Latent Monotonic Feature Discovery for Structural Health Monitoring.

Area of Science:

• Civil Engineering
• Data Science
• Structural Health Monitoring (SHM)

Background:

• Quantifying civil infrastructure health from sensor data is challenging due to weak degradation signals obscured by environmental and operational effects.
• Structural health monitoring (SHM) data often exhibits periodic or intermittent patterns, masking subtle long-term degradation trends.

Purpose of the Study:

• To develop methods for extracting meaningful health proxies from complex sensor data in structural health monitoring.
• To address the challenge of identifying subtle degradation signals masked by noise and operational variability.

Main Methods:

• Proposing monotonicity as a guiding principle, operationalized via Spearman's rank correlation between sensor values and time.
• Employing subgroup discovery to identify monotonic sensor groups for robust health proxy aggregation.
• Introducing Latent Monotonic Feature Discovery (LMFD) to find monotonic combinations of non-monotonic sensor data.

Main Results:

• Meaningful monotonic health proxies were successfully derived from both naturally monotonic sensor subgroups and composite features.
• The methods demonstrated the ability to extract latent degradation trends from a two-year bridge monitoring dataset.
• Interpretable, indirect indicators of structural health were uncovered using the proposed approach.

Conclusions:

• Monotonicity-based methods offer a principled approach to uncovering latent degradation trends in long-term SHM data.
• The developed techniques provide robust and interpretable proxies for assessing civil infrastructure health.
• This study enhances the capability to quantify infrastructure health using challenging, real-world sensor datasets.