Industrial RL Retraining With White-Spot Guided Simulation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Reinforcement learning agents face challenges in industrial process systems due to high sample complexity, high costs of high-fidelity simulators, suboptimal performance, and the need for safe exploration, particularly in complex and dynamic environments with high-dimensional state and action spaces.

Innovation Solution

A method and system for training reinforcement learning agents using a combination of historical closed-loop data and low-fidelity simulators, with safety verifiers and guarantors, to ensure safe and efficient training and deployment, leveraging low-fidelity simulators for sample-efficient exploration and adaptation to changing processes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If reinforcement learning agents are trained using high-fidelity process simulators, then training performance and accuracy are improved, but development time, monetary cost, and computational cost increase significantly

Engineering Contradiction:
Improvetraining accuracyVSAvoiddevelopment time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent uses low-fidelity simulators as simplified copies of the actual industrial process system for training purposes. These low-fidelity models capture essential process dynamics without the computational complexity of high-fidelity simulators, enabling faster and cheaper training while still providing sufficient learning signal for the reinforcement learning agent

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The training process is segmented into multiple phases: initial training with low-fidelity simulators to achieve basic competence, followed by fine-tuning with high-fidelity simulators or real system data to improve accuracy. This segmentation allows the system to benefit from both the speed of low-fidelity models and the accuracy of high-fidelity models without paying the full cost of high-fidelity training throughout

Inventive Principle:
Principle #1Segmentation

2Productivity

If reinforcement learning agents explore a larger region of the industrial process system during training, then learning effectiveness is improved, but safety risks and potential for costly failures increase

Engineering Contradiction:
Improvelearning effectivenessVSAvoidsafety risks
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The patent introduces low-fidelity simulators as intermediary environments that allow the reinforcement learning agent to explore extensively and learn from failures without affecting the actual industrial process system. The agent first learns in this safe intermediary environment before being deployed to the real system, where exploration is constrained

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system performs preliminary exploration and learning in low-fidelity simulators before deploying the agent to the actual industrial process. This preliminary action in a safe environment prepares the agent for real-world deployment while preventing harmful exploration of the actual system

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If reinforcement learning agents are trained from scratch in the actual operational environment, then adaptation to real process dynamics is improved, but safety concerns and operational disruptions increase

Engineering Contradiction:
Improveadaptation to real process dynamicsVSAvoidoperational safety
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent uses historical operational data as a copy of real process behavior to train the reinforcement learning agent. This allows the agent to learn from actual process dynamics without requiring extensive exploration in the real system, thereby maintaining operational safety while achieving good adaptation

Inventive Principle:
Principle #26Copying

4Loss of information

If reinforcement learning algorithms interact numerous times with the process to learn effective policies, then learning completeness is improved, but sample efficiency deteriorates

Engineering Contradiction:
Improvelearning completenessVSAvoidnumber of samples required
Core Design Contradiction:
Loss of informationVSQuantity of substance

Solution Approach 1:

The patent uses low-fidelity simulators as computational copies that can be interacted with extensively without consuming real process resources. These synthetic environments provide abundant training samples that would be expensive or impossible to obtain from the actual industrial process system

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system performs preliminary learning in low-cost synthetic environments before deploying to the real system. This preliminary action accumulates sufficient learning completeness in advance, reducing the number of samples needed from the actual process

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4651024A1Method for training a reinforcement learning agent for an industrial process system and system for training a reinforcement learning agent for an industrial process system
Publication Date: 2025.11.19 ABB (SCHWEIZ) AG
  • EP4651024A1 patent drawingFigure 1
  • EP4651024A1 patent drawingFigure 2
  • EP4651024A1 patent drawingFigure 3

AI summary

A method for training a reinforcement learning, RL, agent for an industrial process system is provided. The method includes, training the RL agent with plant historical data of the industrial process system, and retraining the RL agent using plant historical data and a low-fidelity simulator of the industrial process system. The retraining the RL agent includes analyzing the plant historical data to identify white spots as regions of process states and dynamic behavior that have not been explored during the training the RL agent, and retraining the RL agent by prioritized exploration with information gained from the white spots and with simulated data provided by simulating the industrial process system with the low-fidelity simulator.