Finding reliable signals in noisy and uncertain data.
Problem
In real-world systems, data is noisy, incomplete, and influenced by changing operating conditions.
This makes it difficult to detect early signs of degradation reliably.
Approach
I developed data-driven methods to detect deviations in system behavior, taking into account variability in operating conditions and limited labeled data.
Outcome
The approach enables early detection of changes in system behavior, even under realistic field conditions, supporting timely intervention and analysis.
What I did
• Analyzed time-series data from real systems
• Developed detection methods for system-level deviations
• Considered operating conditions and context in the analysis
• Developed detection methods for system-level deviations
• Considered operating conditions and context in the analysis
What was difficult
• High noise and variability in field data
• Limited labeled degradation cases
• Separating real signals from normal system behavior
• Limited labeled degradation cases
• Separating real signals from normal system behavior
What I learned
• Real-world data requires context, not just models
• Detection is more challenging than classification
• Robust methods matter more than complex ones
• Detection is more challenging than classification
• Robust methods matter more than complex ones
