Gear Train Topology Synthesis Using Deep Reinforcement Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine design methods face challenges in deriving complex mechanical structures efficiently, as traditional artificial intelligence models lack accuracy and are limited to specific forms, making it difficult to explore and synthesize optimal gear train topologies.
Innovation Solution
A method and device utilizing a Markov decision process (MDP) for deep reinforcement learning, specifically through a deep Q-network (DQN), to generate and synthesize gear train topologies by defining actions, states, and rewards, converting structure synthesis into a tree search process, and applying convolutional layers and Rectified Linear Unit (ReLU) activation functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If traditional AI models are used for machine design, then the design process can be automated, but the accuracy of mechanical structures based on physical phenomena cannot be achieved
Solution Approach 1:
The patent introduces a graph-based intermediate representation that bridges traditional AI automation and physical accuracy requirements. The graph structure serves as a mediator that encodes mechanical components and their connections in a way that preserves physical relationships while enabling automated processing through reinforcement learning algorithms.
Solution Approach 2:
The patent replaces traditional mechanical design methodologies with a reinforcement learning-based automated system. The DQN agent learns optimal mechanical structure configurations through iterative exploration of the design space, substituting human expert mechanical knowledge with AI-driven decision-making while maintaining physical accuracy through proper graph representation.
2Adaptability or versatility
If all possible structures are explored to find desired structure, then comprehensive design options can be found, but the number of cases becomes too large to be practical
Solution Approach 1:
The patent segments the complex design space into manageable graph-based representations where components and connections are broken down into discrete nodes and edges. This segmentation allows the reinforcement learning agent to explore the design space systematically through incremental graph modifications rather than evaluating all possible structures at once.
Solution Approach 2:
The patent employs dynamic graph modification where the graph structure evolves iteratively through the reinforcement learning process. The DQN agent dynamically adds, removes, or modifies graph elements based on learned policies, enabling adaptive exploration of the design space that focuses computational resources on promising regions rather than uniformly exploring all cases.
3Productivity
If deep reinforcement learning with DQN is used for gear train topology synthesis, then flexible and efficient topology generation is achieved, but the computational complexity of defining MDP and training increases
Solution Approach 1:
The patent performs preliminary action by pre-defining the graph structure, nodes, and edges that represent mechanical components and their connections before the reinforcement learning training begins. This preliminary setup establishes the design space boundaries and constraints, allowing the DQN agent to focus learning on optimal configuration selection rather than learning basic representation from scratch, thereby reducing overall computational complexity.
Data Source
AI summary
A method performed by an apparatus may comprise defining a Markov decision process (MDP) for use in deep reinforcement learning associated with gear train topology, representing the gear train topology based on the MDP, generating the gear train topology through deep reinforcement learning of a deep Q-network (DQN), and constructing a gear train based on the generated gear train topology.