A method and system for intelligent structural design of manufacturing equipment based on meta-motion

By constructing a database of standard meta-action processes and mapping relationships for manufacturing equipment, and combining it with a cascaded recommendation model, automated and intelligent design of manufacturing equipment was achieved, solving the problems of design consistency and efficiency, and improving the rationality and reproducibility of the design.

CN123116846APending Publication Date: 2026-11-17WUXI XIJIAN TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611401287.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The existing manufacturing equipment design process lacks standardized models, resulting in poor design consistency and reproducibility. It relies on human experience, leading to omissions or redundancies in actions, long design iteration cycles, and a lack of automated end-to-end design tools.

Method used

Establish a standard meta-action process and meta-action structure library for manufacturing equipment, construct a mapping relationship database of design requirements-meta-actions-meta-action structures, and adopt a cascaded recommendation model for intelligent design, including a three-level closed-loop collaboration of semantic matching, graph attention, and twin architecture to generate a three-dimensional assembly model.

Benefits of technology

It has enabled automated and intelligent design of manufacturing equipment, improved design consistency and efficiency, reduced reliance on human judgment, and ensured the rationality and reproducibility of design schemes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116846A_ABST
    Figure CN123116846A_ABST
Patent Text Reader

Abstract

This invention relates to the field of manufacturing equipment design technology, specifically to a method and system for intelligent structural design of manufacturing equipment based on meta-actions. The method includes: establishing standard meta-action processes and meta-action structure libraries for manufacturing equipment according to its characteristics; constructing a mapping relationship database between design requirements, meta-actions, and meta-action structures; constructing and training a cascaded recommendation model for manufacturing equipment structural design based on meta-actions. The cascaded recommendation model consists of three layers: the first layer is the generation layer of meta-action components and sequences; the second layer is the generation layer of meta-action structures and layouts; and the third layer is the result comparison and judgment layer. These three layers collaborate through hierarchical condition transmission and feedback correction, forming a three-level closed-loop collaboration of meta-action components, structure generation, and comparison and judgment; importing design requirements into the cascaded recommendation model to generate manufacturing equipment design schemes. This invention enables automated intelligent design of intelligent manufacturing equipment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of manufacturing equipment design technology, and in particular to a method and system for intelligent structural design of manufacturing equipment based on meta-motions. Background Technology

[0002] In the design process of non-standard intelligent manufacturing equipment, engineers first need to determine which actions the equipment should perform based on the design requirements, and then plan the corresponding meta-action structure for each action, that is, the smallest complete mechanical structure unit that completes the function of each independent action. Multiple meta-action structures are combined and assembled according to the process sequence and spatial layout, and finally the construction of the three-dimensional digital model of the whole machine is completed.

[0003] Currently, this design process of "requirements-action determination-meta-action structure planning-assembly-3D modeling" heavily relies on engineers' personal experience, resulting in the following prominent problems: First, equipment actions lack standardized models; different engineers may identify different sets of actions for the same design requirement, and the granularity, naming, and functional definition of actions vary from person to person, making it difficult to guarantee design consistency and reproducibility; second, the mapping from design requirements to required actions relies on manual experience judgment, requiring engineers to deduce which actions the equipment should perform and the timing and logical relationships between these actions based on experience, easily leading to the omission of key actions or the introduction of redundant actions; third, the design... The expression of design requirements is often vague and incomplete, and key engineering constraints require multiple rounds of communication to complete; fourth, there is a lack of systematic mapping relationship data between functional / performance requirements, meta-actions, and meta-action structures, making it impossible to transform historical design experience into reusable structural recommendations; fifth, the meta-action structure planning stage lacks a comparison and verification mechanism with historical mapping data, and the rationality of the generated structural and layout schemes relies solely on manual judgment, lacking automatic verification of whether structural performance is progressively adapted and whether spatial layout is compatible; sixth, the planning, calculation, and modeling stages are fragmented, and the planning results cannot be automatically converted into a three-dimensional assembly model, resulting in a long design iteration cycle.

[0004] Existing design support tools primarily target single stages: engineering calculation software like MITcalc only provides parameter verification at the component level and lacks the ability to plan and recommend at the action level; 3D CAD software such as SolidWorks and CATIA provide modeling and assembly functions but do not understand the semantic logic of equipment design and cannot automatically determine the required actions and recommend corresponding meta-action structures from design requirements. Furthermore, existing tools have not established standardized models of manufacturing equipment actions, cannot construct a systematic mapping relationship between functional and performance requirements, meta-actions, and meta-action structures, and lack the ability to cascade recommendations based on mapping relationship comparison and verification, thus failing to achieve end-to-end automated output from design requirements to 3D assembly models.

[0005] For the reasons mentioned above, there is an urgent need for a design method and system that integrates standardized modeling of manufacturing equipment actions, construction of functional and performance requirements-meta-action-meta-action structure mapping relationships, generation of meta-action components and structure recommendation and layout planning based on cascade recommendation models, comparison and verification of generated schemes with historical mapping data, and automatic output of 3D assembly models, in order to realize the automated intelligent design of intelligent manufacturing equipment. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for intelligent structural design of manufacturing equipment based on meta-actions, which can realize the automated intelligent design of intelligent manufacturing equipment.

[0007] To achieve the above objectives, in a first aspect, the present invention provides a method for intelligent structural design of manufacturing equipment based on meta-actions, comprising: S1 establishes standard meta-action processes and meta-action structure libraries for manufacturing equipment based on the characteristics of the manufacturing equipment. S2 constructs a database of mapping relationships between the design requirements, meta-actions, and meta-action structures of manufacturing equipment; S3 constructs and trains a cascaded recommendation model for manufacturing equipment structural design based on meta-actions; S4 imports design requirements into the cascaded recommendation model to generate manufacturing equipment design schemes.

[0008] In step S1, The standard meta-action process adopts at least a two-level architecture: the first level is the action type, which includes at least main material loading, auxiliary material loading, transfer, positioning and clamping, processing, unloading and detection; the second level is several meta-actions under each action type, and the meta-action is the smallest complete mechanical structure unit that completes an independent action function.

[0009] In step S1, The meta-action structure is represented by a triplet of power source, power transmission, and end effector; each meta-action structure includes a structure type identifier, key performance parameter specifications, and an association index with the meta-action identifier.

[0010] In step S2, The design requirement-meta-action-meta-action structure-layout mapping relationship is a three-layer mapping. The first layer is the mapping from design requirements to the set of meta-actions and their order; the second layer is the mapping from meta-actions to the combination constraints of power source structure, power transmission structure and end execution structure types; the third layer is the mapping from meta-action structure type to specific design scheme and typical layout. The mapping relationship is constructed from historical design case data after structured extraction and verification.

[0011] In step S3, The cascaded recommendation model consists of three layers: the first layer is the generation layer of the equipment's meta-action composition and sequence; the second layer is the generation layer of the meta-action structure and layout; and the third layer is the result comparison and judgment layer. The three layers cooperate through step-by-step condition transmission and feedback correction to form a three-level closed-loop collaboration of meta-action composition, structure generation, and comparison and judgment.

[0012] In the cascaded recommendation model: the first layer is a meta-action composition and sequence generation layer based on semantic matching cross-attention network, which matches functional requirements with standard meta-action processes to filter action types, performs constraint verification according to performance requirement parameters, determines the arrangement order according to manufacturing process logic, generates meta-action composition and sequence scheme, and injects the output meta-action composition and sequence into the second layer driving structure and layout generation. The second layer is a conditional graph generation layer based on graph attention network. According to the mapping relationship in step S2, it determines the three-element structure form for each element action, and plans the layout according to the substructure space parameters and assembly interface to generate the structure and layout scheme. The output scheme is injected into the third layer to drive the rationality evaluation. The third layer is a comparison and evaluation layer based on a twin architecture. It compares the generated scheme with the mapping relationship of step S2, verifies the adaptability of the performance transmission link and the compatibility of the spatial layout, outputs the recommended scheme when the threshold is met, and generates an evaluation residual signal to feed back to the second layer to trigger correction when the threshold is not met. If the threshold is still not met, it feeds back to the first layer to trigger the adjustment of the meta-action composition, forming a three-level closed-loop collaboration.

[0013] Specifically, step S4 includes: The recommended solution is input into the 3D generated model. The graph decoder encodes and decodes each element action structure into positioning and mating constraint parameters in the assembly coordinate system. The standard part 3D model is called and the non-standard part 3D model is generated, thus completing the automatic assembly and generating of the 3D digital model.

[0014] Secondly, the present invention also provides a manufacturing equipment structure intelligent design system based on meta-actions, including a standard library module, a mapping data module, a cascaded recommendation module, an interaction module, and a three-dimensional assembly module; the standard library module, the mapping data module, the cascaded recommendation module, the interaction module, and the three-dimensional assembly module are connected in sequence; The standard library module is used to store standard meta-action processes and meta-action structure libraries; The mapping data module is used to store the mapping relationship between function and performance requirements, meta-actions, and meta-action structures. The mapping relationship includes three layers of mapping data: requirement to meta-action, meta-action to structure type, and structure type to specific solution. The cascaded recommendation module is used to execute the cascaded recommendation model, including a meta-action composition planning unit, a structure and layout generation unit, and a comparison and evaluation unit. The three units work together through step-by-step condition transmission and feedback correction. The interaction module is used to receive design requirements and call the cascaded recommendation module to pass the recommended solution to the 3D assembly module; The three-dimensional assembly module is used to input the recommended solution into the three-dimensional generation model, call standard parts and generate three-dimensional models of non-standard parts, complete automatic assembly and output the design results.

[0015] The present invention provides a method and system for intelligent structural design of manufacturing equipment based on meta-actions. It integrates standardized modeling of manufacturing equipment actions, construction of functional and performance requirements-meta-action-meta-action structural mapping relationship, generation of meta-action components and structural recommendation and layout planning based on cascade recommendation model, comparison and verification of generated schemes with historical mapping data, and automatic output of three-dimensional assembly models, which can realize the automated intelligent design of intelligent manufacturing equipment. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0017] Figure 1 This is a flowchart of an intelligent structural design method for manufacturing equipment based on meta-actions, according to the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of an intelligent design system for manufacturing equipment based on meta-actions, according to the present invention.

[0019] 101-Standard Library Module, 102-Mapping Data Module, 103-Cascade Recommendation Module, 104-Interaction Module, 105-3D Assembly Module. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0021] Firstly, please refer to Figure 1 This invention provides a method for intelligent structural design of manufacturing equipment based on meta-actions, comprising: S1 establishes standard meta-action processes and meta-action structure libraries for manufacturing equipment based on the characteristics of the manufacturing equipment. In this step, the standard meta-action flow adopts at least a two-level architecture: the first level is the action type, which includes at least main material loading, auxiliary material loading, positioning and clamping, positioning, processing, unloading, and detection; the second level consists of several meta-actions under each action type, where each meta-action is the smallest complete mechanical structural unit that completes an independent action function. The meta-action structure is represented by a triplet of power source, power transmission, and end effector; each meta-action structure includes a structure type identifier, key performance parameter specifications, and an association index with the meta-action identifier. The structured design scheme includes standard part selection information and parametric modeling specifications for non-standard parts.

[0022] S2 constructs a database of mapping relationships between the design requirements, meta-actions, and meta-action structures of manufacturing equipment; In this step, the design requirement-meta-action-meta-action structure-layout mapping relationship is a three-layer mapping. The first layer is the mapping from design requirements to the set of meta-actions and their order; the second layer is the mapping from meta-actions to the combination constraints of power source structure, power transmission structure and end execution structure types; the third layer is the mapping from meta-action structure type to specific design scheme and typical layout. The mapping relationship is constructed from historical design case data after structured extraction and verification.

[0023] S3 constructs and trains a cascaded recommendation model for the structural design of manufacturing equipment based on meta-actions. The cascaded recommendation model consists of three layers: the first layer is the generation layer of meta-actions of the equipment and their sequence; the second layer is the generation layer of meta-action structure and layout; and the third layer is the result comparison and judgment layer. The three layers are coordinated through step-by-step condition transmission and feedback correction to form a three-level closed-loop coordination of meta-action composition, structure generation, and comparison and judgment. In this step, the cascaded recommendation model consists of: the first layer being a meta-action composition and sequence generation layer based on a semantic matching cross-attention network, which matches functional requirements with standard meta-action processes to filter action types, performs constraint verification based on performance requirement parameters, determines the arrangement order according to manufacturing process logic, generates meta-action composition and sequence schemes, and injects the output meta-action composition and sequence into the second layer driving structure and layout generation. The second layer is a conditional graph generation layer based on graph attention network. According to the mapping relationship in step S2, it determines the three-element structure form for each element action, and plans the layout according to the substructure space parameters and assembly interface to generate the structure and layout scheme. The output scheme is injected into the third layer to drive the rationality evaluation. The third layer is a comparison and evaluation layer based on a twin architecture. It compares the generated scheme with the mapping relationship of step S2, verifies the adaptability of the performance transmission link and the compatibility of the spatial layout. If the threshold is met, a recommended scheme is output; if not, an evaluation residual signal is generated and fed back to the second layer to trigger correction. If it is still not met, it is fed back to the first layer to trigger the adjustment of the meta-action composition, forming a three-level closed-loop collaboration. Specifically, the hierarchical condition transmission and feedback correction collaboration mechanism is as follows: the meta-action composition output by the first layer is injected into the second layer to drive the generation of the structure and layout; the scheme output by the second layer is injected into the third layer to drive the rationality evaluation; the evaluation residual signal generated by the third layer is first fed back to the second layer to trigger correction; if it is still not met, it is fed back to the first layer to trigger the adjustment of the meta-action composition, forming a three-level closed-loop collaboration. The first layer matches the functional requirements with the standard meta-action process to filter the action type, performs constraint verification according to the performance requirement parameters, determines the arrangement order according to the manufacturing process logic, and generates the meta-action composition and sequence scheme. The second layer determines the structural type combination for each meta-action and selects the design scheme according to the mapping relationship of step S2, plans the layout according to the substructure spatial parameters and assembly interface, and generates the structure and layout scheme. The third layer uses structural rationality and layout rationality as evaluation dimensions, compares the mapping relationship between the generated scheme and step S2, and verifies the performance transmission link adaptability and spatial layout compatibility; when the threshold is met, a recommended scheme is output, and when it is not met, an evaluation residual signal is generated to trigger correction.

[0024] S4 imports design requirements into the cascaded recommendation model to generate manufacturing equipment design schemes.

[0025] Specifically, this includes: inputting the recommended solution into the 3D generated model, encoding and decoding each element action structure into positioning and mating constraint parameters in the assembly coordinate system through a graph decoder, calling the 3D model of standard parts and generating the 3D model of non-standard parts, and completing the automatic assembly to generate a 3D digital model.

[0026] To better understand the intelligent design method for manufacturing equipment structure based on meta-actions of the present invention, specific embodiments are used below to illustrate the method. In this embodiment, the overall design of an automatic milling machine for a certain type of aluminum alloy shell part is taken as an example to explain in detail the specific implementation of the intelligent design method for manufacturing equipment structure based on meta-actions of the present invention. This equipment needs to complete the following manufacturing tasks: grabbing and transferring the aluminum alloy shell blank from the hopper to the processing station, clamping and fixing it at the processing station, milling the top and side surfaces of the shell, pushing the finished product to the unloading station, and inspecting the appearance and dimensional quality of the finished product; that is, a complete automatic milling machine with a "loading-positioning-processing-unloading-inspection" process. The specific steps of the intelligent design method for manufacturing equipment structure based on meta-actions are as follows: Step S1: Based on the characteristics of the manufacturing equipment, establish standard meta-action flows and meta-action structure libraries for the manufacturing equipment. Specifically, based on the idea of ​​atomized decomposition of actions, all executable operations of the manufacturing equipment in the complete manufacturing process are hierarchically decomposed and uniformly encoded according to functional semantics to establish standard meta-action flows. In this embodiment, the standard meta-action flows adopt a two-level architecture: the first level is the action type, and the second level is the specific meta-actions under each type. The specific division is as follows: The main material feeding category includes gripping feeding, pushing feeding, and adsorption feeding; the auxiliary material feeding category includes auxiliary material pushing, auxiliary material adsorption, and auxiliary material sorting; the transfer category includes horizontal transfer, vertical transfer, rotary transfer, and composite transfer; the positioning category includes clamping positioning, pin hole positioning, and vision positioning; the processing category includes milling, drilling, stamping, laser processing, and gluing; the unloading category includes pushing unloading, gripping unloading, and flipping unloading; and the inspection category includes vision inspection, dimensional inspection, and force inspection.

[0027] Each meta-action is defined with a unique identifier code. For example, "gripping and loading" is identified as "MA-FE-001" and "milling" is identified as "MA-MC-001". Each meta-action includes a functional description, key performance parameter specifications, and a mapping relationship with the meta-action structure type. Taking "gripping and loading" as an example: the functional description is "transferring the workpiece from the external hopper to the machining station through a clamping mechanism"; the key performance parameter specifications include the force parameter range (gripping force 550N), speed parameter range (moving speed 0.1-1.0m / s), accuracy parameter range (positioning accuracy ±0.05~±0.5mm), and stroke parameter range (moving stroke 100~500mm); the mapping relationship with the meta-action structure type specifies that the applicable power source structure type for this meta-action is electric or pneumatic, the power transmission structure type is gear-driven or pneumatic direct drive, and the end effector structure type is mechanical gripper or pneumatic gripper.

[0028] Simultaneously, a meta-action structure library is established to store structured design schemes corresponding to various meta-actions. Each meta-action structure is represented by a triplet model consisting of a power source structure, a power transmission structure, and an end-effector structure. Taking a meta-action structure corresponding to "grabbing and loading" as an example: the power source structure is a stepper motor (model 28BYJ-48, rated torque 0.35 N·m), the power transmission structure is a gear reduction transmission (reduction ratio 1:64), and the end-effector structure is a two-finger mechanical gripper (gripping force range 5~30 N). Each meta-action structure includes a structure type identifier, key performance parameter specifications, and an association index with the meta-action identifier. The structured design scheme includes standard part selection information and parametric modeling specifications for non-standard parts.

[0029] Step S2: Construct a database of mapping relationships between design requirements, meta-actions, and meta-action structures for manufacturing equipment. Specifically, this involves collecting data from 800 historical design cases, extracting and validating the functional / performance requirements, executed meta-actions, and corresponding meta-action structure planning results from each case, and constructing a three-layer mapping relationship: The first layer maps functional / performance requirements to meta-actions. For example, the functional requirement "feeding aluminum alloy shell blanks" and the performance requirement "gripping force ≥10N, positioning accuracy ≤±0.1mm" are mapped to the meta-action set {gripping and feeding}, with the timing sequence being step 1; the functional requirement "shell clamping and fixing" and the performance requirement "clamping force ≥50N" are mapped to the meta-action set {clamping and positioning}, with the timing sequence being step 2.

[0030] The second layer is the mapping from meta-actions to meta-action structure types. For example, the meta-action "grabbing and loading" is mapped to the combined constraints of the power source structure type {electric type, pneumatic type}, the power transmission structure type {gear transmission type, pneumatic direct drive type}, and the end effector structure type {mechanical gripper type, pneumatic gripper type}.

[0031] The third layer maps the meta-action structure types to specific structured design schemes and typical layouts. For example, the combination of meta-action structure types, "electric type + gear transmission type + mechanical gripper type", is mapped to specific structural schemes and optimal ordering, as well as to typical spatial layout parameters and descriptions of assembly relationships between substructures.

[0032] Step S3: Construct and train a cascaded recommendation model for manufacturing equipment structural design based on meta-actions. Specifically, the cascaded recommendation model consists of three layers, which collaborate through hierarchical condition transmission and feedback correction to form a three-level closed-loop collaboration of composition, structure, and evaluation. The specific model architecture and construction method used in each layer, as well as the overall training steps of the cascaded recommendation model, are described below.

[0033] The first layer of the network is the meta-action composition planning layer, which adopts a BERT-based semantic matching cross-attention network. This layer receives functional requirement text and performance requirement parameters as input. The functional requirement text is encoded into a 768-dimensional requirement semantic vector through a pre-trained BERT-base-Chinese model (using the Chinese BERT pre-trained weights jointly released by Harbin Institute of Technology and iFlytek). The performance requirement parameters are composed of parameter vectors formed by the values ​​of various performance indicators (milling accuracy, gripping force, clamping force, production cycle time, external dimensions, etc.) in a fixed order. The main body of the network is a two-layer cross-attention matching module: The first layer of cross-attention uses the requirement semantic vector as the query and the functional description vectors of each action type in the standard meta-action flow as the key and value, calculates the cosine similarity attention weight, and sorts and filters the matching action type subsets according to the weight; The second layer of cross-attention, within the matched action type subset, uses the parameter vector as the query and the key performance parameter specification vector of each meta-action as the key value, calculates the constraint satisfaction, and removes meta-actions that do not meet the performance threshold. Finally, the process logic sorting module determines the order of the selected meta-actions based on the topological sorting rules of the manufacturing process knowledge graph, and outputs the meta-action composition scheme and its timing information. The functional description vector and performance parameter specification vector of this layer are both derived from the BERT encoding representation of the standard meta-action flow constructed by S1.

[0034] The second layer is the structure and layout generation layer, which adopts a conditional graph generation network based on Graph Attention Network (GAT). This layer takes the meta-action components and their temporal information output from the first layer as conditional inputs to construct a heterogeneous graph with candidate meta-action structure types as nodes. The node features in the graph are formed by concatenating the structure type encoding and performance parameter encoding from the S2 mapping relationship, and edge connections represent the compatibility relationships between nodes. The network uses the GATv2 attention mechanism to calculate the attention coefficients between adjacent nodes and aggregates neighbor node information for message passing through multi-head attention (8 attention heads): First, based on the mapping from meta-actions to meta-action structure types in the S2 mapping relationship, candidate structure type combinations are determined for each meta-action as candidate nodes in the graph; then, based on the mapping from meta-action structure types to specific structured design schemes, combined with current performance parameter constraints, a conditional selection mechanism selects a suitable structured design scheme from the candidate schemes for each meta-action. The layout generation section takes the spatial bounding box parameters and assembly interface codes of each substructure as input. It uses a layout planning network based on a self-attention mechanism to calculate the positioning coordinates and assembly docking relationships of each meta-action structure in the overall coordinate system, outputting a complete scheme containing the meta-action structures and their layout descriptions. The mapping reference data for this layer comes from the three-layer mapping relationship constructed by S2.

[0035] The third layer is the comparison and evaluation layer, employing a comparison and evaluation network based on a Siamese architecture. This layer uses the meta-action structure and layout scheme generated in the second layer as the input to be evaluated, and the corresponding structured design scheme and typical layout parameters in the S2 mapping relationship as the reference benchmark. The network uses two weighted encoding branches (Siamese structure): the branch to be evaluated performs graph convolutional encoding on the scheme to be evaluated to extract structural and layout feature vectors, and the reference branch performs the same encoding on the S2 mapping reference scheme. The feature vectors output from the two branches are input to the comparison layer, and the difference score is calculated jointly using cosine similarity and Euclidean distance. The comparison and evaluation includes two sub-modules: the structural rationality evaluation module calculates the adaptation score of the performance parameters of each link along the performance transmission link from power source to power transmission to end-execution, based on the rated parameter adaptation constraints of each link in the S2 mapping relationship; the layout rationality evaluation module calculates the space utilization score of each substructure combined bounding box and the compatibility score of the assembly interface between adjacent substructures based on the typical spatial layout parameters corresponding to each structural scheme in the S2 mapping relationship and the description of the assembly relationship between substructures. The scores of the two sub-modules are weighted and merged to output a comprehensive evaluation score. When the score is lower than the reasonableness threshold (the threshold is set to 80 points), an evaluation residual signal is generated.

[0036] The cascaded recommendation model employs a hierarchical conditional propagation and feedback correction mechanism: The three-layer network collaborates through hierarchical conditional propagation and feedback correction. The meta-action components and their temporal information vectors output from the first-layer BERT cross-attention network serve as conditional inputs, injecting them into the initial node features of the second-layer GAT conditional graph generation network. This drives the second layer to generate structure and layout schemes sequentially according to the meta-action order. The scheme encoding vectors generated by the second layer serve as conditional inputs, injecting them into the third-layer Siamese comparison evaluation network. This drives the third layer to perform rationality evaluation. After the third layer completes its evaluation, it generates an evaluation residual signal. This residual signal is preferentially fed back to the attention coefficient matrix of the second-layer GAT network, triggering iterative structure and layout corrections through gradient backpropagation. When the correction iterations still do not meet the evaluation threshold (the maximum number of iterations is set to 5), the correction residual is further fed back to the attention weight matrix of the first-layer BERT cross-attention network, triggering adjustments to the meta-action components, forming a three-level closed-loop collaboration of composition, structure, and evaluation.

[0037] Training steps for a cascaded recommendation model: The training of the cascaded recommendation model adopts a phased strategy: Phase 1: Independent pre-training of each network layer.

[0038] The first layer of pre-training: Using the first layer mapping data (functional / performance requirements → meta-action set) in the S2 mapping relationship as training samples, the functional requirement text and performance requirement parameters are concatenated as input, and the corresponding meta-action identifier set and its temporal order are used as labels. The BERT encoder loads the Chinese BERT-base pre-trained weights and freezes the first 6 layers, only fine-tuning the last 6 layers, the cross-attention module, and the process logic sorting module. The cross-entropy loss function is used, and the Adam optimizer is employed (learning rate 1e-4, β1=0.9, β2=0.999), with a batch size of 32, and training for 50 epochs. This enables the network to learn to match and sort the meta-actions in the standard meta-action flow according to the requirement input.

[0039] The second layer of pre-training uses the second and third layer mapping data (meta-action → structure type combination → specific structure scheme) in the S2 mapping relationship as training samples. Meta-action identifiers and performance parameter constraints are used as input, while the corresponding structured design schemes and layout parameters are used as labels. The GAT network employs 8 attention heads, each with 64 dimensions, outputting 256-dimensional node features. The self-attention layout planning network uses a 6-layer Transformer encoder structure. A weighted conditional generation loss function (structure generation loss + layout planning loss, weight ratio 3:1) is used, along with the Adam optimizer (learning rate 1e-4), a batch size of 32, and 50 training iterations, enabling the network to learn to generate suitable structure schemes and layouts for each meta-action.

[0040] The third layer of pre-training: Using the complete three-layer mapping data in the S2 mapping relationship as a reference, manually labeled structural rationality scores and layout rationality scores (scoring range 0-100 points) are used as labels. Both encoding branches of the Siamese contrast evaluation network employ a 3-layer GCN (Graph Convolutional Network) encoder, outputting a 128-dimensional feature vector. The mean squared error (MSE) loss function is used, along with the Adam optimizer (learning rate 1e-4), a batch size of 32, and training for 50 epochs, enabling the network to learn to evaluate the rationality of generated schemes.

[0041] Phase Two: Cascaded Joint Training The three-layer network is cascaded into a complete recommendation model. The training samples (approximately 800 sets) are the complete mapping link of function / performance requirement-meta-action in the S2 mapping relationship, divided into training, validation, and test sets in a 7:1:2 ratio. During training, the input function / performance requirement is processed by the first-layer BERT cross-attention network to generate meta-action components, the second-layer GAT network to generate the structure and layout scheme, and the third-layer Siamese evaluation network to compare and evaluate against a benchmark. The total cascaded loss function is then calculated. The total cascaded loss L_total = L_layer1 + L_layer2 + L_layer3 + λ·L_feedback Wherein, L_layer1 is the first layer matching loss (cross-entropy between meta-action components and the reference set); L_layer2 is the second layer generation loss (reconstruction error between generated structure and reference scheme, including structure encoding reconstruction error and layout coordinate regression error); L_layer3 is the third layer evaluation loss (mean square error between evaluation score and reference score); L_feedback is the feedback correction convergence loss (convergence rate penalty term after feedback correction iteration of evaluation residual signal), with a weight coefficient λ of 0.1 to encourage the model to converge to above the reasonableness threshold within a finite number of iterations.

[0042] The joint training employs a strategy combining teacher forcing and curriculum learning: In the early stages of training (the first 20 rounds), the output of the first layer is fixed as the label value (referencing the meta-action set), ensuring a stable conditional input for the second layer; from rounds 21 to 50, the actual output of the first layer is gradually replaced with the label value as the conditional input for the second layer, achieving end-to-end joint optimization. A feedback correction mechanism is activated synchronously during the joint training phase: the evaluation residual signal of the third layer is fed back to the second and first layers via gradient backpropagation along the cascade path. Each layer corrects its attention weight matrix based on the residual signal, and the total cascade loss converges through multiple iterations. The joint training consists of 100 rounds, using the Adam optimizer (learning rate 5e-5), a batch size of 16, cosine annealing scheduling, a minimum learning rate of 1e-6, and an NVIDIA A100 GPU. Early stopping is triggered when the total loss on the validation set no longer decreases, the optimal model weights are saved, and the training of the cascaded recommendation model is completed.

[0043] The inference process of the cascaded recommendation model, taking the overall design of this embodiment as an example: Receive functional / performance requirements input: The functional requirement is "automatic milling machine for aluminum alloy shell parts", and the performance requirements are "milling accuracy ≤ ±0.02mm, gripping force ≥10N, clamping force ≥50N, production cycle ≤60s / piece, and overall machine dimensions ≤1500×1200×800mm".

[0044] The first layer of reasoning: The BERT encoder encodes the functional requirement text to generate a requirement semantic vector. The cross-attention module calculates the cosine similarity between the requirement semantic vector and the functional description vector of each action type in the standard meta-action flow, and matches it to the action type subset {main material loading, positioning and clamping, processing, unloading, and detection}. The second layer of cross-attention performs performance constraint verification, retains the meta-actions {grabbing and loading, clamping and positioning, milling, pushing and unloading, and visual inspection}, and the process logic sorting module outputs the whole machine meta-action composition scheme after sorting according to the manufacturing process knowledge graph topology.

[0045] The second layer of reasoning: The GAT conditional graph generation network constructs a heterogeneous graph from the candidate structure type nodes of the five meta-actions. After 8-head GATv2 attention calculation and 3-layer message passing, it outputs the optimal structure type combination and the adapted structured design scheme for each meta-action. The self-attention layout planning network calculates the optimal layout position of each meta-action structure in the whole machine coordinate system based on the spatial bounding box parameters of each substructure and the assembly interface code. The five meta-action structures are arranged sequentially along the longitudinal direction of the whole machine, occupying a total space of approximately 1200×900×650mm.

[0046] The third layer of reasoning: The scheme features generated by the code of the branch to be evaluated in the twin comparison evaluation network are mapped to the reference scheme features by the reference branch code S2, and the difference is calculated by the comparison layer. Structural rationality evaluation: The performance transmission link verification of the five meta-action structures all pass (e.g., milling: servo motor rated power 2.2kW ≥ load requirement 1.8kW, belt drive rated torque 5N·m ≥ motor output 4.5N·m); Layout rationality evaluation: The overall assembly enclosure of 1200×900×650mm does not exceed the usable boundary of 1500×1200×800mm, and the assembly interfaces of adjacent structures are compatible. The comprehensive evaluation score is 87 points (threshold 80 points), and the recommended scheme for the whole machine is output.

[0047] If the evaluation does not meet the threshold, the evaluation residual signal is generated and fed back to the second-layer trigger structure and layout correction iteration first. If it still does not meet the threshold, it is further fed back to the first-layer trigger element action composition adjustment, forming a three-level closed-loop collaboration of composition-structure-evaluation.

[0048] Step S4: Import design requirements into the cascaded recommendation model to generate a manufacturing equipment design scheme. Specifically, import the design requirements into the cascaded recommendation model, generate the manufacturing equipment design results based on the recommended scheme, analyze the design requirements, extract functional requirement semantics and performance requirement parameters, input them into the cascaded recommendation model, and obtain the overall recommended scheme. Input the recommended scheme into the 3D generation model. The 3D generation model is an end-to-end assembly generation model based on a Graph Convolutional Network (GCN) decoder. The GCN decoder uses three layers of graph convolutional layers, with output dimensions of 256, 128, and 64 for each layer, and finally outputs the 6-DOF pose parameters (translation vector t and rotation quaternion q) and mating constraint parameters for each part. The model uses the structured encoding of the five meta-action structures in the recommended scheme as graph node inputs. According to the overall assembly sequence and mating constraints defined in the layout description, the GCN decoder sequentially decodes the features of each node into the positioning parameters of each standard and non-standard part and the mating constraint parameters between substructures in the overall assembly coordinate system. Accordingly, the standard 3D models of the power source structure, power transmission structure, and end effector structure in each element action structure are sequentially located in the assembly coordinate system using the model library index. Based on the parametric modeling specifications for non-standard parts, the SolidWorks API parametric modeling module is used to generate the 3D models of the non-standard parts, which are then located in the assembly coordinate system. Within each element action structure, the automatic assembly of the power source-power transmission-end effector substructures is completed using surface fit constraints (Mate), axis alignment constraints (Concentric), and offset constraints (Offset). Between element action structures, adjacent element action structures are docked and assembled along the overall machine process sequence using surface fit constraints and pin hole positioning constraints. Finally, a 3D digital model of the entire element action structure assembly of the automatic milling equipment for aluminum alloy housing is generated and output along with the overall machine structure specification.

[0049] The present invention provides an intelligent design method for manufacturing equipment structure based on meta-actions, which integrates standardized modeling of manufacturing equipment actions, construction of functional and performance requirements-meta-action-meta-action structure mapping relationship, generation of meta-action components and structure recommendation and layout planning based on cascade recommendation model, comparison and verification of generated schemes with historical mapping data, and automatic output of three-dimensional assembly model, thereby realizing automated intelligent design of intelligent manufacturing equipment.

[0050] Secondly, please refer to Figure 2The present invention also provides an intelligent design system for manufacturing equipment structure based on meta-actions, including a standard library module 101, a mapping data module 102, a cascaded recommendation module 103, an interaction module 104, and a three-dimensional assembly module 105; the standard library module 101, the mapping data module 102, the cascaded recommendation module 103, the interaction module 104, and the three-dimensional assembly module 105 are connected in sequence; The standard library module 101 is used to store standard meta-action processes and meta-action structure libraries; The mapping data module 102 is used to store the mapping relationship between function and performance requirements, meta-actions, and meta-action structures. The mapping relationship includes three layers of mapping data: requirement to meta-action, meta-action to structure type, and structure type to specific solution. The cascaded recommendation module 103 is used to execute the cascaded recommendation model, which includes a meta-action composition planning unit, a structure and layout generation unit, and a comparison and evaluation unit. The three units work together through hierarchical condition transmission and feedback correction. The interaction module 104 is used to receive design requirements and call the cascaded recommendation module 103 to pass the recommended solution to the three-dimensional assembly module 105. The three-dimensional assembly module 105 is used to input the recommended scheme into the three-dimensional generation model, call standard parts and generate three-dimensional models of non-standard parts, complete automatic assembly and output the design results.

[0051] The present invention discloses a manufacturing equipment structure intelligent design system based on meta-actions. By employing a standard library module 101, a mapping data module 102, a cascaded recommendation module 103, an interaction module 104, and a three-dimensional assembly module 105, the system integrates standardized modeling of manufacturing equipment actions, construction of functional and performance requirements-meta-action-meta-action structure mapping relationships, generation of meta-action components based on the cascaded recommendation model and structure recommendation and layout planning, comparison and verification of generated schemes with historical mapping data, and automatic output of three-dimensional assembly models, it can realize the automated intelligent design of intelligent manufacturing equipment.

[0052] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for intelligent structural design of manufacturing equipment based on meta-actions, characterized in that, include: S1 establishes standard meta-action processes and meta-action structure libraries for manufacturing equipment based on the characteristics of the manufacturing equipment. S2 constructs a database of mapping relationships between the design requirements, meta-actions, and meta-action structures of manufacturing equipment; S3 constructs and trains a cascaded recommendation model for manufacturing equipment structural design based on meta-actions; S4 imports design requirements into the cascaded recommendation model to generate manufacturing equipment design schemes.

2. The intelligent design method for manufacturing equipment structure based on meta-actions as described in claim 1, characterized in that, In step S1, The standard meta-action process adopts at least a two-level architecture: the first level is the action type, which includes at least main material feeding, auxiliary material feeding, positioning and clamping, processing, unloading and detection; the second level is several meta-actions under each action type, and the meta-action is the smallest complete mechanical structure unit that completes an independent action function.

3. The intelligent design method for manufacturing equipment structure based on meta-actions as described in claim 1, characterized in that, In step S1, The meta-action structure is represented by a triplet of power source, power transmission, and end effector; each meta-action structure includes a structure type identifier, key performance parameter specifications, and an association index with the meta-action identifier.

4. The intelligent design method for manufacturing equipment structure based on meta-actions as described in claim 1, characterized in that, In step S2, The design requirement-meta-action-meta-action structure-layout mapping relationship is a three-layer mapping. The first layer is the mapping from design requirements to the set of meta-actions and their order; the second layer is the mapping from meta-actions to the combination constraints of power source structure, power transmission structure and end execution structure types; the third layer is the mapping from meta-action structure type to specific design scheme and typical layout. The mapping relationship is constructed from historical design case data after structured extraction and verification.

5. The intelligent design method for manufacturing equipment structure based on meta-actions as described in claim 1, characterized in that, In step S3, The cascaded recommendation model consists of three layers: the first layer is the generation layer of the equipment's meta-action composition and sequence; the second layer is the generation layer of the meta-action structure and layout; and the third layer is the result comparison and judgment layer. The three layers cooperate through step-by-step condition transmission and feedback correction to form a three-level closed-loop collaboration of meta-action composition, structure generation, and comparison and judgment.

6. The intelligent design method for manufacturing equipment structure based on meta-actions as described in claim 5, characterized in that, In the cascaded recommendation model: the first layer is a meta-action composition and sequence generation layer based on semantic matching cross-attention network. It matches functional requirements with standard meta-action processes to filter action types, performs constraint verification according to performance requirement parameters, determines the arrangement order according to manufacturing process logic, generates meta-action composition and sequence scheme, and injects the output meta-action composition and sequence into the second layer driving structure and layout generation. The second layer is a conditional graph generation layer based on graph attention network. According to the mapping relationship in step S2, it determines the three-element structure form for each element action, and plans the layout according to the substructure space parameters and assembly interface to generate the structure and layout scheme. The output scheme is injected into the third layer to drive the rationality evaluation. The third layer is a comparison and evaluation layer based on a twin architecture. It compares the generated scheme with the mapping relationship of step S2, verifies the adaptability of the performance transmission link and the compatibility of the spatial layout, outputs the recommended scheme when the threshold is met, and generates an evaluation residual signal to feed back to the second layer to trigger correction when the threshold is not met. If the threshold is still not met, it feeds back to the first layer to trigger the adjustment of the meta-action composition, forming a three-level closed-loop collaboration.

7. The intelligent design method for manufacturing equipment structure based on meta-actions as described in claim 1, characterized in that, Step S4 specifically includes: The recommended solution is input into the 3D generated model. The graph decoder encodes and decodes each element action structure into positioning and mating constraint parameters in the assembly coordinate system. The standard part 3D model is called and the non-standard part 3D model is generated, thus completing the automatic assembly and generating of the 3D digital model.

8. A manufacturing equipment structure intelligent design system based on meta-actions, employing the manufacturing equipment structure intelligent design method based on meta-actions as described in any one of claims 1-7, characterized in that, It includes a standard library module, a mapping data module, a cascaded recommendation module, an interaction module, and a 3D assembly module; the standard library module, the mapping data module, the cascaded recommendation module, the interaction module, and the 3D assembly module are connected in sequence; The standard library module is used to store standard meta-action processes and meta-action structure libraries; The mapping data module is used to store the mapping relationship between function and performance requirements, meta-actions, and meta-action structures. The mapping relationship includes three layers of mapping data: requirement to meta-action, meta-action to structure type, and structure type to specific solution. The cascaded recommendation module is used to execute the cascaded recommendation model, including a meta-action composition planning unit, a structure and layout generation unit, and a comparison and evaluation unit. The three units work together through step-by-step condition transmission and feedback correction. The interaction module is used to receive design requirements and call the cascaded recommendation module to pass the recommended solution to the 3D assembly module; The three-dimensional assembly module is used to input the recommended solution into the three-dimensional generation model, call standard parts and generate three-dimensional models of non-standard parts, complete automatic assembly and output the design results.