MACHINE LEARNING BASED STRUCTURAL HEALTH MONITORING OF ROCKING BRIDGE SYSTEM UNDER SEISMIC EVENTS

Resilience is often defined as the ability to prepare and plan for, absorb, recover from, and more successfully adapt to adverse events. Resilience is assessed using 4R-Methodology, including robustness, rapidity, redundancy, and resourcefulness. Machine learning-based structural health monitoring (ML-SHM) can enhance resilience by reducing impeding factors caused by mobilizing equipment, inspection crew, and in-house structural assessment that may lead to shorter recovery time after an event and continuous monitoring of structures for enhanced resourcefulness throughout life. This paper discusses machine learning approaches for structural health monitoring using recorded time-series data from structures during seismic events using convolution neural networks (CNNs). Oftentimes, training CNNs with time-series data requires conversion methods such as reconstructing a random size matrix, converting time-series data into the frequency domain, or converting it into histograms. Even though the reported literature showed acceptable accuracies, the abovementioned methods still ignore the temporal nature of the time series data. ML-SHM approach was developed by converting the input time-series data into input images using methods such as Gramian Angular Summation Field (GASF), Gramian Angular Difference Field (GADF), and Markov Transition Field (MTF). Then, the encoded images were used in CNN models. This paper discusses the training and testing results of CNNs trained with images obtained from time-series data (acceleration, displacement, and strain) that were collected from shake table tests at the University of Nevada, Reno. The CNNs were trained using acceleration and displacement data and using different image encoding techniques (GASF, GASF, MTF). Convolutional neural network models trained on MTF-encoded images from acceleration data performed with 100% accuracy during the training phase and more than 94% for the testing phase for all types of data.