AI Reinforcement Learning for Distributed Production Routing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing industrial factory designs face scalability issues and inflexibility due to static layouts that require redesign and retooling for product changes, and centralized dynamic solutions struggle with computational complexity and inefficiency when dealing with large numbers of production units.
Innovation Solution
A method and system utilizing AI with reinforcement learning capabilities in distributed cyber-physical production units that make independent decisions based on feedback, optimizing production processes without requiring global knowledge of the factory layout, allowing for dynamic product routing and adaptation to changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a static factory layout is used with units placed and connected to facilitate execution of one product's Bill of Process, then the production process is optimized for that specific product, but changing the targeted product requires costly redesign, re-setup, and retooling
Solution Approach 1:
The patent implements dynamic product routing where the path a product takes through the factory is not fixed but determined at runtime based on current unit status, product type, and learned optimization patterns. Each CPPU can dynamically decide whether to process a product or transfer it to another unit, enabling the system to adapt to different product configurations without physical reconfiguration of the factory layout
Solution Approach 2:
The factory is divided into independent cyber-physical production units (CPPUs) that can autonomously make decisions about product processing. Each unit operates semi-autonomously and can be dynamically selected or bypassed based on real-time conditions, allowing the system to reconfigure production paths software-defined rather than requiring physical reconfiguration
2Adaptability or versatility
If centralized algorithms are used to dynamically determine product transportation through the factory at runtime, then flexibility is provided to produce different products with minimal intervention, but the algorithms require global knowledge of factory layout and state which becomes computationally intractable when the shopfloor size and unit interconnections increase
Solution Approach 1:
The centralized control problem is segmented into distributed decision-making at each CPPU. Instead of one central algorithm needing global knowledge, each unit independently makes local decisions based on its own state and simple communication with neighboring units. This distributes the computational burden and eliminates the need for any single entity to have complete global knowledge of the entire factory state
Solution Approach 2:
Each CPPU makes routing decisions based on local information about its own state, the product being processed, and immediate neighbors, rather than requiring global factory state knowledge. The reinforcement learning at each unit learns optimal local policies that collectively achieve system-wide optimization without centralized coordination
3Productivity
If production units are arranged in a layout tailored for a specific product to optimize targeted KPIs, then excellent performance is achieved for that product, but the solution loses efficiency when product changes occur and requires production down-time for redesign
Solution Approach 1:
The system uses dynamic routing decisions made at runtime rather than static pre-configured paths. When product changes occur, the reinforcement learning policies at each CPPU can adapt to new product types and routing requirements without requiring physical reconfiguration or production stoppage for redesign, maintaining continuous optimization across product transitions
Data Source
Figure 1
Figure 2
Figure 3~5
AI summary
Method for producing a product, wherein skills of a bill of process are executed by production units (166). The production units (166) are associated with an AI (102) with RL capabilities and are adapted to only handle the product within a respective subgroup. The method comprises executing (S110), in a first production cycle, a first process skill by a first production unit to a first product, transferring (S120), the first product from the first production unit to a second production unit within the respective subgroup and executing (S130) a second process skill by the second production unit to the first product. Then, the first production unit receives (S140) feedback information (106) based on at least one key performance indicator. By using the feedback information (106) and by using the AI (102) of the first production unit the next production units (166) from the respective subgroup is selected (S160).