A method for internal wave forecasting based on wall sensor

Abstract

This invention discloses an internal wave prediction method based on wall sensors. First, it acquires parameters such as the thickness of the density cascade layer, the location of the density cascade center, the density values ​​of the upper and lower layers, wave amplitude, and buoyancy frequency. These parameters are then used in a numerical solver to solve the DJL equations, constructing the initial flow field for the internal wave. Calculation settings are configured in Fluent. Sensor measurement points are placed on the wall of the vehicle and data is transmitted. In Fluent, the pressure and wall shear stress data at the nodes are stored in ASCII format. A convolutional neural network is used to convert the input data into a dual-channel image format, with one channel storing pressure data and the other storing wall shear stress data. The coordinates and amplitude of the vehicle relative to the wave trough are used as labels. Finally, the convolutional neural network is trained, enabling the prediction of internal waves. This invention enables real-time prediction of internal isolated waves using sensors, which has significant engineering application value for ensuring the safe operation of vehicles.

Description

Technical Field

[0001] This invention belongs to the field of underwater detection technology, specifically relating to an internal wave prediction method. Background Technology

[0002] Internal solitary waves are ubiquitous in the ocean, with amplitudes reaching 150 meters and wavelengths extending for kilometers. However, they only appear as faint undulations on the water's surface and are difficult to detect. When an underwater vehicle encounters an internal solitary wave, it will be pulled down or lifted up within a short time, altering its trajectory and easily leading to its "crash," thus posing a serious threat to the operation of underwater vehicles.

[0003] Currently, there are two main methods for predicting internal solitary waves: satellite remote sensing and regional observation. Using synthetic aperture radar (SAR) for internal solitary wave prediction is the most commonly used method. The principle is as follows: the shear flow field of the internal solitary wave modulates the sea surface wave, causing convergence and divergence on the ocean surface, thus altering the sea surface roughness. This appears as alternating bright and dark stripes in SAR remote sensing images. Therefore, the key to internal wave prediction lies in identifying these stripes in the images. For regional observation, thermo-depth meters and various rapid, dense sampling survey instruments are often used, with fixed-point observations conducted by research vessels. Since the formation of internal waves is related to topography and tides, long-term observation can reveal the generation patterns of waves within the region, providing a certain reference for internal solitary wave prediction. However, both of these prediction methods have the following shortcomings:

[0004] (1) For remote sensing image forecasting, due to the vast area of ​​the global sea, it is difficult to process massive amounts of image data at the same time. Secondly, due to the difficulty in guaranteeing the quality of the captured images, the bright and dark stripes are often not obvious, which poses a challenge to the image processing methods.

[0005] (2) For regional observations, the time required is too long, resulting in a huge investment of financial and material resources. In addition, the area studied by this method is limited, and it can only provide the generation law of internal waves at fixed locations. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides an internal wave prediction method based on wall sensors. First, it acquires parameters such as the thickness of the density cascade layer, the location of the density cascade center, the density values ​​of the upper and lower layers, wave amplitude, and buoyancy frequency. These internal wave parameters are then input into a numerical solver to solve the DJL equations, constructing the initial flow field for the internal wave. Calculation settings are configured in Fluent. Sensor measurement points are placed on the wall of the vehicle and data is transmitted. In Fluent, the pressure and wall shear stress data at the nodes are saved in ASCII format. A convolutional neural network is used to convert the input data into a dual-channel image format, with one channel storing pressure data and the other storing wall shear stress data. The coordinates and amplitude of the vehicle relative to the wave trough are used as labels. Finally, the convolutional neural network is trained, and after training, the prediction of internal waves is achieved. This invention enables real-time prediction of internal isolated waves through sensors, which has significant engineering application value for ensuring the safe operation of vehicles.

[0007] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0008] Step 1: Obtain the thickness of the density cascade layer, the location of the density cascade center, the density values ​​of the upper and lower layers, the wave amplitude, and the buoyancy frequency parameters of the internal wave;

[0009] Step 2: Substitute the internal wave parameters obtained in Step 1 into the numerical solver to solve the DJL equations and construct the initial flow field of the internal wave;

[0010] Step 3: Divide the computational domain into a mesh using mesh generation software;

[0011] Step 4: Create Fluent process files;

[0012] Step 5: Configure the calculation settings in Fluent, including selecting the physical model and setting the parameters;

[0013] Step 6: Arrange sensor measurement points on the surface of the aircraft body and transmit data. In Fluent, save the data on pressure at the nodes and shear stress on the wall as ASCII data.

[0014] Step 7: Write a batch file (.bat) to automatically perform the calculation of the next calculation after the calculation of one calculation is completed;

[0015] Step 8: Convert the ASCII data obtained in Step 6 into a list, extract the parameters at the required positions, and convert it into a two-dimensional matrix; for multiple cases, integrate them into a three-dimensional tensor as the input data for the convolutional neural network.

[0016] Step 9: Using a convolutional neural network, the input data is converted into a dual-channel image format, with one channel storing pressure data and the other channel storing wall shear stress data; the coordinates and amplitude of the vehicle relative to the trough are used as labels.

[0017] Step 10: Train the convolutional neural network to predict internal waves after training.

[0018] Preferably, step 2 is as follows:

[0019] Solving the DJL equations yields the internal wave flow field data. The fully nonlinear DJL equations under continuous density stratification are expressed as follows:

[0020]

[0021] Where: η is the vertical displacement, H represents the water depth of the numerical flume, c represents the wave velocity, z represents the longitudinal coordinate, and N represents the buoyancy frequency;

[0022]

[0023]

[0024] Where ρ1 and ρ2 represent the densities of the upper and lower layers, respectively, z pyc It is the center of the mezzanine layer, d pyc It is half the thickness of the mezzanine layer; when the above parameters are determined, the effective potential energy (APE) must also be considered during wave generation. When the APE is changed, the amplitude and wavelength of the internal wave also change accordingly.

[0025]

[0026] Where l represents the length of the computational domain, H1 represents the width of the computational domain, and g represents the gravitational acceleration.

[0027] Preferably, the mesh generation software is ICEM.

[0028] Preferably, the selection of the physical model in step 5 is as follows:

[0029] The standard k-ε model is used for the internal wave turbulence model; the standard wall function is used; a two-phase flow model, mixture, is used instead for the simulation of continuous density stratification; a 6DOF model is enabled for the simulation of the six-DOF motion of the vehicle; the least-squares-cell-based method is used for the gradient model of convection and diffusion terms; the momentum, volume fraction, turbulent kinetic energy, and turbulent dissipation rate of turbulence are solved using a second-order upwind scheme; and a second-order implicit scheme is used for time discretization.

[0030] Preferably, the convolutional neural network is one of ResNet, AlexNet, VGGNet, and GoogleLeNet.

[0031] Preferably, the optimizer used during the training of the convolutional neural network is the Adam optimizer.

[0032] The beneficial effects of this invention are as follows:

[0033] This invention enables real-time prediction of internal isolated waves by using hydrodynamic signals monitored by sensors when an underwater vehicle encounters such waves, which has significant engineering application value in ensuring the safe operation of the vehicle. Attached Figure Description

[0034] Figure 1 This is a flowchart of the present invention.

[0035] Figure 2 The flow field diagram obtained by solving the DJL equation is shown in the embodiment of the present invention.

[0036] Figure 3 This is a computational mesh diagram showing the surrounding structure of the vehicle in an embodiment of the present invention.

[0037] Figure 4 This is the predicted x-position of the internal wave relative to the vehicle in this embodiment of the invention.

[0038] Figure 5 This is the predicted y-position of the internal wave relative to the vehicle in this embodiment of the invention. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] To address the shortcomings of existing forecasting methods, this invention proposes a novel sensor-based internal wave forecasting method to solve the problem of real-time forecasting of internal solitary waves, which is an important supplement to existing internal wave forecasting methods.

[0041] An internal wave prediction method based on a wall sensor includes the following steps:

[0042] Step 1: Obtain the thickness of the density cascade layer, the location of the density cascade center, the density values ​​of the upper and lower layers, the wave amplitude, and the buoyancy frequency parameters of the internal wave;

[0043] Step 2: Substitute the internal wave parameters obtained in Step 1 into the numerical solver to solve the DJL equations and construct the initial flow field of the internal wave;

[0044] Step 3: Divide the computational domain into a mesh using mesh generation software;

[0045] Step 4: Create Fluent process files;

[0046] Step 5: Configure the calculation settings in Fluent, including selecting the physical model and setting the parameters;

[0047] Step 6: Arrange sensor measurement points on the surface of the aircraft body and transmit data. In Fluent, save the data on pressure at the nodes and shear stress on the wall as ASCII data.

[0048] Step 7: Write a batch file (.bat) to automatically perform the calculation of the next calculation after the calculation of one calculation is completed;

[0049] Step 8: Convert the ASCII data obtained in Step 6 into a list, extract the parameters at the required positions, and convert it into a two-dimensional matrix; for multiple cases, integrate them into a three-dimensional tensor as the input data for the convolutional neural network.

[0050] Step 9: Using a convolutional neural network, the input data is converted into a dual-channel image format, with one channel storing pressure data and the other channel storing wall shear stress data; the coordinates and amplitude of the vehicle relative to the trough are used as labels.

[0051] Step 10: Train the convolutional neural network to predict internal waves after training. Specific implementation examples:

[0053] 1. Because ocean internal waves vary randomly in time and space, rapid and intensive sampling is required over a relatively long period. The most commonly used methods are anchored self-contained current meters that can simultaneously measure temperature, conductivity, and depth, or instruments that simultaneously measure temperature, salinity, and depth. Observations typically last for several months, and this method can obtain internal wave data under actual ocean conditions. In this invention, the observation data can be obtained from relevant institutions or from existing published literature. The actual stratified ocean structure is considered as a three-layered sandwich structure: upper and lower relatively stable density regions and a middle density gradient layer. In data processing, the most important parameters to measure are the density values ​​of the upper and lower layers, the thickness of the density gradient layer, and the location of the density gradient layer's center. The thickness of the density gradient layer can be obtained from the distribution of buoyancy frequencies.

[0054] 2. After obtaining the internal wave parameters of the local sea area based on observations or literature, the flow field data of the internal waves can be obtained by solving the DJL equations. The fully nonlinear DJL equations under continuous density stratification are expressed as follows:

[0055]

[0056] η=0 at|x|→∞

[0057] Where: η is the vertical displacement, H represents the water depth of the numerical flume, c represents the wave velocity, z represents the longitudinal coordinate, and N represents the buoyancy frequency.

[0058]

[0059]

[0060] The ρ1 and ρ2 mentioned above represent the densities of the upper and lower layers, respectively, z pyc It is the center of the mezzanine layer, d pyc It is half the thickness of the mezzanine layer. When the above parameters are determined, the effective potential energy (APE) must also be considered during wave generation. When the APE is changed, the amplitude and wavelength of the internal wave will change accordingly.

[0061]

[0062] Where l represents the length of the computational domain, H represents the width of the computational domain, g represents the gravitational acceleration, and z represents the Cartesian coordinates in the vertical direction.

[0063] 3. After completing step 2, multiple required internal solitary wave flow field data are obtained. For subsequent CFD calculations, a mesh needs to be generated according to the set computational domain size. During CFD calculations, since the motion of the vehicle needs to be simulated, overlapping meshes must be considered. Finally, the mesh generation software ICEM is used to generate the foreground and background meshes.

[0064] 4. Fluent's journal file has the .jou extension. It's a collection of Fluent commands written in Scheme. This method involves setting various solver parameters in the graphical interface. The system automatically records the solver settings and generates the journal file through operations and command input. Click File → Write → Start Journal in the Fluent interface. The system will then begin recording the journal file. Make the necessary settings during this time, and click Stop Journal to terminate the process. This step enables subsequent batch calculations; when changing parameters for simulation, only a single value in the .jou file needs to be modified.

[0065] 5. After the process file starts recording, relevant calculation settings are then configured in Fluent. In this calculation, the standard k-ε model is used for turbulence. The standard k-ε model is widely used in industrial applications, and its convergence and accuracy meet the requirements of engineering calculations. Since this problem is a high Reynolds number problem, the standard wall function is used. For continuous density stratification simulation, a two-phase flow model (mixture) is used instead. For six-DOF motion simulation of the vehicle, the 6DOF model is enabled. During initialization, the flow field data is selected from the DJL internal wave field data obtained in step 2, and interpolated into Fluent using an IP file format recognizable by Fluent for flow field initialization. For the calculation method, the gradient model for convection and diffusion terms uses the least-squares-cell-based method. The PRESTO! format is used to interpolate pressure at the cell surface. A second-order upwind scheme is used to solve for the momentum, volume fraction, turbulent kinetic energy, and turbulent dissipation rate of the turbulence. A second-order implicit scheme is used for time discretization. The above scheme ensures second-order accuracy in the calculation.

[0066] 6. This step requires deploying corresponding sensors on the surface of the vehicle and transmitting data. The pressure data at the nodes and the wall shear stress data are saved in ASCII format using Fluent's export command. Saving is done once at each time step. Since ASCII files occupy very little memory, this method is feasible. Because each node has corresponding Cartesian coordinates during the saving process, the corresponding sensor positions are easily located. Furthermore, the relative positions of these nodes in the ASCII file do not change when the vehicle deflects; therefore, the positions only need to be determined at the beginning, and subsequent extraction can be performed using a Python post-processing program.

[0067] 7. The CFD calculation process described above only applies to a single case. Machine learning requires large-scale data for training, thus necessitating the calculation of multiple cases. This is where batch processing (.bat) files become crucial. They can automatically start calculating the next case after one case is completed, enabling uninterrupted computer operation and significantly saving time. When a batch file is double-clicked, the system calls cmd.exe to run the commands in the file sequentially.

[0068] 8. After the batch processing program finishes running, all the ASCII files generated by the case calculations are saved in the folder. At this point, the data is messy and redundant, so a corresponding Python program needs to be written for post-processing of the results. Python has powerful data integration capabilities. This step uses the pandas library in Python to convert each ASCII file into a list, extract the parameters at the required positions, convert it into a two-dimensional matrix, and for multiple cases, integrate them into a three-dimensional tensor as input data for the neural network and store it.

[0069] 9. In the previous step, we obtained the input data. Since this method uses a supervised learning process, the data needs corresponding labels. The coordinates of the vehicle relative to the wave trough, wave amplitude, etc., are given as labels. A fixed seed is given before training, and the data is randomly shuffled in NumPy. When building the neural network, since convolutional neural networks have excellent performance in extracting image features, this invention adopts the convolutional neural network method. The input data is converted into a two-channel image format, with one channel storing pressure data and the other channel storing wall shear stress data. The convolutional neural network includes convolutional layers, linear activation layers, pooling layers, and fully connected layers, etc. This method can use existing mature deep learning frameworks for construction, such as ResNet, AlexNet, VGGNet, GoogleLeNet, etc.

[0070] 10. The training process of a neural network requires the selection of a suitable optimizer. Analysis revealed that Adam is simple to implement, computationally efficient, and capable of automatically adjusting the learning rate. Therefore, in many cases, Adam is considered a default optimizer with relatively good performance. This invention therefore adopts the Adam optimizer. Regarding the number of training epochs, too many epochs can easily lead to overfitting, while too few epochs can result in underfitting. Both of these will affect the model's predictions. Therefore, sensitivity analysis is necessary when selecting an optimizer, i.e., choosing a reasonable evaluation criterion, analyzing the prediction results under different numbers of epochs, and selecting the optimal number of epochs.

[0071] Through the above steps, a method for predicting internal waves of underwater vehicles based on wall sensors has been proposed. When an underwater vehicle encounters an internal solitary wave, it can achieve real-time prediction of the internal solitary wave through hydrodynamic signals monitored by sensors. This has great engineering application value for ensuring the safe operation of the vehicle.

Claims

1. A wall sensor-based internal wave forecasting method, characterized by, Includes the following steps: Step 1: Obtain the thickness of the density cascade layer, the location of the density cascade center, the density values ​​of the upper and lower layers, the wave amplitude, and the buoyancy frequency parameters of the internal wave; Step 2: Substitute the internal wave parameters obtained in Step 1 into the numerical solver to solve the DJL equations and construct the initial flow field of the internal wave; Step 3: Divide the computational domain into a mesh using mesh generation software; Step 4: Create Fluent process files; Step 5: Configure the calculation settings in Fluent, including selecting the physical model and setting the parameters; The physical model was selected as follows: The turbulence model for internal waves adopts the standard The model uses the standard wall function; a two-phase flow model (mixture) is used instead of the standard wall function for continuous density stratification simulation; a 6DOF model is enabled for six-DOF motion simulation of the aircraft; the gradient model for convection and diffusion terms uses the least-squares-cell-based method; a second-order upwind scheme is used to solve for the momentum, volume fraction, turbulent kinetic energy, and turbulent dissipation rate of the turbulence; and a second-order implicit scheme is used for time discretization. Step 6: Arrange sensor measurement points on the surface of the aircraft body and transmit data. In Fluent, save the data on pressure at the nodes and shear stress on the wall as ASCII data. Step 7: Write a batch file (.bat) to automatically perform the calculation of the next calculation after the calculation of one calculation is completed; Step 8: Convert the ASCII data obtained in Step 6 into a list, extract the parameters at the required positions, and convert it into a two-dimensional matrix; for multiple cases, integrate them into a three-dimensional tensor as the input data for the convolutional neural network. Step 9: Using a convolutional neural network, the input data is converted into a dual-channel image format, with one channel storing pressure data and the other channel storing wall shear stress data; the coordinates and amplitude of the vehicle relative to the trough are used as labels. Step 10: Train the convolutional neural network to predict the internal waves of newly acquired sensor data.

2. The internal wave prediction method based on a wall sensor according to claim 1, characterized in that, Step 2 is described in detail below: Solving the DJL equations yields the internal wave flow field data. The fully nonlinear DJL equations under continuous density stratification are expressed as follows: （1） in: η For vertical displacement, H This represents the water depth in the numerical water tank. c Represents wave speed. z Represents the vertical axis. N Represents the floating frequency; （2） （3） in, and These represent the densities of the upper and lower layers, respectively. It is the center of the density jump layer. It is half the thickness of the density gradient layer; when the above parameters are determined, the effective potential energy must also be considered during wave generation. APE When change APE As time progresses, the amplitude and wavelength of the internal wave change accordingly. The formula for calculating APE is as follows: （4） in l Represents the length of the computational domain. H 1 represents the width of the computational domain. g It represents gravitational acceleration.

3. The internal wave prediction method based on a wall sensor according to claim 1, characterized in that, The mesh generation software is ICEM.

4. The internal wave prediction method based on a wall sensor according to claim 1, characterized in that, The convolutional neural network is one of ResNet, AlexNet, VGGNet, or GoogleLeNet.

5. The internal wave prediction method based on a wall sensor according to claim 1, characterized in that, The optimizer used during the training of the convolutional neural network is the Adam optimizer.

Images