3D Assembly Planning With Graph Neural Sequence Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current assembly planning methods are inefficient and costly due to the complexity of finding an optimal assembly sequence for products with many components, often requiring manual trial and error or specific algorithms that focus on validation rather than generation, and lack of available assembly-specific data.
Innovation Solution
A system utilizing a recurrent graph neural network that generates a feasible assembly plan by creating an assembly graph from 3D CAD models, updating node embeddings, and predicting sequential connections between components, incorporating modules like assembly descriptors, graph encoders, and contact generation models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual trial and error methods are used to develop assembly plans, then complete exploration of assembly sequences is possible, but the process becomes time-consuming and very challenging
Solution Approach 1:
The patent replaces manual trial-and-error mechanical planning with an automated neural network system that processes 3D CAD models and generates assembly sequences algorithmically, eliminating human time investment while maintaining comprehensive sequence exploration
Solution Approach 2:
The system creates a digital replica (assembly graph) of the physical product structure from 3D CAD models, allowing virtual experimentation and validation of assembly sequences without physical prototyping time
2Extent of automation
If algorithmic approaches with manually-specified cost functions are used to validate assembly sequences, then validation can be automated, but the algorithms cannot generate assembly sequences from product information
Solution Approach 1:
The neural network system performs both generation and validation functions in one unified model, eliminating the need for separate manual cost function specification and enabling the system to both create and verify assembly sequences automatically
Solution Approach 2:
The system uses its own predicted assembly sequences as input for self-validation, where the network's output feeds back into the validation process without requiring external manual intervention or separate validation algorithms
3Manufacturing precision
If the search space from within the optimal plan is searched to find the optimal assembly plan, then the optimal sequence can be found, but the process becomes very costly due to the exponential search space
Solution Approach 1:
The patent transforms the discrete combinatorial search problem into a continuous probability space where the neural network outputs likelihood scores for different assembly operations, allowing optimization without exhaustive search of the exponential state space
Solution Approach 2:
The system performs preliminary encoding of the entire product structure into an assembly graph before search begins, pre-organizing all possible connections and constraints so that the subsequent optimization process works with structured data rather than raw component lists
4Manufacturing precision
If complete details of individual parts including geometry, alignment, and movement constraints are required to determine assembly sequence, then accurate assembly planning is possible, but assembly-specific data may not be available to develop the assembly plan
Solution Approach 1:
The system performs preliminary extraction of geometric and constraint information from standard 3D CAD models before assembly planning begins, pre-processing the available data into an assembly graph structure that captures all necessary relationships without requiring additional assembly-specific input data
Solution Approach 2:
The assembly graph serves as an intermediary data structure that translates standard CAD geometry into assembly-planning-relevant information, bridging the gap between available CAD data and the detailed constraints needed for accurate sequencing
Data Source
Figure 1
Figure 2A~2B
Figure 2C~2D
AI summary
Aspects of the present disclosure provide systems, methods, and computer-readable storage media that support mechanisms for generating a feasible assembly plan for a product based on data analytics. In aspects, information on components of a product is obtained from one or more product models (e.g., a three-dimensional (3D) computer aided design (CAD) model) that define the individual components of the product. The individual component information may be used to represent the assembly of the product as an assembly graph, in which each node of the assembly graph represents one of the components of the product to be assembled. The assembly graph is passed through a set of data analytics modules to generate the feasible assembly plan, or assembly sequence, as a series of sequential contact predictions, wherein each contact prediction identifies a component to be connected to one or more other components of the product.