Graph Neural Network for Industrial Control Adaptability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Reinforcement learning in industrial control systems using fully connected neural networks often fails to accurately follow the actual state of control targets and incurs excessive processing loads, requiring large-scale relearning for design changes.
Innovation Solution
A data processing apparatus that acquires graph-structured data to generate a neural network with a reinforcement learning unit, which derives parameters to optimize the feature quantities of output layers, using a graph attention network to set coefficients and propagate feature quantities, thereby improving control accuracy and reducing processing load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a fully connected neural network is used for reinforcement learning in industrial control systems, then the network can process various control targets, but the processing load becomes excessive and control accuracy decreases
Solution Approach 1:
The patent segments the fully connected neural network into multiple localized neural networks, each responsible for specific control targets or regions. This segmentation reduces the processing load on each network while maintaining overall system versatility through the coordinated operation of multiple specialized networks.
Solution Approach 2:
The patent implements local quality by creating neural networks with specialized structures optimized for specific control targets rather than using a generic fully connected network for all targets. Each localized network has tailored connectivity and parameters that match the characteristics of its assigned control target, improving both accuracy and efficiency.
2Adaptability or versatility
If a fully connected neural network is used for reinforcement learning, then the network can be applied to various control targets, but the control accuracy fails to properly follow the actual state of control targets
Solution Approach 1:
The patent divides the control system into multiple localized neural networks, each specialized for specific control targets. This segmentation allows each network to focus on and accurately model the specific dynamics and characteristics of its assigned target, thereby improving measurement precision and control accuracy while maintaining versatility through the collective capability of all networks.
Solution Approach 2:
The patent applies local quality by designing each neural network with structure and parameters optimized for its specific control target. This localized optimization enables each network to capture the precise relationships and dynamics of its target system, resulting in higher control accuracy compared to a generic fully connected network.
3Extent of automation
If reinforcement learning is applied to social infrastructure systems, then automated control can be achieved, but large-scale relearning is required for design changes
Solution Approach 1:
The patent segments the reinforcement learning system into multiple localized networks, each handling specific control targets. When design changes occur, only the affected localized networks need to be relearned or updated, rather than requiring large-scale relearning of the entire system. This segmentation significantly improves adaptability and reduces the burden of relearning.
Solution Approach 2:
The patent implements dynamics by enabling the system to adapt and reconfigure itself in response to design changes. The modular structure allows dynamic updates to specific networks without disrupting the entire system, and the reinforcement learning mechanism continuously adapts parameters to maintain optimal automated control performance.
Data Source
AI summary
A data processing apparatus according to an embodiment includes a data acquisition unit, a setting unit, and a reinforcement learning unit. The data acquisition unit acquires graph-structured data describing a connection relation between nodes. The setting unit sets a first network representing the graph-structured data acquired by the data acquisition unit. The reinforcement learning unit derives a parameter of the first network such that a feature quantity of an output layer of an evaluation target node in the first network approaches a reward and a feature quantity of an output layer of an operation node becomes a feature quantity causing the feature quantity of the output layer of the evaluation target node to approach the reward.