2D-CNN network wake flow electromagnetic field detection method based on wavelet transform time-frequency spectrum characteristics

Abstract

The invention discloses a 2D-CNN network wake flow electromagnetic field detection method based on wavelet transform time-frequency spectrum characteristics. The method comprises the steps of (1) generating an underwater vehicle wake flow electromagnetic field calculation data set through a multi-physics field coupling method of hydrodynamics and electromagnetism, (2) extracting time-frequency spectrum features of a wake flow electromagnetic field through wavelet transform, and (3) developing a wake flow electromagnetic field detection model based on a 2D-CNN deep neural network. And training the model by using the data after the wavelet transform for wake flow electromagnetic field detection. According to the 2D-CNN network wake flow electromagnetic field detection method based on the wavelet transform time-frequency spectrum characteristics, the detection capability of the wake flow electromagnetic field can be effectively improved, a new technical means is provided for detection of the wake flow electromagnetic field of an underwater vehicle, and the method has a good application prospect in the field of marine target electromagnetic field monitoring and early warning.

Description

Technical Field

[0001] The invention relates to the field of electromagnetism, and in particular to a 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics. Background Art

[0002] The wake formed by the underwater vehicle when it sails underwater can last for a long time and propagate over a long distance. The wake is only limited by the hydrodynamic behavior of the underwater vehicle and has nothing to do with its own construction materials. Seawater itself contains a large number of charged ions. When the underwater vehicle causes the seawater to move, the charged ions receive external excitation. Under the action of the geomagnetic field, the local electrical neutrality will be broken, and an induced current will be formed within the range of the entire flow field distribution, thereby stimulating the electromagnetic field distribution in space. Wake electromagnetic field detection is a new type of underwater vehicle detection technology and a cutting-edge development direction, and it is also an important supplement to the existing non-acoustic detection technology. Therefore, exploring the wake electromagnetic field detection method is of great significance and potential application value.

[0003] Although electromagnetic field sensing technology has made great progress, and ultra-low noise electromagnetic field sensors have been realized at home and abroad, it is impossible to directly apply advanced electromagnetic field sensors to underwater vehicle wake electromagnetic field detection. This is because the extremely weak wake electromagnetic field of underwater vehicles will be submerged by the noise due to the influence of the background electromagnetic field noise of the ocean, making it difficult to detect targets. Research on wake electromagnetic fields has just started at home and abroad, and there are few public reports on its detection methods. Traditional underwater vehicle electromagnetic field detection technology mainly relies on simplified models, such as equating underwater vehicles to magnetic dipole or electric dipole models, and using the difference between this model and the noise model to extract the weak electromagnetic field signal of underwater vehicles from the background noise through algorithms. However, the generation mechanism of the wake electromagnetic field of underwater vehicles is relatively complex, and it is necessary to use the mutual coupling of hydrodynamics and electromagnetics for calculation and solution. There is no effective simplified electromagnetic field model, so the traditional underwater vehicle electromagnetic field detection technology cannot be applied to wake electromagnetic field signal detection. Summary of the invention

[0004] In order to solve the international problem of underwater vehicle wake electromagnetic field detection, the present invention proposes a 2D-CNN (two-dimensional convolutional neural network) wake electromagnetic field detection method based on wavelet transform time-frequency characteristics.

[0005] The method of the present invention uses deep learning methods as detection means, velocity field simulation as simulation basis, and electromagnetic field simulation as core data source, which greatly increases the amount of data that can be used for model training and improves detection accuracy.

[0006] More specifically, the present invention improves the detection capability of wake electromagnetic fields from three aspects: first, by training and optimizing deep learning models with a large amount of underwater vehicle wake electromagnetic field calculation data, the problem of constructing complex wake electromagnetic field detection models is solved; second, by generating underwater vehicle wake electromagnetic field calculation data sets through a multi-physics field coupling method of hydrodynamics and electromagnetics, the problem of creating deep learning training samples is solved; third, by extracting the time-frequency spectrum characteristics of the underwater vehicle wake electromagnetic field through wavelet transform, the problem of feature representation of deep learning in wake electromagnetic field detection is solved.

[0007] Specifically, the present invention provides a 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics, the method comprising:

[0008] Step (1) generating a computational dataset of electromagnetic field of the wake of a submersible by means of a multi-physics coupling method of hydrodynamics and electromagnetics,

[0009] The step (1) comprises:

[0010] (1.1) Construct a geometric model in the dynamic field simulation software and set the calculation domain of the geometric model;

[0011] (1.2) Mesh the set computational domain;

[0012] (1.3) setting boundary conditions for the geometric model, wherein the boundary conditions at least include a velocity inlet, a pressure outlet, a symmetry surface, and a wall surface;

[0013] (1.4) Solve the velocity field distribution in the calculation domain in the dynamic field simulation software;

[0014] (1.5) Based on the obtained velocity field distribution, the wake electric field E and wake magnetic field B at any point in the calculation domain are calculated according to the following formula:

[0015] E(r)=v0(r)×B(r)=v0(r)×(B E (r)+B i (r))≈v0(r)×B E (r)(1)

[0016]

[0017] In equations (1) and (2), the position vector r' and the position vector r respectively represent the position of the source point and the position of the point to be measured, where J(r') represents the current density at position r', V represents the integrated volume, v0 represents the water flow velocity of the point to be measured calculated according to the velocity field, μ represents the magnetic permeability of seawater, σ represents the electrical conductivity of seawater, and B E represents the Earth's magnetic field;

[0018] (1.6) Modify the input parameters and model parameters, repeat the above steps, obtain the wake electric field E and wake magnetic field B distribution under different input parameters and model parameter conditions, obtain multiple sets of wake electric field E and wake magnetic field B data, and form a data set;

[0019] Step (2) takes the entire computational domain as the measurement area, selects a number of measurement points from the measurement area, obtains data on the change of the electric and magnetic fields at each measurement point over time under different conditions from the data set, performs wavelet transformation on the obtained electric and magnetic field data, and obtains a wavelet transformed data set;

[0020] Step (3) constructs a 2D-CNN deep neural network, divides the constructed data set into a training set and a test set, inputs the data in the training set into the 2D-CNN deep neural network, trains the 2D-CNN deep neural network and tests it using the test set to obtain the trained 2D-CNN deep neural network, and uses the obtained trained 2D-CNN deep neural network for wake electromagnetic field detection.

[0021] In a preferred implementation, the power field simulation software is CFD software.

[0022] In another preferred implementation, the computational domain is a rectangular region, and a SUBOFF model is set in the computational domain to simulate an underwater vehicle.

[0023] In another preferred implementation, the current density J is calculated by the following formula:

[0024] J=σv0×B E .

[0025] Where σ represents the conductivity of seawater, v0 represents the water velocity at the test point calculated based on the velocity field, and B E Represents the Earth's magnetic field.

[0026] In another preferred implementation, the wavelet transform is performed using the following formula:

[0027]

[0028] Where Ψ(t) represents the wavelet basis function; Ψ ω,τ (t) represents the sub-wavelet generated by scaling and translating the wavelet basis function; W f (ω,τ) is the result of wavelet transform of signal f(t).

[0029] In another preferred implementation, the computational domain uses local meshing and interpolation methods to calculate the wake electromagnetic field.

[0030] In another preferred implementation, the network model includes a convolution module and a fully connected module, the convolution module includes 7 convolution layers and RELU layers, 3 batch normalization layers, and 4 pooling layers; the fully connected classifier module includes 1 flattening layer, 3 fully connected layers, and 2 RELU layers.

[0031] On the other hand, the present invention provides a data set acquisition method for wake electromagnetic field detection model training,

[0032] The method comprises:

[0033] (1.1) Construct a geometric model in the dynamic field simulation software and set the calculation domain of the geometric model;

[0034] (1.2) In the constructed geometric model, mesh the computational domain;

[0035] (1.3) setting boundary conditions for the geometric model, wherein the boundary conditions at least include a velocity inlet, a pressure outlet, a symmetry surface, and a wall surface;

[0036] (1.4) Solve the velocity field distribution in the calculation domain in the dynamic field simulation software;

[0037] (1.5) Based on the obtained velocity field distribution, the wake electric field E and wake magnetic field B at any point in the calculation domain are calculated according to the following formula:

[0038] E(r)=v0(r)×B(r)=v0(r)×(B E (r)+B i (r))≈v0(r)×B E (r)(4)

[0039]

[0040] In equations (4) and (5), the position vector r' and the position vector r represent the position of the source point and the position of the measured point, respectively. J(r') represents the current density at position r', V represents the volume, v0 represents the velocity calculated according to the velocity field, μ represents the magnetic permeability of seawater, σ represents the electrical conductivity of seawater, and B E Represents the Earth's magnetic field.

[0041] (1.6) Repeat the above steps to modify the input parameters and model parameters, obtain the distribution of the wake electric field E and the wake magnetic field B under different input parameters and model parameter conditions, and obtain multiple sets of wake electric field E and wake magnetic field B data to form a data set.

[0042] The present invention aims to realize the detection of wake electromagnetic field and has the following advantages:

[0043] First: The present invention adopts the multi-physics field coupling method of hydrodynamics and electromagnetism to generate a calculation data set of the electromagnetic field of the underwater vehicle wake, successfully solving the problem of creating deep learning training samples.

[0044] First, multi-physics coupling provides a more comprehensive and realistic simulation environment, so that the generated data set can more accurately reflect the real characteristics of the wake electromagnetic field of underwater vehicles, thereby improving the training effect of deep learning models. Secondly, this method can flexibly adjust various parameters and conditions to generate a variety of scenarios and samples, which improves the wide applicability of wake electromagnetic field training data and the generalization ability of the model. In addition, this automated generation technology significantly reduces the time and cost of data acquisition, while ensuring the high quality and consistency of the data, further promoting the development and optimization of wake electromagnetic field detection models.

[0045] Second, the present invention extracts the time-frequency spectrum characteristics of the wake electromagnetic field of the underwater vehicle through wavelet transform, which effectively solves the problem of feature representation of deep learning in wake electromagnetic field detection. Its advantages are: wavelet transform can simultaneously analyze the time and frequency information of the signal, thereby capturing subtle changes and local features in the wake electromagnetic field signal, making feature extraction more accurate and rich. Its multi-resolution analysis capability not only improves the perception and recognition ability of the deep learning model for complex wake electromagnetic field signals, but also reduces the dependence on traditional feature extraction methods. In addition, the features generated by wavelet transform can enhance the robustness and stability of the model, which helps to improve the accuracy and reliability of wake electromagnetic field detection.

[0046] Third, the present invention successfully solves the problem of building a complex wake electromagnetic field detection model by training and optimizing the deep learning model with a large amount of underwater vehicle wake electromagnetic field calculation data.

[0047] The present invention adopts a new deep learning model with powerful feature extraction capabilities, which can automatically identify and capture the complex patterns and subtle changes of wake electromagnetic fields, reducing the reliance on traditional expert knowledge and manual feature engineering. The scalability and flexibility of deep learning methods make it easier to update and expand the model to meet the needs of wake electromagnetic field detection and environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The calculation domain setting of the underwater vehicle SUBOFF model, where (a) is the top view and (b) is the side view;

[0049] Figure 2 Schematic diagram of the relationship between Y+the thickness of the first layer of grid cells;

[0050] Figure 3 Schematic diagram of high Y+ wall treatment method;

[0051] Figure 4 The meshing results of the SUBOFF model for underwater vehicles;

[0052] Figure 5 Schematic diagram of the integration domain: total integration domain (left) and local integration domain (right);

[0053] Figure 6 It is a schematic diagram of linear interpolation of the local integration domain;

[0054] Figure 7 This is the flow chart for wake electromagnetic field calculation;

[0055] Figure 8 This is a schematic diagram of simulating the placement of electromagnetic sensors;

[0056] Fig. 9 Comparison results of pure noise electric field data (left) and signal-noise mixed electric field data (right);

[0057] Fig.10 It is a 2D-CNN deep neural network structure used for wake electromagnetic field detection;

[0058] Fig.11 Comparison of signal detection evaluation results of three algorithms. DETAILED DESCRIPTION

[0059] The present invention will be described in detail below in conjunction with the accompanying drawings and embodiments thereof, but the protection scope of the present invention is not limited to the scope of the embodiments described herein.

[0060] In general, the 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency spectrum features of the present invention first obtains a large number of data sets that can be used for simulation training based on hydrodynamics and electromagnetic simulations, and then extracts the underwater vehicle wake electromagnetic field signal into a time-frequency spectrum feature map through wavelet transform, and then processes and analyzes it through 2D-CNN. Wavelet transform effectively captures the detailed changes of the signal at different times and frequencies, while 2D-CNN performs deep learning and pattern recognition on these time-frequency spectrum features, thereby achieving accurate detection of the wake electromagnetic field. This method not only improves the accuracy of feature extraction, but also enhances the reliability and efficiency of signal recognition, providing a new technical means for underwater vehicle wake electromagnetic field detection.

[0061] The method of the present invention includes three parts, namely: (i) generating a calculation data set of the wake electromagnetic field of an underwater vehicle through a multi-physics field coupling method of hydrodynamics and electromagnetics, (ii) extracting the time-frequency spectrum characteristics of the wake electromagnetic field through wavelet transform, and (iii) developing a wake electromagnetic field detection model based on a 2D-CNN deep neural network. The applicant describes the three parts in detail below.

[0062] (a) Generate a computational dataset of the electromagnetic field of the underwater vehicle wake using a multi-physics coupling method of hydrodynamics and electromagnetics.

[0063] Principle

[0064] The seawater environment contains a large number of charged particles. Under the influence of Coulomb force, any positively or negatively charged ion will attract a number of ions with opposite charges. When the seawater is in equilibrium, these ions are electrically neutral. When the seawater is stimulated (such as the wake formed by underwater vehicles), these charged particles will be partially separated due to their own mass and friction coefficient, generating an electric field between ion clusters and forming ion polarization current.

[0065] Taking ion polarization current as the research object, we first analyze the forces acting on ions in seawater. Ions in seawater are affected by a variety of forces, and the forces that dominate the situation vary in different scenarios. But in general, the force equation for ions in seawater can be expressed as follows:

[0066]

[0067] In formula (1), the subscript p represents the p-type ion. There are seven items on the right side of the equation from left to right, and their meanings are as follows:

[0068] 1) Electric field force: the charge is e p The electric force E on the ions

[0069] 2) Magnetic field force: The Lorentz force on moving ions in magnetic field B

[0070] 3) Friction: The friction force exerted on the solute due to its viscosity can be described by the Debye-Huckel theory.

[0071] 4) Relaxation force: The force caused by the displacement of ion atmosphere and the disturbance of ion mobility. This force can be explained by Debye-Huckel-Onsanger theory.

[0072] describe.

[0073] 5) Electrophoretic force: The perturbation of the viscous drag caused by different ions moving at different speeds, also described by the Debye-Hückel-Onsanger theory.

[0074] 6) Gradient force: It is a thermodynamic force described by Fick's second law of diffusion.

[0075] 7) Pressure: The pressure exerted on ions when they move from a high pressure area to a low pressure area.

[0076] Where V p, v0, ρ0 represent the effective ion volume, fluid velocity and fluid density, respectively.

[0077] Analyzing formula (1), the polarization current density J can be obtained as:

[0078]

[0079] in,

[0080]

[0081] By analyzing the current densities in equation (2), it can be found that the first term comes from the Lorentz term of the geomagnetic field, which can be considered as the ion-induced polarization current density. The second term comes from the total differential of the perturbation velocity field itself. Its essence is that different ions have different molar masses. Because of the inertial effect, polarization separation occurs, forming a polarized current. This term can be considered as the ion inertial polarization current density. The third term is a coupling term of the first two terms. By analyzing its magnitude, it can be found that it is much smaller than the first two terms and can be ignored, resulting in the following equation:

[0082] J=σv0×B+αω (3)

[0083] Generally speaking, the ion polarization current caused by any external excitation perturbation velocity field contains ion induction term and ion inertia term. However, when an underwater vehicle is sailing underwater, the frequency of the perturbation velocity field caused by the wake is low, and the polarization current density contribution of the ion inertia term is very weak, so the ion induced polarization current will occupy the main part.

[0084] Introducing the active Maxwell equations and assuming that seawater is a homogeneous medium, we have the following equation:

[0085]

[0086]

[0087] in represents the curl operator, represents the divergence operator, represents the partial derivative with respect to time. H is the wake magnetic field intensity, D is the electric displacement vector, E is the wake electric field intensity, J is the ion polarization current density vector in seawater, and μ represents the seawater magnetic permeability. From the current density expression of formula (3), the magnetic induction intensity B can be regarded as the geomagnetic field B E and the wake magnetic field B i The sum of the two parts, namely:

[0088] J=σv0×(B E +B i ) (8)

[0089] Because B E >>Bi , and there is no displacement current in seawater, combined with the following constitutive relationship:

[0090] B=μH (9)

[0091] Then the Maxwell-Ampere loop formula in equation (7) can be transformed into:

[0092]

[0093] The problem of studying the wake electromagnetic field is a low-frequency problem. The phase difference of the electromagnetic field at different positions within the range of interest is almost zero. Therefore, the time harmonic term can be ignored when solving the problem. According to formula (10), the relationship between the wake electromagnetic field can be obtained as follows:

[0094] E(r)=v0(r)×B(r)=v0(r)×(B E (r)+B i (r))≈v0(r)×B E (r)(11)

[0095]

[0096] In equations (11) and (12), the position vector r' and the position vector r represent the position of the source point and the position of the point to be measured, respectively.

[0097] When the physical and chemical parameters of seawater and the geomagnetic region are given, only the wake velocity field v0 is an unknown parameter in equations (12) and (13). By dividing the velocity field into discrete grid units, the discrete solution of the wake electromagnetic field can be completed.

[0098] Therefore, the solution of the wake electromagnetic field of the present invention is divided into two steps: the first step is to use CFD fluid simulation software and adopt the k-ε turbulence model to solve the wake velocity field v0 (this step is completed using general simulation software); the second step is to use the obtained wake velocity field v0 as a parameter input, and use numerical calculation to obtain the wake electric field E and the wake magnetic field B.

[0099] The specific process of solving the wake electromagnetic field and acquiring the data set is as follows:

[0100] (1) Use CFD software for fluid simulation: including geometric modeling, mesh generation, simulation material property setting, boundary conditions and initial conditions setting, solver solution, etc. The key steps are as follows:

[0101] 1. Geometric modeling and computational domain setting

[0102] As attached Figure 1As shown, the computational domain of the present invention is a rectangular parallelepiped area, the underwater vehicle adopts the SUBOFF model, the diving depth is h, the air layer height is 20D-h (D is the SUBOFF model diameter), the seawater layer height is 20D+h, the computational domain width is 2L+2wavelength (L is the SUBOFF model length), and the computational domain length is 7L+2wavelength. Figure 1 In (a), the computational domain is divided into the solution domain and the damping zone. The purpose of the damping zone is to prevent wave reflections at the upstream, downstream and side boundaries of the computational domain during the simulation. The width of the damping zone is 1.wavelength, which represents the maximum wavelength λ of the SUBOFF generated wave, which is determined by the dispersion relation of the linear surface wave under deep water conditions and can be calculated by the relationship between the wavelength λ and the SUBOFF model velocity U. It should be noted that the SUBOFF model is simulated in a uniform incoming flow with a velocity of U, rather than the model itself moving forward at a velocity of U.

[0103] 2. Grid division

[0104] From the perspective of CFD fluid simulation, meshing is a key step to determine whether the simulation can converge, whether the calculation accuracy meets the requirements, and whether the calculation time is short enough. For the SUBOFF model, the mesh type selected by the present invention is a hexahedral mesh. In order to effectively simulate the turbulent flow near the outer wall of the SUBOFF model, the thickness X of the first layer of meshing near the wall is very important. In order to calculate it, it is necessary to know the target Y of the problem. + Value, Y + The value is a dimensionless distance related to the turbulence model selection, which represents the distance from the center of the first layer of grid cells to the wall (such as the attached Figure 2 Use Y + When the wall is treated according to the standard, high Y + Wall processing uses a relatively coarse grid to divide the wall surface and uses the standard wall function to process the viscous area near the wall. The value is 30 to 500. Figure 3 shown.

[0105] Take Y + = 100 to perform grid division. The division results are shown in the attached figure. Figure 4 At this time, the volume grid near the underwater vehicle wall and the wall surface grid are finely divided, and the grid size of the rest is relatively large, achieving a good balance between calculation accuracy and calculation speed.

[0106] 3. Boundary condition setting and initial condition setting

[0107] The boundary conditions are set as follows:

[0108] 1) Velocity inlet: The incoming flow end, the top of the air layer and the bottom of the seawater layer are set as velocity inlet boundary conditions, corresponding to the SUBOFF speed.

[0109] 2) Pressure outlet: The flow end face, set as the pressure outlet boundary condition.

[0110] 3) Symmetry: The left and right sides of the calculation domain in the SUBOFF motion direction are set as symmetry boundary conditions, and the flow field is considered to be symmetrical.

[0111] 4) Wall: The SUBOFF model itself is set to a no-slip wall boundary condition.

[0112] The parameters for setting the initial conditions include: speed, underwater depth, and scale of the SUBOFF model.

[0113] 4. Solver solution

[0114] For the SUBOFF model, there is an air-seawater two-phase medium. The present invention adopts VOF multiphase flow calculation, the turbulence model adopts the k-ε turbulence model, the additional model selects gravity, VOF wave and unit mass repair, and the steady-state solution is used to perform simulation and solution in the CFD software, and the velocity field v0 under the current simulation conditions.

[0115] When solving the wake magnetic field, it is assumed that the grid size of the wake velocity field to be calculated is N×N×M (N represents the grid size of the horizontal plane, and M represents the grid size of the depth plane). A local grid of size n×n×n is selected from it, as shown in the attached figure. Figure 5 shown.

[0116] For any intermediate point P, according to the definition of wake magnetic field calculation integral, a discrete sum should be made in the full integral domain. However, due to the attenuation characteristics, only a local area around point P is selected for discrete summation to obtain the approximate value of the wake magnetic field. For each point in the integral range (each discrete point in the selected local area), the result of the velocity field v0 is substituted into formula (12) to obtain the magnetic field at the corresponding point (considering the boundary effect, it can be iterated several times), and the electric field at the corresponding point is calculated using formula (11).

[0117] Optionally, in addition to the above operations, the current density of the nearest neighboring grid of the intermediate point P is interpolated to further improve the calculation accuracy of the wake electric field, as shown in the attached Figure 6 Based on the above calculation ideas, a flow chart for calculating the wake induced electromagnetic field is given, as shown in the attached figure. Figure 7 shown.

[0118] Change the input conditions or model parameters of the simulation process, repeat the above process, and obtain electromagnetic field data under different conditions.

[0119] (II) Extracting the time-frequency spectrum characteristics of the wake electromagnetic field through wavelet transform

[0120] From the entire measurement area, select several measurement points, obtain the data (or change function) of the electric and magnetic fields changing with time at each measurement point, and perform wavelet transformation on the obtained electromagnetic field data. Selecting different measurement point positions here can further expand the electromagnetic field data set.

[0121] While retaining the advantages of short-time Fourier transform, wavelet transform overcomes the defect that time resolution and frequency resolution cannot be improved at the same time. It uses wavelet basis functions to represent signals, and can match the input signal by scaling and translating the wavelet basis functions, as expressed as follows:

[0122]

[0123] Where Ψ(t) represents the wavelet basis function, also known as the mother wavelet, and the present invention adopts the Morse wavelet basis function; Ψ ω,τ (t) represents the sub-wavelet generated by scaling and translating the wavelet basis function; W f (ω,τ) is the result of wavelet transform of signal f(t). For time sampling signal, it contains time domain information and frequency domain information. The present invention uses wavelet method to extract time-frequency information of wake electromagnetic field.

[0124] As attached Figure 8 As shown, N simulated electromagnetic sensors are placed in front of the underwater vehicle and all on one side, and the distance between two adjacent sensors is 10m.

[0125] Attach Figure 8 Take the electric field signal of the S1 measurement point as an example, the sampling rate is 3Hz, and the sampling time is 146s. Gaussian white noise is superimposed on the S1 data, and the signal-to-noise ratio is -5dB. Wavelet transform is performed to compare the time domain and wavelet transform results of the signal-noise mixed electric field data and the pure noise electric field data, as shown in the attached figure. Fig. 9 As shown in the figure, it can be found that when the signal-to-noise ratio is -5dB, the peak-to-peak value and waveform of the pure noise electric field data and the signal-to-noise mixed electric field data are basically the same, and the target signal can no longer be identified at this time. However, the time-frequency spectrum of the wavelet transform can extract the frequency characteristics and duration of the target signal.

[0126] (III) Development of wake electromagnetic field detection model based on 2D-CNN deep neural network The 2D-CNN deep neural network structure designed by the present invention is shown in the attached figure. Fig.10As shown, it includes two modules: a convolutional layer and a fully connected classifier layer. The convolutional module includes 7 convolutional layers and RELU layers, 3 batch normalization layers, 4 pooling layers, etc. The fully connected classifier module includes 1 flattening layer, 3 fully connected layers, 2 RELU layers, and uses Dropout processing.

[0127] The target data is generated by selecting data from several electromagnetic field sensors, with the signal-to-noise ratio set to -10 to 0 dB and a step size of 1 dB. The total length of the time domain signal is 440, of which the target signal length is 195. In order to make the signal detection more in line with the actual situation and increase the diversity of the data set features, the target signal appears randomly in the entire signal.

[0128] Then, the model is trained using the training set. In the training set setting, for a given electromagnetic sensor and signal-to-noise ratio, there are 200 data in total, with half of the target signal data and half of the noise data. For example, when the signal-to-noise ratio is -5dB, there are 200 data for the S1 electric field sensor, 100 of which have target signals and the other 100 do not. Therefore, the training set data has a total of 22,000 time domain data. Wavelet transform is performed on it to obtain 22,000 two-dimensional time-frequency spectrograms. Similarly, the test set has 4,400 time-frequency spectrograms.

[0129] The cross entropy function is used as the loss function during training, Adam is used as the optimization method, the learning rate is lr=0.0001, the batch size is batchsize=44, and the training round is epoch=20.

[0130] The trained model is used to test electromagnetic field wake detection.

[0131] The deep neural network proposed in this invention is compared with other neural networks, including FCN (fully connected neural network) and 1D-CNN (one-dimensional convolutional neural network). The results are shown in the attached figure. Fig.11 As shown in the figure, it can be found that in the wake electromagnetic field signal time period, the three methods all give alarm indications, indicating that the wake electromagnetic field signal is detected. However, in some pure noise time periods, the latter two algorithms have false alarms. This shows that the deep neural network proposed in the present invention is superior to the FCN neural network and the 1D-CNN neural network in detection performance.

[0132] Compared with 1D-CNN, 2D-CNN targets images, while 1D-CNN targets curves; the structure of the present invention is 2D-CNN, which is different from the FCN model result. The typical 2D-CNN model includes a convolution layer, a pooling layer and a fully connected layer, while FCN removes the fully connected layer and the global average pooling layer in 2D-CNN and introduces a transposed convolution layer.

[0133] In summary, the 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics proposed in the present invention can effectively improve the detection capability of wake electromagnetic fields, provide a new technical means for underwater vehicle wake electromagnetic field detection, and has good application prospects in the field of marine target electromagnetic field monitoring and early warning.

[0134] Although the principles of the present invention are described in detail above in conjunction with the preferred embodiments of the present invention, those skilled in the art should understand that the above embodiments are merely explanations of the exemplary implementations of the present invention, and are not intended to limit the scope of the present invention. The details in the embodiments do not constitute limitations on the scope of the present invention, and any obvious changes such as equivalent transformations, simple replacements, etc. based on the technical solution of the present invention, without departing from the spirit and scope of the present invention, fall within the protection scope of the present invention.

Claims

1. A 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics, characterized in that: The method comprises: Step (1) generating a computational dataset of electromagnetic field of the wake of a submersible by means of a multi-physics coupling method of hydrodynamics and electromagnetics, The step (1) comprises: (1.1) Construct a geometric model in the dynamic field simulation software and set the calculation domain of the geometric model; (1.2) Mesh the set computational domain; (1.3) setting boundary conditions for the geometric model, wherein the boundary conditions at least include a velocity inlet, a pressure outlet, a symmetry surface, and a wall surface; (1.4) Solve the velocity field distribution in the calculation domain in the dynamic field simulation software; (1.5) Based on the obtained velocity field distribution, the wake electric field E and wake magnetic field B at any point in the calculation domain are calculated according to the following formula: E(r)=v0(r)×B(r)=v0(r)×(B E (r)+B i (r))≈v0(r)×B E (r)(1) In equations (1) and (2), the position vector r' and the position vector r respectively represent the position of the source point and the position of the point to be measured, where J(r') represents the current density at position r', V represents the integrated volume, v0 represents the water flow velocity of the point to be measured calculated according to the velocity field, μ represents the magnetic permeability of seawater, σ represents the electrical conductivity of seawater, and B E represents the Earth's magnetic field; (1.6) Modify the input parameters and model parameters, repeat the above steps, obtain the wake electric field E and wake magnetic field B distribution under different input parameters and model parameter conditions, obtain multiple sets of wake electric field E and wake magnetic field B data, and form a data set; Step (2) takes the entire computational domain as the measurement area, selects a number of measurement points from the measurement area, obtains data on the change of the electric and magnetic fields at each measurement point over time under different conditions from the data set, performs wavelet transformation on the obtained electric and magnetic field data, and obtains a wavelet transformed data set; Step (3) constructs a 2D-CNN deep neural network, divides the constructed data set into a training set and a test set, inputs the data in the training set into the 2D-CNN deep neural network, trains the 2D-CNN deep neural network and tests it using the test set to obtain the trained 2D-CNN deep neural network, and uses the obtained trained 2D-CNN deep neural network for wake electromagnetic field detection.

2. The 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics according to claim 1 is characterized in that: The power field simulation software is CFD software.

3. The 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics according to claim 1 is characterized in that: The computational domain is a rectangular parallelepiped region, and a SUBOFF model is set in the computational domain to simulate an underwater vehicle.

4. The 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics according to claim 1 is characterized in that: The current density J is calculated by the following formula: J=σv0×B E ; Where σ represents the conductivity of seawater, v0 represents the water velocity at the test point calculated based on the velocity field, and B E Represents the Earth's magnetic field.

5. The 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics according to claim 1 is characterized in that: Wavelet transform is performed using the following formula: Where Ψ(t) represents the wavelet basis function; Ψ ω,τ (t) represents the sub-wavelet generated by scaling and translating the wavelet basis function; W f (ω,τ) is the result of wavelet transform of signal f(t).

6. The 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics according to claim 1 is characterized in that: The computational domain uses local meshing and interpolation methods to calculate the wake electromagnetic field.

7. The 2D-CNN network wake electromagnetic field detection method based on wavelet transform time-frequency characteristics according to claim 1 is characterized in that: The network model includes a convolution module and a full connector module. The convolution module contains 7 convolution layers and RELU layers, 3 batch normalization layers, and 4 pooling layers; the fully connected classifier module contains 1 flattening layer, 3 fully connected layers, and 2 RELU layers.

8. A method for acquiring a data set for training a wake electromagnetic field detection model, characterized in that: The method comprises: (1.1) Construct a geometric model in the dynamic field simulation software and set the calculation domain of the geometric model; (1.2) In the constructed geometric model, mesh the computational domain; (1.3) setting boundary conditions for the geometric model, wherein the boundary conditions at least include a velocity inlet, a pressure outlet, a symmetry surface, and a wall surface; (1.4) Solve the velocity field distribution in the calculation domain in the dynamic field simulation software; (1.5) Based on the obtained velocity field distribution, the wake electric field E and wake magnetic field B at any point in the calculation domain are calculated according to the following formula: E(r)=v0(r)×B(r)=v0(r)×(B E (r)+B i (r))≈v0(r)×B E (r)(4) In equations (4) and (5), the position vector r' and the position vector r represent the position of the source point and the position of the measured point, respectively. J(r') represents the current density at position r', V represents the volume, v0 represents the velocity calculated according to the velocity field, μ represents the magnetic permeability of seawater, σ represents the electrical conductivity of seawater, and B E represents the Earth's magnetic field; (1.6) Repeat the above steps to modify the input parameters and model parameters, obtain the distribution of the wake electric field E and the wake magnetic field B under different input parameters and model parameter conditions, and obtain multiple sets of wake electric field E and wake magnetic field B data to form a data set.

Images