Underwater acoustic communication concealment performance evaluation method, device and equipment based on neural network

Abstract

The invention relates to the field of bionic hidden underwater acoustic communication, in particular to an underwater acoustic communication hidden performance evaluation method, device and equipment based on a neural network, and the method comprises the steps: carrying out the feature extraction of a to-be-evaluated underwater acoustic signal, and obtaining a feature sequence of the to-be-evaluated underwater acoustic signal; converting the feature sequence into a two-dimensional matrix with a preset size; the two-dimensional matrix is predicted according to a pre-trained underwater acoustic signal classification neural network model, a probability matrix is obtained, the probability matrix comprises a first probability and a second probability, the first probability represents the probability that the to-be-evaluated underwater acoustic signal belongs to the bionic modulation signal, and the second probability represents the probability that the to-be-evaluated underwater acoustic signal belongs to the bionic modulation signal; the second probability represents the probability that the underwater acoustic signal to be evaluated belongs to the real marine organism signal, and the underwater acoustic signal classification neural network model comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a global pooling layer and a full-connection output layer which are connected in sequence. The dependence on artificial feature design in a traditional recognition method is reduced, and the recognition stability in a complex underwater environment is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bionic covert underwater acoustic communication, and in particular to a method, device and equipment for evaluating the performance of bionic covert underwater acoustic communication based on a neural network. BACKGROUND

[0002] The method of bionic covert underwater acoustic communication is a low probability of detection communication, which uses the calling sound of marine organisms as a communication signal, loads information onto the calling sound signal of the organisms, so that the non-cooperative party will misjudge the communication signal as the calling sound of marine organisms after detecting the communication signal, and finally achieve the purpose of covert underwater acoustic communication. At present, the calling sound signals of cetaceans, including whistle signals and click signals, are mostly used in bionic covert underwater acoustic communication. Among them, the whistle signal of cetaceans is similar to a chirp signal, which is a frequency-modulated signal with regular changes in frequency over time, and has good time-frequency characteristics. The duration of the whistle signal is usually several hundred milliseconds to several seconds, and the frequency range is between several hundred hertz and several ten thousand hertz.

[0003] The performance analysis of bionic modulation schemes mainly includes detection performance, communication performance and concealment performance. With the rapid development of bionic covert communication, there is a lack of objective and effective evaluation methods for the concealment of underwater acoustic communication systems. Therefore, it is particularly important to quickly and accurately evaluate the concealment of bionic modulation signals and judge their similarity to real dolphin whistle signals. However, most of the current domestic and foreign researches only evaluate the concealment from a single or a small number of features, resulting in a large error in the results. SUMMARY

[0004] The present application aims to provide a method, device and equipment for evaluating the performance of bionic covert underwater acoustic communication based on a neural network, which solves the problems in the prior art.

[0005] The present application is achieved by the following technical solutions:

[0006] In a first aspect, the present application provides a method for evaluating the performance of bionic covert underwater acoustic communication based on a neural network, which includes:

[0007] extracting features from the underwater acoustic signal to be evaluated to obtain a feature sequence of the underwater acoustic signal to be evaluated;

[0008] converting the feature sequence into a two-dimensional matrix of a predetermined size;

[0009] According to a pre-trained underwater acoustic signal classification neural network model, the two-dimensional matrix is predicted to obtain a probability matrix, the probability matrix comprising a first probability and a second probability, the first probability representing a probability that the to-be-evaluated underwater acoustic signal belongs to a bionic modulated signal, and the second probability representing a probability that the to-be-evaluated underwater acoustic signal belongs to a real marine biological signal, the underwater acoustic signal classification neural network model comprising an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a global pooling layer and a fully connected output layer connected in sequence.

[0010] Preferably, the feature extraction of the to-be-evaluated underwater acoustic signal to obtain the feature sequence of the to-be-evaluated underwater acoustic signal comprises:

[0011] The to-be-evaluated underwater acoustic signal is preprocessed to obtain a plurality of preprocessed signal frames, and the preprocessing comprises pre-emphasis, framing and windowing;

[0012] The fast Fourier transform is performed on each signal frame to obtain a power spectrum of each signal frame;

[0013] According to a pre-set mel filter bank, the energy of each frequency point in the power spectrum of each signal frame is weighted and summed with the transfer function of each mel filter to obtain the output energy of each signal frame on each mel filter;

[0014] According to the output energy of each signal frame on each mel filter, the feature sequence of each signal frame is obtained.

[0015] Preferably, the preprocessing of the to-be-evaluated underwater acoustic signal to obtain a plurality of preprocessed signal frames comprises:

[0016] The to-be-evaluated underwater acoustic signal is pre-emphasized by a first-order finite impulse response high-pass filter to obtain a pre-emphasized signal;

[0017] According to pre-set frame length and frame shift parameters, the pre-emphasized signal is divided into a plurality of overlapping short-time frames to obtain a time-domain signal of each frame;

[0018] According to a window function, a windowing operation is performed on the time-domain signal of each frame to obtain a plurality of windowed signal frames.

[0019] Preferably, the feature sequence of each signal frame obtained according to the output energy of each signal frame on each mel filter satisfies:

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] in, For the first The first signal frame One characteristic, The number of filter banks. For the first The first signal frame Output energy after a Mel filter For the first The power spectral energy of a signal frame at frequency point k Let be the transfer function of the m-th Mel filter. For the first The complex spectrum obtained by performing FFT on each signal frame For the first The time-domain sampling points of each signal frame.

[0025] Preferably, the step of predicting the two-dimensional matrix based on a pre-trained underwater acoustic signal classification neural network model to obtain a probability matrix includes:

[0026] Based on the pre-trained underwater acoustic signal classification neural network model, the two-dimensional matrix is ​​input into the input layer of the underwater acoustic signal classification neural network model, and forward propagation calculation is performed sequentially through the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer and the global pooling layer to obtain the high-level feature vector output by the global pooling layer.

[0027] Based on the high-level feature vector and the weight parameters and biases of the fully connected output layer, the high-level feature vector is input into the fully connected output layer to calculate the original output values ​​corresponding to different categories, thus obtaining the original output value vector.

[0028] The original output value vector is normalized to obtain the probability matrix.

[0029] Preferably, the underwater acoustic signal classification neural network model is trained in the following manner:

[0030] Obtain a training dataset, which contains multiple labeled real marine biological signal samples and biomimetic modulation signal samples;

[0031] For each signal sample in the training dataset, feature extraction and preprocessing are performed to obtain a standardized set of feature matrices;

[0032] The preset initial model is iteratively trained using the set of feature matrices, and the model parameters are adjusted through optimization algorithms.

[0033] When the training process meets a preset stop condition, the training is stopped, and a trained underwater acoustic signal classification neural network model is obtained; wherein the underwater acoustic signal classification neural network model is configured to receive a feature matrix of an underwater acoustic signal and output a probability matrix.

[0034] Preferably, the underwater acoustic signal classification neural network model comprises:

[0035] an input layer configured to receive an input matrix of 64x64 size;

[0036] a first convolutional layer configured with 32 3x3 convolutional kernels;

[0037] a first pooling layer configured to perform 2x2 max pooling;

[0038] a second convolutional layer configured with 64 3x3 convolutional kernels;

[0039] a second pooling layer configured to perform 2x2 max pooling;

[0040] an L2 regularization layer for imposing constraints on model parameters;

[0041] a Dropout layer configured to randomly drop neurons with a probability of 0.3 during training;

[0042] a global pooling layer;

[0043] a fully connected output layer configured with two neurons and using a Softmax activation function.

[0044] In a second aspect, an embodiment of the present application provides a device for evaluating the concealment performance of underwater acoustic communication based on a neural network, comprising:

[0045] a feature extraction module for extracting features from a to-be-evaluated underwater acoustic signal to obtain a feature sequence of the to-be-evaluated underwater acoustic signal;

[0046] a conversion module for converting the feature sequence into a two-dimensional matrix of a preset size;

[0047] a prediction module for predicting the two-dimensional matrix according to a pre-trained underwater acoustic signal classification neural network model to obtain a probability matrix, wherein the probability matrix comprises a first probability and a second probability, the first probability representing the probability that the to-be-evaluated underwater acoustic signal belongs to a bionic modulated signal, and the second probability representing the probability that the to-be-evaluated underwater acoustic signal belongs to a real marine biological signal, and the underwater acoustic signal classification neural network model comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a global pooling layer, and a fully connected output layer connected in sequence.

[0048] In a third aspect, an electronic device is provided, comprising at least one processor, at least one memory, and computer program instructions stored in the memory that, when executed by the processor, implement the method of the first aspect in the above-described embodiments.

[0049] In a fourth aspect, a storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the method of the first aspect in the above-described embodiments.

[0050] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0051] The technical scheme provided by the embodiments of the present application realizes end-to-end automatic processing of the underwater acoustic communication concealment performance evaluation process. By converting the underwater acoustic signal into a standardized two-dimensional matrix, the convolutional neural network can effectively extract the local feature pattern of the signal. The combination of the convolutional layer and the pooling layer in the network structure realizes hierarchical extraction and compression of the signal features, and the global pooling layer ensures the adaptability of the model to different lengths of input. The final output probability matrix intuitively reflects the possibility of the signal to be identified belonging to the bionic modulation signal and the real marine biological signal in a quantitative form, providing an objective and reliable basis for the underwater acoustic communication concealment performance evaluation. This method reduces the dependence on artificial feature design in traditional recognition methods and improves the recognition stability in complex underwater environments. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical scheme of the exemplary embodiments of the present application, the following will briefly introduce the drawings needed in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be considered as limiting the scope. For those skilled in the art, other related drawings can also be obtained without creative labor. In the drawings:

[0053] Figure 1 A flowchart of the neural network-based underwater acoustic communication concealment performance evaluation method provided by the present application;

[0054] Figure 2 A Loss fitting curve diagram of the trained model provided by the present application;

[0055] Figure 3 The network model provided by the present application identifies 500 unknown modulation signals and obtains the feature similarity probability graph with bionic modulation signals;

[0056] Figure 4 A structure diagram of the neural network-based underwater acoustic communication concealment performance evaluation device provided by the present application;

[0057] Figure 5 The structural schematic diagram of the electronic device provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0058] In order to make the objectives, technical solutions and advantages of the present application clearer, further detailed description will be given to the present application in combination with embodiments and drawings, and the illustrative embodiments of the present application and the description thereof are only used to explain the present application, and do not limit the present application.

[0059] It should be noted that, in this document, the terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitation, the elements defined by the statement "include" do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0060] It should be noted that all the actions of acquiring signals, information or data in the present application are carried out in compliance with the corresponding data protection regulations and policies of the place, and with the authorization given by the owner of the corresponding device.

[0061] Embodiment 1

[0062] Please refer to Figure 1 The embodiment of the present application provides a neural network-based underwater acoustic communication concealment performance evaluation method, which comprises the following steps:

[0063] S1, feature extraction is performed on a to-be-evaluated underwater acoustic signal to obtain a feature sequence of the to-be-evaluated underwater acoustic signal;

[0064] Specifically, the feature sequence can be an MFCC feature sequence.

[0065] MFCC (Mel Frequency Cepstral Coefficient, Mel Frequency Cepstral Coefficient) is used to analyze the basis of the auditory mechanism, which has two points:

[0066] 1. The height of the sound perceived by the human ear and the frequency of the sound wave are nonlinearly related, so the Mel frequency scale is used to describe the frequency spectrum characteristics of the audio signal, which is more in line with the auditory characteristics of the human ear. The specific relationship is expressed as

[0067] ;

[0068] wherein, is the frequency of the sound wave.

[0069] 2. The auditory masking effect of human ear has the phenomenon of critical band. The characteristics of speech signal in frequency domain can be described by a series of frequency groups, which are obtained by dividing the critical band, and vary with the frequency and are consistent with the trend of the increase of the mel frequency. In the frequency range below 1000 Hz, the critical band is approximately linearly distributed; while in the frequency range above 1000 Hz, it shows a logarithmic growth trend. In order to divide the speech signal in frequency domain, the characteristics of the critical band can be used to design a series of triangular band-pass filters, which form a mel filter bank. The transfer function of each band-pass filter is:

[0070] ;

[0071] wherein, f(m) is the center frequency of the filter. Its definition is:

[0072] ;

[0073] wherein, and are the highest frequency and the lowest frequency of the filter bank respectively, f s is the sampling frequency, M f is the number of the filter bank, N FFT is the number of FFT points, f 梅尔 -1 (x) = 700(e x / 1125 -1).

[0074] In some possible implementation manners, the feature sequence can also be a hybrid time-frequency feature sequence, and the feature extraction on the water acoustic signal to be evaluated to obtain the feature sequence of the water acoustic signal to be evaluated comprises:

[0075] performing wavelet packet transform on the water acoustic signal to be evaluated to obtain time-frequency detail features;

[0076] calculating the spectral kurtosis of the signal in a preset frequency band according to the water acoustic signal to be evaluated to obtain transient feature key frequency band information;

[0077] guiding a preset mel filter bank to perform adaptive offset and aggregation according to the transient feature key frequency band information to obtain enhanced auditory perception features;

[0078] splicing the time-frequency detail features, the enhanced auditory perception features, and time-domain statistical features of the water acoustic signal to be evaluated to obtain a hybrid time-frequency feature sequence.

[0079] Specifically, the feature extraction method realizes deep feature mining of underwater acoustic signals through a multi-signal processing mechanism. First, wavelet packet transform is used to finely decompose the signal in the time-frequency domain. By using different time window lengths in different frequency bands, the high-frequency transient components and low-frequency persistent structures are captured synchronously, thereby obtaining time-frequency detail features representing the details of the local time-frequency energy distribution of the signal. At the same time, the spectral kurtosis of the signal in each preset frequency band is calculated to quantify the strength of the non-stationary impact component. According to the kurtosis peak positioning mechanism, the core frequency band containing the key transient information is identified, and the transient feature key frequency band information is generated. Based on this key frequency band information, the parameters of the standard mel filter bank are dynamically adjusted, so that the center frequency of the mel filter bank is adaptively shifted and the density is concentrated in the key frequency band area, thereby extracting enhanced auditory perception features focused on the discriminative region of the signal. Finally, the time-frequency detail features, enhanced auditory perception features, and time-domain statistical features directly extracted from the original signal are spliced in multiple dimensions to form a hybrid time-frequency feature sequence that combines time-frequency local characteristics, enhanced auditory perception, and global statistical rules, providing a more complete feature representation for the subsequent classification model.

[0080] S2, converting the feature sequence into a two-dimensional matrix of a preset size;

[0081] Specifically, this step structures and reshapes the mel frequency cepstral coefficient feature sequence into a tensor format suitable for neural network processing. The feature sequence is essentially a sequence of cepstral coefficient vectors arranged in time frames. This one-dimensional time sequence structure cannot fully utilize the advantages of convolutional neural networks in two-dimensional local pattern recognition. By reconstructing it into a two-dimensional matrix of a preset size, the time sequence features are essentially mapped into spatial features, where the rows and columns of the matrix correspond to specific feature dimensions and time dimensions, respectively. This conversion not only preserves the topological relationship between features, but also meets the requirements of convolutional operations on input data formats. In specific implementation, interpolation, truncation, or padding can be used to unify feature sequences of different lengths to a fixed size, ensuring dimensional consistency during batch processing. This operation enables the subsequent neural network to extract local correlations in the feature matrix in an image processing manner, laying the foundation for deep feature learning.

[0082] S3, predicting the two-dimensional matrix according to a pre-trained underwater acoustic signal classification neural network model to obtain a probability matrix, the probability matrix including a first probability and a second probability, the first probability representing a probability that the underwater acoustic signal to be evaluated belongs to a biomimetic modulated signal, and the second probability representing a probability that the underwater acoustic signal to be evaluated belongs to a real marine biological signal, the underwater acoustic signal classification neural network model including an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a global pooling layer, and a fully connected output layer connected in sequence.

[0083] Specifically, this step realizes the probabilistic inference of signal categories by constructing a multi-layer feature transformation network. The two-dimensional matrix first enters the input layer for data normalization, then performs primary feature extraction through the first convolutional layer, capturing local spectral patterns by sliding the convolution kernel over the matrix space; the first pooling layer reduces the dimension of the feature map, preserving significant features while improving computational efficiency; the second convolutional layer performs deep feature abstraction based on this, integrating acoustic features of different frequency bands through multi-channel convolution operations; the second pooling layer further compresses the feature dimension, enhancing the spatial invariance of the features; the global pooling layer aggregates the two-dimensional feature map into a compact feature vector, eliminating the effects of input length differences; the fully connected output layer finally maps the high-level features to the category confidence. Through hierarchical feature transformation, the discriminative features that distinguish between biomimetic modulated signals and real marine biological signals are gradually extracted from the bottom acoustic features, and finally the output is converted to a probability distribution through a normalization exponential function, where the first probability reflects the artificial modulation characteristics of the signal, and the second probability reflects the natural biological characteristics of the signal. The comparison of the two probability values provides a quantitative basis for concealment evaluation.

[0084] In some embodiments, S1, feature extraction is performed on the water acoustic signal to be evaluated to obtain a feature sequence of the water acoustic signal to be evaluated, including:

[0085] S11, pre-processing the water acoustic signal to be evaluated to obtain a plurality of pre-processed signal frames, the pre-processing including pre-emphasis, framing, and windowing;

[0086] Specifically, this step converts the original water acoustic signal into a short-time stationary signal frame suitable for spectral analysis through signal preprocessing. Pre-emphasis uses a first-order high-pass filter to boost high-frequency components of the signal, compensating for the attenuation of high frequencies in the underwater channel, making the signal spectrum flat, for example, increasing components above 2000Hz by 3-6dB. Framing divides the continuous signal into overlapping short time periods with a fixed time length, usually with a frame length of 20-40 milliseconds and a 1 / 2 frame shift overlap, converting the non-stationary signal into multiple quasi-stationary processing units. The windowing operation uses a smoothing window function such as the Hamming window to weight each frame of signal, suppressing spectral leakage by reducing signal discontinuities at the frame boundaries, making the spectral profile after Fourier transform clearer. These three consecutive operations together ensure that subsequent feature extraction is based on stable and reliable short-time signals, providing a necessary prerequisite for obtaining representative acoustic features.

[0087] In some embodiments, the pre-processing of the water acoustic signal to be evaluated to obtain a plurality of pre-processed signal frames includes:

[0088] According to a first-order finite impulse response high-pass filter, the water acoustic signal to be evaluated is pre-emphasized to obtain a pre-emphasized signal;

[0089] Based on the preset frame length and frame shift parameters, the pre-emphasized signal is divided into multiple overlapping short-time frames to obtain the time-domain signal of each frame.

[0090] The time-domain signal of each frame is windowed according to the window function to obtain multiple windowed signal frames.

[0091] S12. Perform a fast Fourier transform on each signal frame to obtain the power spectrum of each signal frame.

[0092] Specifically, this step converts the time-domain signal frame into a frequency-domain energy distribution representation. A Fast Fourier Transform (FFT) is performed on each windowed signal frame, reconstructing the signal from its time-domain representation to its frequency-domain representation, revealing the signal's compositional characteristics at different frequency components. The transform result is then subjected to modulo-squaring to obtain the power spectrum, which reflects the distribution density of signal energy along the frequency axis. The power value corresponding to each frequency point represents the contribution of that frequency component to the signal energy. This time-frequency conversion allows the spectral characteristics of the signal to be explicitly presented, laying the foundation for subsequent feature extraction based on auditory perception. Simultaneously, the power spectrum calculation eliminates the influence of phase information, focusing more on characterizing the signal's essential energy distribution pattern.

[0093] S13. Based on the preset Mel filter bank, the energy of each frequency point in the power spectrum of each signal frame is weighted and summed with the transfer function of each Mel filter to obtain the output energy of each signal frame on each Mel filter.

[0094] Specifically, this step maps the linear spectrum to a Mel-scale spectrum by simulating the characteristics of human hearing. The pre-defined Mel filter bank consists of a set of overlapping triangular bandpass filters, uniformly distributed along the Mel frequency axis and logarithmically distributed along the Hertzian scale, conforming to the nonlinear nature of human pitch perception. The transfer function of each Mel filter defines its weighting coefficients at different frequencies, forming a continuous coverage from low to high frequencies. During calculation, the energy value at each frequency point in the power spectrum is multiplied by the corresponding filter transfer function coefficient and summed to achieve the mapping from the linear spectrum to the Mel band. This process is equivalent to passing the signal through a set of filters simulating the frequency-resolving characteristics of the cochlea, ultimately outputting the integrated energy of the signal within each Mel band. This representation method both compresses the data dimensionality and highlights the spectral characteristics related to auditory perception.

[0095] S14. Based on the output energy of each signal frame on each Mel filter, obtain the characteristic sequence of each signal frame.

[0096] Specifically, this step extracts cepstral features through logarithmic compression and discrete cosine transform (DCT). First, the logarithm of the output energy of each Mel filter is taken to simulate the nonlinear response of the human ear to sound intensity, converting multiplicative noise into additive noise. Then, a DCT is applied to the logarithmic Mel spectrum, transforming the spectral domain features to the cepstral domain, thus separating the acoustic excitation source from the transmission characteristics of the vocal tract. The DCT has excellent energy compression properties, concentrating signal information in the first few coefficients. Typically, the first 12-14 coefficients are retained to form a feature sequence. These coefficients characterize the shape of the signal's spectral envelope, with lower-order coefficients reflecting the smooth contour of the spectrum and higher-order coefficients describing the fine structure of the spectrum. The resulting feature sequence preserves the essential characteristics of sound while exhibiting robustness to noise and channel variations, providing an effective feature representation for subsequent pattern recognition.

[0097] The characteristic sequence of each signal frame is obtained based on the output energy of each signal frame on each Mel filter, satisfying the following:

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] in, For the first The first signal frame One characteristic, The number of filter banks. For the first The first signal frame Output energy after a Mel filter For the first The power spectral energy of a signal frame at frequency point k Let be the transfer function of the m-th Mel filter. For the first The complex spectrum obtained by performing FFT on each signal frame For the first The time-domain sampling points of each signal frame.

[0103] In some embodiments, S3, predicting the two-dimensional matrix based on a pre-trained underwater acoustic signal classification neural network model to obtain a probability matrix, includes:

[0104] S31. Based on the pre-trained underwater acoustic signal classification neural network model, the two-dimensional matrix is ​​input into the input layer of the underwater acoustic signal classification neural network model, and forward propagation calculation is performed sequentially through the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer and the global pooling layer to obtain the high-level feature vector output by the global pooling layer.

[0105] S32. Based on the high-level feature vector and the weight parameters and biases of the fully connected output layer, the high-level feature vector is input to the fully connected output layer to calculate the original output values ​​corresponding to different categories, thus obtaining the original output value vector.

[0106] S33. Normalize the original output value vector to obtain the probability matrix.

[0107] Specifically, this step transforms the original output values ​​into a standard probability distribution form through mathematical transformation. The normalization process uses the Softmax function to transform the original output value vector. This function amplifies the relative differences between output values ​​of each category through exponential operations, and then uses a normalization factor to constrain all exponential operation results within the range of 0 to 1, ensuring that the sum of the probabilities of all categories is 1. This transformation process converts the original output values ​​of arbitrary recognition into confidence scores with clear probabilistic meaning. The probability value of each category independently represents the model's confidence in predicting that category, while the relative magnitude of all category probabilities clearly reflects the model's judgment tendency among different categories. The final probability matrix presents the complete evaluation result of the neural network on the category to which the input signal belongs in a standardized mathematical form, providing a directly interpretable quantitative basis for subsequent covertness judgments.

[0108] In some embodiments, the underwater acoustic signal classification neural network model is trained in the following manner:

[0109] Obtain a training dataset, which contains multiple labeled real marine biological signal samples and biomimetic modulation signal samples;

[0110] For each signal sample in the training dataset, feature extraction and preprocessing are performed to obtain a standardized set of feature matrices;

[0111] The preset initial model is iteratively trained using the set of feature matrices, and the model parameters are adjusted through optimization algorithms.

[0112] When the training process meets the preset stopping condition, the training stops, and a trained underwater acoustic signal classification neural network model is obtained; wherein, the underwater acoustic signal classification neural network model is configured to receive the feature matrix of underwater acoustic signals and output the probability matrix.

[0113] Specifically, the underwater acoustic signal classification neural network model is trained using a training dataset containing multiple labeled real marine biological signals and biomimetic modulation signals. First, standardized feature extraction and preprocessing are performed on each signal sample to obtain a unified set of feature matrices. Then, the initial model is iteratively trained using this set. The model parameters are continuously adjusted through optimization algorithms to enable it to learn the feature differences between the two types of signals. The training process is completed when the stopping condition is set based on the performance of the validation set or the number of iterations. The final model can map the input underwater acoustic signal feature matrix into a probability matrix representing whether it belongs to a real marine biological signal or a biomimetic modulation signal.

[0114] Furthermore, the underwater acoustic signal classification neural network model includes:

[0115] The input layer is configured to receive an input matrix of size 64×64;

[0116] The first convolutional layer has 32 3×3 convolutional kernels;

[0117] The first pooling layer is configured to perform 2×2 max pooling;

[0118] The second convolutional layer is configured with 64 3×3 convolutional kernels;

[0119] The second pooling layer is configured to perform 2×2 max pooling;

[0120] L2 regularization layer is used to impose constraints on model parameters;

[0121] The Dropout layer is configured to randomly drop neurons with a probability of 0.3 during training.

[0122] Global pooling layer;

[0123] The fully connected output layer has two neurons and uses the Softmax activation function.

[0124] Specifically, the network parameters are shown in the table below. The stride is always the same, the activation function for convolutional layers is ReLU, and the activation function for classification layers is softmax.

[0125]

[0126] The following example illustrates this embodiment. A dataset was collected, consisting of 600 dolphin whistles and 600 biomimetic modulated signals. Six whistle signals and six biomimetic signals were randomly selected as the test set, and the rest were used as the training set. First, the MFCC of the signals was extracted, converting the signals into feature matrices. These feature matrices were then used as input to the neural network.

[0127] Then, the network parameters are updated through forward and backpropagation training to obtain a well-fitted neural network model suitable for classification. The Adam optimizer is used, with a learning rate of 0.001, training for 50 epochs, a batch size of 16, and randomized data. The SCE loss function is chosen.

[0128] ;

[0129] Where C represents the total number of categories. The final softmax classification layer calculates y, which represents the true label. The loss fitting curve of the trained model is shown below. Figure 2 As shown in the figure, blue represents the loss on the training set, and orange represents the loss on the validation set. The loss curves indicate that the network has achieved a good fit.

[0130] Finally, the trained model was applied to the test set for classification, yielding the classification results. The 1×2 probability matrix (P1, P2) output by the softmax classification function of the last layer of the network was used as the classification criterion, where P1 represents the feature similarity between the input data and the biomimetic modulation signal, and P2 represents the feature similarity between the input data and the real dolphin whistle. When the value of P1 in the first column of the probability matrix is ​​greater than 0.5, it indicates that the input data features are more similar to the biomimetic signal, and the model considers the input data to belong to the biomimetic modulation signal; when the value of P2 in the second column is greater than 0.5, it indicates that the input data features are more similar to the real dolphin whistle, and the model considers the data to belong to the real dolphin signal. In the final test results, the probability of successfully classifying the biomimetic signal as a biomimetic signal was 93.9%, and the probability of successfully classifying the real dolphin whistle as a dolphin whistle was 92.3%, demonstrating good differentiation between dolphin whistles and biomimetic modulation signals. In addition, 500 different modulation signals were generated for testing, and their feature similarity probabilities with the biomimetic modulation signals were obtained as follows: Figure 3 As shown, when the network inputs a biomimetic signal, if the output P1 is greater than 0.5, the model is considered to have successfully recognized the biomimetic modulation signal. Therefore, it is evident that the model also has good recognition capabilities for unseen data and possesses a certain degree of robustness. In conclusion, the algorithm of this patent invention is feasible.

[0131] Example 2

[0132] Please see Figure 4 This invention provides a neural network-based underwater acoustic communication covert performance evaluation device, comprising:

[0133] Feature extraction module 401 is used to extract features from the underwater acoustic signal to be evaluated, and obtain the feature sequence of the underwater acoustic signal to be evaluated;

[0134] The conversion module 402 is used to convert the feature sequence into a two-dimensional matrix of a preset size;

[0135] The prediction module 403 is used to predict the two-dimensional matrix based on a pre-trained underwater acoustic signal classification neural network model to obtain a probability matrix. The probability matrix includes a first probability and a second probability. The first probability represents the probability that the underwater acoustic signal to be evaluated belongs to a biomimetic modulation signal, and the second probability represents the probability that the underwater acoustic signal to be evaluated belongs to a real marine biological signal. The underwater acoustic signal classification neural network model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a global pooling layer, and a fully connected output layer connected in sequence.

[0136] It should be noted that each module and unit in the neural network-based underwater acoustic communication covert performance evaluation device in this embodiment corresponds one-to-one with each step in the neural network-based underwater acoustic communication covert performance evaluation method in the aforementioned embodiment. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned neural network-based underwater acoustic communication covert performance evaluation method, which will not be repeated here.

[0137] Example 3

[0138] Please see Figure 5 This embodiment provides an electronic device, including at least one processor 501 and a memory 502. Optionally, the device further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0139] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0140] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0141] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0142] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0143] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0144] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0145] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0146] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0147] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0148] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0149] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0150] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0151] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0153] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the concealment performance of underwater acoustic communication based on a neural network, characterized in that, The method comprises the following steps: characteristic extraction is performed on the to-be-evaluated underwater acoustic signal to obtain a characteristic sequence of the to-be-evaluated underwater acoustic signal; the characteristic sequence is converted into a two-dimensional matrix of a preset size; a probability matrix is obtained by predicting the two-dimensional matrix according to a pre-trained underwater acoustic signal classification neural network model, the probability matrix comprising a first probability and a second probability, the first probability representing a probability that the to-be-evaluated underwater acoustic signal belongs to a bionic modulated signal, and the second probability representing a probability that the to-be-evaluated underwater acoustic signal belongs to a real marine biological signal, the underwater acoustic signal classification neural network model comprising an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a global pooling layer and a fully-connected output layer connected in sequence.

2. The method of claim 1, wherein, The characteristic extraction performed on the to-be-evaluated underwater acoustic signal to obtain the characteristic sequence of the to-be-evaluated underwater acoustic signal comprises the following steps: preprocessing is performed on the to-be-evaluated underwater acoustic signal to obtain a plurality of preprocessed signal frames, the preprocessing comprising pre-emphasis, framing and windowing; fast Fourier transform is performed on each signal frame to obtain a power spectrum of each signal frame; the energy of each frequency point in the power spectrum of each signal frame is weighted and summed with the transfer function of each Mel filter according to a pre-set Mel filter bank to obtain the output energy of each signal frame on each Mel filter; a characteristic sequence of each signal frame is obtained according to the output energy of each signal frame on each Mel filter.

3. The method of claim 2, wherein, The preprocessing performed on the to-be-evaluated underwater acoustic signal to obtain a plurality of preprocessed signal frames comprises the following steps: pre-emphasis is performed on the to-be-evaluated underwater acoustic signal according to a first-order finite impulse response high-pass filter to obtain a pre-emphasized signal; the pre-emphasized signal is divided into a plurality of overlapping short-time frames according to pre-set frame length and frame shift parameters to obtain a time-domain signal of each frame; windowing is performed on the time-domain signal of each frame according to a window function to obtain a plurality of windowed signal frames.

4. The method of claim 2, wherein, The characteristic sequence of each signal frame obtained according to the output energy of each signal frame on each Mel filter satisfies the following condition: ； ； ； ； in, For the first The first signal frame One characteristic, The number of filter banks. For the first The first signal frame Output energy after a Mel filter For the first The power spectral energy of a signal frame at frequency point k Let be the transfer function of the m-th Mel filter. For the first The complex spectrum obtained by performing FFT on each signal frame For the first The time-domain sampling points of each signal frame.

5. The method of claim 1, wherein, The prediction of the two-dimensional matrix according to the pre-trained underwater acoustic signal classification neural network model to obtain the probability matrix comprises the following steps: the two-dimensional matrix is input into the input layer of the underwater acoustic signal classification neural network model, and the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer and the global pooling layer are sequentially passed through for forward propagation calculation to obtain a high-level feature vector output by the global pooling layer; the high-level feature vector is input into the fully-connected output layer according to the high-level feature vector and the weight parameters and bias of the fully-connected output layer, and original output values corresponding to different categories are calculated to obtain an original output value vector; the original output value vector is normalized to obtain the probability matrix.

6. The method according to any one of claims 1 to 5, characterized in that, The underwater acoustic signal classification neural network model is obtained by training in the following manner: a training data set is obtained, the training data set containing a plurality of labeled real marine biological signal samples and bionic modulated signal samples; Feature extraction and preprocessing are performed on each signal sample in the training data set to obtain a set of standardized feature matrices; The feature matrix set is used to iteratively train a pre-set initial model, and the model parameters are adjusted through an optimization algorithm; When the training process meets the pre-set stopping condition, the training is stopped, and a trained underwater acoustic signal classification neural network model is obtained; wherein the underwater acoustic signal classification neural network model is configured to receive a feature matrix of an underwater acoustic signal and output a probability matrix.

7. The method of claim 6, wherein, The underwater acoustic signal classification neural network model includes: An input layer configured to receive an input matrix of 64x64 size; A first convolutional layer configured with 32 3x3 convolutional kernels; A first pooling layer configured to perform 2x2 max pooling; A second convolutional layer configured with 64 3x3 convolutional kernels; A second pooling layer configured to perform 2x2 max pooling; An L2 regularization layer for applying constraints to model parameters; A Dropout layer configured to randomly drop neurons with a probability of 0.3 during training; A global pooling layer; A fully connected output layer configured with two neurons and using a Softmax activation function.

8. A neural network-based underwater acoustic communication concealment performance evaluation device, characterized in that, It includes: A feature extraction module for extracting features from a to-be-evaluated underwater acoustic signal to obtain a feature sequence of the to-be-evaluated underwater acoustic signal; A conversion module for converting the feature sequence into a two-dimensional matrix of a pre-set size; A prediction module for predicting the two-dimensional matrix according to a pre-trained underwater acoustic signal classification neural network model to obtain a probability matrix, the probability matrix including a first probability and a second probability, the first probability representing the probability that the to-be-evaluated underwater acoustic signal belongs to a biomimetic modulated signal, and the second probability representing the probability that the to-be-evaluated underwater acoustic signal belongs to a real marine biological signal, the underwater acoustic signal classification neural network model including an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a global pooling layer, and a fully connected output layer connected in sequence.

9. An electronic device, comprising: It includes: At least one processor, at least one memory, and computer program instructions stored in the memory, which, when executed by the processor, implement the method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, When the computer program instructions are executed by the processor, the method of any one of claims 1-7 is implemented.

Images