair construction unit is in communication connection with the preprocessing unit; the feature extraction unit is in communication connection with the sample pair construction unit, the feature extraction unit comprises a twin convolutional neural network, and the twin convolutional neural network has two branches sharing weights; the loss calculation unit is in communication connection with the feature extraction unit; the model training unit is in communication connection with the feature extraction unit and the loss calculation unit; and the distance prediction unit is in communication connection with the model training unit, and solves the problems of low precision and poor generalization caused by environment mismatch and data insufficiency of a traditional vertical array passive distance estimation method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of estimation method, and particularly relates to a vertical array target distance estimation system and method based on a twin neural network. BACKGROUND

[0002] In shallow sea and complex ocean environment, passive distance estimation of underwater targets by using vertical hydrophone array is a key link in anti-submarine warfare, submarine pipeline inspection and ocean resource investigation. Most of the traditional methods are based on the matching field processing idea, that is, an acoustic propagation model containing seabed topography, sound speed profile, and sediment layer parameters is first established, then the sound pressure data received by the array is correlated with the sound field at each candidate distance predicted by the model, and the corresponding distance of the correlation peak is taken as the estimation result. However, such model-driven methods have very high requirements for the completeness and accuracy of the environmental prior information. Once the sea depth, sound speed profile or bottom parameters are mismatched, the correlation peak will be widened, drifted or even disappeared, resulting in a sharp increase in positioning error. In actual operation, the ocean environment changes with tides, temperature and salinity fronts and internal waves, and accurate models are often difficult to obtain in real time, so the robustness of the traditional method in unknown or dynamic environment is insufficient. In addition, the matching field algorithm needs to calculate the forward sound field for a large number of candidate distance-depth grids, which has a large amount of calculation and is difficult to meet the real-time processing demand.

[0003] In recent years, deep learning technology represented by convolutional neural network has been introduced into the field of underwater acoustic positioning, which directly maps the sound pressure data to the target distance through data-driven method, avoiding the dependence on accurate environmental model. However, the conventional single tower network needs large-scale labeled data covering all working conditions to avoid overfitting; the cost of sea trial is high, and the measured data is difficult to meet this requirement. When the training set and test set come from different seasons, different sea areas or different bottom environments, the network performance often shows a cliff-like decline. SUMMARY

[0004] Therefore, the vertical array target distance estimation system and method based on the twin neural network provided by the application solve the problem of low precision and poor generalization of the traditional passive distance estimation method of vertical array due to environmental mismatch and data deficiency under the condition of unknown complex ocean environment and small sample measured data.

[0005] The application is implemented as follows:

[0006] The application provides a vertical array target distance estimation system based on a twin neural network, which comprises:

[0007] A data acquisition unit, the data acquisition unit comprises an L-element hydrophone array, and is used for synchronously receiving sound pressure signals of underwater sound sources;

[0008] The preprocessing unit is connected in communication with the data acquisition unit, and is configured to normalize the received sound pressure signal, and calculate a sound pressure covariance matrix based on the normalized signal, and output a divided training set, a verification set and a test set;

[0009] The sample pair construction unit is connected in communication with the preprocessing unit, and is configured to construct positive sample pairs and negative sample pairs from the training set and the test set respectively according to whether the sound source distances are the same.

[0010] The feature extraction unit is connected in communication with the sample pair construction unit, and includes a twin convolutional neural network having two branches sharing weights, each branch including three convolutional layers and pooling layers cascaded in sequence, for extracting high-dimensional features of two input samples respectively and mapping the high-dimensional features to the same dimensional space.

[0011] The loss calculation unit is connected in communication with the feature extraction unit, and is configured to calculate the similarity of the sample pairs in the feature space based on a contrastive loss function.

[0012] The model training unit is connected in communication with the feature extraction unit and the loss calculation unit.

[0013] The distance prediction unit is connected in communication with the model training unit, and is configured to load the trained model weights and predict the distance between the underwater sound source and the L-element hydrophone array based on the input sound pressure covariance matrix.

[0014] The technical effects of the vertical array target distance estimation system based on the twin neural network provided by the application are as follows: through the complete hardware link of "data acquisition-preprocessing-sample pair-twin network-contrastive loss-migration training-distance prediction", the sound pressure data normalization, covariance matrix construction, positive and negative sample screening, network training and reasoning that originally need human intervention are all solidified into the system unit; not only the end-to-end delay from the original sound pressure signal to the distance result is significantly shortened, but also the network parameter quantity is halved while maintaining the accuracy by using the twin convolutional structure to share the weights, thereby reducing the system power consumption and storage demand, and the system is particularly suitable for power-constrained shipborne or long-term deployment scenarios of a submarine buoy.

[0015] On the basis of the above technical solution, the vertical array target distance estimation system based on the twin neural network of the application can be further improved as follows:

[0016] The positive sample pairs constructed by the sample pair construction module are composed of two samples with the same distance but different depths, and the negative sample pairs are composed of two samples with different distances.

[0017] The beneficial effects of adopting the above-mentioned improved scheme are as follows: After clarifying the criterion of "different depths at the same distance are positive and different distances are negative" in the sample pair construction unit, the system can automatically generate a balanced number of positive and negative samples with complementary information without any manual annotation. This avoids overfitting caused by class imbalance in traditional methods, and enables the system to maintain high generalization ability even when there is only a small amount of measured data, significantly reducing the time and cost of marine calibration operations.

[0018] Furthermore, the feature extraction module uses a contrastive loss function to measure feature similarity. For positive sample pairs, the features extracted by the network are highly similar; for negative sample pairs, the features extracted by the network show significant differences.

[0019] The beneficial effects of adopting the above-mentioned improvement scheme are as follows: by directly constraining the "same kind close and different kind far away" in the feature space through comparison loss function, the network exhibits high robustness to channel multipath, noise and array position disturbances. Even when the signal-to-noise ratio drops sharply in shallow sea, it can still output reliable distance estimates stably, thereby improving the success rate of combat or detection in complex underwater acoustic environments.

[0020] Furthermore, the model training module uses a ResNet-18 network and initializes the convolutional layer weights through transfer learning, using a small number of samples for training and fine-tuning.

[0021] The beneficial effects of adopting the above-mentioned improvement scheme are as follows: by using ResNet-18 and only unfreezing the deep weights during the training phase, fine-tuning can be completed in less than 5 minutes with less than 100 small samples; transfer learning enables the system to get rid of its dependence on large-scale marine measurement data, significantly shortening the deployment cycle, while ensuring the model's ability to adapt quickly in unknown sea areas.

[0022] This invention provides a method for estimating the target distance of a vertical array based on a Siamese neural network, comprising the aforementioned system for estimating the target distance of a vertical array based on a Siamese neural network, and the method comprising the following steps:

[0023] S10: Normalize the sound pressure data received by the L-element hydrophone array, calculate the sound pressure covariance matrix based on the normalized sound pressure data, and use the sound pressure covariance matrix as a sample to divide it into training set, validation set and test set according to a preset ratio.

[0024] S20: Based on whether the distances of the underwater sound sources are consistent, construct independent sample pairs in the training set and the test set respectively. Positive sample pairs consist of two samples with the same distance but different depths, while negative sample pairs consist of two samples with different distances.

[0025] S30: Construct a feature extraction network with a convolutional neural network as its core, extract features from the input samples through the feature extraction network, map the high-dimensional features of two input samples to the same dimensional space, and judge their similarity based on the sample distance in the feature space, and use a contrastive loss function to measure feature similarity.

[0026] S40: Train convolutional neural network distance estimation model and residual neural network distance estimation model respectively based on transfer learning, and use the trained models to predict the distance between the underwater sound source and the array.

[0027] The technical effects of the vertical array target distance estimation method based on Siamese neural network provided by this invention are as follows: Steps S10-S40 of the method transform the system function into an executable process. Users only need to input the original hydrophone array sound pressure data, and the system can automatically complete normalization, matrix calculation, sample pair generation, Siamese network training and distance output. The entire process requires zero manual intervention, which greatly reduces the operation threshold and enables non-acoustic professionals to obtain accurate distance estimates within minutes.

[0028] Based on the above technical solution, the vertical array target distance estimation method based on Siamese neural network of the present invention can be further improved as follows:

[0029] Furthermore, in step S20, constructing the sample pair includes the following steps:

[0030] The water area to be measured is divided into discrete grids in the depth dimension and the distance dimension, and each grid corresponds to a sound pressure covariance matrix sample.

[0031] Using grid coordinates as an index, a distance coordinate comparator is used to pair all grid points. When the distance coordinates of two grid points are equal but their depth coordinates are unequal, the pair of samples is marked as a positive sample pair; when the distance coordinates of two grid points are unequal, the pair of samples is marked as a negative sample pair.

[0032] The labeled positive and negative sample pairs are written into the first-in-first-out buffer to form a training sequence.

[0033] The beneficial effects of adopting the above-mentioned improved scheme are as follows: through the three-step hardware design of "discrete grid-coordinate comparator-FIFO buffer", the system can complete the automatic labeling and queuing of tens of thousands of sample pairs in milliseconds, avoiding the memory jitter and CPU usage peaks of traditional script-based processing, ensuring the real-time or near-real-time processing requirements, and providing a stable and continuous data stream for subsequent neural networks.

[0034] Furthermore, in S30, the loss function of the Siamese neural network is a contrastive loss function, which is used to measure the similarity between sample pairs. For positive sample pairs, the features extracted by the network are highly similar; for negative sample pairs, the features extracted by the network show significant differences.

[0035] The beneficial effects of adopting the above-mentioned improvement scheme are as follows: the size of the convolution kernel, the number of channels and the pooling stride of each layer of the Siamese network are clearly given, so that the hardware implementation can be directly mapped to the FPGA or ASIC pipeline, the network size is fixed and the latency is controllable; the 128-dimensional embedding vector combined with L2 normalization not only compresses storage but also maintains discriminative power, and can run on embedded boards without GPU acceleration, further reducing the overall size and power consumption.

[0036] Furthermore, in S30, the convolutional neural network is a twin structure, including a first branch and a second branch. The two branches have the same structure and share weights. Each branch includes three convolutional layers and a pooling layer. Each convolutional layer is followed by a pooling layer, and finally the sound source distance is output through a regression layer.

[0037] Each branch includes, in turn:

[0038] The first convolutional layer has a kernel size of 3×3, a stride of 1, padding of 1, and 32 output channels.

[0039] The first max pooling layer has a pooling kernel size of 2×2 and a step size of 2.

[0040] The second convolutional layer has a kernel size of 3×3, a stride of 1, padding of 1, and 64 output channels.

[0041] The second max pooling layer has a pooling kernel size of 2×2 and a step size of 2.

[0042] The third convolutional layer has a kernel size of 3×3, a stride of 1, padding of 1, and 128 output channels.

[0043] The third max pooling layer has a pooling kernel size of 2×2 and a step size of 2.

[0044] The fully connected layer, with an input dimension of 2048 and an output dimension of 128, is used to map convolutional features into 128-dimensional embedding vectors.

[0045] The L2 normalization layer is used to normalize the embedding vector by a unit length.

[0046] Furthermore, in step S40, the residual neural network is a ResNet-18, and the training process includes:

[0047] Load the pre-trained ResNet-18 convolutional layer weights through read-only memory;

[0048] During the first N iterations, the weights of the first to third residual blocks of ResNet-18 are frozen, and only the fourth residual block and subsequent fully connected layers are updated, where N is a preset integer;

[0049] Starting from the N+1th iteration, unfreeze the third set of residual blocks and continue training until convergence.

[0050] The trained weights are written to non-volatile memory for use during distance prediction.

[0051] In the ResNet-18 network, the specific meaning is as follows:

[0052] The first set of residual blocks: The first set of residual blocks following the initial 7×7 convolution and 3×3 max pooling layers, usually consisting of 2 BasicBlocks, with the output feature map size being 1 / 4 of the input and 64 channels.

[0053] The second set of residual blocks consists of two BasicBlocks, with the output feature map size being 1 / 8 of the input and 128 channels.

[0054] The third set of residual blocks consists of two BasicBlocks, with the output feature map size being 1 / 16 of the input and 256 channels.

[0055] The fourth set of residual blocks consists of two BasicBlocks, with the output feature map size being 1 / 32 of the input and 512 channels.

[0056] The layered unfreezing strategy enables rapid convergence in the early stages of training and fine-tuning in the later stages, avoiding the oscillations caused by updating the entire network at once. After the final weights are solidified in the non-volatile memory, the system can immediately restore its distance prediction capability after a power outage and restart, meeting the tactical requirement of "ready to use" for shipborne or underwater moorings.

[0057] Furthermore, the training dataset consists of simulated sound field data, which is generated in the following manner:

[0058] Set up single-frequency sound sources at different distances and depths in the simulation environment;

[0059] The complex sound pressure of each sound source on the L-element hydrophone array was calculated using the normal mode model.

[0060] The sound pressure covariance matrix is ​​calculated based on complex sound pressure, and this matrix is ​​used as the input feature sample.

[0061] The generated samples are divided into training set, validation set and test set in a 7:2:1 ratio.

[0062] The beneficial effects of adopting the above-mentioned improved scheme are as follows: the sound field data can be generated in batches in the simulation environment using the normal mode model, which can cover the entire depth range from thousands to tens of kilometers, and solve the boundary conditions that are difficult to exhaustively cover by measured data in one go; the 7:2:1 automatic partitioning ensures the statistical consistency of the training-verification-test set, so that the system can achieve the design accuracy on the first deployment in the real sea area, and significantly reduce the number of on-site calibration voyages.

[0063] Compared with existing technologies, the beneficial effects of the vertical array target distance estimation system and method based on Siamese neural network provided by this invention are as follows: This invention deeply couples the contrastive learning idea of ​​Siamese neural network with the vertical array signal processing hardware system, and achieves systematic immunity to environmental mismatch, data scarcity and channel interference through the complete link of "array-preprocessing-sample pair-Siamese network-contrast loss-transfer fine-tuning-distance output". At the data level, the depth-distance discrete grid and coordinate comparator circuit automatically construct positive and negative sample pairs, generating a balanced and complementary training set without manual annotation, fundamentally suppressing the overfitting risk caused by class imbalance. At the feature level, the weighted twin convolutional structure merges two forward propagations into one, extracting discriminative features from the two sound pressure covariance matrices, and amplifying the subtle difference of "same distance - different depth" with a 128-dimensional embedding vector, providing a highly separable metric space for subsequent distance regression. At the training level, the layered unfreezing ResNet-18 transfer learning strategy requires only dozens of target sea area samples to complete model fine-tuning, compressing the number of maritime calibration voyages from "days" to "hours," significantly reducing operational costs. After system-level solidification, all computations can be completed in real time on low-power FPGAs or embedded GPU boards, with a single-frame latency of less than 100 milliseconds, meeting the online ranging requirements of unmanned underwater vehicles, AUVs, and shipborne sonars. Attached Figure Description

[0064] Figure 1 A schematic diagram illustrating the construction of positive and negative sample pairs in the problem of underwater sound source ranging;

[0065] Figure 2 This is a schematic diagram of feature extraction based on Siamese neural networks;

[0066] Figure 3 This is a ResNet-18 distance estimation model based on Siamese network feature extraction;

[0067] Figure 4 The SWellEX-96 experimental simulation environment;

[0068] Figure 5 A flowchart of the model training process;

[0069] Figure 6To evaluate the distance estimation errors of different models, (a) S-CNN method; (b) NS-CNN method; (c) S-ResNet method; (d) NS-ResNet method;

[0070] Figure 7 Visualization results of feature extraction from Siamese neural networks; (a) 2000 iterations; (b) 4000 iterations; (c) 6000 iterations; (d) 8000 iterations; (e) 10000 iterations; (f) 12000 iterations;

[0071] Figure 8 To assess the distance estimation accuracy of Siamese neural networks at different iteration numbers; (a) S-CNN method; (b) S-ResNet method;

[0072] Figure 9 Statistical results of distance estimation for training model under different signal-to-noise ratios on a noisy test set; (a) PCL of model ranging under different signal-to-noise ratios; (b) MAPE of model ranging under different signal-to-noise ratios;

[0073] Figure 10 This is a flowchart of a vertical array target distance estimation system based on a Siamese neural network. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0075] like Figure 10 The diagram shows a flowchart of the vertical array target distance estimation system based on Siamese neural network provided by the present invention, including:

[0076] The data acquisition unit includes an L-element hydrophone array for synchronously receiving sound pressure signals from underwater sound sources.

[0077] The preprocessing unit, which is connected in communication with the data acquisition unit, is used to normalize the received sound pressure signal, calculate the sound pressure covariance matrix based on the normalized signal, and output the divided training set, validation set and test set.

[0078] The sample pair construction unit, which is connected in communication with the preprocessing unit, is used to construct positive sample pairs and negative sample pairs from the training set and the test set respectively, based on whether the distances of the sound sources are the same or not.

[0079] The feature extraction unit is communicatively connected to the sample pair construction unit. The feature extraction unit includes a Siamese convolutional neural network with two branches sharing weights. Each branch includes three convolutional layers and pooling layers cascaded in sequence to extract high-dimensional features from the two input samples and map the high-dimensional features to the same dimensional space.

[0080] The loss calculation unit, which is connected in communication with the feature extraction unit, is used to calculate the similarity of sample pairs in the feature space based on the contrastive loss function.

[0081] The model training unit is communicatively connected to the feature extraction unit and the loss calculation unit.

[0082] The distance prediction unit communicates with the model training unit and is used to load the trained model weights and predict the distance between the underwater sound source and the L-element hydrophone array based on the input sound pressure covariance matrix.

[0083] The following is a detailed description of the specific implementation method of the present invention, "A Vertical Array Target Range Estimation Method Based on Siamese Neural Networks". The implementation process is divided into five stages, from on-site data acquisition to final distance output. All steps can be completed by a single person carrying portable equipment on a deck, shore station, or unmanned underwater vehicle, without the need for additional large-scale experimental support.

[0084] Phase 1: Array Deployment and Raw Data Acquisition

[0085] First, based on the depth of the mission area and the estimated target distance, a 21-element vertical hydrophone array was selected. The array elements were suspended from top to bottom on a rigid cable with equal or unequal spacing. After the cable was lowered to the predetermined depth range by a winch, the GPS / BeiDou synchronization clock in the deck unit was used to time all array elements, ensuring phase consistency during subsequent covariance matrix calculations. Subsequently, a multi-channel synchronous acquisition card was activated to continuously record the target radiated noise or active acoustic pulse echo for 30 seconds at a sampling rate of 25kHz, forming the raw time-domain data file.

[0086] Phase Two: Data Preprocessing and Feature Matrix Construction

[0087] The original time-domain data was transformed to the frequency domain using FFT, and complex sound pressure levels within a ±5Hz bandwidth of the dominant frequency were selected. Then, norm normalization was performed on each complex sound pressure level to eliminate sensitivity differences between different array elements. After normalization, the real and imaginary parts of the 21 complex sound pressure levels were concatenated into a 21×21×2 three-dimensional feature matrix, which served as the input sample for subsequent networks. This matrix physically preserves the target range-dependent multipath interference structure while suppressing interference caused by amplitude fluctuations.

[0088] Phase 3: Automatic generation of sample pairs:

[0089] In the ground station computer, a discrete grid is established with the depth of the area to be measured ranging from 5m to 200m and the distance from 1km to 11km as boundaries. All samples are paired using a grid coordinate comparator: if two samples have the same distance coordinate but different depth coordinates, they are marked as a positive sample pair; if the distance coordinates are different, they are marked as a negative sample pair. Positive and negative sample pairs are written into the first-in-first-out (FIFO) buffers of the training, validation, and test sets in a 7:2:1 ratio, forming an automatic training sequence that requires no manual annotation.

[0090] Phase 4: Training the Siamese Neural Network

[0091] The training program was run on a portable GPU workstation. The network adopted a Siamese architecture, with the two branches sharing weights. Each branch contained three convolutional-pooling cascades: 3×3 convolutional kernels with 1 padding and 1 stride, and the number of channels were 32, 64, and 128 respectively; the pooling kernels were 2×2 with a stride of 2. A fully connected layer followed the convolutions to output a 128-dimensional embedding vector, which was then normalized using L2. A contrastive loss function was used for optimization, with a margin of 0.8, a batch size of 120, an initial learning rate of 0.0001, and a maximum of 12,000 iterations. During training, the embedding space distribution was monitored in real-time using TensorBoard to ensure that similar samples clustered together and dissimilar samples were separated.

[0092] Phase 5: ResNet-18 Transfer Fine-tuning and Distance Output

[0093] The trained Siamese convolution weights are transferred to the front end of the ResNet-18, replacing its 7×7 convolutions and the first three residual blocks. Then, a small sample set (usually less than 100 samples) collected in the second stage is loaded. In the first N iterations, only the fourth residual block and the fully connected layer of the ResNet-18 are unfrozen; after N iterations, the third residual block is unfrozen for further fine-tuning until convergence. Finally, the single neuron at the network's end directly outputs the distance estimate. After the system is solidified, only a new sound pressure covariance matrix needs to be input in the field to complete the entire process from feature extraction to distance output within 1 second, without further manual intervention.

[0094] In the above technical solution, the positive sample pair constructed by the sample pair construction module consists of two samples that are equidistant but have different depths, while the negative sample pair consists of two samples that are equidistant.

[0095] In the above technical solution, the feature extraction module uses a contrastive loss function to measure feature similarity. For positive sample pairs, the features extracted by the network are highly similar; for negative sample pairs, the features extracted by the network show significant differences.

[0096] In the above technical solution, the model training module uses the ResNet-18 network and initializes the convolutional layer weights through transfer learning, using a small number of samples for training and fine-tuning.

[0097] The target distance estimation method based on the vertical array of Siamese neural networks includes the following steps:

[0098] S10: Normalize the sound pressure data received by the L-element hydrophone array, calculate the sound pressure covariance matrix based on the normalized sound pressure data, and use the sound pressure covariance matrix as a sample to divide it into training set, validation set and test set according to a preset ratio.

[0099] S20: Based on whether the distances of the underwater sound sources are consistent, construct independent sample pairs in the training set and the test set respectively. Positive sample pairs consist of two samples with the same distance but different depths, while negative sample pairs consist of two samples with different distances.

[0100] S30: Construct a feature extraction network with a convolutional neural network as the core, extract features from input samples through the feature extraction network, map the high-dimensional features of two input samples to the same dimensional space, and judge their similarity based on the sample distance in the feature space. Use the contrastive loss function to measure feature similarity.

[0101] S40: Train convolutional neural network distance estimation model and residual neural network distance estimation model respectively based on transfer learning, and use the trained models to predict the distance between the underwater sound source and the array.

[0102] The technical solution of the present invention includes the following specific steps:

[0103] Step 1: Data preprocessing and feature construction.

[0104] Let L be the number of elements in the vertical array. The time-domain data received by the array is transformed to the frequency domain using a Fast Fourier Transform (FFT) to obtain the frequency-domain complex sound pressure level (SPL). Norm normalization is then performed on the received frequency-domain SPL data to eliminate the influence of differences in data units and value ranges on model training. The specific formula is as follows:

[0105] ;

[0106] in, For the first The hydrophone received a frequency of The complex sound pressure at that time, This is the normalized complex sound pressure.

[0107] The normalized complex sound pressure is separated into real and imaginary parts to form the initial feature matrix of the network input. Specifically, it is expressed as:

[0108] ;

[0109] in, The dimension is , This represents the number of frequency points to select.

[0110] Step 2: Design and feature extraction of the twin neural network structure.

[0111] This invention uses a Siamese neural network for feature extraction. Siamese neural networks typically process input data in pairs and optimize network weights by comparing the similarity between two samples.

[0112] First, based on whether the distances to the underwater sound sources are consistent, independent sample pairs are randomly constructed in both the training and test sets. The specific method for constructing the sample pairs is as follows:

[0113] Let the set of values ​​for the depth dimension of the two-dimensional region to be measured be . The set of values ​​after discretization of the distance dimension is: .

[0114] A single grid point can be represented as a coordinate tuple. ,in , The grid points corresponding to the samples in the training / test set constitute a set. (Right now It is all possibilities (The actual set of points containing samples).

[0115] Define the set of positive sample pairs : That is, when two grid points are equidistant but at different depths, these two samples form a positive sample pair. This represents the Cartesian product of all possible sample pairs in the sample set S. Figure 1 The two grid points marked with red dashed boxes are typical positive sample pairs.

[0116] Define the set of negative sample pairs : That is, if two grid points are at different distances, regardless of whether their depths are the same, these two samples constitute a negative sample pair. Figure 1 The two grid points marked with blue dashed boxes are typical negative sample pairs.

[0117] Let M be the number of grid points in the depth direction (x-axis direction) and N be the number of grid points in the distance direction (y-axis direction). Then, the sample pairs constructed in the above manner can form positive sample pairs. Yes, negative sample pairs Yes, in total This significantly expands the training sample size, effectively solving the problem of model overfitting caused by insufficient single-class samples, thereby improving the model's generalization ability and the robustness of feature extraction.

[0118] A feature extraction network based on a CNN is constructed. This feature extraction network consists of two sub-networks with shared weights. This network can process two inputs simultaneously (i.e., two samples in an input sample pair). Features are extracted from the input samples using the CNN, and a fully connected structure maps the high-dimensional features of the two input samples to the same dimensional space. Similarity is determined by comparing the distance between samples in the feature space. The Siamese neural network feature extraction structure designed in this paper is as follows: Figure 2 As shown, a contrastive loss function is used to measure feature similarity.

[0119] Step 3: Use a contrastive loss function to measure feature similarity. The contrastive loss function is as follows:

[0120] ;

[0121] in, This indicates that for data pairs with labels, if two samples are close in distance (belonging to the same distance interval), then... ,otherwise . Euclidean distance: Margin is a set hyperparameter that controls the minimum interval between dissimilar samples in the feature space. This loss function can describe the degree of matching between sample pairs well.

[0122] When two samples are similar or match, loss function If the Euclidean distance of the extracted features is large, the network will have a large output loss, indicating that the model needs further training.

[0123] When two samples are not similar or do not match loss function If the Euclidean distance of the extracted features is small, the final output loss of the network will be large, indicating that the model needs further training.

[0124] The model trained using the above iterative optimization method can automatically determine whether input sample pairs have the same sound source distance label.

[0125] In the above technical solution, step S20, constructing the sample pair includes the following steps:

[0126] The water area to be measured is divided into discrete grids in the depth dimension and the distance dimension, and each grid corresponds to a sound pressure covariance matrix sample.

[0127] Using grid coordinates as an index, a distance coordinate comparator is used to pair all grid points. When the distance coordinates of two grid points are equal but their depth coordinates are unequal, the pair of samples is marked as a positive sample pair; when the distance coordinates of two grid points are unequal, the pair of samples is marked as a negative sample pair.

[0128] The labeled positive and negative sample pairs are written into the first-in-first-out buffer to form a training sequence.

[0129] In the above technical solution, in S30, the loss function of the Siamese neural network is a contrastive loss function, which is used to measure the similarity between sample pairs. For positive sample pairs, the features extracted by the network are highly similar; for negative sample pairs, the features extracted by the network show significant differences.

[0130] In the above technical solution, in S30, the convolutional neural network is a twin structure, including a first branch and a second branch. The two branches have the same structure and share weights. Each branch includes three convolutional layers and a pooling layer. Each convolutional layer is followed by a pooling layer, and finally the sound source distance is output through a regression layer.

[0131] In the above technical solution, the residual neural network in S40 is ResNet-18, and the training process includes:

[0132] Load the pre-trained ResNet-18 convolutional layer weights through read-only memory;

[0133] During the first N iterations, the weights of the first to third residual blocks of ResNet-18 are frozen, and only the fourth residual block and subsequent fully connected layers are updated, where N is a preset integer;

[0134] Starting from the N+1th iteration, unfreeze the third set of residual blocks and continue training until convergence.

[0135] The trained weights are written to non-volatile memory for use during distance prediction.

[0136] In the above technical solution, the training dataset consists of simulated sound field data, which is generated in the following way:

[0137] Set up single-frequency sound sources at different distances and depths in the simulation environment;

[0138] The complex sound pressure of each sound source on the L-element hydrophone array was calculated using the normal mode model.

[0139] The sound pressure covariance matrix is ​​calculated based on complex sound pressure, and this matrix is ​​used as the input feature sample.

[0140] The generated samples are divided into training set, validation set and test set in a 7:2:1 ratio.

[0141] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below. The technical solution of the present invention includes the following specific steps:

[0142] Step 1: Data preprocessing and feature construction.

[0143] Let L be the number of elements in the vertical array. The time-domain data received by the array is transformed to the frequency domain using a Fast Fourier Transform (FFT) to obtain the frequency-domain complex sound pressure level (SPL). Norm normalization is then performed on the received frequency-domain SPL data to eliminate the influence of differences in data units and value ranges on model training. The specific formula is as follows:

[0144] ;

[0145] in, For the first The hydrophone received a frequency of The complex sound pressure at that time, This is the normalized complex sound pressure.

[0146] The normalized complex sound pressure is separated into real and imaginary parts to form the initial feature matrix of the network input. Specifically, it is expressed as:

[0147] ;

[0148] in, The dimension is , This represents the number of frequency points to select.

[0149] Step 2: Design and feature extraction of the twin neural network structure.

[0150] This invention uses a Siamese neural network for feature extraction. Siamese neural networks typically process input data in pairs and optimize network weights by comparing the similarity between two samples.

[0151] First, based on whether the distances to the underwater sound sources are consistent, independent sample pairs are randomly constructed in both the training and test sets. The specific method for constructing the sample pairs is as follows:

[0152] Let the set of values ​​for the depth dimension of the two-dimensional region to be measured be . The set of values ​​after discretization of the distance dimension is: .

[0153] A single grid point can be represented as a coordinate tuple. ,in , The grid points corresponding to the samples in the training / test set constitute a set. (Right now It is all possibilities (The actual set of points containing samples) is used to construct sample pairs as follows: Figure 1 As shown.

[0154] Define the set of positive sample pairs : That is, when two grid points are equidistant but at different depths, these two samples form a positive sample pair. This represents the Cartesian product of all possible sample pairs in the sample set S. Figure 1 The two grid points marked with red dashed boxes are typical positive sample pairs.

[0155] Define the set of negative sample pairs : That is, if two grid points are at different distances, regardless of whether their depths are the same, these two samples constitute a negative sample pair. Figure 1 The two grid points marked with blue dashed boxes are typical negative sample pairs.

[0156] Let M be the number of grid points in the depth direction (x-axis direction) and N be the number of grid points in the distance direction (y-axis direction). Then, the sample pairs constructed in the above manner can form positive sample pairs. Yes, negative sample pairs Yes, in total This significantly expands the training sample size, effectively solving the problem of model overfitting caused by insufficient single-class samples, thereby improving the model's generalization ability and the robustness of feature extraction.

[0157] A feature extraction network based on a CNN is constructed. This feature extraction network consists of two sub-networks with shared weights. This network can process two inputs simultaneously (i.e., two samples in an input sample pair). Features are extracted from the input samples using the CNN, and a fully connected structure maps the high-dimensional features of the two input samples to the same dimensional space. Similarity is determined by comparing the distance between samples in the feature space. The Siamese neural network feature extraction structure designed in this paper is as follows: Figure 2 As shown.

[0158] Step 3: Use a contrastive loss function to measure feature similarity. The contrastive loss function is as follows:

[0159] ;

[0160] in, This indicates that for data pairs with labels, if two samples are close in distance (belonging to the same distance interval), then... ,otherwise . Euclidean distance: Margin is a set hyperparameter that controls the minimum interval between dissimilar samples in the feature space. This loss function can describe the degree of matching between sample pairs well.

[0161] When two samples are similar or match, loss function If the Euclidean distance of the extracted features is large, the network will have a large output loss, indicating that the model needs further training.

[0162] When two samples are not similar or do not match loss function If the Euclidean distance of the extracted features is small, the final output loss of the network will be large, indicating that the model needs further training.

[0163] The model trained using the above iterative optimization method can automatically determine whether input sample pairs have the same sound source distance label.

[0164] During training, the network dynamically adjusts the feature extraction weights to enable the model to focus on learning the key discriminative features for the sound source distance estimation task. This mechanism makes the feature representation extracted by the network strongly correlated with the sound source distance, while also helping to suppress interference factors. This enables the model to decouple and predict sound source location information in the sound field. The basic parameters of the feature extraction network are shown in Table 1.

[0165] Table 1. Parameter settings for the feature extraction model based on Siamese neural network

[0166] Step 4: Establish a ResNet-18 distance estimation model based on Siamese network feature extraction.

[0167] like Figure 3 As shown, the ResNet-18 network structure is connected based on the feature extraction results. Features are integrated through two fully connected layers, and finally, the distance between the sound source and the vertical matrix is ​​directly output through a fully connected layer containing a single neuron. In addition to the feature extraction network, the ResNet-18 model contains one independent 7×7 convolutional layer and eight residual blocks, each containing two 3×3 convolutional layers. The residual blocks are connected via skip connections to form an "identity mapping." The fully connected layers have 512, 256, and 1 neuron respectively. The overall structure of the ResNet distance estimation model is as follows: Figure 2 As shown.

[0168] The specific implementation process is as follows:

[0169] First, a feature extraction model based on a Siamese neural network is trained, which mainly includes the following steps:

[0170] (1) Randomly construct independent sample pairs in the training set and the test set respectively. Two samples with the same distance label form a sample pair and form a new label "1". Two samples with different distance labels form a sample pair and form a new label "0".

[0171] (2) Construct the model gradient function. In this study, the hyperparameter margin is set to 0.8.

[0172] (3) Setting training parameters: A smaller batch size can speed up the update frequency, but the computational cost may be higher. Therefore, considering the hardware level of this study, the training sample batch size is set to 120, the gradient decay coefficient is set to 0.9, and the maximum number of training iterations is specified. Since the network only determines the termination condition by the maximum number of iterations, the learning rate can be set to a small value to avoid the training not converging due to the model parameters being too large. In this experiment, the learning rate is set to 0.0001.

[0173] After training the feature extraction model based on Siamese neural networks, the CNN distance estimation model and the ResNet distance estimation model are trained respectively using transfer learning. This mainly includes the following steps:

[0174] (1) The convolutional layer parameters of the Siamese neural network are directly transferred to the CNN distance estimation model and the ResNet-18 distance estimation model as feature extraction structures;

[0175] (2) The distance estimation model was trained using the training set. In this experiment, the initial learning rate was set to 0.001 and the learning rate decay factor was 0.1. The learning rate in the feature extraction structure was set to 10% of that in the other network layers to achieve the effect of parameter fine-tuning. The Adam optimization algorithm was used in the overall training process, and the batch size of the training samples was set to 120.

[0176] This study generates simulation data based on the environment model provided in the SWellEX-96 experiment; detailed experimental information is provided in Section 4. The simulation environment is as follows: Figure 5 As shown, assuming the water layer depth is 216.5 meters and the density is 1... Below the water layer is a 23.5m thick sedimentary layer with a density of 1.76. The attenuation is 0.2 dB / kmHz. The sound velocities at the top and bottom are 1572.3 m / s and 1593.0 m / s, respectively. Below the sedimentary layer is an 800 m thick layer of mudstone with a density of 2.061. The attenuation is 0.06 dB / kmHz. The sound velocities at the top and bottom are 1881 m / s and 3245 m / s, respectively. The seabed has a density of 2.661... The SWellEX-96 experimental simulation environment consists of a half-space with an attenuation of 0.02 dB / kmHz and a sound velocity of 5200 m / s. Figure 4 As shown.

[0177] The sound pressure received by the hydrophone array in this environment was simulated using KRAKEN. For Sound pressure received by the hydrophone array Normalize according to the following formula:

[0178] ;

[0179] Calculate the normalized sound pressure covariance matrix as a sample:

[0180] ;

[0181] The simulation frequency in this study was set to 112Hz. The hydrophone vertical receiving array has a total of 21 array elements. The array element depth is set according to the parameters given in the SWellEX-96 experiment, as shown in Table 2.

[0182] Table 2. Vertical receiver array element depth

[0183] The sound source depth was sampled at 5m intervals, ranging from 5m to 200m, for a total of 40 groups. The array distance from the sound source was sampled at 10m intervals, ranging from 1km to 11km, for a total of 1001 groups. The simulation dataset contains 40 × 1001 = 40040 samples. Since the sound pressure covariance is a complex matrix, the real and imaginary parts were separated during the dataset construction phase, resulting in each sample having a dimension of 21 × 21 × 2.

[0184] This study randomly partitioned the training, validation, and test sets in a 7:2:1 ratio. The training set contained 28,028 samples, the validation set contained 8,008 samples, and the test set contained 4,004 samples. It should be noted that, except for analyzing the accuracy of model distance estimation under different signal-to-noise ratios, both model training and evaluation were performed on a noise-free dataset. Model training consisted of two independent stages. The first stage involved training a Siamese neural network on a benchmark training dataset for feature extraction, such as... Figure 7 As shown, in the second stage, the two types of models constructed in the study are trained independently using datasets with noise or environmental changes. The parameters of the feature extraction network are directly transferred from the training results of the first stage, and only parameter fine-tuning is performed during the second stage training. The specific process is as follows: Figure 5 As shown.

[0185] After training, the model is evaluated using a test set.

[0186] Two objective metrics, Mean Absolute Percentage Error (MAPE) and Probability of Credible Localization (PCL), are used to measure the distance estimation performance of different methods. The Mean Absolute Percentage Error is calculated as follows:

[0187] ;

[0188] in Represents the number of test samples. This is the predicted distance value for the sample. This represents the true distance between the samples.

[0189] The confidence probability (PCL) for distance estimation is calculated as follows:

[0190] ;

[0191] ;

[0192] in Represents the number of test samples. This is the predicted distance value for the sample. This represents the true distance of the sample. The confidence probability (PCL) of the distance estimate reflects the sample situation where the distance estimation error is within 10%.

[0193] The contrastive loss function is:

[0194] .

[0195] Distance estimation errors of different methods, such as Figure 6 , Figure 8 As shown in Table 3, where MAE represents the mean absolute error, the accuracy assessment of distance estimation using different methods is presented in Table 3.

[0196] Table 3. Accuracy Assessment of Distance Estimation by Different Methods

[0197] Wherein, S-CNN and S-ResNet represent the CNN distance estimation model (S-CNN) and the ResNet distance estimation model (S-ResNet) based on Siamese neural network feature extraction; NS-CNN and NS-ResNet represent the CNN distance estimation model (NS-CNN) and the ResNet distance estimation model (NS-ResNet) trained directly on the training set without using Siamese neural network feature extraction; MFP represents the matching field processing model; the statistical results of distance estimation of the trained model under different signal-to-noise ratios on a noisy test set are as follows: Figure 9 As shown.

[0198] Specifically, the principle of this invention is as follows: The core of this invention lies in the three-in-one technical architecture of "contrast loss function - embedding mapping - transfer fine-tuning". First, the complex sound pressure output from the vertical array is normalized by FFT and norm, and then concatenated into a sound pressure covariance matrix with joint real and imaginary parts. This matrix physically contains the coupling information of sound source distance-depth and multipath interference. The two branches of the Siamese network are concatenated with three convolution-pooling operations on this matrix with shared weights, gradually compressing the high-dimensional space-frequency information into a low-dimensional embedding vector. The shared weights ensure that the two signals undergo the same nonlinear transformation, thereby eliminating the interference of network parameter differences on the distance metric. Subsequently, the contrast loss function imposes geometric constraints on the embedding space: the vector distance of positive sample pairs is shortened, and the vector distance of negative sample pairs is pushed far apart while maintaining the minimum interval. This geometric constraint is equivalent to maximizing the inter-class distance and minimizing the intra-class distance under the implicitly defined distance kernel function, so that the network automatically focuses on the interference structure that changes monotonically with distance, rather than the absolute amplitude that changes drastically with the environment. To further address the "small sample size" challenge, the system integrates a ResNet-18 residual network after the Siamese network, and transfers the Siamese convolutional layer weights as pre-trained parameters to the ResNet-18 front end. During initial training, shallow residual blocks are frozen, with only deep and fully connected layers fine-tuned. This preserves a general representation of cross-environment shared features while rapidly adapting to local sound field characteristics using limited target domain data. Finally, the ResNet-18 output layer is directly mapped to a distance scalar, achieving end-to-end ranging. Mathematically, this architecture is equivalent to the joint optimization of "environment-independent feature extraction + task-specific regression," enabling the entire system to combine the approximation capabilities of deep networks with the metric robustness of Siamese structures, thus achieving highly robust and highly generalized passive distance estimation of vertical matrices.

[0199] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A target distance estimation system for a vertical array based on a Siamese neural network, characterized in that, include: The data acquisition unit includes an L-element hydrophone array for synchronously receiving sound pressure signals from underwater sound sources. The preprocessing unit, which is communicatively connected to the data acquisition unit, is used to normalize the received sound pressure signal, calculate the sound pressure covariance matrix based on the normalized signal, and output the divided training set, validation set, and test set. The sample pair construction unit, which is communicatively connected to the preprocessing unit, is used to construct positive sample pairs and negative sample pairs from the training set and the test set respectively, based on whether the distances between the sound sources are the same or not. The feature extraction unit is communicatively connected to the sample pair construction unit. The feature extraction unit includes a Siamese convolutional neural network with two branches sharing weights. Each branch includes three convolutional layers and pooling layers cascaded in sequence, used to extract high-dimensional features of two input samples respectively and map the high-dimensional features to the same dimensional space. The loss calculation unit is communicatively connected to the feature extraction unit and is used to calculate the similarity of sample pairs in the feature space based on the contrastive loss function. The model training unit is communicatively connected to the feature extraction unit and the loss calculation unit. The distance prediction unit is communicatively connected to the model training unit. It is used to load the trained model weights and predict the distance between the underwater sound source and the L-element hydrophone array based on the input sound pressure covariance matrix.

2. The vertical array target distance estimation system based on Siamese neural network according to claim 1, characterized in that, The positive sample pairs constructed by the sample pair construction module consist of two samples that are equidistant but have different depths, while the negative sample pairs consist of two samples that are equidistant.

3. The vertical array target distance estimation system based on Siamese neural network according to claim 2, characterized in that, The feature extraction module uses a contrastive loss function to measure feature similarity. For positive sample pairs, the features extracted by the network are highly similar; for negative sample pairs, the features extracted by the network show significant differences.

4. The vertical array target distance estimation system based on Siamese neural network according to claim 3, characterized in that, The model training module uses a ResNet-18 network and initializes the convolutional layer weights through transfer learning, using a small number of samples for training and fine-tuning.

5. A method for estimating the target distance of a vertical array based on a Siamese neural network, characterized in that, Including the vertical array target distance estimation system based on Siamese neural networks as described in any one of claims 1-4, the method includes the following steps: S10: Normalize the sound pressure data received by the L-element hydrophone array, calculate the sound pressure covariance matrix based on the normalized sound pressure data, and use the sound pressure covariance matrix as a sample to divide it into training set, validation set and test set according to a preset ratio. S20: Based on whether the distances of the underwater sound sources are consistent, construct independent sample pairs in the training set and the test set respectively. Positive sample pairs consist of two samples with the same distance but different depths, while negative sample pairs consist of two samples with different distances. S30: Construct a feature extraction network with a convolutional neural network as its core, extract features from the input samples through the feature extraction network, map the high-dimensional features of two input samples to the same dimensional space, and judge their similarity based on the sample distance in the feature space, and use a contrastive loss function to measure feature similarity. S40: Train convolutional neural network distance estimation model and residual neural network distance estimation model respectively based on transfer learning, and use the trained models to predict the distance between the underwater sound source and the array.

6. The method for estimating the target distance of a vertical array based on a Siamese neural network according to claim 5, characterized in that, In step S20, constructing the sample pair includes the following steps: The water area to be measured is divided into discrete grids in the depth dimension and the distance dimension, and each grid corresponds to a sound pressure covariance matrix sample. Using grid coordinates as an index, a distance coordinate comparator is used to pair all grid points. When the distance coordinates of two grid points are equal but their depth coordinates are unequal, the pair of samples is marked as a positive sample pair; when the distance coordinates of two grid points are unequal, the pair of samples is marked as a negative sample pair. The labeled positive and negative sample pairs are written into the first-in-first-out buffer to form a training sequence.

7. The method for estimating the target distance of a vertical array based on a Siamese neural network according to claim 6, characterized in that, In step S30, the loss function of the Siamese neural network is a contrastive loss function, which is used to measure the similarity between sample pairs. For positive sample pairs, the features extracted by the network are highly similar; for negative sample pairs, the features extracted by the network show significant differences.

8. The target distance estimation method based on Siamese neural network according to claim 7, characterized in that, In S30, the convolutional neural network is a twin structure, including a first branch and a second branch. The two branches have the same structure and share weights. Each branch includes three convolutional layers and a pooling layer. Each convolutional layer is followed by a pooling layer, and finally the sound source distance is output through a regression layer.

9. The method for estimating the target distance of a vertical array based on a Siamese neural network according to claim 8, characterized in that, In step S40, the residual neural network is a ResNet-18, and the training process includes: Load the pre-trained ResNet-18 convolutional layer weights through read-only memory; During the first N iterations, the weights of the first to third residual blocks of ResNet-18 are frozen, and only the fourth residual block and subsequent fully connected layers are updated, where N is a preset integer; Starting from the N+1th iteration, unfreeze the third set of residual blocks and continue training until convergence. The trained weights are written to non-volatile memory for use during distance prediction.

10. The method for estimating the target distance of a vertical array based on a Siamese neural network according to claim 9, characterized in that, The training dataset consists of simulated sound field data, which is generated in the following manner: Set up single-frequency sound sources at different distances and depths in the simulation environment; The complex sound pressure of each sound source on the L-element hydrophone array was calculated using the normal mode model. The sound pressure covariance matrix is ​​calculated based on complex sound pressure, and this matrix is ​​used as the input feature sample. The generated samples are divided into training set, validation set and test set in a 7:2:1 ratio.

Images