Protection film packaging design optimization method and system

Abstract

The invention discloses a protective film packaging design optimization method and system, and relates to the technical field of computer aided design. The protection film packaging design optimization method comprises the following steps: receiving a natural language design demand and a basic constraint condition input by a user, and converting the natural language design demand and the basic constraint condition into an initial design parameter set through a pre-trained demand analysis model; based on the initial design parameter set, calling a preset packaging component library and a topology generation algorithm, automatically generating a plurality of initial packaging structure design schemes, and performing evaluation screening through a pre-trained performance evaluation model to obtain a plurality of candidate schemes; according to the method, a multi-objective optimization function is established, a multi-objective optimization algorithm is adopted to carry out iterative optimization, a group of Pareto optimal solution schemes are obtained, and visual output is carried out, and through integration of natural language understanding, automatic scheme generation, intelligent performance evaluation and multi-objective optimization, a whole-process intelligent design closed loop from requirements to the optimal scheme is realized, and the method has the advantages of being high in practicability and easy to popularize. And the packaging design efficiency and quality are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of computer-aided design technology, specifically to a method and system for optimizing the design of protective film packaging. Background Technology

[0002] In today's commodity manufacturing and logistics sectors, protective film packaging plays a crucial role. It not only needs to effectively cushion impacts and prevent product scratches and damage during transportation and warehousing, but also must consider material costs, warehouse space utilization, and environmental requirements to achieve a comprehensive balance between protection, economy, and space efficiency. As market demands for product personalization, efficient logistics, and refined cost management continue to rise, the complexity and importance of packaging design are becoming increasingly prominent.

[0003] The limitations of existing technologies include at least the following problems: Current protective film packaging design relies heavily on the personal experience of designers and repeated trial and error, resulting in a lengthy design cycle and difficulty in achieving the best balance. For example, user needs, such as "drop protection and low cost," can often only be translated through simple form filling. Designers then manually model and simulate based on their experience. This process is not only inefficient, but also difficult to balance multiple conflicting goals such as packaging volume, cost, and cushioning performance simultaneously. The final solution is often a compromise that sacrifices one aspect for another, making it difficult to quickly provide a series of Pareto solutions that achieve optimal balance in multiple dimensions for decision-making reference. This restricts the overall optimization efficiency of packaging design and the comprehensive competitiveness of the product. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for optimizing the design of protective film packaging, which solves the problems of low design efficiency and difficulty in systematically balancing multiple conflicting objectives to obtain the globally optimal solution caused by fragmented processes and reliance on manual experience in existing protective film packaging design technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing the design of protective film packaging, comprising the following steps: receiving natural language design requirements and basic constraints input by a user, and converting the natural language requirements into a set of initial design parameters that meet the basic constraints using a pre-trained requirement parsing model, wherein the initial design parameter set includes size parameters, material type parameters, and cushioning structure type parameters; based on the initial design parameter set, calling a pre-set packaging component library and topology generation algorithm to automatically generate multiple initial packaging structure design schemes, and evaluating and screening the multiple initial packaging structure design schemes using a pre-trained performance evaluation model to obtain multiple candidate schemes; establishing a multi-objective optimization function with packaging volume, material cost, and cushioning performance of key parts as optimization objectives, and using a multi-objective optimization algorithm to iteratively optimize the multiple candidate schemes to obtain a set of Pareto optimal solutions; and visually outputting the set of Pareto optimal solutions.

[0006] Furthermore, the pre-trained requirement parsing model is specifically a neural network model based on the Transformer architecture, which includes an input embedding layer, a multi-head self-attention encoding layer, a feedforward neural network layer, and an output layer.

[0007] Furthermore, the specific steps for transforming natural language requirements into a set of initial design parameters that meet the basic constraints are as follows: In the input embedding layer of the requirement parsing model, the natural language design requirements are segmented and vectorized to generate a sequence of word vectors; in the multi-head self-attention encoding layer of the requirement parsing model, parallel self-attention calculation is performed on the word vector sequence to capture the global dependencies between design elements and constraints in the natural language design requirements, and the context-enhanced feature vectors are output; in the feedforward neural network layer of the requirement parsing model, the context-enhanced feature vectors are subjected to nonlinear transformation and feature dimensionality reduction to extract high-dimensional abstract features; in the output layer of the requirement parsing model, the high-dimensional abstract features are processed by parameter vector mapping to form the initial design parameter set.

[0008] Furthermore, the specific steps for automatically generating multiple initial packaging structure design schemes are as follows: Based on the initial design parameter set, sampling is performed within the design space to generate multiple sets of basic size combinations; a pre-set packaging component library is called to instantiate parameterized components for each set of dimensions; a topology optimization algorithm is executed on the surfaces that need buffering to generate lightweight buffer structures; the buffer structures and basic components are automatically assembled and conflict detected to generate and output multiple initial packaging structure design schemes.

[0009] Furthermore, the pre-trained performance evaluation model is specifically a proxy model based on a graph attention network, which includes a graph data construction layer, a multi-level graph convolutional coding layer, and a multi-task prediction output layer.

[0010] Furthermore, the specific steps for obtaining multiple candidate solutions are as follows: In the graph data construction layer of the performance evaluation model, each initial packaging structure design scheme is transformed into graph structure data; in the multi-level graph convolutional coding layer of the performance evaluation model, the graph structure data is subjected to multi-level feature aggregation and encoding to extract graph feature vectors; in the multi-task prediction output layer of the performance evaluation model, multiple performance prediction values ​​of each initial scheme are predicted in parallel based on the graph feature vectors, and the initial schemes are filtered and sorted according to preset rules, and the top-ranked initial schemes are output as candidate schemes.

[0011] Furthermore, a multi-objective optimization function is established with packaging volume, material cost, and cushioning performance of key components as optimization objectives. Specifically, each candidate solution is encoded as an optimization variable; packaging volume calculation function, material cost calculation function, and cushioning performance evaluation function are constructed respectively; and combined to form the multi-objective vector function to be optimized.

[0012] Furthermore, the specific steps for iteratively optimizing multiple candidate schemes using a multi-objective optimization algorithm are as follows: initialize multiple candidate schemes to form a first-generation population; in each generation iteration, calculate the multi-objective optimization function value for each individual in the population; based on the multi-objective optimization function value, perform selection, crossover, and mutation operations to generate a offspring population; merge the parent and offspring populations, perform environmental selection based on Pareto dominance and distribution evaluation, and select a new generation population; repeat the iteration until the termination condition is met.

[0013] Furthermore, a set of Pareto optimal solutions is obtained, specifically: when the optimization iteration terminates, all non-dominated solutions are extracted from the final population; a set of solutions that achieve Pareto optimality in terms of packaging volume, material cost, and cushioning performance is formed; and the set of Pareto optimal solutions is output as the set of Pareto optimal solutions.

[0014] A protective film packaging design optimization system includes: a requirement parsing and parameterization unit, used to receive user-input natural language design requirements and basic constraints, and to transform the natural language requirements into an initial design parameter set through a pre-trained requirement parsing model, the initial design parameter set including size parameters, material type parameters, and cushioning structure type parameters; a scheme generation and evaluation screening unit, used to automatically generate multiple initial packaging structure design schemes based on the initial design parameter set, by calling a pre-set packaging component library and topology generation algorithm, and to evaluate and screen the multiple initial packaging structure design schemes through a pre-trained performance evaluation model to obtain multiple candidate schemes; a multi-objective optimization solution unit, used to establish a multi-objective optimization function with packaging volume, material cost, and cushioning performance of key parts as optimization objectives, and to iteratively optimize the multiple candidate schemes using a multi-objective optimization algorithm to obtain a set of Pareto optimal solutions; and a visualization output unit, used to visualize and output the set of Pareto optimal solutions.

[0015] The present invention has the following beneficial effects:

[0016] (1) The protective film packaging design optimization method adopts a neural network model based on the Transformer architecture as the core technology. This model can understand the complex needs described by users in natural language and automatically identify and associate key elements in the needs through its internal multi-head self-attention mechanism. Finally, the model outputs a structured set of initial design parameters, including size, material type and cushioning structure type. This automated requirement parsing method avoids the tedious manual parameter conversion process in manual design, greatly improving design efficiency and accuracy. Designers only need to provide simple requirement input to generate multiple initial packaging structure design schemes that meet the requirements, which greatly shortens the packaging design cycle and improves the market response speed of the product.

[0017] (2) The protective film packaging design optimization method applies a performance evaluation model based on graph attention network in multiple initial design schemes. It can comprehensively consider various performance indicators of packaging design, including cushioning structure performance, packaging volume and material cost. Through automatic evaluation and screening of multiple schemes, the system can select the design scheme with the best performance. This multi-dimensional evaluation method not only ensures the functionality of the packaging, but also considers cost-effectiveness, so that the final design scheme can minimize unnecessary resource waste and achieve the best design effect while ensuring quality.

[0018] (3) The protective film packaging design optimization method adopts a multi-objective optimization algorithm. By comprehensively optimizing the packaging volume, material cost and cushioning performance, it can achieve a good balance among multiple objectives. In traditional packaging design, it is often necessary to make trade-offs between different design objectives. However, this invention introduces Pareto optimal solution and provides a set of multiple simultaneously optimized solutions, which can meet the needs of different application scenarios. This multi-objective optimization not only improves the overall performance of packaging design, but also provides users with more choices, making the packaging solution more flexible.

[0019] (4) The protective film packaging design optimization system organically integrates demand analysis and parameterization, scheme generation and evaluation screening, multi-objective optimization solution and visualization output unit, realizing the intelligent design process. Users only need to input basic design requirements and constraints, and the system can automatically complete the generation, evaluation, optimization and output of packaging design. The systematic process not only reduces human intervention errors and deviations, but also makes the design process more standardized, easy to manage and control, and is particularly suitable for large-scale production and rapid response to market demands.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a protective film packaging design optimization method according to the present invention.

[0022] Figure 2 This is a flowchart illustrating the specific steps involved in automatically generating multiple initial packaging structure design schemes in a protective film packaging design optimization method according to the present invention.

[0023] Figure 3 This is a block diagram of a protective film packaging design optimization system according to the present invention. Detailed Implementation

[0024] Please see Figure 1, an embodiment of the present invention provides a technical solution: an optimization method for protective film packaging design, including the following steps: receiving the natural language design requirements and basic constraint conditions input by the user, and converting the natural language requirements into an initial design parameter set that meets the basic constraint conditions through a pre-trained requirement parsing model. The initial design parameter set includes size parameters, material type parameters, and buffer structure type parameters; based on the initial design parameter set, call the preset packaging component library and topology generation algorithm to automatically generate multiple (for example, N) initial packaging structure design schemes, and evaluate and screen the multiple initial packaging structure design schemes through a pre-trained performance evaluation model to obtain multiple (for example, M) candidate schemes, where M < N; establish a multi-objective optimization function with packaging volume, material cost, and buffer performance of key parts as optimization objectives, and use a multi-objective optimization algorithm to iteratively optimize the multiple candidate schemes to obtain a set of Pareto optimal solution schemes; visually output a set of Pareto optimal solution schemes; associate and store the finally output Pareto optimal solution schemes, corresponding user design requirements, initial parameters, optimization history, and key performance data to generate a structured design version record, supporting retrieval, comparison, and reuse based on design elements.

[0025] Specifically, the pre-trained requirement parsing model is specifically a neural network model based on the Transformer architecture, which includes an input embedding layer, a multi-head self-attention encoding layer, a feed-forward neural network layer, and an output layer.

[0026] Among them, the pre-training steps of the requirement parsing model are as follows:

[0027] Collect pure text content from the enterprise's internal historical design documents, packaging national standard documents (such as the GB / T4857 series), supplier material data sheets, and publicly available packaging engineering papers;

[0028] Use the regular expression r'[^\u4e00-\u9fa5a-zA-Z0-9\s]' to remove all non-Chinese, English, and digital characters, and segment by full stops, semicolons, and line breaks to ensure that the length of each text segment is between 128 and 512 characters, and construct the original domain corpus D raw ; for D raw Apply the WordPiece tokenizer based on the BERT-base-chinese vocabulary to perform tokenization to generate a set of token sequences D tokenized ;

[0029] For each token sequence T = [t1, t2,..., t tokenized in D n , generate the corresponding mask sequence T masked and label sequence T label , and the specific rules are as follows:

[0030] For the position index i of all non-special words in the sequence, generate a random number r. i~Uniform(0,1) ;

[0031] If r i If the value is less than 0.15, then generate another random number s. i~Uniform(0,1) ;

[0032] If s i If T < 0.8, then let T masked[i] =[MASK];

[0033] If 0.8≤s i If the value is less than 0.9, then a word t is randomly sampled from the vocabulary. random Let T masked[i] =t random ;

[0034] If s i If ≥0.9, then keep T. masked[i] =t i At the same time, let T label[i] =t i ;

[0035] If r i If ≥0.15, then let T masked[i] =t_ i And T label[i] =-100 (ignoring the loss);

[0036] Finally, the pre-trained dataset D is generated. pretrain ={(T masked ,T label )};

[0037] Initialize a Transformer encoder with the following structure parameters: word embedding dimension d model =768, hidden dimension d of the feedforward network ff =3072, number of attention heads h=12, number of encoder layers L=12, attention dropout rate drop attn =0.1, dropout rate of feedforward network ff =0.1;

[0038] Define the loss function as the cross-entropy loss at the masked position: Loss = -∑ i∈M log(P(T label[i] |T masked ));

[0039] Where P is the probability distribution predicted by the model, and M represents the set of all masked position indices in the current sequence;

[0040] The optimizer is configured as AdamW, and its hyperparameters are set as β1 = 0.9, β2 = 0.999, ε = 1e-8, and the weight decay rate λ = 0.01;

[0041] A cosine annealing learning rate schedule with linear warm-up is adopted. The warm-up steps W = 10000, the total training steps S = 100000, the maximum learning rate lr max = 5e-5, and the minimum learning rate lr min = 1e-7;

[0042] The training batch size B = 32, the gradient accumulation steps G = 8, and the effective batch size is 256;

[0043] In the first stage, the initialized model is trained for S steps using a random 1 / 100 subset (about 10 billion tokens) of the general Chinese corpus WuDaoCorpora - 2.0. Checkpoints are saved every 5000 steps, and the perplexity is calculated on a 10% held-out validation set. The checkpoint with the lowest perplexity is selected as the stage-one model M general ;

[0044] In the second stage, use D pretrain (about 50 million tokens) to continue training on the weights of M general , with the total number of steps S domain = 50000, the learning rate schedule is reset and warmed up for W domain = 5000 steps, and the other configurations are the same as in S43;

[0045] The accuracy is calculated on a held-out validation set constructed based on package design Q&A every 2500 steps. The checkpoint with the highest accuracy is selected as the pre-trained requirement parsing model for final release, and its file format is.pt for PyTorch or.ckpt for TensorFlow.

[0046] The specific steps to transform natural language requirements into an initial set of design parameters that meet the basic constraint conditions are as follows:

[0047] In the input embedding layer of the requirement parsing model, the natural language design requirements are tokenized and vectorized to generate a sequence of word vectors, which are specifically:

[0048] The input text is tokenized using a domain dictionary based on Jieba tokenizer, and the domain dictionary contains packaging professional terms such as "BOPP film", "buffer corners", and "drop test";

[0049] After tokenization, words such as "de", "le", and "he" in the stop word list are removed;

[0050] Each retained word is transformed into a 300-dimensional word vector x by looking up the embedding matrix of the pre-trained Chinese word vector model sgns.weibo.bigram-char. i ;

[0051] For words not registered in VOC, the zero vector and the mean of the vectors of adjacent words are used to fill the gaps.

[0052] Finally, the sequence of length L is converted into a word vector matrix X∈R^(L×300);

[0053] In the multi-head self-attention encoding layer of the requirement parsing model, parallel self-attention computation is performed on the word vector sequence to capture the global dependencies between design elements and constraints in natural language design requirements, and output context-enhanced feature vectors, specifically:

[0054] The word vector matrix X is passed through three independent trainable weight matrices W. Q W K W V ∈R^(300×64) linear projection yields query matrix Q, key matrix K, and value matrix V, each with dimension R^(L×64);

[0055] Calculate the scaled dot product attention score matrix S=(Q·K^T) / 8, and apply the Softmax function to normalize by row to obtain the attention weight matrix A=softmax(S);

[0056] In weight matrix A, the weight 'a' of business keywords such as "fallproof" to the structural word "corner" is... {防摔,边角} After training, the value reached 0.85, while the weight of "cost" a_{fallproof, cost} was 0.12, thus modeling semantic association;

[0057] Finally, the single-head attention output head = A·V is calculated;

[0058] Eight such attention heads are executed in parallel, and their outputs are concatenated and then passed through a weight matrix W. O Linear projection of ∈R^(512×300) yields the multi-head attention output MHA(X)∈R^(L×300);

[0059] In the feedforward neural network layer of the demand parsing model, nonlinear transformation and feature dimensionality reduction are performed on the context-enhanced feature vectors to extract high-dimensional abstract features. Specifically:

[0060] Each 300-dimensional vector (L in total) at each position in the multi-head attention output MHA(X) is independently input into a two-layer feedforward network;

[0061] The first layer of this network is a linear transformation followed by the GELU activation function:

[0062] F intermediate =GELU(MHA(X) i ·W1+b1);

[0063] Where W1∈R^(300×1200), b1∈R^(1200);

[0064] The second layer is a linear transformation:

[0065] F output =F intermediate ·W2+b2;

[0066] Where W2∈R^(1200×300), b2∈R^(300);

[0067] This operation is performed at each position i, and through the nonlinear mapping of the 1200-dimensional hidden layer, it achieves the compression and refinement of semantic features and filters out noise that is irrelevant to the design.

[0068] Finally, mean pooling is performed on the L positions to obtain the global 300-dimensional reduced feature vector F. global ;

[0069] In the output layer of the requirements analysis model, the high-dimensional abstract features are processed by parameter vector mapping to form the initial design parameter set, which is as follows:

[0070] The 300-dimensional global feature vector F global Input three parallel fully connected prediction heads:

[0071] (1) Size prediction head: Perform linear regression F global ·W size +b size W size ∈R^(300×3),b size ∈R^(3), output a three-dimensional vector [length, width, height] as the size parameter, in millimeters;

[0072] (2) Material classification header: Perform linear transformation F global ·W mat +b mat Followed by Softmax, where W mat ∈R^(300×4),b mat ∈R^(4), output the probability distribution P on the material category set {"BOPP film", "PE film", "corrugated cardboard", "bubble film"}. mat Take argmax(P) mat ) is used as a material type parameter;

[0073] (3) Structural classification header: Perform linear transformation F global ·W str +b tr Followed by Softmax, where W str ∈R^(300×5),b str ∈R^(5), output the probability distribution P on the set of buffer structures {"X-shaped folds", "cellular arrays", "column-shaped", "wavy", "none"}. str Take argmax(P) str ) is used as a parameter for the buffer structure type;

[0074] The final merged output is a structured parameter set: Params = {size: [length, width, height], material: argmax(P mat ), structure: argmax(P str )}.

[0075] In this implementation scheme, the combination of Jieba word segmentation and domain dictionary effectively handles technical terms and common stop words, ensuring that the input text can be accurately understood and converted. Then, a multi-head self-attention mechanism is used to capture the relationship between key information and constraints in the design requirements, enabling the model to comprehensively consider the semantic dependencies between various design elements. The features are further processed and refined through a feedforward neural network, which can effectively filter out key features related to the design goals. Finally, the extracted features are mapped to specific design parameters, such as size, material type, and buffer structure, to generate an initial set of design parameters that meet the requirements. This method not only improves design efficiency but also reduces manual intervention and ensures design accuracy.

[0076] Specifically, such as Figure 2 As shown, the specific steps for automatically generating multiple initial packaging structure design schemes are as follows:

[0077] Based on the initial design parameter set, sampling is performed within the design space to generate multiple sets of basic dimension combinations, specifically:

[0078] Let the nominal dimensions in the initial design parameter set be (L0, W0, H0), and the tolerance zone be ±15%. Construct a three-dimensional design space Ω = [0.85L0, 1.15L0] × [0.85W0, 1.15W0] × [0.85H0, 1.15H0].

[0079] The Latin hypercube sampling algorithm is used to generate N=100 sampling points in this space, ensuring that the edge distribution of each dimension is uniform and the points projected onto any two-dimensional plane are uniformly distributed. The specific algorithm is as follows:

[0080] For each dimension, generate a random permutation of [1,2,…,N], combining them into N three-dimensional coordinates (L). i W i H i );

[0081] Among them, L i =0.85L0+(πL (i)-1 )*(0.3L0) / (N-1), where πL is a random permutation of the length dimension, and W i With H i Similarly, the final output size combination set S = {(L i W i H i )|i=1 to N};

[0082] The pre-built packaging component library is invoked to instantiate parameterized components for each size group, specifically as follows:

[0083] The pre-built packaging component library is stored in XML format, defining parameterized templates for basic box types (such as "Type 0427 top and bottom cover") and cushioning components (such as "45° angle protective wings").

[0084] For each group of dimensions in S, execute:

[0085] Parse the feature parameters in the template (such as "length L", "width W", "height H", "cardboard thickness t");

[0086] (L) i W i H i The material type parameters (such as "BOPP film, t=0.5mm") are bound to the corresponding parameters;

[0087] Call the geometry engine based on ACIS or Parasolid kernel to execute the construction sequence in the template (such as "Extrude(Sketch, H)" or "Shell(Thickness=t)") to generate an accurate boundary representation model;

[0088] The following properties were added to the model: material density ρ = 0.92 g / cm³ (BOPP), Young's modulus E = 2.1 GPa, Poisson's ratio ν = 0.35;

[0089] A topology optimization algorithm is performed on the surfaces requiring buffering to generate a lightweight buffer structure, specifically as follows:

[0090] For the surface Γ that needs buffering (such as the inner bottom surface), it is discretized into a Q4 planar stress element mesh with element number Ne=5000, and the pseudo-density variable ρ is defined for each element. e ∈[0.001,1]. Construct a topology optimization problem:

[0091] ;

[0092] ;

[0093] KU=F;

[0094] Where p=3 is the penalty factor, k0 is the element stiffness matrix, and u e Let Γ be the element displacement, and F be the unit normal load applied at the center point of Γ. The optimization criterion method is used for iterative solution:

[0095] Finite element analysis: Assembly global stiffness matrix K(ρ)=Σ(ρ) e Solving the equilibrium equation, we get U.

[0096] Sensitivity analysis: calculation .

[0097] Update density: The Lagrange multiplier λ satisfies the volume constraint through the bisection method.

[0098] After 200 iterations, the algorithm converges, and ρe is binarized with a threshold of 0.3. The contours of cells with ρe > 0.3 are extracted, and polygonal boundaries are generated using the α-shape algorithm. These boundaries are then subdivided using Catmull-Clark to generate smooth NURBS surfaces, which serve as the geometric definition of the buffer structure.

[0099] The buffer structure and basic components are automatically assembled and conflict detected, generating and outputting multiple initial packaging structure design schemes, specifically:

[0100] Convert the generated NURBS buffer surface into a triangular mesh M. buffer (Target side length ≤ 1mm);

[0101] Assemble it with the BREP model of the basic components:

[0102] Calculate M buffer The center of the bounding box Cb and the center of the target assembly surface Γ CΓ are given. The transformation matrix T is calculated to make Cb coincide with CΓ and their normals aligned.

[0103] After applying T, perform precise triangular mesh-surface intersection detection: for M buffer For each vertex v, calculate the signed distance d(v) from it to the target surface Γ. If d(v) < 0, then penetration occurs.

[0104] If penetration is detected, translate M along the Γ normal direction. buffer Until min(d(v)) > 0.1 mm;

[0105] Ultimately, M bufferThe basic component model is merged into an assembly, and the output is a STEPAP242 file conforming to ISO10303-21 standard, which includes complete geometry, material properties and assembly tree structure, as the i-th initial packaging structure design scheme.

[0106] In this implementation scheme, uniform sampling within the three-dimensional design space ensures comprehensive coverage of size combinations, providing diverse options for design solutions. Subsequently, based on a pre-built packaging component library, combined with specific size and material information, accurate component models are automatically generated, improving design efficiency. By applying a topology optimization algorithm, lightweight design is performed on surfaces requiring cushioning, which not only improves the cushioning effect but also optimizes material usage and reduces costs. Finally, a precise geometry engine and conflict detection algorithm ensure accurate assembly between design elements, avoiding potential design conflicts.

[0107] Specifically, the pre-trained performance evaluation model is a proxy model based on graph attention networks, which includes a graph data construction layer, a multi-level graph convolutional coding layer, and a multi-task prediction output layer.

[0108] The pre-training steps for the performance evaluation model are as follows:

[0109] Explicit dynamics simulations were performed using Abaqus / Explicit. 5000 packaging models with different geometries, materials, and cushioning structures were randomly generated.

[0110] Apply standard drop conditions to each model:

[0111] It falls from a height of 1.2m onto a rigid plane in three ways: bottom, edge, and corner.

[0112] Extract from simulation results:

[0113] Maximum equivalent stress σ max (Unit: MPa), Energy absorption rate η=E absorbed / E impact Material cost C = Σ(ρ i *v i *price i );

[0114] Form a dataset D={(G j ,y j )}, where y j =(σ max ^j,η^j,C^j).

[0115] For each packaging model, construct a heterogeneous graph G=(V,E);

[0116] Each component (such as a panel or buffer block) is treated as node vi, and its feature vector is:

[0117] x i =[type,E,ν,ρ,thickness,area]∈R^6;

[0118] The contact or connection relationship between components is used as edge e ij Its characteristics are:

[0119] e ij =[contact type ,contact area [,angle]∈R^3;

[0120] The graph attention network consists of 4 layers, each of which computes:

[0121] ;

[0122] Where || represents concatenation, K = 8 headers, , For activation functions;

[0123] Global pooling after layer 4: h G =[max pool (H),mean pool (H)];

[0124] Connect the three output heads:

[0125] ŷ=[MLP stress (h G MLP efficiency (h G MLP cost (h G )];

[0126] The loss function is weighted MSE: L = 0.5 * MSE(σ max )+0.3*(1-MSE(η))+0.2*MSE(C);

[0127] Using the Adam optimizer (lr=0.001), train for 500 epochs on 80% of the data, with an early stopping policy of patience=30, and save the model with the minimum loss on the validation set as the pre-trained performance evaluation model.

[0128] The specific steps to obtain multiple candidate solutions are as follows:

[0129] In the graph data construction layer of the performance evaluation model, each initial packaging structure design scheme is transformed into graph structure data, specifically as follows:

[0130] For each STEP format design file, parse its assembly tree and traverse each part instance;

[0131] For the m-th part, extract its geometric center c. m The diagonal length d of the bounding box m Surface area A m Its material properties E m ,ν m ,ρ m splicing to the original feature f m ∈R^7;

[0132] For any two parts m and n, if their bounding boxes intersect or their minimum distance is less than 2mm, then an edge is created. The edge is characterized by a one-hot concatenation of the two part type codes.

[0133] Finally, the j-th design scheme is represented as G. j =(V j E j ), where V j ={f m}, E j ={(m,n,e mn )};

[0134] In the multi-level graph convolutional coding layer of the performance evaluation model, multi-level feature aggregation and encoding are performed on the graph structure data to extract graph feature vectors, specifically as follows:

[0135] G j Input a pre-trained 4-layer graph attention network;

[0136] The first layer projects the node features fm into a d=128 dimensional space: h m ^{(1)}=ReLU(W0f m +b0);

[0137] Subsequent multi-head attention aggregation is performed at each layer:

[0138] h m ^{(l+1)}=BN(h m ^{(l)}+Dropout(|| {k=1} ^{8}Attn k ^{(l)}(h m ^{(l)})));

[0139] Where BN is batch normalization and the Dropout rate is 0.1;

[0140] After four layers of encoding, each node obtains h. m Context-aware features of ^{(4)}∈R^{128};

[0141] Perform global max pooling and average pooling on all nodes in the graph and then concatenate the results: h Gj =[max({h m ^{(4)}}),mean({h m ^{(4)}})]∈R^{256}, which is the graph eigenvector characterizing the performance of the entire scheme;

[0142] In the multi-task prediction output layer of the performance evaluation model, multiple performance prediction values ​​for each initial scheme are predicted in parallel based on graph feature vectors. These predictions are then filtered and ranked according to preset rules, and the top-ranked initial schemes are output as candidate schemes. Specifically:

[0143] The graph feature vector h Gj Input the three pre-trained output heads respectively:

[0144] Stress head: ŷ σj =Sigmoid(W σ2 *ReLU(W σ1 *h Gj +b σ1 )+b σ2 )*100, output the predicted stress value (MPa);

[0145] Efficiency Head: ŷ ηj =Sigmoid(W η2 *ReLU(W η1 *h Gj +b η1 )+b η2 The output prediction efficiency η is calculated.

[0146] Cost Head: ŷ Cj =ReLU(W C2 *ReLU(W C1 *h Gj +b C1 )+b C2 Output the predicted cost (in yuan).

[0147] For all N=100 schemes, the prediction matrix Y is obtained. pred ∈R^{100×3}. First, a feasibility screening is performed:

[0148] If ŷ σj >σ yield (e.g., the yield strength σ of BOPP film) yield If the pressure is 55MPa, then eliminate scheme j;

[0149] Assume there are P remaining solutions. For the remaining solutions... η With ŷ CPerform min-max normalization. Set the weight vector w = [0.6, 0.4] (preference performance) and calculate the overall score S. j =0.6*ŷ ηjnorm +0.4*(1-ŷ Cjnorm );

[0150] Press S j Sort the solutions in descending order, select the first M = min(P, 20), and set their indices to {j1, j2, ..., j...}. M The corresponding original design scheme is output as a candidate scheme.

[0151] In this implementation scheme, each design scheme is first transformed into graph structure data. Graph nodes and edges are constructed based on geometric features and spatial relationships between parts. This effectively represents the interrelationships between various design elements. This graph-structured representation not only improves the expressive power of the design scheme but also allows for the extraction of deeper structural features through graph convolutional networks. Next, through multi-layer encoding of a graph attention network, the global features of the design scheme are aggregated to obtain the context-aware features of each design scheme, thereby comprehensively evaluating the performance of different designs. Based on this, a multi-task prediction output layer is used to predict key performance indicators such as stress, efficiency, and cost. Finally, through comprehensive evaluation and ranking of candidate schemes, it is ensured that the output design scheme meets performance requirements while effectively controlling costs and improving efficiency. This method not only improves the accuracy of design scheme selection but also greatly enhances optimization efficiency.

[0152] Specifically, a multi-objective optimization function is established with packaging volume, material cost, and cushioning performance of key components as optimization objectives, as follows:

[0153] Each candidate solution is encoded as an optimization variable, specifically as follows:

[0154] Encode the i-th candidate solution as an optimization variable vector X. i =[l i ,w i ,h i ,t {i1} ,t {i2} ,...,t {ik} ,d {i1} ,d {i2} ,...,d {im} ];

[0155] Among them (l) i ,w i ,h i ) represents the outer packaging dimensions, t {ik} Let d be the thickness of the k-th component. {im}For the characteristic dimensions of the m-th buffer structure (such as fold spacing, honeycomb aperture);

[0156] The functions for calculating packaging volume, material cost, and cushioning performance are constructed separately, as follows:

[0157] Packaging volume calculation function: V(X) i )=l i ×w i ×h i ;

[0158] Material cost calculation function: C(X) i )=Σ {k} (ρ k ×v k (X i )×p k ), where ρ k v is the material density. k p represents the total volume of the material in the design. k Price per unit weight of materials;

[0159] Buffer performance evaluation function: P(X) i )=f{FEA}(Xi) or P(Xi)=f {surrogate} (X i );

[0160] f {FEA} This indicates the negative value of the maximum acceleration response (amax) of key parts (such as the area corresponding to the corners of the product) under standard drop conditions, calculated directly through finite element simulation (i.e., P = -a). max The smaller the value (the more negative), the better the buffering performance;

[0161] f {surrogate} This represents the buffer performance metric predicted by the trained proxy model.

[0162] The combined functions form a multi-objective vector function to be optimized, specifically:

[0163] Construct a multi-objective optimization problem: minF(X) = [V(X), C(X), -P(X)]^T;

[0164] If the weighted or utility function method is used, it can be constructed as a single objective minU(X)=ω1*V(X)+ω2*C(X)-ω3*P(X), where ω1,ω2,ω3 are weight coefficients, and ω1+ω2+ω3=1.

[0165] The specific steps for iteratively optimizing multiple candidate solutions using a multi-objective optimization algorithm are as follows:

[0166] The first generation population is initialized by selecting multiple candidate solutions, specifically as follows:

[0167] Let the optimization variable vector {X1,X2,...,X} corresponding to the M candidate solutions be... M} is directly used as the initial population P0, with a population size N=M;

[0168] If M is insufficient, it is supplemented to the preset size N (e.g., N=100) by randomly sampling within the range of values ​​of each variable.

[0169] In each iteration, the multi-objective optimization function value of each individual in the population is calculated, specifically as follows:

[0170] For the t-th generation population P t Each individual X in j Call the multi-objective vector function F(X) j ), calculate its three target values ​​[V j C j ,-P j ];

[0171] Based on the multi-objective optimization function value, selection, crossover, and mutation operations are performed to generate the offspring population, specifically as follows:

[0172] Selection: A binary tournament selection method is used. Randomly select from P... t Select two individuals, compare their Pareto rank (obtained from fast non-dominated sorting) and crowding, and select the better one to enter the mating pool. Repeat until the mating pool size is N.

[0173] Crossover: Pair individuals in the mating pool with each other with probability p. c =0.9 Perform simulated binary crossover. For two parent vectors X^p1 and X^p2, generate offspring: X^{c1}=0.5*[(1+β)X^{p1}+(1-β)X^{p2}], X^{c2}=0.5*[(1-β)X^{p1}+(1+β)X^{p2}], where β is the distribution η c =20 generates random numbers.

[0174] Mutation: For each offspring individual after crossover, with probability p m Perform polynomial mutation on variable xk = 1 / n (where n is the dimension of the variable). This generates the mutated value xk. k '=x k +δ k *(x k ^Ux k ^L), where δ k For distribution η m =20 generates random mutations, x k ^U and xk ^L represents the upper and lower bounds of the variable.

[0175] By merging parent and offspring populations, environmental selection is performed based on Pareto dominance and distribution assessment to select a new generation of populations. Specifically:

[0176] Parental population P t (Size N) and offspring population Q t (Size N) merged into R t (Scale 2N);

[0177] For R t Perform a fast nondominated sort and divide it into multiple Pareto front layers F1, F2, ...;

[0178] Starting from F1, individuals are added to the new population P layer by layer. {t+1} Until a certain layer F is added l It will exceed the scale N;

[0179] For F l Layer P is calculated, and the crowding distance of each individual (i.e., the sum of the distances between adjacent individuals at each objective function value) is calculated. Individuals are then selected and added to P in descending order of crowding. {t+1} Continue until N positions are filled;

[0180] Repeat the iteration until the termination condition is met, specifically:

[0181] Set the maximum number of iterations G max =200, when the current iteration number t≥G max The process can be terminated at the appropriate time; or the population dispersion index in the target space (such as hypervolume HV) can be calculated, and if the relative improvement rate of HV for 20 consecutive generations is <1e-4, then the process can be terminated early.

[0182] A set of Pareto optimal solutions is obtained, specifically:

[0183] When the optimization iteration terminates, all non-dominated solutions are extracted from the final population, specifically:

[0184] Let the final population be P. {final} Perform a fast non-dominated sort on them, extracting all individuals in the first Pareto front layer F1. These individuals are non-dominated to each other and are in the P... {final} It is not controlled by any other individual;

[0185] This constitutes a set of solutions that achieve Pareto optimality in terms of packaging volume, material cost, and cushioning performance, specifically as follows:

[0186] The extracted non-dominated solution individuals {X1^*,X2^*,...,X K^*}Decode it back to the corresponding design solution. For any two solutions A and B in this set, it does not satisfy that A is not worse than B in all three objectives and is strictly better than B in at least one objective;

[0187] Output as a set of Pareto-optimal solution schemes, specifically:

[0188] Output the K Pareto-optimal solutions, together with their three objective function values [V, C, P], in the form of structured data (such as a JSON array) or a visualization chart (such as a three-dimensional scatter plot, parallel coordinate plot) for the user to make decisions.

[0189] In this implementation scheme, each candidate solution is first encoded as an optimization variable vector, comprehensively considering factors such as size, material, and buffer structure, providing a clear structure for subsequent optimization calculations. This method constructs a multi-objective optimization problem by setting objective functions such as packaging volume, material cost, and buffer performance, making it possible to trade off and optimize between different solutions. During the optimization process, the multi-objective algorithm used simulates the process of natural evolution through selection, crossover, and mutation operations, and continuously iterates and optimizes to finally obtain a set of Pareto-optimal solutions. In this way, users can not only select the packaging design solution with the best performance but also ensure cost reduction and improved buffer effect in actual production. In addition, through the output of Pareto-optimal solutions, users can intuitively see the performance of each design solution in multiple objectives, providing strong support for decision-making. This method not only improves the design efficiency but also ensures the efficiency and accuracy of the design, meets the actual needs, and provides a comprehensive optimization solution for packaging design.

[0190] Please refer to Figure 3 , an embodiment of the present invention provides a technical solution: a protective film packaging design optimization system, including: a requirement analysis and parameterization unit, configured to receive the natural language design requirements and basic constraint conditions input by the user, and convert the natural language requirements into an initial design parameter set through a pre-trained requirement analysis model. The initial design parameter set includes size parameters, material type parameters, and buffer structure type parameters; a solution generation and evaluation screening unit, configured to automatically generate multiple (for example, N) initial packaging structure design solutions based on the initial design parameter set by calling a preset packaging component library and a topology generation algorithm, and evaluate and screen the multiple initial packaging structure design solutions through a pre-trained performance evaluation model to obtain multiple (for example, M) candidate solutions, where M < N; a multi-objective optimization solving unit, configured to establish a multi-objective optimization function with packaging volume, material cost, and buffer performance at key parts as optimization objectives, and use a multi-objective optimization algorithm to iteratively optimize the multiple candidate solutions to obtain a set of Pareto-optimal solution schemes; a visualization output unit, configured to visually output a set of Pareto-optimal solution schemes.

[0191] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0192] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for optimizing the design of protective film packaging, characterized in that, Includes the following steps: The system receives natural language design requirements and basic constraints input by the user, and transforms the natural language requirements into a set of initial design parameters that meet the basic constraints through a pre-trained requirement parsing model. The set of initial design parameters includes size parameters, material type parameters, and buffer structure type parameters. Based on the initial design parameter set, a pre-set packaging component library and topology generation algorithm are invoked to automatically generate multiple initial packaging structure design schemes. The multiple initial packaging structure design schemes are then evaluated and screened through a pre-trained performance evaluation model to obtain multiple candidate schemes. A multi-objective optimization function was established with packaging volume, material cost, and cushioning performance of key parts as optimization objectives. A multi-objective optimization algorithm was used to iteratively optimize multiple candidate solutions to obtain a set of Pareto optimal solutions. A set of Pareto optimal solutions will be visualized.

2. The method for optimizing the design of protective film packaging according to claim 1, characterized in that, The pre-trained requirement parsing model is specifically a neural network model based on the Transformer architecture, which includes an input embedding layer, a multi-head self-attention encoding layer, a feedforward neural network layer, and an output layer.

3. The method for optimizing the design of protective film packaging according to claim 2, characterized in that, The specific steps to transform natural language requirements into a set of initial design parameters that meet the basic constraints are as follows: In the input embedding layer of the requirement analysis model, the natural language design requirements are segmented and vectorized to generate a sequence of word vectors. In the multi-head self-attention encoding layer of the requirement parsing model, parallel self-attention calculation is performed on the word vector sequence to capture the global dependency relationship between design elements and constraints in the natural language design requirements and output context-enhanced feature vectors. In the feedforward neural network layer of the demand parsing model, nonlinear transformation and feature dimensionality reduction are performed on the context-enhanced feature vectors to extract high-dimensional abstract features. In the output layer of the requirements analysis model, the high-dimensional abstract features are processed by parameter vector mapping and form the initial design parameter set.

4. The method for optimizing the design of protective film packaging according to claim 1, characterized in that, The specific steps for automatically generating multiple initial packaging structure design schemes are as follows: Based on the initial set of design parameters, sampling is performed within the design space to generate multiple sets of basic dimension combinations; Call the pre-built packaging component library to instantiate parameterized components for each size group; A topology optimization algorithm is performed on the surfaces that require buffering to generate lightweight buffer structures; The system automatically assembles and detects conflicts between the buffer structure and the basic components, generating and outputting multiple initial packaging structure design schemes.

5. The method for optimizing the design of protective film packaging according to claim 1, characterized in that, The pre-trained performance evaluation model is specifically a proxy model based on graph attention networks, which includes a graph data construction layer, a multi-level graph convolutional coding layer, and a multi-task prediction output layer.

6. The method for optimizing the design of protective film packaging according to claim 5, characterized in that, The specific steps to obtain multiple candidate solutions are as follows: In the graph data construction layer of the performance evaluation model, each initial packaging structure design scheme is transformed into graph structure data; In the multi-level graph convolutional coding layer of the performance evaluation model, multi-level feature aggregation and encoding are performed on the graph structure data to extract graph feature vectors; In the multi-task prediction output layer of the performance evaluation model, multiple performance prediction values ​​of each initial scheme are predicted in parallel based on the graph feature vector, and the initial schemes are filtered and sorted according to preset rules, and the top-ranked initial schemes are output as candidate schemes.

7. The method for optimizing the design of protective film packaging according to claim 1, characterized in that, A multi-objective optimization function is established, with packaging volume, material cost, and cushioning performance of key components as optimization objectives. Specifically: Encode each candidate solution as an optimization variable; Construct functions for calculating packaging volume, material cost, and cushioning performance, respectively. The combined functions form a multi-objective vector function to be optimized.

8. The method for optimizing the design of protective film packaging according to claim 7, characterized in that, The specific steps for iteratively optimizing multiple candidate solutions using a multi-objective optimization algorithm are as follows: Multiple candidate solutions are initialized to form the first generation population; In each iteration, the multi-objective optimization function value for each individual in the population is calculated; Based on the multi-objective optimization function value, selection, crossover, and mutation operations are performed to generate the offspring population; By merging parent and offspring populations and conducting environmental selection based on Pareto dominance and distribution evaluation, a new generation of populations is selected. Repeat the iteration until the termination condition is met.

9. The method for optimizing the design of protective film packaging according to claim 8, characterized in that, A set of Pareto optimal solutions is obtained, specifically: When the optimization iteration terminates, all non-dominated solutions are extracted from the final population; This constitutes a set of solutions that achieve Pareto optimality in terms of packaging volume, material cost, and cushioning performance. The output is the set of Pareto optimal solutions.

10. A protective film packaging design optimization system, employing the protective film packaging design optimization method according to any one of claims 1-9, characterized in that, include: The requirement parsing and parameterization unit is used to receive the natural language design requirements and basic constraints input by the user, and to transform the natural language requirements into an initial design parameter set through a pre-trained requirement parsing model. The initial design parameter set includes size parameters, material type parameters, and buffer structure type parameters. The scheme generation and evaluation screening unit is used to automatically generate multiple initial packaging structure design schemes based on the initial design parameter set, by calling a pre-set packaging component library and topology generation algorithm, and to evaluate and screen the multiple initial packaging structure design schemes through a pre-trained performance evaluation model to obtain multiple candidate schemes. The multi-objective optimization solution unit is used to establish a multi-objective optimization function with packaging volume, material cost and cushioning performance of key parts as optimization objectives, and to use a multi-objective optimization algorithm to iteratively optimize multiple candidate solutions to obtain a set of Pareto optimal solutions. The visualization output unit is used to visualize a set of Pareto optimal solutions.

Images