An automatic detection method for active sonar echoes based on segmented matched filtering

Abstract

The present invention relates to the field of active sonar detection in underwater acoustic physics, and in particular to a method for automatic detection of active sonar echoes based on segmented matched filtering. First, the method segments the transmitted signal in the time domain, and uses several segmented sub-signals to perform matched filtering on the received data respectively, extracting two-dimensional segmented matching features; then, the extracted two-dimensional feature samples are effectively expanded using a generative adversarial network, and a convolutional neural network classifier is trained as an echo detector; finally, the trained detector is used to detect echoes from the test data and output distance information. The results of marine echo forwarding data processing show that this method achieved a detection rate of more than 87% for all test data with different signal-to-mixing ratios, which is more than 25% higher than the convolutional neural network classifier detection method based on conventional matched filtering features, time-frequency analysis features, and auditory perception features.

Description

Technical Field

[0001] The present invention relates to the field of active sonar detection in underwater acoustic physics, and in particular to an automatic detection method for active sonar echoes based on segmented matched filtering. Background Art

[0002] In recent years, with the rapid development of underwater acoustic target vibration and noise reduction technologies, passive underwater acoustic detection alone has become difficult to meet practical detection needs, and active underwater acoustic detection has gradually become a research hotspot. Active underwater acoustic detection involves actively transmitting a frequency-modulated signal to generate a transmitted echo on the target surface, then detecting the echo to obtain information such as the target's presence and location. In practical applications, active sonar echo detection is challenging due to the low signal-to-noise ratio of the echo signal, interference from reflected clutter from seabed reefs and marine organisms, and complex ocean noise. A commonly used processing method for active sonar echo detection is to use matched filtering (MF) to obtain the energy peak of the echo signal from the time domain data received by the hydrophone. The peak is then detected to determine the presence and location of the target echo. MF processing can improve the signal-to-noise ratio of the echo signal, thereby enhancing the ability to detect echoes from the received signal to a certain extent. However, after the transmitted signal is complexly modulated by the target and the channel, a mismatch between the received and transmitted signals often occurs, significantly reducing the detection performance of conventional MF features. Summary of the Invention

[0003] The purpose of this invention is to extract more stable features from active sonar echo data, effectively expand the echo samples, train a deep learning detector model, and improve the active sonar echo detection performance.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions.

[0005] The present invention proposes an automatic detection method for active sonar echoes based on segmented matched filtering, the method comprising:

[0006] The FM signal emitted by the active sonar is segmented and each sub-FM signal is used to perform matched filtering on the active sonar receiving data to obtain the PMF feature.

[0007] The obtained PMF features are intercepted as test data and post-processed to obtain test samples. The test samples are then input into a pre-established and trained convolutional neural network-based echo detector to obtain the probability of the target echo signal existing in the received data at different times, thereby obtaining the target detection results and their corresponding location information.

[0008] The method also includes training an echo detector based on a convolutional neural network; during training, a generative adversarial network is used to expand the training samples.

[0009] As one of the improvements to the above technical solution, the FM signal s(t) emitted by the active sonar is segmented and the nth segment sub-signal s pn (t) is expressed as:

[0010] s pn (t)=s(t)×w n (t)

[0011] Among them, w n (t) represents the rectangular window function when intercepting the nth sub-signal from the transmitted signal s(t), and the expression is:

[0012]

[0013] Where a represents the time step of the window function in the segmented processing, b represents the time window length of the window function. After the window length and step length of the window function are determined, the total number N of intercepted sub-signals is expressed as:

[0014] N=(T s -b) / a+1

[0015] Among them, T s Indicates the pulse width of the FM signal;

[0016] Use the N sub-signals after segmentation processing to perform matched filtering on the active sonar received data respectively. The time window length of the matched filtering processing is the window length b of the intercepted sub-signal, and the step length of the matched filtering processing is half of the sub-signal window length.

[0017] The N matched filter processing output results are combined to obtain the PMF feature.

[0018] As one of the improvements to the above technical solution, the time domain expression of the output result of the matched filtering process is:

[0019]

[0020] Among them, R s (t) is the autocorrelation function of the FM signal s(t), h(t) and h i (t) are H(f) and H i The inverse Fourier transform of (f) corresponds to the impulse response function of the transmitted signal in the process of propagating from the sound source to the target surface and then reflecting back to the receiver, h i (t) corresponds to the impulse response function of the i-th clutter during the propagation process, n(t) is the ocean environmental noise in the time domain, Represents the convolution operation.

[0021] As one of the improvements to the above technical solution, the acquisition of the training samples specifically includes:

[0022] According to the known echo arrival time in the active sonar received signal, the PMF feature is intercepted from the time region containing the target echo to obtain the target echo positive sample; at the same time, the PMF feature is intercepted from the time region not containing the target echo to obtain the non-target echo negative sample;

[0023] Downsampling the target echo positive samples and non-target echo negative samples;

[0024] The downsampled target echo positive samples and non-target echo negative samples are standardized to obtain training samples, i.e., target echo positive samples and non-target echo negative samples used to input into an echo detector based on a convolutional neural network.

[0025] As one of the improvements to the above technical solution, a sliding window method is adopted when intercepting the test data, and the window length of the sliding window is T s .

[0026] As one of the improvements to the above technical solution, the post-processing of the intercepted test data includes: downsampling and normalization. (The downsampling and normalization processing for the test samples is the same as that for the training samples.)

[0027] As one of the improvements to the above technical solution, the expression for the standardization process is:

[0028]

[0029] Among them, x s represents the sample input to the echo detector based on the convolutional neural network. The min(·) operation represents taking the smallest element in the PMF feature matrix. The max(·) operation represents taking the largest element in the PMF feature matrix. x0 represents the PMF feature sample after downsampling.

[0030] As one of the improvements to the above technical solution, the use of a generative adversarial network to expand the training samples specifically includes: inputting random noise into the trained generative adversarial network to generate false positive samples, and expanding the target echo positive samples;

[0031] The generative adversarial network includes: a generator and a discriminator;

[0032] The training process of the generative adversarial network includes: inputting random noise into the generator to generate false positive samples; inputting the false positive samples and true positive samples into the discriminator respectively to distinguish between true and false; and independently optimizing the parameters of the generator and the discriminator according to the true and false distinction results.

[0033] As one of the improvements to the above technical solution, the generator and the discriminator both use binary cross entropy as the loss function Loss during optimization, and the expression is:

[0034]

[0035] Among them, B represents the number of training samples in a batch, y k represents the true label of the kth sample, p(y k ) represents the probability of target existence output by the discriminator in the generative adversarial network for the kth sample.

[0036] As one of the improvements of the above technical solution, the echo detector based on convolutional neural network includes: an input layer, a first convolution layer, a second convolution layer, a fully connected layer and an output layer;

[0037] The ReLU function is used for activation in the first convolutional layer, the second convolutional layer, and the fully connected layer;

[0038] Add a Dropout layer between the fully connected layer and the output layer;

[0039] The Softmax function is used in the output layer for activation and outputs the probability value.

[0040] The present invention proposes an automatic detection method for active sonar echoes based on piecewise matched filtering (PMF) features. This method segments the transmitted frequency modulated signal and uses each segmented sub-signal to perform MF processing on the received data, thereby obtaining features that are more stable than the energy peak. At the same time, in response to the problem of insufficient active sonar echo data, the present invention uses generative adversarial networks (GAN) to expand the positive samples after PMF processing, and uses the expanded samples to train convolutional neural networks (CNN) as echo detectors. After the detector training is completed, the sliding window method is used to automatically detect the echo from the test data and obtain the distance information of the echo. The overall structure of the present invention is as follows: Figure 1 shown.

[0041] The effects of the present invention include the following aspects:

[0042] 1. The PMF feature proposed in this invention is more stable than the energy peak feature of conventional MF and has better detection performance;

[0043] 2. This paper uses GAN to expand the positive echo samples, effectively solving the problem of imbalance in the number of positive and negative samples in detector training, and can further improve the detection performance of the detector;

[0044] 3. The PMF feature proposed in this paper achieved an echo detection rate of over 87% for test data with different signal-to-noise ratios. Compared with MF features, Short Time Fourier Transform (STFT) features, Mel-Frequency Cepstral Coefficients (MFCC) features, and Wavelet Transform (WT) features, the detection rate was improved by more than 25%.

[0045] 4. The present invention uses GAN to expand the echo samples, which improves the detection rate by more than 10% compared with the repeated method. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of the method of the present invention;

[0047] Figure 2 It is the PMF processing flow;

[0048] Figure 3 is the network structure of GAN;

[0049] Figure 4 is the network structure of the CNN detector;

[0050] Figure 5 The detection process of the trained echo detector on the test data;

[0051] Figure 6(a) shows the experimental scenario, Figure 6(b) shows the depth variation of the sound source, hydrophone, and echo repeater, Figure 6(c) shows the measured sound velocity profile, Figure 6(d) shows the experimental ship's navigation route, and Figure 6(e) shows the comparison between the echo distance and the GPS distance.

[0052] Figure 7(a) is a schematic diagram of the signal-to-noise ratio calculation, and Figure 7(b) is the signal-to-noise ratio result calculated for all data;

[0053] Figure 8(a-1) and Figure 8(a-2) are two groups of time domain echo signals with different signal-to-noise ratios; Figure 8(b-1) and Figure 8(b-2) are the MF features extracted for the above two groups of echo signals; Figure 8(c-1) and Figure 8(c-2) are the PMF features extracted for the above two groups of echo signals; Figure 8(d-1) and Figure 8(d-2) are the STFT features extracted for the above two groups of echo signals; Figure 8(e-1) and Figure 8(e-2) are the MFCC features extracted for the above two groups of echo signals; Figure 8(f-1) and Figure 8(f-2) are the WT features extracted for the above two groups of echo signals.

[0054] Figure 9(a) shows the echo presence probability output by the detector trained using the repeated positive sample augmentation method, Figure 9(b) shows the echo presence probability output by the detector trained using the GAN positive sample augmentation method, and Figure 9(c) compares the detection rates output by the two positive sample augmentation methods.

[0055] Figure 10(a) shows the echo existence probability output by the detector trained based on MF features, Figure 10(b) shows the echo existence probability output by the detector trained based on STFT features, Figure 10(c) shows the echo existence probability output by the detector trained based on MFCC features, Figure 10(d) shows the echo existence probability output by the detector trained based on WT features, Figure 10(e) shows the echo existence probability output by the detector trained based on PMF features, and Figure 10(f) shows a comparison of the detection rates of the detectors trained based on the above features for test data with different signal-to-noise ratio thresholds. DETAILED DESCRIPTION

[0056] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0057] In order to achieve the above object, the present invention adopts the following technical solutions:

[0058] The method consists of three parts: PMF processing, sample generation and expansion, and detector training and testing.

[0059] (1) PMF processing:

[0060] The time domain data x(t) received by active sonar can be written as:

[0061]

[0062] Where s(t) is the frequency modulated signal emitted by the active sonar, h(t) corresponds to the impulse response function of the transmitted signal during the process of propagating from the sound source to the target surface and then reflecting back to the receiver, and h i(t) corresponds to the impulse response function of the i-th clutter during the propagation process, m is the total number of clutter, n(t) is the ocean environment noise, Represents the convolution operation. The spectrum S(f) of the frequency modulated signal s(t) is expressed as:

[0063]

[0064] Among them, T s represents the pulse width of the FM signal. Therefore, the spectrum X(f) of the sonar received time domain data x(t) is:

[0065]

[0066] Among them, H(f) and H i (f) are h(t) and h i (t) corresponds to the transfer function, N(f) is the spectrum of the ocean ambient noise n(t). Based on the spectrum X(f) of the active sonar transmission signal, the frequency domain representation X0(f) output after conventional MF processing is:

[0067]

[0068] Where * represents the conjugate of the complex number. The inverse Fourier transform of the above frequency domain output X0(f) can be further used to obtain the time domain output x0(t) of the conventional MF processing, which is expressed as:

[0069]

[0070] Among them, R s (t) is the autocorrelation function of the transmitted FM signal s(t). Due to the similarity in waveform between the target echo and the transmitted signal, x0(t) will have a matching energy peak at the time of echo arrival. However, after complex modulation by target reflection and channel transmission, the echo signal and the transmitted signal will have varying degrees of waveform mismatch. The peak structure of the MF feature varies significantly for target echo data at different receiving locations, resulting in reduced echo detection performance.

[0071] In this regard, the present invention proposes a PMF processing method, the process of which is as follows: Figure 2 First, the method performs segmented processing on the transmitted FM signal s(t), and the nth segment sub-signal s after segmented processing is pn (t) is expressed as:

[0072] s pn (t)=s(t)×w n (t), (6)

[0073] Among them, w n(t) represents the rectangular window function when intercepting the nth sub-signal from the transmitted signal s(t), and the expression is:

[0074]

[0075] Where a represents the time step of the window function in the segmented processing, b represents the time window length of the window function. After the window length and step length of the window function are determined, the total number N of intercepted sub-signals is expressed as:

[0076] N=(T s -b) / a+1. (8)

[0077] Afterwards, the N sub-signals after segmentation processing are used to perform MF processing on the active sonar reception data. The time window length of the MF processing is the window length b of the intercepted sub-signal, and the step length of the MF processing is half of the sub-signal window length. Finally, the N conventional MF processing output results are combined to obtain the PMF feature. Figure 2 As shown in FIG, at the location of the target echo, the PMF feature is a slanted line with higher energy.

[0078] (2) Sample generation and expansion:

[0079] To facilitate the subsequent training and testing of the CNN classifier, it is necessary to post-process the features output after PMF processing to generate usable training and test samples. To obtain a positive echo sample, it is first necessary to intercept a period of time T at the corresponding position based on the known echo arrival time in the active sonar received signal. s The PMF feature of When the sampling rate of active sonar receiving data is f s When the intercepted time is T s The number of time sampling points corresponding to the PMF feature is f s ×T s , which is usually much larger than the number of sub-signals N in the PMF processing. Subsequently, in order to facilitate the CNN classifier to learn the two-dimensional bright line structure of the echo in the PMF feature, it is necessary to downsample the time domain of the PMF feature. The downsampling factor selected in this invention is 400 (i.e., the time domain sampling period is reduced to 400 times the original). The PMF matrix after downsampling is expressed as x0(n| n＝1,2,3,…,N ,m| m＝1,2,3,…,M ), where M represents the number of time domain sampling points after downsampling, which is 1 / 400 of the original number of sampling points. Finally, the downsampled PMF matrix is ​​normalized to obtain the positive echo samples that can be used for CNN detector training. The normalization expression is:

[0080]

[0081] Among them, x s Denotes PMF positive samples that can be directly used for CNN detector training. The min(·) operation takes the smallest element in the PMF matrix, and the max(·) operation takes the largest element in the PMF matrix. After obtaining available target echo positive samples, non-target echo negative samples can be extracted using the same method from the time region corresponding to the PMF feature that does not contain the target echo. The extraction duration and post-processing methods are the same as for the positive samples.

[0082] In actual active sonar reception data, the data duration corresponding to the target echo is usually much shorter than the data duration corresponding to the non-target echo. Therefore, the number of generated target echo positive samples is much smaller than the number of non-target echo negative samples. In order to be able to train the CNN echo detector, it is necessary to expand the positive samples so that the number of positive and negative samples is balanced. In the field of deep learning, GAN, as a generative model, is often used to generate false samples similar to real samples, thereby expanding the number of training samples and achieving good performance improvement. Therefore, the present invention selects GAN to expand the number of positive echo samples of PMF, and its network structure is as follows: Figure 3 As shown in Figure 1, it mainly consists of a generator and a discriminator. The principle is as follows: first, random noise is input into the generator to generate false positive samples; then, the false positive samples and true positive samples are input into the discriminator for true and false identification; finally, the parameters of the generator and discriminator are continuously optimized through the above iterative process, ultimately generating a large number of false samples that are very similar to real samples.

[0083] The GAN generator of this invention consists of two convolutional layers with kernels of 5×5 and 7×7, respectively. The generator's input random noise is of size 1×200, and its output is of size N×M, the same size as the PMF sample. The GAN discriminator consists of one convolutional layer with a kernel of 3×3 and two fully connected layers with kernel sizes of 128 and 32, respectively. The discriminator's input size is still N×M, and its output is unary data representing the probability that the input sample is a true sample.

[0084] During GAN training, the generator and discriminator perform parameter optimization independently. When the generator is optimizing parameters, the discriminator's parameters are frozen, and the generator parameters are optimized based on the results of its discrimination between true and false samples. When the discriminator is optimizing parameters, the parameters are optimized directly based on its discrimination between true and false samples. Both the generator and the discriminator use binary cross entropy as the loss function during parameter optimization, which is expressed as:

[0085]

[0086] Among them, L represents the calculated binary cross entropy loss function value, B represents the number of training samples in a batch, and y k represents the true label of the kth sample (0 or 1 in the binary classification problem), p(y k ) represents the probability output by the GAN discriminator for the kth sample indicating that it is the target echo.

[0087] (3) Detector training and testing:

[0088] The present invention selects CNN model as the detector to detect active sonar echo. The specific network structure is as follows: Figure 4 As shown in Figure 2, the target echo detector takes as input the PMF feature samples extracted from active sonar data and outputs binary data representing the probabilities of a target echo and a non-target echo. The CNN detector consists of two convolutional layers and one fully connected layer. The convolution kernel size of both convolutional layers is 3×3, the convolution stride is 1×1, and the convolution kernel size of the fully connected layer is 64. Convolution operations are followed by a max pooling layer with a pooling kernel size of 3×3 and a pooling stride of 2×2. Reinforced Luminance (ReLU) activation is used in both convolutional and fully connected layers. To prevent overfitting, a Dropout layer with a regularization factor of 0.3 is added between the fully connected layer and the output layer. The output layer uses a Softmax activation function to convert the output into a probability value. Adam is used for training, with an initial learning rate of 0.0001, a batch size of 64, and 100 epochs for all training samples.

[0089] After completing the training of the CNN detector, it is necessary to use the trained detector to detect the target echo from the test data and estimate the distance information of the echo. The test process is as follows: Figure 5 As shown. For a certain frame of active sonar receiving time domain signal as test data, PMF features are extracted. After the PMF feature extraction is completed, the sliding window method is first used to slide the PMF feature for a time length of T. s The feature matrix of , the total number of feature matrices is cut as g; then, the g feature matrices are post-processed separately to generate test samples that can be directly used as detector input; finally, the g test samples are input into the trained CNN detector one by one to obtain the detection results at the corresponding time. Figure 5 As shown, after the above processing, the detection result of the target echo and the corresponding distance information can be obtained from the processed test data.

[0090] After the detector is trained, it is necessary to use the trained detector to detect target echoes in the test data. First, test samples are extracted from the test data using a sliding window method. The window length of the sliding window is still the pulse width of the FM signal, and the step size of the sliding window is selected based on actual usage. The extracted test samples are then processed using the same method as the training data to generate test samples that can directly serve as detector input.

[0091] An echo repeater experiment simulating active sonar detection was conducted in the northern South China Sea in November 2021. During the experiment, one experimental vessel simultaneously towed a sonar source and a hydrophone while underway; another experimental vessel was anchored and deployed an echo repeater, responsible for forwarding the received frequency-modulated (FM) signals. The towed sonar source emitted a linear FM signal every 20 seconds, with a pulse width of 2 seconds and a frequency band of 700 to 750 Hz. Upon receiving the FM signal from the towed sonar source, the echo repeater randomly added a gain of -3 to 0 dB and forwarded it, simulating the echo generated by an actual target at sea. The hydrophone continuously received underwater acoustic data. A total of 176 complete echo cycles were obtained during the experiment. Figure 6(a) shows the experimental scenario. Figure 6(b) shows the depth variations of the towed sonar source, hydrophone, and echo repeater over time during the experiment. Figure 6(c) shows the sound velocity profile measured at a particular time during the experiment. Figure 6(d) shows the motion paths of the signal transmitting and receiving vessels and the echo repeater during the experiment. Figure 6(e) shows the distances corresponding to the echo signals in all 176 frames of received data. These distances agree well with the actual distances between the echo repeater and the sound source calculated using GPS data, proving that the echo signals in the hydrophone data are indeed the signals forwarded by the echo repeater.

[0092] For each target echo in each frame of echo data, the average energy of the signal and noise regions is calculated as shown in Figure 7(a). The signal-to-noise ratios of the 176 echo frames are then calculated, as shown in Figure 7(b). When training the target detector, the first 10 echo frames are selected as the training set, and the last 166 echo frames are selected as the test set. The training and test sets are separated by the vertical line in Figure 7(b).

[0093] Figure 8 extracts the MF, PMF, STFT, MFCC, and WT features for two frames of received echo data with different signal-to-noise ratios. Figure 8(a) directly shows the time domain received signal corresponding to the two frames of echo data for these two periods, while Figures 8(b) through 8(f) provide comparisons of conventional MF, PMF, STFT, MFCC, and WT features, respectively. The segmented matching duration selected for the PMF is 1 / 2 the FM signal pulse width (1 s), with a step size of 1 / 80 the FM signal pulse width (0.025 s).

[0094] Figure 9 shows a comparison of the detection results of the trained detectors for test data with different signal-to-noise ratio thresholds when using GAN for positive sample expansion and using the repeated method for positive sample expansion. The results show that compared with the repeated expansion method, using GAN for positive sample expansion can increase the echo detection rate by more than 10%, and the false alarm rate does not increase significantly. When using GAN for positive sample expansion, MF features, STFT features, MFCC features, WT features, and PMF features are extracted for detector training. Figure 10 shows the detection results of the test data output with different signal-to-noise ratio thresholds. By comparison, it can be seen that PMF improves the active sonar echo detection rate by more than 25%, and the false alarm rate does not increase significantly.

[0095] The results of processing maritime echo forwarding data show that this method achieved a detection rate of over 87% for all test data with different signal-to-mixture ratios. Compared with the convolutional neural network classifier detection method based on conventional matched filtering features, time-frequency analysis features and auditory perception features, the detection rate is increased by more than 25%.

[0096] From the above specific description of the present invention, it can be seen that the PMF feature proposed in the present invention has stronger robustness against different signal-to-noise ratios, which can enable the trained target detector to have a higher detection rate for active echo signals; the proposed GAN sample expansion method can well expand the echo positive samples, effectively solving the problem of insufficient number of target echo samples.

[0097] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A method for automatic detection of active sonar echoes based on segmented matched filtering, the method comprising: The FM signal emitted by the active sonar is segmented and each sub-FM signal is used to perform matched filtering on the active sonar receiving data to obtain the PMF feature. The obtained PMF features are intercepted as test data and post-processed to obtain test samples. The test samples are then input into a pre-established and trained convolutional neural network-based echo detector to obtain the probability of the target echo signal existing in the received data at different times, thereby obtaining the target detection results and their corresponding location information. The method further includes training an echo detector based on a convolutional neural network; During training, a generative adversarial network is used to expand the training samples.

2. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 1, characterized in that: The FM signal s(t) emitted by the active sonar is processed in segments, and the nth segment signal s after segment processing is pn (t) is expressed as: s pn (t)=s(t)×w n (t) Among them, w n (t) represents the rectangular window function when intercepting the nth sub-signal from the transmitted signal s(t), and the expression is: Where a represents the time step of the window function in the segmented processing, b represents the time window length of the window function. After the window length and step length of the window function are determined, the total number N of intercepted sub-signals is expressed as: N=(T s -b) / a+1 Among them, T s Indicates the pulse width of the FM signal; Use the N sub-signals after segmentation processing to perform matched filtering on the active sonar received data respectively. The time window length of the matched filtering processing is the window length b of the intercepted sub-signal, and the step length of the matched filtering processing is half of the sub-signal window length. The N matched filter processing output results are combined to obtain the PMF feature.

3. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 2, characterized in that: The time domain expression x0(t) of the output result of the matched filtering process is: Among them, R s (t) is the autocorrelation function of the FM signal s(t), h(t) and h i (t) are H(f) and H i The inverse Fourier transform of (f) corresponds to the impulse response function of the transmitted signal in the process of propagating from the sound source to the target surface and then reflecting back to the receiver, h i (t) corresponds to the impulse response function of the i-th clutter during the propagation process, n(t) is the ocean environmental noise in the time domain, Represents the convolution operation.

4. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 3 is characterized in that: The acquisition of the training samples specifically includes: According to the known echo arrival time in the active sonar received signal, the PMF feature is intercepted from the time region containing the target echo to obtain the target echo positive sample; at the same time, the PMF feature is intercepted from the time region not containing the target echo to obtain the non-target echo negative sample; Downsampling the target echo positive samples and non-target echo negative samples; The downsampled target echo positive samples and non-target echo negative samples are normalized to obtain training samples.

5. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 4, characterized in that: When intercepting from the test data, the sliding window method is adopted, and the window length of the sliding window is T s .

6. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 5, characterized in that: The post-processing includes: performing downsampling and standardization processing in sequence.

7. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 4, characterized in that: The expression of the standardization process is: Among them, x s represents the sample input to the echo detector based on the convolutional neural network. The min(·) operation represents taking the smallest element in the PMF feature matrix. The max(·) operation represents taking the largest element in the PMF feature matrix. x0 represents the PMF feature sample after downsampling.

8. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 4, characterized in that: The use of a generative adversarial network to expand the training samples specifically includes: inputting random noise into the trained generative adversarial network to generate false positive samples, and expanding the target echo positive samples; The generative adversarial network includes: a generator and a discriminator; The training process of the generative adversarial network includes: inputting random noise into the generator to generate false positive samples; inputting the false positive samples and true positive samples into the discriminator respectively to distinguish between true and false; and independently optimizing the parameters of the generator and the discriminator according to the true and false distinction results.

9. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 8, characterized in that: The generator and discriminator both use binary cross entropy as the loss function Loss during optimization, and the expression is: Among them, B represents the number of training samples in a batch, y k represents the true label of the kth sample, p(y k ) represents the probability of target existence output by the discriminator in the generative adversarial network for the kth sample.

10. The automatic detection method of active sonar echo based on segmented matched filtering according to claim 1, characterized in that: The echo detector based on convolutional neural network includes: an input layer, a first convolution layer, a second convolution layer, a fully connected layer and an output layer; The ReLU function is used for activation in the first convolutional layer, the second convolutional layer, and the fully connected layer; Add a Dropout layer between the fully connected layer and the output layer; The Softmax function is used in the output layer for activation and outputs the probability value.

Images