Hybrid driving underwater glider shape optimization design method and system based on convolutional neural network

Abstract

The application discloses a hybrid driving underwater glider shape optimization design method and system based on a convolutional neural network. The method firstly generates samples based on an experimental design method and constructs a three-dimensional geometric model, carries out multi-working condition computational fluid dynamics simulation, and constructs a data set; secondly, a multi-task convolutional neural network is trained, a geometric image is taken as input, and fast prediction of a flow field cloud picture and key performance indicators is realized; finally, the prediction model is integrated into an optimization platform, a multi-objective optimization algorithm is combined, a Pareto front layout scheme in a preset range is searched, and an optimal layout is selected based on working condition weighted scoring. The application effectively overcomes problems of high time consumption of traditional CFD simulation and limitation of single model optimization capacity, significantly improves design efficiency and global optimization capacity, realizes overall performance optimization of the underwater glider under multi-working conditions such as gliding, propelling and rotating, and has important engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of underwater glider appendage layout optimization technology, and particularly relates to a hybrid drive underwater glider shape optimization design method based on convolutional neural networks. Background Technology

[0002] In the process of human understanding and development of the ocean, the ever-increasing demand for ocean observation and exploration has driven the rapid iteration of mobile ocean observation platform technology. Among them, hybrid-driven underwater gliders, as a new type of underwater vehicle, have shown broad application prospects in marine environmental monitoring and scientific research due to their high endurance driven by buoyancy and high maneuverability driven by propellers.

[0003] The performance of underwater gliders needs to be comprehensively evaluated under various operating conditions, primarily including gliding, steady propulsion, and steady turning. The relative spatial arrangement of the propeller, wings, and tail rudder is a key factor determining the hydrodynamic performance of the aircraft. Through reasonable layout optimization, the aircraft's long-term endurance, propulsion efficiency, and maneuverability can be effectively balanced, thereby achieving optimal overall performance.

[0004] However, existing underwater glider layout optimization methods still have the following limitations: First, traditional optimization methods heavily rely on computational fluid dynamics (CFD) simulations. Since CFD simulations are extremely time-consuming, the high computational cost severely restricts R&D efficiency and optimization depth when performing global optimization in complex design spaces.

[0005] Secondly, in terms of optimization targets, existing studies mostly focus on independent optimization of individual components such as wings, propellers, or fuselages, lacking a collaborative optimization mechanism from the overall perspective of the hydrodynamic system. This makes it difficult to effectively handle complex flow interference effects between components (such as the interaction between propeller wake and fuselage / tail rudder), resulting in the inability to fully explore the performance potential brought about by the collaborative layout of multiple components.

[0006] Furthermore, in terms of optimization objectives, existing research focuses on improving single performance indicators (such as lift-to-drag ratio or propulsion efficiency), neglecting the complex performance requirements of underwater gliders to perform diverse tasks. This makes it difficult to meet the synergistic requirements of comprehensive indicators such as efficient gliding, rapid propulsion, and flexible maneuverability in practical engineering.

[0007] With the development of artificial intelligence technology, optimization strategies based on surrogate models have been introduced into the field of engineering design to alleviate simulation pressure. However, traditional response surface models or Kriging models suffer from insufficient fitting ability and limited generalization performance when dealing with input features of high-dimensional spaces, strongly nonlinear coupling, and complex multi-component systems. In recent years, deep learning technologies, represented by convolutional neural networks (CNNs), have demonstrated excellent image feature extraction and rapid prediction capabilities, providing a new technical path for complex shape optimization. Although this technology has achieved certain results in the field of aerodynamic shape design, for systems such as hybrid-driven underwater gliders with complex structures, special operating environments, and high coupling between multiple operating conditions, there is still a lack of systematic overall layout optimization methods. Research on multi-component layout collaborative optimization combined with deep learning technology is still in its early stages. Summary of the Invention

[0008] To address the problems of existing hybrid-driven underwater glider layout optimization methods relying on time-consuming CFD simulations, resulting in low design efficiency, difficulty in handling complex coupling interference of multiple components, and the inability of existing single performance index optimization strategies to meet the comprehensive requirements of diverse tasks, as well as the limited ability of conventional surrogate models to handle high-dimensional and strongly nonlinear spaces, this invention provides a hybrid-driven underwater glider shape optimization design method based on convolutional neural networks.

[0009] This invention is implemented as follows: a hybrid-driven underwater glider shape optimization design method based on convolutional neural networks, characterized by the following steps: The key layout parameters of the hybrid-driven underwater glider appendages were determined as design variables. Multiple sets of sample parameter combinations were generated within a preset range using experimental design methods, and the corresponding three-dimensional geometric models were automatically generated. Multi-condition computational fluid dynamics simulations are performed on each of the three-dimensional geometric models to extract flow field cloud maps and calculate key performance indicators. The geometric images, flow field cloud maps and key performance indicators corresponding to the three-dimensional geometric models are paired to construct a multi-condition dataset. Construct and train a multi-task convolutional neural network, using the geometric image as input and the flow field cloud map and the key performance indicators as learning targets, to establish a predictive model from geometric shape to flow field distribution and performance indicators; The trained prediction model is integrated into the optimization platform. Within the preset value range of the design variables, a multi-objective optimization algorithm is used to automatically search for Pareto front layout schemes. The optimal layout scheme is then selected based on the weighted score of multiple operating conditions.

[0010] In the above technical solution, preferably, the hybrid-driven underwater glider appendage includes dual propellers, wings, and a tail rudder; the geometric image is a multi-view orthogonal two-dimensional contour map generated by rendering the three-dimensional geometric model, including a front view, a top view, and a side view, and each view is stitched together in the channel dimension to form a multi-channel composite image.

[0011] In the above technical solution, preferably, the multi-condition computational fluid dynamics simulation includes steady gliding condition, steady propulsion condition and steady rotation condition; the flow field cloud map includes velocity cloud map, pressure cloud map and vorticity cloud map that correspond to the geometric image in terms of view and spatial scale.

[0012] In the above technical solution, preferably, the key performance indicators include the lift-to-drag ratio obtained under steady gliding conditions, and the propeller propulsion efficiency and fuselage drag obtained under steady propulsion conditions; the specific calculation formulas for the lift-to-drag ratio and the propeller propulsion efficiency are as follows: Boost-to-drag ratio: ; Propeller propulsion efficiency: , Among them, the advance ratio is: , Thrust coefficient: , Torque coefficient: ; In the formula, The inlet velocity of the fluid domain. The propeller speed, The diameter of the blade. For the total thrust, This is the sum of torques. This refers to the density of seawater.

[0013] In the above technical solution, preferably, the key performance indicators include the turning radius and unit steering energy consumption obtained under steady-state turning conditions; the specific calculation formulas for the turning radius and the unit steering energy consumption are as follows: Turning radius: ; Energy consumption per unit of steering: , The total input power of the propeller is: , angular velocity of rotation: ; In the formula, and These are the torques of the left and right propellers, respectively. and These represent the rotational speeds of the left and right propellers, respectively.

[0014] In the above technical solution, preferably, the multi-task convolutional neural network adopts an encoder-decoder architecture, and the network structure includes a shared encoder and parallel image reconstruction branch and performance regression branch; the shared encoder is used to extract features layer by layer from the input geometric image, the image reconstruction branch is used to decode the features and reconstruct the predicted flow field cloud map, the performance regression branch is used to decode the features and predict the performance index, and the shared encoder and the two parallel branches fuse multi-scale information through skip connections.

[0015] In the above technical solution, preferably, the loss function used to train the multi-task convolutional neural network includes the pixel-level mean square error of the predicted flow field cloud map, structural similarity loss, and the mean square error of the performance index prediction; the formula of the loss function is: ; in, To predict flow field contour maps, This is a true flow field cloud map. and This is the task loss weighting coefficient. For structural similarity weighting coefficients, For predicting performance metric vectors: , This is a vector of actual performance metrics.

[0016] In the above technical solution, preferably, the multi-objective optimization algorithm is a multi-objective genetic algorithm, which uses the prediction model to predict the flow field cloud map and key performance indicators of candidate solutions, and performs Pareto front search; the optimization objective formula of the multi-objective optimization algorithm is: ; ; The proportion of operating time in steady gliding, steady propulsion, and steady rotation modes is used as a weight to weight the normalized performance indicators of each layout scheme in the Pareto front, and the layout with the highest weighted total score is selected as the optimal layout scheme.

[0017] This invention provides a hybrid-driven underwater glider shape optimization design method based on convolutional neural networks. By constructing an end-to-end intelligent prediction and optimization link of "geometric shape - flow field cloud map - performance indicators," it achieves collaborative automatic design under multiple components and operating conditions. Specific beneficial effects are as follows: This invention achieves real-time intelligent reasoning from geometric shape to full flow field information: It applies convolutional neural networks to the shape optimization of underwater gliders. Through image feature mapping, the core model can directly predict high-resolution velocity, pressure, and vorticity fields from standardized geometric contour diagrams. Compared to traditional surrogate models that can only predict performance scalars, this method captures the physical essence of fluid flow, providing rich flow field distribution features for optimization design, significantly improving the physical interpretability of the design scheme and the accuracy of optimization decisions.

[0018] A highly efficient and accurate multi-task end-to-end prediction system was established: The multi-task convolutional neural network prediction system constructed in this invention can complete the quantitative feedback of key hydrodynamic performance within milliseconds after receiving the geometric image of the candidate layout. This system effectively replaces the time-consuming computational steps in the optimization iteration of traditional CFD simulation, making global optimization possible in complex design spaces. While maintaining computational accuracy, it greatly shortens the R&D cycle and significantly reduces the consumption of computing resources.

[0019] This invention overcomes the challenge of collaborative automatic optimization under multiple objectives and constraints: It uses an intelligent prediction model as an evaluation engine, coupled with global optimization strategies such as multi-objective genetic algorithms, to achieve joint optimization of performance indicators for multiple operating conditions, including gliding, propulsion, and turning. This method can automatically search for the optimal combination of layout parameters that satisfies geometric constraints and propulsion interference constraints, overcoming the limitations of traditional serial design or empirical trial-and-error methods that are prone to getting trapped in local optima and cannot simultaneously address multiple conflicting objectives. It achieves system-level collaborative optimization of the spatial layout of multiple components.

[0020] A closed-loop technology system from data acquisition to verification has been constructed: This invention provides a complete process covering parametric modeling, automated CFD high-fidelity dataset construction, intelligent model training, and final engineering verification. This closed-loop system not only ensures the quality of training samples and the accuracy of learning objectives, but also ensures the engineering feasibility of optimization results through final high-precision CFD verification, achieving a high degree of unity between theoretical models and actual manufacturing requirements, and can directly guide the design and manufacturing of engineering prototypes.

[0021] Possessing broad versatility and technical scalability: The "graphical feature representation - flow field intermediate state constraint - intelligent strategy optimization" framework proposed in this invention has a high degree of modularity. This methodology is not limited to specific models of hybrid-driven underwater gliders; by changing the set of design variables and specific flow field training data, it can be easily transferred to the field of shape optimization for other underwater vehicles and even aircraft, providing a universal technical paradigm for solving the shape mapping and performance optimization problems of complex dynamic fluid systems.

[0022] This invention proposes a hybrid-driven underwater glider shape optimization design system based on convolutional neural networks. The system is characterized by including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned method.

[0023] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method described above. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process of the method described in the embodiment of the present invention; Figure 2 This is a schematic diagram of the parameterized drive layout model of the underwater glider appendages according to an embodiment of the present invention; Figure 3 The surface mesh of the propeller model studied in the embodiments of the present invention; Figure 4 This is a pressure cloud diagram of the underwater glider under steady propulsion conditions as described in an embodiment of the present invention. Figure 5 This is a longitudinal plane pressure cloud diagram of an underwater glider under steady propulsion conditions as described in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] This invention provides a hybrid-driven underwater glider shape optimization design method based on convolutional neural networks. To further illustrate the structure of this invention, a detailed description is provided below in conjunction with the accompanying drawings: A hybrid-driven underwater glider shape optimization design method based on convolutional neural networks is available. Figure 1The process includes the following steps: S1, Parametric Modeling and Sample Generation: Determine the key layout parameters of the underwater glider appendages (twin propellers, wings, and tail rudder) as design variables, and set the value range of each design variable; generate multiple sets of sample parameter combinations through experimental design methods, and automatically generate the corresponding three-dimensional geometric models, realizing the automation of modeling and sample generation. S2, High-Fidelity Flow Field Simulation and Dataset Construction: Perform multi-condition CFD simulations on each geometric model in S1, including steady gliding, steady propulsion, and steady turning conditions; extract the corresponding flow field physical quantity cloud maps from the simulation results, and calculate key performance indicators to form a paired dataset of "geometric image - flow field cloud map - performance label" for subsequent model training. S3, Convolutional Neural Network Model Construction and Training: Construct an image-to-image convolutional neural network, using geometric images as input and flow field cloud maps as learning targets for training, thereby establishing the ability to predict flow field distribution from geometric shape. The network structure includes a shared encoder and two parallel decoding branches: the encoder extracts input features layer by layer, and the decoder reconstructs the features into a predicted flow field. Information at different scales is fused between the encoder and decoder via skip connections. S4, CNN Model Integration and Multi-Objective Optimization: The trained multi-task CNN model is integrated into the optimization platform. Leveraging its ability to quickly predict flow fields and performance indicators, a multi-objective optimization algorithm is used to automatically search within the design space, generating Pareto front layout schemes representing different performance trade-offs. When selecting the final layout scheme, the performance indicators are weighted and scored based on the operating weights of different conditions, selecting the layout scheme with the best overall performance to balance endurance, propulsion efficiency, and maneuverability across multiple conditions and objectives. S5, Optimization Result Verification and Design Application: The optimal layout parameter combination obtained in S4 is verified using high-precision CFD simulation. The prediction accuracy of the surrogate model and the reliability of the optimization results are quantitatively evaluated, providing engineering guidance for the design of actual underwater gliders.

[0027] Taking a hybrid-driven underwater glider with a dual-sided symmetrically arranged propeller as an example, this paper determines the key layout parameters of the glider's appendages as design variables. The appendages include twin propellers, wings, and a tail rudder. Unlike traditional optimization methods based on empirical formulas, low-dimensional parametric response surfaces, or simple performance label regression, this embodiment not only establishes a mapping relationship between geometric layout parameters and performance indicators but also further introduces flow field contour maps as intermediate supervisory information. This creates a learnable and interpretable correlation chain between the shape, flow structure, and performance response, thereby improving the prediction accuracy, generalization ability, and optimization efficiency of the surrogate model under complex configurations. The specific steps are as follows.

[0028] A parametric 3D model of the corresponding structure was created in SolidWorks. Five key design variables are related to the propeller's position: X-axis mounting position, Y-axis mounting position, Z-axis mounting position, and two parameters related to the propeller blade orientation. Five key design variables are related to the wing's position: wing distance from the bow, leading edge angle, wingspan, root length, and tip length. Five key design variables are related to the tail rudder's position: tail rudder distance from the bow, trailing edge angle, wingspan, root length, and tip length. A total of 15 design variables were determined, and their value ranges were defined based on structural layout constraints, manufacturing constraints, and propulsion interference constraints. Clear physical and geometric constraint equations must be established when setting the value ranges for these 15 design variables. Specifically, structural layout constraints require that the wing roots of the wings and tail rudder must fall entirely within the effective load-bearing assembly section of the glider's main body shell; manufacturing constraints limit the minimum cross-sectional thickness and wingtip sweep angle of the wings and tail rudder to ensure structural strength in the high-pressure underwater environment; propulsion interference constraints, through a geometric intersection algorithm, strictly limit that the propeller blades must not interfere with the surfaces of the glider's main body, tail rudder, or antenna at any rotational phase and operating speed. Based on these physical constraints, the system calculates the upper and lower limits of each parameter, thereby determining the effective feasible region of each design variable.

[0029] Furthermore, the aforementioned 15 design variables do not affect the aircraft's performance independently, but rather jointly determine the overall external flow field distribution through the coupling effects between multiple components. For example, changes in the position and orientation of the twin propellers not only alter the spatial diffusion range and velocity deficit distribution of the propulsion wake, but also change the way the wake scours the glider's tail rudder and local fuselage areas; changes in the wing's mounting position, aspect ratio, and leading edge angle further affect the boundary layer development near the main body, the lift generation location, and the downstream wake morphology. Changes in the downstream wake will affect the uniformity of the propeller inflow, thus affecting the propeller's propulsion efficiency; tail rudder parameters will change the local flow deflection, separation, and vortex structure evolution characteristics at the tail. Especially in the complex configuration of hybrid-driven underwater gliders, which simultaneously perform gliding and propulsion missions, there are significant spatial correlations and coupling interference effects between the propeller wake, wing-induced flow, tail rudder control flow, and fuselage flow, resulting in overall performance exhibiting strong nonlinearity, multi-peak values, and cross-component linkage characteristics in response to geometric changes. Therefore, please refer to [link to relevant documentation]. Figure 2 In this embodiment, the twin propellers, wings and tail rudder are modeled in an integrated and coordinated manner, rather than each sub-component is optimized independently and separately, so as to more realistically reflect the flow interference mechanism and the overall performance change law under complex configuration.

[0030] Subsequently, multiple sets of sample parameter combinations were generated within a preset value range using experimental design methods, and the corresponding 3D geometric models were automatically generated. Specifically, the "Optimal Latin Hypercube" experimental design component in Isight was used to automatically generate 1000 sets of sample points within the aforementioned 15-dimensional design space. The automated modeling and simulation process was driven by Isight: for each set of sample parameters, its "Simcode" component called the SolidWorks API to execute a preset macro script, automatically updating the 3D geometric model based on the current design variables and exporting it as an x_t format file. To ensure that the 3D geometric model does not experience topological damage or reconstruction failure during large-scale parameter perturbations, a top-down skeleton modeling strategy was adopted in the parametric modeling stage, binding 15 design variables to the dimensions and geometric relationships of the global skeleton sketch. All entity features of the glider were geometrically stretched, feature arrayed, and subjected to Boolean operations with reference to the global skeleton sketch. When the macro script updated the design variables, the global skeleton sketch was updated first, thus robustly driving the interference-free reconstruction of downstream 3D entities and intersecting boundaries, achieving high-quality automated sample generation.

[0031] Next, multi-condition computational fluid dynamics (CFD) simulations are performed on each 3D geometric model. These simulations include steady gliding, steady propulsion, and steady rotation conditions. Specifically, Isight's "Simcode" component calls ANSYS Workbench, reads the geometry file generated in the previous step, and executes the parametric CFD simulation template. The CFD simulation template pre-sets steady gliding, steady propulsion, and steady rotation conditions. The fluid domain uses a cylindrical computational domain, and steady RANS numerical calculations are performed based on a turbulence model. To ensure the accuracy and convergence of the fluid dynamics calculations under multiple conditions, the fluid computational domain is set as a closed cylindrical space extending forward to 3 times the glider's main dimension, backward to 5 times the main dimension, and radially to 4 times the main dimension. (See also...) Figure 3 The mesh generation employs an unstructured polyhedral mesh combined with a boundary layer prism mesh. At least five boundary layer meshes are generated on the surfaces of the glider body, wings, tail rudder, and propeller to ensure that the height of the first mesh layer near the wall meets the dimensionless wall distance requirement. A high-fidelity solution of approximately 1 is required. To accurately capture the fluid separation and wake vortex structure around the appendages, the SST turbulence model, which has good anti-separation prediction capabilities, is selected. A turbulence model was developed, and the SIMPLEC algorithm was used for pressure-velocity coupled solution to balance the efficiency of large-scale automated computation with the accuracy of capturing flow field details.

[0032] After the simulation, the automated script extracts the key performance indicators corresponding to each operating condition from the results. Specifically, under the steady gliding condition, the propeller is not operating (speed is set to 0), the model angle of attack is set to 5°, and the inlet velocity of the fluid domain is set to the economic speed of 0.5 m / s. During the simulation, its lift is monitored. With resistance Based on the obtained data, calculate the lift-to-drag ratio under the corresponding design parameters: .

[0033] A higher lift-to-drag ratio indicates a higher gliding efficiency of the glider under this condition, and it is used as the core evaluation index for this steady gliding condition.

[0034] Under the stated steady propulsion condition, the propulsion performance and drag characteristics of the underwater glider in a dual-propeller coordinated propulsion state were evaluated. The left and right propellers operated at the same rotational speed of 2800 rpm, and the inlet flow velocity was set to a typical mission speed of 2 m / s for a series of simulations. During the simulations, the total thrust of the two propellers was monitored and recorded. Total torque The drag of the glider body (excluding the propeller) Calculate the propeller propulsion efficiency under the corresponding design parameters based on the obtained data: , Among them, the advance ratio is: , Thrust coefficient: , Torque coefficient: ; In the formula, The inlet velocity of the fluid domain. The propeller speed, The diameter of the blade. For the total thrust, This is the sum of torques. The density is seawater. Optimization focuses on improving propulsion efficiency and reducing overall fuselage drag. In this operating condition, the coupling relationship between the layouts of different components is particularly pronounced. The wake from the twin propellers creates velocity gradient regions, vortex core regions, and localized high dynamic pressure areas on both sides of the fuselage. If the propeller position changes, the way the wake scours the tail rudder and rear fuselage will also change; changes in the wings and tail rudder, in turn, will alter the uniformity of the incoming flow near the propellers and the propulsion efficiency. The performance under this complex configuration exhibits significant spatial correlation and is difficult to effectively represent using traditional response surface models based on explicit parametric function fitting.

[0035] Under the aforementioned steady-state rotation conditions, the rotational maneuverability and energy economy of the underwater glider were evaluated, with the inlet flow velocity set as the reference speed. By changing the speed difference between the left and right propellers A series of computational fluid dynamics simulations were conducted to bring the vehicle to a stable rotational state. During the simulations, lateral forces were monitored. Combined with the mass of the aircraft With the added mass during the motion Calculate the turning radius under steady-state turning conditions: .

[0036] Simultaneously, monitor the torque and rotational speed of the two propellers, and calculate the total input power of the propellers: .

[0037] To characterize the energy economy during the turning process, the energy consumption per unit of turning is calculated: , Wherein the angular velocity of rotation: ; In the formula, and These are the torques of the left and right propellers, respectively. and These represent the rotational speeds of the left and right propellers, respectively.

[0038] The turning condition also exhibits a clear flow structure dominance. The asymmetric wake generated by the differential propulsion of the left and right propellers is coupled with the flow around the tail rudder, affecting the local separation point and determining the lateral force. Therefore, it is suitable to use the intermediate features of the flow field to participate in the modeling to improve the prediction reliability.

[0039] In the aforementioned multi-condition automated simulation, to ensure that the performance indicators extracted from each computational sample have extremely high confidence and effectively filter out numerical oscillations, strict steady-state convergence and data cleaning extraction criteria were configured in the simulation script. Specifically, all computational residuals (such as the residuals of the continuity equation and the residuals of the anisotropic momentum equation) are required to decrease to a certain level. The absolute fluctuation of the key monitored boundary interface physical quantities (such as wing surface lift, main body drag, and propeller thrust) within the last 100 consecutive iterations does not exceed 0.5%. Once the solver determines that the above steady-state convergence criteria are met, the extraction program automatically extracts the monitoring data from the last 50 iterations and calculates the time arithmetic mean, which is used as the final performance numerical label for the sample under this condition. This avoids the adverse effects of random truncation errors in a single iteration step or numerical jumps caused by the shedding of local unsteady vortices on subsequent neural network supervised training.

[0040] Please see Figure 4 and Figure 5 During the simulation, flow field contour maps are simultaneously extracted. These contour maps include velocity, pressure, and vorticity contour maps that correspond to the geometric image in terms of both view and spatial scale. Specifically, the multi-profile velocity contour map reflects the propulsion wake diffusion range, local velocity deficit regions, and induced flow distribution; the pressure contour map reflects the high-pressure and low-pressure regions and pressure differential structures near the fuselage, wing, tail rudder, and propeller; and the vorticity contour map reflects the wake vortex structure, shear layer strength, separated flow, and vortex interactions. In this process, the geometric image is a multi-view orthogonal two-dimensional contour map generated by rendering the three-dimensional geometric model, including a front view, top view, and side view. These views are stitched together in the channel dimension to form a 3-channel multi-channel composite image, named... .

[0041] To ensure consistency between the input and output feature tensors of the convolutional neural network and improve training efficiency, a unified coordinate window clipping and matrix normalization preprocessing mechanism is adopted when extracting flow field cloud maps and stitching multi-view geometric images. First, a fixed physical window boundary is defined to enclose the glider body and the effective wake evolution region behind it, ensuring that the absolute coordinate systems of all geometric schemes with different shapes and sizes are perfectly aligned in physical space during flow field capture and geometric rendering. Second, the extracted color flow field cloud map is converted into a numerical matrix, that is, velocity, pressure, and vorticity are extracted as independent tensor channels, and the Min-Max scaling algorithm is used to linearly map the original physical field values ​​at each pixel to... The dimensionless region is used to eliminate the risk of gradient explosion caused by different physical quantities such as pressure and velocity. All processed two-dimensional geometric images Both the actual flow field contour plot and the actual flow field contour plot were uniformly resampled using a bilinear interpolation algorithm. Standard feature map size of a pixel.

[0042] Subsequently, the geometric image corresponding to the three-dimensional geometric model is... The actual flow field cloud map after size unification and normalization processing and a vector containing the key performance indicators mentioned above, including lift-to-drag ratio, propulsion efficiency, main body drag, turning radius, and steering energy consumption. Precise pairing was performed to construct a multi-condition dataset in the form of "image-flow field-performance" triplet. All 1000 samples, automatically collected and organized by engineering software, were randomly divided into training, validation, and test sets in a ratio of 7:2:1 for efficient training and generalization performance evaluation of subsequent deep learning network models.

[0043] Then, a multi-task convolutional neural network is constructed and trained. This network, based on an encoder-decoder structure (U-Net), achieves an end-to-end mapping from geometric images to multiple physical quantities, using the geometric images... As input, use the actual flow field cloud map. and the vector of real performance metrics that includes the above performance metrics To achieve the learning objective, a predictive model is established to correlate geometric shape with flow field distribution and performance indicators. The network structure mainly includes a shared encoder and parallel image reconstruction and performance regression branches. The shared encoder is responsible for extracting deep features layer by layer from the input geometric image. In constructing the shared encoder, a hierarchical downsampling architecture is adopted, containing at least four convolutional module sequences. Each convolutional module consists of two consecutive... It consists of two-dimensional convolutional layers, batch normalization layers, and non-linear activation functions (such as LeakyReLU), followed by a... Max pooling layers are used to reduce the spatial resolution of the feature map and double the number of feature channels. With this structure, the encoder can adaptively extract high-dimensional abstract feature representations that are highly correlated with local flow curvature, appendage relative positions, and chord length distribution from multi-view geometric contour boundaries.

[0044] Subsequently, the network is divided into an image reconstruction branch and a performance regression branch. The image reconstruction branch uses upsampling and other operations to decode and reconstruct the features into a predicted flow field cloud map with the same size as the target. Specifically, the image reconstruction branch utilizes transposed convolution for layer-by-layer spatial resolution restoration, and introduces skip connections between the feature layers at each upsampled level and the corresponding feature layers of the shared encoder at the same resolution. Skip connections fuse the fine geometric contour information of the shallow layers with the global semantic features of the deeper layers through channel concatenation, effectively preventing the loss of spatial location information in the deep network and ensuring high pixel-level alignment accuracy of the predicted flow field cloud map at fluid separation points and vortex core positions near the appendage boundary. The performance regression branch is used to decode features into predictive performance metrics. Specifically, it compresses the spatial dimension of the deep feature map into a one-dimensional feature vector through a global average pooling layer, then connects two or more layers of multilayer perceptrons (fully connected layers) and Dropout regularization layers, finally outputting the corresponding predictive performance metric vector. .

[0045] Furthermore, in this embodiment, the image reconstruction branch is not merely used to add an additional output, but rather serves as a crucial mechanism to guide the network in learning physically relevant intermediate features. Since performance indicators such as lift-to-drag ratio, propulsion efficiency, drag, turning radius, and unit turning energy consumption are essentially determined by the flow field structure, requiring the network to simultaneously reconstruct the velocity, pressure, and vorticity fields during training encourages the shared encoder to prioritize learning intermediate flow representations closely related to performance changes. These intermediate representations can correspond to the velocity deficit pattern in the propeller wake region, the distribution shape of high-pressure areas near the fuselage or wing surfaces, the location of local flow separation points at the tail, the vortex coupling characteristics between the wing and tail rudder, and the asymmetric wake deflection mode under differential propulsion conditions. These flow field features are difficult to capture stably and directly using simple geometric parameter numerical vectors, but they have a direct impact on various performance characteristics. Therefore, the constraints imposed on shared features by the image reconstruction branch during training can feed back into the performance regression branch, enabling the latter to perform performance predictions based on a more physically meaningful feature space, rather than simply fitting based on black-box statistical relationships. This multi-task convolutional neural network not only achieves end-to-end mapping, but also establishes a joint learning mechanism on the "geometric image - intermediate state of flow field - performance result" link, thereby enhancing the physical interpretability and engineering credibility of the model.

[0046] The multi-task convolutional neural network is trained using supervised training with a composite loss function that combines pixel accuracy and structural similarity. This loss function includes the pixel-level mean squared error of the predicted flow field contour map, structural similarity loss, and the mean squared error of the performance metric prediction, used to balance the contributions of pixel accuracy, flow field structural consistency, and performance prediction during training. Total loss function. The formula is: , in, To predict flow field contour maps, This is a true flow field cloud map. and This is the task loss weighting coefficient. This is the structural similarity weighting coefficient (which can be 0.1 in this embodiment). For predicting performance metric vectors, To correspond to the actual performance metric vector, the Adam Optimizer (Adaptive Moment Estimator) is used for backpropagation network parameter updates during model training. The initial learning rate is set to... A learning rate decay strategy based on validation set loss function monitoring (ReduceLROnPlateau) is introduced. When the total loss on the validation set does not decrease significantly for 10 consecutive epochs, the learning rate decays by 0.5 to ensure smooth convergence and escape from local minima in the later stages of training. The network training process is performed on a graphics processing unit that supports parallel computing and is configured with an early stopping mechanism. If the loss on the validation set does not improve for 20 consecutive iterations, training is terminated to prevent overfitting and to preserve the optimal network weights.

[0047] Furthermore, since flow field images provide higher-dimensional and denser supervision information than single performance scalars, the network can extract richer spatial structure knowledge from each training sample with the same sample size, thereby improving the utilization rate of information carried per unit sample. When there are strong nonlinear couplings, local performance abrupt changes, or uneven distribution of training samples in the design space, the intermediate supervision provided by the flow field branch can effectively alleviate the overfitting of the model to a small number of samples and improve the prediction stability for unseen layout schemes. By adopting this flow field-assisted training method, higher performance prediction accuracy and more robust generalization ability can be obtained with the same amount of computational fluid dynamics samples, or the dependence on the number of computational fluid dynamics samples can be reduced, thus demonstrating the objective technical effect of reducing data generation costs and computational resource consumption.

[0048] Finally, the trained prediction model is integrated into the optimization platform. Within the preset value range of the design variables, a multi-objective optimization algorithm is used for automatic search to generate Pareto front layout schemes. The optimal layout scheme is then selected based on a weighted score across multiple operating conditions. Specifically, the trained multi-task convolutional neural network model is exported as an independent inference script containing network weight parameters and embedded as a "Calculator" proxy model component into the Isight optimization platform. During the optimization iteration process, this proxy model component replaces the time-consuming traditional CFD numerical simulation solver. It can receive candidate layout parameter combinations generated by the optimization algorithm online, convert them into two-dimensional multi-channel geometric images, and perform millisecond-level fast forward inference, directly outputting the multi-operating condition flow field distribution prediction and five key performance index predictions corresponding to the scheme. This allows the optimization design process to not only obtain performance values ​​extremely efficiently but also simultaneously obtain flow field data characterizing local high-pressure areas, wake diffusion trends, and changes in separation structures, thereby significantly improving engineers' understanding of the optimization results and the efficiency of selection. In configuring the multi-objective optimization algorithm, this embodiment selects the Non-Dominated Sorting Genetic Algorithm (NSGA-II) as the core optimization engine. To balance global exploration capability and convergence efficiency, the algorithm's operating parameters are explicitly set as follows: population size of 100, number of generations of evolution of 200, crossover probability of 0.9 with simulated binary crossover (SBX) operator, and mutation probability of 0.05 with polynomial mutation operator. This explicit parameter configuration ensures high repeatability of the optimization process and achieves efficient global search of a 15-dimensional complex geometric space within a very short overall computation time.

[0049] In the Optimization module of the Isight platform, a multi-objective optimization problem is established, with its design variable vector denoted as . Its feasible domain The objective is determined by the physical value range of the aforementioned appendage layout parameters. The specific definition of the optimization objective formula for the multi-objective optimization algorithm is:

[0050]

[0051] In the formula, The lift-to-drag ratio for predicted steady gliding conditions; To predict the propeller propulsion efficiency under steady propulsion conditions, For the predicted drag of the fuselage under steady propulsion conditions, The predicted turning radius for steady-state turning conditions. This refers to the predicted unit turning energy consumption under steady-state turning conditions. Due to the profound physical contradictions and mutual constraints among these performance objectives (for example, increasing the wingspan improves the lift-to-drag ratio in gliding conditions, but inevitably increases the total drag of the fuselage in propulsion conditions and increases the turning radius in turning conditions), traditional single-objective optimization methods cannot provide a solution that balances all performance aspects. Therefore, multi-objective trade-off optimization must be performed within the entire feasible domain. To ensure that the candidate solutions generated by the optimization algorithm during the evolutionary search process fully conform to engineering realities, this embodiment introduces a constraint processing mechanism based on Deb's Feasibility Rules within the NSGA-II algorithm. This mechanism uses the structural layout constraint equations, manufacturing constraint equations, and propulsion interference geometry intersection algorithm described in Part 1 as strong constraint boundary conditions. When the algorithm generates inferior topological individuals that fail to pass geometric intersection or exceed the boundaries of the physical assembly, the constraint processor will impose a large penalty term on them, forcibly guiding the population to evolve towards a safe and feasible region that satisfies all physical constraints, thereby completely preventing the algorithm from converging to an unreasonable configuration that cannot be actually processed and manufactured.

[0052] After a pre-defined algebraic iterative evolution, the algorithm filters individuals in the population layer by layer, ultimately outputting a set of non-dominated solutions representing the optimal trade-offs between different performance objectives, i.e., the Pareto optimal front layout scheme set. After generating the Pareto optimal solution set, to better guide practical engineering applications with the abstract mathematical solution set, it is necessary to comprehensively consider the proportion of runtime under different operating conditions in the actual ocean observation mission profile of the hybrid-driven underwater glider. Specifically, this embodiment determines the runtime weighting coefficients for three operating conditions based on the typical mission characteristics of the underwater glider: the steady gliding condition accounts for 60% of the total runtime (weighting...). ), steady propulsion conditions account for 30% (weight) Steady rotation conditions account for 10% (weight) When making the final decision on each candidate design scheme in the Pareto front, the incomparability of the five performance indicators due to differences in dimensions and orders of magnitude must be eliminated first. Then, based on the weighted scores of multiple working conditions, the layout with the highest weighted total score is selected as the optimal layout scheme for the final engineering application.

[0053] To achieve scientific and objective comprehensive quantitative decision-making, this embodiment employs a classification and normalization method for benefit-type and cost-type indicators to perform dimensionless processing on various indicators within the Pareto front. For benefit-type indicators such as lift-to-drag ratio and propulsion efficiency, where higher is better, the following formula is used:

[0054] Mapped to interval; For cost-related indicators such as main body resistance, turning radius, and unit turning energy consumption, which are all better when they are as small as possible, the following formula is used:

[0055] Perform inverse mapping. Based on this, construct the final weighted total score control equation:

[0056] In the formula, and These are the sub-weights for the internal efficiency and drag of the propulsion system (both are set to 0.5 in this embodiment). and These are the sub-weights for maneuverability and energy consumption within the turning condition (both are set to 0.5 in this embodiment). The multi-objective optimization platform automatically calculates the comprehensive score for each individual in the Pareto solution set. The final selection is based on the overall score. The highest layout scheme is the optimal layout scheme for the appendages of the hybrid-driven underwater glider that takes into account the overall performance requirements of high endurance, high propulsion efficiency and high maneuverability. The corresponding parameterized values ​​are output to guide the subsequent physical structure design and prototype manufacturing.

[0057] The optimal layout parameters were verified again using high-precision CFD simulation. The results showed that the relative errors between the performance indicators predicted by the surrogate model and the CFD simulation results were all within 5%, and the predicted flow field was highly consistent with key flow characteristics. The optimized scheme significantly improved overall performance compared to the initial design. Furthermore, the optimization process only took a few seconds per design evaluation, compared to the hours required for traditional CFD simulations, significantly improving design efficiency. Further, because this embodiment enhances the model's learning ability to physical laws through flow field image-assisted training, it can reduce the number of high-cost CFD training samples required while achieving similar prediction accuracy, or significantly improve the performance prediction quality with the same number of samples. This demonstrates the technical effects of reducing computational resource consumption, shortening the surrogate model construction cycle, and improving overall optimization efficiency. For new layout schemes not involved in training, the model can still output flow field predictions and performance predictions consistent with its flow mechanism based on the spatial configuration features corresponding to the geometric image, indicating that this method has good generalization ability and engineering application value.

[0058] This embodiment fully demonstrates a closed-loop process from automated data generation, data preprocessing, deep learning model training, surrogate model integration and optimization to final high-precision verification. By deeply integrating parametric simulation, convolutional neural networks, and intelligent optimization algorithms, and using flow field cloud maps as an intermediate physical representation connecting shape parameters and performance indicators, this invention not only achieves efficient layout optimization of complex configurations of dual-thrust hybrid-driven underwater gliders, but also explains the sources of performance variations from the perspectives of wake structure, high and low pressure zone distribution, local separation characteristics, and component interference flows. It combines the advantages of high prediction accuracy, strong physical interpretability, high data utilization efficiency, and fast optimization speed, providing an efficient, intelligent, and physically interpretable implementation method for the shape layout optimization of dual-thrust hybrid-driven underwater gliders.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hybrid-driven underwater glider shape optimization design method based on convolutional neural networks, characterized in that, Includes the following steps: The key layout parameters of the hybrid-driven underwater glider appendages were determined as design variables. Multiple sets of sample parameter combinations were generated within a preset range using experimental design methods, and the corresponding three-dimensional geometric models were automatically generated. Multi-condition computational fluid dynamics simulations are performed on each of the three-dimensional geometric models to extract flow field cloud maps and calculate key performance indicators. The geometric images, flow field cloud maps and key performance indicators corresponding to the three-dimensional geometric models are paired to construct a multi-condition dataset. Construct and train a multi-task convolutional neural network, using the geometric image as input and the flow field cloud map and the key performance indicators as learning targets, to establish a predictive model from geometric shape to flow field distribution and performance indicators; The trained prediction model is integrated into the optimization platform. Within the preset value range of the design variables, a multi-objective optimization algorithm is used to automatically search for Pareto front layout schemes. The optimal layout scheme is then selected based on the weighted score of multiple operating conditions.

2. The method according to claim 1, characterized in that, The hybrid-driven underwater glider appendages include twin propellers, wings, and tail rudders; the geometric image is a multi-view orthogonal two-dimensional contour map generated by rendering the three-dimensional geometric model, including a front view, a top view, and a side view, and each view is stitched together in the channel dimension to form a multi-channel composite image.

3. The method according to claim 1, characterized in that, The multi-condition computational fluid dynamics simulation includes steady gliding, steady propulsion, and steady rotation conditions; the flow field cloud map includes velocity cloud map, pressure cloud map, and vorticity cloud map that correspond to the geometric image in terms of view and spatial scale.

4. The method according to claim 3, characterized in that, The key performance indicators include the lift-to-drag ratio obtained under steady gliding conditions, and the propeller propulsion efficiency and fuselage drag obtained under steady propulsion conditions; the specific calculation formulas for the lift-to-drag ratio and the propeller propulsion efficiency are as follows: Boost-to-drag ratio: ； Propeller propulsion efficiency: ， Among them, the advance ratio is: ， Thrust coefficient: ， Torque coefficient: ； In the formula, The inlet velocity of the fluid domain. The propeller speed, The diameter of the blade. For the total thrust, This is the sum of torques. This refers to the density of seawater.

5. The method according to claim 3, characterized in that, The key performance indicators include the turning radius and unit steering energy consumption obtained under steady-state turning conditions; the specific calculation formulas for the turning radius and the unit steering energy consumption are as follows: Turning radius: ； Energy consumption per unit of steering: ， The total input power of the propeller is: ， angular velocity of rotation: ； In the formula, and These are the torques of the left and right propellers, respectively. and These represent the rotational speeds of the left and right propellers, respectively.

6. The method according to claim 1, characterized in that, The multi-task convolutional neural network adopts an encoder-decoder architecture. The network structure includes a shared encoder and parallel image reconstruction and performance regression branches. The shared encoder is used to extract features layer by layer from the input geometric image. The image reconstruction branch is used to decode the features and reconstruct the predicted flow field cloud map. The performance regression branch is used to decode the features and predict the performance index. The shared encoder and the two parallel branches are fused with multi-scale information through skip connections.

7. The method according to claim 6, characterized in that, The loss function used to train the multi-task convolutional neural network includes pixel-level mean square error of predicted flow field cloud maps, structural similarity loss, and mean square error of performance index prediction; the formula for the loss function is: ； in, To predict flow field contour maps, This is a true flow field cloud map. and This is the task loss weighting coefficient. For structural similarity weighting coefficients, For predicting performance metric vectors: ， This is a vector of actual performance metrics.

8. The method according to claim 1, characterized in that, The multi-objective optimization algorithm is a multi-objective genetic algorithm, which uses the prediction model to predict the flow field cloud map and key performance indicators of candidate schemes, and performs Pareto front search; the optimization objective formula of the multi-objective optimization algorithm is: ； ； The proportion of operating time in steady gliding, steady propulsion, and steady rotation modes is used as a weight to weight the normalized performance indicators of each layout scheme in the Pareto front, and the layout with the highest weighted total score is selected as the optimal layout scheme.

9. A hybrid-driven underwater glider shape optimization design system based on convolutional neural networks, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images