Facility Control Model Correction Using Real-Simulation Drift

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing control systems for facilities like distillation apparatuses face challenges in enhancing controllability due to strong mutual interference, long time constants, and non-linear operations, making it difficult to achieve quality assurance, energy savings, and yield enhancement, especially when relying on manual operations that depend on experience and intuition.

Innovation Solution

A control apparatus that uses machine learning models to correct operation model outputs based on indices calculated from real data, comparing simulated and actual operations, and incorporating reinforcement learning to optimize control actions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If manual operations are used for facilities with strong mutual interference and long time constants, then operator experience and intuition can be leveraged, but controllability and response speed are insufficient

Engineering Contradiction:
ImprovecontrollabilityVSAvoidresponse speed
Core Design Contradiction:
Ease of operationVSSpeed

Solution Approach 1:

The patent replaces manual mechanical control operations with an automated control system that uses machine learning models. The operation model trained by reinforcement learning automatically determines control actions, substituting human operators' manual adjustments with algorithm-driven control that responds rapidly to facility state changes without being limited by human reaction time.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The control system performs self-learning and self-optimization through reinforcement learning. The operation model continuously learns from simulated operations and real facility data, automatically improving its control strategies without requiring manual reprogramming or external intervention, enabling the system to adapt to changing conditions autonomously.

Inventive Principle:
Principle #25Self-service

2Productivity

If simulation data is used to train operation models, then training can be performed without affecting actual facility operations, but differences between simulated and real behavior reduce control accuracy

Engineering Contradiction:
Improvetraining efficiencyVSAvoidcontrol accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the operation model is trained on simulation data first, then its predictions are compared against actual facility measurements. The difference between simulated and real behavior serves as feedback to fine-tune the model, creating a closed-loop training process that progressively improves accuracy while maintaining the benefit of simulation-based pre-training.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent performs preliminary training of the operation model using simulation data before deploying it to control the actual facility. This preliminary action allows the model to learn basic control strategies in a risk-free simulated environment, and then these pre-learned strategies are refined using real facility data, combining the advantages of both simulation and real-world training.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If reinforcement learning is used to train operation models, then optimal control actions can be learned, but the learning process requires extensive simulation time and computational resources

Engineering Contradiction:
Improvecontrol optimizationVSAvoidlearning time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent performs preliminary training of the operation model using simulation data before deploying it to control the actual facility. This preliminary action allows the model to learn basic control strategies in a risk-free simulated environment, and then these pre-learned strategies are refined using real facility data, combining the advantages of both simulation and real-world training.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses a hybrid training approach where the operation model is trained partially on simulation data and partially on real facility data. Rather than requiring extensive training solely on real data (which would take too long), the model receives sufficient simulation training to learn basic patterns, then uses limited real data to fine-tune, achieving good performance without excessive learning time.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4270118A1Control apparatus, control method, and control program
Publication Date: 2023.11.01 YOKOGAWA ELECTRIC CORP
  • EP4270118A1 patent drawingFigure 1
  • EP4270118A1 patent drawingFigure 2
  • EP4270118A1 patent drawingFigure 3

AI summary

There is provided a control apparatus including: a model output acquisition unit configured to acquire an operation model output which is output according to inputting state data indicating a state of a facility, to an operation model trained by machine learning to output an action in accordance with the state of the facility by using simulation data from a simulator that simulates an operation in the facility; an index acquisition unit configured to acquire an index which is calculated by using real data from the facility and which is for monitoring a difference between a behavior of the simulator and an actual operation in the facility; a correction unit configured to correct the operation model output based on the index; and a control unit configured to output a manipulated variable for controlling a control target provided in the facility, according to the corrected operation model output.