Neural Network Controller Training From Historical Process Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Industrial processes face challenges in maintaining specific physical conditions due to complex connections between manipulated and disturbance variables and controlled variables, making it difficult and expensive to program effective controllers for optimal operation.
Innovation Solution
A method involving the training of neural networks to predict and control continuous processes using off-line historical data, where a predictor neural network is trained first to estimate controlled variables, and then a controller neural network is trained using the predictor to adjust and optimize the process, allowing for efficient operation without real-time online information or knowledge of the existing controller's policy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional programming methods are used to create controllers for industrial processes, then the controller can be programmed with explicit control logic, but the development time and cost increase substantially requiring substantial human expert involvement
Solution Approach 1:
The system enables self-service by allowing the controller to automatically learn control policies from historical process data without requiring substantial human expert involvement. The neural network trainer autonomously trains predictor and controller neural networks using offline historical data, eliminating the need for manual programming and expert knowledge input while achieving effective control.
2Reliability
If model predictive control methods are used to handle complex variable connections, then control performance improves for complex processes, but the complexity of system modeling and implementation increases
Solution Approach 1:
The patent replaces traditional mechanical modeling approaches with neural network-based learning. Instead of requiring explicit mathematical models of the complex process dynamics, the system uses neural networks to automatically learn the relationships between manipulated variables, disturbance variables, and controlled variables from historical data, thereby reducing modeling complexity while maintaining control performance.
Solution Approach 2:
The system changes the approach from fixed mathematical model parameters to adaptive neural network parameters that are automatically trained on historical data. The neural networks learn optimal control policies by adjusting their internal parameters (weights and biases) based on historical process data, eliminating the need for complex manual system modeling.
3Productivity
If offline historical data is used for training instead of real-time online data, then the controller training can be performed without interrupting process operation and without requiring real-time data, but the training must rely on historical patterns which may not capture current process conditions
Solution Approach 1:
The system performs preliminary action by training the controller offline using historical data before deployment. The neural network trainer completes the entire training process using stored historical process data without requiring real-time data or interrupting current process operation. This preliminary training equips the controller with learned policies that can be directly applied when deployed.
Solution Approach 2:
The system uses copying by training on historical data that replicates normal process conditions. The neural networks learn from copies of past process data stored in databases, capturing the essential patterns and relationships without needing access to real-time process data during training. This allows the controller to be trained efficiently using historical patterns.
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.