Small sample sonar reverberation data enhancement and target detection method and system

Abstract

The invention belongs to the technical field of sonar data processing, and particularly relates to a small sample sonar reverberation data enhancement and target detection method and system. According to the small sample sonar reverberation data enhancement and target detection method, spectrum sensing loss and multi-scale short-time Fourier transform feature constraints are introduced through WGAN-GP; the cosine similarity between the time-frequency domain distribution of the generated reverberation signal and real data is greater than or equal to 0.92, so that the problem of mode collapse of a traditional generative adversarial network model under a small sample is solved, and the data diversity is improved; a sequential structure and a state transition rule of a sonar detection data set are coded into a two-dimensional image through a Gramer angle field and a Markov transition field, the feature discrimination degree of a target and reverberation is improved in combination with a CBAM attention mechanism, the detection accuracy can be improved by a multi-modal convolutional network in a small sample data scene, the multi-modal convolutional network does not depend on a large sample data volume any more, and the detection efficiency is improved. And overfitting caused by less sample data is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of sonar data processing, and specifically to a small sample sonar reverberation data enhancement and target detection method and system. Background Art

[0002] In the field of ocean detection and target identification, active sonar detects and identifies underwater targets by emitting sound waves and receiving echoes. However, the performance of active sonar detection solutions is limited by reverberation interference and small sample data in complex ocean environments, making it less than ideal. Reverberation, the superimposed signal formed by sound waves reflecting off scatterers such as the sea surface, seabed, and plankton, often has much higher energy than the target echo, resulting in a low signal-to-mixing ratio (SRR). This makes it difficult for traditional methods such as matched filtering and adaptive filtering to effectively distinguish between targets and reverberation.

[0003] At the same time, due to the complexity of the marine environment (such as thermohaline jump layers and variability of seabed topography), the high cost of sonar data collection (requiring long-term sea trials on ship platforms) and limited experimental conditions, the actual amount of available sonar reverberation and target echo data is extremely small, forming a typical small-sample learning scenario, resulting in insufficient model training and weak generalization ability, which seriously restricts the practical application of sonar target detection technology.

[0004] In small sample sizes, insufficient data exacerbates the inductive bias of traditional deep learning models (such as CNNs), making it difficult to learn the underlying distribution differences between reverberation and target echoes. Although techniques such as transfer learning and meta-learning have been explored, the time-frequency characteristics of sonar signals differ significantly from those of modalities such as images and text. Existing feature conversion methods (such as direct spectrogram input) fail to fully exploit the signal's temporal structure (such as state transition patterns) and phase information, limiting the model's ability to detect weak and distant targets.

[0005] Traditional signal processing methods (such as fractional Fourier transforms, time-reversal mirrors combined with Wignerhof transforms) suppress reverberation through time-frequency filtering. This approach relies on a priori differences between the target and the reverberation in the time-frequency domain and has limited adaptability to non-stationary and non-Gaussian reverberation. While machine learning-based methods (such as support vector machines and BP neural networks) can automatically extract features, they lack the ability to represent the features of one-dimensional signals, and model parameters are difficult to optimize effectively with small sample sizes. Consequently, detection accuracy decreases significantly as the signal-to-mixture ratio decreases.

[0006] In summary, the existing sonar target detection technology solutions rely on large sample data volumes, which easily leads to overfitting when the sample data is small, seriously restricting the practical application of sonar target detection technology. Summary of the Invention

[0007] The main purpose of the present invention is to provide a small sample sonar reverberation data enhancement and target detection method and system, aiming to solve the problem that the existing sonar target detection technology solutions rely on large sample data volumes and are prone to overfitting when the sample data is small.

[0008] The technical solution proposed by the present invention is: A small sample sonar reverberation data enhancement and target detection method is applied to a small sample sonar reverberation data enhancement and target detection system; the system includes a reverberation data enhancement module, a target echo simulation module, and a feature conversion and fusion detection module; the method includes: The reverberation data enhancement module constructs a generative adversarial network model combined with gradient penalty based on Wasserstein distance; The reverberation data enhancement module obtains a real reverberation signal and random noise, and learns the time-frequency characteristics of the real reverberation signal through a generator of an adversarial network model combined with a gradient penalty to generate enhanced reverberation data consistent with the real ocean environment; The target echo simulation module calculates the time delay, phase shift and amplitude based on the bright spot model and the geometric characteristics and material acoustic properties of the target to be detected to generate a simulated echo signal of the target to be detected; The target echo simulation module fuses the simulated echo signal with the enhanced reverberation data to construct a sonar detection data set; The feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through the Gram angle field and the Markov transfer field; The feature conversion and fusion detection module inputs the two-dimensional feature image into a multimodal convolutional network for feature fusion and target classification, and outputs a target detection result.

[0009] Preferably, the adversarial network model combined with gradient penalty includes a generator and a discriminator; the generator includes: Input layer: used to receive random noise vector; Fully connected layer and residual layer: used for preliminary transformation and feature extraction of noise; Upsampling layer and convolution layer: used to increase the dimension and process the data, in conjunction with batch normalization layer and ReLU activation function; Output layer: used to output the enhanced reverberation signal through the transposed convolution layer and Tanh activation function to approximate the real reverberation distribution.

[0010] Preferably, the discriminator includes: Input layer: used to receive real reverberation signals or generate reverberation signals; Convolutional module: It consists of a convolutional layer, a leaky ReLU activation function, and a random dropout layer for feature extraction; Output layer: used to output the true and false discrimination probability through the flattening layer, linear layer and Sigmoid activation function.

[0011] Preferably, the target echo simulation module calculates the time delay, phase shift and amplitude based on the bright spot model and the geometric features and material acoustic properties of the target to be detected to generate a simulated echo signal of the target to be detected, including: The target echo simulation module calculates the time delay according to the distance between the target to be detected and the sonar and the sound speed of the sonar; The target echo simulation module calculates the phase shift by combining the posture of the target to be detected and the frequency of the incident sound wave; The target echo simulation module calculates the amplitude based on the scattering cross section and the acoustic wave attenuation characteristics of the target to be detected.

[0012] Preferably, the feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through the Gram angle field and the Markov transfer field, including: The feature conversion and fusion detection module normalizes the one-dimensional sonar signal to the interval [-1, 1]; The feature conversion and fusion detection module performs polar coordinate encoding on the normalized one-dimensional sonar signal, converting the value of the one-dimensional sonar signal into an angle or , and convert the timestamp of the one-dimensional sonar signal into a radius , where N is the sequence length; The feature conversion and fusion detection module constructs a Gram matrix to form a two-dimensional feature image, where the elements of the Gram matrix are or .

[0013] Preferably, the feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through the Gram angle field and the Markov transfer field, and further includes: The feature conversion and fusion detection module quantizes the one-dimensional sonar signal into Q quantile units; The feature conversion and fusion detection module constructs a Q×Q Markov transfer matrix; The feature conversion and fusion detection module expands the time step τ and calculates the state transition probability under different lag steps to generate a two-dimensional transition probability matrix containing time information.

[0014] Preferably, the multimodal convolutional network includes: Grammar angular field feature branch: used to perform convolution, batch normalization, activation through the ReLU function, maximum pooling and convolution block attention module on the Grammar angular field image in sequence to extract key features in spatial and channel dimensions; Markov transfer field feature branch: used to process the Markov transfer field image using the same process as the Gram angle field feature branch; Fusion detection layer: used to splice the Gram angular field feature branch and the Markov transfer field feature branch, and then pass through the flattening layer, the fully connected layer, the ReLU function for activation and the random inactivation layer, and the output layer to calculate the target classification probability.

[0015] Preferably, the target echo simulation module fuses the simulated echo signal with the enhanced reverberation data to construct a sonar detection data set, including: The target echo simulation module superimposes the simulated echo signal and the enhanced reverberation data according to a signal-to-mixing ratio.

[0016] Preferably, it also includes: The feature conversion and fusion detection module trains the multimodal convolutional network, including: The feature conversion and fusion detection module uses the cross entropy loss function to calculate the difference between the predicted probability and the true label; The feature conversion and fusion detection module uses a stochastic gradient descent optimizer to dynamically adjust the learning rate; The feature conversion and fusion detection module uses 70% of the sonar detection dataset as a training set, 15% of the sonar detection dataset as a validation set, and 15% of the sonar detection dataset as a test set.

[0017] The present invention also proposes a small sample sonar reverberation data enhancement and target detection system, which applies a small sample sonar reverberation data enhancement and target detection method; the system includes a reverberation data enhancement module, a target echo simulation module and a feature conversion and fusion detection module.

[0018] The above technical solution can achieve the following beneficial effects: The small sample sonar reverberation data enhancement and target detection method proposed in this paper introduces spectrum perception loss and multi-scale short-time Fourier transform feature (STFT) constraints through WGAN-GP, and generates a time-frequency domain distribution of the reverberation signal with a cosine similarity of ≥0.92 with the real data, which solves the mode collapse problem of the traditional generative adversarial network model under small samples, improves data diversity, and provides sufficient samples for model training; the temporal structure and state transition law of the sonar detection dataset are encoded into a two-dimensional image through the Gram angle field and Markov transfer field, and combined with the CBAM attention mechanism, the characteristics of the target and the reverberation are distinguished. The feature separability under low signal-to-mixture ratio conditions is improved compared with the traditional spectrum graph method; this method introduces a multimodal convolutional network, which can improve the detection accuracy in small sample data scenarios, which is better than the single feature model, and thus meets the low signal-to-mixture ratio detection needs in real sea environments, so it no longer relies on large sample data volumes and avoids overfitting when the sample data is small; in addition, by fine-tuning the model parameters with a small amount of real data, cross-domain migration from simulation data to real sea data is achieved. After the migration, the detection accuracy of the model in unknown sea areas decreases, which is significantly better than the existing algorithm and adapts to the diversity of complex marine environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0020] Figure 1 This is a flowchart of the steps of the first embodiment of the small sample sonar reverberation data enhancement and target detection method proposed by the present invention; Figure 2 A schematic diagram of the Gram's angle field characteristics of a fifth embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention; Figure 3 This is a schematic diagram of the Markov transfer field characteristics of the sixth embodiment of the small sample sonar reverberation data enhancement and target detection method proposed by the present invention. DETAILED DESCRIPTION

[0021] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0022] The present invention proposes a small sample sonar reverberation data enhancement and target detection method and system.

[0023] As attached Figure 1As shown, in a first embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention, the small sample sonar reverberation data enhancement and target detection method is applied to a small sample sonar reverberation data enhancement and target detection system; the system includes a reverberation data enhancement module, a target echo simulation module, and a feature conversion and fusion detection module; this embodiment includes the following steps: Step S110: The reverberation data enhancement module constructs a generative adversarial network model (WGAN-GP) combined with gradient penalty based on Wasserstein distance.

[0024] Specifically, Wasserstein distance (i.e., Wasserstein-distance), also known as Earth-Mover's-Distance (EMD), is a distance metric used in metric theory to compare the differences between two probability distributions.

[0025] Traditional generative adversarial network (GAN) models, based on the JS divergence optimization objective, suffer from vanishing gradients and mode collapse. WGAN-GP addresses this issue by introducing the Wasserstein distance as an alternative to the JS divergence. The Wasserstein distance provides a continuous gradient signal, enabling stable optimization even when the generated distribution differs significantly from the true distribution.

[0026] Step S120: The reverberation data enhancement module obtains a real reverberation signal and random noise, and learns the time-frequency characteristics of the real reverberation signal through a generator of an adversarial network model combined with a gradient penalty to generate enhanced reverberation data consistent with the real ocean environment.

[0027] Specifically, by introducing spectrum perception loss and multi-scale short-time Fourier transform (STFT) feature constraints through WGAN-GP, the cosine similarity between the time-frequency domain distribution of the generated real reverberation signal and the real data is ≥0.92.

[0028] Step S130: The target echo simulation module calculates the time delay, phase shift and amplitude based on the bright spot model and the geometric features and material acoustic properties of the target to be detected to generate a simulated echo signal of the target to be detected.

[0029] Step S140: The target echo simulation module fuses the simulated echo signal with the enhanced reverberation data to construct a sonar detection data set.

[0030] Step S150: The feature conversion and fusion detection module converts the sonar detection dataset into a two-dimensional feature image through Gramian-Angular-Field (GAF) and Markov Transition Field (MTF).

[0031] Specifically, by combining the CBAM attention mechanism (a convolutional block attention module consisting of two sub-modules: channel attention and spatial attention), the feature differentiation between the target and enhanced reverberation data is improved. The feature separability under low signal-to-mixture ratio conditions is improved compared to traditional spectrogram methods.

[0032] Step S160: The feature conversion and fusion detection module inputs the two-dimensional feature image into a multimodal convolutional network for feature fusion and target classification, and outputs a target detection result.

[0033] The small sample sonar reverberation data enhancement and target detection method proposed in this paper introduces spectrum perception loss and multi-scale short-time Fourier transform feature (STFT) constraints through WGAN-GP, and generates a time-frequency domain distribution of the reverberation signal with a cosine similarity of ≥0.92 with the real data, which solves the mode collapse problem of the traditional generative adversarial network model under small samples, improves data diversity, and provides sufficient samples for model training; the temporal structure and state transition law of the sonar detection dataset are encoded into a two-dimensional image through the Gram angle field and Markov transfer field, and combined with the CBAM attention mechanism, the characteristics of the target and the reverberation are distinguished. The feature separability under low signal-to-mixture ratio conditions is improved compared with the traditional spectrum graph method; this method introduces a multimodal convolutional network, which can improve the detection accuracy in small sample data scenarios, which is better than the single feature model, and thus meets the low signal-to-mixture ratio detection needs in real sea environments, so it no longer relies on large sample data volumes and avoids overfitting when the sample data is small; in addition, by fine-tuning the model parameters with a small amount of real data, cross-domain migration from simulation data to real sea data is achieved. After the migration, the detection accuracy of the model in unknown sea areas decreases, which is significantly better than the existing algorithm and adapts to the diversity of complex marine environments.

[0034] In a second embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention, based on the first embodiment, the adversarial network model combined with gradient penalty includes a generator and a discriminator; the generator includes: Input layer: used to receive random noise vectors.

[0035] Fully connected layer and residual layer: used for preliminary transformation of noise and feature extraction.

[0036] Upsampling layer and convolution layer: used to increase the dimension and process data, in conjunction with the batch normalization layer (BatchNorm) and ReLU activation function (linear rectification function, also known as rectified linear unit, is a commonly used activation function in artificial neural networks).

[0037] Output layer: used to output the enhanced reverberation signal through the transposed convolution layer and the Tanh activation function (hyperbolic tangent activation function, which is a deformation of the Sigmoid activation function) to approximate the real reverberation distribution.

[0038] Specifically, the generator employs a multi-layer residual network structure. Its input is a uniformly distributed random noise vector. After being mapped to a low-dimensional feature space via a fully connected layer, nonlinear features are extracted using residual blocks (each consisting of a convolutional layer, a batch normalization layer, and a ReLU activation function). Upsampling layers are used to gradually increase the feature dimension to the length of the sonar signal. Finally, a Tanh activation function is used to output enhanced reverberation data in the range [-1, 1], ensuring that the energy distribution of the enhanced reverberation data closely resembles that of real reverberation.

[0039] In a third embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention, based on the first embodiment, the discriminator includes: Input layer: used to receive real reverberation signals or generate reverberation signals.

[0040] Convolutional module: It consists of a convolutional layer, a leaky ReLU activation function (LeakyReLU, i.e., a leaky linear rectification function) and a random dropout layer (Dropout), which is used to extract features.

[0041] Output layer: used to output the true and false discrimination probability through the flattening layer, linear layer and Sigmoid activation function, and introduce the gradient penalty term to constrain the discriminator gradient norm.

[0042] Specifically, the Sigmoid function is a monotonically increasing function with an output range between [0,1]. The more negative the value is, the closer it is to 0, and the slower the approximation speed is; the more positive the value is, the closer it is to 1, and the slower the approximation speed is.

[0043] Specifically, the discriminator is used to: use a deep convolutional network, take the real reverberation signal or the generated signal as input, perform feature downsampling through a convolution module, a leaky ReLU activation function, and a random dropout layer, and extract multi-scale time-frequency features; finally, output a single discriminant probability value (0-1) through a flattening layer and a linear layer. The formula for introducing the gradient penalty term (GP) is: , Where, It is a linear interpolation sample between the real and generated signals, which is used to constrain the discriminator gradient norm to be close to 1 to ensure training stability; : represents the gradient penalty loss, which is a loss term used to constrain the discriminator gradient norm. This loss term can ensure the stability of the generative adversarial network training process; : represents expectation, which is used to calculate the average expectation of subsequent random variable values; : represents the output function of the discriminator. The input sample in the brackets is the real reverberation signal, the generated reverberation signal, or the interpolated sample of the two. The output is the probability value related to the authenticity of the sample. : represents the probability distribution of linear interpolation samples between the real reverberation signal and the generated reverberation signal; is the interpolated sample sampled from this distribution; are interpolated samples sampled from the probability distribution of linearly interpolated samples between the real reverberation signal and the generated reverberation signal.

[0044] In a fourth embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention, based on the first embodiment, step S130 includes the following steps: Step S410: The target echo simulation module calculates the time delay according to the distance between the target to be detected and the sonar and the speed of sound of the sonar.

[0045] Step S420: The target echo simulation module calculates the phase shift based on the posture of the target to be detected and the frequency of the incident sound wave.

[0046] Step S430: the target echo simulation module calculates the amplitude based on the scattering cross section and the acoustic wave attenuation characteristics of the target to be detected.

[0047] Specifically, the target scattering characteristics are parameterized and the emission signal is in the form of , the target response function is , the echo signal output by the network response is , the target echo is composed of the superposition of scattered signals from N independent bright spots, and the simulated echo signal is expressed as: , , in: is the time delay; is the phase shift; is the amplitude; : Represents the spatial position vector, which is used to describe the position information of the target, transmitted signal, echo signal, etc. in space; : represents angular frequency. The relationship between angular frequency and frequency f is ω=2πf; Is an imaginary unit.

[0048] In a fifth embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention, based on the first embodiment, step S150 includes the following steps: Step S510: The feature conversion and fusion detection module normalizes the one-dimensional sonar signal to the interval [-1, 1].

[0049] The schematic diagram of Gram's angle field characteristics is attached. Figure 2 As shown in FIG, the calculation formula for the above normalization processing is: .

[0050] in, Represents a sonar signal sequence The maximum value in ; Represents a sonar signal sequence The minimum value in .

[0051] Step S520: The feature conversion and fusion detection module performs polar coordinate encoding on the normalized one-dimensional sonar signal, converting the value of the one-dimensional sonar signal into an angle. or , and convert the timestamp of the one-dimensional sonar signal into a radius , where N is the sequence length.

[0052] Step S530: The feature conversion and fusion detection module constructs a Gram matrix to form a size of The two-dimensional feature image of , where the elements of the Gram matrix are or .

[0053] Specifically, is the normalized one-dimensional sonar signal Sample values ​​that satisfy: By calculating the trigonometric relationship between the angles corresponding to different sample points: or , to construct a two-dimensional feature image.

[0054] In a sixth embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention, based on the fifth embodiment, step S150 further includes the following steps: Step S610: The feature conversion and fusion detection module quantizes the one-dimensional sonar signal into Q quantile units.

[0055] The characteristic diagram of Markov transition field is shown in the attached figure. Figure 3 shown.

[0056] Specifically, in this embodiment, Q=8 quantile units.

[0057] Map each data point to a state: .

[0058] in, Indicates the state identifier of the quantized one-dimensional sonar signal; in the process of quantizing the one-dimensional sonar signal into Q quantile units, each data point will be mapped to one of these Q states. qk is used to represent the state number corresponding to the k-th data point. Its value range is {1,2,...,Q}, which means that the state of the quantized sonar signal can only be a value in this set.

[0059] Step S620: The feature conversion and fusion detection module constructs a Q×Q Markov transfer matrix, where the element w of the Markov transfer matrix is ij is the transition probability from state i to state j.

[0060] Specifically, the steps of constructing the Q×Q Markov transition matrix include: Statistics of different time steps: The number of state transitions under .

[0061] in, : Indicates the time step, which is used to measure the time interval between two states. The value range is ; : Indicates the maximum time step, its value is ( is the total number of sample points, Indicates rounding down), used to limit the upper limit of the time step; :Indicates that the time step is When the status changes from Transfer to The probability of is calculated as ; :Indicates the time The state of the sonar signal at that time; :Indicates the A quantized state is one of the possible states of the sonar signal after quantization, and its value range is ( is the number of quantile units); :Indicates the quantized state, and Similarly, it is one of the quantized states, and its value range is ; :Indicates the time The state of the sonar signal at that time.

[0062] Step S630: The feature conversion and fusion detection module expands the time step τ and calculates the state transition probabilities under different lag steps to generate a two-dimensional transition probability matrix containing time information.

[0063] Specifically include: Will The tensors are accumulated along the time dimension to generate a Q×Q Markov transfer field matrix, which is normalized to [0,1] and used as a two-dimensional feature image.

[0064] In a seventh embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention, based on the first embodiment, the multimodal convolutional network includes: Gram Angular Field feature branch (GAF / MTF shared structure): used to perform convolution, batch normalization (BatchNorm), activation through the ReLU function, maximum pooling and convolutional block attention module (CBAM) processing on the Gram Angular Field image in sequence to extract key features in spatial and channel dimensions.

[0065] Specifically, the Gram angular field feature branch inputs a two-dimensional image and processes it through three layers of convolution blocks: the first layer: convolution layer → batch normalization → leaky ReLU activation function layer → maximum pooling layer; the second layer: convolutional block attention module (CBAM): used to combine channel attention and spatial attention to enhance target-related features; the third layer: depthwise separable volume → batch normalization → ReLU activation function layer to reduce computational complexity.

[0066] Markov transfer field feature branch: used to apply the same processing flow as the Gram angle field feature branch to the Markov transfer field image.

[0067] Fusion detection layer: used to splice the Gram angular field feature branch and the Markov transfer field feature branch, and then pass through the flattening layer, the fully connected layer, the ReLU function for activation and the random inactivation layer (Dropout) to calculate the target classification probability through the output layer.

[0068] Specifically, the fusion detection layer splices the feature maps output by the GAF / MTF branches, converts them into one-dimensional vectors through the flattening layer, extracts high-order features through two fully connected layers, and uses a random dropout layer to prevent overfitting. Finally, the output layer calculates the classification probability of each type of target.

[0069] In the eighth embodiment of the small sample sonar reverberation data enhancement and target detection method proposed by the present invention, based on the first embodiment, step S140 includes the following steps: Step S810: the target echo simulation module superimposes the simulated echo signal and the enhanced reverberation data according to a signal-to-reverberation ratio (SNR) to cover scenarios from weak targets to strong targets.

[0070] Specifically, the signal-to-mixing ratio ranges from -10 dB to 20 dB. The target echo simulation module superimposes the simulated echo signal and the enhanced reverberation data according to the signal-to-mixing ratio using the following expression: , Where, is the amplitude scaling factor, and The SNR value range is set to -10dB to 20dB, covering weak target (SNR < 0dB), medium target (0dB ≤ SNR ≤ 10dB), and strong target (SNR > 10dB) scenarios. : represents the sonar detection signal obtained by superimposing the simulated echo signal and the enhanced reverberation data according to the signal-to-mixing ratio. It is the input signal for subsequent feature conversion and target detection. : represents enhanced reverberation data, that is, a reverberation signal consistent with the real ocean environment generated by an adversarial network model combined with gradient penalty; : Indicates the index of the independent bright spot in the target echo, which is used to distinguish different bright spot scattering signals. The value range is ( is the total number of independent bright spots contained in the target echo); :Indicates the The scattered signal of an independent bright spot is the basic unit of the simulated echo signal. Multiple such signals are superimposed to form the simulated echo of the target to be detected.

[0071] In a ninth embodiment of a small sample sonar reverberation data enhancement and target detection method proposed by the present invention, based on the first embodiment, this embodiment further includes the following steps: Step S910: The feature conversion and fusion detection module trains the multimodal convolutional network, including the following steps: Step S911: The feature conversion and fusion detection module uses a cross entropy loss function to calculate the difference between the predicted probability and the true label.

[0072] Step S912: The feature conversion and fusion detection module dynamically adjusts the learning rate using a stochastic gradient descent optimizer (SGD).

[0073] The initial value of the learning rate is 0.0002, and after 30 epochs the learning rate drops to 0.00005.

[0074] Step S913: The feature conversion and fusion detection module uses 70% of the sonar detection dataset as a training set, 15% of the sonar detection dataset as a validation set, and 15% of the sonar detection dataset as a test set to prevent model overfitting.

[0075] The present invention also proposes a small sample sonar reverberation data enhancement and target detection system, which applies a small sample sonar reverberation data enhancement and target detection method; the system includes a reverberation data enhancement module, a target echo simulation module and a feature conversion and fusion detection module.

[0076] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0077] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A small sample sonar reverberation data enhancement and target detection method, characterized in that: Applied to a small sample sonar reverberation data enhancement and target detection system; the system includes a reverberation data enhancement module, a target echo simulation module, and a feature conversion and fusion detection module; the method includes: The reverberation data enhancement module constructs a generative adversarial network model combined with gradient penalty based on Wasserstein distance; The reverberation data enhancement module obtains a real reverberation signal and random noise, and learns the time-frequency characteristics of the real reverberation signal through a generator of an adversarial network model combined with a gradient penalty to generate enhanced reverberation data consistent with the real ocean environment; The target echo simulation module calculates the time delay, phase shift and amplitude based on the bright spot model and the geometric characteristics and material acoustic properties of the target to be detected to generate a simulated echo signal of the target to be detected; The target echo simulation module fuses the simulated echo signal with the enhanced reverberation data to construct a sonar detection data set; The feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through the Gram angle field and the Markov transfer field; The feature conversion and fusion detection module inputs the two-dimensional feature image into a multimodal convolutional network for feature fusion and target classification, and outputs a target detection result.

2. The small sample sonar reverberation data enhancement and target detection method according to claim 1, characterized in that: The adversarial network model combined with gradient penalty includes a generator and a discriminator; the generator includes: Input layer: used to receive random noise vector; Fully connected layer and residual layer: used for preliminary transformation and feature extraction of noise; Upsampling layer and convolution layer: used to increase the dimension and process the data, in conjunction with batch normalization layer and ReLU activation function; Output layer: used to output the enhanced reverberation signal through the transposed convolution layer and Tanh activation function to approximate the real reverberation distribution.

3. The small sample sonar reverberation data enhancement and target detection method according to claim 2, characterized in that: The discriminator comprises: Input layer: used to receive real reverberation signals or generate reverberation signals; Convolutional module: It consists of a convolutional layer, a leaky ReLU activation function, and a random dropout layer for feature extraction; Output layer: used to output the true and false discrimination probability through the flattening layer, linear layer and Sigmoid activation function.

4. The small sample sonar reverberation data enhancement and target detection method according to claim 1, characterized in that: The target echo simulation module calculates the time delay, phase shift, and amplitude based on the bright spot model and the geometric features and material acoustic properties of the target to be detected to generate a simulated echo signal of the target to be detected, including: The target echo simulation module calculates the time delay according to the distance between the target to be detected and the sonar and the sound speed of the sonar; The target echo simulation module calculates the phase shift by combining the posture of the target to be detected and the frequency of the incident sound wave; The target echo simulation module calculates the amplitude based on the scattering cross section and the acoustic wave attenuation characteristics of the target to be detected.

5. The small sample sonar reverberation data enhancement and target detection method according to claim 1, characterized in that: The feature conversion and fusion detection module converts the sonar detection dataset into a two-dimensional feature image through the Gram angle field and the Markov transfer field, including: The feature conversion and fusion detection module normalizes the one-dimensional sonar signal to the interval [-1, 1]; The feature conversion and fusion detection module performs polar coordinate encoding on the normalized one-dimensional sonar signal, converting the value of the one-dimensional sonar signal into an angle or , and convert the timestamp of the one-dimensional sonar signal into a radius , where N is the sequence length; The feature conversion and fusion detection module constructs a Gram matrix to form a two-dimensional feature image, where the elements of the Gram matrix are or .

6. The small sample sonar reverberation data enhancement and target detection method according to claim 5, characterized in that: The feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through the Gram angle field and the Markov transfer field, and also includes: The feature conversion and fusion detection module quantizes the one-dimensional sonar signal into Q quantile units; The feature conversion and fusion detection module constructs a Q×Q Markov transfer matrix; The feature conversion and fusion detection module expands the time step τ and calculates the state transition probability under different lag steps to generate a two-dimensional transition probability matrix containing time information.

7. The small sample sonar reverberation data enhancement and target detection method according to claim 1, characterized in that: The multimodal convolutional network comprises: Grammar angular field feature branch: used to perform convolution, batch normalization, activation through the ReLU function, maximum pooling and convolution block attention module on the Grammar angular field image in sequence to extract key features in spatial and channel dimensions; Markov transfer field feature branch: used to process the Markov transfer field image using the same process as the Gram angle field feature branch; Fusion detection layer: used to splice the Gram angular field feature branch and the Markov transfer field feature branch, and then pass through the flattening layer, the fully connected layer, the ReLU function for activation and the random inactivation layer, and the output layer to calculate the target classification probability.

8. The small sample sonar reverberation data enhancement and target detection method according to claim 1, characterized in that: The target echo simulation module fuses the simulated echo signal with the enhanced reverberation data to construct a sonar detection data set, including: The target echo simulation module superimposes the simulated echo signal and the enhanced reverberation data according to a signal-to-mixing ratio.

9. The small sample sonar reverberation data enhancement and target detection method according to claim 1, characterized in that: Also includes: The feature conversion and fusion detection module trains the multimodal convolutional network, including: The feature conversion and fusion detection module uses the cross entropy loss function to calculate the difference between the predicted probability and the true label; The feature conversion and fusion detection module uses a stochastic gradient descent optimizer to dynamically adjust the learning rate; The feature conversion and fusion detection module uses 70% of the sonar detection dataset as a training set, 15% of the sonar detection dataset as a validation set, and 15% of the sonar detection dataset as a test set.

10. A small sample sonar reverberation data enhancement and target detection system, characterized in that: The small sample sonar reverberation data enhancement and target detection method described in any one of claims 1 to 9 is applied; the system includes a reverberation data enhancement module, a target echo simulation module and a feature conversion and fusion detection module.

Images