Automating Structural Health Monitoring: A Machine Learning-Enhanced Bayesian Operational Modal Analysis Framework for Long-Span Bridges

Long-span bridges are critical components of transportation networks, requiring continuous monitoring to ensure their safety and operational efficiency. The advent of structural health monitoring (SHM) systems has enabled the real-time tracking of a structure’s health by analyzing its vibration response to ambient environmental loads. Traditional modal analysis methods, while effective, often suffer from limitations such as high sensitivity to noise and the need for manual intervention in data processing and modal parameter identification, leading to a demand for more autonomous and reliable approaches. Bayesian Operational Modal Analysis (OMA) stands out by providing a probabilistic framework that quantitatively assesses the uncertainty in modal parameter identification, offering a robust alternative to deterministic methods. However, its application in long-term SHM is hindered by the necessity for expert intervention in parameter selection, making the process labor-intensive and prone to subjective biases.

A new study published in Engineering Structures and conducted by Dr. Jianxiao Mao, Mr. Xun Su, Professor Hao Wang from Southeast University, and Dr. Jinyang Li from Dalian University of Technology, the researchers developed an automated framework for modal parameter identification in long-span bridges, integrating machine learning with Bayesian OMA. Their research involved both numerical simulations and real-world data application to a long-span suspension bridge. The team introduced the Cross-Modal Assurance Criterion (CMAC) matrix to globally correlate target modes across the frequency spectrum, to address the limitations of traditional Modal Assurance Criterion (MAC) sequences. They used Convolutional Autoencoder (CAE) to reconstruct the CMAC matrix, effectively denoising the data and identifying modal resonant frequency bands. This step was essential in processing and filtering the inherent noise in SHM data, ensuring accurate modal frequency detection. Moreover, they employed a Kohonen network, or Self-Organizing Map, to automatically classify normalized singular value (SV) sequences, determining the number of target modes within specific frequency bands. This method reduced the need for manual intervention and subjective decision-making in the modal analysis process. Additionally, the team performed extensive numerical simulations to generate structural acceleration data under various modes, noise levels, and excitation conditions. These simulations aimed to replicate the complex conditions a long-span bridge might experience, providing a robust dataset for testing the proposed framework. Furthermore, the framework was applied to real-world SHM data from a full-scale long-span suspension bridge. This step validated the effectiveness of the proposed method in a practical setting, assessing its performance in identifying and tracking modal parameters from actual structural health monitoring data.

The authors found that CAE-reconstructed CMAC matrix significantly improved the accuracy of modal frequency band detection by effectively removing noise. This approach proved more reliable than traditional methods, demonstrating its capability to handle the large volume and complexity of SHM data.  They also demonstrated the Kohonen network successfully automated the process of determining the number of target modes within given frequency bands. This automation was found to be accurate and significantly less labor-intensive compared to manual methods. The numerical simulations confirmed that the proposed framework could accurately estimate modal parameters under various simulated conditions. The method demonstrated robustness against noise and the ability to distinguish closely spaced modes, a common challenge in modal analysis.

The application of the framework to real-world data from a long-span suspension bridge showcased its practical viability. The method was able to process and analyze actual SHM data, accurately estimating modal parameters and thereby validating its effectiveness in a real-world scenario. The authors’ findings suggest that the proposed automated framework is not only applicable to long-span bridges but also has the potential to be adapted for other types of structures requiring SHM. Its integration of machine learning with Bayesian OMA presents a significant advancement towards more autonomous and reliable structural health monitoring systems.

The implications of the new study are far-reaching, offering a scalable, efficient solution for the SHM of long-span bridges and potentially other structures. By reducing the reliance on manual processes and subjective judgment, the proposed method enhances the objectivity and reliability of SHM practices. Moreover, the integration of machine learning techniques opens new avenues for the development of intelligent SHM systems capable of predictive maintenance and real-time health assessment. In conclusion, the new study contributes a novel, automated framework for modal parameter identification, leveraging machine learning to enhance Bayesian OMA. Their experiments, ranging from numerical simulations to real-world application, demonstrate the framework’s accuracy, efficiency, and practical viability, marking a significant step forward in the field of structural health monitoring of bridges and other large-scale civil infrastructure.