TL;DR
This paper introduces a novel physics-guided signal processing method to extract a "damage-sensitive and domain-invariant" (DS & DI) feature from vehicle vibrations for bridge health monitoring. This new feature aims to overcome limitations of existing methods by being sensitive to damage while remaining robust to variations in bridge properties and environmental factors, achieving superior damage detection and localization in simulations.
What Problem Does This Paper Address?
Monitoring the structural health of bridges is crucial for public safety, but traditional methods are often costly and time-consuming. "Indirect" bridge health monitoring (IBHM) using vehicle vibrations offers a promising low-cost, low-maintenance alternative. However, existing IBHM approaches face two major hurdles: 1) manually engineered physics-based features are often not sensitive enough to detect damage, and 2) machine learning-based features tend to overfit to specific bridges and don't generalize well to others. This paper tackles these challenges by developing a feature that is both sensitive to damage and invariant across different bridge types.
Methods
The core of this work is a physics-guided signal decomposition approach. The authors model the vehicle-bridge interaction system and then leverage the Synchrosqueezed Wavelet Transform (SWT). The SWT is chosen for its ability to decompose non-stationary signals (like vehicle vibrations) into components that resemble intrinsic mode functions (IMFs), which are well-suited for representing localized signal characteristics.
The process involves:
- Modeling the Vehicle-Bridge Interaction System (VBIS): A physics-based model is established to understand the dynamics between a moving vehicle and a bridge, incorporating the possibility of local stiffness reduction (damage).
- Signal Decomposition with SWT: The vehicle's acceleration signal is transformed into the time-frequency domain using SWT. This allows for the representation of the signal as a superposition of components with varying instantaneous frequencies.
- Feature Reconstruction with Inverse SWT (ISWT): A specific component, identified as being sensitive to damage and invariant to other factors, is reconstructed from the SWT representation using ISWT within a carefully chosen frequency band. This reconstructed component is then normalized by a "domain-invariant factor" (C51) derived from the VBIS model to create the final DS & DI feature.
This approach aims to extract a feature that is inherently sensitive to changes in the bridge's stiffness (damage) while being robust to variations in vehicle properties, environmental conditions, and even different bridge designs.
Datasets
The paper utilizes simulated data generated from finite element models (FEMs) of five different bridges. These bridges vary in length and natural frequencies. The dataset includes scenarios with varying degrees of stiffness reduction (damage) at different locations, as well as undamaged cases. A total of 2000 vehicle acceleration records were generated for evaluation. No real-world datasets are used in this study.
Key Insights and Ideas
- Physics-Guided Signal Decomposition: Framing IBHM as a signal decomposition problem, rather than purely a machine learning task, allows for the incorporation of physical principles to guide feature extraction.
- SWT for Non-Stationary Signals: The SWT proves effective in decomposing complex, time-varying signals like vehicle accelerations into meaningful components, overcoming limitations of traditional time-frequency methods.
- The "Third Component" is Key: The authors identify a specific component within the vehicle's acceleration signal (related to the interaction of mode shapes and their derivatives) that inherently contains damage-sensitive and domain-invariant properties.
- Domain Invariance through Normalization: Multiplying the reconstructed signal component by a derived "domain-invariant factor" is crucial for making the feature robust to variations in bridge properties.
Key Results
The proposed DS & DI feature demonstrated superior performance in simulated experiments. When compared to conventional time- and frequency-domain features, as well as other signal decomposition techniques (like IMFs from EMD and inverse CWT), the proposed feature achieved the best damage quantification and localization results in five out of six experimental setups. This indicates its high sensitivity to damage and its robustness across different bridge configurations.
Limitations and Future Work
- Boundary Effects: The authors acknowledge that the ISWT reconstruction may exhibit boundary effects, which require further investigation.
- Supervised Learning Comparison: While the proposed feature shows promise in unsupervised/physics-guided learning, its performance is compared against supervised learning models in terms of RMSE. The paper suggests that the proposed feature can be used as an input to multi-task learning models for improved performance.
- Real-World Validation: The current study relies solely on simulated data. Future work should focus on validating the proposed method using real-world bridge monitoring data to confirm its effectiveness in practical scenarios.
- Computational Cost: While not explicitly detailed, the computational cost of SWT and ISWT might be a consideration for real-time applications, which could be an area for future optimization.
Summary updated