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

Abstract

The application 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. The small-sample sonar reverberation data enhancement and target detection method introduced the spectral perception loss and multi-scale short-time Fourier transform feature constraint through WGAN-GP, the cosine similarity of the time-frequency domain distribution of the reverberation signal generated by the method is greater than or equal to 0.92 with the real data, the mode collapse problem of the traditional generative adversarial network model under small samples is solved, and the data diversity is improved. The time sequence structure and state transition law of the sonar detection data set are coded into a two-dimensional image through the Gram angle field and the Markov transition field, and the CBAM attention mechanism is combined to improve the feature distinction degree of the target and the reverberation. The multi-modal convolution network can improve the detection accuracy under the small sample data scene, is no longer dependent on a large sample data volume, and avoids overfitting when the sample data is small.

Description

TECHNICAL FIELD

[0001] The application 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. BACKGROUND

[0002] In the field of ocean exploration and target identification, an active sonar detects an underwater target by emitting a sound wave and receiving a return wave, so as to identify the underwater target; however, the performance of the active sonar detection scheme is limited by reverberation interference in a complex marine environment and a small-sample data problem, and is not ideal. Among them, the reverberation is a superimposed signal formed by the reflection of the sound wave on the sea surface, the seabed and the scattering body such as plankton, and the energy thereof is often much higher than that of the target return wave, resulting in a low signal-reverberation-ratio (SRR), so that traditional methods such as matched filtering and adaptive filtering are difficult to effectively distinguish the target from the reverberation.

[0003] Meanwhile, due to the complexity of the marine environment (such as the thermocline and the variability of seabed topography), the high cost of sonar data collection (which requires a long time of sea trial relying on a ship platform), and the limitation of experimental conditions, the amount of available sonar reverberation and target return wave data is extremely small, forming a typical small-sample learning scenario, which leads to insufficient model training and weak generalization ability, and seriously restricts the practical application of sonar target detection technology.

[0004] In a small-sample scenario, the lack of data leads to an increase in the inductive bias of traditional deep learning models (such as CNN), making it difficult to learn the potential distribution difference between the reverberation and the target return wave. Although transfer learning and meta-learning techniques are tried to be applied, the time-frequency characteristics of the sonar signal are significantly different from those of images, texts and other modalities, and existing feature conversion methods (such as direct spectrum input) fail to fully exploit the time sequence structure (such as state transition law) and phase information of the signal, limiting the detection ability of the model for weak targets and distant targets.

[0005] Traditional signal processing methods (such as fractional Fourier transform, time reversal mirror and Wigner-Hough transform) suppress reverberation through time-frequency domain filtering, which relies on the prior difference between the target and the reverberation in the time-frequency domain, and has limited adaptability to non-stationary and non-Gaussian reverberation. Although machine learning-based methods (such as support vector machines and BP neural networks) can automatically extract features, the feature representation ability of one-dimensional signals is insufficient, and the model parameters are difficult to be effectively optimized under a small sample, and the detection accuracy significantly decreases as the signal-to-reverberation ratio decreases.

[0006] In summary, the existing sonar target detection technology scheme relies on a large amount of sample data, which is prone to overfitting when the sample data is small, and seriously restricts the practical application of sonar target detection technology. SUMMARY

[0007] The main purpose of the present application 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 technical scheme relies on large sample data, which is prone to overfitting when the sample data is small.

[0008] The technical scheme provided by the present application is:

[0009] 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 comprises a reverberation data enhancement module, a target echo simulation module and a feature conversion and fusion detection module; the method comprises:

[0010] The reverberation data enhancement module is based on the Wasserstein distance to construct a generative adversarial network model combined with gradient penalty;

[0011] The reverberation data enhancement module obtains real reverberation signals and random noise, and learns the time-frequency features of the real reverberation signals through the generator of the adversarial network model combined with gradient penalty, to generate enhanced reverberation data consistent with the real marine environment;

[0012] The target echo simulation module is based on a highlight model, and calculates the time delay, phase shift and amplitude according to the geometric features and material acoustic characteristics of the target to be detected, to generate a simulated echo signal of the target to be detected;

[0013] The target echo simulation module fuses the simulated echo signal and the enhanced reverberation data to construct a sonar detection data set;

[0014] The feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through a Gram angle field and a Markov transition field;

[0015] The feature conversion and fusion detection module inputs the two-dimensional feature image into a multi-modal convolution network for feature fusion and target classification, and outputs a target detection result.

[0016] Preferably, the adversarial network model combined with gradient penalty comprises a generator and a discriminator; the generator comprises:

[0017] Input layer: for receiving a random noise vector;

[0018] Fully connected layer and residual layer: for preliminary transformation and feature extraction of noise;

[0019] Up-sampling layer and convolution layer: for dimension lifting and data processing, cooperating with batch normalization layer and ReLU activation function;

[0020] Output layer: used to output the enhanced reverberation signal to approximate the real reverberation distribution through the transpose convolution layer and the Tanh activation function.

[0021] Preferably, the discriminator comprises:

[0022] Input layer: used to receive the real reverberation signal or the generated reverberation signal;

[0023] Convolution module: composed of a convolution layer, a ReLU activation function with leakage, and a random inactivation layer, used to extract features;

[0024] Output layer: used to output the true-false discrimination probability through the flattening layer, the linear layer, and the Sigmoid activation function.

[0025] Preferably, the target echo simulation module is based on a bright spot model, and calculates the time delay, phase shift, and amplitude according to the geometric features and material acoustic properties of the target to be detected to generate the simulated echo signal of the target to be detected, comprising:

[0026] 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;

[0027] The target echo simulation module calculates the phase shift in combination with the attitude of the target to be detected and the frequency of the incident sound wave;

[0028] The target echo simulation module calculates the amplitude based on the scattering cross section of the target to be detected and the sound wave attenuation characteristics.

[0029] Preferably, the feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through a Gram angle field and a Markov transition field, comprising:

[0030] The feature conversion and fusion detection module normalizes the one-dimensional sonar signal to the interval [-1, 1];

[0031] The feature conversion and fusion detection module performs polar coordinate coding on the normalized one-dimensional sonar signal, converts the value of the one-dimensional sonar signal to an angle or , and converts the timestamp of the one-dimensional sonar signal to a radius , wherein N is the sequence length.

[0032] The feature conversion and fusion detection module constructs a Gram matrix to form a two-dimensional feature image, wherein the elements of the Gram matrix are or .

[0033] Preferably, the feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through a Gram angle field and a Markov transition field, further comprising:

[0034] The feature conversion and fusion detection module quantizes one-dimensional sonar signal into Q quantile units;

[0035] The feature conversion and fusion detection module constructs a QxQ Markov transition matrix;

[0036] The feature conversion and fusion detection module expands the time step τ, calculates the state transition probability under different lag steps, and generates a two-dimensional transition probability matrix containing time information.

[0037] Preferably, the multi-modal convolutional network comprises:

[0038] Gram angle field feature branch: used for sequentially performing convolution, batch normalization, activation through ReLU function, maximum pooling and convolution block attention module processing on the Gram angle field image to extract spatial and channel dimension key features;

[0039] Markov transition field feature branch: used for adopting the same processing flow as the Gram angle field feature branch to process the Markov transition field image;

[0040] Fusion detection layer: used for splicing the Gram angle field feature branch and the Markov transition field feature branch, processing through the flattening layer, the fully connected layer, the activation through the ReLU function and the random inactivation layer, and calculating the target classification probability through the output layer.

[0041] Preferably, the target echo simulation module fuses the simulated echo signal and the enhanced reverberation data to construct a sonar detection dataset, comprising:

[0042] The target echo simulation module superimposes the simulated echo signal and the enhanced reverberation data according to the signal-to-mixing ratio.

[0043] Preferably, it further comprises:

[0044] The feature conversion and fusion detection module trains the multi-modal convolutional network, comprising:

[0045] The feature conversion and fusion detection module calculates the difference between the predicted probability and the true label using the cross-entropy loss function;

[0046] The feature conversion and fusion detection module dynamically adjusts the learning rate using the stochastic gradient descent optimizer;

[0047] The feature conversion and fusion detection module takes 70% of the sonar detection dataset as the training set, 15% of the sonar detection dataset as the validation set, and 15% of the sonar detection dataset as the test set.

[0048] The application further provides a small sample sonar reverberation data enhancement and target detection system, and the small sample sonar reverberation data enhancement and target detection method is applied.

[0049] By the technical solution, the following beneficial effects can be achieved:

[0050] The small sample sonar reverberation data enhancement and target detection method provided by the application introduces spectral perception loss and multi-scale short-time Fourier transform features (STFT) constraints through WGAN-GP, generates a time-frequency domain distribution of reverberation signals, and the cosine similarity with real data is greater than or equal to 0.92, solves the mode collapse problem of a traditional generative adversarial network model under a small sample, improves data diversity, and provides sufficient samples for model training; the time sequence structure and state transition law of the sonar detection data set are encoded into a two-dimensional image through a Gram angle field and a Markov transition field, and the CBAM attention mechanism is combined to improve the feature discrimination degree of the target and the reverberation, and the feature separability under a low signal-to-reverberation ratio condition is improved compared with a traditional spectrum graph method; the method introduces a multi-modal convolution network, the multi-modal convolution network can improve the detection accuracy under a small sample data scene, and the improvement is higher than that of a single feature model, thereby meeting the low signal-to-reverberation ratio detection demand under a real sea environment, so that the method no longer depends on a large sample data quantity and avoids overfitting when the sample data is small; in addition, the model parameter is fine-tuned through a small amount of real data, cross-domain migration from simulation data to real sea data is realized, the detection accuracy of the migrated model in an unknown sea area is reduced, is significantly better than an existing algorithm, and is suitable for the diversity of a complex marine environment. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only show some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from the structures shown in the drawings without creative labor.

[0052] Figure 1 The step flow chart of the first embodiment of the small sample sonar reverberation data enhancement and target detection method provided by the application is shown in the figure.

[0053] Figure 2 The Gram angle field feature schematic diagram of the fifth embodiment of the small sample sonar reverberation data enhancement and target detection method provided by the application is shown in the figure.

[0054] Figure 3 The Markov transition field feature schematic diagram of the sixth embodiment of the small sample sonar reverberation data enhancement and target detection method provided by the application is shown in the figure. DETAILED DESCRIPTION

[0055] It should be understood that the specific embodiments described herein are merely exemplary and not intended to limit the application.

[0056] The application provides a small sample sonar reverberation data enhancement and target detection method and system.

[0057] As shown in the drawings, Figure 1 In a first embodiment of the small sample sonar reverberation data enhancement and target detection method provided by the application, 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 comprises a reverberation data enhancement module, a target echo simulation module, and a feature conversion and fusion detection module; the embodiment comprises the following steps:

[0058] Step S110: The reverberation data enhancement module constructs a generative adversarial network model with gradient penalty based on the Wasserstein distance (i.e., WGAN-GP).

[0059] Specifically, the Wasserstein distance (i.e., Wasserstein-distance) is also known as the Earth-Mover's Distance (EMD), and the Wasserstein distance is a distance measure used to compare the differences between two probability distributions in the theory of measure.

[0060] Meanwhile, the traditional generative adversarial network model (GAN) has the problem of gradient disappearance and mode collapse based on the optimization target of JS divergence. WGAN-GP solves this defect by introducing the Wasserstein distance to replace the JS divergence. The Wasserstein distance can provide continuous gradient signals, and can be optimized stably even when the generated distribution and the real distribution are quite different.

[0061] Step S120: The reverberation data enhancement module acquires real reverberation signals and random noise, and learns the time-frequency features of the real reverberation signals through the generator of the adversarial network model with gradient penalty to generate enhanced reverberation data consistent with the real marine environment.

[0062] Specifically, the WGAN-GP introduces spectral perception loss and multi-scale short-time Fourier transform features (STFT) constraints to generate real reverberation signals with a time-frequency domain distribution and a cosine similarity of ≥0.92 with real data.

[0063] Step S130: The target echo simulation module calculates the time delay, phase shift, and amplitude based on the geometric features and material acoustic characteristics of the target to be detected according to the bright spot model to generate a simulated echo signal of the target to be detected.

[0064] Step S140: the target echo simulation module fuses the simulated echo signal with the enhanced reverberation data to construct a sonar detection data set.

[0065] Step S150: the feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through Gramian-Angular-Field (GAF) and Markov Transition Field (MTF).

[0066] Specifically, in combination with the CBAM attention mechanism (Convolutional Block Attention Module, which is a convolutional block attention module composed of channel attention (Channel Attention) and spatial attention (Spatial Attention)), the feature discrimination of the target and the enhanced reverberation data is improved, and the feature separability under the condition of low signal-to-reverberation ratio is improved compared with the traditional frequency spectrum method.

[0067] Step S160: the feature conversion and fusion detection module inputs the two-dimensional feature image into a multi-modal convolutional network for feature fusion and target classification, and outputs a target detection result.

[0068] The small sample sonar reverberation data enhancement and target detection method provided by the application introduces spectral perception loss and multi-scale short-time Fourier transform features (STFT) constraints through WGAN-GP, generates a time-frequency domain distribution of reverberation signals, and the cosine similarity with real data is ≥0.92, solves the mode collapse problem of the traditional generative adversarial network model under small samples, improves the diversity of data, and provides sufficient samples for model training; the time sequence structure and state transition law of the sonar detection data set are encoded into a two-dimensional image through Gramian-Angular-Field and Markov Transition Field, and in combination with the CBAM attention mechanism, the feature discrimination of the target and the reverberation is improved, and the feature separability under the condition of low signal-to-reverberation ratio is improved compared with the traditional frequency spectrum method; the method introduces a multi-modal convolutional network, which can improve the detection accuracy under the small sample data scene, and is improved compared with a single feature model, thereby meeting the low signal-to-reverberation ratio detection demand in real sea environment, so as to no longer rely on large sample data, and avoid 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 realized, the detection accuracy of the migrated model in unknown sea areas is reduced, which is significantly better than the existing algorithm, and is suitable for the diversity of complex marine environments.

[0069] In a second embodiment of the small sample sonar reverberation data enhancement and target detection method proposed in the application, based on the first embodiment, the gradient-penalty-based adversarial network model includes a generator and a discriminator; the generator includes:

[0070] The input layer is used to receive a random noise vector.

[0071] The fully connected layer and the residual layer are used to preliminarily transform and extract features of the noise.

[0072] The up-sampling layer and the convolution layer are used to increase the dimension and process the data, in cooperation with the BatchNorm layer (Batch Normalization) and the ReLU activation function (Rectified Linear Unit, a commonly used activation function in artificial neural networks).

[0073] The output layer is used to output the enhanced reverberation signal through the transpose convolution layer and the Tanh activation function (Hyperbolic Tangent Activation Function, a transformation of the Sigmoid activation function), so as to approximate the real reverberation distribution.

[0074] Specifically, the generator adopts a multi-layer residual network structure, the input is a random noise vector obeying a uniform distribution, after being mapped to a low-dimensional feature space by the fully connected layer, the residual block (each residual block includes a convolution layer, a BatchNorm layer, and a ReLU activation function) is used to extract nonlinear features. The up-sampling layer is used to gradually increase the feature dimension to the length of the sonar signal, and finally the Tanh activation function is used to output the enhanced reverberation data with a value range of [-1, 1], so as to ensure that the energy distribution of the enhanced reverberation data approximates the real reverberation.

[0075] In a third embodiment of the small sample sonar reverberation data enhancement and target detection method proposed in the application, based on the first embodiment, the discriminator includes:

[0076] The input layer is used to receive a real reverberation signal or a generated reverberation signal.

[0077] The convolution module is composed of a convolution layer, a LeakyReLU activation function (LeakyReLU, Leaky Rectified Linear Unit) with leakage, and a Dropout layer, and is used to extract features.

[0078] The output layer is used to output a true-false discrimination probability through a flattening layer, a linear layer, and a Sigmoid activation function, and a gradient penalty term is introduced to constrain the gradient norm of the discriminator.

[0079] Specifically, the Sigmoid function is a monotonically increasing function, the output range is between 0 and 1, and the more negative the increase, the closer to 0, and the slower the approximation speed; the more positive the increase, the closer to 1, and the slower the approximation speed.

[0080] Specifically, the discriminator is configured to: adopt a deep convolutional network, input being a real reverberation signal or a generated signal, the input being subjected to feature down-sampling through a convolution module, a ReLU activation function with leakage, and a random inactivation layer to extract multi-scale time-frequency features; and finally output a single discrimination probability value (0-1) through a flattening layer and a linear layer, and introduce a gradient penalty term (GP) whose expression formula is:

[0081] ,

[0082] In the formula, is a linear interpolation sample between the real and generated signals, used to constrain the gradient norm of the discriminator close to 1, and ensure the stability of the training; : represents a gradient penalty loss (Gradient Penalty Loss), which is a loss term for constraining the gradient norm of the discriminator, and through the loss term, the stability of the training process of the generative adversarial network can be ensured; : represents expectation (Expectation), which is used to calculate the average expectation of the subsequent random variable values; : represents an output function of the discriminator (Discriminator), and the sample (which can be a real reverberation signal, a generated reverberation signal or an interpolation sample between the two) in the bracket is input, and the output is a true or false discrimination probability value of the sample; : represents a probability distribution of a linear interpolation sample between the real reverberation signal and the generated reverberation signal; is an interpolation sample sampled from the distribution; is an interpolation sample sampled from the probability distribution of the linear interpolation sample between the real reverberation signal and the generated reverberation signal.

[0083] In a fourth embodiment of the small-sample sonar reverberation data enhancement and target detection method provided in the application, based on the first embodiment, step S130 includes the following steps:

[0084] 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.

[0085] Step S420: The target echo simulation module calculates the phase shift in combination with the attitude of the target to be detected and the incident sound wave frequency.

[0086] Step S430: The target echo simulation module calculates the amplitude based on the scattering cross section of the target to be detected and the sound wave attenuation characteristics.

[0087] Specifically, the target scattering feature parameters are parameterized, the form of the transmitted signal is , the target response function is , and the echo signal output by the network response is When the target echo is composed of the scattering signals of N independent bright spots, the simulation echo signal is represented as:

[0088] ,

[0089] ,

[0090] wherein: is the time delay; is the phase shift; is the amplitude; represents a spatial position vector, used for describing the position information of a target, a transmission signal, an echo signal and the like in space; represents an angular frequency, and the relationship between the angular frequency and the frequency f is ω = 2πf; is an imaginary unit.

[0091] In a fifth embodiment of the small-sample sonar reverberation data enhancement and target detection method provided in the application, based on the first embodiment, step S150 comprises the following steps:

[0092] Step S510: The feature conversion and fusion detection module normalizes the one-dimensional sonar signal to the [-1, 1] interval.

[0093] A Gram angle field feature diagram is shown in FIG. 2, and the calculation formula of the above normalization processing is: Figure 2

[0094] .

[0095] wherein, represents the maximum value in the sonar signal sequence ; and represents the minimum value in the sonar signal sequence .

[0096] Step S520: The feature conversion and fusion detection module performs polar coordinate coding on the normalized one-dimensional sonar signal, converts the value of the one-dimensional sonar signal into an angle or , and converts the time stamp of the one-dimensional sonar signal into a radius , wherein N is the sequence length.

[0097] Step S530: The feature conversion and fusion detection module constructs a Gram matrix to form a two-dimensional feature image with a size of , wherein the element of the Gram matrix is

[0098] or .

[0099] Specifically,​ The first normalized one-dimensional sonar signal There are sample values ​​that satisfy: By calculating the trigonometric relationships between the angles corresponding to different sample points: or To construct a two-dimensional feature image.

[0100] In the sixth embodiment of the small-sample sonar reverberation data enhancement and target detection method proposed in this invention, based on the fifth embodiment, step S150 further includes the following steps:

[0101] Step S610: The feature conversion and fusion detection module quantizes the one-dimensional sonar signal into Q quantile units.

[0102] A schematic diagram of the Markov transition field characteristics is attached. Figure 3 As shown.

[0103] Specifically, in this embodiment, Q = 8 quantile units.

[0104] Map each data point to a state: .

[0105] in, This represents 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 is mapped to one of these Q states. qk is used to represent the state number corresponding to the k-th data point, and its value range is {1,2,...,Q}, which means that the state of the quantized sonar signal can only be one of the values ​​in this set.

[0106] Step S620: The feature transformation and fusion detection module constructs a Q×Q Markov transition matrix, wherein the elements w of the Markov transition matrix are... ij Let be the transition probability from state i to state j.

[0107] Specifically, constructing a Q×Q Markov transition matrix includes:

[0108] Statistics at different time steps: Calculate the probability based on the number of state transitions. .

[0109] in, : Represents the time step, used to measure the time interval between two states, with a value range of . ; : Represents the maximum time step, its value is ( The total number of sample points. (Indicates rounding down), used to limit the upper limit of the time step; : represents the probability of state transition from to at time step , and the calculation formula is ; : represents the state of the sonar signal at time ; : represents the mth quantization state, which is one of the states that the sonar signal can be in after quantization, and the value range is ; ; ; : represents the mth quantization state, which is one of the states that the sonar signal can be in after quantization, and the value range is ; ; ; : represents the state of the sonar signal at time .

[0110] Step S630: The feature conversion and fusion detection module expands the time step τ, calculates the state transition probability under different lag steps, and generates a two-dimensional transition probability matrix containing time information.

[0111] Specifically, it comprises:

[0112] The tensor is accumulated along the time dimension to generate a QxQ Markov transition field matrix, which is normalized to [0, 1] and used as a two-dimensional feature image. In the seventh embodiment of the small sample sonar reverberation data enhancement and target detection method provided in the application, based on the first embodiment, the multi-modal convolutional network comprises:

[0113] Gram angle field feature branch (GAF / MTF shared structure): used for sequentially performing convolution, batch normalization (BatchNorm), activation through a ReLU function, maximum pooling and convolution block attention module (CBAM) processing on the Gram angle field image, so as to extract key features in the spatial and channel dimensions.

[0114] Specifically, the Gram angle field feature branch inputs a two-dimensional image, which is processed by 3 convolution blocks: the first layer: convolution layer→batch normalization→leaky ReLU activation function layer→maximum pooling layer; the second layer: convolution block attention module (CBAM): used for combining channel attention and spatial attention to enhance target-related features; the third layer: depth separable convolution→batch normalization→ReLU activation function layer, so as to reduce the computational complexity.

[0115]

[0116] ​Markov transition field feature branch: used for adopting the same processing flow as the gram angle field feature branch for the Markov transition field image.

[0117] Fusion detection layer: used for splicing the gram angle field feature branch and the Markov transition field feature branch, performing processing through a flattening layer, a fully connected layer, an activation through a ReLU function and a random inactivation layer (Dropout), and calculating a target classification probability through an output layer.

[0118] Specifically, the fusion detection layer splices the feature maps output by the GAF / MTF branch, converts the feature maps into one-dimensional vectors through a flattening layer (Flatten), extracts high-order features through two fully connected layers, prevents overfitting through a random inactivation layer, and finally calculates a target classification probability through an output layer.

[0119] In an eighth embodiment of the small-sample sonar reverberation data enhancement and target detection method provided in the application, based on the first embodiment, step S140 includes the following steps:

[0120] 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 weak target to strong target scenes.

[0121] Specifically, the range of the signal-to-reverberation ratio is -10dB to 20dB; the expression formula for the target echo simulation module to superimpose the simulated echo signal and the enhanced reverberation data according to the signal-to-reverberation ratio is:

[0122] ,

[0123] In the formula, is an amplitude scaling factor, and ; the value range of the SNR is set to -10dB to 20dB, covering weak target (SNR<0dB), medium target (0dB≤SNR≤10dB), and strong target (SNR>10dB) scenes; : represents a sonar detection signal obtained by superimposing the simulated echo signal and the enhanced reverberation data according to the signal-to-reverberation ratio, which is an input signal for subsequent feature conversion and target detection; : represents enhanced reverberation data, i.e., reverberation signals consistent with real marine environments generated through a gradient-penalty combined generative adversarial network model; : represents the index of an independent bright spot in a target echo, used to distinguish different bright spot scattering signals, and the value range is is the total number of independent bright spots contained in the target echo; : represents the ​The scattering signal of an independent highlight is a basic unit of the analog echo signal, and a plurality of such signals are superimposed to form an analog echo of a target to be detected.

[0124] In a ninth embodiment of the small-sample sonar reverberation data enhancement and target detection method, based on the first embodiment, the method further includes the following steps:

[0125] Step S910: The feature conversion and fusion detection module trains the multi-modal convolutional network, including the following steps:

[0126] Step S911: The feature conversion and fusion detection module calculates the difference between the predicted probability and the true label using the cross-entropy loss function.

[0127] Step S912: The feature conversion and fusion detection module dynamically adjusts the learning rate using the Stochastic Gradient Descent (SGD) optimizer.

[0128] The initial value of the specific learning rate is 0.0002, and the learning rate is reduced to 0.00005 after 30 epochs.

[0129] Step S913: The feature conversion and fusion detection module uses 70% of the sonar detection dataset as the training set, 15% of the sonar detection dataset as the validation set, and 15% of the sonar detection dataset as the test set to prevent model overfitting.

[0130] The present application also provides a small-sample sonar reverberation data enhancement and target detection system, which applies the 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.

[0131] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0132] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative and not limiting, and those skilled in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the scope of protection of the claims, which are all within the scope of protection of the present application.

Claims

1. A method for small sample sonar reverberation data enhancement and target detection, characterized in that, The application is applied to a small sample sonar reverberation data enhancement and target detection system; the system comprises a reverberation data enhancement module, a target echo simulation module and a feature conversion and fusion detection module; the method comprises: The reverberation data enhancement module is based on the Wasserstein distance to construct a generative adversarial network model combined with gradient penalty; The reverberation data enhancement module obtains a real reverberation signal and a random noise, and learns the time-frequency features of the real reverberation signal through the generator of the adversarial network model combined with gradient penalty to generate enhanced reverberation data consistent with the real marine environment; The target echo simulation module is based on a highlight model and calculates the time delay, phase shift and amplitude according to the geometric features and material acoustic characteristics 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 and 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 a Gram angle field and a Markov transition field; The feature conversion and fusion detection module inputs the two-dimensional feature image into a multi-modal convolution network for feature fusion and target classification, and outputs a target detection result.

2. The method of claim 1, wherein, The adversarial network model combined with gradient penalty comprises a generator and a discriminator; the generator comprises: An input layer for receiving a random noise vector; A fully connected layer and a residual layer for preliminary transformation and feature extraction of the noise; An up-sampling layer and a convolution layer for dimension lifting and data processing, cooperating with a batch normalization layer and a ReLU activation function; An output layer for outputting an enhanced reverberation signal through a transposed convolution layer and a Tanh activation function to approximate the real reverberation distribution.

3. The method of claim 2, wherein, The discriminator comprises: An input layer for receiving a real reverberation signal or a generated reverberation signal; A convolution module composed of a convolution layer, a ReLU activation function with leakage and a random inactivation layer for feature extraction; An output layer for outputting a true-false discrimination probability through a flattening layer, a linear layer and a Sigmoid activation function.

4. The method of claim 1, wherein, The target echo simulation module is based on a highlight model and calculates the time delay, phase shift and amplitude according to the geometric features and material acoustic characteristics of the target to be detected to generate a simulated echo signal of the target to be detected, comprising: 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 the sonar; The target echo simulation module calculates the phase shift in combination with the attitude of the target to be detected and the incident sound wave frequency; The target echo simulation module calculates the amplitude based on the scattering cross section of the target to be detected and the sound wave attenuation characteristics.

5. The method of claim 1, wherein, The feature conversion and fusion detection module converts the sonar detection data set into a two-dimensional feature image through a Gram angle field and a Markov transition field, comprising: The feature conversion and fusion detection module normalizes the one-dimensional sonar signal to the [-1, 1] interval; The feature conversion and fusion detection module performs polar coordinate coding on the normalized one-dimensional sonar signal, converts the value of the one-dimensional sonar signal into an angle or , and converts the timestamp of the one-dimensional sonar signal into a radius , wherein N is the sequence length. The feature conversion and fusion detection module constructs a Gram matrix to form a two-dimensional feature image, wherein elements of the Gram matrix are or .

6. The method of claim 5, wherein, The feature conversion and fusion detection module converts the one-dimensional sonar signal into Q quantile units; ​ The feature conversion and fusion detection module constructs a QxQ Markov transition matrix; The feature conversion and fusion detection module expands the time step τ, calculates the state transition probability under different lag steps, and generates a two-dimensional transition probability matrix containing time information.

7. The method of claim 1, wherein, The multi-modal convolutional network comprises: Gram angle field feature branch: used for sequentially performing convolution, batch normalization, activation through a ReLU function, maximum pooling, and convolution block attention module processing on the Gram angle field image to extract spatial and channel dimension key features; Markov transition field feature branch: used for adopting the same processing procedure as the Gram angle field feature branch to process the Markov transition field image; Fusion detection layer: used for splicing the Gram angle field feature branch and the Markov transition field feature branch, performing processing through a flattening layer, a fully connected layer, an activation through a ReLU function, and a random dropout layer, and calculating a target classification probability through an output layer.

8. The method of claim 1, wherein, The target echo simulation module fuses the simulated echo signal and the enhanced reverberation data to construct a sonar detection dataset, comprising: The target echo simulation module superimposes the simulated echo signal and the enhanced reverberation data according to a signal-to-mixing ratio.

9. The method of claim 1, wherein, Further comprising: The feature conversion and fusion detection module trains the multi-modal convolutional network, comprising: The feature conversion and fusion detection module calculates the difference between the predicted probability and the real label by using a cross-entropy loss function; The feature conversion and fusion detection module dynamically adjusts the learning rate by using a stochastic gradient descent optimizer; The feature conversion and fusion detection module takes 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 system comprises a reverberation data enhancement module, a target echo simulation module, and a feature conversion and fusion detection module. The system comprises a reverberation data enhancement module, a target echo simulation module, and a feature conversion and fusion detection module.

Images