Intelligent seabed sediment discrimination method based on active target echoes

Abstract

The invention relates to a seabed sediment intelligent discrimination method based on active target echoes, and the method comprises the following steps: 1, building different types of seabed sediment echo simulation data sets based on a normal wave model; step 2, constructing a seabed sediment classification network model; establishing a mapping relation between the echo data and the type of the seabed sediment by using a neural network; step 3, training the seabed sediment classification network; and 4, preprocessing actual echo data to be predicted according to the same preprocessing mode, and inputting the data into the network to obtain an estimation result of the type of the seabed sediment. According to the method, the normal wave approximation under the distance-independent shallow sea waveguide condition is utilized, the seabed sediment needing to be represented by a plurality of parameters is simplified to be represented by a single parameter of the normal wave number capable of being propagated in the waveguide, the parameter quantity is greatly simplified, and the automatic classification of the seabed sediment is realized by utilizing the deep learning technology.

Description

Technical Field:

[0001] The present invention belongs to the technical field of underwater acoustic engineering, and particularly relates to an intelligent discrimination method for seabed sediment based on active target echoes. Background Art:

[0002] Active sonar is one of the main devices for underwater detection. By transmitting signals and receiving echoes scattered by targets, the received echoes are analyzed to achieve the detection and identification of underwater targets. When sound waves propagate underwater, they will be reflected by the seabed and the sea surface. According to the physical and chemical properties of sediments, the seabed sediment can be divided into nine different types. Different types of seabed sediment will have different effects on active sonar detection. For example, the density and compressibility of the sediment will affect the sound speed, thereby affecting the propagation time of sound waves; different sediments have different acoustic properties, and the degrees of reflection, absorption, and scattering of sound waves are also different. These properties will directly affect the propagation path, signal intensity, and detection effect of sound waves. Therefore, obtaining seabed characteristics and judging the type of seabed sediment play an important role in improving the detection, positioning, and other performances of active sonar systems.

[0003] In practice, it is difficult to obtain seabed characteristics. In the acoustic field, the seabed acoustic parameters such as the density, longitudinal wave speed, transverse wave speed, and sound speed attenuation of the seabed sediment layer are mainly obtained through the method of inverse acoustic parameter inversion, including matched field inversion, arrival time analysis method, modal dispersion technology, and acoustic field interference method, etc. However, there are couplings and correlations among multiple seabed acoustic parameters, which increase the difficulty of inversion. Some inversion methods rely on prior information of environmental parameters, and inaccurate prior information will also cause the inversion results to deviate from the true values. The marine environment is complex and changeable, and the errors of acoustic propagation models will also lead to model mismatches, all of which will affect the judgment of seabed acoustic parameters and seabed sediment types.

[0004] In recent years, deep learning technology has developed rapidly and has been widely applied in the fields of computer vision, natural language processing, speech signal processing, etc., and has achieved excellent performance. Deep learning technology extracts and maps input data through a deep neural network model at multiple layers, extracts high-dimensional features, and can automatically establish the mapping relationship between input and output. How to apply it to sonar signal processing is a technical problem that needs to be solved in this field. Summary of the Invention:

[0005] The technical problem to be solved by the present invention is to provide an intelligent discrimination method for seabed sediment based on active target echoes, which can use deep learning technology to discriminate the type of seabed sediment.

[0006] The technical solution of the present invention is to provide an intelligent discrimination method for seabed sediment based on active target echoes, including the following steps,

[0007] Step 1, construct echo simulation datasets of different types of seabed sediments based on the normal mode model;

[0008] Step 2, construct a seabed sediment classification network model. Use a neural network to establish the mapping relationship between echo data and seabed sediment types;

[0009] Step 3, train the seabed sediment classification network;

[0010] Step 4, for the actual echo data to be predicted, after preprocessing it in the same preprocessing manner and inputting it into the network, the estimation result of the seabed sediment type can be obtained.

[0011] Preferably, the specific operations of Step 1 are as follows.

[0012] Step 1.1, select the seabed sediment with the maximum sound speed, and use the KRAKEN model to calculate the eigenvalues, phase velocities, group velocities, and mode depth functions of all normal modes in this environment;

[0013] Step 1.2, traverse nine types of seabed sediments. For the current seabed sediment D i , i = 0, 1, …, 8, select the normal modes with phase velocities less than the seabed sound speed to participate in the subsequent simulation calculations, and count the number of normal modes that meet this condition as M i ;

[0014] Step 1.3, traverse nine types of seabed sediments, and use the normal mode model to generate echo data corresponding to different seabed sediments. For the current seabed sediment D i , its echo is generated according to the following formula:

[0015]

[0016] where p i represents the complex sound pressure of the echo corresponding to the i-th sediment, r is the distance from the sound source to the receiver, z is the receiving depth, Q is the sound source intensity, z s is the sound source depth, M i is the number of normal modes that can propagate in the waveguide under the condition of the i-th sediment, Ψ m is the mode depth function of the m-th normal mode, is the horizontal beam of the m-th normal mode, and α m is the attenuation coefficient of the m-th normal mode;

[0017] The echo label corresponding to the i-th sediment is Label = [i], i = 0, 1, …, 8;

[0018] After the simulation datasets are generated, divide the training set and the test set in a ratio of 9:1.

[0019] Preferably, the specific operations of Step 2 are as follows.

[0020] Step 2.1: Preprocess the simulated active echo data to obtain the preprocessed active echo feature spectrum X i , i = 1, …, N, where N is the total number of samples;

[0021] Step 2.2: Use the preprocessed feature spectrum as the network input. The seabed sediment classification network includes a feature extraction backbone network and several fully connected layers. After the features output by the backbone network pass through multiple fully connected layers, the seabed sediment is classified through a classification loss function. The cross-entropy loss function is adopted:

[0022]

[0023] Among them, the backbone network can be ViT, ResNet, IncenptionNet or UNet.

[0024] Preferably, the preprocessing can be time-frequency analysis, auditory perception analysis or directly normalizing the original one-dimensional echo sequence.

[0025] Preferably, the time-frequency analysis includes any one of STFT, WVD, and CWT; the auditory perception analysis includes any one of MFCC, PNCC, and GFCC.

[0026] Preferably, the training in Step 3 includes two stages:

[0027] The first-stage training uses the simulated data to train the seabed sediment classification network until the model converges;

[0028] The second-stage training uses the actual data to fine-tune the model parameters. The actual data uses the same preprocessing method as the simulated data.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] (1) The present invention uses the normal mode approximation under the range-independent shallow water waveguide condition to simplify the seabed sediment that needs to be characterized by multiple parameters into being characterized only by a single variable of the normal mode order that can propagate in the waveguide, greatly simplifying the number of physical model parameters required for seabed sediment classification compared with the traditional method.

[0031] (2) The present invention uses the powerful non-linear feature representation ability of deep learning technology to automatically establish the mapping relationship from the echo data to the seabed sediment type, integrates the physical model knowledge into the generation of deep learning training data, and realizes the seabed sediment estimation under the guidance of expert knowledge. Description of the drawings:

[0032] Figure 1 is the flowchart of the present invention.

[0033] Figure 2 It is a calculation example of the propagable normal mode numbers corresponding to different seabed sediments.

[0034] Figure 3 It is a schematic diagram of the network structure for seabed sediment classification. Specific implementation manner:

[0035] The present invention will be further described in detail below in conjunction with the accompanying drawings in terms of the specific implementation manner:

[0036] Figure 1 This is the working flowchart of the intelligent discrimination method for seabed sediments based on the active target echo of the present invention, and its specific steps are as follows:

[0037] Step 1: Construct an echo simulation data set of different types of seabed sediments based on the normal mode model. The normal mode approximation under the shallow water waveguide condition independent of distance can be expressed by the following formula

[0038]

[0039] where p represents the complex sound pressure of the echo, r is the distance from the sound source to the receiver, z is the receiving depth, Q is the sound source intensity, z s is the sound source depth, M is the number of propagable normal modes in the waveguide, Ψ m is the mode depth function of the m-th order normal mode, is the horizontal beam of the m-th order normal mode, and α m is the attenuation coefficient of the m-th order normal mode.

[0040] For the sea area to be detected, the sea depth is known. Under the condition of fixed sea depth, the following approximations are considered:

[0041] (1) The influence of seabed sediment changes on the horizontal beams of each order of normal modes can be ignored;

[0042] (2) The influence of seabed sediment changes on the modal depth functions of each order of normal modes can be ignored.

[0043] Under the above approximations, the change of seabed sediment can be considered to only cause the change of M, that is, the number of accumulated normal modes in the case of long-distance propagation is different. Therefore, different seabed sediments can be characterized by the corresponding normal mode number M.

[0044] The calculation method of the normal mode numbers corresponding to different seabed sediments is as follows:

[0045] 1.1 First, select the seabed sediment with the maximum sound speed, and use the KRAKEN model to calculate the eigenvalues, phase velocities, group velocities, and mode depth functions of all normal modes in this environment.

[0046] 1.2 Traverse nine types of seabed sediments, and for the current seabed sediment D i, for \(i = 0, 1, \cdots, 8\), select the normal modes whose phase velocity is less than the seabed sound speed to participate in the subsequent simulation calculation. The number of normal modes that meet this condition is the \(M\) corresponding to the current seabed sediment i .

[0047] Figure 2 A calculation example is given. Assume that the current seabed sound speed is \(1560m / s\). For the current seabed sediment, the eigenvalues, attenuation coefficients, phase velocities, and group velocities of all normal modes calculated using the KRAKEN model are as Figure 3 shown. The first column is the number of propagable normal modes, the third column is the attenuation coefficient, and the fourth column is the phase velocity. It can be seen that for the first six normal modes, when the phase velocity of the normal mode is less than the seabed sound speed, the attenuation coefficient of the normal mode is relatively small, at the order of \(E - 30\), and the corresponding normal mode can propagate over a long distance; for the seventh normal mode, after the phase velocity of the normal mode is greater than the seabed sound speed, the attenuation coefficient of the normal mode increases significantly, reaching the order of \(E - 10\sim E - 3\), and the corresponding normal mode attenuates quickly and cannot propagate over a long distance. Therefore, for the current seabed sediment, the number of propagable normal modes is 6.

[0048] Get \(M\) i After that, use the above formula to generate the simulation echo data corresponding to different seabed sediments. The echo label corresponding to the \(i\) - th sediment is Label = \([i]\), \(i = 0, 1, \cdots, 8\). After the simulation data set is generated, divide the training set and the test set according to a ratio of 9:1.

[0049] Step 2: Construct a seabed sediment classification network model, and use a neural network to establish a mapping relationship between the echo data and the seabed sediment types. Figure 3 is a schematic diagram of the deep - learning network model for seabed sediment type discrimination provided by the present invention. The specific steps are as follows:

[0050] 2.1 Pre - process the simulated active echo data to obtain the pre - processed active echo feature spectrum \(X\) i , for \(i = 1, \cdots, N\), where \(N\) is the total number of samples. The pre - processing can be time - frequency analysis, such as STFT, WVD, CWT, etc.; or auditory perception analysis, such as MFCC, PNCC, GFCC, etc.; or directly perform normalization processing on the original one - dimensional echo sequence.

[0051] 2.2 Use the pre - processed feature spectrum as the network input. The seabed sediment classification network includes a feature extraction backbone network and several fully - connected layers. The backbone network can be ViT, ResNet, IncenptionNet, UNet, etc. After the features output by the backbone network pass through multiple fully - connected layers, the seabed sediment classification is performed through a classification loss function. The cross - entropy loss function is used:

[0052]

[0053] Step 3: Train the seabed sediment classification network, and the training includes two stages:

[0054] In the first-stage training, use the simulation data to train the seabed sediment classification network until the model converges;

[0055] In the second-stage training, use the actual data to fine-tune the model parameters, and the actual data adopts the same preprocessing method as the simulation data.

[0056] Step 4: For the actual echo data to be predicted, after preprocessing it in the same preprocessing manner and inputting it into the network, the estimation result of the seabed sediment type can be obtained.

[0057] The present invention first generates simulation echoes under different seabed sediment conditions based on the normal mode model, then constructs a seabed sediment classification neural network model, uses deep learning technology to establish a mapping relationship between the echo and the seabed sediment, and uses the simulation data to train the seabed sediment classification network. After the model converges, use the actual data to fine-tune the model parameters.

[0058] The present invention uses the normal mode approximation under the shallow water waveguide condition independent of distance to simplify the seabed sediment, which needs to be characterized by multiple parameters, into a single parameter characterized only by the number of normal modes that can propagate in the waveguide, greatly simplifying the number of parameters, and uses deep learning technology to realize the automatic classification of seabed sediment.

[0059] The above is only an illustration of the preferred embodiments of the present invention, but it should not be construed as a limitation of the claims. All equivalent process transformations made using the specification of the present invention are included in the patent protection scope of the present invention.

Claims

1. An intelligent discrimination method for seabed sediment based on active target echoes, characterized in that: It includes the following steps: Step 1: Construct echo simulation datasets of different types of seabed sediments based on the normal mode model; Step 2: Construct a seabed sediment classification network model and use a neural network to establish a mapping relationship between echo data and seabed sediment types; Step 3: Train the seabed sediment classification network; Step 4: For the actual echo data to be predicted, after preprocessing it in the same preprocessing manner, input it into the network, and the estimated result of the seabed sediment type can be obtained.

2. The intelligent discrimination method of seabed sediment based on active target echo according to claim 1, wherein: The specific operations of Step 1 are as follows: Step 1.1: Select the seabed sediment with the maximum sound speed, and use the KRAKEN model to calculate the eigenvalues, phase velocities, group velocities, and mode depth functions of all normal modes in this environment; Step 1.2, traverse nine types of seabed sediments, and for the current seabed sediment D i , i = 0, 1, …, 8, select the normal modes with phase velocities less than the seabed sound speed to participate in the subsequent simulation calculations, and count the number of normal modes that meet this condition as M i ; Step 1.3, traverse nine types of seabed sediments, use the normal mode model to generate echo data corresponding to different seabed sediments, and for the current seabed sediment D i , its echo is generated according to the following formula: where p i represents the complex sound pressure of the echo corresponding to the i-th bottom substrate, r is the distance from the sound source to the receiver, z is the receiving depth, Q is the sound source intensity, z s is the sound source depth, M i is the number of normal modes that can propagate in the waveguide under the i-th bottom substrate condition, Ψ m is the mode depth function of the m-th normal mode, k rm is the horizontal beam of the m-th normal mode, α m is the attenuation coefficient of the m-th normal mode; The echo label corresponding to the i-th type of sediment is Label = [i], where i = 0, 1, …, 8; After the simulation dataset is generated, divide it into a training set and a test set according to a ratio of 9:

1.

3. The intelligent discrimination method of submarine bottom sediment based on active target echo according to claim 1, characterized in that: The specific operations of Step 2 are as follows: Step 2.1, preprocess the simulated active echo data to obtain the preprocessed active echo feature spectrum X i , i = 1, L, N, where N is the total number of samples; Step 2.2: Use the preprocessed feature spectrum as the network input. The seabed sediment classification network includes a feature extraction backbone network and several fully connected layers. After the features output by the backbone network pass through multiple fully connected layers, the seabed sediment is classified through a classification loss function, and the cross-entropy loss function is adopted: Among them, the backbone network can be ViT, ResNet, IncenptionNet, or UNet.

4. The intelligent discrimination method of seabed sediment based on active target echo according to claim 3, characterized in that: The preprocessing can be time-frequency analysis, auditory perception analysis, or directly normalizing the original one-dimensional echo sequence.

5. The intelligent discrimination method of seabed sediment based on active target echo according to claim 1, characterized in that: The training in Step 3 includes two stages: In the first stage of training, use the simulation data to train the seabed sediment classification network until the model converges; In the second stage of training, use the actual data to fine-tune the model parameters, and the actual data adopts the same preprocessing method as the simulation data.

6. The intelligent discrimination method for seabed sediment based on active target echo according to claim 3, wherein: The time-frequency analysis includes any one of STFT, WVD, and CWT; the auditory perception analysis includes any one of MFCC, PNCC, and GFCC.

Images