Method and system for identifying tick-like underwater acoustic bionic communication signal in complex marine environment

Abstract

The invention relates to the technical field of underwater acoustic signal processing, in particular to a tick-like underwater acoustic bionic communication signal identification method and system in a complex marine environment, and the method comprises the steps: constructing a training set containing real cetella tick signals and tick-like underwater acoustic bionic communication signals; dividing each signal audio sample in the training set into fixed frame sub-signal samples; seven types of features of each frame of sub-signals are extracted, fusion feature vectors are constructed, and then the fusion feature vectors of all the sub-signals form a fusion feature image; building an underwater sound bionic communication signal identification model, and performing model optimization by using the training set; and dividing a to-be-identified signal sample into fixed frame sub-signal samples to obtain a fused feature image, and inputting the fused feature image into the trained underwater acoustic bionic communication signal identification model to obtain an identification result. According to the invention, accurate identification of the tick-like underwater acoustic bionic communication signal can be realized in a small sample length and complex underwater acoustic channel environment.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic signal processing, and in particular to a method and system for identifying ticking-like underwater acoustic bionic communication signals in a complex marine environment. Background Art

[0002] The ticking underwater acoustic bionic communication technology is based on the idea of ​​bionics and uses the ticking sounds of cetaceans in the ocean or artificially synthesized ticking sounds as information carriers. During the identification process, the third party will treat the ticking underwater acoustic bionic communication signal as ocean background noise and exclude it, thus achieving covert communication. This method greatly improves the concealment of underwater acoustic communication and has broad application prospects.

[0003] While underwater bionic acoustic communication technology has attracted much attention, the recognition of underwater bionic acoustic communication signals also occupies an important position in the field of underwater electronic countermeasures. Effective signal recognition methods can bring significant advantages. However, the research on the recognition of biomimetic camouflage click communication train (BCCCT) is still in its infancy. The existing methods are based on the statistical characteristics of the delay difference between adjacent clicks. First, the delay difference interval between adjacent clicks is obtained by accurately locating a single click in the time domain. Then, the signal is recognized by observing whether there is a step distribution statistical characteristic or inputting the delay difference distribution curve into the convolutional neural network. The sample length requirement is high and it has limitations under complex underwater acoustic channel conditions. In the actual marine environment, the underwater acoustic channel is affected by inhomogeneous media, sound reflections from the sea surface and seabed, and the absorption of sound energy by seawater. It has the characteristics of significant multipath effects, large environmental noise, and temporal and spatial frequency variations. It is one of the most complex channels in the field of wireless communications. At the same time, the real click train (RCT) emitted by cetaceans generally lasts for a short time. If BCCCT wants to maintain a high level of concealment performance, it must also control the duration of the signal within a shorter range. Therefore, how to accurately identify small sample, short burst BCCCT in a complex underwater acoustic channel environment has become a practical problem that needs to be solved urgently. Summary of the invention

[0004] The present invention aims to solve the problems existing in the prior art and proposes a method and system for identifying click-like underwater acoustic bionic communication signals in a complex ocean environment, which can realize accurate identification of BCCCT in a small sample length and complex underwater acoustic channel environment.

[0005] To achieve the above purpose, the technical solution adopted is:

[0006] A method for identifying ticking underwater acoustic bionic communication signals in a complex marine environment, comprising:

[0007] Step 1, constructing a training set including real cetacean ticking signals and imitation ticking underwater acoustic bionic communication signals, and dividing each signal audio sample in the training set into fixed frame sub-signal samples;

[0008] Step 2, extract the zero-crossing rate, spectrum centroid, spectrum extension, Mel-frequency cepstrum coefficient, harmonic ratio and fundamental frequency, spectrum entropy, and spectrum roll-off point of each frame sub-signal, construct a fused feature vector, and then combine the fused feature vectors of all sub-signals into a fused feature image;

[0009] Step 3: Build an underwater acoustic bionic communication signal recognition model and use the training set to optimize the model;

[0010] Step 4: After dividing the signal sample to be identified into fixed-frame sub-signal samples, use step 2 to obtain the fused feature image, and input it into the underwater acoustic bionic communication signal recognition model trained in step 3 to obtain the recognition result.

[0011] According to the method for identifying ticking-like underwater acoustic bionic communication signals in a complex marine environment of the present invention, the training set is further expanded: according to transfer learning, the training set composed of real cetacean ticking signals and ticking-like underwater acoustic bionic communication signals is passed through a simulation channel generated by a Bellhop model, and noise of different intensities is added to obtain an expanded training set.

[0012] According to the method for identifying bionic communication signals imitating ticking sound in a complex marine environment of the present invention, further, step 2 extracting seven types of features of sub-signals specifically includes:

[0013] Extract the zero-crossing rate of the sub-signal, which represents the number of zero-crossing times per unit time of the signal's time domain waveform;

[0014] Extract the spectrum centroid of the sub-signal. The spectrum centroid represents the weighted average of the frequency to the signal spectrum and is the center of gravity of the frequency component.

[0015] Calculate the spectrum extension of the sub-signal, which reflects the width or extension of the signal spectrum, that is, the dispersion degree of the signal energy in the frequency domain;

[0016] To calculate the Mel frequency cepstrum coefficients of the sub-signal, first calculate the energy spectrum of each frame of the sub-signal; then calculate the Mel energy spectrum E of the sub-signal based on the energy spectrum n '(m); finally to E n '(m) Perform discrete cosine transform to obtain Mel frequency cepstrum coefficients;

[0017] To calculate the harmonic ratio and fundamental frequency of the sub-signal, first calculate the signal autocorrelation value R n (k), the autocorrelation value R n(k) After smoothing filtering, find the frequency corresponding to the smallest maximum point as the fundamental frequency, and then determine the harmonic ratio by the ratio of other maximum points to the fundamental frequency;

[0018] Calculate the spectral entropy of the sub-signal, which describes the distribution of the signal power spectrum in each frequency band;

[0019] Calculate the spectrum roll-off point of the sub-signal and accumulate the spectrum energy according to the frequency scale. When the accumulated energy is just greater than the specified proportion of the total energy, the corresponding frequency is the spectrum roll-off point.

[0020] According to the method for identifying underwater acoustic bionic communication signals simulating ticking sounds in a complex marine environment of the present invention, further, the underwater acoustic bionic communication signal recognition model includes 4 convolutional layers using Relu activation function, 2 fully connected layers, 4 maximum pooling layers, 3 Dropout layers and a flattening layer; the two-dimensional feature image is input into the convolutional layer to extract the high-dimensional features of the image, and then the spatial dimension is reduced through the maximum pooling layer, and then the features are further extracted through the convolutional layer, 50% of the connections are randomly discarded through the Dropout layer, and the spatial dimension is further reduced through the maximum pooling layer, and the features are further extracted through the convolutional layer, 50% of the connections are randomly discarded through the Dropout layer, and the spatial dimension is further reduced through the maximum pooling layer, and after passing through the convolutional layer, the feature map size is reduced for the last time through the maximum pooling layer; the multidimensional feature map is converted into a one-dimensional vector through the flattening layer; and then the recognition result is output through the Dropout layer that randomly discards 50% of the connections and two fully connected layers.

[0021] Furthermore, the present invention also proposes a bionic communication signal recognition system for imitating ticking sound in a complex marine environment, the system comprising a training set construction module, a fusion feature image acquisition module, a model building module and a recognition result module, wherein:

[0022] A training set construction module is used to construct a training set including real cetacean click signals and imitation click water sound bionic communication signals, and divide each signal audio sample in the training set into fixed frame sub-signal samples;

[0023] The fusion feature image acquisition module is used to extract seven types of features of each frame of sub-signal, including zero-crossing rate, spectrum centroid, spectrum extension, Mel-frequency cepstrum coefficient, harmonic ratio and fundamental frequency, spectrum entropy, and spectrum roll-off point, construct a fusion feature vector, and then combine the fusion feature vectors of all sub-signals into a fusion feature image;

[0024] Model building module, used to build a water acoustic bionic communication signal recognition model and use the training set to optimize the model;

[0025] The recognition result module is used to divide the signal sample to be recognized into fixed-frame sub-signal samples, use the fused feature image acquisition module to obtain the fused feature image, and input it into the trained underwater acoustic bionic communication signal recognition model to obtain the recognition result.

[0026] The beneficial effects achieved by adopting the above technical solution are:

[0027] 1. Existing BCCCT recognition methods are all based on the statistical characteristics of the time delay difference interval between adjacent ticks. The statistical characteristics can only be manifested when the signal reaches a certain sample length. This method is based on signal feature extraction and has a lower requirement for sample length.

[0028] 2. The present invention enhances the robustness of feature extraction by extracting seven types of features commonly used in speech recognition and fusion signal time-frequency domain, so that the fusion feature image contains rich semantic information, which helps to improve the accuracy of signal recognition.

[0029] 3. To address the problem of scarce samples, a data set expansion method based on transfer learning is proposed. The original training set contains the ticking signals of cetaceans in a real ocean environment and the simulated ticking underwater acoustic bionic communication signals. By adding typical simulated channels and noise in different scenarios to expand the training set data, the underwater acoustic bionic communication signal recognition model is trained, which provides rich data support for model training and enables the trained model to adapt to complex underwater acoustic channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0031] Figure 1 It is a flow chart of a method for identifying a bionic communication signal simulating a ticking sound under a complex ocean environment according to an embodiment of the present invention;

[0032] Figure 2 is a flow chart of Mel-frequency cepstral coefficient extraction according to an embodiment of the present invention;

[0033] Figure 3 is a schematic diagram of a Mel filter according to an embodiment of the present invention;

[0034] Figure 4 is a schematic diagram of a fused feature image structure according to an embodiment of the present invention;

[0035] Figure 5 is a structural diagram of an underwater acoustic bionic communication signal recognition model according to an embodiment of the present invention;

[0036] Figure 6It is the recognition accuracy of the signal after passing through the channel h11 by the three methods in the embodiments of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings of specific embodiments of the present invention to clearly and completely describe the exemplary scheme of the embodiment of the present invention. Unless otherwise defined, the technical terms or scientific terms used in the present invention should be the common meanings understood by people with ordinary skills in the field.

[0038] This embodiment discloses a method for identifying a bionic communication signal imitating a ticking sound in a complex marine environment. Figure 1 As shown, the following steps are included:

[0039] Step S101, construct a training set including real cetacean click signals and simulated click water acoustic bionic communication signals, and according to transfer learning, pass the original training set through the simulation channel generated by the Bellhop model and add noise of different intensities to obtain an expanded training set.

[0040] In actual scenarios, in order to ensure high recognition rate and low false alarm rate, it is usually hoped that the imitation ticking underwater acoustic bionic communication signal to be identified has similar distribution characteristics to the training set signal. This is also a basic assumption of many existing machine learning and deep learning methods. But in fact, in practical applications, this assumption may not hold true. The imitation ticking underwater acoustic bionic communication signal to be identified often comes from different sea areas from the training set of the model, has completely unrelated channel characteristics, and is generated based on completely different ticking signal simulation. To solve the above problems, the simulation channels h1~h6 are used as training channels, and the original training set is passed through 6 simulation channels respectively, and noise of different intensities is added, finally generating a sufficient number of expanded training sets to train the underwater acoustic bionic communication signal recognition model.

[0041] Step S102: Divide each signal audio sample in the expanded training set into 81 frame sub-signal samples, with a frame length and frame shift ratio of 25:6, to obtain 81 frame sub-signals.

[0042] Step S103, extracting seven types of features of each frame sub-signal, including zero-crossing rate, spectrum centroid, spectrum stretch, Mel-frequency cepstrum coefficient, harmonic ratio and fundamental frequency, spectrum entropy, and spectrum roll-off point.

[0043] ① Extract the zero-crossing rate of the sub-signal. The zero-crossing rate represents the number of zero-crossing times per unit time of the signal’s time domain waveform. The calculation formula is shown in formula (1).

[0044]

[0045] Among them, Z(n) represents the zero-crossing rate characteristic of the n-th frame sub-signal, sgn[*] represents the sign operation, and x n(m) represents the n-th frame sub-signal.

[0046] ② Extract the spectral centroid of the sub-signal. The spectral centroid represents the weighted average of the frequency to the signal spectrum and is the center of gravity of the frequency component. First, calculate the discrete Fourier transform of the sub-signal X = [X1, X2, ..., X N ] = DFT[x n (m)] = FFT[x n (m)], where FFT[*] represents fast Fourier transform calculation.

[0047] Then the spectrum centroid calculation formula is as follows.

[0048]

[0049] Among them, c n represents the spectral centroid of the n-th frame sub-signal, Indicates the center frequency corresponding to the nth frequency band, f s is the signal sampling rate.

[0050] ③ Calculate the spectrum stretch of the sub-signal. The spectrum stretch reflects the width or extension of the signal spectrum, that is, the dispersion of the signal energy in the frequency domain. The calculation formula is shown in formula (3).

[0051]

[0052] Among them, s n Indicates the spectrum extension of the n-th frame sub-signal.

[0053] ④ Calculate the Mel Frequency Cepstral Coefficients (MFCC) of the sub-signal. MFCC is designed based on the research results of the human hearing characteristics. The human ear's perception of sound frequency is not linear, but presents a logarithmic relationship, that is, the perception of low-frequency sounds is more sensitive, while the perception of high-frequency sounds is relatively weak. The Mel scale is designed to simulate this nonlinear relationship, which enables MFCC to better reflect the actual perception of the human ear to sound frequency. Therefore, in signal classification and recognition, using MFCC as a feature can be more in line with human hearing characteristics, thereby improving the accuracy and robustness of recognition. Figure 2 This is the MFCC extraction process. First, calculate each frame sub-signal x n Energy spectrum E of (m) n (k), the calculation formula is shown in formula (4).

[0054] E n (k) = FFT 2 [x n(m)],0≤k≤N-1 (4)

[0055] Figure 3 This is a schematic diagram of the Mel filter, which is composed of a set of triangular filters superimposed on each other, simulating the nonlinear characteristics of the human ear being less sensitive to high-frequency sounds. The calculation formula of the Mel filter is as follows.

[0056]

[0057] Among them, H m (k) represents the frequency response of the mth triangular filter, M represents the total number of triangular filters, in this embodiment, M=30, E n '(m) represents the Mel energy spectrum of the n-th frame sub-signal after taking the logarithm.

[0058] Finally, E n '(m) is subjected to discrete cosine transform to obtain MFCC. The calculation formula is shown in formula (6).

[0059]

[0060] Wherein, MFCC(n) is the MFCC of the n-th frame sub-signal, r represents the order of MFCC, and in this embodiment, r=36.

[0061] ⑤ Calculate the harmonic ratio and fundamental frequency of the sub-signal. The harmonic frequency is an integer multiple of the fundamental frequency. The signal waveform can generally be decomposed into a sinusoidal wave component containing the fundamental frequency and a series of harmonics that are multiples of the fundamental frequency. The actual cetacean click signal is often mixed with a certain whistle signal. Compared with the imitation click water acoustic bionic communication signal, it is easier to extract the harmonic ratio and fundamental frequency characteristics. Therefore, the harmonic ratio and fundamental frequency can be used as important features to identify the two types of signals. This embodiment uses the time domain autocorrelation method to calculate the harmonic ratio and fundamental frequency of the sub-signal. First, the signal autocorrelation value R is calculated. n (k), the calculation formula is shown in formula (7).

[0062]

[0063] The autocorrelation value R n (k) After smoothing filtering, find the frequency corresponding to the smallest maximum point as the fundamental frequency, and then determine the harmonic ratio by the ratio of other maximum points to the fundamental frequency. If the harmonic ratio and fundamental frequency cannot be legally extracted, both the harmonic ratio and fundamental frequency are set to 0.

[0064] ⑥ Calculate the spectral entropy of the sub-signal. The spectral entropy describes the distribution of the signal power spectrum in each frequency band. It measures the uncertainty or complexity of the energy distribution of the signal in the frequency domain. The calculation formula is shown in formula (8).

[0065]

[0066] Among them, e n represents the spectral entropy of the sub-signal of the nth frame, B represents the number of sub-bands into which the sub-signal spectrum is divided, and in this embodiment, B is set to 10, and E b Represents the energy of the b-th subband.

[0067] ⑦ Calculate the spectrum roll-off point of the sub-signal and accumulate the spectrum energy according to the frequency scale. When the accumulated energy is just greater than the specified proportion of the total energy, the corresponding frequency is the spectrum roll-off point, which marks the boundary between the frequency range of the main energy distribution in the signal and the frequency point where the energy begins to decay rapidly. The calculation formula is shown in formula (9).

[0068]

[0069] Wherein, C is a specified ratio, generally ranging from 0.85 to 0.9, and in this embodiment, C=0.9.

[0070] Step S104: construct a fused feature vector. The seven types of features extracted in step S103 are concatenated to construct a 1×43-dimensional fused feature vector.

[0071] Step S105: Arrange the fused feature vectors of all sub-signals obtained in step S104 in rows in chronological order to form an 81×43-dimensional fused feature image. The structure of the fused feature image is as follows: Figure 4 shown.

[0072] Step S106: building an underwater acoustic bionic communication signal recognition model.

[0073] The underwater acoustic bionic communication signal recognition model uses a convolutional neural network (CNN) to recognize the fused feature image. The model structure is as follows: Figure 5As shown. The model contains 4 convolutional layers using Relu activation function, 2 fully connected layers, 4 maximum pooling layers, 3 Dropout layers and a flattening layer. It extracts high-dimensional features from the signal feature image through convolution, pooling, random inactivation and other operations, and then inputs them into the Softmax classifier for classification. The input of the model is a 81×43 two-dimensional feature image. The first layer is a 3×3 convolutional layer with 32 channels to extract high-dimensional features of the image; the second layer is a maximum pooling layer with a scale of 2×2 and a step size of 2 to reduce the spatial dimension; the third layer is a 64-channel 3×3 convolutional layer to further extract features; the fourth layer is a Dropout layer that randomly discards 50% of the connections to prevent overfitting and enhance the robustness of the network; the fifth layer is a maximum pooling layer with a scale of 2×2 and a step size of 2 to reduce the spatial dimension; the sixth layer is a 128-channel 3×3 convolutional layer to further extract features; the seventh layer is a 50% random dropout layer. Dropout layer; the eighth layer is a maximum pooling layer with a scale of 2×2 and a stride of 2; the ninth layer is a convolutional layer of size 3×3 with 256 channels; the tenth layer is a maximum pooling layer with a scale of 2×2 and a stride of 2 to reduce the size of the feature map for the last time; the eleventh layer is a flattening layer that converts the multi-dimensional feature map into a one-dimensional vector for input to the fully connected layer; the twelfth layer is a Dropout layer that randomly discards 50% of the connections; the thirteenth layer is a fully connected layer with 1024 neurons using the Relu activation function; the last layer is a fully connected layer (output layer) with 2 neurons using the Softmax activation function.

[0074] Step S107, input the training set fusion feature image obtained in step S106 into the constructed underwater acoustic bionic communication signal recognition model for training.

[0075] Step S108, repeat steps S102-S105 for the cetacean-like click signal to be identified to obtain a fused feature image, and input it into the trained underwater acoustic bionic communication signal recognition model to obtain a recognition result, which can accurately identify it as a real cetacean click signal or an imitation click underwater acoustic bionic communication signal.

[0076] Corresponding to the above-mentioned method for identifying ticking-like underwater acoustic bionic communication signals in a complex ocean environment, this embodiment also proposes a system for identifying ticking-like underwater acoustic bionic communication signals in a complex ocean environment, comprising a training set construction module, a fusion feature image acquisition module, a model building module and a recognition result module, wherein:

[0077] The training set construction module is used to construct a training set containing real cetacean click signals and imitation click water acoustic bionic communication signals, and divide each signal audio sample in the training set into fixed frame sub-signal samples.

[0078] The fusion feature image acquisition module is used to extract seven types of features of each frame sub-signal, including zero-crossing rate, spectrum centroid, spectrum extension, Mel-frequency cepstrum coefficient, harmonic ratio and fundamental frequency, spectral entropy, and spectrum roll-off point, construct a fusion feature vector, and then combine the fusion feature vectors of all sub-signals into a fusion feature image.

[0079] The model building module is used to build an underwater acoustic bionic communication signal recognition model and use the training set to optimize the model.

[0080] The recognition result module is used to divide the signal sample to be recognized into fixed-frame sub-signal samples, use the fused feature image acquisition module to obtain the fused feature image, and input it into the trained underwater acoustic bionic communication signal recognition model to obtain the recognition result.

[0081] In order to verify the effectiveness of this solution, further explanation is given below in combination with experimental data.

[0082] (1) Construction of training and testing datasets. In order to verify the effectiveness of this method, a validation dataset was constructed. The dataset includes real cetacean clicks and imitation click underwater acoustic bionic communication signals based on delay difference coding (Liu Songzuo, Liu Bingjie, Yin Yanling, et al. M-element dolphin call covert underwater acoustic communication [J]. Journal of Harbin Engineering University, 2014, 35(01): 119-125.). The real cetacean clicks are from the Watkins marine mammal sound database.

[0083] In order to simulate the actual underwater acoustic channel environment, 7 underwater acoustic channels under different transmission conditions were simulated based on the Bellhop model. Among them, h1~h6 are training channels, h 11 Table 1 gives the parameters of each channel.

[0084] Table 1 Simulation channel parameters

[0085]

[0086] Their channel system functions are:

[0087] H1(z)=1+0.6702z -3331 +0.4795z -10041 +0.4174z -14687 (10)

[0088] H2(z)=1+0.8263z -4692 +0.7759z -7375 +0.6971z -12895 (11)

[0089] H3(z)=1+0.9098z -2372+0.8515z -4980 +0.8020z -7637 (12)

[0090] H4(z)=1+0.7079z -6735 +0.6326z -12266 +0.5533z -21003 (13)

[0091] H5(z)=1+0.3599z -20041 (14)

[0092] H6(z)=1+0.5023z -9547 +0.4403z -14719 +0.3514z -28647 (15)

[0093]

[0094] The signal sampling rate is 48kHz, and the maximum propagation delays of the above channels relative to their direct paths are 0.306s, 0.269s, 0.159s, 0.438s, 0.418s, 0.597s, and 1.619s, respectively.

[0095] The training data set is generated using the above training channels. The tick-like underwater acoustic bionic communication signal is ([1:6]bit)×([-10:5:0 2.5:2.5:25]dB+no noise)×(channel[1:6]+no channel)=6×14×7=588 segments, each segment is about 55s long, and the total length is about 32340s. The real cetacean tick signal is ([-10:5:02.5:2.5:25]dB+no noise)×(channel[1:6]+no channel), the original real tick signal length is about 329s, and the total length of the real tick signal after expansion by the transfer learning model is about 32263s.

[0096] Using the above test channel h 11 Generate a test data set. The signal coding bit number of the test set is 2 bits. The signals are selected in the range of [-15,30]dB with an interval of 3dB. About 1460 samples of each signal are generated for testing at each signal-to-noise ratio, and the length of each sample is 0.7s.

[0097] (2) Each signal audio sample in the training set is fixedly divided into 81 frame sub-signal samples, with a frame length and frame shift ratio of 25:6, to obtain 81 frame sub-signals. Seven types of features are extracted, and fusion feature vectors and fusion feature images are constructed.

[0098] (3) Build Figure 5The underwater acoustic bionic communication signal recognition model shown in the figure is trained using the training set fusion feature image obtained in step (2). Based on the TensorFlow1.14 framework and Python 3.7 environment, the training of the network model relies on an NVIDIA Quadro RTX P5000 graphics card, and the training batch size N T =10, the initial learning rate l0=0.0001, and the Adam optimizer is used to update the parameters until the model converges.

[0099] (4) Extract the fused feature image of the test set signal through step (2), input it into the trained model, and obtain the recognition result. At the same time, it is compared with the method based on time delay statistics (Time Delay Statistics, TDS) (Yao Q, Jiang J, Yu X, et al. Recognition method for biomimetic camouflage communication signalimitating cetacean click in underwater multipath channels[J]. Applied Acoustics, 2024, 222: 110039.) and the method based on statistics and CNN (Statistics and CNN, S-CNN) (Jiang J, Qiao F, Li Y, et al. Recognition method for the bionic camouflage click communication trains modulated by time delay difference[J]. The Journal of the Acoustical Society of America, 2022, 152(1): 491-500.). The experimental results are as follows Figure 6 shown.

[0100] Since both TDS and S-CNN are based on delay difference statistics, they have high requirements for sample length. Under the condition of 0.7s sample length, they both show certain instability. The recognition accuracy fluctuates with the increase of signal-to-noise ratio. Moreover, TDS is not robust to underwater acoustic multipath signals, and the fluctuation is more obvious. This method (FF-CNN) is based on features and can still achieve good results under the conditions of small sample length and underwater acoustic multipath channel. The recognition accuracy increases with the improvement of signal-to-noise ratio, reaching 100% at around 0dB, and is better than the other two methods.

[0101] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0102] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0103] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0104] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0105] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for identifying water acoustic bionic communication signals simulating ticking sounds in a complex marine environment, characterized in that: Include: Step 1, constructing a training set including real cetacean ticking signals and imitation ticking underwater acoustic bionic communication signals, and dividing each signal audio sample in the training set into fixed frame sub-signal samples; Step 2, extract the zero-crossing rate, spectrum centroid, spectrum extension, Mel-frequency cepstrum coefficient, harmonic ratio and fundamental frequency, spectrum entropy, and spectrum roll-off point of each frame sub-signal, construct a fused feature vector, and then combine the fused feature vectors of all sub-signals into a fused feature image; Step 3: Build an underwater acoustic bionic communication signal recognition model and use the training set to optimize the model; Step 4: After dividing the signal sample to be identified into fixed-frame sub-signal samples, use step 2 to obtain the fused feature image, and input it into the underwater acoustic bionic communication signal recognition model trained in step 3 to obtain the recognition result.

2. The method for identifying water acoustic bionic communication signals simulating ticking sounds in a complex marine environment according to claim 1, characterized in that: Expand the training set: Based on transfer learning, the training set consisting of real cetacean click signals and simulated click underwater acoustic bionic communication signals is passed through a simulated channel generated by the Bellhop model, and noise of different intensities is added to obtain an expanded training set.

3. The method for identifying water acoustic bionic communication signals imitating ticking sound in a complex marine environment according to claim 1 is characterized in that: Step 2 extracts seven types of sub-signal features, including: Extract the zero-crossing rate of the sub-signal, which represents the number of zero-crossing times per unit time of the signal's time domain waveform; Extract the spectrum centroid of the sub-signal. The spectrum centroid represents the weighted average of the frequency to the signal spectrum and is the center of gravity of the frequency component. Calculate the spectrum extension of the sub-signal, which reflects the width or extension of the signal spectrum, that is, the dispersion degree of the signal energy in the frequency domain; To calculate the Mel frequency cepstrum coefficients of the sub-signal, first calculate the energy spectrum of each frame of the sub-signal; then calculate the Mel energy spectrum E of the sub-signal based on the energy spectrum n '(m); finally to E n '(m) Perform discrete cosine transform to obtain Mel frequency cepstrum coefficients; To calculate the harmonic ratio and fundamental frequency of the sub-signal, first calculate the signal autocorrelation value R n (k), the autocorrelation value R n (k) After smoothing filtering, find the frequency corresponding to the smallest maximum point as the fundamental frequency, and then determine the harmonic ratio by the ratio of other maximum points to the fundamental frequency; Calculate the spectral entropy of the sub-signal, which describes the distribution of the signal power spectrum in each frequency band; Calculate the spectrum roll-off point of the sub-signal and accumulate the spectrum energy according to the frequency scale. When the accumulated energy is just greater than the specified proportion of the total energy, the corresponding frequency is the spectrum roll-off point.

4. The method for identifying water acoustic bionic communication signals simulating ticking sound in a complex marine environment according to claim 1, characterized in that: The underwater acoustic bionic communication signal recognition model includes 4 convolutional layers using Relu activation function, 2 fully connected layers, 4 maximum pooling layers, 3 Dropout layers and a flattening layer; the two-dimensional feature image is input into the convolutional layer to extract the high-dimensional features of the image, and then the spatial dimension is reduced through the maximum pooling layer, and then the features are further extracted through the convolutional layer, 50% of the connections are randomly discarded through the Dropout layer, and the spatial dimension is further reduced through the maximum pooling layer, and the features are further extracted through the convolutional layer, 50% of the connections are randomly discarded through the Dropout layer, and the spatial dimension is further reduced through the maximum pooling layer. After passing through the convolutional layer, the feature map size is reduced for the last time through the maximum pooling layer; the multidimensional feature map is converted into a one-dimensional vector through the flattening layer; and then the recognition result is output through the Dropout layer that randomly discards 50% of the connections and two fully connected layers.

5. A ticking underwater acoustic bionic communication signal recognition system in a complex marine environment, characterized in that: The system includes a training set construction module, a fusion feature image acquisition module, a model building module and a recognition result module, among which: A training set construction module is used to construct a training set including real cetacean click signals and imitation click water sound bionic communication signals, and divide each signal audio sample in the training set into fixed frame sub-signal samples; The fusion feature image acquisition module is used to extract seven types of features of each frame of sub-signal, including zero-crossing rate, spectrum centroid, spectrum extension, Mel-frequency cepstrum coefficient, harmonic ratio and fundamental frequency, spectrum entropy, and spectrum roll-off point, construct a fusion feature vector, and then combine the fusion feature vectors of all sub-signals into a fusion feature image; Model building module, used to build a water acoustic bionic communication signal recognition model and use the training set to optimize the model; The recognition result module is used to divide the signal sample to be recognized into fixed-frame sub-signal samples, use the fused feature image acquisition module to obtain the fused feature image, and input it into the trained underwater acoustic bionic communication signal recognition model to obtain the recognition result.

6. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images