Underwater Acoustic Preamble Signal Detection Method and Device Based on Preamble Emulator

Abstract

The present invention discloses an underwater acoustic preamble signal detection method and device based on a preamble emulator. The preamble emulator is used to simulate preamble data and interference data, and the data is transformed into a time-frequency image through time-frequency transformation, and the time-frequency image is preprocessed to obtain a training set. The MobileNet network model is pre-trained using the training set. Using a small amount of underwater acoustic preamble data obtained from real experiments, transfer learning is performed on the pre-trained MobileNet network model, and some parameters of the model are optimized and trained to make the model applicable to the corresponding underwater acoustic preamble detection task. The simulation underwater acoustic data set established by the present invention through the preamble emulator solves the problems of less real and available underwater acoustic data sets, non-uniform sample specifications, and few types. The problem of high sensitivity of the STFT time-frequency image to noise and high resolution is solved through image preprocessing. By using the data of the preamble emulator as the training set of the pre-trained network model, the problems that general networks do not have reasonable initial parameters and are prone to overfitting are solved.

Description

Technical Field

[0001] The present application relates to the technical field of underwater acoustic communication systems, and particularly to an underwater acoustic preamble signal detection method and device based on a preamble emulator. Background Art

[0002] Preamble detection is a key technology in underwater acoustic communication systems, which is used to wake up the system at the receiving end and then start the receiving work. Accurate preamble detection is an important guarantee for the success of underwater acoustic communication. The commonly used preamble signals in underwater acoustic communication systems are mainly frequency modulation signals. The commonly used preamble detection methods are mainly detection methods based on matched filtering. These methods mainly use the correlation between the received signal and the transmitted signal for detection, and are easily affected by the signal-to-noise ratio, the double-spread underwater acoustic channel, and underwater interference, resulting in a decline in detection performance. In fact, the frequency modulation signal used as the preamble signal has typical time-frequency characteristics, and its time-frequency image is significantly different from the time-frequency images of various underwater interferences. Therefore, the deep neural network suitable for image recognition has great potential in improving the preamble detection performance. However, the existing deep neural networks require a large amount of training data, the training takes a long time, and the portability is poor, which cannot meet the actual needs of underwater acoustic communication systems. Summary of the Invention

[0003] In view of the above problems, the present invention provides an underwater acoustic preamble signal detection method and device based on a preamble emulator, which generates simulated underwater acoustic preamble data through the preamble emulator for pre-training the MobileNet neural network model, and helps the pre-trained model quickly update the underwater acoustic preamble detection task through transfer learning.

[0004] In the first aspect of the present invention, an underwater acoustic preamble signal detection method based on a preamble emulator includes the following steps:

[0005] Use the preamble emulator to simulate preamble data and interference data, generate time-frequency images of the preamble data and interference data through time-frequency transformation, and preprocess the generated time-frequency images to finally obtain a training set;

[0006] Use the training set to pre-train a deep neural network model, where the deep neural network model is a MobileNet network model;

[0007] Use a small amount of underwater acoustic preamble data obtained from real experiments to perform transfer learning on the pre-trained MobileNet network model, and optimize and train some parameters of the model to make the model suitable for the corresponding underwater acoustic preamble detection task.

[0008] Further, obtaining the training set by using the preamble emulator specifically includes:

[0009] Build the preamble signal and underwater acoustic channel model, and the underwater interference noise model respectively, set the underwater environment parameters in each model, and generate time series signal data by using the preamble signal and underwater acoustic channel model and the underwater interference noise model;

[0010] Convert the generated time series signal data into a time-frequency image by using the short-time Fourier transform, perform normalization preprocessing on the time-frequency image, and obtain a training data set by classifying and labeling the preprocessed video image.

[0011] Furthermore, the establishment and parameter setting of the preamble signal and underwater acoustic channel model specifically include:

[0012] Let the duration of the hyperbolic frequency modulation waveform be T, and define

[0013]

[0014] where f1 represents the start frequency of the hyperbolic frequency modulation waveform, and f2 represents the end frequency of the hyperbolic frequency modulation waveform;

[0015] The generation formula of the preamble signal is:

[0016]

[0017] where A(t) represents a rectangular envelope, t represents time, and the preamble signal parameters are set as follows: the center frequency of the signal is [12, 13] kHz, the bandwidth is [5, 6] kHz, the duration is 100 ms or 15 ms, and the signal-to-noise ratio is [-13, 12] dB;

[0018] In the baseband, the underwater acoustic channel h(τ) is expressed as:

[0019]

[0020] where N pa represents the number of paths associated with the p-th path, A p represents the amplitude associated with the p-th path, τ p represents the delay associated with the p-th path, f c represents the center frequency, δ(t - τ p ) represents the impulse function, and τ represents the time delay;

[0021] Then the received baseband signal x(t) is

[0022]

[0023] where ω(t) is the environmental noise, η(t) is the external interference, and the underwater acoustic channel parameters are set as: the number of paths N pa ~U(15, 25), the arrival interval time τ of consecutive pathsp -τ p-1 It follows an exponential distribution with an average value of 1 ms.

[0024] Furthermore, the underwater interference noise model includes a narrowband interference model, an impulse interference model, and a short-time narrowband interference model. The specific expression of the narrowband interference model is:

[0025]

[0026] where f nb [i] represents the frequency of the i-th tone, A nb [i] represents the amplitude of the i-th tone, φ nb [i] represents the phase shift of the i-th tone. The parameters of the narrowband interference are set as follows: the central frequency f c ~U(11.5, 13.5) kHz, the bandwidth B~U(1, 3000) kHz, the duration is 200 ms, and the interference-to-noise ratio INR~U(-5, 10) dB. N nb represents multiple tones;

[0027] The specific expression of the impulse interference model is:

[0028]

[0029] where CN() represents a complex Gaussian distribution, p represents the mixing ratio of the Gaussian mixture model, represents the energy of the Gaussian noise, represents the energy of the impulse noise. The parameters of the impulse interference are set as follows: p~U(0.001, 0.01), and the duration is 200 ms;

[0030] The specific expression of the short-time narrowband interference model is:

[0031]

[0032] where c l represents the coefficient of the Fourier series, B1 represents the bandwidth of the short-time narrowband interference, and T1 represents the duration of the short-time narrowband interference. The parameters of the short-time narrowband interference are set as follows: the central frequency f c ~U(11.5, 13.5) kHz, the bandwidth B~U(1, 3) kHz, the duration T PBPD ~U(5, 60) ms, and the interference-to-noise ratio INR~U(2, 15) dB.

[0033] Furthermore, the generated time-series signal data is converted into a time-frequency image using the short-time Fourier transform. The two-dimensional data expression v(m, l) of the time-frequency image is:

[0034] v(m, l) = |χ(m, l)| 2

[0035]

[0036] where is a complex discrete-time signal vector, and N x represents the block length. The signal vector is obtained by equidistant sampling according to a given fixed sampling rate fs. w = [w[0], w[1],..., w[N - 1]] T ∈ C N×1 is a sampling window function. N represents the length parameter, H represents the hop size parameter, m ∈ [0:M], represents the maximum frame index, l ∈ [-L + 1:L], and L = N / 2 represents the frequency index corresponding to the Nyquist frequency.

[0037] Furthermore, perform normalization preprocessing on the time-frequency image, and obtain a training data set by classifying and labeling the normalized preprocessed video image, specifically including:

[0038] Convert the time-frequency image to grayscale. The grayscale formula for an RGB image is as follows:

[0039]

[0040] where R, G, and B respectively represent the color brightness of the red, green, and blue image channels;

[0041] Convert the grayscale processed image to double precision and perform gray-scale adjustment through linear mapping;

[0042] Compress the resolution of the gray-scale adjusted time-frequency image, automatically generate samples and labels, and save the samples to the file corresponding to the label in the simulator data set.

[0043] Furthermore, replace the input layer and the fully connected layer in the MobileNet network model, specifically including: Set the input layer in the model to the size of the time-frequency image generated by the leading simulator. The classifier of the fully connected layer uses the Sigmoid algorithm for binary classification tasks and the Softmax algorithm for multi-classification tasks.

[0044] Furthermore, use the underwater acoustic leading data obtained from a small number of real experiments to perform transfer learning on the pre-trained MobileNet network model, and optimize and train some parameters of the model to make the model applicable to the corresponding underwater acoustic leading detection task, specifically including:

[0045] Use x1′, x2′,..., x m ′ ∈ X′ to represent the real leading data samples, y1′, y2′,..., ym '∈Y' represents the label corresponding to the true leading data sample, and the pre-trained MobileNet network model is expressed as:

[0046] y i = f(x i ; μ)

[0047] where μ represents the MobileNet model parameters;

[0048] Taking a certain layer in the MobileNet network model as the demarcation layer, the parameter μ is divided into μ l and μ h , and the MobileNet model is expressed as

[0049] y i = f(x i ; μ l , μ h )

[0050] where μ l and μ h respectively represent the parameter set above the demarcation layer and the parameter set below the demarcation layer. From the characteristics of the MobileNet model, it can be known that μ l extracts shallow basic features, and μ h extracts deep abstract features.

[0051] The transfer learning formula based on the MobileNet model is as follows:

[0052] y′ i = f(x i ′; μ h ′)

[0053] By freezing μ l , and only retraining μ h based on a small amount of real data sets, the MobileNet model after transfer learning can be used for the leading detection task of real underwater acoustic data.

[0054] In the second aspect of the present invention, an underwater acoustic leading signal detection device based on a leading emulator is provided, including:

[0055] A training set acquisition module, configured to use a leading emulator to simulate leading data and interference data, generate time-frequency images from the leading data and interference data through time-frequency transformation, and preprocess the generated time-frequency images to finally obtain a training set;

[0056] A deep neural network model pre-training module, configured to pre-train a deep neural network model using the training set, where the deep neural network model is a MobileNet network model;

[0057] A deep neural network model transfer learning module is used to perform transfer learning on a pre-trained MobileNet network model by using underwater acoustic preamble data obtained from a small number of real experiments, optimize and train some parameters of the model, and make the model applicable to the corresponding underwater acoustic preamble detection task.

[0058] In the third aspect of the present invention, a computer-readable storage medium is provided, on which instructions are stored. When the instructions are executed by a processor, the processor executes the above-mentioned underwater acoustic preamble signal detection method based on a preamble emulator.

[0059] An underwater acoustic preamble signal detection method and device provided by the present invention mainly face the problems of insufficient general training data sets and diverse detection tasks in the process of underwater acoustic communication preamble detection. By establishing a simulated underwater acoustic data set through a preamble emulator, the problems of less real available underwater acoustic data sets, inconsistent sample specifications, and limited types are solved; through the image preprocessing step of the preamble emulator, the problems that STFT time-frequency images are sensitive to noise and have high resolution are solved; by generating simulated underwater acoustic preamble data through the preamble emulator to pre-train the MobileNet neural network model, the problems that general networks do not have reasonable initial parameter settings and are prone to overfitting are solved; and transfer learning is used to help the pre-trained model quickly update the underwater acoustic preamble detection task; the learning method of the neural network in preamble detection is improved based on transfer learning, and the portability of the neural network in preamble detection is improved. The specific advantages of this detection method are as follows:

[0060] 1. Less real underwater acoustic data is required for training: Utilizing the high correlation between the preamble emulator data and real underwater acoustic data, only a small number of key layers in the model need to be trained with real underwater acoustic data, greatly reducing the demand for real underwater acoustic data;

[0061] 2. Good model portability: The simulated underwater acoustic data generated by the preamble emulator is rich in quantity and diverse in type, with good versatility and robustness, and is applicable to most underwater acoustic communication scenarios;

[0062] 3. Fast training convergence and good detection effect: The pre-trained network model has more reasonable initial parameter settings, which can effectively narrow the training range and the number of training parameters, and accelerate the training efficiency;

[0063] 4. Different detection tasks can be switched: When a new preamble detection task appears, through transfer learning, the parameters of the pre-trained model can be selectively optimized and adjusted to help the pre-trained model quickly adapt to the new preamble detection task;

[0064] Through data verification, the underwater acoustic preamble signal detection method and device of the present invention have lower signal detection errors, fewer data restrictions, and lower complexity compared with existing methods, and are more suitable for complex and changeable underwater acoustic environments and preamble detection tasks. Brief Description of the Drawings

[0065] Figure 1 is a flowchart of a method for detecting underwater acoustic preamble signals based on a preamble emulator in an embodiment of the present invention;

[0066] Figure 2 is a flowchart of a method for obtaining a training set of a preamble emulator diagram in an embodiment of the present invention;

[0067] Figure 3 is a schematic structural diagram of a device for detecting underwater acoustic preamble signals based on a preamble emulator in an embodiment of the present invention. Detailed Description of the Embodiment

[0068] The present invention will be further described in detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention, rather than limiting the present invention. Additionally, it should be noted that for the sake of description, only parts related to the present invention rather than all structures are shown in the drawings.

[0069] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts depict the steps as sequential processes, many of the steps can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operations are completed, but it can also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, and so on.

[0070] The embodiments of the present invention provide the following embodiments for a method and device for detecting underwater acoustic preamble signals based on a preamble emulator:

[0071] Based on Embodiment 1 of the present invention

[0072] As Figure 1 shown, it is a flowchart of a method for detecting underwater acoustic preamble signals based on a preamble emulator in Embodiment 1 of the present invention, and the specific steps are as follows:

[0073] Step 1: Use a preamble emulator to simulate preamble data and interference data, generate a time-frequency image from the preamble data and interference data through time-frequency transformation, and perform preprocessing on the generated time-frequency image to finally obtain a training set;

[0074] The specific implementation process is as Figure 2As shown, first, a preamble signal, an underwater acoustic channel, and various underwater interference noise models are established, appropriate underwater environment parameters are set, and then data is generated. After that, the generated time-series signal data is converted into a time-frequency image using the short-time Fourier transform, and these time-frequency images are preprocessed by normalization. After classification and labeling, a training data set is obtained. The above process is encapsulated into a preamble simulator, and the corresponding simulation training set can be obtained by simply adjusting the simulator parameters later.

[0075] Specifically, it includes: establishing a preamble signal model, an underwater acoustic channel model, and an underwater interference noise model respectively, setting underwater environment parameters in each model, and generating time-series signal data using the preamble signal model, the underwater acoustic channel model, and the underwater interference noise model; converting the generated time-series signal data into a time-frequency image using the short-time Fourier transform, performing normalization preprocessing on the time-frequency image, and obtaining a training data set by classifying and labeling the preprocessed video image.

[0076] Among them, the establishment and parameter setting of the preamble signal and underwater acoustic channel model are specifically implemented as follows:

[0077] Common preamble signals are hyperbolic frequency modulation (HFM) or linear frequency modulation (LFM) signals that are insensitive to Doppler. In the embodiment, taking HFM as an example, the corresponding preamble signal model is given. Let the duration of the hyperbolic frequency modulation waveform be T, and define

[0078]

[0079] where f1 represents the start frequency of the hyperbolic frequency modulation waveform, and f2 represents the end frequency of the hyperbolic frequency modulation waveform;

[0080] Let the transmitted signal s(t) be

[0081]

[0082] where A(t) represents a rectangular envelope, t represents time, and the specific preamble signal parameter settings in the preamble simulator are as follows: Generate a preamble signal according to the transmitted signal s(t) formula, the center frequency f of the signal c is [12, 13] kHz, the bandwidth B is [5, 6] kHz, where the start frequency f1 and the cut-off frequency f2 in the b formula are obtained from the center frequency f c and the bandwidth B, the duration T is 100 ms or 15 ms, and the signal-to-noise ratio is [-13, 12] dB. The parameters are uniformly distributed within the above intervals;

[0083] The underwater acoustic channel that varies with time in the passband is expressed as:

[0084]

[0085] Since HFM and LFM are insensitive to Doppler compression or expansion, and since the Doppler effect on the waveform can be approximated by a displacement in delay, a simple channel model is adopted. In the baseband, the channel h(τ) is expressed as

[0086]

[0087] where N pa represents the number of paths associated with the p-th path, A p represents the amplitude associated with the p-th path, τ p represents the delay associated with the p-th path, f c represents the center frequency, δ(t - τ p ) represents the impulse function, and τ represents the time delay;

[0088] Then the received baseband signal x(t) is

[0089]

[0090] where ω(t) is the ambient noise and η(t) is the external interference. The underwater acoustic channel parameters are set as follows: the number of paths N pa ~U(15, 25), the arrival interval time τ p -τ p-1 of consecutive paths is exponentially distributed with an average value of 1 ms. The average channel delay spread is in the range of [15, 25] ms according to the number of paths. The amplitudes are independent and follow a mixed Gaussian distribution, with the average power decreasing exponentially with delay, where the power difference from 0 to 20 ms is 20 dB. The underwater ambient noise ω(t) is a background noise that can be composed of numerous sound sources, approximated as Gaussian noise, not white noise, and is easily affected by changes in time, position, or depth.

[0091] Furthermore, the underwater interference noise model includes a narrowband interference model, an impulse interference model, and a short-time narrowband interference model. Among them, NB (narrowband interference) generally comes from marine mammals or narrowband sonar systems, usually with a long duration but limited frequency. When considering the underwater reception situation, usually the narrowband interference is composed of multiple tones N nb and the specific expression of the model is:

[0092]

[0093] where f nb [i] represents the frequency of the i-th tone, A nb [i] represents the amplitude of the i-th tone, φ nb [i] represents the phase shift of the i-th tone, and the parameters of the narrowband interference are set as: the center frequency fc ~U(11.5, 13.5) kHz, the bandwidth B~U(1, 3000) Hz, the duration is 200 ms, the interference-to-noise ratio INR~U(-5, 10) dB, N nb represents multiple tones. Marine mammals or narrowband sonar systems may emit sounds with multiple tones in a short period of time. Therefore, a narrowband interference may contain multiple tones.

[0094] IMP (impulse interference) is completely different from Gaussian noise. It has a high amplitude but a very short duration. Its empirical amplitude distribution can be accurately represented by a symmetric alpha-stable (SαS) distribution. The stable distribution is described by four parameters: the characteristic exponent α, the scale parameter γ, the location parameter λ, and the symmetry parameter β. In a symmetric stable distribution, λ = 0 and β = 0.

[0095]

[0096] Among them, α is the characteristic exponent that controls the heaviness of the tail. The γ scale parameter (also known as dispersion) determines the spread of the distribution in a way similar to the variance in the Gaussian distribution. When α = 2, γ is half of the variance. For all other values of α, the variance of the stable distribution is infinite. The probability density function of the SαS distribution can be calculated through the characteristic function:

[0097]

[0098] IMP is generated at the baseband and is independent of each other. Under the Gaussian mixture model, the specific expression of the impulse interference model is:

[0099]

[0100] where CN() represents the complex Gaussian distribution, p represents the mixing ratio of the Gaussian mixture model, represents the energy of Gaussian noise, represents the energy of impulse noise. The parameters of impulse interference are set as: p~U(0.001, 0.01), the duration is 200 ms;

[0101] PBPD (short-time narrowband interference) cannot be treated as either impulse noise or narrowband interference. PBPD uses the Fourier series to represent the baseband waveform, and the specific expression is:

[0102]

[0103] where, B1 represents the bandwidth of short-time narrowband interference, T1 represents the duration of short-time narrowband interference, c l represents the coefficient of the Fourier series. Without loss of generality, it can be assumed that and is even. The parameters of the short-time narrowband interference are set as follows: the center frequency f c ~U(11.5, 13.5) kHz, the bandwidth B~U(1, 3) kHz, and the duration T PBPD ~U(5, 60) ms, and the interference-to-noise ratio INR~U(2, 15) dB.

[0104] It should be noted that in the specific implementation process, there is also a similar preamble interference model establishment and parameter setting. The communication equipment of SCI (similar preamble signals) can also use HFM or LFM signals, but the parameters are different. These are represented as s′(t). The channel path parameters are also different, written as (A′ p , τ′ p ). The similar interference has the following forms

[0105]

[0106] The correlation between similar preamble signals is relatively high, and it is the most difficult to detect in underwater acoustic interference. In the preamble simulator, various similar preamble interference data can be covered according to the specific detection environment, mainly involving different modulation types. The relevant parameters are: up and down sweep type, center frequency, bandwidth, and duration.

[0107] Furthermore, the generated time-series signal data is converted into a time-frequency image using the short-time Fourier transform. In the specific implementation process, the data generated by the underwater acoustic preamble simulator is a time series, and its time-frequency characteristics need to be further enhanced to extract the preamble signal characteristics for detection. The short-time Fourier transform (STFT) method is used to process the time-series data. Considering the discrete case of STFT, in the actual system, the continuously input data during the waiting for wake-up is processed in blocks. Define N x to represent the block length. At time n, the data received within the block is x[n] = [x[n - N x + 1],..., x[n]] T . For the convenience of formula expression in the calculation process, x[n] is redefined as y, which is a complex discrete-time signal vector. That is, y is the vector form of a certain determined block of the input discrete complex signal, and this signal vector is obtained by equidistant sampling according to the given fixed sampling rate fs.

[0108] w = [w][0], w[1],..., w[N - 1]] T ∈C N×1 is a sampling window function. In the specific implementation process, the window power is normalized to ||w|| 2 = 1. The length parameter N determines the duration of the considered part, which is equivalent to N / f sSeconds. A hop size parameter H is introduced, which is specified in the sample and determines the step size at which the window moves in the signal. Regarding these parameters, the discrete STFT of the signal is given by:

[0109]

[0110] The two-dimensional data expression v(m, l) of the time-frequency image is:

[0111] v(m, l) = |χ(m, l)| 2

[0112] where, is a complex discrete-time signal vector, N x represents the block length, the signal vector is obtained by equidistant sampling according to a given fixed sampling rate fs, w = [w[0], w[1],..., w[N - 1]] T ∈ C N×1 is the sampling window function, N represents the length parameter, H represents the hop size parameter, m ∈ [0:M], represents the maximum frame index such that the time range of the window is completely contained within the time range of the signal, l ∈ [-L + 1:L], L = N / 2 (assuming N is even) represents the frequency index corresponding to the Nyquist frequency. The complex number χ(m, l) represents the l-th Fourier coefficient of the m-th time frame. The calculation of each such spectral vector is equivalent to a DFT of size N.

[0113] Furthermore, the time-frequency image is preprocessed by normalization, and the preprocessed video image is classified and labeled to obtain a training dataset. In the specific implementation process, since the time-frequency image v(m, l) generated by the short-time Fourier transform is very sensitive to noise, it is not conducive to subsequent image classification, and the resolution of the time-frequency map is too high, which increases the computational burden on the detection neural network. Therefore, a method for preprocessing the time-frequency image is proposed. For the time-frequency Figure 2 Each data value v at each point in the two-dimensional data v(m, l) has a mapping relationship to the color map, and the color differences of v with similar values are not obvious, and the color map usually uses the RGB color model as the standard for analysis. First, the time-frequency image is grayscale processed, and the grayscale formula for the RGB image is as follows:

[0114]

[0115] where R, G, and B represent the color brightness of the red, green, and blue image channels respectively, represented by integers from 0 to 255;

[0116] Next, convert the grayscale image to double precision. In the above formula, Gary is of 8-bit unsigned integer type, with a range of [0, 255]. The double-precision conversion changes the data type of the image to double type, and through normalization, the output pixel values are rescaled to the range [0, 1]. Finally, perform grayscale adjustment through linear mapping. For an image with pixel values in the range [0, 1], set [low_in; high_in] as the grayscale range to be transformed in the original image, and [low_out; high_out] as the transformed grayscale range. The grayscale adjustment formula is as follows:

[0117]

[0118] For pixel values less than low_in, perform linear mapping with the value of low_out. For pixel values greater than high_in, perform linear mapping with the value of high_out. Through the above image processing, some redundant information in the image can be effectively removed, the ability to distinguish information in the image can be enhanced, and information redundancy in the image can be reduced;

[0119] Compress the resolution of the time-frequency image after grayscale adjustment, automatically generate samples and labels, and save the samples to the file corresponding to the label in the simulator dataset. In the specific implementation process, simplify the definition of the time-frequency image v(m, l) as x i , x1, x2,.., x n ∈X are the samples in the training set generated by the leading simulator, and y1, y2,.., y n ∈Y are the corresponding labels (signal or interference type names), and save the samples to the file corresponding to the label in the simulator dataset.

[0120] Through the above leading simulator, the following functions are realized:

[0121] A. Established leading signal, channel, and underwater interference models with unified specifications and diverse types, set appropriate underwater environment parameters, and generated a large amount of simulated underwater acoustic data. Solved the problems of a small amount of real and available underwater acoustic datasets, inconsistent sample specifications, and limited types.

[0122] B. Generated more similar leading interference data, strengthened the learning of the neural network model for similar leading interference, and greatly enhanced the anti-interference ability of underwater leading detection.

[0123] C. Performed preprocessing operations on the time-frequency image, grayscaled and adjusted the grayscale of the RGB image, compressed the image resolution, reduced noise effects and information redundancy. Solved the problems that the STFT time-frequency image is sensitive to noise and has high resolution, facilitating subsequent feature extraction and image classification by the neural network.

[0124] Based on Embodiment 2 of the present invention

[0125] In this embodiment, based on the training set obtained in Step 1 of Embodiment 1, Step 2 is continued:

[0126] Step 2: Use the training set to pre-train a deep neural network model, where the deep neural network model is a MobileNet network model;

[0127] In the specific implementation process, the MobileNet model, as a transitional model before transfer learning, needs to have a certain degree of generality and low complexity. The lightweight nature of the MobileNet network lies in depthwise separable convolutions. The overall framework of the MobileNet model is shown in Table 1. The entire network is composed of multiple depthwise separable modules, and each depthwise separable module is composed of a depthwise convolution and a pointwise convolution.

[0128] Table 1 MobileNet Network Model Architecture Table

[0129]

[0130] In Table 1, the convolutional layer containing "dw" is a depthwise convolution, and the convolutional layer without "dw" is a pointwise convolution. Compared with a general convolutional layer, the depthwise convolution splits the convolutional kernel into a single-channel form and performs convolution operations on each channel without changing the depth of the input feature image, thus obtaining an output feature map with the same number of channels as the input feature map. The pointwise convolution is a 1×1 convolution. Its main function is to increase or decrease the dimension of the feature map.

[0131] Furthermore, since the MobileNet network model still needs to match the simulated data generated by the leading emulator, the input layer and the fully connected layer in the MobileNet network model are replaced. Specifically, the input layer in the model is set to the size of the time-frequency image generated by the leading emulator, and the classifier of the fully connected layer uses the Sigmoid algorithm for binary classification tasks and the Softmax algorithm for multi-classification tasks.

[0132] Among them, Sigmoid is a binary classification algorithm. In logistic regression, the training set consists of m labeled samples {(x (1) , y (1) ),..., (x (m) , y (m) )}. Since the main goal is a binary classification problem, the class label y (i) ∈ {0, 1}. The Sigmoid function that uses the parameter z as the input:

[0133]

[0134] The Softmax regression model is an extension of the Logistic regression model for multi-classification problems, where the class label y (i) can take more than two values.

[0135]

[0136] where z i is the output value of the i-th node, C is the number of output nodes, that is, the number of classification categories. The Softmax function can convert the output values of multiple classifications into a probability distribution within the range of [0,1] and make the sum of probabilities equal to 1.

[0137] The specific implementation process of the pre-training method of the MobileNet model is as follows: The MobileNet model network can be expressed as

[0138] y i = g(Φ(x i ; θ); w

[0139] where the feature extractor is Φ: X → Z, the parameter is θ, the classifier is g: Z → Y, the parameter is w, x and f(x) are the samples and corresponding labels, and the goal is to learn Φ and g simultaneously so that the corresponding x i and y i satisfy the formula y i . During the learning process, the mini-batch gradient descent method is used to train the overall model. The data set X is divided into multiple batches, and the sample size of each batch is much smaller than that of the data set. A group of data in a batch jointly determines the direction of the gradient, and the parameters θ and w are updated according to the batch. The optimizer settings for training the network model in the embodiment are as follows: the learning rate is 0.08, and the learning rate decay value is 1e-6. To accelerate convergence, Nesterov momentum update is used in the embodiment.

[0140] As a lightweight network, the MobileNet model has much less computational complexity than general neural networks, meeting the actual requirements of preamble detection in underwater acoustic communication. After pre-training with the data of the preamble simulator, the MobileNet model has already possessed the ability to extract the features of underwater acoustic data. In the case of no real data (only using the simulation data set), it already has good preamble detection ability. When used as the starting point for subsequent transfer learning, this model has more reasonable initial parameters and can accelerate the convergence speed of subsequent learning.

[0141] Based on Embodiment 3 of the present invention

[0142] This embodiment continues to implement Step 3 after Step 1 and Step 2 in Embodiments 1-2:

[0143] Step 3: Use the underwater acoustic preamble data obtained from a small number of real experiments to perform transfer learning on the pre-trained MobileNet network model, and optimize and train some parameters of the model to make the model applicable to the corresponding underwater acoustic preamble detection task.

[0144] Specifically, it includes:

[0145] Use x1′, x2′,..., x m ′∈X′ to represent the real preamble data samples, and y1′, y2′,..., y m ′∈Y′ to represent the labels corresponding to the real preamble data samples. The pre-trained MobileNet network model in Example 2 can be expressed as: i y

[0146] y i = f(x i ; μ)

[0147] where μ represents the MobileNet model parameters. The MobileNet model is f: X → Y, and at this time, μ already has the ability to extract the characteristics of underwater acoustic data.

[0148] Take a certain layer in the MobileNet network model as the demarcation layer and divide the parameter μ into μ l and μ h , and the MobileNet model is expressed as

[0149] y i = f(x i ; μ l , μ h )

[0150] where μ l and μ h represent the parameter sets above and below the demarcation layer respectively. According to the characteristics of the MobileNet model, μ l extracts shallow basic features, and μ h extracts deep abstract features.

[0151] The transfer learning formula based on the MobileNet model is as follows:

[0152] y′ i = f(x i ′; μ h ′)

[0153] By freezing μ l , only re-train μ h based on a small number of real data sets. After transfer learning, the MobileNet model can be used for the preamble detection task of real underwater acoustic data.

[0154] In the specific implementation process, as shown in Table 1, it is recommended that the layer for fine-tuning settings be near the AvePool layer, and it can be shifted up and down according to the correlation between the real underwater acoustic data and the emulator data. When the correlation is poor, it is recommended to shift the fine-tuning layer upward, that is, increase the number of trainable parameters μ, and vice versa for downward shift. During the transfer learning process, it is recommended to use a smaller learning rate to train the network. Since the parameters of the pre-trained model are reasonable after being trained by the leading emulator, there is no need to update the parameters too quickly, and the initial learning rate of transfer learning can be adjusted to 1 / 10 of the initial learning rate during the pre-training of the entire model. h During the transfer learning process, it is recommended to use a smaller learning rate to train the network. Since the parameters of the pre-trained model are reasonable after being trained by the leading emulator, there is no need to update the parameters too quickly, and the initial learning rate of transfer learning can be adjusted to 1 / 10 of the initial learning rate during the pre-training of the entire model.

[0155] Transfer learning utilizes the correlation between the leading emulator data and the real underwater acoustic data, significantly reducing the number of parameters that the network model needs to train and the demand for real underwater acoustic data. It can quickly switch between different leading detection tasks based on the pre-trained model, and preferably solves the problem of poor portability of neural networks in the leading detection method.

[0156] Based on Embodiment 4 of the present invention

[0157] Hereinafter, with reference to Figure 3 to describe the apparatus corresponding to the method according to Embodiments 1-3 of the present disclosure and Figure 1 , Figure 2 As shown, an underwater acoustic leading signal detection apparatus based on a leading emulator, the apparatus 300 includes a training set acquisition module 301, which is used to simulate leading data and interference data by the leading emulator, generate time-frequency images from the leading data and interference data through time-frequency transformation, and preprocess the generated time-frequency images to finally obtain a training set; a deep neural network model pre-training module 302, which is used to pre-train a deep neural network model using the training set, where the deep neural network model is a MobileNet network model; a deep neural network model transfer learning module 303, which is used to perform transfer learning on the pre-trained MobileNet network model using a small amount of underwater acoustic leading data obtained from real experiments, and optimize and train some parameters of the model to make the model applicable to the corresponding underwater acoustic leading detection task. In addition to the above three modules, the apparatus 300 may further include other components. However, since these components are not related to the content of the embodiments of the present disclosure, their illustrations and descriptions are omitted here.

[0158] The specific working process of an underwater acoustic leading signal detection apparatus 300 based on a leading emulator refers to the description of Embodiments 1-3 of the above-mentioned underwater acoustic leading signal detection method based on a leading emulator, and will not be elaborated here.

[0159] Based on Embodiment 5 of the present invention

[0160] Embodiments of the present invention can also be implemented as a computer-readable storage medium. Computer-readable instructions are stored on the computer-readable storage medium according to Embodiment 5. When the computer-readable instructions are run by a processor, the method for detecting an underwater acoustic preamble signal based on a preamble emulator according to Embodiments 1-3 of the present invention described with reference to the above drawings can be executed.

[0161] In summary, a method and device for detecting an underwater acoustic preamble signal based on a preamble emulator provided by the above embodiments mainly face the problems of insufficient general training data sets and diverse detection tasks in the process of underwater acoustic communication preamble detection. By establishing a simulated underwater acoustic data set through the preamble emulator, the problems of fewer real and available underwater acoustic data sets, non-uniform sample specifications, and limited types are solved; through the image preprocessing step of the preamble emulator, the problems that the STFT time-frequency image is sensitive to noise and has a high resolution are solved; by generating simulated underwater acoustic preamble data through the preamble emulator to pre-train the MobileNet neural network model, the problems that general networks do not have reasonable initial parameters and are prone to overfitting are solved; and transfer learning is used to help the pre-trained model quickly update the underwater acoustic preamble detection task; based on transfer learning, the learning method of the neural network in preamble detection is improved, and the portability of the neural network in preamble detection is improved. The specific advantages of this detection method are as follows: less real underwater acoustic data required for training: using the high correlation between the preamble emulator data and real underwater acoustic data, only a small number of key layers in the model need to be trained with real underwater acoustic data, greatly reducing the demand for real underwater acoustic data; good model portability: the simulated underwater acoustic data generated by the preamble emulator is rich in quantity and diverse in type, with good versatility and robustness, and is applicable to most underwater acoustic communication scenarios; fast training convergence and good detection effect: the pre-trained network model has more reasonable initial parameters, which can effectively narrow the training range and the number of training parameters, and accelerate the training efficiency; different detection tasks can be switched: when a new preamble detection task appears, through transfer learning, the parameters of the pre-trained model can be selectively optimized and adjusted to help the pre-trained model quickly adapt to the new preamble detection task; through data verification, the method and device for detecting an underwater acoustic preamble signal based on a preamble emulator of the present invention have lower signal detection errors, fewer data restrictions, and lower complexity compared with existing methods, and are more suitable for complex and changeable underwater acoustic environments and preamble detection tasks.

[0162] In this article, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a step or method comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such step or method.

[0163] Note that the above is only the preferred embodiment of the present invention and the technical principles applied. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein. Various obvious changes, re-adjustments and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments. Without departing from the concept of the present invention, more other equivalent embodiments can be included, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. An underwater acoustic preamble signal detection method based on a preamble emulator, characterized in that, It includes the following steps: Use a preamble emulator to simulate preamble data and interference data, generate a time-frequency image from the preamble data and interference data through time-frequency transformation, and preprocess the generated time-frequency image to finally obtain a training set; Use the training set to pre-train a deep neural network model, where the deep neural network model is a MobileNet network model; Use the underwater acoustic preamble data obtained from a small number of real experiments to perform transfer learning on the pre-trained MobileNet network model, optimize and train some parameters of the model, and make the model applicable to the corresponding underwater acoustic preamble detection task; Use the preamble emulator to obtain a training set, specifically including: Respectively establish a preamble signal and an underwater acoustic channel model, and an underwater interference noise model, set underwater environment parameters in each model, and use the preamble signal and the underwater acoustic channel model as well as the underwater interference noise model to generate time-series signal data; Use the short-time Fourier transform to convert the generated time-series signal data into a time-frequency image, perform normalization preprocessing on the time-frequency image, and obtain a training data set by classifying and labeling the pre-normalized video image; Replace the input layer and the fully connected layer in the MobileNet network model, specifically including: set the input layer in the model to the size of the time-frequency image generated by the preamble emulator, and the classifier of the fully connected layer uses the Sigmoid algorithm for binary classification tasks and the Softmax algorithm for multi-classification tasks; Use the underwater acoustic preamble data obtained from a small number of real experiments to perform transfer learning on the pre-trained MobileNet network model, optimize and train some parameters of the model, and make the model applicable to the corresponding underwater acoustic preamble detection task, specifically including: Use x1′, x2′,....., x m ′∈X′ to represent the true leading data samples, and y1′, y2′,..., y m ′∈Y′ to represent the labels corresponding to the true leading data samples. The pre-trained MobileNet network model is represented as: y i = f(x i ; μ) where μ represents the MobileNet model parameters and x i represents the model input sample data; Taking a certain layer in the MobileNet network model as the demarcation layer, the parameter μ is divided into μ l and μ h . The MobileNet model is expressed as y i = f(x i ; μ l , μ h ) where μ l and μ h represent the parameter sets above and below the demarcation layer respectively. According to the characteristics of the MobileNet model, μ l extracts shallow basic features, and μ h extracts deep abstract features; The transfer learning formula based on the MobileNet model is as follows: y′ i = f(x i ′; μ h ′) By freezing μ l , only re-training μ h based on a small amount of real data sets, the MobileNet model after transfer learning can be used for the leading detection task of real underwater acoustic data.

2. The underwater acoustic preamble signal detection method based on a preamble emulator according to claim 1, characterized in that, The establishment and parameter setting of the preamble signal and the underwater acoustic channel model, specifically including: Let the duration of the hyperbolic frequency modulation waveform be T, and define where f1 represents the start frequency of the hyperbolic frequency modulation waveform, and f2 represents the end frequency of the hyperbolic frequency modulation waveform; The generation formula of the preamble signal is: Among them A(t) represents a rectangular envelope, t represents time, and the leading signal parameters are set as follows: the center frequency of the signal is [12, 13] kHz, the bandwidth is [5, 6] kHz, the duration is 100 ms or 15 ms, and the signal-to-noise ratio is [-13, 12] dB; In the baseband, the underwater acoustic channel h(τ) is expressed as: where N pa represents the number of paths associated with the p-th path, A p represents the amplitude associated with the p-th path, τ p represents the delay associated with the p-th path, f c represents the center frequency, δ(t - τ p ) represents the impulse function, and τ represents the time delay; Then the received baseband signal x(t) is where ω(t) is the environmental noise and η(t) is the external interference, and the underwater acoustic channel parameters are set as follows: the number of paths N pa ~U(15, 25), and the arrival interval time τ of consecutive paths p -τ p-1 is exponentially distributed with an average value of 1 ms.

3. The underwater acoustic preamble signal detection method based on a preamble emulator according to claim 1, characterized in that, The underwater interference noise model includes a narrowband interference model, an impulse interference model, and a short-time narrowband interference model, where the specific expression of the narrowband interference model is: where f nb [i] represents the frequency of the i-th tone, A nb [i] represents the amplitude of the i-th tone, φ nb [i] represents the phase offset of the i-th tone, and the parameters of the narrowband interference are set as follows: the center frequency f c ~U(11.5, 13.5) kHz, the bandwidth B~U(1, 3000) Hz, the duration is 200 ms, the interference-to-noise ratio INR~U(-5, 10) dB, and N nb represents multiple tones; The specific expression of the impulse interference model is: where CN() represents the complex Gaussian distribution, p represents the mixing ratio of the Gaussian mixture model, represents the energy of Gaussian noise, represents the energy of impulse noise, and the parameters of the impulse interference are set as: p ~ U(0.001, 0.01), with a duration of 200 ms; The specific expression of the short-time narrowband interference model is: Among them, c l represents the coefficient of the Fourier series, B1 represents the bandwidth of the short-time narrowband interference, T1 represents the duration of the short-time narrowband interference, and the parameters of the short-time narrowband interference are set as follows: the center frequency f c ~U(11.5, 13.5) kHz, the bandwidth B~U(1, 3) kHz, the duration T PBPD ~U(5, 60) ms, and the interference-to-noise ratio INR~U(2, 15) dB.

4. The underwater acoustic preamble signal detection method based on a preamble emulator according to claim 1, wherein, Use the short-time Fourier transform to convert the generated time-series signal data into a time-frequency image, and the two-dimensional data expression v(m, l) of the time-frequency image is: v(m, l) = |x(m, l)| 2 wherein, is a complex discrete-time signal vector, and N x represents the block length. The signal vector is obtained by equidistant sampling according to a given fixed sampling rate fs. W = [w[0], w[1],..., w[N-1]] T ∈ C N×1 is a sampling window function, N represents the length parameter, H represents the hop size parameter, m ∈ [0:M], represents the maximum frame index, l ∈ [-L+1:L], L = N / 2 represents the frequency index corresponding to the Nyquist frequency, T is the matrix transpose, m, l represent the l-th Fourier coefficient of the m-th time frame of the complex number x(m, l).

5. The underwater acoustic preamble signal detection method based on a preamble emulator according to claim 4, wherein, Perform normalization preprocessing on the time-frequency image, and obtain a training data set by classifying and labeling the pre-normalized video image, specifically including: Perform grayscale processing on the time-frequency image, and the grayscale formula of the RGB image is as follows: where R, G, and B respectively represent the color brightness of the red, green, and blue image channels; Convert the grayscale processed image to double precision and perform gray adjustment through linear mapping; Compress the resolution of the time-frequency image after gray adjustment, automatically generate samples and labels, and save the samples to the file corresponding to the label in the emulator data set.

6. An underwater acoustic preamble signal detection device based on a preamble emulator, wherein, The device includes: The training set acquisition module is used to simulate preamble data and interference data by using a preamble emulator, generate time-frequency images from the preamble data and interference data through time-frequency transformation, and preprocess the generated time-frequency images to finally obtain a training set; The deep neural network model pre-training module is used to pre-train the deep neural network model by using the training set, where the deep neural network model is a MobileNet network model; The deep neural network model transfer learning module is used to perform transfer learning on the pre-trained MobileNet network model by using the underwater acoustic preamble data obtained from a small number of real experiments, optimize and train some parameters of the model, and make the model applicable to the corresponding underwater acoustic preamble detection task; Replace the input layer and the fully connected layer in the MobileNet network model, specifically including: setting the input layer in the model to the size of the time-frequency image generated by the preamble emulator, and using the Sigmoid algorithm for the classifier in the fully connected layer for binary classification tasks and the Softmax algorithm for multi-classification tasks; Perform transfer learning on the pre-trained MobileNet network model by using the underwater acoustic preamble data obtained from a small number of real experiments, optimize and train some parameters of the model, and make the model applicable to the corresponding underwater acoustic preamble detection task, specifically including: Use \(x_1', x_2', \ldots, x\) m '\(\in X'\) to represent the true leading data samples, and \(y_1', y_2', \ldots, y\) m '\(\in Y'\) to represent the labels corresponding to the true leading data samples. The pre-trained MobileNet network model is represented as: y i = f(x i ; μ) where μ represents the MobileNet model parameters, and x i represents the model input sample data; Divide the parameter μ into μ l and μ h with a certain layer in the MobileNet network model as the demarcation layer. The MobileNet model is expressed as y i = f(x i ; μ l , μ h ) where μ i and μ h represent the parameter sets above and below the demarcation layer respectively. From the characteristics of the MobileNet model, it can be known that μ l extracts shallow basic features, and μ h extracts deep abstract features; The transfer learning formula based on the MobileNet model is as follows: y′ i = f(x i ′; μ h ′) By freezing μ l and only retraining μ h using a small amount of real datasets, the MobileNet model after transfer learning can be used for the leading detection task of real underwater acoustic data.

7. A computer-readable medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the underwater acoustic preamble signal detection method based on a preamble emulator as described in any one of claims 1-5.

Images