A digital twin method for AUVs based on Mamba-enhanced residual super-resolution

Abstract

This invention belongs to the field of underwater vehicle digital twins and discloses an AUV digital twin method based on Mamba-enhanced residual super-resolution. The method includes: determining design parameters for the airfoil and relative positions of the fins and rudders to characterize the AUV, and constructing multiple sets of design variables based on these parameters; constructing an AUV shape hydrodynamic simulation calculation framework to perform hydrodynamic simulation on the AUV shape after inputting the design variables, and calculating the simulated velocity data of the wake field at the AUV's tail; storing each set of design variables and the corresponding simulated wake field velocity data as samples in a dataset; constructing a Mamba-enhanced residual super-resolution network and training the residual super-resolution network using the dataset; for a new set of design variables, inputting it into the trained residual super-resolution network to obtain the corresponding predicted wake field velocity data. This invention can provide rapid data generation and accelerate iteration for AUV fin and rudder shape design optimization.

Description

Technical Field

[0001] This invention belongs to the field of digital twins for unmanned underwater vehicles (AUVs), specifically relating to a digital twin method for AUVs based on Mamba-enhanced residual super-resolution. Background Technology

[0002] Autonomous Underwater Vehicles (AUVs) are widely used in scenarios such as seabed mapping, marine environmental monitoring, marine resource development, and underwater obstacle search and location. The fin and rudder layout of an AUV is a crucial component of the entire design process, directly reflecting its external characteristics. Furthermore, the quality of the fin and rudder layout design directly determines the AUV's hydrodynamic performance and navigation characteristics.

[0003] Optimizing the hydrodynamic shape of an AUV is of great significance for the design and optimization of its propulsion system. Although advances in computational fluid dynamics (CFD) have made ultra-high-precision calculations of underwater vehicle hydrodynamics possible, the use of complex meshes leads to the generation of large amounts of data and huge computational costs, which is detrimental to the optimization of fin and rudder design. Therefore, rapid and accurate prediction and reconstruction technology of the wake field of an AUV is of great importance for achieving rapid optimization design of the AUV's shape. Summary of the Invention

[0004] The purpose of this invention is to provide a digital twin method for AUVs based on Mamba-enhanced residual super-resolution, so as to overcome the limitations of existing methods in the hydrodynamic shape design of AUVs.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A digital twin method for AUVs based on Mamba-enhanced residual super-resolution includes:

[0007] Determine the design parameters for the airfoil and relative positions of the fins and rudders used to characterize the AUV, and construct multiple sets of design variables based on the design parameters;

[0008] A hydrodynamic simulation calculation framework for AUV shape is constructed to perform hydrodynamic simulation of AUV shape after inputting design variables, and to calculate the simulated velocity data of the wake field at the tail of AUV.

[0009] Each set of design variables and the corresponding simulated velocity data of the wake field are stored as samples in the dataset.

[0010] A residual super-resolution network based on Mamba enhancement is constructed, and the residual super-resolution network is trained using a dataset.

[0011] For a new set of design variables, input them into the trained residual super-resolution network to obtain the corresponding predicted velocity data of the wake field.

[0012] Furthermore, the design parameters of the AUV's fin and rudder airfoil and the relative positions of the fin and rudder include the distance from the leading edge of the fin to the AUV's head, the chord length of the rudder, the overall height of the fin and rudder, and the leading edge angle of the rudder.

[0013] Furthermore, multiple sets of design variables are constructed based on the design parameters, including:

[0014] Each design parameter is assigned a corresponding value range, and the value ranges of all design parameters constitute the value data space. The Latin hypercube sampling method is used to sample the value of each design parameter from the value data space. Each sampling yields a set of values ​​for all design parameters, which constitutes a set of design variables.

[0015] Furthermore, after inputting design variables, a hydrodynamic simulation of the AUV's shape is performed, including:

[0016] The mathematical language description of the design variables is performed using programming software to obtain the updated shape value point coordinates of the AUV's fins and rudders.

[0017] The batch file is executed, and the fin and rudder shape of the AUV is updated in the simulation management platform using the input model point coordinates and CAD secondary development technology to obtain the updated AUV hydrodynamic calculation engineering file.

[0018] Import the AUV hydrodynamic calculation project file, call the CFD mesh generation software, and perform mesh generation on the AUV and its external watershed.

[0019] Call the CFD fluid dynamics solver to calculate the fluid dynamic parameters of the AUV;

[0020] The velocity of the wake field is calculated using hydrodynamic parameters; the Kriging interpolation method is used to interpolate the wake field velocity into the background grid represented by polar coordinates on the wake field tail disk surface, which serves as a set of wake field simulation velocity data corresponding to the design variables.

[0021] Furthermore, the structure of the residual super-resolution network is as follows:

[0022] The residual attention block, Mamba block, upsampling block, residual attention block, and Mamba block are sequentially connected to form the first feature extraction unit; the upsampling block, residual attention block, and Mamba block are sequentially connected to form the second feature extraction unit; wherein:

[0023] After the design variables are input into the residual super-resolution network, the first feature is extracted by the first feature extraction unit. The first feature is fused with the design variables and used as the input of the second feature extraction unit. The second feature is then processed by the second feature extraction unit to obtain the second feature. After the second feature is fused with the design variables, it is then upsampled by the upsampling output block and output to obtain the predicted velocity data of the flow field.

[0024] Furthermore, each residual attention block is constructed by stacking multiple residual channel attention blocks.

[0025] Furthermore, during training, the residual super-resolution network uses mean squared error loss as the loss function to calculate the loss value between the predicted velocity data of the trailing flow field output after the design variables are processed by the residual super-resolution network and the simulated velocity data of the trailing flow field corresponding to the design variables. The Adam optimizer is used to adaptively adjust the learning rate of the residual super-resolution network during training. Once the loss value decreases to the preset value or the maximum number of training iterations is reached, optimization is stopped, and the performance of the trained residual super-resolution network is verified on the validation set.

[0026] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the AUV digital twin method based on Mamba-enhanced residual super-resolution.

[0027] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the AUV digital twin method based on Mamba-enhanced residual super-resolution.

[0028] Compared with the prior art, the present invention has the following technical features:

[0029] 1. This invention significantly improves the efficiency and accuracy of underwater vehicle wake field prediction and optimization design through the deep integration of super-resolution methods and hydrodynamic simulation technology. It innovatively combines residual super-resolution networks with Mamba modules to construct an end-to-end mapping model from design variables to the wake field. The Mamba module, with its state-space model characteristics, can efficiently capture the complex nonlinear relationships between fin airfoils, layout parameters, and the wake field, while the residual network preserves key physical features through cross-layer connections, achieving detailed twin reconstruction of the high-resolution wake velocity field. This overcomes the information loss problem of traditional digital twins based on reduced-order models and supports real-time or near-real-time analysis.

[0030] 2. This invention provides end-to-end support for the intelligent design of AUV fin and rudder shapes. Through a closed-loop process of parametric modeling, automated simulation, and AI prediction, it can flexibly adapt to various design variables (such as geometry and relative position), quickly evaluate the fluid performance of different configurations, and directly serve optimization goals such as drag reduction and noise reduction, reducing manual intervention. Its super-resolution network also possesses small-sample learning potential, generating high-precision wake fields using limited high-fidelity CFD data, alleviating the problem of scarce high-quality data in the field of fluid mechanics. In engineering applications, this technology can quickly evaluate the hydrodynamic performance of different AUV fin and rudder shape designs, significantly shortening the AUV design optimization development cycle and accelerating design iteration, especially valuable in scenarios such as rudder effect analysis and propeller layout optimization. Simultaneously, it introduces the Mamba module into the field of fluid digital twin field prediction for the first time, providing a new approach to complex physical field twin modeling. Its technical framework can also be transferred to the optimization design of cross-domain equipment such as ships and aircraft, demonstrating the innovation and practicality of multidisciplinary integration. Attached Figure Description

[0031] Figure 1 This is a schematic flowchart of the method of the present invention;

[0032] Figure 2 This is a simulation diagram of the AUV shape hydrodynamic simulation calculation framework in this invention;

[0033] Figure 3 This is a schematic diagram of the Mamba-enhanced residual super-resolution network used in this invention;

[0034] Figure 4 This is a schematic diagram of the design variables and position parameterization in this invention;

[0035] Figure 5 It is the advance coefficient fitting curve of the predicted velocity data of the wake field obtained based on a new set of design variables in an embodiment of the present invention. Detailed Implementation

[0036] To address the problems of existing technologies, super-resolution technology has been used in the rapid prediction process of wake fields. This technology can effectively extract the implicit features of design variables and establish a mapping relationship from design variables to wake field data. Combined with image processing techniques such as pixel analysis, feature extraction, and point cloud analysis, super-resolution technology can effectively extract one-dimensional input features and accurately restore the nonlinear components in high-resolution material flow fields. It has demonstrated superior performance in the process of rapid prediction of high-dimensional physical fields and provides technical support for realizing digital twins of AUV structural fields.

[0037] See Figure 1 The present invention provides an AUV digital twin method based on Mamba-enhanced residual super-resolution, comprising the following steps:

[0038] Step 1: Determine the design parameters for the airfoil and relative positions of the fins and rudders used to characterize the AUV, and construct multiple sets of design variables based on the design parameters.

[0039] Step 11: Determine the design parameters of the airfoil and relative position of the fin and rudder of the AUV; the design parameters include the distance X from the leading edge of the fin to the head of the AUV, the chord length L of the rudder, the overall height H of the fin and rudder, and the leading edge angle D of the rudder.

[0040] Step 12: Set the corresponding value range for each design parameter. The value ranges of all design parameters constitute the value data space. Use the Latin hypercube sampling method to sample the value of each design parameter from the value data space. Each sampling will obtain a set of values ​​for all design parameters, which constitutes a set of design variables [X,L,H,D]. Multiple sets of design variables can be obtained through multiple samplings.

[0041] Step 2: Construct a hydrodynamic simulation calculation framework for the AUV's shape, which is used to perform hydrodynamic simulation on the AUV's shape after inputting design variables, and calculate the simulated velocity data of the trailing flow field at the tail of the AUV; store each set of design variables and the corresponding simulated velocity data of the trailing flow field as samples in the dataset.

[0042] Among them, after inputting the design variables, a fluid dynamic simulation of the AUV's shape is performed, such as... Figure 2 As shown, it includes:

[0043] Step 21: Use programming software to describe the design variables in mathematical language to obtain the updated shape value point coordinates of the AUV's fins and rudders.

[0044] Step 22: Execute the batch file, use the input model point coordinates and combine them with CAD secondary development technology to automatically update the fin and rudder shape of the AUV in the simulation management platform, and obtain the updated AUV hydrodynamic calculation engineering file.

[0045] Step 23: Import the AUV hydrodynamic calculation project file, call the CFD mesh generation software to perform mesh generation on the AUV and its external flow domain, and perform appropriate mesh refinement in important areas such as fins and rudders.

[0046] Step 24: Call the CFD fluid dynamics solver to calculate the fluid dynamic parameters of the AUV.

[0047] Step 25: Calculate the velocity of the wake field using fluid dynamic parameters; use the Kriging interpolation method to interpolate the wake field velocity into the background grid represented by polar coordinates on the wake field tail disk surface, as a set of wake field simulation velocity data corresponding to the design variables, and the two constitute a set of samples.

[0048] Step 3: Construct a residual super-resolution network based on Mamba enhancement and train the residual super-resolution network using the dataset; for a new set of design variables, input them into the trained residual super-resolution network to obtain the corresponding predicted velocity data of the wake field.

[0049] See Figure 3 This invention provides a residual super-resolution network based on Mamba enhancement, comprising:

[0050] The residual attention block, Mamba block, upsampling block, residual attention block, and Mamba block are sequentially connected to form the first feature extraction unit; the upsampling block, residual attention block, and Mamba block are sequentially connected to form the second feature extraction unit; wherein:

[0051] After the design variables are input into the residual super-resolution network, the first feature is extracted by the first feature extraction unit. The first feature is fused with the design variables and used as the input of the second feature extraction unit. The second feature is then processed by the second feature extraction unit to obtain the second feature. After the second feature is fused with the design variables, it is then upsampled by the upsampling output block and output to obtain the predicted velocity data of the flow field.

[0052] In the residual super-resolution network described above in this scheme, each residual attention block is constructed by stacking multiple residual channel attention blocks (RCABs).

[0053] For the first feature extraction unit, the residual attention block is used to extract feature tensors from the design variables input into the network, and the extracted feature tensors are input into the Mamba block for channel convolution to fully extract the hidden flow field features. The feature map obtained after channel convolution is enlarged by upsampling blocks and then input into the next set of residual attention blocks and Mamba blocks for processing to obtain the first feature.

[0054] For the second feature extraction unit, the upsampling block first upsamples the first fused feature (the first feature fused with the design variable) to increase its size; then it is fed into the residual attention block and the Mamba block for feature tensor extraction and channel convolution processing to obtain the second feature.

[0055] Finally, the second feature undergoes final upsampling processing in the upsampling output block to obtain the predicted velocity data of the wake field, which is consistent with the simulated velocity data format.

[0056] When training the residual super-resolution network using the training set, the dataset is divided into a training set and a validation set in an 8:2 ratio. Mean squared error (MSE) loss is used as the loss function to calculate the loss between the predicted velocity data of the trailing flow field output after processing the design variables by the residual super-resolution network and the simulated velocity data of the trailing flow field corresponding to the design variables. The Adam optimizer is used to adaptively adjust the learning rate of the residual super-resolution network during training. Once the loss value decreases to a preset value or the maximum number of training iterations is reached, optimization stops, and the performance of the trained residual super-resolution network is verified on the validation set. When it meets the requirements, it can be used as the final residual super-resolution network.

[0057] In practical applications, for a set of new design variables [X,L,H,D] to be predicted, inputting them into the residual super-resolution network will yield the predicted velocity data of the trailing flow field output by the residual super-resolution network.

[0058] In summary, this invention ultimately achieves the result of rapidly extracting predicted velocity data of the wake field using a residual super-resolution network while given design variables, and accurately reconstructing the wake field characteristics. This enables accurate and rapid twin reconstruction of AUV wake field data, which can be used for automated design optimization of AUV fin and rudder shapes.

[0059] like Figure 4 and Figure 5 As shown, in one embodiment of the present invention, for the design variable [X,L,H,D], the lower bound of its value range is [3.55m, 0.3m, 0.4m, 100°], and the upper bound of its value range is [3.7m, 0.6m, 0.7m, 140°]. The two together constitute the value data space.

[0060] In the residual super-resolution network of this embodiment, the three residual attention blocks are composed of stacked residual channel attention blocks (RCABs) of 5, 3, and 3, respectively.

[0061] During training, the residual super-resolution network has an initial learning rate of 1e-3. The learning rate is adaptively adjusted during the training process. Once the loss function value drops below 0.004, optimization stops, and the current network parameters are output and saved.

[0062] Finally, taking a new set of design variables X=3.6m, L=0.5m, H=0.6m, D=120° as an example, after inputting them into the residual super-resolution network, the corresponding predicted velocity data of the wake field at 1.5 times the AUV shell radius is obtained as shown in the following figure. Figure 5 As shown, this enables rapid data generation and accelerated iteration for optimizing the AUV fin and rudder shape design.

[0063] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A digital twin method for AUVs based on Mamba-enhanced residual super-resolution, characterized in that, include: Determine the design parameters for the airfoil and relative positions of the fins and rudders used to characterize the AUV, and construct multiple sets of design variables based on the design parameters; A hydrodynamic simulation framework for AUV shape is constructed to perform hydrodynamic simulation on the AUV shape after inputting design variables, and to calculate the simulated velocity data of the wake field at the tail of the AUV. The design variables generated by each set of sampling and the corresponding simulated velocity data of the wake field are stored as samples in the dataset. A residual super-resolution network based on Mamba enhancement is constructed and trained using a dataset. For a new set of design variables, the network is input into the trained residual super-resolution network to obtain the corresponding predicted velocity data of the wake field. The structure of the residual super-resolution network is as follows: The residual attention block, Mamba block, upsampling block, residual attention block, and Mamba block are connected in sequence to form the first feature extraction unit; The upsampling block, residual attention block, and Mamba block are connected sequentially to form the second feature extraction unit; where: After the design variables are input into the residual super-resolution network, the first feature is extracted by the first feature extraction unit. The first feature is fused with the design variables and used as the input of the second feature extraction unit. The second feature is then processed by the second feature extraction unit to obtain the second feature. After the second feature is fused with the design variables, it is then upsampled by the upsampling output block and output to obtain the predicted velocity data of the trailing flow field. Each residual attention block is composed of multiple residual channel attention blocks stacked together.

2. The AUV digital twin method based on Mamba-enhanced residual super-resolution according to claim 1, characterized in that, The design parameters for the airfoil and relative position of the fins and rudders of the AUV include the distance from the leading edge of the fin to the head of the AUV, the chord length of the rudder, the overall height of the fins and rudders, and the leading edge angle of the rudder.

3. The AUV digital twin method based on Mamba-enhanced residual super-resolution according to claim 1, characterized in that, Multiple sets of design variables are constructed based on design parameters, including: Each design parameter is assigned a corresponding value range, and the value ranges of all design parameters constitute the value data space. The Latin hypercube sampling method is used to sample the value of each design parameter from the value data space. Each sampling yields a set of values ​​for all design parameters, which constitutes a set of design variables.

4. The AUV digital twin method based on Mamba-enhanced residual super-resolution according to claim 1, characterized in that, After inputting design variables, perform hydrodynamic simulation of the AUV's shape, including: The mathematical language description of the design variables is performed using programming software to obtain the updated shape value point coordinates of the AUV's fins and rudders. The batch file is executed, and the fin and rudder shape of the AUV is updated in the simulation management platform using the input model point coordinates and CAD secondary development technology to obtain the updated AUV hydrodynamic calculation engineering file. Import the AUV hydrodynamic calculation project file, call the CFD mesh generation software, and perform mesh generation on the AUV and its external watershed. Call the CFD fluid dynamics solver to calculate the fluid dynamic parameters of the AUV; The velocity of the wake field is calculated using hydrodynamic parameters; the Kriging interpolation method is used to interpolate the wake field velocity into the background grid represented by polar coordinates on the wake field tail disk surface, which serves as a set of wake field simulation velocity data corresponding to the design variables.

5. The AUV digital twin method based on Mamba-enhanced residual super-resolution according to claim 1, characterized in that, During training, the residual super-resolution network uses mean square error loss as the loss function to calculate the loss value between the predicted velocity data of the trailing flow field output after the design variables are processed by the residual super-resolution network and the simulated velocity data of the trailing flow field corresponding to the design variables. The Adam optimizer is used to adaptively adjust the learning rate during the training process of the residual super-resolution network. Once the loss value decreases to the preset value or the maximum number of training iterations is reached, the optimization stops, and the performance of the trained residual super-resolution network is verified on the validation set.

6. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes the computer program, it implements the AUV digital twin method based on Mamba-enhanced residual super-resolution as described in any one of claims 1-5.

7. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the AUV digital twin method based on Mamba-enhanced residual super-resolution as described in any one of claims 1-5.

Images