Weak Monotonicity With Trend Analysis for Unsupervised Feature Evaluation.

Area of Science:

• Engineering
• Data Science
• Machine Learning

Background:

• Engineering system performance degrades over time due to wear and aging.
• Supervisory controlled processes generate numerous signals for monitoring degradation.
• Unsupervised feature selection is crucial for predictive analysis with large, unlabeled datasets.

Purpose of the Study:

• To propose a novel unsupervised feature selection method robust to measurement disturbances.
• To identify key features for predictive analysis in degrading systems using weak monotonicity (WM).
• To develop a suitability indicator for feature evaluation in the presence of noise and population variation.

Main Methods:

• Utilized the concept of weak monotonicity (WM) for feature selection.
• Developed a novel suitability indicator based on WM to evaluate feature performance.
• Applied the framework to select key features contributing to WM across a family of processes.
• Conducted a comparative study against nine state-of-the-art unsupervised methods.

Main Results:

• The proposed framework demonstrated promising performance in unsupervised feature evaluation.
• Effectively identified key features in the presence of measurement noises and population variations.
• Outperformed existing methods on benchmark datasets in handling noisy and varied data.

Conclusions:

• The weak monotonicity (WM) based feature selection framework is effective for degrading engineering systems.
• The method offers robust performance in unsupervised learning scenarios with noisy and varied data.
• Provides a valuable tool for predictive maintenance and system health monitoring.