Intelligent design method and device for pressure-resistant shell based on full-convolution deep-convolution generative adversarial network fusion process constraint

Abstract

The invention relates to a full-convolution deep-convolution generative adversarial network fusion process constraint-based artificial intelligence pressure-resistant shell design method, which comprises the steps of constructing a simulation model, the method comprises the steps of setting a pressure-resistant shell shape parameter range and step length and a pressure-resistant shell material, setting a rib parameter range and step length, performing simulation software modeling, and calculating the volume, the weight and the maximum pressure-resistant value of the pressure-resistant shell. Constructing a data set, and generating a plurality of basic samples according to the parameters of the simulation model; an AI model is constructed, the AI model is a full-convolution deep-convolution generative adversarial network architecture, the AI model comprises a forward prediction model and a reverse design model, the forward prediction model comprises three convolution layers of a full-convolution structure, and the reverse design model comprises three transposed convolution layers of the full-convolution structure; training and optimizing the AI model; and intelligently designing the pressure-resistant shell by using the AI model. The design efficiency is improved, the feasibility of the design scheme is high, and the withstand voltage value prediction precision and the parameter generation accuracy are improved.

Description

Technical Field

[0001] This application relates to the field of underwater vehicles, and in particular to an artificial intelligence pressure hull design method and device based on the constraints of a fully convolutional depth convolutional generative adversarial network fusion process. Background Technology

[0002] Underwater vehicles are devices capable of autonomous or remotely controlled navigation in water, primarily used in fields such as ocean exploration, resource discovery, and military reconnaissance. Their core structure requires pressure resistance, sealing, and maneuverability. The pressure hull is the main pressure-bearing structure of the underwater vehicle, typically made of high-strength materials, used to resist external water pressure and protect internal equipment; its design must balance strength, weight, and manufacturing cost. Ribs are ring-shaped reinforcing structures on the outer side of the pressure hull, preventing deformation or collapse of the hull under water pressure by distributing loads and increasing hull rigidity. Existing technologies for underwater vehicles include the following literature:

[0003] Patent document 1, CN119611684 A, published on March 14, 2025, entitled "A compartment structure for verifying the deformation coordination capability of a large floating raft vibration isolation system", provides a solution to avoid high stress concentration, simplify the bulkhead structure, ensure maximum radial deformation, and thus assess the deformation compensation capability.

[0004] Patent document 2, CN119442798 A, published on February 14, 2025, entitled "Design Method and System for a Similar Model of Ultimate Bearing Capacity of Cylindrical Ring-Rib Pressure Shell," provides a method and system for determining the prototype size parameters of the pressure shell and establishing a finite element model of the pressure shell prototype. Based on the calculation formula for the critical buckling strength of an unreinforced pressure shell, it defines the relationship between the scaling ratio Cl of the plate width of the cylindrical ring-rib pressure shell and the scaling ratio CR of the radius and the scaling ratio Ct of the plate thickness. It determines the principal scale scaling ratio of the cylindrical ring-rib pressure shell and calculates the number of ring ribs NS in the scaled model. The method and system are based on the calculation formula for the critical buckling strength of an unreinforced pressure shell. Formulas were used to calculate the relationship between the geometric length scaling ratio CL, radius scaling ratio CR, plate thickness scaling ratio Ct, number of ring ribs Np in the original model, and number of ring ribs NS in the scaled-down model of the cylindrical ring-ribbed pressure shell. Based on the parameters determined in S3 and S4, the dimensional parameters of the scaled-down model of the cylindrical ring-ribbed pressure shell were determined, and a finite element model of the scaled-down model was established. By analyzing the buckling strength and ultimate strength of the pressure shell prototype and the scaled-down model, the accuracy of the scaled-down model was verified, and the design of a similar model of the ultimate bearing capacity of the cylindrical ring-ribbed pressure shell was completed.

[0005] Patent document 3, CN118439126 A, published on August 6, 2024, entitled "A Modular Cabin Structure for Underwater Self-Propelled Models," describes a method for manufacturing cylindrical shell sections of different specifications based on a predetermined reference length, multiples of that reference length. The main scale scaling ratio of the cylindrical ring-ribbed pressure hull is then determined according to the reference length and specifications, and the number of ring ribs in the scaled-down model is calculated. Next, the geometric length parameters of the cylindrical ring-ribbed pressure hull are determined using the critical buckling strength calculation formula for unribbed pressure hulls, and a finite element model of the scaled-down model is established. Finally, the accuracy of the scaled-down model is verified by analyzing the buckling strength and ultimate strength of the prototype and the scaled-down model, thus completing the design of the modular cabin structure for underwater self-propelled models.

[0006] Patent document 4, CN117754329A, published on March 26, 2024, entitled "A Deep-Sea Pressure Shell Forming Device," describes a pressure shell forming device comprising a main shaft, a cylindrical body, and a tensioning mechanism. The main shaft, through a structure of multiple support shafts, enables the forming of the cylindrical body, and locking nuts and clamping nuts ensure the stability of the forming process. The tensioning mechanism, composed of multiple sets of planar support mechanisms, effectively controls the tension of the cylindrical body. Furthermore, the hinged design ensures the entire device remains balanced and stable during the forming process.

[0007] Patent document 5, CN116238189A, published on June 9, 2023, entitled "A composite material reinforced pressure-resistant shell core mold design and shell integral molding method", discloses the core mold material selection (foam or plastic), core mold structure design (multiple arc surface segments assembled, with preferred and non-preferred items alternately set, and the inclined plane deflection angle is 30°-45°), core mold assembly (the core mold is assembled sequentially on the steel shaft, and the rib width is controlled by positioning blocks), reinforced pressure-resistant shell integral molding (winding yarn to form ribs, winding skin and then curing), and core mold demolding (first pulling out the steel shaft, and then disassembling the arc surface segments).

[0008] Patent document 6, CN108674576A, published on October 19, 2018, entitled "A segmented cylindrical pressure-resistant cabin and assembly method for a deep-sea unmanned vehicle", discloses the pressure-resistant cabin structure, shell composition, connection method, material selection and assembly method.

[0009] The existing pressure hull design methods have the following problems:

[0010] Low level of intelligence: Existing technologies focus on traditional structural design (such as compartment structure, ring-ribbed pressure hull, segmented pressure hull), forming process (such as deep-sea pressure hull forming device, composite material integrated molding) or similar model design methods, without involving the application of artificial intelligence technology, lacking intelligent design means such as intelligent parameter optimization and automatic scheme generation, and design efficiency relies on human experience and traditional calculation, resulting in low efficiency.

[0011] Fragmentation in the design process: Existing technologies focus on single aspects (such as forming devices, assembly methods, and scaling up of similar models), lacking integrated design of the entire process of pressure hull (from material selection and structural parameter design to performance verification), making it difficult to achieve collaborative optimization of multiple aspects.

[0012] Insufficient multi-objective coordination: Existing technology designs focus on single performance (such as ultimate load-bearing capacity, sealing performance, and deformation coordination ability), without fully considering the coordinated optimization of multiple objectives such as pressure resistance, lightweight, buoyancy balance, and material properties, making it difficult to adapt to the comprehensive needs of complex underwater working conditions.

[0013] Low verification efficiency: Existing technologies rely on costly scaled-down tests and repetitive finite element simulations (such as cabin design cycles that can last for several weeks), lacking rapid virtual verification methods.

[0014] Weak cross-scenario versatility: Existing technical methods are limited to specific structures (such as cylindrical ring ribs) and are difficult to transfer to spherical / conical pressure hulls or different water depth scenarios (such as incompatible design parameters between deep sea and shallow water).

[0015] Existing artificial intelligence (AI) design approaches for pressure hull design have the following problems:

[0016] Application limitations: Existing AI design methods are mostly targeted at specific fields (such as transmissions, mechanical structures, stainless steel components, metamaterials, energetic materials, earth-rock dams, MEMS (Micro-Electro-Mechanical Systems) sensors, 3D printers, etc.), and do not involve the special object of underwater vehicle pressure hulls. They lack targeted model design for the core requirements of pressure hulls (such as pressure resistance, buckling strength, sealing structure and adaptability to the underwater environment).

[0017] Insufficient integration of physical properties: Some AI methods (such as purely data-driven generative adversarial networks and variational autoencoders) do not adequately integrate the physical and mechanical properties of the design object (such as the critical buckling strength of the pressure shell and the strain hardening characteristics of the material), which can easily lead to the generated scheme failing to meet structural safety requirements in actual engineering. For example, patent document 7: CN119623153A, published on March 14, 2025, entitled "Intelligent Design Method and System for MEMS Sensors Based on Generative Networks", ignores manufacturing constraints and is therefore unable to meet structural safety requirements.

[0018] Lack of integration across multiple stages: Existing AI design focuses on single processes (such as parameter prediction and solution generation) and has not formed a closed loop of the entire process of "requirement input - parameter optimization - performance simulation - solution iteration". In particular, it lacks the intelligent integration capability for the multi-dimensional coupling of "structure-material-working conditions" in the design of pressure hulls.

[0019] Parameter information loss: Patent document 8, CN118862683 A, published on October 29, 2024, entitled "A Smart Design Method for Composite Material Layup Based on Generative Neural Network", addresses the loss of parameter information caused by pooling in pressure shell design. Summary of the Invention

[0020] The purpose of this application is to solve the following technical problems:

[0021] Addressing the lack of intelligent design in pressure hull design: Overcoming the shortcomings of existing pressure hull designs that rely on traditional methods and are inefficient, thereby improving design efficiency.

[0022] Solving the challenge of multi-objective collaborative optimization: Integrating pressure resistance and weight into a collaborative global optimal design to avoid overall performance shortcomings caused by single performance optimization.

[0023] Addressing the shortcomings of the "structure-material-operating condition" coupled design: Combining the physical characteristics of the pressure hull of underwater vehicles, an AI model integrating physical constraints is constructed to ensure that the generated scheme achieves refined design while meeting safety specifications.

[0024] To address the fragmentation of the design process: Construct an intelligent closed loop covering the entire process of "requirements analysis - parameter generation - finite element simulation - solution iteration", integrating the scattered forming, assembly, and performance verification stages in the existing pressure shell design, and realizing end-to-end intelligent design from initial requirements to final solution, thus shortening the R&D cycle.

[0025] This addresses the issue of parameter information loss in existing AI models for pressure hull design.

[0026] To achieve the above-mentioned objectives, this application provides an artificial intelligence-based pressure-resistant shell design method constrained by a fusion process of fully convolutional depthwise generative adversarial networks, characterized by the following steps:

[0027] The simulation model construction includes: setting the range and step size of the pressure shell shape parameters and the pressure shell material, setting the range and step size of the rib parameters, modeling with simulation software, and calculating the volume and weight of the pressure shell and the maximum pressure resistance value.

[0028] Dataset construction involves generating multiple base samples based on the parameters of the simulation model.

[0029] AI model construction: The AI ​​model is a fully convolutional deep convolutional generative adversarial network architecture. The AI ​​model includes a forward prediction model and a reverse design model. The forward prediction model includes three convolutional layers with a fully convolutional structure, and the reverse design model includes three transposed convolutional layers with a fully convolutional structure.

[0030] Train and optimize the AI ​​model using an optimizer;

[0031] The AI ​​model is used to intelligently design a pressure-resistant shell.

[0032] As an improvement, the reverse design model includes a physical constraint module, which embeds process feasibility constraints. If the parameters generated by the reverse design model violate the constraints, a penalty loss is applied to the reverse design model network.

[0033] Training and optimizing the AI ​​model includes: setting optimizer parameters, setting batch size, setting total number of iterations; performing pre-training of the forward prediction model and the reverse design model alternately updating in the first iteration before the total number of iterations; and performing adversarial training of the forward prediction model and the reverse design model synchronously updating in the second iteration after the total number of iterations. Evaluation metrics are set, whereby the parameters generated by the forward prediction model must meet the set prediction accuracy, and the parameters generated by the reverse design model network must meet the set process constraint satisfaction rate and pressure resistance value compliance rate.

[0034] As another improvement, the dataset construction includes generating a set number of base samples with material labels to cover the main parameter combinations of all rib parameter combinations, wherein the base samples include enhanced samples generated for high pressure resistance and lightweight requirements;

[0035] The intelligent design of pressure-resistant shells using the AI ​​model includes: inputting design requirement parameters, generating a set number of candidate rib parameters using the reverse design model, verifying the performance of the candidate rib parameters using the forward prediction model, selecting design schemes that meet the requirement parameters, optimizing the target through an optimization algorithm, and outputting the optimal design scheme and simulation verification model.

[0036] As an improvement, the three convolutional layers of the fully convolutional structure of the forward prediction model include: a first convolutional layer with a 5×5 convolutional kernel, a stride of 2, and 64 output channels, followed by Batch Normalization and an activation function; a second convolutional layer with a 5×5 convolutional kernel, a stride of 2, and 128 output channels, followed by Batch Normalization and an activation function; and a third convolutional layer with a 5×5 convolutional kernel, a stride of 1, and 256 output channels, followed by Batch Normalization and an activation function. The input of the three convolutional layers is a 5-dimensional vector composed of 4-dimensional rib parameters and 1-dimensional material labels, and the output is a 2-dimensional normalized prediction value composed of weight and pressure resistance.

[0037] As an improvement, the three transposed convolutional layers of the fully convolutional structure include a first transposed convolutional layer, a second transposed convolutional layer, and a third transposed convolutional layer. The first transposed convolutional layer has a 5×5 kernel, a stride of 2, and outputs 64 channels, followed by Batch Normalization and an activation function. The second transposed convolutional layer has a 5×5 kernel, a stride of 2, and outputs 32 channels, followed by Batch Normalization and an activation function. The third transposed convolutional layer has a 5×5 kernel, a stride of 2, and outputs 1 channel, followed by an activation function. The input of the three transposed convolutional layers is a 3D vector composed of 2D performance requirement parameters and 1D material labels, and the output is normalized rib parameters.

[0038] As another improvement, the dataset construction includes: generating a set number of base samples with material labels to cover the main parameter combinations of all rib parameter combinations, wherein the base samples include enhanced samples generated for high pressure resistance and lightweight requirements; and sample labeling, wherein each base sample is labeled with input parameters, output labels, material type labels, and operating condition labels.

[0039] As another improvement, training and optimizing the AI ​​model further includes setting a loss function: the loss of the positive prediction model is...

[0040] LD=E[log(D(x))]+E[log(1-D(G(z)))];

[0041] Where E[log(D(x))] represents the expected probability that the discriminator D classifies the real sample x as true, and E[log(1-D(G(z)))] represents the expected probability that the discriminator classifies the generated sample G(z) as false;

[0042] And the loss of the reverse design model is

[0043] LG=E[log(D(G(z)))]+Lpenalty;

[0044] Where E[log(D(G(z)))] represents the probability that the generated sample is judged as "real" by the D network; Lpenalty represents the penalty added when the generation parameters violate the process constraints.

[0045] As another improvement, the input parameters include rib thickness, width, number, and spacing; the output markings include the weight of the pressure shell and the maximum pressure resistance value; and the operating condition label includes the water pressure rating.

[0046] As another improvement, the dataset construction also includes: normalizing the input parameters; standardizing the output labels; and partitioning the dataset by material type.

[0047] To achieve the aforementioned objectives, this application also provides an AI-based pressure-resistant shell design device constrained by a fully convolutional depthwise convolutional generative adversarial network fusion process, used to implement the aforementioned pressure-resistant shell design method, comprising:

[0048] The simulation model building module is used to set the range and step size of the pressure hull shape parameters and the pressure hull material, set the range and step size of the rib parameters, model the simulation software, and calculate the volume and weight of the pressure hull and the maximum pressure resistance value.

[0049] A dataset construction module is used to generate multiple basic samples based on the parameters of the simulation model;

[0050] The AI ​​model building module is an AI model that is a fully convolutional deep convolutional generative adversarial network architecture. The AI ​​model includes a forward prediction model and a reverse design model. The forward prediction model includes three convolutional layers with a fully convolutional structure, and the reverse design model includes three transposed convolutional layers with a fully convolutional structure.

[0051] A module for training and optimizing the AI ​​model, used to train the AI ​​model using an optimizer;

[0052] A module that uses the AI ​​model to intelligently design pressure-resistant shells.

[0053] The advantages of this application are:

[0054] This invention improves design efficiency by over 80%. Existing pressure shell designs (such as the aforementioned reference 2, "A design method and system for a similar model of the ultimate bearing capacity of a cylindrical ring-ribbed pressure shell") rely on manual modeling and simulation, which takes 2-4 hours per iteration. This invention achieves rapid parameter-performance mapping through a fully convolutional deep convolutional generative adversarial network (DCGAN), greatly reducing the time required to <0.1 seconds.

[0055] This invention achieves multi-objective collaborative optimization. Existing technologies (such as the aforementioned Reference 1, "A Cabin Structure for Verifying the Deformation Coordination Capacity of a Large Floating Raft Vibration Isolation System") only optimize a single performance, while this invention integrates pressure resistance, weight, and process feasibility to achieve global optimization.

[0056] This invention features cross-material compatibility. Existing composite material pressure shell designs (such as the aforementioned document 5, "A Composite Material Reinforced Pressure Shell Mandrel Design and Integral Molding Method") are disconnected from metal pressure shell designs. This invention achieves cross-material compatibility of the same model through material labels.

[0057] The pressure-resistant shell design of this invention is highly feasible. Existing artificial intelligence designs (such as the aforementioned reference 7, "Intelligent Design Method and System for MEMS Sensors Based on Generative Networks") neglect manufacturing constraints, while the solution of this invention has a process constraint satisfaction rate of ≥98%, which is in line with engineering practice.

[0058] The AI ​​model provided by this invention solves the problem of parameter information loss caused by pooling, and improves the accuracy of withstand voltage prediction and parameter generation. Attached Figure Description

[0059] Figure 1 This is an overall flowchart of the artificial intelligence pressure-resistant shell design method according to an embodiment of the present invention.

[0060] Figure 2 This is a flowchart of the positive prediction model algorithm according to an embodiment of the present invention.

[0061] Figure 3 This is a flowchart of the reverse design model algorithm according to an embodiment of the present invention.

[0062] In the attached image:

[0063] Batch Normalization (BN)

[0064] LeakyReLU is the activation function for deep convolutional generative adversarial network architectures;

[0065] ReLU is the activation function for deep convolutional generative adversarial network architectures;

[0066] Tanh is the activation function for deep convolutional generative adversarial network architectures. Detailed Implementation

[0067] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. In these drawings, the same or similar components are represented by the same or similar reference numerals. Those skilled in the art should understand that these embodiments are merely for explaining the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. It should be noted that in the description of the present invention, terms such as "upper," "lower," "left," "right," "inner," and "outer," indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is merely for ease of description and does not indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of the present invention.

[0068] This invention overcomes the shortcomings of existing pressure hull designs, which rely on traditional methods and are inefficient. It improves design efficiency by using artificial intelligence algorithms (generative neural networks that integrate physical constraints and multi-objective optimization models) to achieve intelligent reasoning and automatic generation from design requirements to structural parameters (such as the number of ring ribs, plate thickness, and material selection).

[0069] This invention solves the problem of multi-objective collaborative optimization: it integrates the pressure resistance value and weight into a collaborative global optimal design, avoiding the overall performance shortcomings caused by single performance optimization.

[0070] This invention addresses the shortcomings of the "structure-material-operating condition" coupled design: by combining the physical characteristics of the pressure hull of underwater vehicles (such as deep-sea pressure environment and material mechanical properties), an AI model is constructed that integrates physical constraints (such as critical buckling strength formula and material strain characteristics) to ensure that the generated scheme achieves refined design (such as the adaptation of composite material layup to pressure-resistant structure and accurate calculation of ring rib scaling ratio) while meeting safety specifications.

[0071] This invention addresses the problem of fragmented design processes by constructing a fully intelligent closed loop encompassing "requirements analysis - parameter generation - finite element simulation - solution iteration." It integrates the scattered forming, assembly, and performance verification stages in existing pressure shell designs, achieving end-to-end intelligent design from initial requirements to the final solution, thus shortening the R&D cycle.

[0072] The AI ​​model provided by this invention is a fully convolutional deep convolutional generative adversarial network (DCGAN) architecture. Its technical advantages are: eliminating the FC layer (fully connected layer) and pooling layer, and adopting a fully convolutional structure + transposed convolutional upsampling, which solves the problem of parameter information loss caused by pooling in existing AI models (such as the VAE of the aforementioned reference 8 "A Smart Design Method for Composite Material Layup Based on Generative Neural Network"), and improves the accuracy of pressure resistance value prediction (mean absolute error MAE ≤ 5%) and parameter generation accuracy.

[0073] Overall, the present invention provides an artificial intelligence-based pressure-resistant shell design method constrained by a fully convolutional depthwise generative adversarial network (DCGAN) fusion process, comprising:

[0074] Simulation model construction: Set the shape of the pressure shell, the range of rib parameters and the step size, and select PMI foam reinforced composite material or titanium alloy as the material; use simulation software to model and apply gradient water pressure, and calculate the volume, weight and maximum pressure resistance of the pressure shell;

[0075] Dataset construction: Generate basic samples with material labels (covering 75% of rib parameter combinations) and high pressure resistance and lightweight enhancement samples, label input parameters, output labels and material and working condition labels, complete data preprocessing and divide the data into subsets according to materials;

[0076] AI Model Construction: Build a fully convolutional DCGAN architecture, including a forward prediction model (3 convolutional layers, no FC layers and pooling layers) and a reverse design model (3 transposed convolutional layers, with embedded physical constraint modules);

[0077] AI model training and optimization: Using the Adam optimizer, set the batch size and total number of iterations (first 50 rounds of pre-training, last 150 rounds of adversarial training), define the loss function and evaluation metrics, and optimize the model;

[0078] Intelligent design application: Input design requirements, reverse model to generate candidate parameters, forward model to verify, and output the optimal solution and verification model after being filtered by optimization algorithm.

[0079] A detailed description of each of the above steps is provided below:

[0080] Simulation model construction: Set the range and step size of the pressure hull shape parameters (e.g., cylindrical diameter 800-2500mm, length 1500-4000mm), and the range and step size of the rib parameters (thickness 8-40mm, width 15-80mm, number 6-18, spacing 80-400mm). Select PMI foam-reinforced composite material or titanium alloy as the pressure hull material. Model the pressure hull using simulation software (e.g., Abaqus) and solid elements (e.g., C3D8I). Apply a gradient water pressure of 0-80MPa (corresponding to water depth of 0-8000m) to the outer surface and calculate the volume, weight, and maximum pressure resistance value (critical water pressure for first buckling).

[0081] Dataset Construction: Based on the simulation model parameters, generate basic samples with material labels (e.g., 12,000 samples, covering 75% of rib parameter combinations), including enhanced samples (e.g., 3,000 samples) for high pressure resistance (>60MPa) and lightweight (<500kg) requirements, for a total sample size of 15,000; label each sample with input parameters (rib thickness, width, number, spacing), output labels (weight, maximum pressure resistance), material type labels, and operating condition labels (water pressure rating); normalize the input parameters to [0,1], standardize the output labels, and divide them into subsets according to material type.

[0082] AI Model Construction: The AI ​​model adopts a fully convolutional DCGAN architecture, including a forward prediction model (D network) and a reverse design model (G network). The forward prediction model is a fully convolutional structure containing three convolutional layers: (1st convolutional layer: 5×5 kernel, stride 2, 64 channels, followed by Batch Normalization and LeakyReLU activation function; 2nd convolutional layer: 5×5 kernel, stride 2, 128 channels, followed by Batch Normalization and LeakyReLU activation function; 3rd convolutional layer: 5×5 kernel, stride 1, 256 channels, followed by Batch Normalization and LeakyReLU activation function). The input is a 5-dimensional vector (4-dimensional rib parameters + 1-dimensional material label), and the output is a 2-dimensional normalized predicted value (weight, pressure resistance). The reverse design model is a fully convolutional structure containing 3 transposed convolutional layers (first transposed convolutional layer: 5×5 transposed convolutional kernel, stride 2, 64 channels, followed by BatchNormalization and ReLU activation function; second transposed convolutional layer: 5×5 transposed convolutional kernel, stride 2, 32 channels, followed by BatchNormalization and ReLU activation function; third transposed convolutional layer: 5×5 transposed convolutional kernel, stride 2, 1 channel, followed by Tanh activation function), embedding a physical constraint module (such as rib thickness ≥5mm, spacing ≥50mm), the input is a 3-dimensional vector (2-dimensional performance requirements + 1-dimensional material label), and the output is normalized rib parameters.

[0083] AI Model Training and Optimization: The Adam optimizer was selected (learning rate 0.0002, β1 = 0.5), with a batch size of 64 and a total of 200 iterations. The first 50 iterations were for pre-training, alternately updating the forward prediction model and the reverse design model. The last 150 iterations were for adversarial training, updating both models simultaneously. Loss functions were set (forward prediction model loss LD = E[log(D(x))] + E[log(1-D(G(z)))]; reverse design model loss LG = E[log(D(G(z)))] + Lpenalty, where Lpenalty is the process constraint penalty term). The model was optimized using the prediction accuracy of the forward prediction model, the process constraint satisfaction rate of the reverse design model, and the withstand voltage value compliance rate as evaluation indicators.

[0084] Intelligent design of pressure hull using AI model: Input design requirements parameters (target water depth / pressure resistance value, weight limit, material type); reverse design model generates a set number of candidate rib parameters (e.g., 10 groups); forward prediction model verifies the performance of candidate parameters and selects the scheme that meets the requirements; optimization algorithm (e.g., NSGA-II) is used to optimize the target (e.g., minimize weight, constrain pressure resistance value ≥ requirement value); output the optimal design scheme and the corresponding simulation verification model (e.g., Abaqus model).

[0085] Example 1

[0086] like Figure 1 As shown, this invention provides an artificial intelligence-based pressure-resistant shell design method constrained by a fusion process of fully convolutional depthwise generative adversarial networks, comprising the following steps:

[0087] (1) Simulation model construction, including: setting the range and step size of the shape parameters of the pressure hull and the material of the pressure hull, setting the range and step size of the rib parameters, modeling with simulation software, calculating the volume and weight of the pressure hull and the maximum pressure resistance value; the simulation software modeling may also include applying gradient water pressure to the outer surface of the pressure hull;

[0088] The simulation model of this invention is an improved parametric multi-condition simulation design, specifically including:

[0089] Define basic parameters: Set the shape parameter range and step size of the pressure shell, for example, the main body of the pressure shell is cylindrical, with a diameter range of 800-2500mm (step size of 100mm) and a length of 1500-4000mm (step size of 200mm); and select the pressure shell material, using carbon fiber composite materials mentioned in the prior art (such as the PMI foam (i.e., polymethacrylimide foam) reinforced composite material in the aforementioned literature 5 "A Composite Material Reinforced Pressure Shell Core Mold Design and Shell Integrated Molding Method", with an elastic modulus of 140GPa and a density of 1.5g / cm³). 3) or titanium alloy (such as TC4 titanium alloy, refer to the commonly used materials for existing pressure shells).

[0090] Set the rib parameter range and step size: thickness 8-40mm (step size 2mm), width 15-80mm (step size 5mm), quantity 6-18 (incrementing to integers), rib distribution spacing 80-400mm (step size 20mm), covering the typical parameter range of "cylindrical ring rib pressure shell" in the prior art.

[0091] Simulation software modeling: In this embodiment, the simulation software Abaqus is used for modeling. The modeling details are as follows: the solid element C3D8I (8-node linear hexahedral non-conforming mode element to improve the buckling analysis accuracy) element in Abaqus is used, and the mesh size is 8-15mm (the mesh size is refined to 8mm at the connection between the rib and the shell); the boundary condition is that one end is fixed and the other end is simply supported (to simulate the actual installation state of the underwater vehicle), and a gradient water pressure (0-80MPa, corresponding to a water depth of 0-8000m, with each water pressure step of 10MPa as one working condition) is applied to the outer surface.

[0092] Performance calculations include calculating the volume and weight of the pressure hull and its maximum pressure resistance; Volume = hull volume (π × diameter × length × hull thickness) + rib volume (cross-sectional area of ​​a single rib × length × number); Weight = volume × material density; Maximum pressure resistance is the critical water pressure at which the pressure hull structure first buckles (using Abaqus).

[0093] The “Buckling Analysis” module calculates the minimum value of the first 5 buckling modes.

[0094] (2) Dataset construction: Generate multiple basic samples based on the parameters of the simulation model;

[0095] This invention provides an improved multi-material-multi-condition fusion dataset, which includes multiple base samples generated based on the parameters of the simulation model, as detailed below:

[0096] Sample size: Generate a set number of basic samples covering a set percentage or more of all rib parameter combinations, including multiple materials, and cover the main parameter combinations of all rib parameter combinations (e.g., generate 12,000 basic samples, covering 75% of all rib parameter combinations). The basic samples of the first material are a first specific number (e.g., 7,000 composite material samples), and the basic samples of the second material are a second specific number (e.g., 5,000 titanium alloy samples). Preferably, for the requirements of high pressure resistance (pressure resistance > 60 MPa) and lightweight (weight < 500 kg), an additional third specific number (e.g., 3,000) of reinforcement samples are generated (reinforcement samples are expanded by parameter interpolation), for a total sample size of 15,000.

[0097] Data labeling: Each sample is labeled with 4 input parameters (e.g., rib thickness / width / quantity / spacing), 2 output labels (e.g., weight, maximum pressure resistance), and an additional material type label (e.g., composite material / titanium alloy) and operating condition label (e.g., water pressure rating).

[0098] Preprocessing: (a) Input parameters are normalized to [0,1], where the normalization formula is:

[0099] xnorm=(x-xmin) / (xmax-xmin);

[0100] Where 'x' is the original parameter value, xmin is the minimum value of the parameter, xmax is the maximum value of the parameter, and xnorm is the normalized value;

[0101] (b) Output label normalization, where the normalization formula is:

[0102] ystd=(y-μ) / σ;

[0103] Where μ is the mean, σ is the standard deviation, y is the output label, and ystd is the standardized value;

[0104] (c) Divide the data into subsets based on material type to avoid cross-material interference.

[0105] (3) AI Model Construction

[0106] An improvement to the AI ​​model in this invention is a fully convolutional DCGAN architecture design. This AI model includes: a forward prediction model (D network) and a reverse design model (G network).

[0107] like Figure 2 As shown, the forward prediction model (D network) includes:

[0108] Input: 4D rib parameters + material type label (encoded as 1D) → concatenate into a 5D vector, and convert into a 32-channel feature map through 1×1 convolution.

[0109] Network structure: Fully convolutional (no fully connected layers or pooling layers), three convolutional layers: the first convolutional layer (corresponding to...) Figure 2 The first convolutional layer ("convolutional layer 1") has a 5×5 convolutional kernel with a stride of 2 and 64 output channels, followed by Batch Normalization (abbreviated as BN in the attached diagram) and LeakyReLU activation function (slope 0.2); the second convolutional layer (corresponding to...) Figure 2The second convolutional layer ("convolutional layer 2") has a 5×5 convolutional kernel, a stride of 2, and 128 output channels, followed by Batch Normalization and LeakyReLU activation (slope 0.2); the third convolutional layer has a 5×5 convolutional kernel, a stride of 1, and 256 output channels, followed by Batch Normalization and LeakyReLU activation. This network structure avoids gradient vanishing and enhances sensitivity to subtle parameter changes.

[0110] Output: A 2D vector (standardized predicted values ​​of weight and pressure resistance), which is converted to actual values ​​through destandardization.

[0111] like Figure 3 As shown, the reverse engineering model (G network) includes:

[0112] Input: 2D performance requirement parameters (weight upper limit, pressure resistance lower limit) + material type label → concatenate into a 3D vector, and convert into a 128-channel feature map through 1×1 convolution.

[0113] Network Structure: Fully convolutional structure, including 3 transposed convolutional layers (5×5 transposed kernel, stride 2, number of channels 128→64→32→1). Specifically, it includes a first transposed convolutional layer, a second transposed convolutional layer, and a third transposed convolutional layer. For example, the first transposed convolutional layer (the 1*1 convolutional layer in the diagram) has a 5×5 transposed kernel, a stride of 2, and outputs 64 channels, followed by Batch Normalization with ReLU activation function; the second transposed convolutional layer (corresponding to convolution 1 in the diagram) has a 5×5 transposed kernel, a stride of 2, and outputs 32 channels, followed by Batch Normalization. Normalization, with ReLU activation function; the third transposed convolutional layer (corresponding to convolution 2 in the figure) has a 5×5 transposed convolutional kernel, stride 2, and 1 output channel, followed by the activation function Tanh (which serves as the output layer and constrains the parameters to the range [0,1]) + LeakyReLU activation function (slope 0.2, consistent with the first two layers).

[0114] Physical constraint module: embeds process feasibility constraints (e.g., rib thickness ≥ 5mm, spacing ≥ 50mm). If the generated parameters violate the constraints, such as the generated parameters rib thickness < 5mm or spacing < 50mm, a penalty loss is imposed on the G network (e.g., the number of violations is one or two).

[0115] (4) Model training and optimization: training and optimizing the AI ​​model;

[0116] Another improvement in this invention is the dual-loop training strategy that integrates process constraints:

[0117] Set the training parameters: optimizer is Adam (learning rate 0.0002, β1 = 0.5); batch size = 64; set the total number of iterations (e.g., 200 iterations), the first iteration (e.g., 50 iterations) is for pre-training (fixed G network trains D network, fixed D network trains G network, alternating), and the second iteration (e.g., 150 iterations) is for adversarial training (D network and G network are updated synchronously).

[0118] The loss function is set as follows: Network Loss LD = E[log(D(x))] + E[log(1-D(G(z)))]. This loss function is used to distinguish between real and generated samples: the first term, E[log(D(x))], represents the expected probability (to be maximized) that the discriminator D classifies a real sample x as true, and the second term, E[log(1-D(G(z)))], represents the expected probability that the discriminator classifies a generated sample G(z) as false. Overall, LD guides D to better distinguish between real and generated data. Network Loss G

[0119] LG = E[log(D(G(z)))] + Lpenalty; This loss function aims to maximize the probability that the generated sample is judged as "real" by the D network (i.e., the E[log(D(G(z)))] term), while introducing a constraint penalty term Lpenalty (such as an additional penalty when the generation parameters violate process constraints) to ensure that the generation result meets the feasibility requirements. Here, Lpenalty represents the process constraint penalty term, which is usually the loss value calculated based on the severity or number of constraint violations.

[0120] Lpenalty calculation method: Constraint checks are performed on each set of candidate rib parameters (thickness t, spacing s) generated by the G network. The loss is calculated by "quantification of violation degree + weighting of penalty coefficient", and the formula is L_{penalty}=

[0121] \lambda\times\left[\max(0,5-t)+\max(0,50-s)\right]where,

[0122] \(\lambda\) is the penalty coefficient (recommended value is 10-20, which can be adjusted according to the training convergence situation to ensure that the penalty is strong enough to correct the violation parameters); \(\max(0,5-t)\) is the thickness violation penalty term (0 when \(t≥5mm\), and "5-t" when \(t<5mm\), the larger the violation difference, the stronger the penalty); \(\max(0,50-s)\) is the spacing violation penalty term (0 when \(s≥50mm\), and "50-s" when \(s<50mm\), the larger the violation difference, the stronger the penalty).

[0123] When the G network generates non-compliant parameters, Lpenalty will significantly increase the total loss of the G network. Through backpropagation, it forces the G network to adjust its parameters to meet process constraints, ultimately ensuring that the generated pressure shell design meets both performance requirements and can be manufactured.

[0124] (5) Intelligent design application: intelligently designing pressure-resistant shells using the trained AI model;

[0125] Another improvement in this invention is the multi-scheme selection mechanism, which specifically includes:

[0126] Input: Design requirements parameters, such as target water depth (or corresponding pressure resistance value), weight limit, and material type.

[0127] The calculation process is as follows: The G network generates a set number of candidate rib parameters (e.g., 10 groups) → The D network quickly verifies the performance of the candidate parameters → The scheme that meets the required parameters is selected → The optimization is performed by an optimization algorithm (e.g., non-dominated sorting genetic algorithm (NSGA-II) (e.g., the optimization objective is to minimize the weight, and the constraint is that the pressure resistance value is ≥ the required value) → The optimal design scheme and the corresponding simulation verification model and Abaqus verification model are output.

[0128] Combination Figure 2 The flowchart of the forward prediction model (D-network) algorithm is shown below, and the algorithm is explained as follows:

[0129] The D network is a fully convolutional structure (without fully connected layers or pooling layers). Downsampling is achieved through convolutions with a stride of 2 (in the three fully convolutional layers of the network, the first two convolutional layers use a stride of 2 for downsampling, while the third convolutional layer uses a stride of 1 without downsampling. The three layers work together to achieve the functions of "feature extraction - downsampling compression - deep feature refinement", and there are no fully connected layers or pooling layers throughout the process to avoid loss of parameter information). This avoids information loss from pooling layers. Batch Normalization stabilizes the training process. The LeakyReLU activation function solves the problem of vanishing gradients on the negative half-axis and improves the sensitivity to subtle changes in rib parameters (such as the effect of rib thickness ±2mm on the pressure resistance value).

[0130] Combination Figure 3 The flowchart of the reverse design model (G-network) algorithm is shown below, and the algorithm of the reverse design model is explained as follows:

[0131] The G network uses transposed convolution to achieve upsampling, which can better preserve the nonlinear mapping relationship between performance and parameters compared with traditional interpolation methods; the Tanh activation function constrains the output parameters to [0,1], which is convenient for inverse normalization to the actual range; the physical constraint module ensures that the generated parameters meet the assembly process requirements of the "segmented pressure cabin" in the prior art (such as the structural strength constraint in the aforementioned reference 6 "A segmented cylindrical pressure cabin of a deep-sea unmanned vehicle and its assembly method").

[0132] As another embodiment of the present invention, the pressure shell body can also be of other shapes, such as spherical or conical pressure shells, with the basic parameters defined according to the shape of the pressure shell. The simulation software Abaqus can be replaced with ANSYS; the present invention is not limited to any particular simulation software. The core of the present invention is parametric modeling and buckling analysis, and software replacement does not affect data validity. The activation function LeakyReLU can be replaced with ELU; both can solve the gradient vanishing problem, requiring only minor adjustments and not changing the fully convolutional DCGAN architecture. The optimization algorithm NSGA-II can be replaced with particle swarm optimization; both are multi-objective optimization algorithms, and as long as the "minimum weight" objective is achieved, it will not affect the design process. The dataset size of 15,000 can be replaced with 12,000; the sample size fluctuates within a reasonable range, as long as it covers the main parameter combinations, it will not affect the model training effect.

[0133] Example 2

[0134] This application also provides an artificial intelligence pressure-resistant shell design device based on the constraints of a fully convolutional depthwise convolutional generative adversarial network fusion process, used to implement the pressure-resistant shell design method in Embodiment 1, including:

[0135] The simulation model building module is used to set the range and step size of the pressure hull shape parameters and the pressure hull material, set the range and step size of the rib parameters, model the simulation software, calculate the volume and weight of the pressure hull and the maximum pressure resistance value; in particular, the simulation software modeling also includes applying gradient water pressure to the outer surface of the pressure hull;

[0136] A dataset construction module is used to generate multiple basic samples based on the parameters of the simulation model;

[0137] The AI ​​model building module is an AI model that is a fully convolutional deep convolutional generative adversarial network architecture. The AI ​​model includes a forward prediction model and a reverse design model. The forward prediction model includes three convolutional layers with a fully convolutional structure, and the reverse design model includes three transposed convolutional layers with a fully convolutional structure.

[0138] A module for training and optimizing the AI ​​model, used to train the AI ​​model using an optimizer;

[0139] A module that uses the AI ​​model to intelligently design pressure-resistant shells.

[0140] Through the detailed description of the above embodiments, at least three improvements of the technical solution of the present invention have been clarified:

[0141] Improvement 1: Fully Convolutional Deep Convolutional Generative Adversarial Network (DCGAN) Architecture

[0142] Technical advantages: By eliminating the FC layer and pooling layer and adopting a fully convolutional structure + transposed convolution upsampling, the problem of parameter information loss caused by pooling in existing AI models (such as the VAE of the aforementioned reference 8 "Intelligent Design Method for Composite Material Layup Based on Generative Neural Network") is solved, thereby improving the accuracy of pressure resistance value prediction (MAE≤5%) and parameter generation accuracy.

[0143] Improvement 2: Dual-loop training strategy incorporating process constraints

[0144] Technical advantages: By embedding process constraints of rib thickness ≥5mm and spacing ≥50mm into the G network, and combining a dual-loop strategy of pre-training (50 rounds) + adversarial training (150 rounds), the problem of existing AI design (such as document 9 published on August 11, 2023: patent CN 116579105 A, titled "Intelligent Design Method for Stainless Steel Structural Components Based on Adversarial Generative Network") is ignored, resulting in unusable solutions. The process constraint satisfaction rate is ≥98%.

[0145] Improvement Point 3: Multi-material, multi-condition datasets and selection mechanism

[0146] Technical advantages: A dataset of 15,000 samples with material and working condition labels is constructed. Combined with the NSGA-II algorithm, the minimum weight scheme that meets the pressure resistance requirement is selected from 10 candidate schemes. This solves the problem that existing pressure shell designs (such as the aforementioned reference 5 "A composite material reinforced pressure shell core mold design and shell integral molding method") only target a single material and cannot adapt to multiple water depth working conditions. It supports the design requirements of composite materials / titanium alloys and water depths of 0-8000m.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. An AI-based pressure-resistant shell design method constrained by a fusion process of fully convolutional depthwise convolutional generative adversarial networks, characterized in that, Includes the following steps: The simulation model construction includes: setting the range and step size of the pressure shell shape parameters and the pressure shell material, setting the range and step size of the rib parameters, modeling with simulation software, and calculating the volume and weight of the pressure shell and the maximum pressure resistance value. Dataset construction involves generating multiple base samples based on the parameters of the simulation model. AI model construction: The AI ​​model is a fully convolutional deep convolutional generative adversarial network architecture. The AI ​​model includes a forward prediction model and a reverse design model. The forward prediction model includes three convolutional layers with a fully convolutional structure, and the reverse design model includes three transposed convolutional layers with a fully convolutional structure. Train and optimize the AI ​​model using an optimizer; The AI ​​model is used to intelligently design a pressure-resistant shell.

2. The artificial intelligence pressure-resistant shell design method according to claim 1, characterized in that, The reverse design model includes a physical constraint module, which embeds process feasibility constraints. If the parameters generated by the reverse design model violate the constraints, a penalty loss is applied to the reverse design model network. Training and optimizing the AI ​​model includes: setting optimizer parameters, setting batch size, setting total number of iterations, performing pre-training of alternating updates of the forward prediction model and the reverse design model in the first iteration before the total number of iterations, and performing adversarial training of synchronous updates of the forward prediction model and the reverse design model in the second iteration after the total number of iterations. The evaluation indicators are set, and the parameters generated by the forward prediction model must meet the set prediction accuracy, while the parameters generated by the reverse design model network must meet the set process constraint satisfaction rate and pressure resistance value compliance rate.

3. The artificial intelligence pressure-resistant shell design method according to claim 1, characterized in that, The dataset construction includes generating a set number of basic samples with material labels to cover the main parameter combinations of all rib parameter combinations, wherein the basic samples include enhanced samples generated for high pressure resistance and lightweight requirements; The intelligent design of pressure-resistant shells using the AI ​​model includes: inputting design requirement parameters, generating a set number of candidate rib parameters using the reverse design model, verifying the performance of the candidate rib parameters using the forward prediction model, selecting design schemes that meet the requirement parameters, optimizing the target through an optimization algorithm, and outputting the optimal design scheme and simulation verification model.

4. The artificial intelligence pressure-resistant shell design method according to claim 1, characterized in that, The three convolutional layers of the fully convolutional structure of the forward prediction model include: a first convolutional layer with a 5×5 kernel, a stride of 2, and 64 output channels, followed by BatchNormalization and an activation function; a second convolutional layer with a 5×5 kernel, a stride of 2, and 128 output channels, followed by BatchNormalization and an activation function; and a third convolutional layer with a 5×5 kernel, a stride of 1, and 256 output channels, followed by BatchNormalization and an activation function. The input to the three convolutional layers is a 5-dimensional vector composed of 4-dimensional rib parameters and 1-dimensional material labels, and the output is a 2-dimensional normalized prediction value composed of weight and pressure resistance.

5. The artificial intelligence pressure-resistant shell design method according to any one of claims 1-4, characterized in that, The three transposed convolutional layers of the fully convolutional structure include a first transposed convolutional layer, a second transposed convolutional layer, and a third transposed convolutional layer. The first transposed convolutional layer has a 5×5 kernel, a stride of 2, and outputs 64 channels, followed by Batch Normalization and an activation function. The second transposed convolutional layer has a 5×5 kernel, a stride of 2, and outputs 32 channels, followed by Batch Normalization and an activation function. The third transposed convolutional layer has a 5×5 kernel, a stride of 2, and outputs 1 channel, followed by an activation function. The input of each of the three transposed convolutional layers is a 3D vector composed of 2D performance requirement parameters and 1D material labels, and the output is a normalized rib parameter.

6. The artificial intelligence pressure-resistant shell design method according to claim 1, characterized in that, The dataset construction includes: generating a set number of basic samples with material labels to cover the main parameter combinations of all rib parameter combinations, wherein the basic samples include enhanced samples generated for high pressure resistance and lightweight requirements; and sample labeling, wherein each basic sample is labeled with input parameters, output labels, material type labels and operating condition labels.

7. The artificial intelligence pressure-resistant shell design method according to claim 2, characterized in that, The training and optimization of the AI ​​model also includes setting a loss function: the loss of the positive prediction model is... LD=E[log(D(x))]+E[log(1-D(G(z)))]; Where E[log(D(x))] represents the expected probability that the discriminator D classifies the true sample x as true. E[log(1-D(G(z)))] represents the expected probability that the discriminator classifies the generated sample G(z) as false; And the loss of the reverse design model is LG=E[log(D(G(z)))]+Lpenalty; Where E[log(D(G(z)))] represents the probability that the generated sample is judged as "real" by the D network; Lpenalty represents the penalty added when the generation parameters violate the process constraints.

8. The artificial intelligence pressure-resistant shell design method according to claim 6, characterized in that, The input parameters include rib thickness, width, number, and spacing; the output markings include the weight of the pressure shell and the maximum pressure resistance value; and the operating condition label includes the water pressure rating.

9. The artificial intelligence pressure-resistant shell design method according to claim 6 or 8, characterized in that, The dataset construction also includes: normalizing the input parameters; standardizing the output labels; and dividing the dataset into subsets according to material type.

10. An AI-based pressure-resistant shell design device constrained by a fusion process of fully convolutional depthwise convolutional generative adversarial networks, used to implement the pressure-resistant shell design method of claim 1, comprising: The simulation model building module is used to set the range and step size of the pressure hull shape parameters and the pressure hull material, set the range and step size of the rib parameters, model the simulation software, and calculate the volume and weight of the pressure hull and the maximum pressure resistance value. A dataset construction module is used to generate multiple basic samples based on the parameters of the simulation model; The AI ​​model building module is an AI model that is a fully convolutional deep convolutional generative adversarial network architecture. The AI ​​model includes a forward prediction model and a reverse design model. The forward prediction model includes three convolutional layers with a fully convolutional structure, and the reverse design model includes three transposed convolutional layers with a fully convolutional structure. A module for training and optimizing the AI ​​model, used to train the AI ​​model using an optimizer; The module uses the AI ​​model to intelligently design pressure-resistant shells.

Images