ML Control Training Using Correlated Time-Window Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing control devices for technical systems require large amounts of representative training data to optimize their performance, but often struggle with limited coverage of operating conditions, leading to inefficient training and potential negative impacts on success.
Innovation Solution
A method that records and correlates temporal changes in control action data with state data within specific time windows, extracting training data from these windows to focus the learning process on highly informative data, thereby determining a resulting time window for efficient training, using techniques like numerical optimization to enhance correlation values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If all available training data is used for training, then the training process covers more operating conditions, but the training efficiency decreases and causal relationships are learned slower
Solution Approach 1:
The patent segments the training data by dividing it into multiple time windows based on temporal correlations between control actions and system states. Instead of treating all training data uniformly, the system identifies and separates time windows where causal relationships are strongest, allowing efficient training on high-quality segments while maintaining coverage of diverse operating conditions across different windows.
Solution Approach 2:
The patent applies local quality by assigning different weights or priorities to different time windows based on their correlation values. Time windows with higher temporal correlations between control actions and state changes are identified as having higher local quality and are selected for training, while lower quality windows are excluded or downweighted, thereby improving overall training efficiency.
2Productivity
If training data is pre-filtered based on feature correlation, then training efficiency improves, but the selection of relevant features becomes more complex
Solution Approach 1:
The patent uses feedback by calculating temporal correlation values between control actions and system states to automatically identify high-quality training time windows. This feedback mechanism guides the selection process, where the correlation metric provides objective criteria for filtering training data, reducing the need for manual feature selection while improving training efficiency through data-driven window selection.
Data Source
Figure 1~2
Figure 3
AI summary
A control device (CTL) according to the invention for a technical system (TS) is trained using a machine learning method. For this purpose, a temporal sequence of training data (TD) is acquired for the machine learning method, wherein the training data (TD) comprises both state data (SD) and control action data (AD) of the technical system (TS). According to the invention, a temporal change (ΔAD) of the control action data (AD) is specifically acquired and correlated with temporal changes (ΔSD) of the state data (SD) within different time windows (TF), whereby a time-window-specific correlation value (CC) is determined in each case. Depending on the determined correlation values (CC), a resulting time window (RTF) is then determined, and the training data (FTD) located within the resulting time window (RTF) are extracted in a time-window-specific manner.The extracted training data (FTD) is used to train the control unit (CTL) using the machine learning method and thus configured to control the technical system (TS).