Stone Layout Using PPO-Clip for Irregular Part Nesting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing stone arrangement methods face challenges in achieving optimal spatial layout with irregular parts, particularly concave or hole-containing parts, and heuristic algorithms often result in local optimal solutions with inaccurate results due to incorrect parameter settings.
Innovation Solution
An automatic stone arrangement method using image and deep reinforcement learning, employing an Actor-Critic convolutional network model with a PPO-Clip algorithm, which eliminates the need for critical polygon calculation and allows continuous optimization through trial-and-error learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If critical polygon calculation is used for irregular parts, then part positioning can be achieved, but the calculation becomes very complex and cannot handle concave or hole-containing parts
Solution Approach 1:
The patent extracts the complex critical polygon calculation step from the arrangement process and replaces it with a simplified bounding box-based overlap detection method. This allows the system to handle irregular parts including concave and hole-containing parts without the computational complexity of critical polygon calculations.
Solution Approach 2:
The patent substitutes the geometric calculation mechanism (critical polygon calculation) with a machine learning-based mechanism (deep reinforcement learning model). The model learns to directly predict optimal arrangement positions and orientations, replacing the need for complex geometric computations and enabling handling of arbitrary part geometries.
2Productivity
If heuristic algorithms are used for sequencing, then arrangement order can be optimized, but the algorithm may fall into local optimal solution and cannot find global optimal solution
Solution Approach 1:
The patent changes the fundamental parameters of the optimization approach by transitioning from heuristic algorithms with manually tuned parameters to deep reinforcement learning. The learning rate, discount factor, and network architecture parameters are automatically optimized through training, enabling the system to escape local optima and find better global solutions.
Solution Approach 2:
The reinforcement learning model learns arrangement strategies through self-play and trial-and-error in a simulated environment. It automatically improves its own performance without human intervention, continuously optimizing arrangement sequences to achieve better material utilization and avoid local optima.
3Ease of operation
If heuristic algorithms with multiple parameters are used, then arrangement sequencing can be performed, but incorrect parameter settings lead to inaccurate results or low computational efficiency
Solution Approach 1:
The patent removes the need for manual parameter tuning by extracting the parameter optimization process into the training phase of the reinforcement learning model. During deployment, the model uses learned parameters that have been automatically optimized, eliminating the trade-off between ease of operation and accuracy.
4Ease of operation
If traditional manual arrangement methods are used, then arrangement can be performed, but efficiency is low and arrangement accuracy is low
Solution Approach 1:
The patent replaces manual arrangement operations with an automated deep reinforcement learning system. The model processes part geometry images and directly outputs optimal arrangement positions and orientations, achieving both high efficiency through automated processing and high accuracy through learned optimization strategies.
Data Source
AI summary
An automatic stone arrangement method including: (1) a stone arrangement simulation environment is built; an action space and a return function are designed based on rules that location of parts must be inside raw materials, the parts must not overlap with each other, and utilization of the raw materials is maximized; (2) a PPO-Clip algorithm model including an Actor-Critic network is constructed; both Actor neural network and Critic neural network are trained at the same time in an Actor-Critic algorithm; (3) based on a new production order of stone parts, images, which are obtained through preprocessing a part image and a raw material image corresponding to the new production order, are used as an input to a trained PPO-Clip algorithm model; and action instruction for arranging the parts are outputted to complete an automatic arrangement task under the new production order.