Active sonar anti-reverberation method and system based on neural network

Abstract

The present application provides an active sonar anti-reverberation method and system based on a neural network. The method uses the active sonar emission pulse width as the window length, and uses a trained convolutional neural network to perform sliding window filtering on the echo signal to obtain a result graph; according to the set threshold value, the result graph is threshold judged to obtain a filtered result graph. The training process of the convolutional neural network includes: collecting a large amount of reverberation and real target data in active sonar detection, and pre-processing the data such as beamforming; performing Fourier transform on the target signal, extracting auditory perception features, and training the convolutional neural network using the pre-processed data and the extracted auditory perception features; evaluating the training effect according to the indicators of accuracy, recall rate and false alarm rate. The advantage of the present application is that: based on the theory of human ear recognition, an anti-reverberation algorithm that uses the auditory perception features of the signal and its changes is proposed, which can bring more reliable detection effects for active sonar detection.

Description

Technical Field

[0001] The present application belongs to the field of active sonar signal processing, and specifically relates to an active sonar anti-reverberation method and system based on a neural network. Background Art

[0002] Reverberation signals are different from ocean noise. They have strong coherence with signals and are colored, non-stationary noise. Traditional matched filters can only achieve optimal detection performance in the case of Gaussian noise and cannot cope with complex reverberation environments. In order to solve the problems of high false alarm rate and low target detection rate caused by reverberation signals in active sonar detection and recognition processing, domestic and foreign scholars have made extensive attempts in array design, transmission signal design and signal post-processing. Among them, the array and transmission signal design are generally costly and complex, and effective signal processing is the key to solving the problem.

[0003] Traditional signal processing methods including pre-whitening and subspace methods are limited by prior knowledge and conditional assumptions and are difficult to adapt to complex ocean environments. Summary of the invention

[0004] The purpose of this application is to overcome the shortcomings of traditional signal processing methods that are limited by prior knowledge and conditional assumptions and are difficult to adapt to complex marine environments.

[0005] In order to achieve the above objectives, the present application proposes an active sonar anti-reverberation method based on a neural network, comprising:

[0006] Taking the pulse width of active sonar transmission as the window length, the trained convolutional neural network is used to perform sliding window filtering on the echo signal to obtain the result image; according to the set threshold value, the result image is threshold judged to obtain the filtered result image;

[0007] The training process of the convolutional neural network includes:

[0008] Step 101: Collect active sonar detection data, and organize and pre-process the data;

[0009] Step 102: Perform sliding window segmentation, Fourier transform and filtering on the data to extract auditory perception features of the data and use them as input features of the convolutional neural network;

[0010] Step 103: Divide the data into three parts: a training set, a validation set, and a test set, and train a convolutional neural network using the preprocessed data and the extracted auditory perception features;

[0011] Step 104: Evaluate the training effect according to the indicators of accuracy, recall rate and false alarm rate to obtain the optimal convolutional neural network.

[0012] As an improvement of the above method, in step 101, the data preprocessing method is:

[0013] The non-directional array data is converted into a directional beam time signal through the beamforming method, and the interference signal outside the working frequency is filtered out by a bandpass filter, and the data is normalized using the zero mean normalization method:

[0014] s=(x-μ) / σ

[0015] Among them, s is the preprocessed data; x is the sample; μ is the sample mean; σ is the sample variance.

[0016] As an improvement of the above method, in step 102, the sliding window division method for data is:

[0017] During training, the duration of the input training data is the same as the pulse width of the transmitted signal; taking the point to be measured as the starting point, 1 / 5 of the pulse width of the transmitted signal as the window length, 1 / 20 of the pulse width as the step length, the data after beamforming is intercepted frame by frame as training data.

[0018] As an improvement of the above method, in step 102, the data is filtered as follows:

[0019] According to the frequency band range of the transmitted signal, the corresponding input signal spectrum is selected, the input signal starting point is set as the spectrum change starting point, and the selected frequency band interval is changed according to the change of the broadband FM signal spectrum over time.

[0020] As an improvement of the above method, in step 102, the auditory perception features of the extracted data include: spectral centroid, spectral centroid bandwidth, spectral decline rate, spectral irregularity, spectral flatness, noisiness and Mel frequency cepstrum coefficients.

[0021] As an improvement of the above method, the loss function when training the convolutional neural network in step 103 is a cross entropy loss function, and the calculation formula is as follows:

[0022] loss(y,T)=-[Tlog(y)+(1-T)log(1-y)]

[0023] Among them, y is the predicted output probability and T is the true label.

[0024] As an improvement to the above method, the hyperparameters for training the convolutional neural network in step 103 are set as follows:

[0025] Batchsize is set to 32, epoch is set to 100, the initial learning rate is 0.001; it decreases by 50% every 20 epochs, and Adam is used as the gradient updater.

[0026] As an improvement of the above method, the sliding window filtering of the echo signal includes:

[0027] When sliding window filtering is performed, the window length is set to the pulse width τ of the transmitted signal, and the step length is set to τ / 100.

[0028] As an improvement to the above method, the set threshold value is selected according to different application scenarios;

[0029] When the recall rate is guaranteed, the threshold λ=min(Ω(x1)), Ω is the optimal convolutional neural network, and x1 is the target echo input;

[0030] When ensuring a low false alarm rate, the threshold λ=max(Ω(x2)), which is the non-target echo input;

[0031] The threshold determination method is as follows:

[0032]

[0033] Among them, θ represents the angular direction of the signal to be processed; t represents the delay distance; S1(θ, t) is the output result after processing by the convolutional neural network; S2(θ, t) is the output result after threshold filtering.

[0034] The present application also provides an active sonar anti-reverberation system based on a neural network, which is implemented based on the above method, and the system includes:

[0035] The anti-reverberation module is used to use the active sonar emission pulse width as the window length, use the trained convolutional neural network to perform sliding window filtering on the echo signal to obtain a result image, and perform threshold judgment on the result image according to the set threshold value to obtain a filtered result image;

[0036] Training module, used to train convolutional neural networks.

[0037] Compared with the prior art, the advantages of this application are:

[0038] The method proposed in this paper uses Fourier transform to extract the signal spectrum, and trains a convolutional neural network based on the extracted auditory perception features to build a target signal recognition model. The model outputs the probability value of the true echo as the filtering result for different input signals, which effectively improves the dereverberation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Shown is a data processing block diagram of an active sonar anti-reverberation method based on a neural network;

[0040] Figure 2 Shown is a schematic diagram of the model for convolutional neural network training;

[0041] Figure 3 Shown is a distribution plot of target and reverberation output results. DETAILED DESCRIPTION

[0042] The technical solution of the present application is described in detail below with reference to the accompanying drawings.

[0043] Since deep learning has strong adaptive, autonomous learning, and nonlinear approximation capabilities, it can abstract different scale information of original data into effective feature expressions, and has a wide range of applications in complex data processing and analysis. Among them, convolutional neural network CNN is usually used in the field of computer vision. Its main feature is to extract features of images or other two-dimensional data through convolution operations, thereby realizing the processing of tasks such as image classification, target detection, and semantic segmentation.

[0044] In the task of active sonar target detection and recognition, the target and reverberation signals have significant differences in auditory perception characteristics. Therefore, we select a convolutional neural network model to build an active sonar dereverberation model. The model outputs the probability value of the real echo for different input signals as the filtering result.

[0045] For broadband signals, their frequency composition and the strength and change characteristics of each frequency component constitute the auditory perception characteristics of the underwater acoustic signal. The frequency distribution and energy distribution of the reverberation and noise signal are significantly different from those of the transmitted signal. By quantizing the extracted auditory features and combining them with a convolutional neural network, the target signal and the reverberation can be effectively separated. Based on the theory of human ear recognition, the present invention proposes an active sonar anti-reverberation method and system based on a neural network, which is an anti-reverberation method that utilizes the auditory perception characteristics of the signal and its changes. Verified by experimental data, the reverberation elimination is achieved well.

[0046] Example 1

[0047] refer to Figure 1 The processing flow of the active sonar anti-reverberation method based on neural network provided by the present invention is as follows:

[0048] Step 101: Collect a large amount of active sonar detection data, and organize and pre-process the data.

[0049] The data preprocessing method is: the non-directional array data is used to generate a directional beam time signal through the beamforming method, and the interference signal outside the working frequency is filtered out by the bandpass filter. The data is normalized using the zero mean normalization method:

[0050] s=(x-μ) / σ

[0051] Among them, x is the sample, μ is the sample mean, and σ is the sample variance.

[0052] Step 102: Sliding window segmentation is performed on the signal. During training, the duration of the input training data is the same as the pulse width of the transmitted signal. Taking the point to be measured as the starting point, 1 / 4 of the pulse width of the transmitted signal is the window length, 1 / 20 of the pulse width is the step length, and the beamformed data is intercepted frame by frame to obtain 16 frames of signals as training data.

[0053] Perform Fourier transform and filtering on the input signal, select the corresponding input signal spectrum according to the frequency band range of the transmitted signal, set the input signal starting point as the spectrum change starting point, and change the selected frequency band interval according to the change of the broadband FM signal spectrum over time.

[0054] Extract spectral centroid SC:

[0055]

[0056] Where f is the signal frequency, [f min , f max ] is the frequency band range, and X(f) is the signal spectrum.

[0057] In the field of music, the spectral centroid is one of the important physical parameters that describe the properties of timbre. It is important information about the frequency distribution and energy distribution of sound signals. In the field of subjective perception, the spectral centroid describes the brightness of the sound. The noise signal with a "dark" timbre characteristic that the sonar operator listens to usually has a relatively low spectral centroid, while the noise signal with a "bright" timbre characteristic usually has a higher spectral centroid. The spectral centroid is the center of gravity of the frequency component of the sound signal. It is the frequency weighted by energy within a certain frequency range, indicating the characteristics that change with intensity. Its unit is Hz.

[0058] Extract the spectrum centroid bandwidth SBW:

[0059]

[0060] Among them, SC low for [f min , SC] spectral centroid within the frequency band, SC high is [SC,f max ] spectral centroid within the frequency band.

[0061] Spectral centroid bandwidth refers to the frequency band width where energy is concentrated, that is, the difference between the spectral centroid above the SC band and the spectral centroid below the SC band. Spectral centroid bandwidth can be understood as the amplitude-weighted average of the difference between the frequency band component and the corresponding centroid. Relative to the spectral centroid, it reflects the concentrated area of ​​sound energy, and the unit is Hz.

[0062] Extract the spectrum drop rate SR:

[0063]

[0064] Among them, P F is the maximum value of the spectrum, P nF is the spectrum amplitude when it is n octave.

[0065] The spectrum drop rate describes the rate of spectrum drop. It is generally believed that the ship radiated noise drops at 6dB / octave in the high frequency band, but in fact there are differences for different noise sources. The main reason is the different degree of cavitation. For example, the spectrum drop rates of merchant ships sailing on the surface, submarines sailing underwater, and torpedoes are significantly different.

[0066] Extract spectral irregularities SI:

[0067]

[0068] Spectral irregularity describes the shape of the spectral envelope and is a coefficient of how much the amplitudes of adjacent partials of a polyphonic tone differ in the spectrum. Thus, large amplitude differences produce notched envelopes, while smaller differences produce smoother envelopes.

[0069] Extract spectral flatness SFM:

[0070]

[0071] Spectral flatness is defined as the ratio of the geometric mean (Gm) of the spectrum to its arithmetic mean (Am), which is used to describe the flatness of the spectrum in dB. When SFM is close to 0, it means that the signal is more like a sine curve, and when SFM is close to 1, it means that the signal is flatter and less correlated.

[0072] Extract signal noise:

[0073]

[0074] Among them, X2(f) is the line spectrum within the frequency band, and X1(f) is the continuous spectrum within the frequency band.

[0075] Noise level is also called annoyance level. It is a psychoacoustic parameter that describes the degree of noise disturbance of a sound. In the process of listening recognition, when the loudness is equivalent, the degree of noise disturbance of different target noises is different. Generally speaking, the noise of a torpedo target is fine, the most musical, and the noise level is small; the sound of a submarine sailing with an underwater motor is clear, and the sense of disorder of its noise is often lower than that of a surface ship; various surface ships often sound more chaotic due to their strong mechanical noise and cavitation noise. It can be seen that different target noise levels have a certain degree of separability. The main physical factors that affect the perception of noise level include the energy level of spectral components, the complexity of the spectrum and the pure tone components, the frequency and amplitude of amplitude modulation, and the rise time of the pulse sound.

[0076] Extract the Mel frequency cepstrum coefficients and determine the center frequency range of the filter bank [fmin , f max ], divide it into 10 frequency spaces, and use the Mel frequency scale conversion formula to convert the center frequency range into the corresponding Mel frequency value:

[0077] Mel(f)=2595lg(1+f / 700)

[0078] Mel frequency cepstral coefficients simulate the human ear's perception characteristics of speech at different frequencies. The human ear has different perception capabilities for speech at different frequencies. Experiments have found that below 1000Hz, the perception capability is linearly related to the frequency, while above 1000Hz, the perception capability is logarithmically related to the frequency.

[0079] The frequency values ​​on the Mel frequency scale are converted back to the linear frequency scale using the inverse conversion formula to obtain the filter center frequency on the linear frequency scale. For each filter center frequency, a bandpass filter group with a triangular filter response is created. For each filter, the spectrum is multiplied by the filter response to obtain the filter output. All filter outputs are further logarithmically operated, and then discrete cosine transform (DCT) is further performed to obtain the Mel frequency cepstrum coefficient (MFCC).

[0080] The auditory perception features are concatenated to obtain a 16×16 two-dimensional matrix of auditory perception features for the input signal. The original input signal is divided into 16 segments of signals to be processed by sliding windows. For each segment of the signal, the first six features are extracted to obtain a 1×6-dimensional feature vector, and then the 10-dimensional MFCC is extracted to obtain a 1×10-dimensional feature vector. After concatenation, a 1×16-dimensional vector is obtained. The feature vectors of each segment of the signal are concatenated to obtain a 16×16-dimensional two-dimensional matrix of auditory perception features.

[0081] Step 103: The data set is divided into three parts: training set, validation set, and test set. A convolutional neural network is trained using the sorted data and the initially extracted prior features.

[0082] Among them, the training criterion is the cross entropy loss function, and the binary classification calculation formula is as follows:

[0083] loss(y,T)=-[Tlog(y)+(1-T)log(1-y)]

[0084] Among them, y is the model prediction output probability, and T is the true label.

[0085] The overall framework of the convolutional neural network is shown in the figure Figure 2 The input layer is the segmented short-time Fourier transform spectrogram, which passes through multiple convolutional layers and pooling layers, and finally outputs the result through the fully connected layer. The training settings are as follows:

[0086] Batchsize is set to 32, epoch is set to 100, the initial learning rate is 0.001, and it is dropped by 50% every 20 epochs, and Adam is used as the gradient updater.

[0087] Step 104: Evaluate the system effect according to the indicators of accuracy, recall rate and false alarm rate, which are defined as:

[0088] Accuracy: (number of correct target recognition + number of correct non-target recognition) / (number of target samples + number of non-target samples)

[0089] Recall rate: (number of correct target recognition) / (number of target samples)

[0090] False alarm rate: (Number of non-targets identified as targets) / (Number of targets identified correctly + Number of non-targets identified as targets)

[0091] Step 105: Select the active sonar transmission pulse width as the window length, and use the optimal model to perform sliding window filtering on the echo signal with a step length of τ / 100. The decision threshold λ needs to be selected according to different application scenarios. When ensuring the recall rate, λ=min(Ω(x1)), where Ω is the optimal classification model and x1 is the target echo input; when ensuring a low false alarm rate, λ=max(Ω(x2)), where Ω is the optimal classification model and x2 is the non-target echo input. The specific threshold decision method is as follows:

[0092]

[0093] Among them, for the signal S to be processed, the horizontal axis is θ, representing the angle direction, and the vertical axis is t, representing the delay distance; S1(θ, t) is the output result after model processing, and S2(θ, t) is the output result after threshold filtering.

[0094] The performance of the neural network-based S-spectrum enhanced anti-reverberation algorithm in improving the target signal-to-mixture ratio (SRR) and distinguishing between the target and the reverberation is analyzed by combining actual data with simulation data.

[0095] Example 2

[0096] The present application also provides an active sonar anti-reverberation system based on a neural network, which is implemented based on the above method, and the system includes:

[0097] The anti-reverberation module is used to use the active sonar emission pulse width as the window length, use the trained convolutional neural network to perform sliding window filtering on the echo signal to obtain a result image, and perform threshold judgment on the result image according to the set threshold value to obtain a filtered result image;

[0098] Training module, used to train convolutional neural networks.

[0099] Based on the active sonar echo data obtained in a real near-shore strong reverberation environment, this application generates active sonar target echo data under different conditions through the following simulation settings: (1) Simulate the shallow sea sound velocity gradient distribution map and the sound line propagation path to generate the multipath pulse response of the target at different distances; (2) Simulate the scattering function form of the sound wave at different incident angles of the target; (3) Randomly set a group of target movement speeds within a reasonable range, and generate the Doppler deformation signal of the original signal at the corresponding speed, and convolve the signal with the scattering function and the multipath pulse response and add a certain amount of random noise to obtain the simulated target echo signal; (4) Add ocean reverberation levels of different intensities to the array data to ensure that the energy of the reverberation and the target in the matched filtering results is equivalent. This scheme focuses on the proposed scheme's ability to distinguish reverberation when the reverberation level is equivalent to the target echo. After all the points detected in the matched filter map are passed through the proposed anti-reverberation algorithm, the distribution diagram of the target and reverberation output results is as follows: Figure 3 When the signal-to-mix ratio is 0dB, the output results of the target and reverberation after passing through the S filter algorithm are close to normal distribution. If the probability density of the two is equal, the method of this application can achieve a 95% accuracy rate in distinguishing the target from the reverberation while the recall of the target reaches 93%.

[0100] The present application may also provide a computer device, comprising: at least one processor, a memory, at least one network interface and a user interface. The various components in the device are coupled together through a bus system. It is understood that the bus system is used to achieve connection and communication between these components. In addition to the data bus, the bus system also includes a power bus, a control bus and a status signal bus.

[0101] The user interface may include a display, a keyboard or a pointing device, such as a mouse, a trackball, a touch pad or a touch screen.

[0102] It is understood that the memory in the embodiments disclosed in the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memories.

[0103] In some embodiments, the memory stores the following elements, executable modules or data structures, or a subset thereof, or an extended set thereof: an operating system and applications.

[0104] The operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., which are used to implement various basic services and process hardware-based tasks. The application includes various application programs, such as a media player (Media Player), a browser (Browser), etc., which are used to implement various application services. The program for implementing the method of the embodiment of the present disclosure can be included in the application.

[0105] In the above embodiment, the processor may also call a program or instruction stored in the memory, specifically, a program or instruction stored in an application program, and is used to:

[0106] Execute the steps of the above method.

[0107] The above method can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor may be a general processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The above-disclosed methods, steps and logic block diagrams can be implemented or executed. The general processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the above-disclosed method can be directly embodied as a hardware decoding processor to execute, or the hardware and software modules in the decoding processor are combined to execute. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0108] It is understood that the embodiments described in the present application can be implemented by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in the present application or a combination thereof.

[0109] For software implementation, the technology of the present application can be implemented by executing the functional modules (such as procedures, functions, etc.) of the present application. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0110] The present application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, each step in the above method embodiment can be implemented.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present application and are not intended to limit it. Although the present application is described in detail with reference to the embodiments, a person skilled in the art should understand that any modification or equivalent replacement of the technical solution of the present application does not depart from the spirit and scope of the technical solution of the present application and should be included in the scope of the claims of the present application.

Claims

1. An active sonar anti-reverberation method based on a neural network, comprising: Taking the active sonar transmission pulse width as the window length, the trained convolutional neural network is used to perform sliding window filtering on the echo signal to obtain the result graph; According to the set threshold value, the result graph is judged by threshold value to obtain the filtered result graph; The training process of the convolutional neural network includes: Step 101: Collect active sonar detection data, and organize and pre-process the data; Step 102: Perform sliding window segmentation, Fourier transform and filtering on the data to extract auditory perception features of the data and use them as input features of the convolutional neural network; Step 103: Divide the data into three parts: a training set, a validation set, and a test set, and train a convolutional neural network using the preprocessed data and the extracted auditory perception features; Step 104: Evaluate the training effect according to the indicators of accuracy, recall rate and false alarm rate to obtain the optimal convolutional neural network; In step 101, the data preprocessing method is: The non-directional array data is converted into a directional beam time signal through the beamforming method, and the interference signal outside the working frequency is filtered out by a bandpass filter, and the data is normalized using the zero mean normalization method: s=(x-μ) / σ Among them, s is the preprocessed data; x is the sample; μ is the sample mean; σ is the sample variance.

2. The active sonar anti-reverberation method based on neural network according to claim 1 is characterized in that: In step 102, the method of sliding window partitioning the data is as follows: During training, the duration of the input training data is the same as the pulse width of the transmitted signal; taking the point to be measured as the starting point, 1 / 5 of the pulse width of the transmitted signal as the window length, 1 / 20 of the pulse width as the step length, the data after beamforming is intercepted frame by frame as training data.

3. The active sonar anti-reverberation method based on neural network according to claim 1, characterized in that: In step 102, the data is filtered as follows: According to the frequency band range of the transmitted signal, the corresponding input signal spectrum is selected, the input signal starting point is set as the spectrum change starting point, and the selected frequency band interval is changed according to the change of the broadband FM signal spectrum over time.

4. The active sonar anti-reverberation method based on neural network according to claim 1, characterized in that: In step 102, the auditory perception features of the extracted data include: spectral centroid, spectral centroid bandwidth, spectral drop rate, spectral irregularity, spectral flatness, noisiness and Mel frequency cepstrum coefficients.

5. The active sonar anti-reverberation method based on neural network according to claim 1, characterized in that: The loss function when training the convolutional neural network in step 103 is a cross entropy loss function, and the calculation formula is as follows: loss(y,T)=-[Tlog(y)+(1-T)log(1-y)] Among them, y is the predicted output probability and T is the true label.

6. The active sonar anti-reverberation method based on neural network according to claim 1, characterized in that: The hyperparameters for training the convolutional neural network in step 103 are set as follows: Batchsize is set to 32, epoch is set to 100, the initial learning rate is 0.001; it decreases by 50% every 20 epochs, and Adam is used as the gradient updater.

7. The active sonar anti-reverberation method based on neural network according to claim 1, characterized in that: The sliding window filtering of the echo signal comprises: When sliding window filtering is performed, the window length is set to the pulse width τ of the transmitted signal, and the step length is set to τ / 100.

8. The active sonar anti-reverberation method based on neural network according to claim 1, characterized in that: The set threshold value is selected according to different application scenarios; When the recall rate is guaranteed, the threshold λ=min(Ω(x1)), Ω is the optimal convolutional neural network, and x1 is the target echo input; When ensuring a low false alarm rate, the threshold λ=max(Ω(x2)), where x2 is the non-target echo input; The threshold determination method is as follows: Among them, θ represents the angular direction of the signal to be processed; t represents the delay distance; S1(θ, t) is the output result after processing by the convolutional neural network; S2(θ, t) is the output result after threshold filtering.

9. An active sonar anti-reverberation system based on a neural network, implemented based on the method described in any one of claims 1 to 8, characterized in that: The system comprises: The anti-reverberation module is used to use the active sonar transmission pulse width as the window length, use the trained convolutional neural network to perform sliding window filtering on the echo signal to obtain a result image, and perform threshold judgment on the result image according to the set threshold value to obtain a filtered result image; and Training module, used to train convolutional neural networks.

Images