A method, system and related device for evaluating the realism of a received signal of a generative active sonar targeting an unmanned underwater vehicle
By combining deep learning technology with wavelet analysis and self-attention mechanism, the problems of capturing time-frequency features and environmental adaptability in the evaluation of the authenticity of active sonar received signals of unmanned underwater vehicles were solved. This enabled multi-dimensional feature evaluation of the signals, improving the accuracy and adaptability of the evaluation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2025-05-30
- Publication Date
- 2026-07-24
AI Technical Summary
Existing methods for assessing the authenticity of active sonar received signals from unmanned underwater vehicles (UUVs) cannot fully evaluate the non-stationary characteristics of the signal in the joint time-frequency domain and its degree of matching with the real underwater environment. In particular, they are easily affected by background noise under low signal-to-noise ratio conditions, resulting in a large deviation between the assessment results and human subjective judgment.
By employing deep learning techniques combined with wavelet analysis, successive variational mode decomposition, and self-attention mechanisms, and through hybrid position coding and feature enhancement modules, segmented position coding, rotational position coding, denoising, mode decomposition, feature extraction, and fusion of active sonar received signals are performed to achieve joint evaluation of multi-dimensional features of the signal.
It improves the classification accuracy of target echo signals and noise signals from unmanned underwater vehicles, enhances the feature modeling capability of active sonar received signals, and provides a more accurate method for evaluating the authenticity of generative active sonar received signals.
Smart Images
Figure CN120595273B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of unmanned underwater vehicles, specifically relating to a method, system, and related apparatus for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets. Background Technology
[0002] In the field of active sonar signal simulation for unmanned underwater vehicles (UUVs), the realism of the generated active sonar signals directly affects the training effect of underwater target recognition algorithms and the reliability of system testing. Due to the complexity of the underwater acoustic propagation environment, the active sonar signals of UUVs are subject to multipath effects, time-varying channel characteristics, and nonlinear noise interference, making it difficult for traditional signal generation models to accurately reproduce the time-frequency characteristics of real underwater signals. Existing signal realism assessment methods mainly rely on human experience or traditional evaluation indicators, such as mean square error (MSE), signal-to-noise ratio (SNR), and power spectral density similarity (PSD Similarity). However, these methods can only reflect some characteristics of the signal and cannot comprehensively assess the non-stationary characteristics of the signal in the joint time-frequency domain and its degree of matching with the real underwater environment.
[0003] In recent years, deep learning techniques such as Generative Adversarial Networks (GANs) have been used to generate active sonar signals for unmanned underwater vehicles (UUVs), but significant limitations remain in terms of realism assessment. Traditional evaluation methods employ convolutional neural networks (CNNs) and autoencoders, which, while performing well in image generation quality assessment, struggle to effectively capture the unique sparsity and transient pulse structures of active sonar signals from UUV targets. Existing evaluation models generally suffer from a single feature extraction dimension, failing to simultaneously consider the signal's time-domain waveform details, frequency-domain energy distribution, and nonlinear characteristics in the time-frequency joint domain. Particularly when dealing with low signal-to-noise ratio (SNR) generated signals, conventional evaluation methods are easily affected by background noise, leading to significant discrepancies between evaluation results and subjective human judgment.
[0004] Most deep learning-based evaluation methods employ fixed-weight multi-index fusion strategies, lacking adaptability to complex underwater propagation environments. Existing models, when processing active sonar signals received by unmanned underwater vehicles (UUVs), often neglect the correlation between signal dynamics and carrier motion parameters, resulting in evaluation results that fail to accurately reflect signal fidelity in real-world application scenarios. Therefore, there is an urgent need to construct a novel deep learning evaluation model capable of deeply integrating underwater acoustic field physical characteristics, establishing a multi-dimensional feature joint evaluation mechanism, and achieving intelligent and accurate evaluation of the fidelity of generative active sonar received signals. Summary of the Invention
[0005] The purpose of this invention is to provide a method, system, and related apparatus for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets, thereby solving the problem of low authenticity of existing generative active sonar received signals for unmanned underwater vehicle targets.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] In a first aspect, the present invention provides a method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets, comprising the following steps:
[0008] The generative active sonar received signal of the unmanned underwater vehicle target is preprocessed to obtain the preprocessed active sonar received signal.
[0009] The preprocessed active sonar received signal is subjected to segmented position coding and rotational position coding to obtain a hybrid position coding sequence.
[0010] The mixed position-coded sequence is denoised to obtain a denoised signal sequence;
[0011] The denoised signal sequence is subjected to mode decomposition, feature extraction, and feature fusion to obtain a feature-enhanced sequence.
[0012] Feature extraction and fusion processing are performed on the feature enhancement sequence to obtain the fused feature sequence;
[0013] The fused characteristic sequences are classified, and the probability of the echo signal classified as an unmanned underwater vehicle target is used as the authenticity evaluation result of the generative active sonar received signal of the unmanned underwater vehicle target.
[0014] Preferably, the preprocessed active sonar received signal is subjected to segmented position coding and rotational position coding to obtain a hybrid position coding sequence. Specifically, the method is as follows:
[0015] The preprocessed active sonar received signal is divided using one-dimensional convolution, and the positional information between each segment is embedded to obtain a segmented coded sequence.
[0016] Encode the segmented coded sequence to obtain the rotated coded sequence;
[0017] The rotated encoded sequence is normalized to obtain the hybrid position encoded sequence.
[0018] Preferably, the preprocessed active sonar received signal is divided using the following formula, and the positional information between each segment is embedded to obtain a segmented coded sequence:
[0019]
[0020] In the formula: This represents the number of segments in the segmented position encoding. The sequence length of the active sonar received signal; The length of each segment in the segmented position encoding; This represents the step size for each segment in the segmented position encoding.
[0021] The segmented coded sequence is encoded using the following formula to obtain the rotated coded sequence:
[0022]
[0023]
[0024]
[0025] In the formula: Index for feature dimensions; For the first Rotation parameters on each feature index; The characteristic dimension of the signal received by active sonar; The position index of each segment after rotational position encoding; For the rotation matrix in the th... The feature dimension, the first The value at each segment index; This is the result of rotating the position encoding of the segmented position encoding sequence; This is to convert the rotated position encoding into a complex sequence.
[0026] Preferably, the denoised signal sequence is subjected to mode decomposition, feature extraction, and feature fusion to obtain a feature-enhanced sequence. Specific methods include:
[0027] Modal decomposition is performed on the mixed position-coded sequence to obtain signals in multiple frequency bands;
[0028] Multiple feature sequences are obtained by extracting features from signals across multiple frequency bands.
[0029] Multiple feature sequences are fused to obtain a feature-enhanced sequence.
[0030] Preferably, the following formula is used to extract features from signals across multiple frequency bands:
[0031]
[0032] In the formula: This is the final weighted feature output; Score for attention; The product of the transposes of the Query matrix and the Key matrix represents the feature similarity at different positions. The dimension of the Key vector; The Value matrix;
[0033] Feature extraction and fusion processing are performed on the feature enhancement sequence to obtain the fused feature sequence. The specific method is as follows:
[0034] Two-dimensional convolution is used to perform global feature extraction and local feature extraction on the feature enhancement sequence, respectively, to obtain the local feature attention extraction result and the global feature attention extraction result;
[0035] Calculate feature weights using the results of local feature attention extraction and global feature attention extraction;
[0036] The features are fused using the obtained feature weights to obtain a fused feature sequence.
[0037] Preferably, the feature enhancement sequence is subjected to feature extraction and fusion processing to obtain a fused feature sequence. Specific methods include:
[0038] Feature extraction is performed on the feature enhancement sequence using one-dimensional and three-dimensional convolutions to obtain different convolution sequences. The different convolution sequences are then activated and fused together to obtain a joint convolution sequence.
[0039] The joint convolution sequence obtained from the first layer is fused using a one-dimensional convolution kernel to obtain a fused feature sequence.
[0040] Preferably, the feature enhancement sequence is used to extract features to obtain different convolutional sequences. These different convolutional sequences are then activated and fused together to obtain a joint convolutional sequence:
[0041]
[0042]
[0043] In the formula: This is the result of the first type of convolutional feature extraction; Use the GELU activation function; The input feature sequence for the joint feature extraction module; This is a one-dimensional convolution operation; This is a 3D convolution operation; This is the result of the second type of convolution feature extraction;
[0044] The following formula is used to fuse the features of the joint convolutional sequence obtained in the first layer, resulting in a fused feature sequence:
[0045]
[0046] In the formula: This is the output sequence of the joint feature extraction module.
[0047] Secondly, the present invention provides a generative active sonar received signal authenticity evaluation system for unmanned underwater vehicle targets, comprising:
[0048] The data preprocessing module is used to preprocess the generative active sonar received signal of the unmanned underwater vehicle target to obtain the preprocessed active sonar received signal.
[0049] The hybrid position coding module is used to perform segmented position coding and rotational position coding on the preprocessed active sonar received signal to obtain a hybrid position coding sequence.
[0050] The wavelet analysis denoising module is used to denoise the mixed position-coded sequence to obtain a denoised signal sequence.
[0051] The feature enhancement module is used to perform mode decomposition, feature extraction, and feature fusion on the denoised signal sequence to obtain the feature-enhanced sequence;
[0052] The joint feature extraction module is used to extract and fuse features from the feature-enhanced sequence to obtain a fused feature sequence.
[0053] The authenticity assessment module is used to classify the fused characteristic sequence and use the probability of the echo signal classified as an unmanned underwater vehicle target as the authenticity assessment result of the generative active sonar received signal of the unmanned underwater vehicle target.
[0054] Thirdly, the present invention provides an electronic device including a processor and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed by the processor, the electronic device causes the method described thereon.
[0055] Fourthly, the present invention provides a computer program product, the computer program product including computer-executable instructions, which, when executed, implement the method described.
[0056] Compared with the prior art, the beneficial effects of the present invention are:
[0057] This invention provides a generative active sonar signal authenticity assessment method for unmanned underwater vehicle (UUV) targets. Traditional assessment methods typically extract selected features in the time and frequency domains and calculate similarity to complete the authenticity assessment. This invention employs deep learning technology for assessment, enabling analysis of the overall signal characteristics. Furthermore, compared to existing deep learning assessment methods, this invention combines wavelet analysis, successive variational mode decomposition, and self-attention mechanisms to enhance the feature modeling capability of active sonar received signals and improve the classification accuracy of UUV target echo and noise signals.
[0058] This invention designs a hybrid position coding module. By fusing information from rotational position coding and segmented position coding, this module obtains a hybrid coded sequence. It can not only obtain the relative position features of the signal sequence from rotational position coding, but also obtain the long-range dependence features of the signal from segmented coding, thus enabling better learning of the position information of the active sonar received signal.
[0059] This invention designs a feature enhancement module. This module uses successive variational mode decomposition to decompose the active sonar received signal into different frequency bands. After reconstructing the signals of different frequency bands, a self-attention extraction method is used to extract the signal features of each frequency band. Convolution is used to fuse the various feature vectors. This not only extracts features from different frequency bands of the signal separately, but also fuses different features, thereby better enhancing the feature information of the active sonar received signal.
[0060] In summary, this invention provides the ability to extract features of echo and noise signals from unmanned underwater vehicle (UUV) targets, applicable to the classification of echo and noise signals from UUV targets, and provides a more accurate method for assessing the authenticity of generative active sonar received signals from UUV targets. Attached Figure Description
[0061] Figure 1 This is a schematic diagram of the overall network structure of the present invention.
[0062] Figure 2 This is a schematic diagram of the hybrid position encoding module structure of the present invention.
[0063] Figure 3 This is a schematic diagram of the feature enhancement module structure of the present invention.
[0064] Figure 4 This is a schematic diagram of the joint feature enhancement module structure of the present invention. Detailed Implementation
[0065] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.
[0066] Example 1
[0067] This embodiment provides a method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets, comprising the following steps:
[0068] Step 1: Preprocess the generative active sonar received signal of the unmanned underwater vehicle target to obtain the preprocessed active sonar received signal.
[0069] Step 2: The pre-constructed hybrid position coding module is used to perform segmented position coding and rotational position coding on the pre-processed active sonar received signal. The long-range dependence features and relative position features of the signal in the active sonar received signal are extracted respectively to obtain the hybrid position coding sequence.
[0070] The pre-built hybrid location coding module extracts location features from the input active sonar received signal, which aims to assist in time series modeling.
[0071] The pre-built hybrid position coding module includes a segmented position coding layer, a rotational position coding layer, and a normalization layer. This pre-built hybrid position coding module is trained using a dataset of generative active sonar received signals from autonomously collected unmanned underwater vehicle targets.
[0072] Step 3: Use a pre-built wavelet analysis denoising module to analyze and calculate the mixed position coding sequence to reduce noise in the active sonar received signal, and obtain a denoised signal sequence.
[0073] The wavelet analysis denoising module uses wavelet analysis to perform denoising analysis on the mixed position-coded sequence, with the aim of filtering and denoising high-frequency signals and low-frequency signals respectively.
[0074] Step 4: Use the pre-built feature enhancement module to perform mode decomposition, feature extraction and feature fusion on the denoised signal sequence to obtain the feature-enhanced sequence.
[0075] The pre-constructed feature enhancement module includes a successive variational mode decomposition layer, a self-attention feature extraction layer, and a feature fusion layer, wherein:
[0076] The successive variational mode decomposition layer is used to calculate the frequency domain signal of the active sonar received signal from the denoised signal sequence, select a preset number of frequency bands with the highest frequency domain intensity, and reconstruct the active sonar received signal from the preset number of frequency domain information, thereby preserving the feature information.
[0077] The self-attention feature extraction layer is used to extract features from a preset number of frequency band signals to obtain dependent feature information in different frequency bands.
[0078] The feature fusion layer uses different convolution kernels to fuse a preset number of frequency domain feature sequences after feature extraction to obtain a feature enhancement sequence; thus achieving feature enhancement of the active sonar received signal of the unmanned underwater vehicle target.
[0079] The pre-built feature enhancement module is trained using a dataset of generative active sonar received signals from autonomously collected unmanned underwater vehicle targets.
[0080] Step 5: Use the pre-built joint feature extraction module to extract and fuse features from the feature enhancement sequence to obtain the fused feature sequence.
[0081] The pre-built joint feature extraction module includes two convolutional computation layers: a one-dimensional convolutional computation layer and a three-dimensional convolutional computation layer.
[0082] The first convolutional computation layer is used to extract features from the feature enhancement sequence using one-dimensional and three-dimensional convolutions to obtain different convolutional sequences. The different convolutional sequences are activated and fused together to obtain a joint convolutional sequence.
[0083] The second convolutional computation layer is used to perform feature fusion on the joint convolutional sequence obtained from the first layer using a one-dimensional convolutional kernel to obtain a fused feature sequence, thereby realizing global and long-range feature extraction of the active sonar received signal of the unmanned underwater vehicle target.
[0084] The pre-built joint feature extraction module is trained using a dataset of generative active sonar received signals from autonomously collected unmanned underwater vehicle targets.
[0085] Step 6: Use the pre-built authenticity assessment module to classify the fused characteristic sequence, and use the probability of the echo signal classified as an unmanned underwater vehicle target as the authenticity assessment result of the generative active sonar received signal of the unmanned underwater vehicle target.
[0086] Example 2
[0087] Based on Example 1, this example provides a method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle (UUV) targets. It trains a pre-constructed hybrid position coding module using a dataset of autonomously collected generative active sonar received signals from UUV targets. Specifically, the method includes:
[0088] S1, Data Preparation.
[0089] The dataset of active sonar received signals from unmanned underwater vehicle targets collected in autonomous experiments is preprocessed to obtain preprocessed active sonar received signals.
[0090] The dataset includes echo signals from unmanned underwater vehicle targets and environmental noise signals.
[0091] Preprocessing can include, but is not limited to, sequentially performing data cleaning, label generation, and dataset partitioning. Data cleaning removes potential missing values, outliers, data representation anomalies, or other interfering factors. This can be achieved through methods such as missing value handling, outlier detection and handling, and data standardization. Clean data helps the model better learn from real-world scenarios.
[0092] S2, set the segmented position encoding layer.
[0093] The segmented location coding layer uses one-dimensional convolutions to divide the preprocessed active sonar received signal and embeds the location information between each segment to obtain a segmented coding sequence; the expression for this segmented location coding layer is:
[0094]
[0095] In the formula: This represents the number of segments in the segmented position encoding. The sequence length of the active sonar received signal; The length of each segment in the segmented position encoding; This represents the step size that each segment spans in the segmented position encoding.
[0096] S3, set the rotation position encoding layer.
[0097] The rotation position coding layer encodes the segmented coding sequence and embeds the positional dependencies between the segments to obtain the rotational coding sequence. The expression for the rotation position coding layer is:
[0098]
[0099]
[0100]
[0101] In the formula: Index for feature dimensions; For the first Rotation parameters on each feature index; The feature dimension of the segmented encoded sequence; The position index of each segment after rotational position encoding; For the rotation matrix in the th... The feature dimension, the first The value at each segment index; This is the result of rotating the position encoding of the segmented position encoding sequence; This is to convert the rotated position encoding into a complex sequence.
[0102] S4, set the normalization layer.
[0103] To avoid the hybrid position coding causing significant changes to the amplitude of the active sonar received signal and thus affecting model training, the hybrid position coding module employs a normalization operation. The expression for the normalization layer is as follows:
[0104]
[0105] In the formula: This is the position encoding result of the output after layer normalization; The result of segmented position encoding; Segmented encoding for rotational position encoding The result after rotation encoding.
[0106] S5, Set the loss function.
[0107] To optimize the encoding quality of the hybrid positional coding module, it is necessary to calculate the module's training loss. The training loss is used to optimize the encoding capability of the hybrid positional coding, and the loss function is:
[0108]
[0109] In the formula: The optimized loss value for the hybrid location coding module; These are the signal characteristics after processing by the hybrid position coding module; Characteristics of active sonar received signals; This represents the restoration loss of the hybrid position encoding module.
[0110] S7, Optimization and Training.
[0111] The training loss and restoration loss are calculated using the hybrid positional encoding module in S6, and the learnable parameters in the model are optimized using the backpropagation algorithm.
[0112] Example 3
[0113] Based on Example 1, this example provides a method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets. It utilizes a pre-built wavelet analysis denoising module to analyze and calculate the mixed position coding sequence to reduce noise in the active sonar received signal, resulting in a denoised signal sequence, specifically including:
[0114] S1, signal filtering.
[0115] One-dimensional convolution is used to perform low-pass and high-pass filtering on the mixed position-coded sequence, with the stride of the one-dimensional convolution set to 2. The filtering operation is as follows:
[0116] Low-pass filter:
[0117]
[0118] In the formula: This is the result after low-pass filtering; This is a one-dimensional convolution operation; For mixed positional encoding sequences; This is a low-pass filter with a filter threshold set to [0.7071, 0.7071].
[0119] High-pass filter:
[0120]
[0121]
[0122] In the formula: This is the result after high-pass filtering; This is a high-pass filter, with the filter threshold set to [-0.7071, 0.7071]; This is the result after thresholding the high-frequency portion; This is a soft thresholding denoising function; This is the threshold parameter.
[0123] S2, signal sampling.
[0124] Upsampling of low-pass and high-pass signals is performed using nearest-neighbor interpolation. The upsampling operation is as follows:
[0125]
[0126]
[0127] In the formula: This is the result of upsampling the low-frequency portion; For interpolation operations, nearest neighbor interpolation is used, and the value of each newly added sampling point is taken from the value of the nearest known sampling point; This is the result after low-pass filtering; This is the upsampling scaling factor, set to a value of 2; This is the upsampling result for the high-frequency portion; This is the result after thresholding the high-frequency portion.
[0128] S3, signal reconstruction.
[0129] Signal reconstruction is achieved by using one-dimensional convolution on the upsampled result. The reconstruction operation is as follows:
[0130]
[0131]
[0132]
[0133] In the formula: The signal is reconstructed by filtering the low-pass signal; This is the result of upsampling the low-frequency portion; This is a low-pass filter, and the filter threshold is set to [0.7071, 0.7071]. The signal is reconstructed by filtering the high-pass signal; This is the upsampling result for the high-frequency portion; This is a high-pass filter, with the filter threshold set to [-0.7071, 0.7071]; This is a denoised signal sequence.
[0134] Example 4
[0135] Based on Example 1, this example provides a method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle (UUV) targets. It trains a pre-constructed feature enhancement module using a dataset of autonomously collected generative active sonar received signals from UUV targets. Specifically, the method includes:
[0136] S1, set the successive variational mode decomposition layer.
[0137] Modal decomposition is performed on the denoised signal sequence, which requires decomposing the signal and filtering it to complete the reconstruction.
[0138] The signal decomposition operation is as follows:
[0139]
[0140]
[0141] In the formula: The signal embedding matrix is constructed from the denoised signal. It is a left singular vector matrix; It is a singular value diagonal matrix; It is a right singular vector matrix.
[0142] The operation to reconstruct the component matrix is as follows:
[0143]
[0144] In the formula: For the reconstruction of the first Each component matrix; For the first One left singular vector matrix component; For the first 1 singular value diagonal matrix component; For the first The right singular vector matrix components.
[0145] The reconstructed component matrix is mapped back to the signal components by diagonal averaging. The mapping reconstruction operation is as follows:
[0146]
[0147] In the formula: For the component matrix at time... Restore the numerical values of the signal sequence; This refers to the size of the mode decomposition window; The number of columns in the embedding matrix is determined by the signal sequence length and the window size; For the first Component matrices in The value at the location; For indicator functions; To recover the moments in the signal sequence; This represents the number of times the time point is accumulated.
[0148] S2, set up a self-attention feature extraction layer.
[0149] A self-attention mechanism is used to extract features from the signal reconstructed by successive variational mode decomposition, resulting in multiple feature sequences. The self-attention feature extraction operation is as follows:
[0150]
[0151] In the formula: This is the final weighted feature output; Score for attention; The product of the transposes of the Query matrix and the Key matrix represents the feature similarity at different positions. The dimension of the Key vector; The Value matrix; It is a normalized exponential function.
[0152] S3, set the feature fusion layer.
[0153] Two-dimensional convolution is used to extract global and local features separately, and feature fusion is then achieved. The feature fusion operation is as follows:
[0154] Local feature attention extraction:
[0155]
[0156] In the formula: This represents the results of local feature attention extraction. This is a two-dimensional convolution operation; For activation functions; For batch regularization operations; The input feature sequence.
[0157] Global feature attention extraction:
[0158]
[0159] In the formula: This represents the result of global feature attention extraction. This is a two-dimensional average pooling operation.
[0160] Attention weight fusion:
[0161]
[0162]
[0163] In the formula: The result includes both global and local features; It is the sigmoid function; These are the feature weights.
[0164] Feature fusion:
[0165]
[0166] In the formula: Output the feature fusion result; Residual connections for self-attention feature sequences; The input feature sequence.
[0167] S4, Set the loss function.
[0168] To optimize the encoding quality of the feature enhancement module, it is necessary to calculate the module's training loss. The training loss is used to optimize the feature enhancement modeling capability of the feature enhancement module, and the loss function is:
[0169]
[0170] In the formula: The loss value after optimization for the feature enhancement module; These are the features after processing by the feature enhancement module; For target features; This represents the restoration loss of the feature enhancement module.
[0171] S5, Optimization and Training.
[0172] The training loss and restoration loss, calculated in S4, are used to optimize the learnable parameters in the model using the backpropagation algorithm.
[0173] Example 5
[0174] Based on Example 1, this example provides a method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle (UUV) targets. It trains a pre-constructed joint feature extraction module using a dataset of autonomously collected generative active sonar received signals from UUV targets. Specifically, the method includes:
[0175] S1 sets the first convolutional computation layer.
[0176] Joint feature extraction is achieved by using one-dimensional and three-dimensional convolutional kernels on the feature-enhanced signal, thereby extracting global features and long-range dependency features of the feature signal. The joint feature extraction operation is as follows:
[0177] Convolutional feature extraction:
[0178]
[0179]
[0180] In the formula: This is the result of the first type of convolutional feature extraction; Use the GELU activation function; The input feature sequence for the joint feature extraction module; This is a one-dimensional convolution operation; This is a 3D convolution operation; This is the result of the second type of convolution feature extraction.
[0181] S2 sets up the second convolutional computation layer.
[0182] One-dimensional convolution is used to fuse global features and long-range dependent features. The feature fusion operation is as follows:
[0183]
[0184] In the formula: The output sequence of the joint feature extraction module; It is a one-dimensional convolution kernel.
[0185] S3, Set the loss function.
[0186] To optimize the feature extraction quality of the joint feature extraction module, it is necessary to calculate the module's training loss. The training loss is used to optimize the feature extraction capability of the joint feature extraction module, and the loss function is:
[0187]
[0188] In the formula: The optimized loss value for the joint feature extraction module; These are the features after processing by the joint feature extraction module; For target features; This represents the restoration loss of the joint feature extraction module.
[0189] S4, Optimization and Training.
[0190] The training loss and restoration loss are calculated using the joint feature extraction module in S3, and the learnable parameters in the model are optimized using the backpropagation algorithm.
[0191] Example 6
[0192] Based on Example 1, this example provides a method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle (UUV) targets. It utilizes a pre-built authenticity evaluation module to classify the fused characteristic sequence and uses the probability of the echo signal classified as an UUV target as the authenticity evaluation result of the UUV target's generative active sonar received signal. The specific method includes the following steps:
[0193] A linear classification layer is used to classify the fused feature sequence to obtain the probability of classifying the generative active sonar received signal of the unmanned underwater vehicle (UUV) target as the echo signal of the UUV target; this probability is used as the authenticity evaluation result of the generative active sonar received signal of the UUV target.
[0194] Example 7
[0195] Based on Example 1, this example provides a method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets. The specific method for decomposing the number of modes in successive variational mode decomposition includes:
[0196] S1, Envelope Entropy Analysis.
[0197] The envelope entropy evaluation method calculates the average envelope entropy of each modal component and selects the modal decomposition number that minimizes the average envelope entropy. The envelope entropy calculation operation is as follows:
[0198] Calculate the analytic signal:
[0199]
[0200] In the formula: To analyze the signal; It is a signal sequence; The result of applying a Hilbert transform to the signal.
[0201] Calculate the signal envelope:
[0202]
[0203] In the formula: It is the envelope of the signal, which is numerically equal to the modulus of the analytic signal.
[0204] Calculate modal energy:
[0205]
[0206] In the formula: This refers to the modal energy value. This represents the number of samples in the modality. The signal is decomposed into non-overlapping modes.
[0207] Calculate the envelope entropy:
[0208]
[0209] In the formula: Let be the entropy of the signal envelope.
[0210] S2, Energy Proportion Analysis.
[0211] The energy proportion assessment method calculates the proportion of the new mode's energy in the total signal energy after each new mode is added. Mode decomposition is considered to stop when the proportion is less than a set threshold. The energy proportion calculation operation is as follows:
[0212] Modal component energy calculation:
[0213]
[0214] In the formula: The energy value for a single mode; The length of the signal sequence; For the first One modal component.
[0215] Total energy calculation:
[0216]
[0217] In the formula: The sum of the modal energies; This represents the number of modes in the signal decomposition.
[0218] Energy percentage calculation:
[0219]
[0220] In the formula: For the first The energy percentage of each modal component.
[0221] S3, Select the number of decomposition modes.
[0222] The decomposed mode number was obtained using the envelope entropy analysis evaluation method. To minimize the envelope entropy, the number of decomposed modes is obtained using the energy proportion analysis evaluation method. This results in a smaller energy percentage after further mode decomposition. (Comparison) , Are they equal? , If they are equal, then select As the modality decomposition number; if , If they are not equal, choose the larger one. The value is used as the modality decomposition number.
[0223] Example 8
[0224] This embodiment provides a generative active sonar received signal authenticity assessment system for unmanned underwater vehicle targets, comprising:
[0225] The data preprocessing module is used to preprocess the generative active sonar received signal of the unmanned underwater vehicle target to obtain the preprocessed active sonar received signal.
[0226] The hybrid position coding module is used to perform segmented position coding and rotational position coding on the preprocessed active sonar received signal to obtain a hybrid position coding sequence.
[0227] The wavelet analysis denoising module is used to denoise the mixed position-coded sequence to obtain a denoised signal sequence.
[0228] The feature enhancement module is used to perform mode decomposition, feature extraction, and feature fusion on the denoised signal sequence to obtain the feature-enhanced sequence;
[0229] The joint feature extraction module is used to extract and fuse features from the feature-enhanced sequence to obtain a fused feature sequence.
[0230] The authenticity assessment module is used to classify the fused characteristic sequence and use the probability of the echo signal classified as an unmanned underwater vehicle target as the authenticity assessment result of the generative active sonar received signal of the unmanned underwater vehicle target.
[0231] Example 9
[0232] This embodiment also provides a computing device. The computing device includes a bus, a processor, a memory, and a communication interface. The processor, memory, and communication interface communicate with each other via the bus. The computing device can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memory in the computing device.
[0233] A bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, a bus can include a path for transmitting information between various components of a computing device (e.g., memory, processor, communication interfaces).
[0234] The processor may include any one or more of the following: central processing unit (CPU), graphics processing unit (GPU), tensor processing unit (TPU), application specific integrated circuit (ASIC), field-programmable gate array (FPGA), microprocessor (MP), or digital signal processor (DSP).
[0235] The memory may include volatile memory, such as random access memory (RAM). The processor may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).
[0236] The memory stores executable program code, which the processor executes to implement the functions of the aforementioned units, thereby achieving, for example, the method described in Embodiment 1. That is, the memory may store instructions for the methods and functions relating to the computing device in any of the above embodiments.
[0237] The communication interface uses transceiver modules such as, but not limited to, network interface cards and transceivers to enable communication between computing devices and other devices or communication networks.
[0238] Example 10
[0239] This embodiment also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, cause the processor to perform the methods and functions of the computing device involved in any of the above embodiments.
[0240] Generally, the various embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software, which can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are shown and described as block diagrams, flowcharts, or represented using some other illustration, it should be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as, as non-limiting examples, in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.
[0241] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets, characterized in that, Includes the following steps: The generative active sonar received signal of the unmanned underwater vehicle target is preprocessed to obtain the preprocessed active sonar received signal. The preprocessed active sonar received signal is subjected to segmented position coding and rotational position coding to obtain a hybrid position coding sequence. Specifically, the preprocessed active sonar received signal is divided using one-dimensional convolution, and the position information between each segment is embedded to obtain the segmented coding sequence. Encode the segmented coded sequence to obtain the rotated coded sequence; The rotated encoded sequence is normalized to obtain the hybrid position encoded sequence; The mixed position-coded sequence is denoised to obtain a denoised signal sequence; The denoised signal sequence is processed by mode decomposition, feature extraction and feature fusion to obtain a feature-enhanced sequence. Successive variational mode decomposition is used to decompose the active sonar received signal into different frequency bands. After reconstructing the signals of different frequency bands, the self-attention extraction method is used to extract the signal features of each frequency band. Convolution is used to complete the fusion of each feature vector. Feature extraction and fusion processing are performed on the feature-enhanced sequence to obtain a fused feature sequence; specific methods include: Feature extraction is performed on the feature enhancement sequence using one-dimensional and three-dimensional convolutions to obtain different convolution sequences. The different convolution sequences are activated and fused together to obtain a joint convolution sequence. The joint convolution sequence obtained in the first layer is then fused using a one-dimensional convolution kernel to obtain a fused feature sequence. The fused feature sequences are classified, and the probability of the echo signal classified as an unmanned underwater vehicle target is used as the authenticity evaluation result of the generative active sonar received signal of the unmanned underwater vehicle target.
2. The method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets according to claim 1, characterized in that, The preprocessed active sonar received signal is divided using the following formula, and the positional information between each segment is embedded to obtain a segmented coded sequence: In the formula: This represents the number of segments in the segmented position encoding. The sequence length of the active sonar received signal; The length of each segment in the segmented position encoding; This represents the step size for each segment in the segmented position encoding. The segmented coded sequence is encoded using the following formula to obtain the rotated coded sequence: In the formula: Index for feature dimensions; For the first Rotation parameters on each feature index; The characteristic dimension of the signal received by active sonar; The position index of each segment after rotational position encoding; For the rotation matrix in the th... The feature dimension, the first The value at each segment index; This is the result of rotating the position encoding of the segmented position encoding sequence; This is to convert the rotated position encoding into a complex sequence.
3. The method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets according to claim 1, characterized in that, Modality decomposition, feature extraction, and feature fusion are performed on the denoised signal sequence to obtain the feature-enhanced sequence. Specific methods include: Modal decomposition is performed on the mixed position-coded sequence to obtain signals in multiple frequency bands; Multiple feature sequences are obtained by extracting features from signals across multiple frequency bands. Multiple feature sequences are fused to obtain a feature-enhanced sequence.
4. The method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets according to claim 3, characterized in that, Feature extraction is performed on signals across multiple frequency bands using the following formula: In the formula: This is the final weighted feature output; Score for attention; The product of the transposes of the Query matrix and the Key matrix represents the feature similarity at different positions. The dimension of the Key vector; The Value matrix; Two-dimensional convolution is used to perform global feature extraction and local feature extraction on the feature enhancement sequence, respectively, to obtain the local feature attention extraction result and the global feature attention extraction result; Calculate feature weights using the results of local feature attention extraction and global feature attention extraction; The features are fused using the obtained feature weights to obtain a feature-enhanced sequence.
5. The method for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets according to claim 1, characterized in that, Feature extraction is performed on the feature enhancement sequence using the following formula to obtain different convolutional sequences. These different convolutional sequences are then activated and fused together to obtain a joint convolutional sequence: In the formula: This is the result of the first type of convolutional feature extraction; Use the GELU activation function; The input feature sequence for the joint feature extraction module; This is a one-dimensional convolution operation; This is a 3D convolution operation; This is the result of the second type of convolution feature extraction; The following formula is used to fuse the features of the joint convolutional sequence obtained in the first layer, resulting in a fused feature sequence: In the formula: This is the output sequence of the joint feature extraction module.
6. A system for evaluating the authenticity of generative active sonar received signals for unmanned underwater vehicle targets, based on the evaluation method according to any one of claims 1 to 5, characterized in that, include: The data preprocessing module is used to preprocess the generative active sonar received signal of the unmanned underwater vehicle target to obtain the preprocessed active sonar received signal. The hybrid position coding module is used to perform segmented position coding and rotational position coding on the preprocessed active sonar received signal to obtain a hybrid position coding sequence. The wavelet analysis denoising module is used to denoise the mixed position-coded sequence to obtain a denoised signal sequence. The feature enhancement module is used to perform mode decomposition, feature extraction, and feature fusion on the denoised signal sequence to obtain the feature-enhanced sequence; The joint feature extraction module is used to extract and fuse features from the feature-enhanced sequence to obtain a fused feature sequence. The authenticity assessment module is used to classify the fused characteristic sequence and use the probability of the echo signal classified as an unmanned underwater vehicle target as the authenticity assessment result of the generative active sonar received signal of the unmanned underwater vehicle target.
7. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer instructions that, when executed by the processor, cause the electronic device to perform the method of any one of claims 1 to 5.
8. A computer program product, characterized in that, The computer program product includes computer-executable instructions that, when executed, implement the method according to any one of claims 1 to 5.