Neural Network Controller Training From Historical Process Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Programming a controller for industrial processes is a long and expensive task due to the complexity of connections between manipulated and disturbance variables and controlled variables, requiring substantial human expert involvement.
Innovation Solution
A method involving the training of neural networks to predict and control industrial processes using off-line historical data, where a first neural network is trained as a predictor and a second neural network is trained as a controller, allowing for offline data usage without relying on existing controller policies, and employing a reward function to adjust the controller based on target variables and economic factors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If model predictive control (MPC) methods are used to control industrial processes, then the ability to handle complex connections between variables is improved, but the programming time and cost increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training neural network controllers offline using historical process data before actual deployment. The controller is trained in advance on a dataset containing sequences of process variables, allowing it to learn optimal control policies beforehand. This eliminates the need for time-consuming programming and commissioning during actual process operation, while still achieving high adaptability to complex process dynamics.
Solution Approach 2:
The patent uses copying by training the neural network controller to replicate the behavior of an ideal controller using historical data. The network learns to copy optimal control decisions from past operations, storing this knowledge in its weights and biases. This allows the controller to inherit expertise from historical operations without requiring manual programming of control logic for each new process.
2Adaptability or versatility
If model predictive control (MPC) methods are used to control industrial processes, then the ability to handle complex connections between variables is improved, but the cost and human expert involvement increase substantially
Solution Approach 1:
The patent applies self-service by enabling the controller to train itself autonomously using historical process data without requiring continuous human expert intervention. The neural network automatically learns control policies by processing past operational data, adjusting its internal parameters through backpropagation. This self-training capability dramatically reduces dependency on expensive human experts while maintaining high adaptability to complex process relationships.
Solution Approach 2:
The patent substitutes mechanical expert knowledge with an automated neural network system. Instead of relying on human experts to manually program control logic based on their experience, the system automatically learns control strategies from historical data. This replacement of human expertise with an automated learning system reduces both cost and human involvement while achieving comparable or superior adaptability to complex process dynamics.
3Ease of manufacture
If neural networks are trained using offline historical data without knowledge of the original controller policy, then the need for human expertise is reduced and costs are lowered, but the training data requirements and computational resources increase
Solution Approach 1:
The patent applies universality by designing a training methodology that works with generic historical process data without requiring knowledge of the specific controller policy that generated it. The neural network is trained to learn control policies from any sufficiently rich historical dataset, making the approach universally applicable across different processes and controllers. This multi-functionality allows the same training framework to reduce human expertise needs across various industrial applications while efficiently utilizing available data.
Data Source
AI summary
A method of generating a controller for a continuous process. The method includes receiving from a storage memory, off-line stored values of one or more controlled variables and one or more manipulated variables of the continuous process over a plurality of time points. The off-line stored values are used to train a first neural network to operate as a predictor of the controlled variables. Then, the method includes training a second neural network to operate as a controller of the continuous process using the first neural network after it was trained to operate as the predictor for the continuous process and employing the second neural network as a controller of the continuous process.