Sensor Index Selection Using PCA Migration Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for failure prognosis in equipment maintenance, such as machine learning approaches, face challenges in accurately selecting relevant sensors due to issues like multi-collinearity and the difficulty in identifying local features among vast quantities of data, leading to decreased prediction accuracy.
Innovation Solution
An index selection device that partitions sensor data into blocks, applies principal component analysis to each block, calculates migration distances between principal components, and detects migration factor indices with large migration distances to automatically select indexes with local features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If numerous sensors are installed and all sensor data is input to machine learning, then comprehensive monitoring is achieved, but prediction accuracy deteriorates due to multi-collinearity and dimensional differences
Solution Approach 1:
The patent extracts and removes redundant sensors from the system. By calculating correlation coefficients between sensor readings and removing sensors with high correlation to already-selected sensors, the system extracts only the essential, non-redundant sensors needed for accurate prediction, thereby reducing complexity while maintaining reliability
Solution Approach 2:
The patent changes the selection criterion from domain expert experience to an automated correlation-based metric. By using correlation coefficient calculations and automated selection algorithms, the system objectively identifies and removes redundant sensors based on statistical parameters rather than subjective judgment, improving both accuracy and efficiency
2Measurement precision
If domain experts manually select indexes based on experience, then local features may be identified, but automatic index selection becomes difficult when there are several thousand types of indexes
Solution Approach 1:
The patent introduces correlation coefficients as an intermediary metric between sensor data and selection decisions. Instead of directly comparing complex sensor patterns, the system uses correlation coefficients as a mediator to quantify the relationship between sensors and target variables, making automated selection both accurate and scalable to thousands of indexes
Solution Approach 2:
The patent replaces the mechanical process of expert manual review with an automated computational system. By substituting human expert analysis with algorithmic correlation calculations and automated selection processes, the system maintains the ability to detect local features while scaling to handle thousands of indexes that would be impossible for experts to review manually
3Device complexity
If principal component analysis is used to integrate indexes, then mutually independent indexes are selected, but local features are ignored because PCA captures only overall tendency
Solution Approach 1:
The patent applies local quality by treating different data segments differently. By dividing data into training sets and verification sets, and by identifying sensors that show significant performance differences across these sets, the system preserves local features that vary across different conditions while maintaining the overall structure provided by PCA
Solution Approach 2:
The patent segments the data into multiple parts (training set, verification set) and segments the sensor selection process into multiple stages. This segmentation allows the system to identify sensors with local features that may not be apparent in the overall data, capturing condition-specific important sensors that PCA alone would miss
Data Source
AI summary
Indexes having local features are automatically selected from sensor data of a plurality of sensors. Sensor data of the plurality of sensors, each associated with the plurality of indexes, is partitioned into a plurality of blocks. A principal component analysis is applied to the sensor data of each of the partitioned blocks and a plurality of principal components are extracted from each of the blocks. A migration distance evaluation unit extracts, from two different blocks, two principal components that form a principal component pair, and calculates a migration distance between each of the principal components regarding the extracted principal component pair. A migration factor index detection unit detects, as a migration factor index, an index among the plurality of indexes configuring the principal components having a large migration distance among the migration distances between each of the principal components calculated by the migration distance evaluation unit.