A Passive Sonar Intelligent Detection Method Based on Multi-Feature Fusion

Abstract

The present invention discloses a passive sonar intelligent detection method based on multi-feature fusion, which relates to the field of underwater acoustic detection and includes: First, based on the ASEMD decomposition algorithm, DEMON spectral features of the target radiated noise signal are extracted and MFCC features based on the human ear auditory characteristics of the target radiated noise signal are extracted; Then, based on the multi-feature collaborative fusion network, the DEMON spectral features and the MFCC features are fused; Finally, the fused features are fed into the Ret residual detection network for target detection; The present invention solves the problem that it is difficult to determine the parameters of the traditional DEMON spectral detection band-pass filter and it cannot adaptively change with the change of the input signal; At the same time, it realizes the collaborative learning and feature fusion of multi-dimensional features, breaks through the limitation of relying on single-feature detection, improves the passive detection performance of underwater targets, as well as the generalization and robustness in the complex ocean noise background.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic detection, and particularly to a passive sonar intelligent detection method based on multi-feature fusion. Background Art

[0002] The statements in this section only provide background information related to the present disclosure and may not constitute prior art.

[0003] In future naval battles, underwater acoustic countermeasure plays an important role, and target detection is the basis of underwater acoustic countermeasure. Only by accurately detecting enemy targets can many subsequent tasks be completed; underwater target detection mainly uses acoustic means and can be divided into active and passive working modes, which respectively target the target echo or target radiated noise, and realize acoustic-electric conversion through a transducer array. Usually, array and time correlation operations are used to obtain spatial and temporal gains to achieve the purpose of increasing the detection range of underwater targets; among them, the passive working mode has the advantages of long detection range, good concealment, and being not easily attacked by the enemy, and is an important means for underwater acoustic detection of various underwater targets; however, since the target noise becomes very weak after long-distance propagation as a signal, passive sonars often work in a low signal-to-noise ratio situation. Therefore, how to take more signal processing measures to realize the detection of weak targets is the key problem of passive detection.

[0004] Since the line spectrum has relatively stable energy, traditional passive detection mainly detects the line spectrum components of the target. Among them, the Detection of Envelope Modulation on Noise (DEMON) detection method is a commonly used classical method; the DEMON detection method is an algorithm that demodulates the received broadband high-frequency signal to calculate the low-frequency modulation spectrum, and can obtain the strong modulation characteristic spectrum at the low-frequency end; through DEMON analysis, invariant target physical field characteristics such as target shaft frequency and blade frequency can also be obtained, providing a new means for target classification and recognition.

[0005] However, the traditional DEMON algorithm has certain limitations in processing. The received radiated noise needs to be processed by band-pass filters with different bandwidths before envelope analysis, and the bandwidth and the number of band-pass filters are unknown, and it is difficult to determine the filter parameters, which cannot adaptively change with the change of the input signal; and with the development of noise reduction technology, the radiated noise of underwater targets is easily submerged in the ambient noise, and the performance of the above traditional methods will drop significantly. Therefore, it is very difficult to realize target detection at low signal-to-noise ratio only relying on the detection of a single feature. Summary of the Invention

[0006] The object of the present invention is to provide a passive sonar intelligent detection method based on multi - feature fusion in view of the limitations of the current traditional underwater target passive detection DEMON algorithm in processing. Before envelope analysis, the received radiated noise needs to be processed by band - pass filtering with different bandwidths, and the bandwidth and the number of band - pass filters are unknown, making it difficult to determine the filter parameters and unable to adaptively change with the change of the input signal. Moreover, with the development of noise reduction technology, the radiated noise of underwater targets is easily submerged in environmental noise. The method uses the mode components IMF obtained by the adaptive selection empirical mode decomposition algorithm (ASEMD) to replace the band - pass filtering in the DEMON method, avoiding the deficiency of relying on experience to preset the bandwidth of the band - pass filter and the number of band - pass filters in advance, realizing the adaptive division of frequency bands and solving the problem of difficult determination of filter parameters. By extracting the Mel - frequency cepstral coefficients (MFCC) features of the radiated noise and using multi - feature fusion based on collaborative learning for passive detection of underwater targets, it improves the detection ability of weak targets with low radiated noise levels, enhances the target detection performance in complex marine environments, and breaks through the bottleneck of "inaccurate detection" of underwater targets, thus solving the above problems.

[0007] The technical solution of the present invention is as follows:

[0008] A passive sonar intelligent detection method based on multi - feature fusion, comprising:

[0009] Step S1: Based on the ASEMD decomposition algorithm, extract the DEMON spectral features of the target radiated noise signal;

[0010] Step S2: Extract the MFCC features of the target radiated noise signal based on the human ear auditory characteristics;

[0011] Step S3: Based on the multi - feature collaborative fusion network, fuse the DEMON spectral features and the MFCC features;

[0012] Step S4: Send the fused features into the Ret residual detection network for target detection.

[0013] Further, the step S1 includes:

[0014] Step S11: Decompose the target radiated noise signal into n mode components by the ASEMD decomposition algorithm;

[0015] Step S12: Demodulate the n mode components respectively to obtain the DEMON spectra corresponding to each mode component;

[0016] Step S13: Calculate the weighted coefficients corresponding to each DEMON spectrum;

[0017] Step S14: Fuse each DEMON spectrum with its corresponding weighting coefficient to obtain the fused DEMON spectrum.

[0018] Further, the ASEMD decomposition algorithm includes:

[0019] Step A: Perform EMD modal decomposition on the input target radiated noise signal;

[0020] Step B: Extract modal components from the decomposition result;

[0021] Step C: Determine whether the modal component is the first-order modal component IMF1; if so, jump to Step D; if not, jump to Step E;

[0022] Step D: Determine whether the modal component contains high-frequency gap components; if not, jump to Step E; if so, perform the CEEMD algorithm;

[0023] Step E: Calculate the remaining component; when the remaining component meets the cut-off condition, the algorithm ends; when the remaining component does not meet the cut-off condition, repeat Step A and Step B until the remaining component meets the cut-off condition.

[0024] Further, the CEEMD algorithm includes:

[0025] Step D1: Add I pairs of noise components with opposite polarities to the target radiated noise signal to form 2I groups of noise-added signals;

[0026] Step D2: Perform EMD decomposition on the noise-added signals respectively to obtain 2I groups of modal components;

[0027] Step D3: Add the corresponding 2I groups of modal components and perform lumped averaging to obtain the nth-order modal component;

[0028] Step D4: Calculate the remaining component; when the remaining component meets the cut-off condition, the algorithm ends; when the remaining component does not meet the cut-off condition, repeat Step D2 and Step D3 until the remaining component meets the cut-off condition.

[0029] Further, the cut-off condition includes:

[0030] The remaining component is a constant or a monotonic function.

[0031] Further, Step S12 includes:

[0032] Perform square detection and low-pass filtering on the n modal components to obtain the envelope signal, and perform FFT processing on the envelope signal to obtain the DEMON spectrum P of each modal component IMFi , i = 1, 2, …, n.

[0033] Further, the step S13 includes:

[0034] Pass the DEMON spectra of each modal component through an α bidirectional filter to obtain the DEMON spectrum weighting coefficients of each modal component;

[0035] The step S14 includes:

[0036] Calculate the line spectra of each modal component through the DEMON line spectrum estimation formula;

[0037] Multiply the line spectra of each modal component by their corresponding DEMON spectrum weighting coefficients, and then sum them to obtain the fused DEMON spectrum.

[0038] Further, the step S2 includes:

[0039] Perform pre-emphasis, frame segmentation, and windowing on the target radiated noise signal;

[0040] After performing discrete Fourier transform on the time-domain segmented signal, obtain the energy spectrum:

[0041] Filter the energy spectrum through a Mel frequency filter bank to obtain the Mel spectrum:

[0042] Take the logarithmic energy of the Mel spectrum and perform discrete cosine transform to obtain the MFCC coefficients.

[0043] Further, the multi-feature collaborative fusion network includes:

[0044] The backbone network consists of two parallel residual networks to form a dual-branch distributed network, which are respectively used as the DEMON spectrum feature and MFCC feature extractors;

[0045] The extractor contains 4 convolutional blocks and an attention module. Each convolutional block is composed of a BN layer, a convolutional layer, an average pooling layer, and an activation function. The convolutional layer contains 128 convolutional kernels with a size of [5×5] and a stride of 2;

[0046] At the same time, a 1×1 convolutional layer is in parallel with the fourth convolutional block to complete the extraction of the two features;

[0047] The extracted feature vectors enter the attention module, and are converted into numbers from 0 to 1 using the softmax activation function as the weights of the features. The two weights are connected to form a feature selection matrix and multiplied by the features to achieve feature selection, thereby obtaining the fused features.

[0048] Further, the Ret residual detection network includes:

[0049] The Ret residual detection network consists of two residual modules, a fully connected layer, and an output layer;

[0050] The residual module consists of a BN layer, a convolutional layer with 256 convolutional kernels, a size of [3×3], a stride of 2, a max pooling layer, and a ReLU activation function.

[0051] Compared with the existing technologies, the beneficial effects of the present invention are as follows:

[0052] 1. A passive sonar intelligent detection method based on multi-feature fusion, which solves the problems existing in traditional underwater target passive detection; improves the traditional DEMON spectrum detection by using the ASEMD algorithm, solves the problem that it is difficult to determine the parameters of the band-pass filter in traditional DEMON spectrum detection and it cannot adaptively change with the change of the input signal; analyzes the MFCC features of the underwater target radiation noise, uses a multi-feature collaborative fusion network to deeply mine the DEMON and MFCC features, realizes the collaborative learning and feature fusion of multi-dimensional features, and uses a residual network to intelligently detect the target, breaks through the limitation of relying on single-feature detection, improves the passive detection performance of underwater targets, as well as the generalization and robustness in the complex ocean noise background; can be widely applied to military and civilian fields, and provides accurate target detection information for combat and civilian applications.

[0053] 2. A passive sonar intelligent detection method based on multi-feature fusion, which adaptively selects the type of EMD algorithm (ASEMD algorithm) by judging the presence or absence of the high-frequency intermittent component of the first-order intrinsic mode function of the signal. When there is no intermittent component, the EMD algorithm is selected, and when there is an intermittent component, the CEEMD algorithm is selected, which can not only effectively solve the mode mixing problem but also reduce the calculation amount.

[0054] 3. A passive sonar intelligent detection method based on multi-feature fusion, which improves the traditional DEMON spectrum detection algorithm by using the ASEMD algorithm, replaces the band-pass filter for DEMON spectrum demodulation with the IMF components decomposed by the ASEMD algorithm, avoids the situation that the bandwidth and the number of band-pass filters need to be preset by experience before DEMON spectrum demodulation, and solves the problem that it is difficult to determine the filter parameters and it cannot adaptively change with the change of the input signal.

[0055] 4. A passive sonar intelligent detection method based on multi-feature fusion, aiming at the problem that it is difficult to detect targets under low signal-to-noise ratio by relying on traditional single-feature detection. While using DEMON spectrum feature detection, the Mel-frequency cepstral coefficient features (MFCC) based on human ear auditory characteristics are extracted from the target radiation noise.

[0056] 5. A passive sonar intelligent detection method based on multi-feature fusion realizes underwater target detection by using a multi-feature collaborative fusion network. A dual-branch distributed network is used as an extractor for the DEMON spectrum and MFCC features of radiated noise to obtain the unique features of the modality and the common features across modalities. An attention mechanism is introduced to screen the features using this mechanism, increasing the weight of important features during fusion, thereby achieving the effect of multi-dimensional information complementarity, improving the utilization rate of target information, and finally realizing the intelligent detection of targets through a residual network. Description of the Drawings

[0057] Figure 1 It is a flowchart of a passive sonar intelligent detection method based on multi-feature fusion;

[0058] Figure 2 It is a power spectrum diagram of radiated noise;

[0059] Figure 3 It is a DEMON spectrum detection diagram based on ASEMD;

[0060] Figure 4 It is a flowchart of the ASEMD decomposition algorithm;

[0061] Figure 5 It is a DEMON spectrum diagram of the first 5 order IMF components;

[0062] Figure 6 It is a flowchart for extracting MFCC features;

[0063] Figure 7 It is a multi-feature collaborative fusion network diagram. Detailed Embodiment

[0064] It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.

[0065] The features and performance of the present invention will be further described in detail below in conjunction with the embodiments.

[0066] Embodiment 1

[0067] Since the line spectrum has relatively stable energy, traditional passive detection mainly focuses on detecting the line spectrum components of the target. Among them, the Detection of Envelope Modulation on Noise (DEMON) detection method is a commonly used classical method. The DEMON detection method is an algorithm that demodulates the received broadband high-frequency signal to calculate the low-frequency modulation spectrum, and can obtain the strong modulation characteristic spectrum at the low-frequency end. Through DEMON analysis, invariant target physical field characteristics such as the target shaft frequency and blade frequency can also be obtained, providing a new means for the classification and recognition of targets.

[0068] However, the traditional DEMON algorithm has certain limitations in processing. The received radiated noise needs to be processed by band-pass filtering with different bandwidths before envelope analysis, and the bandwidth and the number of band-pass filters are unknown. It is difficult to determine the filter parameters and they cannot adaptively change with the change of the input signal. And with the development of noise reduction technology, the radiated noise of underwater targets is easily submerged in the environmental noise, and the performance of the above traditional methods will drop significantly. Therefore, it is very difficult to achieve target detection under low signal-to-noise ratio only relying on the detection of a single feature.

[0069] In view of the above problems, this embodiment proposes a passive sonar intelligent detection method based on multi-feature fusion. The mode component IMF obtained by the Adaptive Selection Empirical Mode Decomposition algorithm (ASEMD) is used to replace the band-pass filtering in the DEMON method, avoiding the deficiency of relying on experience to preset the bandwidth of the band-pass filter and the number of band-pass filters in advance, and can realize the adaptive division of frequency bands, solving the problem of difficult determination of filter parameters. And by extracting the Mel Frequency Cepstral Coefficient (MFCC) features of the radiated noise, passive detection of underwater targets is realized by using multi-feature fusion based on collaborative learning, improving the detection ability of weak targets with low radiated noise level, enhancing the target detection performance in complex marine environments, and breaking through the bottleneck of "difficult to accurately detect" underwater target detection.

[0070] In this embodiment, it should be noted first that in a complex marine environment, the generation and propagation mechanism of the radiated noise of underwater acoustic targets such as ships is very complex, with diverse components, including both broadband continuous spectrum components and narrowband line spectrum components, as well as obvious modulation components. Therefore, the radiated noise of ships can be modeled as:

[0071] x(t) = (1 + m(t))·N(t) + A·L(t)

[0072] Where:

[0073] x(t) is the ship noise;

[0074] N(t) is the broadband continuous spectrum, corresponding to the propeller cavitation noise signal;

[0075] m(t) is the modulation spectrum component, and the modulation frequency and multiple relationship correspond to the propeller characteristics of the ship target;

[0076] L(t) is the line spectrum component corresponding to the mechanical noise of the ship target;

[0077] A can be corrected by the decibel number by which the line spectrum is higher than the continuous spectrum;

[0078] Figure 2 The power spectrum diagram of the radiated noise is given.

[0079] In the actual ocean environment, the underwater acoustic channel is complex and changeable, and the signal-to-noise ratio is low. The single feature of the underwater acoustic target is easy to be weakened and interfered, resulting in the inability to effectively detect the target. Therefore, the present invention adopts a detection method based on the fusion of improved DEMON spectral features and MFCC features, and the process is as Figure 1 shown.

[0080] Among them, the extraction method of the improved DEMON spectral feature is to use the ASEMD modal decomposition algorithm to replace the band-pass filter of the traditional DEMON spectral demodulation to solve the problem that the parameters of the band-pass filter cannot adaptively change with the change of the input signal;

[0081] The MFCC feature is to extract the Mel cepstral coefficients of the target radiated noise;

[0082] After obtaining the DEMON envelope spectrum and Mel-domain features of the target noise, feature fusion is carried out through a multi-feature collaborative fusion network, and target detection can be realized by training and testing optimization of the residual detection network.

[0083] Please refer to Figure 1-3 , a passive sonar intelligent detection method based on multi-feature fusion, specifically including the following steps:

[0084] Step S1: Based on the ASEMD decomposition algorithm, extract the DEMON spectral features of the target radiated noise signal; that is, obtain the DEMON line spectrum data set S1;

[0085] Step S2: Extract the MFCC features of the target radiated noise signal based on the human ear auditory characteristics; that is, obtain the logarithmic Mel energy spectrum data set S2;

[0086] Preferably, the DEMON line spectrum data set S1 and the logarithmic Mel energy spectrum data set S2 can be divided into a training set and a test set according to a ratio of 7:3, and it is determined whether there is an underwater target through manual annotation; among them, the training set is used to train the multi-feature collaborative fusion network and the Ret residual detection network, and the test set is used to test the multi-feature collaborative fusion network and the Ret residual detection network;

[0087] Step S3: Based on the multi-feature collaborative fusion network, fuse the DEMON spectral features and the MFCC features;

[0088] Step S4: Send the fused features into the Ret residual detection network for object detection.

[0089] In this embodiment, specifically, the step S1 includes:

[0090] Step S11: Decompose the target radiated noise signal into n modal components by the ASEMD decomposition algorithm; preferably, n is generally taken as 3 - 5;

[0091] Step S12: Demodulate the n modal components respectively to obtain the DEMON spectra corresponding to each modal component;

[0092] Step S13: Calculate the weighting coefficients corresponding to each DEMON spectrum;

[0093] Step S14: Fuse each DEMON spectrum with its corresponding weighting coefficient to obtain the fused DEMON spectrum.

[0094] In this embodiment, specifically, first, it should be noted that in the step S11:

[0095] Traditional DEMON spectrum fusion detection requires band - pass filtering of the target radiated noise signal in different frequency bands, but the parameters and number of band - pass filters are unknown and need to be set manually by experience, and cannot adaptively change with the input signal; while the EMD decomposition is theoretically equivalent to a binary filter bank, which is equivalent to screening the target radiated noise signal through a group of band - pass filters. Therefore, the band - pass filter part in the traditional DEMON method can be replaced by the IMF extracted by EMD, and EMD is suitable for analyzing non - linear and non - stationary signal sequences and has a high signal - to - noise ratio, so it is suitable for processing underwater acoustic detection signals with complex channels; in view of the possible existence of high - frequency intermittent components during underwater detection, the present invention adopts an improved ASEMD decomposition algorithm to extract DEMON spectrum features of the target radiation signal, and the flow chart is as Figure 3 shown.

[0096] In practical applications, when there are high - frequency intermittent components in the original signal, modal aliasing occurs in the EMD decomposition, which will contaminate the subsequent decomposition results; while the complementary ensemble empirical mode decomposition algorithm (CEEMD) can effectively solve this problem. By adding (a positive number and a negative number) Gaussian white noise in pairs to fill the entire time - frequency space and then taking the ensemble average after EMD decomposition, it not only reduces modal aliasing but also can quickly cancel the residual noise doped in the reconstructed signal with only a few average times.

[0097] On this basis, the present invention provides an Adaptive Selective EMD decomposition algorithm (ASEMD) that can adaptively select a suitable EMD decomposition algorithm according to the input signal. By judging the presence or absence of high-frequency intermittent components in the first-order intrinsic mode function of the signal, the type of EMD algorithm (ASEMD algorithm) is adaptively selected. When there are no intermittent components, the conventional EMD algorithm is selected; when there are intermittent components, the CEEMD algorithm is selected. This can not only effectively solve the mode mixing problem and reduce the computational complexity, but also avoid another mode mixing caused by the noise-assisted algorithm when the signal does not have high-frequency intermittent components (the randomness of the noise causes the signal components at the same time scale to be decomposed into different modes).

[0098] Among them, the detection of high-frequency intermittent components is performed on the first-order intrinsic mode component IMF1. The first-order intrinsic mode component of empirical mode decomposition represents the high-frequency signal component in the signal, and its extreme points correspond to the extreme points of the original signal in time. Therefore, it is possible to check whether there are high-frequency intermittent components in the original signal by examining the change of the extreme points of IMF1. When there are high-frequency intermittent components, the distance between the extreme points of the signal will significantly decrease, that is, at the starting point of the intermittent component, there is a jump in the distance between the extreme points. By detecting the fluctuation of the distance between the extreme points of the signal and using the mean prediction method, it is possible to determine whether there is an intermittent signal in the signal.

[0099] In this embodiment, specifically, the ASEMD decomposition algorithm includes:

[0100] First, the ASEMD decomposition algorithm needs to initialize the number of iterations. Let n = 0, and the remaining component r n (t) = x(t);

[0101] Step A: Perform EMD mode decomposition on the input target radiated noise signal; that is, perform conventional EMD mode decomposition on r n (t), that is, obtain the maximum and minimum envelope curves m n (t) of r 11 (t), and calculate the candidate component c n+1,1 (t) = r n (t) - m n+1,1 (t);

[0102] Step B: Extract the mode components from the decomposition results; preferably, if the candidate component c n+1,1 (t) does not meet the two conditions of IMF, then go back to Step A and repeat k times, that is, c n+1,k (t) = c n+1,k-1 (t) - m n+1,k (t), until c n+1,k (t) is an IMF component, denoted as imf n+1 (t) = c n+1,k (t);

[0103] Step C: Determine whether the modal component is the first-order modal component IMF1; if so, jump to Step D; if not, jump to Step E;

[0104] Step D: Determine whether the modal component contains a high-frequency gap component; if not, jump to Step E; if so, perform the CEEMD algorithm; that is, determine whether imf 1 (t) contains a high-frequency gap component; preferably, by detecting the fluctuation of the distance between the extreme points of imf 1 (t), determine whether there is a jump in the distance between the extreme points;

[0105] Step E: Calculate the residual component; when the residual component meets the cut-off condition, the algorithm ends; when the residual component does not meet the cut-off condition, repeat Step A and Step B until the residual component meets the cut-off condition; preferably, the cut-off condition includes: the residual component is a constant or a monotonic function; therefore, Step E is detailed as: calculate the residual component Let n = n + 1, and repeat Steps A and B until the residual component r n+1 (t) is a constant or a monotonic function, and the algorithm ends.

[0106] In this embodiment, specifically, the CEEMD algorithm includes:

[0107] Step D1: Add I pairs of noise components with opposite polarities to the target radiated noise signal to form 2I groups of noisy signals; that is, add white noise components to r n (t) in pairs,

[0108] Step D2: Perform EMD decomposition on the noisy signals respectively to obtain 2I groups of modal components, that is, obtain the corresponding modal components; that is, obtain the IMF components and

[0109] Step D3: Add the corresponding 2I groups of modal components and then perform lumped averaging to obtain the nth-order modal component; that is, perform lumped averaging on and after adding them to obtain the nth-order IMF component

[0110] Step D4: Calculate the residual component; when the residual component meets the cut-off condition, the algorithm ends; when the residual component does not meet the cut-off condition, repeat Step D2 and Step D3 until the residual component meets the cut-off condition; that is, calculate the residual component Repeat Step D2 and Step D3 until the residual component r n+1 (t) is a constant or a monotonic function, and the algorithm ends.

[0111] During the screening process of the ASEMD decomposition algorithm, it is necessary to limit the number of screenings. The commonly used screening termination criterion is the Cauchy-like convergence criterion:

[0112]

[0113] where T is the signal time length; when SD is between 0.2 and 0.3, the screening process terminates.

[0114] In this embodiment, specifically, the step S12 includes:

[0115] Demodulation: Square detection and low-pass filtering are performed on n modal components to obtain an envelope signal, and the envelope signal is subjected to FFT processing to obtain the DEMON spectrum P of each modal component IMFi , i = 1, 2,..., n.

[0116] In this embodiment, specifically, the step S13 includes:

[0117] Calculating the weighting coefficient: The DEMON spectra of each modal component are passed through an α bidirectional filter to obtain the DEMON spectrum weighting coefficients of each modal component; that is, the DEMON spectrum of the i-th IMF component is passed through an α bidirectional filter, that is

[0118]

[0119]

[0120]

[0121] where α is selected as 0.01, the adaptive threshold c 1 is a constant, then the number of line spectra L exceeding the threshold is obtained i , with L i 2 being the DEMON spectrum weighting coefficient of this IMF component.

[0122] In this embodiment, specifically, the step S14 includes:

[0123] DEMON spectrum weighted fusion; The line spectra of each modal component are calculated through the DEMON line spectrum estimation formula; the DEMON line spectrum estimation formula is: Here c 2 is a constant, and 1 ≤ c 2 ≤ c 1 ;

[0124] The line spectra of each modal component are multiplied by their corresponding DEMON spectrum weighting coefficients, and then summed to obtain the fused DEMON spectrum; that is The line spectrum and its corresponding coefficient L i 2 Multiply them and then sum to obtain the fused DEMON spectrum P.

[0125] Figure 5 The simulation results of the improved DEMON spectrum based on ASEMD are given. There are 6 obvious spectral lines in the figure. For the simulated ship target signal, IMF1, IMF2, and IMF3 have strong modulation effects. Therefore, IMF1, IMF2, and IMF3 can be selected as the main sources for obtaining the DEMON spectrum. Compared with the traditional DEMON method, it is no longer necessary to design the demodulation bandwidth, the number of demodulation filters, and other filter parameters.

[0126] In this embodiment, specifically, in step S2, it is first necessary to explain that:

[0127] With the development of noise reduction technology, the radiated noise of underwater targets is easily submerged in the ambient noise. It is very difficult to detect targets at low signal-to-noise ratios relying only on the detection of a single feature. While adopting the DEMON spectrum feature detection, the present invention extracts the Mel-frequency cepstral coefficient feature (MFCC) based on the human ear auditory characteristics for the target radiated noise signal.

[0128] The MFCC feature is a classic human ear auditory perception feature. According to relevant research on the auditory mechanism, it is found that the human ear has different sensitivities to sound waves of different frequencies. The Mel frequency scale reflects the non-linear characteristics of human ear frequency perception and is more in line with the human ear auditory characteristics. MFCC is the cepstral parameter extracted in the Mel frequency domain.

[0129] Among them, the specific relationship between the Mel frequency and the actual frequency can be expressed as:

[0130] Mel(f) = 2595lg(1 + f / 700)

[0131] In the formula, f is the actual frequency. The critical frequency bandwidth changes with the frequency and is consistent with the growth of the Mel frequency. It is roughly linearly distributed below 1000 Hz with a bandwidth of about 100 Hz, and grows logarithmically above 1000 Hz, which makes the human ear more sensitive to low-frequency signals than high-frequency signals.

[0132] The calculation process of MFCC is as Figure 6 shown. In this embodiment, specifically, step S2 includes:

[0133] Perform pre-emphasis, framing, and windowing processing on the target radiated noise signal;

[0134] The time-domain framed signal is subjected to discrete Fourier transform to obtain the energy spectrum: that is, the time-domain framed signal x(n) is subjected to N-point discrete Fourier transform to obtain the energy spectrum:

[0135] p(f) = |X(f) 2 | = |FFT(x(n))| 2 ;

[0136] The energy spectrum is filtered through a Mel frequency filter bank to obtain the Mel spectrum; that is:

[0137]

[0138] Take the logarithmic energy of the Mel spectrum and perform discrete cosine transform to obtain the MFCC coefficients; it should be noted that, in order to make the result more robust to noise and spectral estimation errors, generally take the logarithmic energy of the Mel spectrum, and its formula can be expressed as:

[0139] E′(m) = lnE(m)

[0140] Then perform discrete cosine transform on the above logarithmic energy spectrum to obtain the MFCC coefficients, and its formula is expressed as follows:

[0141]

[0142] In this embodiment, specifically, the multi-feature collaborative fusion network, as Figure 7 shown, includes:

[0143] Two parallel residual networks (ResNet) in the backbone network form a dual-branch distributed network, which are respectively used as the DEMON spectral feature and MFCC feature extractors;

[0144] The extractor contains 4 convolutional blocks and an attention module. Each convolutional block consists of a BN layer, a convolutional layer, an average pooling layer, and an activation function. The convolutional layer contains 128 convolutional kernels with a size of [5×5] and a stride of 2;

[0145] At the same time, a 1×1 convolutional layer is in parallel with the fourth convolutional block to complete the extraction of the two features;

[0146] The extracted feature vectors enter the attention module, and the softmax activation function is used to convert them into numbers between 0 and 1 as the weights of the features. The two weights are connected to form a feature selection matrix and multiplied by the features to achieve feature selection, so as to obtain the fused features.

[0147] The key modules are as follows:

[0148] 1) Multi-scale feature collaboration

[0149] Considering that the global features and local features have their respective advantages and disadvantages, a multi-scale feature collaborative learning strategy is adopted, and the cross-modal target features are processed by means of horizontal multi-scale segmentation. The multi-scale feature vector of the target radiated noise is obtained by means of multi-scale block pooling, and more discriminative target information can be obtained by means of multi-scale feature collaborative learning.

[0150] 2) Multi-level feature collaboration

[0151] The present invention adopts a multi-level feature collaborative learning method. In order to avoid increasing a large amount of computational complexity and feature dimensions, only the feature map obtained in the Stage3 stage is considered for the shallow-layer information. The number of channels of this feature map is increased using 1×1 convolution, and then this shallow-layer feature is cascaded with the deep-layer feature in the Stage4 stage and sent into the subsequent network.

[0152] 3) Distributed multi-dimensional information collaborative fusion

[0153] Based on the collaborative feature extraction of the dual-branch neural network, an attention mechanism is introduced. This mechanism is used to screen the features, and the weight of the important features is increased during fusion, thereby achieving the effect of multi-dimensional information complementarity and improving the utilization rate of target information.

[0154] In this embodiment, specifically, the Ret residual detection network includes:

[0155] The Ret residual detection network is composed of two residual modules, a fully connected layer, and an output layer;

[0156] The residual module is composed of a BN layer, a convolutional layer containing 256 convolutional kernels with a size of [3×3] and a stride set to 2, a max pooling layer, and a ReLU activation function.

[0157] The above embodiments only represent the specific implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the protection scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the technical solution of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application.

[0158] This background technology section is provided to generally present the context of the present invention. The work of the currently named inventors, to the extent described in this background technology section, and aspects of this section that are not prior art at the time of filing this application are neither expressly nor impliedly admitted to be prior art to the present invention.

Claims

1. A passive sonar intelligent detection method based on multi-feature fusion, characterized in that, it includes: Step S1: Based on the ASEMD decomposition algorithm, extract the DEMON spectrum features of the target radiated noise signal; Step S2: Extract the MFCC features of the target radiated noise signal based on the human ear auditory characteristics; Step S3: Based on the multi-feature collaborative fusion network, fuse the DEMON spectrum features and the MFCC features; Step S4: Send the fused features into the Ret residual detection network for target detection; The said Step S1 includes: Step S11: Decompose the target radiated noise signal into n modal components by the ASEMD decomposition algorithm; Step S12: Demodulate each of the n modal components respectively to obtain the DEMON spectrum corresponding to each modal component; Step S13: Calculate the weighted coefficients corresponding to each DEMON spectrum; Step S14: Fuse each DEMON spectrum with its corresponding weighted coefficient to obtain the fused DEMON spectrum; The said ASEMD decomposition algorithm includes: Step A: Perform EMD modal decomposition on the input target radiated noise signal; Step B: Extract the modal components from the decomposition result; Step C: Determine whether the modal component is the first-order modal component IMF1; if so, jump to Step D; if not, jump to Step E; Step D: Determine whether the modal component contains high-frequency gap components; if not, jump to Step E; if so, perform the CEEMD algorithm; Step E: Calculate the remaining component; when the remaining component meets the cut-off condition, the algorithm ends; when the remaining component does not meet the cut-off condition, repeat Steps A and B until the remaining component meets the cut-off condition; The said multi-feature collaborative fusion network includes: Two parallel residual networks in the backbone network form a double-branch distributed network, which are respectively used as the DEMON spectrum feature and MFCC feature extractors; The said extractor contains 4 convolutional blocks and an attention module, where each convolutional block consists of a BN layer, a convolutional layer, an average pooling layer and an activation function. The convolutional layer contains 128 convolutional kernels with a size of [5×5] and a stride of 2; At the same time, a 1×1 convolutional layer is in parallel with the fourth convolutional block to complete the extraction of the two features; The extracted feature vectors enter the attention module, and are converted into numbers from 0 to 1 using the softmax activation function as the weight of the feature. The two weights are connected to form a feature selection matrix and multiplied by the feature to realize the selection of the feature, so as to obtain the fused feature.

2. A passive sonar intelligent detection method based on multi-feature fusion according to claim 1, characterized in that, the said CEEMD algorithm includes: Step D1: Add a pair of noise components with opposite polarities to the target radiated noise signal to form a group of noise-added signals; Step D2: Perform EMD decomposition on the noisy signal respectively to obtain groups of modal components; Step D3: After adding the corresponding group modal components and performing lumped averaging, the n n-th order modal component is obtained; Step D4: Calculate the remaining component. When the remaining component meets the cut-off condition, the algorithm ends; when the remaining component does not meet the cut-off condition, repeat Steps D2 and D3 until the remaining component meets the cut-off condition.

3. A passive sonar intelligent detection method based on multi-feature fusion according to claim 2, characterized in that, the said cut-off condition includes: The remaining component is a constant or a monotonic function.

4. A passive sonar intelligent detection method based on multi-feature fusion according to claim 1, characterized in that, the said Step S12 includes: Square detect and low-pass filter the n modal components to obtain the envelope signal, and perform FFT processing on the envelope signal to obtain the DEMON spectrum of each modal component .

5. A passive sonar intelligent detection method based on multi-feature fusion according to claim 1, It is characterized in that The step S13 includes: After passing the DEMON spectra of each modal component through a two-way filter, the weighted coefficients of the DEMON spectra of each modal component are obtained; The step S14 includes: Calculating the line spectra of each modal component through the DEMON line spectrum estimation formula; Multiplying the line spectra of each modal component by their corresponding DEMON spectrum weighting coefficients, and then summing them to obtain the fused DEMON spectrum.

6. A passive sonar intelligent detection method based on multi-feature fusion according to claim 1, It is characterized in that The step S2 includes: Performing pre-emphasis, framing, and windowing processing on the target radiated noise signal; Obtaining the energy spectrum after performing discrete Fourier transform on the time-domain framed signal; Filtering the energy spectrum through a Mel frequency filter bank to obtain the Mel spectrum; Taking the discrete cosine transform of the logarithmic energy of the Mel spectrum to obtain the MFCC coefficients.

7. A passive sonar intelligent detection method based on multi-feature fusion according to claim 1, It is characterized in that The Ret residual detection network includes: The Ret residual detection network consists of two residual modules, a fully connected layer, and an output layer; The residual module consists of a BN layer, a convolutional layer with 256 convolutional kernels, a size of [3×3], a stride set to 2, a max pooling layer, and a ReLU activation function.

Images