Session: 16-01-01: Poster Session: NSF-Funded Research (Grad & Undergrad)

Paper Number: 99419

99419 - Smart Bridge: Machine Learning to Identify Structural Defects of a Bridge and a Moving Vibration Source in Real-Time 

In the field of vibration-based structural health monitoring, there is a high need for engineers to employ vibration-based methods to precisely identify the reduction in structural stiffness due to corrosion and/or cracks in structural members. It is also crucial for engineers to measure corresponding vibration responses on a structure such as a bridge, tunnel, or roadway itself. Theoretical studies and computational algorithms exist in attempting to tackle such problems, but all the existing methods deploy a lot of limitations of which a major one is a difficulty in real-setting implementation due to its computational cost. The most recent approach to tackle this problem was using the genetic algorithm (GA). Using a GA-based inversion method, researchers were able to predict stiffness parameters and the material properties of a one-dimensional Timoshenko beam. While the GA-based inversion method does yield inspiring results in the prediction of stiffness values and the material properties, the computational cost of prediction and time is concerning. Even a one-dimensional problem takes an extremely long time to finish making predictions.

The following work presents the implementation of a one-dimensional convolutional neural network (CNN) for the simultaneous identification of the material properties of the Timoshenko beam and the material properties of a moving vibration source. This work incorporates the use of the finite element method (FEM) to solve wave equations in a Timoshenko beam exposed to a moving vibrating source. Using the FEM wave solver, we generate synthetic data across multiple different stiffness parameters of the Timoshenko beam and five properties of the moving vibrational source in diverse ranges: the position, velocity, acceleration, frequency, and amplitude. Using modern machine learning (ML) algorithms, we utilize the powerful feature learning capabilities between the input- and output-layer datasets providing a robust prediction of the stiffness values of the different sections of the beam and the material properties of the source waves. The utilization of such modern architectures makes it a desirable method to generate accurate predictions for both stiffness parameters and the material properties of the source waves in real-time. It is true that a neural network requires an extensive computational process to train the network, but once trained, the trained model can predict the desired properties fast and without compromising prediction performance. The authors are aware that in a real-life setting, the sensor recorded data might employ a certain amount of noise. In this work, we show that our trained CNN is extremely effective in handling noise-riddled sensor data.

The presented work focuses on showing the readers that we are able to deploy a one-dimensional CNN to tackle a one-dimensional problem such as this, and the extension of such an approach to a higher-dimensional problem is straightforward and highly desirable.

Presenting Author: Shashwat Maharjan Central Michigan University

Presenting Author Biography: Shashwat Maharjan is a recent graduate from Central Michigan University (CMU) with a major in Mechanical Engineering and a minor in Mathematics. He currently works at the Solids, Waves, Intelligence, and Mechanics lab at CMU under the mentorship of Dr. Chanseok Jeong. Recently, Shashwat was awarded the CMU President&#39;s Award for outstanding undergraduate research, the Rictmeyer-Foust award for the best Mathematics senior, and the Engineering Society of Detroit’s 2022 Outstanding Undergraduate Students of the Year award.<br/><br/>Shashwat plans to continue his educational journey as he is an incoming graduate student at Central Michigan University&#39;s Masters of Engineering program at CMU.

Authors:

Shashwat Maharjan Central Michigan University
Bruno Guidio Central Michigan University
Chanseok Jeong Central Michigan University