.png){kind=icon}
.png){kind=icon}
Return
Automatic modification and repair of three-dimensional models in intelligent manufacturing based on reinforcement learning
.png){kind=icon}
PRE
AI
T
L
DOI：10.1051/meca/2026005.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In intelligent manufacturing, the complex defects of three-dimensional models (such as holes, self-intersections, non-manifold edges/vertices, topological breaks, surface wrinkles, and normal anomalies) lead to low repair efficiency and poor manufacturing adaptability. This paper constructs an automatic repair method based on the PPO (proximal policy optimization) algorithm. Through the trust-domain constraint in PPO, the policy update divergence is avoided. Combined with a GNN (graph neural network), multi-scale geometric feature extraction, environmental perception, and dynamic decision-making are realized, thereby improving repair automation levels and manufacturing feasibility. The three-dimensional grid structure is topologically modeled using a GNN; local geometric features are extracted; and defect areas are identified. The reinforcement learning framework is applied to model the repair process as a state-action sequence decision problem, and the PPO algorithm is used to optimize the policy function and generate adaptive repair operations. The self-iteration mechanism is designed to realize multiple rounds of repair optimization and enhance the system's robust processing capabilities for different types of defects. The experimental results show that when PPO+GNN is used to process typical complex defects such as holes, self-intersection overlaps, non-manifold edges and vertices, topological fractures, surface wrinkles, and normal anomalies, the boundary closure rate is between 0.7 and 0.9; the surface smoothness error is between 0.05 and 0.09; and there is high repair accuracy and geometric consistency. The printing success rate is between 0.85 and 0.95; the material utilization rate is between 0.77 and 0.82; and the manufacturing adaptability is good. The average number of convergence steps is 15; the total modification time is 5.8 s; the peak memory usage is 2.1 GB; and the repair efficiency and resource consumption are well balanced. The experimental data verify the effectiveness of the research presented in this paper on the automatic modification and repair of intelligent manufacturing 3D models.
Keywords:
Reinforcement learning
graph neural network
3D model repair
intelligent manufacturing
proximal policy optimization
geometric consistency
Journal
M
IF:
1.2
Papers:
23
Citations:
0
Organization
Cited Papers