Research

The overarching theme of my research is making built infrastructure systems smart and resilient through the synergistic integration of data analytics and machine intelligence with engineering domain expertise. This requires high-fidelity digital replicas of engineered assets known as Digital Twins. My work facilitates such digital representations by extracting and analyzing raw big data into accurate actionable information via:

Theme 1: Information extraction in the natural and built environment

Theme 2: Augmented engineering data visualization via mixed reality

Theme 3: Computational modeling for damage Identification, monitoring

Effective preparedness and timely response to natural disasters requires the integration of artificial intelligence with remote sensing to build scalable data harnessing and information extraction solutions. Many such applications have been hampered by the shortage and heterogeneity of field-extracted ground truth data. We are building a scalable semi-supervised wildfire fuel mapping framework based on propagating the sparse ground truth labels, which will enable physics-based computational fire models.

This research is part of a larger multi-disciplinary NSF-funded research effort that introduces a data-informed, physics-based computational framework for combatting wildfire hazard, For more information about this project, please refer to https://packpages.unr.edu/wildfireproject/

This work leverages Internet-scale sources of data, namely web images and Google Street View scenes as a source for urban infrastructure assessment. To maximize scalability and minimize costs and human supervision, a deep omni-supervised learning scheme was employed to automatically label images and reuse them for re-training models. The resulting models can be used by city officials for monitoring of urban infrastructure, or in a crowd-sourced citizen-driven framework.

Read our paper in the Journal of Civil Structural Health Monitoring

In this work, a large collection of inspection report narratives was modeled using a deep hierarchical attention neural network to establish a mapping between qualitative inspection narratives and quantitative condition ratings. We further created a context-aware recurrent neural network that accounted for the context and dependencies of words to dissect the narratives into structural defects, their severity, and locations. This system can be used for maintenance planning and historical change detection.

• Li, T., Alipour, M., Harris, D.K. "Context-aware condition information extraction from bridge inspection reports using bi-directional LSTM-CRF sequence labeling". Advanced engineering informatics (in review)

• Li, T., Alipour, M., Harris, D.K. "Mapping textual descriptions to condition ratings to assist bridge inspection and condition assessment using hierarchical attention". Automation in construction (in review)

This work also proposes a semi-supervised framework that uses simple inexpensive texture images to train a deep learning-based corrosion damage identification model, and transfer its knowledge to complicated field imagery (e.g., a database of real inspection reports). This eliminates the need for costly training data procurement and annotation and introduces a scalable solution for large-scale model generation. We also use conditional random fields to explicitly model the interdependencies between the identified damage pixels, resulting in further uncertainty reduction in the predictions.

This work created the first pixel-level crack detection system using deep learning. Fully-convolutional neural networks were leveraged to identify cracks at the pixel level. The resulting dense probability maps were then used for sub-pixel measurements. This work significantly improved the state of the art in terms of the level of detail in detecting cracks and enabled the measurement of crack properties (e.g., width, length). Unlike the state-of-the-art sliding window approaches, the fully convolutional neural network model created in this work classified each pixel while differentiating between a real crack and extraneous crack-like distractors.

Detailed and accurate damage maps can not only assist in objective maintenance decisions, but can facilitate the creation of more realistic digital twins of existing aged infrastructure. This will in turn lead to uncertainty reduction in remaining life estimation via better damage-integrated numerical modeling.

Read our papers in Computing in Civil Engineering and Engineering Structures.