III-type composite material hydrogen storage container liner-composite layer integrated design method based on deep learning

Abstract

The invention discloses an III type composite material hydrogen storage container liner-composite layer integrated design method based on deep learning. The method comprises the following steps: firstly, determining a design target, design parameters and a value range and constraint conditions thereof according to design requirements, and generating an initial sample data set by adopting Latin hypercube sampling; performance parameters corresponding to all samples are obtained through numerical simulation, and design parameters are input into a Transform model in a structured matrix form for feature extraction and correlation modeling; performing data enhancement by using the trained model to generate an extended sample set; and finally, inputting the samples into a conditional constraint enhanced generative adversarial network, and generating a design scheme which meets various design constraints and achieves optimal balance on various performance indexes through adversarial training. According to the method, all design parameters of the inner container and the composite layer are fully considered, the generated design scheme ensures the optimal balance of the strength performance, the fatigue performance and the material cost of the container, and the design efficiency of similar products is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of hydrogen storage container design, and more specifically, to a deep learning-based integrated design method for the inner liner and composite layer of a Type III composite hydrogen storage container. Background Technology

[0002] Hydrogen energy, as an efficient, clean, and sustainable energy carrier, is a key pillar for achieving global energy transition and carbon neutrality goals. In the hydrogen energy industry chain, hydrogen refueling stations, as crucial infrastructure, rely on stationary hydrogen storage containers—core equipment—which possess significant characteristics such as large capacity, high storage pressure, and long fatigue life. Compared to traditional Type I and Type II containers with high-strength chromium-molybdenum steel liners, Type III containers, employing seamless stainless steel or aluminum liners with high nickel content, exhibit significant advantages in strength, fatigue performance, and lightweight design. These containers utilize stainless steel or aluminum alloys with hydrogen embrittlement resistance and excellent fatigue performance as the liner, combined with a high-specific-modulus and high-specific-strength carbon fiber composite layer, enabling them to withstand higher hydrogen storage pressures and meet longer fatigue life requirements.

[0003] The main failure modes of hydrogen storage containers include strength failure and fatigue failure. Type III hydrogen storage containers employ a fully covered carbon fiber composite layer structure, exhibiting significantly superior strength and fatigue performance compared to single-layer metal containers. The internal pressure of the container is primarily borne by the composite layer; strength failure depends mainly on the composite layer design, while fatigue failure is primarily determined by the inner liner design. To ensure the container's strength and fatigue performance, the parameters of the inner liner and composite layer must be designed as a single unit. However, existing design methods typically consider burst failure and fatigue failure separately. Furthermore, since existing Type III and Type IV gas cylinders are mainly used in vehicle applications, their hydrogen storage pressure and fatigue life requirements are relatively low. Additionally, to meet lightweight requirements, the inner liner thickness is usually thin, and designs generally only focus on the composite layer parameters, with less consideration given to the impact of the inner liner parameters on the overall strength and fatigue performance of the container.

[0004] In summary, numerous parameters influence the strength and fatigue performance of containers, including the outer diameter, wall thickness, and end cap shape of the inner liner, as well as the thickness, angle, and arrangement sequence of the composite layers. The complex relationships between these parameters are difficult to quantify directly using traditional theoretical formulas. Therefore, artificial intelligence methods are needed to uncover their underlying patterns and quantitatively guide container design. Relying solely on existing calculation formulas, design experience, and finite element simulations for the design of the inner liner and composite layers not only fails to comprehensively consider the impact of numerous design parameters but may also lead to over-design, resulting in material waste and increased costs. Therefore, establishing an integrated design method for the parameters of the inner liner and composite layers that comprehensively considers container strength, fatigue performance, and economic requirements is of significant practical importance for improving the design efficiency and performance analysis capabilities of hydrogen storage containers. Summary of the Invention

[0005] This invention addresses the problem that existing technologies struggle to integrate the design of the inner liner and composite layer parameters of hydrogen storage containers while simultaneously considering strength, fatigue performance, and economic requirements. It proposes a deep learning-based integrated design method for the inner liner and composite layer of a Type III composite hydrogen storage container.

[0006] The technical solution adopted in this invention is: an integrated design method for the inner liner and composite layer of a Type III composite hydrogen storage container based on deep learning, which includes the following steps:

[0007] Step 100: Determine the design parameters and performance parameters of the container based on the structure of the Type III composite hydrogen storage container; the design parameters include the inner liner parameters and the composite layer parameters, and the performance parameters include strength performance parameters, fatigue performance parameters and cost parameters;

[0008] Step 101: Determine the value range and design constraints of each design parameter according to the design requirements and production process, and use Latin hypercube sampling to generate an initial sample set with uniform spatial distribution;

[0009] Step 102: Use a random generation algorithm to generate random layering schemes for the sampling results, and then visualize and extract features;

[0010] Step 103: Establish finite element models for all samples to calculate strength and fatigue performance, extract mass parameters, and calculate material costs;

[0011] Step 200: Establish an attention-based Transformer deep learning model, using the feature matrix containing all container design parameters as input, to directly predict the container performance parameters obtained from finite element analysis, and establish a mapping relationship from the design space to the performance space; search for the optimal hyperparameter combination and use all simulation samples to complete the training of the Transformer deep learning model;

[0012] Step 300: Use the trained Transformer deep learning model to perform data augmentation to expand the training samples, establish and train a conditional constraint-enhanced generative adversarial network, and achieve end-to-end generation from performance requirements to the optimal design scheme through a composite loss function of adversarial training, performance matching and multi-objective optimization. This ensures that the output scheme automatically balances the three key performance parameters of strength performance, fatigue performance and cost while meeting all engineering constraints.

[0013] Step 301: Establish a finite element model for the optimized design parameters and calculate the strength and fatigue performance. Compare the prediction results with the prediction results of the Transformer deep learning model to compare the prediction errors and ensure that they meet the design requirements. Then the generated design scheme has engineering feasibility.

[0014] Furthermore, in step 100, the inner liner parameters include four parts: the cylinder body, the end cap, the neck, and the material; the cylinder body is cylindrical, and the design parameters include the outer diameter d. l Wall thickness t l The length L; the head adopts an ellipsoidal shape, and the design parameters include the ellipsoid ratio ε and the top thickness t. h The bottleneck structure consists of an outer diameter dimension d. o and inner diameter d i The inner liner material is determined jointly based on the characteristics of the high-pressure hydrogen environment. When stainless steel is selected, it is denoted as k1, and when aluminum alloy is selected, it is denoted as k2. The composite layer parameters are characterized by the layup thickness, winding angle, and arrangement.

[0015] Furthermore, in step 101, the bottleneck thickness and the top thickness of the end cap should be greater than the cylinder wall thickness. The outer diameter of the cylinder must meet the minimum size requirements while also considering structural rationality. The water volume of the inner liner must meet the design specifications while also considering manufacturing errors. The number of composite layers is calculated based on grid theory, allowing for a certain offset, and must be an even number. A fixed number of hole expansions is used during spiral winding, thus limiting the number of winding angles, whose values ​​are randomly selected within a certain range. The main constraints are expressed as follows:

[0016] t h -t l ≥0

[0017]

[0018] d l -d design ≥0

[0019]

[0020] α0=arcsin(d o / d l )

[0021]

[0022] In the formula: d design V represents the minimum outer diameter of the cylinder as required by the design. design For the required inner liner volume, Δd l ΔL represents the manufacturing deviation of the outer diameter of the cylinder, ΔL represents the manufacturing deviation of the cylinder length; α0 represents the spiral winding angle tangential to the pole hole; n α n is the number of spiral layers. θ P represents the number of circumferential layers. b To design the burst pressure, K is the helical fiber strength utilization coefficient, σ b t represents the tensile strength of the composite material along the fiber direction. s Z represents the thickness of a single layer of the composite layer, and Z is a set of integers.

[0023] Furthermore, in step 102, each sampling result constitutes a sample data point, corresponding to all the parameters required for a container design. This data is first represented in a list, with list elements categorized into inner liner parameters and composite layer parameters, represented as [d]. l ,t l ,L,ε,t h ,d o ,d i ,k i [θ1, θ2, θ3, θ4...], the list length is fixed at the maximum value of the composite layer number plus eight, where θ i This represents the actual layup angle of the i-th winding layer. The spiral layers appear in alternating positive and negative directions. Each angle represents two actual layups. Redundant angles are filled with 0. The graph is converted into a bar chart, with the horizontal axis representing parameter order and the vertical axis representing parameter magnitude. The value ranges of all inner liner parameters are uniformly mapped according to the bar height of the winding angle. Different bar colors represent inner liner parameters, spiral winding composite layers, and circumferential winding composite layers, respectively. The graph size and bar width are fixed. For zero-filled θ... i The height of the bar is also 0, but it still occupies the corresponding position. For this bar chart, the feature matrix is ​​extracted. The number of columns in the matrix is ​​7 to represent the features, and the number of rows is the list length to represent the feature parameters corresponding to each bar. The 7 features are bar height, bar center X coordinate, bar center Y coordinate, bar color, and bar order. The height and coordinate parameters are determined by pixel coordinates and normalized according to the pixel size of the graphic. The bar order is also normalized according to the total number.

[0024] Further, in step 103, for each sample data, an axisymmetric finite element model of the hydrogen storage container is constructed in the finite element software using a parametric modeling method. Simultaneously, material properties are assigned, a mesh is generated, boundary conditions are established, and the calculation is submitted. The boundary conditions refer to the actual working environment of the hydrogen storage container. Internal pressure is applied to the inner surface of the inner liner, one end of the bottleneck is fixed, and an equivalent tensile stress is applied to the other end. Binding constraints are added between the inner liner and the composite layer to simulate the interface connection characteristics. The calculation conditions are divided into two categories: strength conditions and fatigue conditions. In the strength condition, only the design burst pressure is applied. The proportion of fiber tensile failure elements η is calculated by transforming the composite layer element coordinate system to the fiber direction and based on the three-dimensional Hashin failure criterion, thus characterizing the container's strength performance. In the fatigue condition, self-tightening pressure and fatigue cycle pressure are applied. The circumferential stress distribution along the cylinder wall thickness path is extracted, and the fatigue life N of the inner liner is calculated using fracture mechanics methods, thus characterizing the container's fatigue performance. The mass parameters of the inner liner and the composite layer are extracted through the finite element model, and the total material cost P is calculated based on the material unit price, expressed as:

[0025]

[0026] σ θ =A0 + A1(x / a) + A2(x / a) 2 +A3(x / a) 3

[0027]

[0028] P = c α M l +c β M c

[0029] Where: σ xyz For the stress in the global coordinate system, σ 123 The stress is in the fiber orientation coordinate system. X is the coordinate transformation matrix; N' is the total number of elements in the finite element model, X T S represents the tensile strength of the composite layer in the fiber direction. L σ is the shear strength of the composite layer; θ For the circumferential stress of the inner liner, A i Here are the fitting coefficients, x is the fitting variable, a is the crack depth, K1 is the stress intensity factor, and G is the crack depth. i As a correction factor, A p t is the fatigue upper limit pressure, l is the crack length; a' is the initial crack depth, t l For the cylinder wall thickness, K Imax and K Imin These are the stress intensity factors ΔK and ΔK, respectively, under the working conditions corresponding to the upper and lower fatigue pressure limits. I For K Imax and K Imin The difference, where C and m are fracture mechanics parameters in the Paris formula for materials; M l and M c The quality of the inner liner and the composite layer are respectively, c α and c β These are the average material prices for the inner liner and the composite layer, respectively.

[0030] Further, in step 200, the Transformer deep learning model first maps the input features to a high-dimensional space through a linear projection layer, and uses learnable parameterized position encoding to capture the spatial relationships between features. The preprocessed data is then input into a core network consisting of three stacked encoder layers. Each encoder layer contains two key substructures: a multi-head self-attention mechanism to effectively capture the complex dependencies between input features, and a feedforward neural network to realize nonlinear feature transformation. All sublayers use residual connections and layer normalization to alleviate the gradient vanishing problem, and introduce Dropout operation to prevent overfitting. Finally, the model performs global average pooling on the encoder output and simultaneously outputs the three key performance parameters of the container through a multi-task learning framework.

[0031] Further, in step 200, the training process of the Transformer deep learning model is as follows: all performance parameters are normalized; the AdamW optimizer is used, and the update magnitude of each parameter is dynamically adjusted through an adaptive learning rate mechanism and explicit weight decay regularization to suppress overfitting; gradient clipping is introduced to strictly limit the gradient norm; the learning rate scheduling scheme adopts the OneCycleLR strategy, keeping the maximum learning rate synchronized with the base learning rate; and a two-stage dynamic adjustment mechanism integrating linear warm-up and cosine annealing is integrated—gradual warm-up is used in the early stage of training to ensure parameter update stability, while periodic changes in the learning rate are used in the middle and later stages to break through local optima and achieve the best balance between convergence speed and model accuracy; different weighted loss terms are assigned to different performance indicators to adapt to their statistical characteristics; and an early stopping mechanism is introduced, terminating the training process when the average validation set loss over multiple training cycles fails to show a significant improvement. This is expressed as:

[0032]

[0033] In the formula: η base Let η be the initial learning rate. max T represents the maximum learning rate, t represents the current training step number, and T represents the maximum learning rate. up y represents the number of steps during the warm-up phase, and T represents the total number of training steps; i and Let represent the predicted and actual values ​​of the i-th target parameter, respectively. and These are the Huber loss function, SmoothL1 loss function, and LogCosh loss function, respectively, ω i S represents the weights corresponding to the loss function for each performance parameter; t It is an early stopping function. The validation set loss is w, where w is the sliding window size, k is the number of consecutive tolerances, and ∈ is the threshold for significant improvement.

[0034] Furthermore, in step 200, fine-tuning is performed on key hyperparameters including maximum learning rate, batch size, number of network channels, number of attention heads, and total number of training steps. The optimization process includes two stages: first, a global exploration of the hyperparameter space is conducted through a random search strategy to quickly locate potential high-performance regions; then, a high-precision grid search is performed within the identified advantageous subspace to accurately lock in the optimal hyperparameter combination through denser sampling and finer granularity.

[0035] Further, in step 300, the data augmentation specifically involves: firstly, reasonably scaling and randomly perturbing the inner liner design parameters, with a maximum perturbation amplitude of 20% but not exceeding the parameter value range; simultaneously, ensuring through constraint conditions that all perturbed parameters still strictly meet the physical limitations in the design requirements; then, for the angle parameters in the composite layer, only the spiral winding angle of the expanded hole is perturbed, and the order of some lay-up layers is randomly exchanged; finally, a predetermined number of sample data are generated, and all perturbed design parameters are input into the trained Transformer deep learning model to predict their corresponding performance parameters.

[0036] Further, in step 300, the conditionally constrained enhanced generative adversarial network comprises two parts: a generator and a discriminator. The generator takes a Gaussian distributed noise vector and a condition vector representing the performance requirements of the container as joint inputs, and generates a design parameter feature matrix that meets the constraints through generator mapping. The discriminator simultaneously receives the real or generated design parameter feature matrix and its corresponding condition vector, on the one hand evaluating the authenticity of the design parameters, and on the other hand verifying their matching degree with the performance conditions. Both the generator and the discriminator adopt a fully connected layer to construct the core architecture. In each hidden layer, LeakyReLU is used as the activation function, and a batch normalization layer is introduced to enhance training stability and prevent overfitting. The output layer of the generator is designed with a parameterized constraint activation function to ensure that the output design parameter feature matrix strictly meets the constraints. The loss function of the discriminator comprises four parts: the basic adversarial loss to ensure the authenticity of the generated samples, the performance matching loss to force the design to meet the target performance index, the multi-objective optimization term to promote the synergistic optimization of strength performance, fatigue performance and material cost, and the constraint penalty term to strengthen physical feasibility in the form of regularization. These four parts are weighted and summed to form the final optimization objective.

[0037] The beneficial effects of this invention are reflected in:

[0038] This invention overcomes the limitations of traditional composite hydrogen storage containers that rely on empirical design and grid theory calculations, and innovatively proposes a deep learning-based intelligent design method. This method, by establishing a deep learning model, can predict performance and automatically generate compliant design schemes based on design requirements, while simultaneously balancing key performance indicators such as container strength, fatigue life, and material cost. Compared with conventional design methods, this invention has significant technical advantages: it can significantly improve the strength and fatigue performance of the container while ensuring that design constraints are met, effectively reduce material costs and avoid redundant design, greatly shorten the design cycle and improve design efficiency, and automatically generate multiple sets of compliant design schemes, providing a better selection space for engineering design decisions. Attached Figure Description

[0039] Figure 1 A flowchart illustrating the implementation of an integrated design method for the inner liner and composite layer of a Type III composite hydrogen storage container based on deep learning, provided by this invention.

[0040] Figure 2 This is a schematic diagram illustrating the sample data collection process provided in an embodiment of the present invention;

[0041] Figure 3 A schematic diagram illustrating data type conversion and the Transformer model provided in an embodiment of the present invention;

[0042] Figure 4 The prediction results of the model provided in the embodiments of the present invention;

[0043] Figure 5 A comparison of the computation results of the Transformer model provided in this embodiment of the invention with other conventional machine learning models;

[0044] Figure 6 Numerical simulation results for the final design scheme provided in the embodiments of the present invention. Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. The accompanying drawings are only schematic representations related to specific embodiments and do not represent the entirety of the invention.

[0046] Figure 1 A flowchart illustrating the implementation of an integrated design method for the inner liner and composite layer of a Type III composite hydrogen storage container based on deep learning, provided for embodiments of the present invention, includes the following steps:

[0047] Step 100: Determine the design and performance parameters of the Type III composite hydrogen storage container based on its structure. The design parameters include two categories: inner liner parameters and composite layer parameters. The inner liner parameters are further subdivided into four parts: cylinder, end caps, neck, and materials. The cylinder is cylindrical, and its main design parameters include the outer diameter d. l Wall thickness t l The length L; the head adopts an ellipsoidal shape, and the design parameters include the ellipsoid ratio ε (the ratio of the major axis to the minor axis) and the top thickness t. h The bottleneck structure is determined by the outer diameter d. o and inner diameter d i The inner liner material is selected from stainless steel (k1) or aluminum alloy (k2) based on the characteristics of the high-pressure hydrogen environment. The composite layer parameters are characterized by features such as layup thickness, winding angle and arrangement. An actual layup scheme that meets engineering constraints is constructed based on an adaptive random generation algorithm. The performance evaluation system focuses on three core indicators: strength performance parameters, fatigue performance parameters and cost parameters.

[0048] Step 101: Determine the value range and design constraints of each design parameter according to the design requirements and production process. The thickness of the bottleneck and the top of the end cap should be greater than the wall thickness of the cylinder. The outer diameter of the cylinder must meet the minimum size requirements and take into account the structural rationality. The water volume of the inner liner must meet the design indicators and also take into account the manufacturing error. The number of composite layers is calculated according to the grid theory and a certain offset is allowed. To simplify the design, a fixed number of hole expansions is used when spiral winding, thus limiting the number of winding angles. The value of the angle is randomly selected within a certain range. The layup scheme of the composite layer is randomly generated based on these parameters.

[0049] Specifically, the water volume of the inner tank must be no less than 1m³. 3(The maximum deviation in cylinder length is 10mm, and the maximum deviation in outer diameter is 2mm). The outer diameter of the cylinder should not be less than 406mm, and based on the available steel pipe specifications and actual production conditions, the outer diameter can only be 426, 457, 480, 508, and 530mm. The wall thickness should be 15-25mm. The cylinder length is constrained by the inner liner volume. The ellipsoid ratio of the end cap should be 1.0-2.0, and the thickness of the top of the end cap can be 1-2 times the wall thickness. The inner diameter of the bottleneck should be 40-80mm, and the outer diameter of the bottleneck should be 120-160mm. In addition, to ensure the sealing at the bottleneck, the bottleneck thickness should not be less than 30mm. The inner liner material can be 316L stainless steel or 6061-T6 aluminum alloy. The design pressure is 102 MPa, and the design burst pressure is 253 MPa. The inner liner is required to have a fatigue life of no less than 20,000 cycles under full-amplitude pressure cycling. The composite layer material is T700 carbon fiber, with a single layer thickness of 0.386 mm. The number of composite layers is calculated based on mesh theory, allowing for a certain offset, and must be an even number. To simplify the design, the number of hole enlargements during spiral winding is fixed at 6, meaning the number of spiral winding angles is 7. Except for the smallest winding angle α0 tangent to the pole hole, the remaining spiral winding angles are randomly selected between [α0+5, 60]. The composite layer layup scheme is randomly generated based on these parameters. The main constraints are as follows:

[0050] t h -t l ≥0

[0051]

[0052] d l -d design ≥0

[0053]

[0054] α0=arcsin(d o / d l )

[0055]

[0056] In the formula: d design V represents the minimum outer diameter of the cylinder as required by the design. design For the required inner liner volume, Δd l ΔL represents the manufacturing deviation of the cylinder's outer diameter, ΔL represents the manufacturing deviation of the cylinder's length, and deviations of other parameters have a minor impact on volume and are not considered here; α0 is the helical winding angle tangential to the pole hole; n α n is the number of spiral layers. θ P represents the number of circumferential layers. b To determine the burst pressure, K is the helical fiber strength utilization coefficient, typically taken as 0.6–0.8, and σb t represents the tensile strength of the composite material along the fiber direction. s Let Z be the thickness of a single layer of the composite layer, and Z be a set of integers.

[0057] Step 102: Based on the value range of each design parameter and design constraints, perform Latin hypercube sampling, generating a total of 1000 results. Each sampling result constitutes a sample data point, corresponding to all the parameters required for a container design. First, represent these results in a list. The list elements are mainly divided into two categories: inner liner parameters and composite layer parameters, represented as [d...]. l ,t l ,L,ε,t h ,d o ,d i ,k i [θ1, θ2, θ3, θ4...], the list length is fixed at the maximum value of the number of composite layers plus eight (62 in this embodiment), where θ i This represents the actual layup angle of the i-th winding layer. Spiral layers often appear in alternating positive and negative directions, so each angle here represents two actual layups. Redundant angles are filled with 0. The graph is converted into a bar chart, with the horizontal axis representing the parameter order and the vertical axis representing the parameter magnitude. The bar height is 0-90 to match the winding angles. All inner liner parameters are mapped to a range of 10-80, where material type k... i There are only two possible values, hence they are represented by 30 and 60 respectively. The bars have three colors: gray, blue, and red, representing the inner liner parameters, the spiral winding composite layer, and the circumferential winding composite layer, respectively. The graphic size and bar width are fixed. For zero-fill θ... i The height of the bar is also 0, but it still occupies the corresponding position. For this bar chart, the feature matrix is ​​extracted. The number of columns in the matrix is ​​7 to represent the features, and the number of rows is the list length to represent the feature parameters corresponding to each bar. The 7 features are bar height, bar center X coordinate, bar center Y coordinate, bar color (represented by one-hot encoding, with three colors corresponding to 3 features), and bar order. The height and coordinate parameters are determined by pixel coordinates and normalized according to the pixel size of the graphic. The bar order is also normalized according to the total number.

[0058] Step 103: For each sample data, an axisymmetric finite element model of the hydrogen storage container is constructed in the finite element software Abaqus using a parametric modeling method. Material properties are assigned, meshes are generated, boundary conditions are established, and the calculation is submitted. The inner liner material can be either 316L stainless steel or 6061-T6 aluminum alloy, and the composite layer material is T700 carbon fiber. The material parameters are shown in Tables 1-3.

[0059] Table 1. Mechanical parameters of 316L stainless steel

[0060] E / MPa μ <![CDATA[R p / MPa]]> <![CDATA[R m / / MPa]]> <![CDATA[ρ / g cm -3 ]]> m C <![CDATA[K Ic / MPa*m 1 / 2 ]]> 200 0.3 279 546 7.98 3.61 <![CDATA[9.72*10 -13 ]]> 197.4

[0061] Table 2. Mechanical parameters of 6061-T6 aluminum alloy

[0062] E / MPa μ <![CDATA[R p / MPa]]> <![CDATA[R m / / MPa]]> <![CDATA[ρ / g cm -3 ]]> m C <![CDATA[K Ic / MPa*m 1 / 2 ]]> 69 0.33 283 312 2.70 2.3 <![CDATA[6.53*10 -10 ]]> 121.6

[0063] Table 3. Mechanical parameters of T700 carbon fiber

[0064] <![CDATA[E1 / GPa]]> <![CDATA[E2 / GPa]]> <![CDATA[G 12 / GPa]]> <![CDATA[G 23 / GPa]]> <![CDATA[ν 12 ]]> <![CDATA[ρ / g cm -3 ]]> <![CDATA[X T / MPa]]> <![CDATA[S L / MPa]]> 134 7.42 3.71 4.79 0.28 1.8 2530 50

[0065] Boundary conditions were based on the actual working environment of the hydrogen storage container. Internal pressure was applied to the inner surface of the liner, one end of the bottleneck was fixed, and equivalent tensile stress was applied to the other end. Binding constraints were added between the liner and the composite layer to simulate the interface connection characteristics. The calculation conditions were divided into two categories: strength conditions and fatigue conditions. In the strength condition, only the design burst pressure was applied. The proportion of fiber tensile failure elements η was calculated based on the three-dimensional Hashin failure criterion by transforming the coordinate system of the composite layer elements to the fiber direction, thus characterizing the strength performance of the container. In the fatigue condition, self-tightening pressure and fatigue cycle pressure were applied. The circumferential stress distribution along the wall thickness path of the cylinder was extracted and the fatigue life N of the liner was calculated using fracture mechanics methods, serving as a key indicator for evaluating the fatigue performance of the container. In addition, the mass parameters of the liner and the composite layer were extracted using the finite element model, and the total material cost P was calculated based on the unit price of the materials, achieving a quantitative assessment of the container's economic efficiency. The entire data collection process was as follows: Figure 2 As shown, it is represented as:

[0066]

[0067] σ θ =A0 + A1(x / a) + A2(x / a) 2 +A3(x / a) 3

[0068]

[0069] P = c α M l +c β M c

[0070] Where: σ xyz The stress in the global coordinate system, σ 123 The stress is in the fiber orientation coordinate system. X is the coordinate transformation matrix; N' is the total number of elements in the finite element model, X T S represents the tensile strength of the composite layer in the fiber direction. L For the shear strength of the composite layer, and Let be the stress at a certain element in the fiber direction coordinate system. This represents an indicator function, which takes a value of 1 when the unit meets the failure condition and 0 otherwise, and is used to count the number of failed units; σ θ For the circumferential stress of the inner liner, A i The fitting coefficients are calculated based on the extracted circumferential stress distribution along the cylinder wall thickness path. x is the fitting variable, a is the crack depth, K1 is the stress intensity factor, and G... i As a correction factor, A p t is the fatigue upper limit pressure, l is the crack length; a' is the initial crack depth, t l For the cylinder wall thickness, K Imax and K Imin These are the stress intensity factors ΔK and ΔK, respectively, under the working conditions corresponding to the upper and lower fatigue pressure limits. I For K Imax and K Imin The difference, C and m are the fracture mechanics parameters in the Paris formula for materials, and N is the fatigue life of the inner liner; M l and M c The quality of the inner liner and the composite layer are respectively, c α and c β The values ​​are the average material prices of the inner liner and the composite layer, respectively. The crack is assumed to be an axial-radial semi-elliptical crack on the inner surface of the inner liner, with a depth-to-length ratio of 1:3. The initial crack depth is taken as 0.5 mm based on the level of non-destructive testing.

[0071] Step 200: Establish a deep learning model based on the Transformer architecture. Using a feature matrix containing all container design parameters as input, directly predict the container performance parameters obtained from finite element analysis. The model first maps the input features to a high-dimensional space through a linear projection layer and employs learnable parameterized positional encoding to capture the spatial relationships between features. The preprocessed data is then input into a core network consisting of three stacked encoder layers. Each encoder layer contains two key substructures: a multi-head self-attention mechanism to effectively capture the complex dependencies between input features, and a feedforward neural network to implement nonlinear feature transformation. All sublayers use residual connections and layer normalization to alleviate the gradient vanishing problem and introduce Dropout to prevent overfitting. Finally, the model performs global average pooling on the encoder output and simultaneously outputs the three key performance indicators of the container through a multi-task learning framework. The model structure and data transformation process are as follows: Figure 3 As shown;

[0072] Step 201: Train the model based on the established database. Normalize all performance parameters and divide all samples into training and test sets in a 7:3 ratio. Use the AdamW optimizer and dynamically adjust the update magnitude of each parameter through an adaptive learning rate mechanism and explicit weight decay regularization to suppress overfitting. Furthermore, to ensure numerical stability during training, gradient clipping is introduced to strictly limit the gradient norm. The learning rate scheduling scheme adopts the OneCycleLR strategy, keeping the maximum learning rate synchronized with the base learning rate. A two-stage dynamic adjustment mechanism integrating linear warm-up and cosine annealing is integrated—gradual warm-up ensures parameter update stability in the early training phase (first 15% of training steps), while periodic learning rate changes in the mid-to-late stages help overcome local optima and achieve the best balance between convergence speed and model accuracy. Different weighted loss terms are assigned to different performance metrics to adapt to their statistical characteristics, and an early stopping mechanism is introduced. When the average validation set loss over multiple training cycles fails to show significant improvement, the training process is terminated, expressed as:

[0073]

[0074] In the formula: η base Let η be the initial learning rate. max T represents the maximum learning rate, t represents the current training step number, and T represents the maximum learning rate. up y represents the number of steps during the warm-up phase, and T represents the total number of training steps; i and Let represent the predicted and actual values ​​of the i-th target parameter, respectively. and These are the Huber loss function, SmoothL1 loss function, and LogCosh loss function, respectively, ω i Here, we assign weights to the loss functions for each performance parameter, taking values ​​of 1, 0.8, and 0.6 respectively; S t This is an early stopping function that stops training prematurely when certain constraints are met. For the validation set loss, w is the sliding window size, which is 3 here; k is the number of consecutive tolerances, which is 10 here; and ∈ is the threshold for significant improvement, which is the average of the three performance parameters R. 2 The coefficients are used for evaluation, with a value of 0.001. m, n, i, j, and t are all index variables.

[0075] Step 202: To maximize model performance, fine-tuning is performed on key hyperparameters including maximum learning rate, batch size, number of network channels, number of attention heads, and total training steps. The optimization process consists of two stages. First, a random search strategy is used to globally explore the hyperparameter space to quickly locate potential high-performance regions. Then, a high-precision grid search is performed within the identified advantageous subspace. More dense sampling and finer granularity are used to precisely pinpoint the optimal hyperparameter combination. This hierarchical optimization method ensures both the comprehensiveness of the search process and significantly improves optimization efficiency, ensuring the best model configuration is obtained within a reasonable computational cost. The final model hyperparameters are: maximum learning rate 3*102 -4 The initial learning rate is 1.5 * 10^6 -5 The final learning rate is 1.5 * 10^6 -8 The batch size was 256, the network channels were 1024, the number of attention heads was 4, and the total number of training steps was 400. The model's predictive performance was evaluated using root mean square error (RMSE) and R² coefficient of determination. The final training results are as follows: Figure 4 As shown;

[0076] Step 203: For this training task, conventional machine learning models are used for training, including Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Tree (XGBoost), and Artificial Neural Network (ANN). The model input is directly a sequence of design parameters, and the output is the model's three corresponding performance parameters. The same hyperparameter optimization strategy is used to optimize the hyperparameters of each model. The optimization results are shown in Table 4. After processing and partitioning the data using the same data preprocessing method, the model is trained. The root mean square error (RMSE) and R² coefficient of determination are used to evaluate the model's prediction performance. The prediction results are compared with those of the Transformer. Figure 5 As shown, the Transformer model demonstrates significant advantages in performance prediction tasks, with its prediction accuracy being significantly improved compared to traditional machine learning methods. This performance leap is mainly attributed to the Transformer's unique self-attention mechanism, which can efficiently capture the complex nonlinear relationship between design parameters and performance indicators. At the same time, by modeling long-distance dependencies, it significantly improves the ability to represent container design parameters.

[0077] Table 4. Hyperparameter Optimization of Conventional Machine Learning Models

[0078]

[0079] Step 300: Use the trained Transformer model to perform data augmentation to generate more high-quality sample data. First, the inner liner design parameters are reasonably scaled and randomly perturbed. The perturbation amplitude is up to 20% but does not exceed the parameter value range. At the same time, the constraints ensure that all perturbed parameters still strictly meet the physical limitations in the design requirements. Then, for the angle parameters in the composite layer, only the spiral winding angle of the hole is perturbed and the order of some lay-up (angle parameters) is randomly swapped. Finally, a total of 3000 sample data are generated. All perturbed design parameters are input into the pre-trained Transformer performance prediction model to predict their corresponding performance parameters.

[0080] Step 301: Based on the enhanced data, establish a conditionally constrained enhanced generative adversarial network (CC-GAN). The model consists of a generator and a discriminator. The generator takes a 128-dimensional Gaussian noise vector and a condition vector representing the performance requirements of the container as joint inputs. The condition vector sets basic requirements for the performance parameters, which are: the proportion of failed cells η is less than 0.25, the fatigue life of the inner liner N is greater than 22,000, and the total material cost P is less than 200,000. Other value ranges and basic design constraints remain the same as in the initial design. The generator maps and generates a feature matrix of design parameters that meets the constraints. The discriminator simultaneously receives the feature matrix of the actual or generated design parameters and their corresponding condition vectors. On the one hand, it evaluates the authenticity of the design parameters, and on the other hand, it verifies their matching degree with the performance conditions. The entire system ensures that all output design parameters meet the constraints of the design requirements through a strict constraint processing mechanism. On the basis of meeting the minimum performance requirements specified by the condition vector, it further optimizes the key indicators: minimizing the proportion of failed cells, maximizing the fatigue life of the inner liner, and minimizing the material cost of the container.

[0081] Step 302: Both the generator and discriminator employ fully connected layers to construct their core architecture. LeakyReLU is used as the activation function in each hidden layer, and a batch normalization layer is introduced to enhance training stability and prevent overfitting. The generator's output layer features a specially designed parameterized constraint activation function to ensure that the output's design parameter feature matrix strictly satisfies the constraints. The discriminator's loss function comprises four parts: a basic adversarial loss to ensure the authenticity of generated samples; a performance matching loss to force the design to meet the target performance indicators; a multi-objective optimization term to promote the synergistic optimization of strength performance, fatigue performance, and material cost; and a constraint penalty term to strengthen physical feasibility in a regularized form. These four losses are weighted and summed to form the final optimization objective, expressed as:

[0082]

[0083] In the formula: G represents the generator, D represents the discriminator, D(x|c) represents the discriminator's score on the real design parameter x under condition c, D(G(z|c)) represents the discriminator's score on the generated design parameter G(z|c), λ1 is the adversarial loss weight with a value of 1.0, T(G(z|c)) is the performance prediction of the Transformer model for the generated design parameter, c is the target performance condition vector, λ2 is the performance matching loss weight with a value of 10.0, and the deviation between the predicted value and the target value is quantified using the square of the L2 norm, T i (G(z|c) are three predicted performance parameters, α, β, γ are the weights of each parameter with a weight of 0.5, 1.0 and 0.8 respectively, λ3 is the weight of the multi-objective optimization term with a weight of 5.0, used to further optimize the performance index under the premise of satisfying the basic requirement of c, ∑φ(G(z|c)) is the constraint violation measurement function, λ4 is the constraint penalty term with a weight of 2.0, which forces the design constraints to be met;

[0084] Step 303: Train the model to obtain the final design parameters, establish the finite element model corresponding to the design parameters and perform numerical simulation calculations. First, compare the consistency between the performance parameters predicted by the Transformer model and the actual simulation results to evaluate the accuracy of the prediction model. Second, comprehensively check whether the design parameters and their corresponding performance strictly meet the preset technical requirements. When the simulation results match the predicted values ​​and all parameters meet the design constraints, it can be confirmed that the intelligently generated design scheme has engineering feasibility.

[0085] Specifically, the final design parameters obtained by training the model are shown in Table 5. These parameters fully meet all design constraints. Numerical simulations of the finite element model based on these parameters show that the failure element ratio is 0.2346, the liner fatigue life is 221,369 cycles, and the material cost is 181,310.13 yuan. These performance indicators all meet the design requirements and are highly consistent with the predicted values ​​of the Transformer model (failure element ratio 0.2270, liner fatigue life 216,895 cycles, material cost 179,303.76 yuan), with relative errors all less than 5%. This verifies the excellent accuracy of the Transformer model in performance prediction and also confirms the engineering feasibility of the generated design scheme. The numerical simulation results of the final design scheme are shown in Table 5. Figure 6 As shown;

[0086] Table 5. Final Design Parameters

[0087]

[0088] The above is a further detailed description of the present invention in conjunction with preferred embodiments, and is not intended to limit the present invention. It should be noted that for those skilled in the art, any simple deductions and optimizations made to the present invention based on the core ideas of the present invention should be considered within the protection scope of the present invention.

Claims

1. A deep learning-based integrated design method for the inner liner and composite layer of a Type III composite hydrogen storage container, characterized in that, Includes the following steps: Step 100: Determine the design parameters and performance parameters of the container based on the structure of the Type III composite hydrogen storage container; the design parameters include the inner liner parameters and the composite layer parameters, and the performance parameters include strength performance parameters, fatigue performance parameters and cost parameters; Step 101: Determine the value range and design constraints of each design parameter according to the design requirements and production process, and perform Latin hypercube sampling; Step 102: Use a random generation algorithm to generate random layering schemes for the sampling results, and then visualize and extract features; Step 103: Establish finite element models for all samples to calculate strength and fatigue performance, extract mass parameters, and calculate material costs; Step 200: Build a Transformer deep learning model, using the feature matrix containing all container design parameters as input, to directly predict the container performance parameters obtained from finite element analysis; search for the optimal combination of hyperparameters and train the Transformer deep learning model. Step 300: Use the trained Transformer deep learning model to perform data augmentation, establish and train a conditionally constrained augmented generative adversarial network, so that it can achieve synergistic optimization of the three key performance parameters of strength performance, fatigue performance and cost while meeting the basic performance requirements; Step 301: Establish a finite element model for the optimized design parameters and calculate the strength and fatigue performance. Compare the prediction results with the prediction results of the Transformer deep learning model to compare the prediction errors and ensure that they meet the design requirements. Then the generated design scheme has engineering feasibility.

2. The method according to claim 1, characterized in that, In step 100, the inner liner parameters include four parts: the cylinder, the end cap, the neck, and the material; the cylinder is cylindrical, and the design parameters include the outer diameter d. l Wall thickness t l The length L; the head adopts an ellipsoidal shape, and the design parameters include the ellipsoid ratio ε and the top thickness t. h The bottleneck structure consists of an outer diameter dimension d. o and inner diameter d i The inner liner material is determined jointly based on the characteristics of the high-pressure hydrogen environment. When stainless steel is selected, it is denoted as k1, and when aluminum alloy is selected, it is denoted as k2. The composite layer parameters are characterized by the layup thickness, winding angle, and arrangement.

3. The method according to claim 2, characterized in that, In step 101, the bottleneck thickness and the top thickness of the end cap should be greater than the cylinder wall thickness. The outer diameter of the cylinder must meet the minimum size requirements while also considering structural rationality. The water volume of the inner liner must meet the design specifications while also considering manufacturing errors. The number of composite layers is calculated based on grid theory, allowing for a certain offset, and must be an even number. A fixed number of hole expansions is used during spiral winding, thus limiting the number of winding angles, whose values ​​are randomly selected within a certain range. The main constraints are expressed as follows: t h -t l ≥0 d l -d design ≥0 α0=arcsin(d o / is l ) In the formula: d design V represents the minimum outer diameter of the cylinder as required by the design. design For the required inner liner volume, Δd l ΔL represents the manufacturing deviation of the outer diameter of the cylinder, ΔL represents the manufacturing deviation of the cylinder length; α0 represents the spiral winding angle tangential to the pole hole; n α n is the number of spiral layers. θ P represents the number of circumferential layers. b To design the burst pressure, K is the helical fiber strength utilization coefficient, σ b t represents the tensile strength of the composite material along the fiber direction. s Z represents the thickness of a single layer of the composite layer, and Z is a set of integers.

4. The method according to claim 2, characterized in that, In step 102, each sampling result constitutes a sample data point, corresponding to all the parameters required for a container design. This data is first represented in a list, with list elements categorized into inner liner parameters and composite layer parameters, represented as [d...]. l ,t l ,L,ε,t h ,d o ,d i ,k i [θ1, θ2, θ3, θ4...], the list length is fixed at the maximum value of the composite layer number plus eight, where θ i This represents the actual layup angle of the i-th winding layer. The spiral layers appear in alternating positive and negative directions. Each angle represents two actual layups. Redundant angles are filled with 0. The graph is converted into a bar chart, with the horizontal axis representing parameter order and the vertical axis representing parameter magnitude. The value ranges of all inner liner parameters are uniformly mapped according to the bar height of the winding angle. Different bar colors represent inner liner parameters, spiral winding composite layers, and circumferential winding composite layers, respectively. The graph size and bar width are fixed. For zero-filled θ... i The height of the bar is also 0, but it still occupies the corresponding position. For this bar chart, the feature matrix is ​​extracted. The number of columns in the matrix is ​​7 to represent the features, and the number of rows is the list length to represent the feature parameters corresponding to each bar. The 7 features are bar height, bar center X coordinate, bar center Y coordinate, bar color, and bar order. The height and coordinate parameters are determined by pixel coordinates and normalized according to the pixel size of the graphic. The bar order is also normalized according to the total number.

5. The method according to claim 4, characterized in that, In step 103, for each sample data, an axisymmetric finite element model of the hydrogen storage container is constructed in the finite element software using a parametric modeling method. Material properties are assigned, a mesh is generated, boundary conditions are established, and the calculation is submitted. The boundary conditions refer to the actual working environment of the hydrogen storage container. Internal pressure is applied to the inner surface of the liner, one end of the bottleneck is fixed, and an equivalent tensile stress is applied to the other end. Binding constraints are added between the liner and the composite layer to simulate the interface connection characteristics. The calculation conditions are divided into strength conditions and fatigue conditions. In the strength condition, only the design burst pressure is applied. The proportion of fiber tensile failure elements η is calculated by transforming the composite layer element coordinate system to the fiber direction and based on the three-dimensional Hashin failure criterion, thus characterizing the container's strength performance. In the fatigue condition, self-tightening pressure and fatigue cycle pressure are applied. The circumferential stress distribution along the cylinder wall thickness path is extracted, and the fatigue life N of the liner is calculated using fracture mechanics methods, thus characterizing the container's fatigue performance. The mass parameters of the liner and composite layer are extracted using the finite element model, and the total material cost P is calculated based on the material unit price, expressed as: σ θ =A0+A1(x / a)+A2(x / a) 2 +A3(x / a) 3 P=c α M l +c β M c Where: σ xyz The stress in the global coordinate system, σ 123 The stress is in the fiber orientation coordinate system. X is the coordinate transformation matrix; N' is the total number of elements in the finite element model, X T S represents the tensile strength of the composite layer in the fiber direction. L σ is the shear strength of the composite layer; θ For the circumferential stress of the inner liner, A i Here are the fitting coefficients, x is the fitting variable, a is the crack depth, K1 is the stress intensity factor, and G is the crack depth. i As a correction factor, A p t is the fatigue upper limit pressure, l is the crack length; a' is the initial crack depth, t l For the cylinder wall thickness, K Imax and K Imin These are the stress intensity factors ΔK and ΔK, respectively, under the working conditions corresponding to the upper and lower fatigue pressure limits. I For K Imax and K Imin The difference, where C and m are fracture mechanics parameters in the Paris formula for materials; M l and M c The quality of the inner liner and the composite layer are respectively, c α and c β These are the average material prices for the inner liner and the composite layer, respectively.

6. The method according to claim 1, characterized in that, In step 200, the Transformer deep learning model first maps the input features to a high-dimensional space through a linear projection layer and uses learnable parameterized position encoding to capture the spatial relationships between features. The preprocessed data is then input into the core network consisting of three stacked encoder layers. Each encoder layer contains two key substructures: a multi-head self-attention mechanism to effectively capture the complex dependencies between input features and a feedforward neural network to achieve nonlinear feature transformation. All sublayers use residual connections and layer normalization to alleviate the gradient vanishing problem and introduce Dropout operations to prevent overfitting. Finally, the model performs global average pooling on the encoder output and simultaneously outputs the three key performance parameters of the container through a multi-task learning framework.

7. The method according to claim 1, characterized in that, In step 200, the training process of the Transformer deep learning model is as follows: all performance parameters are normalized; the AdamW optimizer is used, and the update magnitude of each parameter is dynamically adjusted through an adaptive learning rate mechanism and explicit weight decay regularization to suppress overfitting; gradient clipping is introduced to strictly limit the gradient norm; the learning rate scheduling scheme adopts the OneCycleLR strategy, keeping the maximum learning rate synchronized with the base learning rate; and a two-stage dynamic adjustment mechanism of linear warm-up and cosine annealing is integrated—gradual warm-up is used in the early stage of training to ensure the stability of parameter updates, while periodic changes in the learning rate are used in the middle and later stages to break through the local optimum dilemma and achieve the best balance between convergence speed and model accuracy; different weighted loss terms are assigned to different performance indicators to adapt to their statistical characteristics. An early stopping mechanism is introduced, which terminates the training process when the average validation set loss over multiple training epochs fails to show a significant improvement. This is expressed as: In the formula: η base Let η be the initial learning rate. max T represents the maximum learning rate, t represents the current training step number, and T represents the maximum learning rate. up y represents the number of steps during the warm-up phase, and T represents the total number of training steps; i and Let represent the predicted and actual values ​​of the i-th target parameter, respectively. and These are the Huber loss function, SmoothL1 loss function, and LogCosh loss function, respectively, ω i The weights corresponding to the loss function for each performance parameter; S t For early stopping functions, The validation set loss is w, where w is the sliding window size, k is the number of consecutive tolerances, and ∈ is the threshold for significant improvement.

8. The method according to claim 1, characterized in that, In step 200, fine-tuning is performed on key hyperparameters including maximum learning rate, batch size, number of network channels, number of attention heads, and total number of training steps. The optimization process includes two stages: first, a global exploration of the hyperparameter space is carried out through a random search strategy to quickly locate potential high-performance regions; then, a high-precision grid search is performed within the identified advantageous subspace to accurately lock in the best combination of hyperparameters through denser sampling and finer granularity.

9. The method according to claim 1, characterized in that, In step 300, the data augmentation specifically involves: firstly, reasonably scaling and randomly perturbing the inner liner design parameters, with a maximum perturbation amplitude of 20% but not exceeding the parameter value range; simultaneously, ensuring through constraint conditions that all perturbed parameters still strictly meet the physical limitations in the design requirements; then, for the angle parameters in the composite layer, only the spiral winding angle of the expanded hole is perturbed, and the order of some lay-up layers is randomly exchanged; finally, a predetermined number of sample data are generated, and all perturbed design parameters are input into the trained Transformer deep learning model to predict their corresponding performance parameters.

10. The method according to claim 1, characterized in that, In step 300, the condition-constrained enhanced generative adversarial network includes two parts: a generator and a discriminator. The generator takes a Gaussian distributed noise vector and a condition vector representing the performance requirements of the container as joint inputs. The generator maps and generates a feature matrix of design parameters that meets the constraints. The discriminator simultaneously receives the real or generated feature matrix of design parameters and its corresponding condition vector. On the one hand, it evaluates the authenticity of the design parameters, and on the other hand, it verifies the degree of matching between them and the performance conditions. Both the generator and discriminator employ fully connected layers to construct their core architecture. LeakyReLU is used as the activation function in each hidden layer, and a batch normalization layer is introduced to enhance training stability and prevent overfitting. The generator's output layer is designed with a parameterized constraint activation function to ensure that the output design parameter feature matrix strictly meets the constraint conditions. The discriminator's loss function consists of four parts: a basic adversarial loss to ensure the authenticity of the generated samples; a performance matching loss to force the design to meet the target performance indicators; a multi-objective optimization term to promote the synergistic optimization of strength performance, fatigue performance, and material cost; and a constraint penalty term to strengthen physical feasibility in the form of regularization. These four parts are weighted and summed to form the final optimization objective.

Images