Representative Data Selection for Adaptive Process Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing monitoring systems face challenges in selecting an optimal training set from historic sensor data for empirical modeling, which can lead to false alarms, missed alarms, and computational inefficiencies, particularly in dynamic systems where structural changes occur frequently.
Innovation Solution
A method for selecting a representative training set by segmenting the magnitude axis into equal or non-uniform intervals and choosing snapshots closest to these intervals, ensuring more data points are included from dynamic ranges while minimizing the overall set size, allowing for customizable distribution and emphasis on specific variables.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large volume of historic sensor data is used to train the empirical model, then the model fidelity and representativeness improve, but the computational overhead and training time increase
Solution Approach 1:
The patent segments the magnitude axis of sensor data into multiple intervals (e.g., quintiles) and selects representative snapshots from each interval. This segmentation approach reduces the total number of snapshots needed for training while ensuring comprehensive coverage of the dynamic range, thereby decreasing training time without sacrificing model fidelity.
Solution Approach 2:
The patent applies different selection criteria to different regions of the data distribution. By identifying and emphasizing snapshots from dynamic ranges (regions with significant parameter changes) versus stable ranges, the method ensures higher quality training data in critical regions while reducing overall data volume, thus improving model fidelity efficiently.
2Reliability
If more snapshots are selected from dynamic ranges, then the model's ability to detect incipient changes improves, but the overall training set size increases
Solution Approach 1:
The patent applies partial action by selecting a disproportionate number of snapshots from dynamic ranges compared to stable ranges. This selective emphasis ensures sufficient representation of critical dynamic behavior in the training set without including excessive data from less informative stable regions, thereby maintaining detection capability while controlling training set size.
Solution Approach 2:
The patent changes the selection parameter from uniform random sampling to stratified sampling based on magnitude intervals. By defining selection probability as a function of the interval position and local dynamics, the method optimizes the balance between detection capability and training set size through parameter-driven selective sampling.
3Ease of operation
If uniform segmentation of the magnitude axis is used, then the selection process is simplified and computationally efficient, but the distribution of selected snapshots may not optimally represent regions of high dynamics
Solution Approach 1:
The patent introduces dynamics into the selection process by making the selection probability dependent on local data characteristics within each interval. While the segmentation itself remains uniform and simple, the selection criterion adapts to dynamic regions by favoring snapshots where parameter changes are most significant, thus maintaining simplicity while improving representation accuracy.
Solution Approach 2:
The method incorporates feedback from the data distribution analysis into the selection process. By evaluating local dynamics metrics (such as variance or rate of change) within each magnitude interval, the system adjusts snapshot selection to emphasize regions with higher dynamics, ensuring accurate data representation while building upon the simple uniform segmentation framework.
Data Source
AI summary
System and method for selection of appropriate modeling data from a general data set to characterize a modeled process. The data is typically correlated sensor data, representing a multitude of snapshots of a sensed machine or process. The invention accommodates selection of greater amounts of general data for inclusion in the modeling data where that data exhibits greater dynamics, and selects less data from regions of little change. The system can comprise a computer running a software program, or a microprocessor.