Method and device for identifying target azimuth based on phase space reconstruction of active sonar

Abstract

The application relates to a method and device for identifying the target direction of an active sonar based on phase space reconstruction. The method comprises the following steps: obtaining a detection echo sample, extracting the recurrent graph feature, the laminar flow feature, the time domain feature and the time-frequency feature corresponding to each angle sequence of the detection echo sample; constructing a multi-feature fusion network, which comprises a sequence processing module, an image processing module, a multi-feature fusion module and a target direction identification module; training the multi-feature fusion network; obtaining a detection echo to be detected, inputting the recurrent graph feature, the laminar flow feature, the time domain feature and the time-frequency feature corresponding to each angle sequence of the detection echo to be detected into the trained multi-feature fusion network, and obtaining the target direction identification result of the detection echo to be detected. The method can improve the accuracy of active sonar detection in a low signal-to-noise ratio environment.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic target recognition technology, and in particular to a method and apparatus for active sonar target location recognition based on phase space reconstruction. Background Technology

[0002] Underwater target sonar detection includes active sonar detection and passive sonar detection. Due to the development of target stealth technology, the effectiveness of passive sonar detection has been limited, so active sonar detection has become one of the focuses of current research.

[0003] In the field of active sonar detection, beamforming combined with matched filtering is commonly used. This method can maximize the output signal-to-noise ratio (SNR) when the detected waveform is known. However, in underwater environments, various interferences exist, such as ambient noise, ship noise, and multiple reflections from the water surface and seabed, causing distortion in the underwater acoustic channel. These interference factors reduce the SNR of the received detection echo signal, exhibiting nonlinear characteristics, thus posing a challenge to the performance of active sonar detection. For detection echoes with low SNR, it is difficult to accurately extract target signal features, thereby affecting the accurate determination of target direction and position information. Therefore, considering the nonlinearity of the echo, studying the nonlinear dynamic characteristics of the detection echo and analyzing its internal patterns is of greater significance for target feature extraction. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and apparatus for active sonar target location identification based on phase space reconstruction to address the aforementioned technical problems.

[0005] A method for identifying the azimuth of an active sonar target based on phase space reconstruction, the method comprising:

[0006] Acquire probe echo samples, and extract recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the probe echo samples; the probe echo samples are labeled.

[0007] A multi-feature fusion network is constructed, comprising a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation recognition module. The sequence processing module processes the preprocessed temporal features and laminar flow features, respectively, and outputs a first classification result and a second classification result. The image processing module processes the preprocessed recursive graph features and time-frequency features, respectively, and outputs a third classification result and a fourth classification result. The multi-feature fusion module fuses the first, second, third, and fourth classification results to obtain a fusion result. The target orientation recognition module obtains the target orientation recognition result corresponding to the detected echo sample based on the fusion result.

[0008] The multi-feature fusion network is trained using the detected echo samples and a pre-set loss function to obtain a trained multi-feature fusion network.

[0009] The target location identification result of the target detection echo is obtained by acquiring the recursive graph features, laminar flow features, time domain features and time frequency features corresponding to each angle sequence of the target detection echo.

[0010] In one embodiment, the system further includes: the sequence processing module includes two convolutional neural networks with the same structure, and the image processing module includes two long short-term memory neural networks with the same structure; the convolutional neural network includes ResNet50.

[0011] In one embodiment, the loss function is further defined as:

[0012]

[0013] Where Loss represents the loss function, N represents the number of probe echo samples taken in each training iteration, and y i Let x represent the true label of the i-th probe echo sample. WVDi x represents the fourth classification result corresponding to the i-th probe echo sample. RPi Let x represent the third classification result corresponding to the i-th probe echo sample. AFi Let x represent the first classification result corresponding to the i-th probe echo sample. RLi α1 represents the second classification result corresponding to the i-th probe echo sample, and α2, α3 and α4 represent weighting coefficients.

[0014] In one embodiment, the method further includes: acquiring the echo beam of the probe echo sample, reconstructing the phase space of each angle sequence of the echo beam, and determining whether the currently reconstructed angle sequence has nonlinear characteristics; if the echo beam has nonlinear characteristics, then using a recursive matrix analysis method to calculate the repeatability of different state points in the phase space to obtain a recursive graph corresponding to each angle sequence, and performing quantitative calculations based on the recursive graph to obtain laminar flow characteristics.

[0015] In one embodiment, the method further includes: obtaining a target identification result corresponding to the probe echo sample based on the fusion result; if the target identification result indicates the presence of a target, then drawing an azimuth history map based on the laminar flow characteristics of each angle sequence corresponding to the probe echo sample, and determining the target azimuth based on the azimuth history map.

[0016] In one embodiment, the method further includes: obtaining a time period in which the laminar flow characteristic is greater than a threshold, taking the starting segment of the time period as the target point, and obtaining a location history map.

[0017] In one embodiment, the method further includes: performing preprocessing operations on the recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the probe echo sample; the preprocessing operations include data alignment operations, discretization operations, and type conversion operations.

[0018] An active sonar target location identification device based on phase space reconstruction, the device comprising:

[0019] The feature extraction module is used to acquire probe echo samples and extract recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the probe echo samples; the probe echo samples are labeled.

[0020] A network construction module is used to construct a multi-feature fusion network. The multi-feature fusion network includes a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation recognition module. The sequence processing module processes the preprocessed temporal features and laminar features respectively, outputting a first classification result and a second classification result. The image processing module processes the preprocessed recursive graph features and time-frequency features respectively, outputting a third classification result and a fourth classification result. The multi-feature fusion module fuses the first classification result, the second classification result, the third classification result, and the fourth classification result to obtain a fusion result. The target orientation recognition module obtains the target orientation recognition result corresponding to the detected echo sample based on the fusion result.

[0021] The network training module is used to train the multi-feature fusion network using the probe echo samples and a pre-set loss function to obtain the trained multi-feature fusion network.

[0022] The output module is used to acquire the detection echo to be detected. It inputs the recursive graph features, laminar flow features, time-domain features and time-frequency features corresponding to each angle sequence of the detection echo to the trained multi-feature fusion network to obtain the target orientation recognition result of the detection echo.

[0023] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0024] Acquire probe echo samples, and extract recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the probe echo samples; the probe echo samples are labeled.

[0025] A multi-feature fusion network is constructed, comprising a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation recognition module. The sequence processing module processes the preprocessed temporal features and laminar flow features, respectively, and outputs a first classification result and a second classification result. The image processing module processes the preprocessed recursive graph features and time-frequency features, respectively, and outputs a third classification result and a fourth classification result. The multi-feature fusion module fuses the first, second, third, and fourth classification results to obtain a fusion result. The target orientation recognition module obtains the target orientation recognition result corresponding to the detected echo sample based on the fusion result.

[0026] The multi-feature fusion network is trained using the detected echo samples and a pre-set loss function to obtain a trained multi-feature fusion network.

[0027] The target location identification result of the target detection echo is obtained by acquiring the recursive graph features, laminar flow features, time domain features and time frequency features corresponding to each angle sequence of the target detection echo.

[0028] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0029] Acquire probe echo samples, and extract recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the probe echo samples; the probe echo samples are labeled.

[0030] A multi-feature fusion network is constructed, comprising a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation recognition module. The sequence processing module processes the preprocessed temporal features and laminar flow features, respectively, and outputs a first classification result and a second classification result. The image processing module processes the preprocessed recursive graph features and time-frequency features, respectively, and outputs a third classification result and a fourth classification result. The multi-feature fusion module fuses the first, second, third, and fourth classification results to obtain a fusion result. The target orientation recognition module obtains the target orientation recognition result corresponding to the detected echo sample based on the fusion result.

[0031] The multi-feature fusion network is trained using the detected echo samples and a pre-set loss function to obtain a trained multi-feature fusion network.

[0032] The target location identification result of the target detection echo is obtained by acquiring the recursive graph features, laminar flow features, time domain features and time frequency features corresponding to each angle sequence of the target detection echo.

[0033] The aforementioned active sonar target location identification method and apparatus based on phase space reconstruction acquires detection echo samples and extracts recursive graph features, laminar flow features, temporal features, and time-frequency features corresponding to each angle sequence of the detection echo samples. Then, a multi-feature fusion network is constructed, comprising a sequence processing module, an image processing module, a multi-feature fusion module, and a target location identification module. The sequence processing module processes the preprocessed temporal and laminar flow features, and the image processing module processes the preprocessed recursive graph features and time-frequency features. The feature fusion module fuses the outputs of the sequence processing module and the image processing module to obtain a fusion result. The target orientation recognition module then uses this fusion result to obtain the target orientation recognition result corresponding to the detected echo sample. Next, the multi-feature fusion network is trained using the detected echo sample and a pre-set loss function to obtain a trained multi-feature fusion network. Finally, the detected echo to be detected is acquired, and the recursive graph features, laminar flow features, temporal features, and time-frequency features corresponding to each angle sequence of the detected echo to be detected are input into the trained multi-feature fusion network to obtain the target orientation recognition result of the detected echo to be detected. This embodiment of the invention can improve the accuracy of active sonar detection in low signal-to-noise ratio environments. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating an active sonar target location identification method based on phase space reconstruction in one embodiment.

[0035] Figure 2 This is an overall framework diagram of the PWFR network model in one embodiment;

[0036] Figure 3 This is a flowchart illustrating the nonlinear feature extraction algorithm for sounding echoes in one embodiment.

[0037] Figure 4 This is a schematic diagram of the training of a PWFR network model in one embodiment;

[0038] Figure 5 Here are flowcharts of two methods in a simulation experiment in one embodiment, where (a) is a flowchart of the matched filtering method and (b) is a flowchart of the nonlinear feature extraction method.

[0039] Figure 6Here are the echo waveforms of the 20th element under different signal-to-noise ratios in one embodiment, where (a) is a schematic diagram of the echo waveform when the signal-to-noise ratio is 0dB, (b) is a schematic diagram of the echo waveform when the signal-to-noise ratio is -10dB, and (c) is a schematic diagram of the echo waveform when the signal-to-noise ratio is -20dB.

[0040] Figure 7 Here are two different simulation results obtained by different methods in one embodiment: (a) is a schematic diagram of the result of the matched filtering method when the signal-to-noise ratio is 0dB; (b) is a schematic diagram of the result of the matched filtering method when the signal-to-noise ratio is -10dB; (c) is a schematic diagram of the result of the matched filtering method when the signal-to-noise ratio is -20dB; (d) is a schematic diagram of the result of the nonlinear feature processing when the signal-to-noise ratio is 0dB; (e) is a schematic diagram of the result of the nonlinear feature processing when the signal-to-noise ratio is -10dB; and (f) is a schematic diagram of the result of the nonlinear feature processing when the signal-to-noise ratio is -20dB.

[0041] Figure 8 This is a schematic diagram of the training set loss function curve in one embodiment;

[0042] Figure 9 This is a schematic diagram of the accuracy curve in one embodiment;

[0043] Figure 10 This is a structural block diagram of an active sonar target location identification device based on phase space reconstruction in one embodiment;

[0044] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0046] In one embodiment, such as Figure 1 As shown, an active sonar target location identification method based on phase space reconstruction is provided, including the following steps:

[0047] Step 102: Obtain the probe echo samples and extract the recursive graph features, laminar flow features, time domain features and time frequency features corresponding to each angle sequence of the probe echo samples.

[0048] The time-domain features of the probe echo are extracted by calculating its autocorrelation function, while the time-frequency features are obtained using WVD (Wigner-Ville distribution). The recurrence plot (RP) and recursive laminarity (RL) features are nonlinear features. The impact of feature extraction from the time domain, frequency domain, and nonlinear methods on signal processing is discussed. Simulations were performed to extract the three features of the probe echo under different signal-to-noise ratios (SNR). Time-domain analysis shows that as the SNR decreases, the correlation peaks become blurred, posing a challenge to accurate signal extraction. Frequency-domain analysis demonstrates that under low SNR, the target bright spots in the original signal are no longer obvious, requiring the assistance of other methods. Nonlinear analysis, by combining recurrence plot features and RL features, emphasizes the feature differences and extraction reliability under different SNR conditions. The quantization features of the recurrence plot are more obvious than those of the recurrence plot itself, and can still effectively extract target echo features even at low SNR. Furthermore, the RL feature exhibits a large feature difference across various SNR conditions; selecting this feature helps reduce the complexity of network construction. In summary, feature extraction using time-domain, frequency-domain, and nonlinear methods complement each other, providing multi-level and multi-angle data support for marine active sonar detection systems and improving the accuracy and reliability of target detection.

[0049] After obtaining these four features, data preprocessing is required to effectively input them into the deep learning network and minimize the amount of data. This includes data alignment, discretization, and type conversion. These preprocessing steps help to input the features into the network model in a standardized manner.

[0050] The detected echo samples are labeled. Specifically, each sample contains four different types of features and one label. These features are AF (autocorrelation) features, WVD (wooden volume decomposition) features, RP (reactive power) features, and RL (reactive energy) features. Among them, WVD and RP features are image features, while AF and RL features are time series features. The label is used to indicate whether the sample contains a target, where 0 indicates no target and 1 indicates a target.

[0051] Step 104: Construct a multi-feature fusion network.

[0052] The multi-feature fusion network includes a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation recognition module. The sequence processing module processes the preprocessed temporal features and laminar features respectively, outputting a first classification result and a second classification result. The image processing module processes the preprocessed recursive graph features and time-frequency features respectively, outputting a third classification result and a fourth classification result. The multi-feature fusion module fuses the first classification result, the second classification result, the third classification result, and the fourth classification result to obtain the fusion result. The target orientation recognition module obtains the target orientation recognition result corresponding to the detected echo sample based on the fusion result.

[0053] After preprocessing the feature data, it is input into the PWFR network model (Multi-Feature Fusion Network). This model includes a time series processing module and an image processing module to handle different types of features. Time-domain features and laminar flow features are input into the time series processing module, while WVD features and recursive graph features are input into the image processing module. The image processing network is based on the ResNet50 network framework and improved upon it; the time series processing network uses an LSTM network. Its overall framework diagram is shown below. Figure 2 As shown.

[0054] Step 106: Train the multi-feature fusion network using the probe echo samples and the pre-set loss function to obtain the trained multi-feature fusion network.

[0055] During network training, WVD features and recurrent graph features are input into the image processing network, while temporal and RL features are input into the time series processing network, along with their corresponding labels. Then, by initializing the network parameters, sequence data is input into the network, and the network output is calculated through forward propagation. Next, a loss function is calculated to measure the difference between the model's classification result and the true label, and a backpropagation algorithm is used to update the network parameters to reduce the loss. Through multiple iterations of training, the network model's parameters are optimized. This network can extract features from each input and produce four different outputs. Different coefficients are assigned to each output, and the final result is calculated, outputting a decision on whether the target is contained.

[0056] Through model training, these four features produce different outputs. To better integrate these features, different coefficients are assigned to each feature, with the coefficients of the nonlinear features determined by the LLE method (Largest Lyapunov Exponents). Therefore, this method can more comprehensively describe the characteristics of the probe echo, and by fusing these characteristics, a final judgment result is obtained.

[0057] Step 108: Obtain the detection echo to be detected. Input the recursive graph features, laminar flow features, time domain features and time frequency features corresponding to each angle sequence of the detection echo to be detected into the trained multi-feature fusion network to obtain the target orientation recognition result of the detection echo to be detected.

[0058] In the aforementioned active sonar target orientation identification method based on phase space reconstruction, the method acquires detection echo samples and extracts the recursive graph features, laminar flow features, temporal features, and time-frequency features corresponding to each angle sequence of the detection echo samples. Then, a multi-feature fusion network is constructed, including a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation identification module. The sequence processing module processes the preprocessed temporal and laminar flow features, and the image processing module processes the preprocessed recursive graph features and time-frequency features. The multi-feature fusion module fuses the outputs of the sequence processing module and the image processing module to obtain a fusion result. The target orientation recognition module then uses this fusion result to obtain the target orientation recognition result corresponding to the detected echo sample. Next, the multi-feature fusion network is trained using the detected echo sample and a pre-set loss function to obtain a trained multi-feature fusion network. Finally, the detected echo to be detected is acquired, and the recursive graph features, laminar flow features, temporal features, and time-frequency features corresponding to each angle sequence of the detected echo to be detected are input into the trained multi-feature fusion network to obtain the target orientation recognition result of the detected echo to be detected. This embodiment of the invention can improve the accuracy of active sonar detection in low signal-to-noise ratio environments.

[0059] In one embodiment, the steps of extracting the recursive graph features and laminar flow features corresponding to each angle sequence of the probe echo sample include: acquiring the echo beam of the probe echo sample, reconstructing the phase space of each angle sequence of the echo beam, and determining whether the reconstructed angle sequence has nonlinear features; if the echo beam has nonlinear features, then using the recursive matrix analysis method to calculate the repeatability of different state points in the phase space to obtain the recursive graph corresponding to each angle sequence, and performing quantitative calculations based on the recursive graph to obtain the laminar flow features.

[0060] In this embodiment, as Figure 3 The flowchart shown illustrates the nonlinear feature extraction algorithm for underwater acoustic detection echoes. First, the detection echo is obtained. After calculating the embedding dimension and delay time of the sequence, its phase space is reconstructed. The Wolf algorithm is used to calculate the maximum Lyapunov exponent of the echo beam, confirming that the echo has nonlinear characteristics. If the echo exhibits nonlinearity, its nonlinear characteristics are further analyzed. A recursion graph is obtained through recursion matrix analysis. To further analyze the quantitative relationships between recursion points on the recursion graph, the recursion graph features need to be calculated, thereby achieving the extraction of nonlinear features from underwater acoustic detection echoes.

[0061] In the phase space reconstruction of probe echoes, a crucial step in accurately reconstructing a one-dimensional probe echo into a multi-dimensional space is determining the delay time *t* and the embedding dimension *m*. Methods for calculating *t* include the average mutual information function method, the autocorrelation function method, and the average displacement method. To better suit nonlinear probe echoes and improve computational accuracy, this invention employs the average mutual information function method to calculate *t*. Methods for calculating *m* include the pseudo-nearest neighbor method, singular value decomposition (SVD), and the Cao method. Considering the advantages of computational simplicity and avoiding information loss, this invention uses the pseudo-nearest neighbor method to calculate *m*. The main idea of ​​the average mutual information function method is to calculate the mutual information of the signal at different delay times and then select the delay time of the first local minimum of the average mutual information as the optimal delay time. This method effectively captures the dynamic characteristics and complexity of the signal, thereby determining a suitable delay time for more accurate phase space reconstruction and analysis. The main idea of ​​the pseudo-nearest neighbor method is to gradually increase the embedding dimension during phase space reconstruction, observing the evolution of the proximity of adjacent trajectory points in different dimensions. By observing whether points originally considered nearby exhibit greater dispersion as the dimension increases, this method can be applied to the reconstruction. If these points maintain their adjacency in high-dimensional space, they are considered true neighbors; conversely, if they are scattered in high-dimensional space, they are considered "false neighbors." By analyzing the number or proportion of false neighbors, the optimal embedding dimension is selected as the embedding dimension that makes the number or proportion of false neighbors stable. At this point, the geometric structure of the system is fully revealed.

[0062] In determining the nonlinear characteristics of echoes, the maximum Lyapunov exponent describes the maximum separation rate between trajectory points in phase space. In probe echo signals, this exponent reflects the nonlinear characteristics of the signal. By constructing a recursive graph and extracting its features, the nonlinear structure and complex dynamic characteristics in probe echo data can be effectively captured, thereby achieving effective target identification and classification. Selecting appropriate recursive graph quantization features helps to accurately describe the nonlinear characteristics of probe echo data. In the phase space reconstruction process, according to Takens' theorem, when m ≥ 2k + 1, the reconstructed phase space is equivalent to the original phase space only when the embedding dimension extends k-dimensional data to m-dimensional space. If a certain angle sequence is 1-dimensional, then its system embedding dimension should be greater than or equal to 3 dimensions. If the reconstruction dimension of the phase space exceeds three dimensions, the reconstructed phase space cannot be directly presented through visualization. To quantitatively evaluate the morphology and evolution characteristics of phase space trajectories, a recursive matrix analysis method was used to calculate the repeatability of different state points in the phase space. The calculation of the recursive matrix is ​​as follows:

[0063] D ij =H(e-||S i -S j ||)

[0064] Where i,j represent different state points in phase space, i,j=1,2,…,N, N=n-(m-1)τ, and H is the Heaviside step function used to determine the similarity between state points. When state point S i and S j The Euclidean distance between them is less than a given threshold e, i.e., e > ||S i -S j ||At that time, D ij =1,S i and S j A closer distance between two points indicates that their states are similar; conversely, a greater distance between two points indicates that their states are similar. i and S j The Euclidean distance between them is greater than a given threshold e, i.e., e < ||S i -S j ||,D ij =0, S i and S j The distance between them is large, and the states of the two points are not similar.

[0065] To further analyze the characteristics of each node in a recursive graph, Recursive Quantitative Analysis (RQA) was employed. RQA aims to reveal the characteristics of a recursive graph quantitatively through mathematical and statistical analysis, thereby providing a more comprehensive understanding of the nonlinear structure and dynamics of the data. RQA can calculate several metrics of a recursive graph, including the entropy of the diagonalline lengths (ENTR), recurrence laminrity (RL), determination (DET), and transitivity (Trans). These metrics obtained through RQA can more accurately describe the complexity and dynamic behavior of the data, providing important reference for further data analysis. Quantitative features of recursive graphs play a crucial role in the study of complex systems: ENTR measures the uncertainty and diversity of diagonal structure lengths in a recursive graph, RL describes the frequency of layered structures, DET quantifies the proportion of deterministic trajectories, and Trans reflects the degree of triangular connections between nodes, revealing the tightness of the local network structure. These features are significant for understanding the structure and dynamic behavior of recursive graphs and for in-depth analysis of the nonlinear evolution of probe-echo systems. Given that recursive quantification features exhibit similar performance on probe-echo systems, choosing RL features helps reduce the complexity of subsequent network construction while still meeting research requirements.

[0066] Nonlinear feature extraction can describe complex underwater acoustic signals more comprehensively and accurately. By extracting features through recursion graphs and recursion graph quantization, the intrinsic structure of the signal can be analyzed in depth, thereby enhancing the understanding and detection capabilities of signal characteristics.

[0067] In one embodiment, the sequence processing module includes two convolutional neural networks with the same structure, and the image processing module includes two long short-term memory neural networks with the same structure; the convolutional neural network includes ResNet50.

[0068] In this embodiment, the core concept of ResNet is that each additional network layer should be able to easily fit the residual mapping. The central component of the network is the residual block, which effectively facilitates the flow of gradients within the network by introducing residual connections across layers, thereby achieving faster information transfer. This design strategy improves the model's fitting ability and training efficiency, allows the network to learn feature representations at deeper levels, and mitigates gradient vanishing and gradient exploding phenomena, providing effective support for training deeper networks. Figure 4 The diagram shown illustrates the training of the PWFR network model. ResNet50 is a convolutional neural network model based on a deep residual network architecture. It consists of multiple residual blocks containing 50 convolutional layers, as well as global average pooling layers and fully connected layers. The initial convolutional layers (Conv1) and pooling layers (POOL) are used to extract basic features of the image, reducing its size and complexity. Furthermore, ResNet adds a batch normalization (BN) layer and an activation function (ReLU) after each convolutional layer. The BN layer accelerates the convergence of the neural network training by reducing internal covariate shifts, improving network performance and stability, and allowing for higher learning rates. The final fully connected layer maps the output of the global average pooling layer to the final result.

[0069] This invention redefines the first convolutional layer and the last fully connected layer of ResNet50 based on the input image feature size and output requirements. The input channels of the first convolutional layer are set to 2, and the fully connected layer is replaced with a linear layer and a LogSoftmax activation function. Therefore, the network's output can be converted into a normalized probability distribution, enabling the model to output the probability of each class for classification prediction.

[0070] LSTM incorporates multiple memory modules, each containing three multiplication control units: an input gate, an output gate, and a forget gate, providing functions similar to read, write, and reset. At each time step, the LSTM network receives the input of the current time step and the state of the previous time step, producing the output and the state of the next time step. A fully connected layer is added after the LSTM layers to learn the features extracted from the LSTM layers and perform classification tasks.

[0071] In one embodiment, the loss function is:

[0072]

[0073] Where Loss represents the loss function, N represents the number of probe echo samples taken in each training iteration, and y i Let x represent the true label of the i-th probe echo sample. WVDi x represents the fourth classification result corresponding to the i-th probe echo sample. RPi Let x represent the third classification result corresponding to the i-th probe echo sample. AFi Let x represent the first classification result corresponding to the i-th probe echo sample. RLi α1 represents the second classification result corresponding to the i-th probe echo sample, and α2, α3 and α4 represent weighting coefficients.

[0074] In this embodiment, four different features are input into the neural network, and the entire dataset is trained 50 times. Due to memory limitations, the batch size for each training iteration is set to 128, the optimal learning rate is adjusted to 0.001, and the network optimizer is Adam. After training, the network obtains four different outputs, named x. WVD x RP x AF and x RL α1x corresponds to the outputs of WVD, RP, AF, and RL features, respectively. To measure the difference between the network model's predicted values ​​and the actual values, α1x... WVDi +α2x RPi +α3x AFi +α4x RLiThis represents the combined probability after weighting the four output features with different coefficients, where the sum of α1, α2, α3, and α4 is 1. Since this network model aims to analyze complex underwater acoustic signals, focusing on their nonlinear characteristics, after parameter optimization, we set α1 = 0.216, α2 = 0.331, α3 = 0.181, and α4 = 0.272 in the network. The loss function quantifies the true value and the model's predicted value into a single numerical value, thus measuring the model's prediction accuracy and the difference between the predicted and true values. During training, it is necessary to continuously adjust the hyperparameters, monitor model performance, and evaluate the model's performance on the validation set until a better performance level is achieved.

[0075] In one embodiment, obtaining the target orientation identification result corresponding to the probe echo sample based on the fusion result includes: obtaining the target identification result corresponding to the probe echo sample based on the fusion result; if the target identification result indicates the existence of a target, then drawing an orientation history map based on the laminar flow characteristics of each angle sequence corresponding to the probe echo sample, and determining the target orientation based on the orientation history map.

[0076] In one embodiment, drawing an azimuth history map based on the laminar flow characteristics of each angle sequence corresponding to the probe echo sample includes: obtaining the time period where the laminar flow characteristics are greater than a threshold, taking the starting segment of the time period as the target point, and obtaining the azimuth history map.

[0077] In one embodiment, the method further includes: preprocessing the recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the probe echo samples; the preprocessing operations include data alignment, discretization, and type conversion. In this embodiment, data alignment aims to ensure consistent data lengths in the input network, discretization aims to reduce the amount of data in the input network while preserving feature information as much as possible, and type conversion converts various feature data into single-precision data for input into the network.

[0078] In one specific embodiment, to further ensure the reliability of the algorithm and the credibility of the results, this invention considers comparison with the classic matched filtering algorithm when constructing simulation data, and performs result analysis under different signal-to-noise ratio conditions. The simulation data consists of multi-element data, used to simulate the complex detection echoes received by the hydrophone array in a real underwater environment. Then, the multi-element echoes are processed using both the matched filtering method and the nonlinear feature extraction method proposed in this invention, and the processing results of the two methods are compared and analyzed in detail. Simultaneously, this invention also specifically selects lake trials as the noise source to more realistically simulate the noise characteristics of the underwater environment, improving the reliability and comparability of the simulation results. Furthermore, this simulation sets 32 array elements, an array frequency of 500Hz, and a sound wave propagation speed of 1500m / s. The sampling rate of the simulation data is 4000Hz, the frequency of the target signal is 480Hz, the total duration of the multi-element detection echo is 2s, the length of the target echo is 0.4s (time period 0.6s-1s), and the actual angle is 80°. Finally, the flowcharts for the two different algorithms are as follows: Figure 5 As shown.

[0079] exist Figure 5 In this study, a fractional-delay time-domain beamforming method was employed to generate an optimized signal beam. This beam exhibits higher gain in a specific direction, enabling directional reception of the target signal and effectively suppressing interference signals in other directions. Figure 5 In (a), the matched filtering algorithm directly processes the beamforming matrix, ultimately obtaining the azimuth history map. However, Figure 5 In (b), the nonlinear feature extraction algorithm extracts features from each angle sequence after beamforming. After traversing 180 azimuths, it obtains the azimuth history map represented by the feature values ​​of each angle. The target endpoint is determined by the feature value map to obtain the final azimuth history map.

[0080] In simulation experiments, this invention thoroughly investigated the performance of two algorithms under different signal-to-noise ratio (SNR) conditions: nonlinear feature extraction and matched filtering methods were applied to three different SNR conditions (0dB, -10dB, and -20dB), and their effects were compared. The echo waveforms of the 20th element under different SNR conditions are shown below. Figure 6 As shown, where, Figure 6 (a) is a schematic diagram of the echo waveform when the signal-to-noise ratio is 0dB. Figure 6 (b) is a schematic diagram of the echo waveform when the signal-to-noise ratio is -10dB. Figure 6 (c) is a schematic diagram of the echo waveform at a signal-to-noise ratio of -20dB. The final azimuth history diagrams obtained by simulation using two different methods are shown below. Figure 7 As shown. By Figure 6When the signal-to-noise ratio is 0dB, the waveform amplitude is relatively large and there is a target echo in the range of 0.6s to 1s. However, at -10dB and -20dB, the time period in which the target echo is located cannot be observed at all on the time domain waveform diagram.

[0081] Figure 7 In the images (a), (b), and (c), the results of the matched filtering algorithm are shown for signal-to-noise ratios of 0dB, -10dB, and -20dB, respectively. The target can be observed in all three images. As the signal-to-noise ratio decreases, the target becomes increasingly faint. Figure 7 Images (d), (e), and (f) respectively illustrate the processing results of the nonlinear feature extraction algorithm at signal-to-noise ratios of 0dB, -10dB, and -20dB. Compared to the matched filtering method, the bright spots are more pronounced, and the azimuth is more accurate. The reason for this phenomenon lies in the fact that, unlike the matched filtering method where the brightest energy point represents the target, the nonlinear feature extraction method effectively extracts features that distinguish the target from the noise. In the echo sequence, the time period containing the target will show bright spots. Selecting the endpoint where the change begins as the target location makes the azimuth history map clearer and the azimuth more accurate.

[0082] In one specific embodiment, this invention constructs a dataset containing 6000 sets of probe echo samples. Of these, 3000 sets contain simulated target probe echoes, while the other 3000 sets do not. The sampling rate of these probe echoes is 2000Hz, and each set contains 4096 sampling points. The target echo is a Costas signal with a center frequency of 100Hz, a frequency interval of 10Hz, and 1024 sampling points randomly distributed within the probe echo sequence. Feature extraction was performed on these 6000 sets of probe echo data, resulting in four different types of features. Each sample contains four different types of features and one label. This invention performs data alignment, discretization, and type conversion on all features. Ultimately, the WVD feature has a size of 2×205×205, containing both real and imaginary parts; the RP feature has a size of 1×430×430; the AF feature has a size of 1×4096; and the RL feature has a size of 1×3440. In the multi-feature fusion network, the improved ResNet50 network used for training image features has the same number of input channels as the number of feature channels in its Conv1 layer, while the number of output channels is 64. This means that each convolutional kernel generates 64 feature maps. The kernel size is 7×7, which determines the receptive field size of each convolution operation, i.e., the size of the input feature map region covered by the convolution operation. This invention chooses a smaller stride of 2 to help extract more accurate features. The amount of padding is usually determined based on the kernel size and stride to ensure that the convolution operation can cover the entire region of the input feature map. The parameter is set to 3×3 to pad with 3 elements around the edges to keep the feature map size constant. In addition, no bias term is set in ResNet50 to reduce the number of model parameters, reduce the risk of overfitting, and simplify the model training and tuning process.

[0083] For training time series features, an LSTM network was used in the model, with its input and hidden layer sizes consistent with the feature sizes. Simultaneously, an iterative parameter initialization method was used to initialize the model parameters, including initializing the weights of linear layers to a constant 1 and the biases to -100, and performing constant initialization and Kaiming initialization on the weights and biases of convolutional and batch normalized layers.

[0084] Based on the above training process, the training set loss function curve was obtained, as shown below. Figure 8As shown in the figure, after training, it can be observed that the error between the network model's predicted values ​​and the true values ​​is close to zero, indicating that the network exhibits good performance in classifying detection echoes with and without targets. This result further verifies the effectiveness and generalization ability of the network model. Through analysis of the loss function curve, it can be observed that the model gradually converges during training and exhibits a low loss on the training set, which means that the model can fit the training data well. In addition, by observing that the error is close to zero, it can be confirmed that the model achieves high accuracy in the prediction task and effectively classifies target and non-target echoes. Therefore, based on the training set loss function curve and error analysis, this invention draws conclusions on the effectiveness and reliability of the network model in classifying detection echoes with and without targets.

[0085] In one specific embodiment, the present invention also evaluated the model's performance on the test set and obtained an accuracy curve, such as... Figure 9 As shown, after 18 epochs of training, the accuracy on the test set gradually approaches 100%. This demonstrates the model's excellent generalization ability when predicting unseen data. The steady rise in accuracy and the near-perfect accuracy further confirm the model's superior performance on the test set. This result is significant for validating the model's effectiveness and reliability, further strengthening its credibility in target and non-target echo detection classification tasks. By observing the test set accuracy curve, we can more comprehensively evaluate the model's generalization ability and performance stability, thus providing important support for its reliability in practical applications.

[0086] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in this invention, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0087] In one embodiment, such as Figure 10 As shown, an active sonar target location identification device based on phase space reconstruction is provided, comprising:

[0088] The feature extraction module 1002 is used to acquire the probe echo samples and extract the recursive graph features, laminar flow features, time-domain features and time-frequency features corresponding to each angle sequence of the probe echo samples; the probe echo samples are labeled with tags;

[0089] Network construction module 1004 is used to construct a multi-feature fusion network. The multi-feature fusion network includes a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation recognition module. The sequence processing module processes the preprocessed temporal features and laminar features respectively, and outputs the first classification result and the second classification result. The image processing module processes the preprocessed recursive graph features and time-frequency features respectively, and outputs the third classification result and the fourth classification result. The multi-feature fusion module fuses the first classification result, the second classification result, the third classification result, and the fourth classification result to obtain the fusion result. The target orientation recognition module obtains the target orientation recognition result corresponding to the detected echo sample based on the fusion result.

[0090] The network training module 1006 is used to train a multi-feature fusion network using probe echo samples and a pre-set loss function to obtain a trained multi-feature fusion network.

[0091] The output module 1008 is used to acquire the detection echo to be detected. It inputs the recursive graph features, laminar flow features, time domain features and time frequency features corresponding to each angle sequence of the detection echo to the trained multi-feature fusion network to obtain the target orientation recognition result of the detection echo.

[0092] In one embodiment, the sequence processing module includes two convolutional neural networks with the same structure, and the image processing module includes two long short-term memory neural networks with the same structure; the convolutional neural networks include ResNet50.

[0093] In one embodiment, the loss function is further used as:

[0094]

[0095] Where Loss represents the loss function, N represents the number of probe echo samples taken in each training iteration, and y i Let x represent the true label of the i-th probe echo sample. WVDi x represents the fourth classification result corresponding to the i-th probe echo sample. RPi Let x represent the third classification result corresponding to the i-th probe echo sample. AFi Let x represent the first classification result corresponding to the i-th probe echo sample. RLi α1 represents the second classification result corresponding to the i-th probe echo sample, and α2, α3 and α4 represent weighting coefficients.

[0096] In one embodiment, it is also used to acquire the echo beam of the probe echo sample, reconstruct the phase space of each angle sequence of the echo beam, and determine whether the current reconstructed angle sequence has nonlinear characteristics; if the echo beam has nonlinear characteristics, the recursive matrix analysis method is used to calculate the repeatability of different state points in the phase space to obtain the recursive graph corresponding to each angle sequence, and the laminar flow characteristics are obtained by quantitative calculation based on the recursive graph.

[0097] In one embodiment, it is also used to obtain the target identification result corresponding to the probe echo sample based on the fusion result; if the target identification result is that a target exists, then draw an azimuth history map based on the laminar flow characteristics of each angle sequence corresponding to the probe echo sample, and determine the target azimuth based on the azimuth history map.

[0098] In one embodiment, it is also used to obtain the time period in which the laminar flow characteristics are greater than a threshold, and take the starting segment of the time period as the target point to obtain the azimuth history map.

[0099] In one embodiment, the method is further used to perform preprocessing operations on the recursive graph features, laminar flow features, time-domain features and time-frequency features corresponding to each angle sequence of the probe echo sample; the preprocessing operations include data alignment operations, discretization operations and type conversion operations.

[0100] Specific limitations regarding the phase-space reconstruction-based active sonar target location identification device can be found in the above-described limitations of the phase-space reconstruction-based active sonar target location identification method, and will not be repeated here. Each module in the aforementioned phase-space reconstruction-based active sonar target location identification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0101] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 11As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a phase-space reconstruction-based active sonar target location identification method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0102] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0103] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0104] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0105] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for identifying the azimuth of an active sonar target based on phase space reconstruction, characterized in that, The method includes: Acquire probe echo samples, and extract recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the probe echo samples; the probe echo samples are labeled. A multi-feature fusion network is constructed, comprising a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation recognition module. The sequence processing module processes the preprocessed temporal features and laminar flow features, respectively, and outputs a first classification result and a second classification result. The image processing module processes the preprocessed recursive graph features and time-frequency features, respectively, and outputs a third classification result and a fourth classification result. The multi-feature fusion module fuses the first, second, third, and fourth classification results to obtain a fusion result. The target orientation recognition module obtains the target orientation recognition result corresponding to the detected echo sample based on the fusion result. The multi-feature fusion network is trained using the detected echo samples and a pre-set loss function to obtain a trained multi-feature fusion network. The target location identification result of the detection echo is obtained by acquiring the recursive graph features, laminar flow features, time domain features and time frequency features corresponding to each angle sequence of the detection echo. The steps of extracting the recursive graph features and laminar flow features corresponding to each angle sequence of the detected echo samples include: Acquire the echo beam of the probe echo sample, reconstruct the phase space of each angle sequence of the echo beam, and determine whether the current reconstructed angle sequence has nonlinear characteristics; If the echo beam has nonlinear characteristics, the recursive matrix analysis method is used to calculate the repeatability of different state points in the phase space to obtain the recursive diagram corresponding to each angle sequence. Based on the recursive diagram, the laminar flow characteristics are obtained by quantitative calculation. The step of obtaining the target orientation identification result corresponding to the probe echo sample based on the fusion result includes: The target identification result corresponding to the probe echo sample is obtained based on the fusion result; If the target identification result indicates that a target exists, then an azimuth history map is drawn based on the laminar flow characteristics of each angle sequence corresponding to the detected echo sample, and the target azimuth is determined based on the azimuth history map. The step of drawing an azimuth history map based on the laminar flow characteristics of each angle sequence corresponding to the detected echo sample includes: The time period in which the laminar flow characteristic is greater than the threshold is obtained, and the starting segment of the time period is taken as the target point to obtain the azimuth history map.

2. The method according to claim 1, characterized in that, The sequence processing module includes two convolutional neural networks with the same structure, and the image processing module includes two long short-term memory neural networks with the same structure; the convolutional neural network includes ResNet50.

3. The method according to claim 1 or 2, characterized in that, The loss function is: in, Represents the loss function. N This indicates the number of probe echo samples taken in each training iteration. y i Indicates the first i The true label of a probe echo sample Indicates the first i The fourth classification result corresponding to each probe echo sample Indicates the first i The third classification result corresponding to each probe echo sample Indicates the first i The first classification result corresponding to each probe echo sample Indicates the first i The second classification result corresponding to each probe echo sample , , and This represents the weighting coefficient.

4. The method according to claim 1, characterized in that, The method further includes: Preprocessing operations are performed on the recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the detected echo samples; the preprocessing operations include data alignment, discretization, and type conversion.

5. A target location identification device based on phase space reconstruction for active sonar detection, applied to the method described in any one of claims 1-4, characterized in that, The device includes: The feature extraction module is used to acquire probe echo samples and extract recursive graph features, laminar flow features, time-domain features, and time-frequency features corresponding to each angle sequence of the probe echo samples; the probe echo samples are labeled. A network construction module is used to construct a multi-feature fusion network. The multi-feature fusion network includes a sequence processing module, an image processing module, a multi-feature fusion module, and a target orientation recognition module. The sequence processing module processes the preprocessed temporal features and laminar features respectively, outputting a first classification result and a second classification result. The image processing module processes the preprocessed recursive graph features and time-frequency features respectively, outputting a third classification result and a fourth classification result. The multi-feature fusion module fuses the first classification result, the second classification result, the third classification result, and the fourth classification result to obtain a fusion result. The target orientation recognition module obtains the target orientation recognition result corresponding to the detected echo sample based on the fusion result. The network training module is used to train the multi-feature fusion network using the probe echo samples and a pre-set loss function to obtain the trained multi-feature fusion network. The output module is used to acquire the detection echo to be detected. It inputs the recursive graph features, laminar flow features, time-domain features and time-frequency features corresponding to each angle sequence of the detection echo to the trained multi-feature fusion network to obtain the target orientation recognition result of the detection echo.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images