Engine Hood Structure Optimization Method and Device

Abstract

This invention discloses a method and apparatus for optimizing engine hood structure, belonging to the technical field of vehicle body structure optimization. First, based on the hard point gap distribution in the engine compartment, the functional zoning of the energy-absorbing core layer is completed, and honeycomb, negative Poisson's ratio, and honeycomb composite sandwich structures are customized. Then, a dual-stream encoder multimodal deep learning proxy model is constructed, introducing uncertainty quantification and active learning closed-loop to ensure accuracy. Finally, the optimal solution is output through dual-objective multi-objective optimization and simulation verification. While meeting regulatory and mass production requirements, it achieves synergistic optimization of pedestrian protection and lightweighting, significantly improving optimization efficiency and result reliability.

Description

Technical Field

[0001] This invention relates to the field of technology, and in particular to a method and apparatus for optimizing engine hood structure. Background Technology

[0002] The hood of a car should absorb energy effectively during a collision to reduce the Head Injury Criterion (HIC) and minimize the risk of head injuries to pedestrians.

[0003] Existing engine hood energy-absorbing structures mostly use uniform honeycomb, foam, or coarse-grained reinforcing components, and structurally they are mostly single energy-absorbing platforms without a graded energy management mechanism; the optimization process relies on large-scale finite element scanning to participate in empirical debugging and uses conventional proxy models to carry out optimization.

[0004] Currently, the low-clearance area of ​​the engine hood is prone to insufficient energy absorption stroke and high HIC, while the high-clearance area has redundant energy absorption and increased weight. The force and displacement characteristics make it difficult to achieve soft start and bottom collision prevention. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for optimizing engine hood structure, so as to alleviate the technical problem that existing automotive engine hood energy-absorbing structures cannot simultaneously achieve low HIC values ​​for pedestrian head protection and lightweight structure under the dual constraints of limited gaps at engine compartment hard points and mass production.

[0006] In a first aspect, the present invention provides a method for optimizing an engine hood structure, comprising: Based on the geometric relationship between the hard point gap in the engine compartment and the engine cover, the composite sandwich structure of the energy-absorbing core layer in the cover is determined, and the energy-absorbing core layer is partitioned to output the structural design variables corresponding to the composite sandwich structure of each functional partition. The structural design variables were experimentally designed and finite element simulations were performed according to pedestrian protection conditions to obtain multimodal input data and simulation indicators for model training, and the trained optimized surrogate model was determined. Based on the aforementioned optimized proxy model, multi-objective optimization and verification are completed according to the aforementioned structural design variables to determine the optimal structural scheme of the engine cover.

[0007] In an optional embodiment, the composite sandwich structure consists of an upper honeycomb unit layer, an intermediate negative Poisson's ratio unit layer, and a lower honeycomb unit layer stacked sequentially along the length direction; both the upper and lower honeycomb unit layers are constructed using multiple interconnected hexagonal honeycomb units to form a load-bearing and energy-absorbing frame, and the intermediate negative Poisson's ratio unit layer is constructed using multiple axisymmetric concave arc-shaped re-entry negative Poisson's ratio units connected to each other, and the negative Poisson's ratio units undergo lateral outward expansion deformation when subjected to vertical compression.

[0008] In an optional implementation, the steps of determining the composite sandwich structure of the energy-absorbing core layer in the engine cover based on the geometric relationship between the hard point gap in the engine compartment and the engine cover, partitioning the energy-absorbing core layer, and outputting the structural design variables corresponding to the composite sandwich structure of each functional partition include: Based on the vertical gap distribution between the hard points in the engine compartment and the engine cover, and the geometric contours of the inner and outer plates in the cover, the composite sandwich structure of the energy-absorbing core layer in the cover is determined; wherein, the cover includes an outer plate, the composite sandwich structure and the inner plate connected sequentially along the thickness direction; The energy-absorbing core layer is divided in the plane into a low-gap region directly above the hard point, a transition region connecting the low-gap region and the high-gap region, and a high-gap region away from the hard point. Output the unit geometric parameters corresponding to each functional area and the engineering feasible range of the unit geometric parameters as structural design variables.

[0009] In an optional implementation, the steps of experimentally designing the structural design variables and performing finite element simulations under pedestrian protection conditions to obtain multimodal input data and simulation indices for model training, and determining the trained optimized surrogate model, include: Using the structural design variables as input, parameter samples and corresponding multimodal input data are generated through experimental design; wherein, the parameter samples are the discretized values ​​of the structural design variables of each functional area, the multimodal input data includes multimodal image data and the structured design variables corresponding to the parameter samples, and the multimodal image data includes geometric schematic diagrams of the inner and outer plates of the cover, a cross-sectional view of the head impact point, and a heat map of the structural parameter field distribution; Using the parameter samples as input, finite element simulation is performed according to pedestrian protection conditions, and simulation indicators are output to characterize pedestrian protection performance and structural quality. Using the multimodal input data as features and the simulation indicators as labels, a multimodal deep learning prediction model is trained to obtain an optimized surrogate model with the required accuracy.

[0010] In an optional implementation, the step of training a multimodal deep learning prediction model using the multimodal input data as features and the simulation indicators as labels to obtain an optimized surrogate model with achieved accuracy includes: The multimodal deep learning prediction model adopts a cascaded architecture consisting of a dual-stream encoder, a cross-modal feature fusion module, and a multi-task decoder. The multimodal input data is input into the multimodal deep learning prediction model and processed by the dual-stream encoder to determine the image features characterizing the local stiffness differences and gap distribution of the cover body and the parameter features characterizing the nonlinear coupling and long-range dependence between the parameters. After the image features and parameter features are fused by the cross-modal feature fusion module, the multi-task decoder outputs the prediction results and corresponding uncertainty estimates of the engine cover performance indicators under pedestrian protection impact conditions. Based on the prediction results, the total loss function, composed of the uncertainty estimate and the simulation index, iterates the weights of the multimodal deep learning prediction model until the total loss function converges, thus obtaining the trained optimized surrogate model.

[0011] In an optional implementation, before the step of determining the optimal structural scheme of the engine cover by performing multi-objective optimization and verification based on the optimized proxy model and the structural design variables, the method further includes: The prediction accuracy of the optimized surrogate model after training is verified using reserved test set samples outside the training set. If the accuracy of the optimized proxy model does not reach the preset accuracy threshold, then based on the uncertainty estimate, high-value samples with prediction uncertainty higher than the preset uncertainty threshold and located in the preset key area and their corresponding label data are added to the training set, and the optimized proxy model is incrementally trained and its accuracy is re-verified until the accuracy of the optimized proxy model reaches the preset accuracy threshold.

[0012] In an optional implementation, the step of determining the optimal structural scheme of the engine cover by performing multi-objective optimization and verification based on the optimized surrogate model and the structural design variables includes: Using the optimized proxy model as a prediction tool and the upper and lower limits of the structural design variables as the optimization boundaries, a multi-objective optimization problem that balances pedestrian protection performance and lightweight design is constructed to obtain the Pareto optimal solution set. The frontier representative solution is extracted from the compliant and valid solution set corresponding to the Pareto optimal solution set, and then the full-condition finite element simulation is performed for verification. If the verification result deviates from the prediction result of the optimized surrogate model by more than the preset accuracy threshold, the parameter samples and simulation indicators corresponding to the current Pareto optimal solution set are fed back into the training set, and the optimized surrogate model is incrementally trained and then multi-objective optimization is re-executed. After the Pareto front converges iteratively, the engineering compromise solution is selected from the set of compliant and valid solutions, and the optimal structural scheme of the engine cover is output.

[0013] In a second aspect, the present invention provides an engine hood structure optimization device, comprising: The design module determines the composite sandwich structure of the energy-absorbing core layer in the engine cover based on the geometric relationship between the hard point gap in the engine compartment and the engine cover, and performs partitioning processing on the energy-absorbing core layer, outputting the structural design variables corresponding to the composite sandwich structure of each functional partition. The module determines the structural design variables, conducts experimental design and performs finite element simulation according to pedestrian protection conditions, obtains multimodal input data and simulation indicators for model training, and determines the trained optimized surrogate model. The optimization module, based on the optimization proxy model, performs multi-objective optimization and verification according to the structural design variables to determine the optimal structural scheme of the engine cover.

[0014] Thirdly, the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method as described in any of the foregoing embodiments.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed, implements the method described in any of the foregoing embodiments.

[0016] This invention provides a method and apparatus for optimizing engine hood structures. First, a composite sandwich energy-absorbing structure is customized based on the hard point gaps in the engine compartment and the geometric relationship of the hood body, and functional partitioning is completed. The structural design variables corresponding to each partition are output, achieving precise adaptation between the energy-absorbing structure and the hard point gap distribution from the design source. This solves the root cause problems of poor protection performance in low-gap areas and redundant stiffness in high-gap areas in traditional uniform structures. Second, a multimodal dataset is constructed through experimental design and finite element simulation, and a multimodal deep learning optimization surrogate model is trained. This overcomes the limitation of traditional surrogate models in integrating geometric gap features and structured parameters, providing a high-precision and high-reliability prediction tool for subsequent efficient optimization. Finally, multi-objective optimization and verification are completed based on the optimization surrogate model to determine the optimal structural scheme, forming a complete closed loop from structural design to mass production scheme output. This achieves synergistic optimization of pedestrian protection performance and lightweighting, providing a replicable and scalable end-to-end solution for engine hood energy-absorbing structure design.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart of an engine hood structure optimization method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a partitioned gradient honeycomb sandwich structure for an engine hood provided in an embodiment of the present invention; Figure 3 A schematic diagram of a finished product of a partitioned gradient honeycomb sandwich layer for an engine hood provided in an embodiment of the present invention; Figure 4 A collision diagram of an engine hood provided as an embodiment of the present invention; Figure 5 This is a schematic diagram of a unit structure of a partitioned gradient honeycomb sandwich layer for an engine hood, provided in an embodiment of the present invention. Figure 6 A flowchart of another engine hood structure optimization method provided in an embodiment of the present invention; Figure 7 A schematic diagram of a deep learning network architecture applied to an engine hood structure optimization method provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the functional modules of an engine hood structure optimization device provided in an embodiment of the present invention; Figure 9 A schematic diagram of the hardware architecture of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Research has revealed that as the automotive industry increasingly demands lightweight design and collision safety, the hood, as the first line of defense for pedestrian protection, is receiving growing attention for its energy-absorbing structure and optimized design. Currently, hoods commonly employ a composite structure of thin metal sheets and honeycomb / foam sandwich layers, dissipating collision energy through plastic deformation. Traditional designs treat the honeycomb core layer as a homogeneous material or a simple partitioned structure, failing to adequately consider the differences in gaps between various components within the engine compartment (such as the engine block, hinge brackets, and locking mechanisms) and the inner surface of the hood. This homogeneous design leads to hard contact in low-gap areas (typically 3-5mm) during the initial stages of a collision. Due to insufficient effective deformation stroke and hindered contact patch expansion, this easily results in excessively high peak acceleration and head injury index (HIC). Conversely, high-gap areas (up to 15-20mm) suffer from material redundancy and additional weight due to excessive deformation space.

[0023] Based on this, the engine hood structure optimization method and device provided in this embodiment of the invention can achieve a significant improvement in the head protection performance of the engine hood and a lightweight structure at the same time, under the premise of meeting the requirements of pedestrian protection and mass production manufacturing, by customizing the composite sandwich structure design, constructing a multimodal deep learning agent model adapted to the structural characteristics, and an uncertainty-driven active learning and optimization closed loop. The optimization calculation efficiency and result reliability have achieved a leap in magnitude, and it has strong engineering feasibility.

[0024] To facilitate understanding of this embodiment, a method for optimizing the engine hood structure disclosed in this embodiment of the invention will first be described in detail. This method can be applied to intelligent control devices such as host computers, controllers, and servers.

[0025] Figure 1 A flowchart of an engine hood structure optimization method provided in an embodiment of the present invention.

[0026] Reference Figure 1 The method may include the following steps: Step S102: Based on the geometric relationship between the hard point gap in the engine compartment and the engine cover, determine the composite sandwich structure of the energy-absorbing core layer in the cover, and perform partitioning of the energy-absorbing core layer, outputting the structural design variables corresponding to the composite sandwich structure of each functional partition. This step uses the vertical clearance distribution between the engine compartment hard points and the hood as the core design input, breaking through the empirical design approach of the traditional engine hood with a uniform structure across the entire area. First, it customizes and adapts a composite sandwich energy-absorbing core layer for the differentiated distribution of hard point clearances. Then, it completes planar functional partitioning according to the impact risk level, and finally outputs the parametric design variables corresponding to each partition. Specifically, it utilizes the mapping and matching design of hard point clearance structural stiffness, differentiated partitioning based on impact risk, and partition-level parametric modeling to provide clear input boundaries for subsequent optimization. In this way, it fundamentally solves the core pain points of poor pedestrian protection performance in the low clearance area directly above the hard points of traditional engine hoods and redundant stiffness in the high clearance area, achieving pre-matching of pedestrian protection, bending stiffness, and lightweighting. At the same time, it transforms the complex structural optimization into an optimization problem of a finite number of partition design variables, significantly reducing the dimensionality and computational cost of subsequent optimization.

[0027] It should be noted that the microstructure of the composite sandwich energy-absorbing core layer can be as follows: Figure 2 As shown; firstly, Figure 2 The composite sandwich structure consists of an upper honeycomb unit layer 13, a middle negative Poisson's ratio unit layer 12, and a lower honeycomb unit layer 11 stacked sequentially along the length direction; The sandwich-style layered architecture of the energy-absorbing core layer is evident. It employs two upper and lower hexagonal honeycomb structures as the load-bearing framework, with a negative Poisson's ratio structure in the middle layer serving as the core energy-absorbing unit. This overcomes the limitations of traditional single-cell honeycomb sandwich structures, which suffer from low energy absorption efficiency and poor impact stability. Utilizing the interlayer functional division of labor between honeycomb load-bearing and negative Poisson's ratio energy absorption, the upper and lower honeycomb layers ensure in-plane bending stiffness and out-of-plane load-bearing capacity, while the middle negative Poisson's ratio layer is responsible for gradient energy absorption during impact. This structure addresses the pain point of traditional single-cell honeycomb sandwich structures where bending stiffness and energy absorption performance cannot be simultaneously achieved. The upper and lower honeycomb layers ensure the modal, stiffness, and dent resistance of the hood, while the middle negative Poisson's ratio layer significantly improves impact energy absorption efficiency. Furthermore, the layered architecture boasts strong compatibility in molding processes, adapting to various mass production processes such as aluminum alloy stamping, composite material molding, and 3D printing, demonstrating exceptional engineering feasibility.

[0028] Based on this, the upper cellular unit layer 11 and the lower cellular unit layer 13 both adopt multiple interconnected hexagonal cellular units to form a load-bearing and energy-absorbing frame. The middle negative Poisson's ratio unit layer 12 is composed of multiple axisymmetric concave arc reentry negative Poisson's ratio units connected to each other. When the negative Poisson's ratio unit is subjected to vertical compression, it undergoes lateral outward expansion deformation.

[0029] Here, the upper and lower layers employ mechanically stable hexagonal honeycomb units, while the middle layer uses axisymmetric, concave, arc-shaped re-entry negative Poisson's ratio units. Each unit consists of an upper unit plate, a lower unit plate, a first unit side plate, a second unit side plate, a third unit side plate, and a fourth unit side plate, forming a cavity. The left and right connecting parts are recessed into the cavity, creating a rotational-outward extension mechanism that undergoes lateral outward tensile deformation upon vertical impact. Utilizing the negative Poisson's ratio effect, the impact contact area is rapidly expanded within a short stroke, reducing peak acceleration and HIC value, and preventing hard-point bottoming.

[0030] For impact scenarios in the low-gap zone directly above the hard point, the lateral outward expansion deformation of the negative Poisson's ratio unit can effectively prevent the impact head from directly hitting the hard point, significantly reducing head injury value and solving the industry pain point that pedestrian protection performance in the low-gap zone is difficult to meet standards; compared with the traditional linear negative Poisson's ratio unit, the concave arc reentry unit eliminates stress concentration at sharp corners, improving the fatigue performance, impact stability and energy absorption efficiency of the structure.

[0031] In practical application examples, the definition of composite sandwich structures begins with material selection. The materials for composite sandwich structures can be selected based on vehicle positioning and cost requirements. Mass-produced vehicles prioritize materials such as 5052 aluminum alloy sheets, while high-end or new energy vehicles may use carbon fiber reinforced composite materials or glass fiber reinforced polypropylene materials.

[0032] Then, the unit parameters are designed. The upper and lower honeycomb unit layers adopt a regular hexagonal honeycomb structure, such as... Figure 5 As shown, the cell wall thickness of the upper and lower honeycomb units (d) is a, and the cell-to-side distance is b. The arrangement direction of the honeycomb units is consistent with the main load-bearing direction of the engine hood. The middle negative Poisson's ratio unit layer adopts an axisymmetric concave arc-shaped re-entry unit. The distance c between the two vertical reference lines represents the net distance at the waist of the unit. The letter d represents the unit wall thickness, which is a uniform wall thickness for the entire unit. The letter e represents the radius of the outermost circular arc on the left and right. The letters f and g are the vertices of the upper and lower re-entry triangles, respectively, which can be equilateral or determined according to the target curve. The circumscribed height and circumscribed width of the negative Poisson's ratio unit are determined by the upper and lower cusps and the outer contour of the outer circular arc. Together with c, d, e, f, and g, they are used to define the geometry of the unit. The lateral pitch of the negative Poisson's ratio unit is consistent with the cell-to-side distance of the honeycomb unit to ensure the alignment of the interlayer connections.

[0033] In addition, the forming and connection process design is carried out. The composite sandwich structure of aluminum alloy is formed by integral stamping and vacuum brazing to ensure the interlayer bonding strength. The composite material is formed by integral molding to avoid the performance degradation caused by secondary bonding. In order to improve the structural robustness, micropores can be set on the side wall of the negative Poisson unit and covered with threshold diaphragms to achieve rate-related damping, or shearable microbridges can be set at the connecting limbs to achieve two-stage load bearing for soft start and hard finish.

[0034] Finally, the energy absorption performance of the composite sandwich structure was verified through uniaxial compression tests, and the bending stiffness was verified through three-point bending tests to ensure that the structural performance meets the design requirements.

[0035] In some embodiments, such as Figure 3 As shown, the composite sandwich structure, after being assembled with the upper edge sealing plate 21 and the lower edge sealing plate 22, forms a composite sandwich structure 23 that can be used for partitioning; based on the energy-absorbing core layer of this composite sandwich structure, further refined partitioning is performed, including: Step 1.1: Based on the vertical clearance distribution between the hard points in the engine compartment and the engine cover, and the geometric contours of the inner and outer plates in the cover, determine the composite sandwich structure of the energy-absorbing core layer in the cover. Among them, such as Figure 4 As shown, the cover includes an outer plate 31 connected sequentially along the thickness direction, a partitioned composite sandwich structure 32, and an inner plate 33. An impact 34, simulated by a pedestrian head model, can be applied to the outer plate 31. Here, the design input of the composite sandwich structure and the stacking relationship of the engine hood assembly are known. With the vertical gap distribution of hard points and the geometric contours of the inner and outer plates as the core design boundaries, the structure can achieve precise matching with the overall vehicle layout. By utilizing the positive mapping design of the thickness energy absorption performance of the gap distribution structure, the total thickness and interlayer parameters of the composite sandwich structure are customized according to the gap size of different areas, thereby achieving precise adaptation of the composite sandwich structure with the overall vehicle engine compartment layout, avoiding the risk of hard point bottoming out from the design source; at the same time, clarifying the stacking architecture of the engine hood assembly can provide clear geometric boundaries for subsequent finite element simulation and structural optimization.

[0036] Step 1.2: Divide the energy-absorbing core layer in the plane into a low-gap region directly above the hard point, a transition region connecting the low-gap region and the high-gap region, and a high-gap region away from the hard point; Here, based on the level of impact risk at hard points, the planar functional zoning of the energy-absorbing core layer is completed. The core logic is that areas with higher impact risk have higher priority in energy absorption performance design, which differs from the traditional design approach based on shape and block division. Utilizing zoning rules based on hard point gaps and impact injury risk, the high-risk area directly above the hard point is separately designated as a low-gap zone. A transition zone is used to achieve a smooth transition in performance, avoiding stress concentration caused by abrupt stiffness changes. This achieves differentiated zoning design based on impact risk, concentrating design resources in the high-risk low-gap zone, significantly improving the efficiency of pedestrian protection optimization. Through the connection design of the transition zone, abrupt stiffness changes between different zones are avoided, ensuring the overall modal, dent resistance, and fatigue performance of the hood.

[0037] Step 1.3: Output the unit geometric parameters and the engineering feasible range of the unit geometric parameters corresponding to each functional area, as structural design variables.

[0038] Here, for each functional zone, the geometric parameters of the corresponding cellular units and negative Poisson's ratio units, as well as the engineering feasible range that meets the requirements of mass production manufacturing, are output, transforming the complex structural design into quantifiable optimization variables. Utilizing zone-level parametric modeling, the design variables for each zone are independent and can be adjusted individually, achieving precise control over the performance of each zone. This achieves a fully parametric definition of the engine hood structure, providing clear input variables for subsequent experimental design, surrogate model training, and multi-objective optimization steps. By limiting the engineering feasible range, parameter schemes that cannot be mass-produced are avoided from the design source, ensuring the manufacturability of subsequent optimization results.

[0039] In practical application examples, for the partitioning of composite sandwich structures, the first step is to acquire hard point gap data. Using a three-dimensional digital model of the entire vehicle, the minimum vertical gap between all hard points in the engine compartment and the inner panel of the engine hood is measured, generating a gap distribution heat map of the entire engine hood area to clarify the hard point gap value at each location. Then, partition boundaries are defined. Based on the gap distribution heat map, the area directly above the hard points with specific vertical gaps, such as ≤10mm, is designated as a low-gap zone. This area is a high-risk area for pedestrian head impact and has the highest design priority. The area with gaps of 10mm to 25mm is designated as a transition zone to connect the low-gap zone and the high-gap zone, achieving a smooth transition in stiffness. The area with gaps ≥25mm and far from the hard points is designated as a high-gap zone. This area has a low risk of impact bottoming out and its design priority is lightweighting and stiffness assurance.

[0040] Next, a zonal parametric design is carried out. For the low-gap region, energy absorption performance is prioritized, and the design variables are based on the geometric parameters of the negative Poisson's ratio unit. For the transition region, energy absorption performance and stiffness are considered simultaneously, and the design variables are the combined parameters of the honeycomb unit and the negative Poisson's ratio unit. For the high-gap region, lightweighting and stiffness are prioritized, and the design variables are based on the parameters of the honeycomb unit. At the same time, the engineering feasible range of each parameter is set, and a 10~20mm wide gradient band is formed at the boundary of the zone through continuous parameter interpolation to avoid abrupt changes in stiffness.

[0041] Finally, a fully parameterized composite sandwich structure model of the engine hood was constructed in 3D software to achieve real-time driven updates of design variables, providing a model foundation for subsequent simulation and optimization.

[0042] Step S104: Experimentally design the structural design variables and perform finite element simulation according to the pedestrian protection working condition to obtain multimodal input data and simulation indicators for model training, and determine the trained optimized surrogate model; Based on the structural design variables partitioned in the aforementioned embodiments, samples are generated by experimental design covering the entire design space. Finite element simulations are performed based on the pedestrian protection regulations in GB24550-2009 to obtain true performance values. Finally, a multimodal deep learning optimization surrogate model is constructed and trained. Specifically, this can be achieved through optimal Latin hypercube experimental design for multi-objective optimization, compliant finite element simulations, and multimodal heterogeneous data-driven surrogate model construction. The surrogate model replaces time-consuming finite element simulations to achieve rapid performance prediction. This overcomes the limitations of traditional surrogate models, which can only handle structured parameters and cannot integrate geometric gap information. Multimodal input data simultaneously covers structural parameters and hard-point gap geometric features, laying a data foundation for high-precision surrogate model construction. The performance evaluation time for a single structural scheme is reduced from several hours to milliseconds, solving the industry pain point of extremely low computational efficiency in traditional multi-objective optimization.

[0043] In some embodiments, the step of determining the trained optimized agent model in step S104 includes: Step 2.1: Using structural design variables as input, generate parameter samples and corresponding multimodal input data through experimental design; Among them, the parameter samples are the discretized values ​​of the structural design variables of each functional zone, the multimodal input data includes the multimodal image data and the structured design variables corresponding to the parameter samples, and the multimodal image data includes the geometric schematic diagram of the inner and outer plates of the cover, the head impact point profile, and the structural parameter field distribution heat map. This approach breaks through the limitations of traditional surrogate models that only use structured parameters as input. It constructs a dual-branch input system of structured parameters and multimodal image data. Through optimal Latin hypercube experimental design, it generates parameter samples covering the entire design space, simultaneously generating corresponding geometric image data. Core technologies include experimental design for multi-objective optimization, simultaneous generation and alignment of multimodal heterogeneous data, and image data construction that integrates geometric features and parameter fields. This addresses the pain point of traditional surrogate models in failing to characterize unstructured features such as hard point gap distribution, local stiffness differences, and geometric contours. The multimodal image data fully covers the core geometric factors affecting pedestrian protection performance, significantly improving the prediction accuracy of the surrogate model. In other words, it covers the entire design space with the fewest possible samples, greatly reducing the sample size and time cost of simulation.

[0044] Step 2.2: Using the parameter samples as input, perform finite element simulation according to the pedestrian protection working conditions, and output simulation indicators to characterize the pedestrian protection performance and structural quality; This step provides ground truth labels for training the surrogate model. Simulation conditions are strictly set according to national standards for pedestrian protection, and full-condition finite element simulations are performed on each set of parameter samples, outputting core performance and quality indicators. Strict adherence to the required simulation conditions ensures the accuracy and authority of the simulation ground truth, providing a ground truth guarantee for the surrogate model's prediction accuracy. The output simulation indicators directly correspond to the optimization objectives, achieving complete alignment between the simulation ground truth model training and optimization objectives, avoiding the problem of a disconnect between training and optimization goals.

[0045] Step 2.3: Using multimodal input data as features and simulation indicators as labels, train a multimodal deep learning prediction model to obtain an optimized surrogate model with the required accuracy.

[0046] Here, multimodal input data is used as features, and the simulation ground truth is used as labels to train a multimodal deep learning prediction model. This ultimately yields an optimized surrogate model with satisfactory accuracy, replacing traditional finite element simulation for rapid performance prediction. Utilizing multimodal data-driven deep learning model training provides stronger nonlinear fitting and high-dimensional data processing capabilities. It overcomes the limitations of traditional response surface methodology and Kriging surrogate models in fitting high-dimensional, strongly nonlinear problems, accurately fitting the strong nonlinear coupling relationship between structural parameters, geometric features, and pedestrian protection performance, significantly improving prediction accuracy. The trained surrogate model can achieve millisecond-level performance prediction, improving computational efficiency by more than five orders of magnitude.

[0047] In practical applications, the steps for determining the trained optimized surrogate model are as follows: First, experimental design and sample generation are performed. The optimal Latin hypercube experimental design method is adopted to generate several, such as 300, uniformly distributed parameter samples within the engineering feasible range of the design variables in each partition. For each sample, corresponding multimodal image data is generated simultaneously, including schematic diagrams of the geometric contours of the inner and outer panels of the engine hood, cross-sectional diagrams of the head impact point, and heat maps of the structural parameter field distribution. At the same time, the structured design variable vectors corresponding to each sample are output, thus completing the construction of the multimodal dataset.

[0048] Then, finite element simulation and truth acquisition were performed. In LSDYNA finite element software, a refined simulation model of the engine hood composite sandwich structure was constructed. The honeycomb element and negative Poisson's ratio element were modeled using shell elements. According to GB245502009, the head impactor (pedestrian head model) was set as a hemispherical rigid body with a weight of 4.5 kg, and the impact velocity was set as 11.1 m / s. The impact point covered all zones. Simulation was performed on each set of parameter samples, and several core simulation indicators were extracted and output: head injury value HIC, residual gap after impact, and total mass of the engine hood sandwich structure.

[0049] In addition, the dataset is divided into 300 sample datasets, which can be divided into training set, validation set and reserved test set in a ratio of 8:1:1. The training set is used for model weight iteration, the validation set is used for overfitting monitoring, and the reserved test set is used for final accuracy verification.

[0050] Finally, model training and accuracy verification are performed to construct a multimodal deep learning prediction model. The training set data is used as input and the simulation index is used as label. The AdamW optimizer is used to iteratively train the model. After training, the model accuracy is verified using a reserved test set. The preset accuracy threshold is the coefficient of determination R² ≥ 0.95 between the prediction result and the simulation true value. The model that meets the accuracy threshold is used as a surrogate model for subsequent optimization.

[0051] For example, step 2.3 can also be refined through the following steps, including: Step 3.1: The multimodal deep learning prediction model adopts a cascaded architecture consisting of a dual-stream encoder, a cross-modal feature fusion module, and a multi-task decoder; The dual-stream encoder includes an image encoder and a parametric encoder; Here, the surrogate model employs a cascaded architecture. For heterogeneous inputs of multimodal image data and structured design variables, a dual-stream encoder with a dual-branch processing architecture is designed. This architecture adapts feature extraction from image data and structured parameters separately, then achieves feature alignment through a cross-modal fusion module, and finally enables simultaneous prediction of multiple indicators through a multi-task decoder. This dual-stream branch architecture design, utilizing heterogeneous data, overcomes the limitation of traditional single-branch models that cannot simultaneously process image and structured data.

[0052] This dual-stream encoder features a dual-branch architecture with customized feature extraction networks for the characteristics of two types of data, achieving optimal feature extraction for heterogeneous data and solving the pain point of insufficient heterogeneous data processing capability of traditional single-branch models. The cascaded architecture has a clear logic and well-defined functional division of each module, possessing strong scalability and maintainability.

[0053] Step 3.2: Input the multimodal input data into the multimodal deep learning prediction model. The image encoder uses a CNN combined with a feature pyramid network and introduces channel attention and spatial attention mechanisms to extract multi-scale features from the multimodal image data to determine the image features used to characterize the local stiffness differences and gap distribution of the cover. At the same time, the parametric encoder uses an embedding layer combined with a Transformer structure to encode the structured design variables and capture the parametric features used to characterize the nonlinear coupling and long-range dependence between parameters. Here, the specific network structures and functions of the two branches of the dual-stream encoder are clearly defined: the image encoder branch uses CNN, feature pyramid network, and attention mechanism to achieve multi-scale image feature extraction and accurately capture core geometric features such as hard point gap distribution and local stiffness differences; the parametric encoder branch uses embedding layers combined with a Transformer structure to achieve feature encoding of structured parameters and accurately capture the nonlinear coupling and long-range dependencies between parameters in each partition. Dedicated feature extraction networks are customized for the characteristics of the two types of data; the multi-scale feature extraction of the image encoder combined with the attention mechanism can accurately focus on the features of key areas such as hard point gaps and impact points, significantly improving the representation ability of image features; the Transformer structure of the parametric encoder can accurately capture the long-range dependencies between parameters in each partition, solving the pain point of insufficient fitting ability of traditional fully connected networks for high-dimensional parameter coupling relationships.

[0054] Step 3.3: Input the image features and parametric features into the cross-modal feature fusion module. After the geometric and parametric dimensions are aligned by the cross-attention mechanism and feature modulation, the feature fusion module is passed to the multi-task decoder. The module outputs the predicted results of the engine cover under the impact of pedestrian protection conditions in parallel. At the same time, the uncertainty estimate of the predicted results is output through evidence regression or deep integration. The prediction results include head injury value, impact residual gap, and total mass of interlayer. The uncertainty estimate is used to quantify the confidence level of the prediction results. Here, the core function of cross-modal feature fusion and multi-task decoder is to achieve cross-modal alignment and fusion of image features and parametric features through a cross-attention mechanism, solving the problem of dimensionality mismatch of heterogeneous features. The multi-task decoder simultaneously outputs two core contents: first, the performance prediction results of head injury value, residual gap, and total mass; and second, the uncertainty estimate of the prediction results based on evidence regression, quantifying the model's confidence in the current prediction results. The core innovation lies in the simultaneous output of prediction results and uncertainty estimates, breaking through the limitation of traditional deterministic surrogate models that cannot assess prediction reliability.

[0055] This invention achieves deep fusion of image geometric features and parametric features through a cross-modal feature fusion module, accurately capturing the coupling relationship between the impact performance of geometric gap structural parameters and significantly improving the prediction accuracy of the model. For the first time, uncertainty quantification is introduced into the surrogate model for engine hood optimization, providing a core basis for subsequent active learning loop closure, optimized sample selection, and simulation verification.

[0056] Step 3.4: Based on the prediction results, the total loss function, composed of the uncertainty estimate and the simulation index used as labels, iterates the weights of the multimodal deep learning prediction model until the total loss function converges, thus obtaining the trained optimized surrogate model.

[0057] The total loss function integrates the fitting error of the prediction results, the uncertainty calibration error, and the physical prior regularization term, achieving simultaneous optimization of prediction accuracy, uncertainty assessment reliability, and physical consistency. The total loss function, which incorporates multi-task loss and physical prior regularization, is constructed as follows: Definition of HIC loss:

[0058] in, It is the predicted HIC value of the i-th sample. This is the actual HIC value, and N is the number of samples.

[0059] Definition of remaining gap δ:

[0060] in, It is the predicted δ value of the i-th sample. This is the actual δ value, and N is the number of samples.

[0061] The final total loss function is:

[0062] The regularization terms include multi-task physical consistency and manufacturing prior regularization, including HIC and acceleration consistency constraints, local monotonic constraints of thickness-plateau stress, gradual smoothness constraints of the parametric field, and minimum feature / formula sparsity penalty.

[0063] The multi-task total loss function simultaneously optimizes prediction accuracy and uncertainty calibration, solving the problem that traditional models predict accurately but fail to assess confidence, and ensuring the reliability of uncertainty estimates. The introduction of physical prior regularization terms makes the model learning results conform to physical laws and manufacturing constraints, greatly improving the model's generalization ability and engineering applicability.

[0064] As a detailed embodiment of a specific implementation, such as Figure 7 As shown, the input layer (multimodal) is the core data input terminal of the model. It first receives two types of heterogeneous data: one is the image encoding stream, which includes multimodal image data such as outer plate image, inner plate image, head impact profile, and parametric field heat map; the other is the parameter encoding stream, which covers structured design variables such as cell thickness, cell edge distance, reference line spacing, waist clearance, reentry triangular angle, and arc radius, providing a unified data foundation for feature extraction of the dual-stream encoder.

[0065] Building upon this foundation, a dual-stream encoder network was first constructed, splitting into two main branches: an image encoder and a parametric encoder. The image encoder branch uses ResNet18 as its backbone CNN network, combined with a feature pyramid network to achieve multi-scale feature extraction. Furthermore, channel attention and spatial attention modules are introduced in the last two layers of the network to focus on key region features such as hard point gaps and local stiffness, ultimately outputting a 256-dimensional image feature vector. The parametric encoder branch first maps structured design variables into high-dimensional feature vectors through an embedding layer, then captures the long-range dependencies between parameters in each partition through two Transformer encoder layers, finally outputting a 256-dimensional parametric feature vector, thus completing independent feature extraction from heterogeneous data.

[0066] Then, a cross-modal feature fusion module is built, with cross-attention mechanism as the core of fusion. Image features and parametric features are used as Query, Key and Value respectively to realize cross-modal interaction and dimensional alignment of image geometric features and structured parametric features. Then, the two features are fused into a 512-dimensional unified fusion feature vector through the feature modulation layer to complete the deep fusion of multimodal features.

[0067] In addition, a multi-task decoder is built, using three parallel fully connected layers as independent prediction heads to output the prediction results of head injury value HIC, impact residual gap δ, and total mass m of the cell interlayer, respectively. At the same time, an evidence regression branch is added to the end of the decoder, and the evidence distribution of each prediction result is output through the evidence deep neural network to calculate the corresponding uncertainty estimate, that is, the cognitive uncertainty of the prediction result, so as to realize the simultaneous prediction and reliability assessment of multiple indicators.

[0068] Next, the total loss function is constructed, which consists of three core parts: first, the multi-task prediction loss, which is obtained by weighted summation of HIC prediction loss, residual gap prediction loss, and quality prediction loss; second, the uncertainty calibration loss, which adopts negative log-likelihood loss to ensure the matching degree between the uncertainty estimate and the actual prediction deviation; and third, the physical prior regularization term, which includes core rules such as HIC and acceleration consistency constraints and parameter field gradual smoothness constraints. At the same time, an L2 regularization term is added to suppress model overfitting and ensure the stability and generalization of training.

[0069] Finally, model training was performed, using the AdamW optimizer as the training core. The initial learning rate was set to a preset parameter, the batch size to 16, and the training epochs to 200. An early stopping mechanism and a learning rate scheduling strategy were configured to ensure stable convergence. During training, enhancement strategies such as brightness / contrast jittering and noise / edge enhancement were applied to the image data, and enhancement strategies such as affine perturbation were applied to the parameter data to further improve the model's generalization ability. After training, the model weights were saved as optimization surrogate models for subsequent multi-objective optimization.

[0070] After determining the trained optimized agent model, before implementing step S106, the optimized agent model is also validated: Step 4.1: After the optimized proxy model training is completed, use the reserved test set samples outside the training set to perform prediction accuracy verification; This step provides a benchmark for accuracy verification of the active learning loop. Using a reserved test set independent of the training set, the generalization accuracy of the initially trained surrogate model is verified, avoiding inflated accuracy caused by overfitting within the training set. Utilizing model generalization capability verification based on an independent test set provides a clear criterion for initiating the active learning loop. Verification using an independent reserved test set accurately reflects the model's generalization ability, avoiding inflated accuracy caused by verification within the training set, and ensuring the authenticity of the accuracy verification results. It also provides a clear initiation condition for the active learning loop; the active learning process is only initiated when the model accuracy does not reach a threshold, avoiding unnecessary simulation cost waste.

[0071] Step 4.2: If the accuracy does not reach the preset accuracy threshold, based on the uncertainty estimate of the model output, prioritize the selection of high-value samples whose prediction uncertainty is higher than the preset uncertainty threshold and are located in the design space boundary or low gap area, re-execute the finite element simulation to obtain the corresponding label data, and then supplement the high-value samples carrying label data into the training set to perform incremental training and accuracy re-verification of the optimized surrogate model until the accuracy of the optimized surrogate model reaches the preset accuracy threshold. The preset accuracy threshold is determined based on the prediction results output by the optimized surrogate model and the determination coefficient of the simulation index.

[0072] Here, based on uncertainty estimates, high-value samples with the least certainty in the model are selectively chosen, rather than through traditional random sampling. The core logic is to supplement only the samples that provide the greatest improvement to the model's accuracy. This involves a high-value sample selection rule based on uncertainty, a sample priority design for hard-point, low-gap regions, and a closed-loop process of incremental training and accuracy re-verification. This approach addresses the pain points of low efficiency and poor accuracy improvement in core sensitive areas caused by traditional active learning random sampling. Targeted selection based on uncertainty achieves the maximum improvement in model accuracy with the fewest simulation samples, reducing simulation costs by over 70%. Prioritizing high-uncertainty samples in low-gap regions precisely improves prediction accuracy in high-risk areas, solving the industry pain points of inaccurate predictions and high defined risks in low-gap regions by traditional surrogate models.

[0073] In practical applications, this verification process includes: First, set the accuracy threshold. The preset accuracy threshold is set to the coefficient of determination R² between the predicted result and the simulated true value ≥ 0.95. The preset uncertainty threshold is set to the 75th quantile of the uncertainty estimate, that is, the samples with the highest uncertainty are selected as high-value samples.

[0074] Then, an initial accuracy check is performed. After the initial model training is completed, the accuracy check is performed using 30 reserved test set samples. The multimodal data of the test set is input, the model outputs the prediction results, and the R² value between the prediction results and the simulation true value is calculated. If R² ≥ 0.95, the model accuracy meets the standard and there is no need to start active learning; if R² < 0.95, the active learning closed loop is started.

[0075] Next, high-value samples are screened. Based on the uncertainty estimate output by the model, high-value samples that meet the following conditions are selected from the candidate samples in the design space: first, the prediction uncertainty is higher than the preset uncertainty threshold; second, it is located in the low gap area at the boundary of the design space or directly above the hard point. The number of samples screened in one session is controlled to be 20 to 30 groups to avoid excessive simulation costs.

[0076] Then, simulation labeling and incremental training are performed. For the selected high-value samples, finite element simulation is performed according to the standardized process to obtain the corresponding simulation indicators as true value labels. The labeled high-value samples are added to the original training set, and incremental training is performed on the model. The initial learning rate of incremental training is set to a preset parameter, and the training rounds are 50 to avoid destroying the features already learned by the original model.

[0077] Finally, a closed-loop iteration is performed. After the incremental training is completed, the accuracy is re-verified using the reserved test set. If the accuracy is still not up to standard, the above process of screening and simulation incremental training is repeated until the model accuracy reaches the preset threshold of R²≥0.95, thus completing the active learning closed loop.

[0078] Step S106: Based on the optimized surrogate model, complete multi-objective optimization and verification according to the structural design variables to determine the optimal structural scheme of the engine cover.

[0079] Here, an optimized surrogate model with high accuracy is used as the core prediction tool. The range of structural design variables output from the aforementioned embodiments is used as the optimization boundary. A multi-objective optimization problem that balances pedestrian protection performance and lightweight design is constructed. Through iterative optimization, constraint verification, and simulation verification, the optimal structural solution that is feasible for engineering is finally output. Among them, the efficient multi-objective optimization based on the surrogate model, the compliance verification of regulations and manufacturing constraints, and the closed-loop optimization of simulation verification and model iteration ensure the engineering feasibility of the optimization results. It also solves the core pain points of traditional open-loop optimization, such as the large number of false optimal solutions and the inability to implement optimization results in mass production. The simulation verification and iterative closed-loop ensure the engineering feasibility of the optimal solution. It realizes the dual-objective collaborative optimization of maximizing pedestrian protection performance and minimizing structural mass, breaking through the technical bottleneck of the incompatibility between pedestrian protection and lightweight design in traditional design.

[0080] In some embodiments, step S106 may be implemented by the following steps: Step 5.1: Using the optimized surrogate model as the prediction tool and the upper and lower limits of the structural design variables as the optimization boundary, construct a multi-objective optimization problem that balances pedestrian protection performance and lightweight design, and obtain the Pareto optimal solution set; This approach utilizes a surrogate model with verified accuracy as the core prediction tool, and uses the upper and lower limits of structural design variables as the optimization boundaries. It constructs a dual-objective multi-objective optimization problem that minimizes the maximum HIC value under each working condition and minimizes the total structural mass. The NSGAII multi-objective optimization algorithm iteratively optimizes the solution, eliminating highly unreliable samples based on uncertainty estimates during the iteration process. Finally, it outputs a Pareto optimal solution set. Replacing traditional finite element simulation with a surrogate model reduces the performance evaluation time for a single set of samples from several hours to milliseconds, improving the computational efficiency of multi-objective optimization by more than five orders of magnitude. This addresses the industry pain points of extremely high computational costs and long cycles in traditional multi-objective optimization. Furthermore, the elimination of highly unreliable samples based on uncertainty during the iteration process fundamentally avoids the spurious optimal solution problem commonly found in traditional optimization, ensuring the reliability of the Pareto optimal solution set.

[0081] Step 5.2: Perform constraint compliance verification on the Pareto optimal solution set, remove invalid solutions to obtain a compliant and valid solution set; prioritize extracting the frontier representative solutions with high prediction uncertainty from the compliant and valid solution set, and then perform full-condition finite element simulation verification. If the verification result deviates from the prediction result of the optimized surrogate model by more than the preset accuracy threshold, then feed the parameter samples with the deviation exceeding the threshold and the corresponding simulation index true values ​​back into the training set, incrementally train the optimized surrogate model, and re-execute multi-objective optimization. This process consists of two core components: first, constraint compliance verification, which eliminates invalid solutions that do not meet regulations and mass production requirements through dual verification of safety and manufacturing constraints; second, targeted simulation verification and iterative closed-loop, which prioritizes high-risk samples for simulation verification based on uncertainty estimates. If the deviation exceeds the limit, the samples with excessive deviation are fed back into the training set for incremental model training and re-optimization, forming a complete closed loop of optimization verification and model iteration. Dual constraint compliance verification eliminates invalid solutions that do not meet regulations and cannot be mass-produced from the design source, ensuring the engineering feasibility of subsequent optimization results; targeted simulation verification based on uncertainty only performs simulation verification on samples with low model prediction confidence, completing the reliability verification of optimization results with minimal simulation cost and avoiding the high cost of full-scale simulation; the iterative closed-loop of deviation feedback enables dynamic improvement of model accuracy during the optimization process, completely solving the core pain point of traditional open-loop optimization where the optimization results deviate greatly from the simulation results and cannot be implemented.

[0082] Step 5.3: After the Pareto front converges iteratively, select an engineering compromise solution from the set of compliant and valid solutions that takes into account the head protection performance, the level of lightweighting and the prediction confidence, and output the optimal structural scheme of the engine cover.

[0083] For example, a stability convergence criterion can be set by using a hypervolume index change rate ≤1% after three consecutive iterations, ensuring the stability of the optimization results. Finally, the TOPSIS multi-attribute decision method is employed to select an engineering compromise solution that balances head protection performance, lightweight level, and prediction confidence. This utilizes convergence criteria based on the hypervolume index and multi-dimensional optimal solution selection rules.

[0084] Here, the convergence criterion is quantified by using the rate of change of the Pareto front hypervolume index, thus avoiding infinite loops in optimization iteration and ensuring the stability of the optimization results. The selection of the optimal solution also takes into account the prediction confidence, ensuring that the final output solution not only has the best performance and lightweight design, but also the highest reliability of model prediction, which greatly reduces the regulatory risks of engineering implementation.

[0085] In practical applications, a multi-objective optimization problem can be constructed first. An optimization surrogate model with satisfactory accuracy can be used as a prediction tool, and the upper and lower limits of the structural design variables of each partition can be used as the optimization boundary to construct a multi-objective optimization mathematical model. The optimization objectives are to minimize the maximum HIC value under each head impact condition and to minimize the total mass of the engine cover interlayer. The constraints include safety regulations such as remaining gap ≥ 10 mm and HIC15 ≤ 1000, as well as manufacturing constraints such as minimum wall thickness ≥ 0.18 mm, maximum cell-to-edge distance ≤ 20 mm, and parameter gradient smoothness meeting the requirements.

[0086] Then, iterative optimization and Pareto solution set output are performed. The NSGAII multi-objective optimization algorithm is used, with the population size set to 100, the number of iterations set to 200 generations, the crossover probability to 0.9, and the mutation probability to 0.1. During the iteration, the performance indicators of each generation of the population are quickly predicted by optimizing the surrogate model, while high unreliable samples with uncertainty higher than the preset threshold are removed. After the iteration is completed, the Pareto optimal solution set is output.

[0087] Next, constraint compliance checks are performed. For each set of solutions in the Pareto optimal solution set, compliance checks are performed on safety constraints, regulatory constraints, and manufacturing constraints in sequence. Invalid solutions that do not meet any of the constraints are eliminated, and the remaining solutions form a set of compliant and valid solutions.

[0088] In addition, targeted simulation verification and iterative closed-loop are performed. From the set of compliant and valid solutions, 5 to 10 representative solutions with the highest prediction uncertainty on the Pareto front are extracted and full-condition finite element simulation verification is performed according to the standardized process. If the relative deviation between the verification result of any set of solutions and the model prediction result exceeds the preset threshold of 10%, the parameter samples with the deviation exceeding the standard and the simulation true values ​​are fed back into the training set. After incremental training is performed on the surrogate model, the multi-objective optimization process is restarted.

[0089] Finally, convergence judgment and optimal solution output are performed. The stability convergence criterion is that the rate of change of the hypervolume index of the Pareto front is ≤1% after three consecutive iterations. After the front stabilizes and converges, the TOPSIS multi-attribute decision method is used to select the optimal engineering compromise solution that takes into account the head injury value, structural quality and prediction confidence from the set of compliant and valid solutions. Finally, the corresponding three-dimensional model of the engine hood, performance simulation report, manufacturability analysis report and mass production process documents are output to complete the entire optimization process.

[0090] In a preferred embodiment of practical application, such as Figure 6 As shown, the engine hood structure optimization method of this application can also be achieved through the following steps: First, complete the preliminary preparations. This method is applicable to the structural optimization of steel / aluminum alloy / composite material engine hoods for fuel vehicles and pure electric vehicles. The preliminary preparations require the completion of the overall layout design of the vehicle's engine compartment, obtaining the three-dimensional coordinates of all hard points in the engine compartment, the vertical clearance data between the hard points and the inner panel of the engine hood, as well as the geometric models of the outer and inner panels of the engine hood, laying the foundation for subsequent partition design and simulation modeling.

[0091] Then the partition structure design is executed, corresponding to... Figure 6 The initial step involves dividing the engine hood plane into three functional zones—a low-gap zone, a transition zone, and a high-gap zone—based on the distribution of vertical gaps at hard points. A customized composite sandwich structure is then developed for each zone, and the core design variables and engineering feasibility range for each zone are output to achieve precise matching between the structure and the gaps at hard points.

[0092] Next, sample generation and model training were performed. The optimal Latin hypercube experimental design method was adopted to generate parameter samples covering the entire design space within the feasible range of design variables, and corresponding multimodal input data were generated simultaneously. In accordance with pedestrian protection regulations, full-condition finite element simulation was performed on each group of samples, and core performance and quality indicators such as head injury value HIC, impact residual gap, and total mass of interlayer were output as simulation true values. Using multimodal data as features and simulation indicators as labels, a multimodal deep learning surrogate model was trained to build an efficient performance prediction tool.

[0093] Subsequently, model accuracy verification and active learning iteration are carried out. In the accuracy judgment branch corresponding to the attached figure, if the model prediction accuracy does not meet the standard, high-value samples are selected based on the uncertainty estimate, the sample is expanded, and the experimental design and sample generation stage is returned to, and finite element simulation and model training are re-executed until the accuracy meets the requirements; if the accuracy meets the standard, the multi-objective optimization stage is entered.

[0094] Finally, multi-objective optimization and solution output are performed. Using the trained surrogate model as a prediction tool, a bi-objective optimization problem is constructed to minimize the HIC value and the total structural mass. With regulatory safety requirements and mass production requirements as constraints, the NSGA-II algorithm is used to perform iterative optimization. The output Pareto optimal solution set is sequentially checked for constraint compliance and verified by finite element simulation. If the verification result deviates from the standard, active learning backfeeding is initiated to iteratively optimize the process. After the Pareto front stabilizes, the optimal solution that balances performance, lightweight and manufacturability is selected, and the corresponding geometric model, performance report and mass production process documents are output.

[0095] In some embodiments, such as Figure 8 As shown, an embodiment of the present invention provides an engine hood structure optimization device, comprising: The design module determines the composite sandwich structure of the energy-absorbing core layer in the engine cover based on the geometric relationship between the hard point gap in the engine compartment and the engine cover, and performs partitioning processing on the energy-absorbing core layer, outputting the structural design variables corresponding to the composite sandwich structure of each functional partition. The module determines the structural design variables, conducts experimental design and performs finite element simulation according to pedestrian protection conditions, obtains multimodal input data and simulation indicators for model training, and determines the trained optimized surrogate model. The optimization module, based on the optimization proxy model, performs multi-objective optimization and verification according to the structural design variables to determine the optimal structural scheme of the engine cover.

[0096] The present invention provides an embodiment for implementing an electronic device. In this embodiment, the electronic device may be, but is not limited to, a personal computer (PC), a laptop computer, a monitoring device, a server, or other computer device with analysis and processing capabilities.

[0097] As an exemplary embodiment, see [reference]. Figure 9 The electronic device 110 includes a communication interface 111, a processor 112, a memory 113, and a bus 114. The processor 112, the communication interface 111, and the memory 113 are connected via the bus 114. The memory 113 is used to store a computer program that supports the processor 112 in executing the above-described method. The processor 112 is configured to execute the program stored in the memory 113.

[0098] The machine-readable storage medium mentioned in this article can be any electronic, magnetic, optical, or other physical storage device that can contain or store information such as executable instructions, data, etc. For example, machine-readable storage media can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0099] Non-volatile media can be non-volatile memory, flash memory, storage drives (such as hard disk drives), any type of storage disk (such as optical discs, DVDs, etc.), or similar non-volatile storage media, or combinations thereof.

[0100] It is understood that the specific operation methods of each functional module in this embodiment can be referred to the detailed description of the corresponding steps in the above method embodiment, and will not be repeated here.

[0101] The computer-readable storage medium provided in the embodiments of the present invention stores a computer program. When the computer program code is executed, it can implement the method described in any of the above embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0102] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0103] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0104] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0105] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing the structure of an engine hood, characterized in that, include: Based on the geometric relationship between the hard point gap in the engine compartment and the engine cover, the composite sandwich structure of the energy-absorbing core layer in the cover is determined, and the energy-absorbing core layer is partitioned to output the structural design variables corresponding to the composite sandwich structure of each functional partition. The structural design variables were experimentally designed and finite element simulations were performed according to pedestrian protection conditions to obtain multimodal input data and simulation indicators for model training, and the trained optimized surrogate model was determined. Based on the aforementioned optimized proxy model, multi-objective optimization and verification are completed according to the aforementioned structural design variables to determine the optimal structural scheme of the engine cover.

2. The method according to claim 1, characterized in that, The composite sandwich structure consists of an upper honeycomb unit layer, a middle negative Poisson's ratio unit layer, and a lower honeycomb unit layer stacked sequentially along the length direction. The upper and lower honeycomb unit layers each use multiple interconnected hexagonal honeycomb units to form a load-bearing and energy-absorbing frame. The middle negative Poisson's ratio unit layer is composed of multiple axisymmetric concave arc-shaped re-entry negative Poisson's ratio units connected to each other. The negative Poisson's ratio units undergo lateral outward expansion deformation when subjected to vertical compression.

3. The method according to claim 1, characterized in that, The steps include determining the composite sandwich structure of the energy-absorbing core layer in the engine cover based on the geometric relationship between the hard point gap in the engine compartment and the engine cover, partitioning the energy-absorbing core layer, and outputting the structural design variables corresponding to the composite sandwich structure of each functional partition, including: Based on the vertical gap distribution between the hard points in the engine compartment and the engine cover, and the geometric contours of the inner and outer plates in the cover, the composite sandwich structure of the energy-absorbing core layer in the cover is determined; wherein, the cover includes an outer plate, the composite sandwich structure and the inner plate connected sequentially along the thickness direction; The energy-absorbing core layer is divided in the plane into a low-gap region directly above the hard point, a transition region connecting the low-gap region and the high-gap region, and a high-gap region away from the hard point. Output the unit geometric parameters corresponding to each functional area and the engineering feasible range of the unit geometric parameters as structural design variables.

4. The method according to claim 1, characterized in that, The steps for designing the structural design variables experimentally and performing finite element simulations under pedestrian protection conditions to obtain multimodal input data and simulation indices for model training, and determining the trained optimized surrogate model, include: Using the structural design variables as input, parameter samples and corresponding multimodal input data are generated through experimental design; wherein, the parameter samples are the discretized values ​​of the structural design variables of each functional area, the multimodal input data includes multimodal image data and the structured design variables corresponding to the parameter samples, and the multimodal image data includes geometric schematic diagrams of the inner and outer plates of the cover, a cross-sectional view of the head impact point, and a heat map of the structural parameter field distribution; Using the parameter samples as input, finite element simulation is performed according to pedestrian protection conditions, and simulation indicators are output to characterize pedestrian protection performance and structural quality. Using the multimodal input data as features and the simulation indicators as labels, a multimodal deep learning prediction model is trained to obtain an optimized surrogate model with the required accuracy.

5. The method according to claim 4, characterized in that, The steps for training a multimodal deep learning prediction model using the multimodal input data as features and the simulation metrics as labels, to obtain an optimized surrogate model with achieved accuracy, include: The multimodal deep learning prediction model adopts a cascaded architecture consisting of a dual-stream encoder, a cross-modal feature fusion module, and a multi-task decoder. The multimodal input data is input into the multimodal deep learning prediction model and processed by the dual-stream encoder to determine the image features characterizing the local stiffness differences and gap distribution of the cover body and the parameter features characterizing the nonlinear coupling and long-range dependence between the parameters. After the image features and parameter features are fused by the cross-modal feature fusion module, the multi-task decoder outputs the prediction results and corresponding uncertainty estimates of the engine cover performance indicators under pedestrian protection impact conditions. Based on the prediction results, the total loss function, composed of the uncertainty estimate and the simulation index, iterates the weights of the multimodal deep learning prediction model until the total loss function converges, thus obtaining the trained optimized surrogate model.

6. The method according to claim 5, characterized in that, Before the step of determining the optimal structural scheme of the engine cover by performing multi-objective optimization and verification based on the optimized surrogate model and the structural design variables, the method further includes: The prediction accuracy of the optimized surrogate model after training is verified using reserved test set samples outside the training set. If the accuracy of the optimized proxy model does not reach the preset accuracy threshold, then based on the uncertainty estimate, high-value samples with prediction uncertainty higher than the preset uncertainty threshold and located in the preset key area and their corresponding label data are added to the training set, and the optimized proxy model is incrementally trained and its accuracy is re-verified until the accuracy of the optimized proxy model reaches the preset accuracy threshold.

7. The method according to claim 1, characterized in that, Based on the aforementioned optimized surrogate model, and according to the aforementioned structural design variables, the steps for completing multi-objective optimization and verification to determine the optimal structural scheme of the engine cover include: Using the optimized proxy model as a prediction tool and the upper and lower limits of the structural design variables as the optimization boundaries, a multi-objective optimization problem that balances pedestrian protection performance and lightweight design is constructed to obtain the Pareto optimal solution set. The frontier representative solution is extracted from the compliant and valid solution set corresponding to the Pareto optimal solution set, and then the full-condition finite element simulation is performed for verification. If the verification result deviates from the prediction result of the optimized surrogate model by more than the preset accuracy threshold, the parameter samples and simulation indicators corresponding to the current Pareto optimal solution set are fed back into the training set, and the optimized surrogate model is incrementally trained and then multi-objective optimization is re-executed. After the Pareto front converges iteratively, the engineering compromise solution is selected from the set of compliant and valid solutions, and the optimal structural scheme of the engine cover is output.

8. An engine hood structure optimization device, characterized in that, include: The design module determines the composite sandwich structure of the energy-absorbing core layer in the engine cover based on the geometric relationship between the hard point gap in the engine compartment and the engine cover, and performs partitioning processing on the energy-absorbing core layer, outputting the structural design variables corresponding to the composite sandwich structure of each functional partition. The module determines the structural design variables, conducts experimental design and performs finite element simulation according to pedestrian protection conditions, obtains multimodal input data and simulation indicators for model training, and determines the trained optimized surrogate model. The optimization module, based on the optimization proxy model, performs multi-objective optimization and verification according to the structural design variables to determine the optimal structural scheme of the engine cover.

9. An electronic device, characterized in that, It includes a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed, implements the method described in any one of claims 1-7.

Images