[Paper Review] System-reliability based multi-ensemble of GAN and one-class joint Gaussian distributions for unsupervised real-time structural health monitoring

Mohammad Hesam Soleimani‐Babakamali, Reza Sepasdar+2 more|arXiv (Cornell University)|Feb 1, 2021

Structural Health Monitoring Techniques28 references4 citations

TL;DR

This paper proposes a novel unsupervised real-time structural health monitoring (SHM) framework that combines multi-ensemble GANs and one-class joint Gaussian models using both low- and high-dimensional FFT features. By employing system-reliability-based Monte Carlo histogram sampling to tune detection thresholds, the method achieves robust, parameter-insensitive novelty detection with no false alarms on high-class datasets like Yellow Frame (21 classes) and Z24 Bridge (15 classes).

ABSTRACT

Unsupervised health monitoring has gained much attention in the last decade as the most practical real-time structural health monitoring (SHM) approach. Among the proposed unsupervised techniques in the literature, there are still obstacles to robust and real-time health monitoring. These barriers include loss of information from dimensionality reduction in feature extraction steps, case-dependency of those steps, lack of a dynamic clustering, and detection results' sensitivity to user-defined parameters. This study introduces an unsupervised real-time SHM method with a mixture of low- and high-dimensional features without a case-dependent extraction scheme. Both features are used to train multi-ensembles of Generative Adversarial Networks (GAN) and one-class joint Gaussian distribution models (1-CG). A novelty detection system of limit-state functions based on GAN and 1-CG models' detection scores is constructed. The Resistance of those limit-state functions (detection thresholds) is tuned to user-defined parameters with the GAN-generated data objects by employing the Monte Carlo histogram sampling through a reliability-based analysis. The tuning makes the method robust to user-defined parameters, which is crucial as there is no rule for selecting those parameters in a real-time SHM. The proposed novelty detection framework is applied to two standard SHM datasets to illustrate its generalizability: Yellow Frame (twenty damage classes) and Z24 Bridge (fifteen damage classes). All different damage categories are identified with low sensitivity to the initial choice of user-defined parameters with both introduced dynamic and static baseline approaches with few or no false alarms.

Motivation & Objective

• Address the limitations of existing unsupervised SHM methods, including information loss from dimensionality reduction and sensitivity to user-defined parameters.
• Overcome case-dependent feature extraction by using a simple, universal FFT-based feature extraction method applicable to any instrumented structure.
• Develop a dynamic and static baseline novelty detection system that identifies multiple damage classes without prior labeling or expert input.
• Ensure robustness to user-defined parameters by integrating system-reliability analysis via Monte Carlo histogram sampling on GAN-generated data.
• Demonstrate generalizability and scalability of the framework on real-world, high-class SHM datasets with minimal false alarms.

Proposed method

• Extract both low-dimensional (mean, variance) and high-dimensional (half-spectrum FFT) features from sensor data without case-specific preprocessing.
• Train a multi-ensemble of Generative Adversarial Networks (GANs) on high-dimensional FFT features to model normal structural behavior.
• Train one-class joint Gaussian (1-CG) models on low-dimensional features to capture baseline statistical distributions.
• Construct a novelty detection system using limit-state functions where detection scores from GAN and 1-CG models are treated as 'loads' and thresholds as 'resistance'.
• Use Monte Carlo histogram sampling on GAN-generated data to calibrate detection thresholds (resistance) based on user-defined reliability levels (e.g., $V_L$), ensuring parameter insensitivity.
• Apply both static and dynamic baseline approaches: static uses a fixed baseline, while dynamic adapts the baseline to detect multiple damage classes as novelties.

Experimental results

Research questions

• RQ1 Can a unified, unsupervised SHM framework detect multiple damage classes across diverse structures without prior labeling or case-specific feature engineering?
• RQ2 How can detection thresholds be tuned to be robust against user-defined parameters like $V_L$ in real-time SHM?
• RQ3 To what extent does combining low- and high-dimensional features improve detection sensitivity and generalizability in unsupervised SHM?
• RQ4 Can a system-reliability-based approach using Monte Carlo histogram sampling reduce false alarms and improve robustness in real-time structural monitoring?
• RQ5 How well does the proposed method generalize to high-class datasets such as Yellow Frame (21 classes) and Z24 Bridge (15 classes) with minimal false alarms?

Key findings

• The static baseline approach achieved zero false alarms on both the Yellow Frame (21 classes) and Z24 Bridge (15 classes) datasets, successfully identifying all damage classes.
• The dynamic baseline approach detected all 20 distinct damage conditions in the Yellow Frame dataset, with only 3, 1, and 2 false alarms for $V_L$ values of 10, 20, and 40, respectively.
• The method demonstrated insensitivity to user-defined parameters ($V_L$) due to reliability-based threshold tuning via Monte Carlo histogram sampling, confirming robustness.
• The framework successfully detected different damage types (e.g., pier lowering, tendon rupture) but could not distinguish between different intensities of the same damage type, likely due to similar FFT responses.
• The use of a simple half-spectrum FFT for feature extraction enabled direct application to new structures without retraining, supporting scalability for smart city infrastructure networks.
• The multi-ensemble of GANs and 1-CG models effectively captured complex normal behavior across diverse structural responses, enabling reliable anomaly detection in real-time.