ML Controller Training Using Correlated Time-Window Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing control devices for complex technical systems, such as gas turbines and wind turbines, require large volumes of representative training data to optimize their operation, but often struggle with inefficient training due to insufficient data coverage of operating conditions.
Innovation Solution
A method is introduced to efficiently configure control devices by capturing and correlating temporal sequences of state and control action data within specific time windows, extracting training data that show strong correlations between control actions and system states, and using optimization methods to select the most informative time windows for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning methods are used to train control devices with large volumes of training data, then the control device can be configured to optimize system performance, but the training process becomes computationally expensive and time-consuming
Solution Approach 1:
The patent extracts only the most relevant training data segments by identifying time windows where control actions actually caused state changes. Instead of using all available training data, the method extracts subsets containing causal relationships between control interventions and system effects, significantly reducing the data volume required for effective training while maintaining or improving training efficiency
Solution Approach 2:
The patent applies local quality by differentiating between relevant and irrelevant training data segments. It identifies specific time windows with high correlation values where control actions had meaningful impacts on system states, and focuses training computation on these localized high-value segments rather than uniformly processing all data, thereby improving training efficiency
2Productivity
If training data is reduced to improve training efficiency, then training time and computational resources are reduced, but the coverage of operating conditions may become insufficient
Solution Approach 1:
The patent changes the parameter of data selection from random or uniform sampling to correlation-based selective sampling. By computing correlation values between control actions and state changes, and selecting time windows with correlation values above a threshold, the method ensures that reduced training data sets maintain representative coverage of meaningful operating conditions while improving training efficiency
Data Source
AI summary
A controller for a technical system is trained using a machine learning method. For this purpose, a chronological sequence of training data is detected for the machine learning method, the training data including both state data as well as control action data of the technical system. A change in the control action data over time is detected specifically and correlated with changes in the state data over time within different time windows, wherein a time window-specific correlation value is ascertained in each case. A resulting time window is then ascertained on the basis of the ascertained correlation values, and the training data which is found within the resulting time window is extracted in a time window-specific manner. The controller is trained by means of the machine learning method using the extracted training data and thereby configured to control the technical system.