Ali Fakhri's M.Sc Thesis defense presentation
Supervision: Zachary Hamida & James-A. Goulet
Thesis:
Bayesian Neural Networks to Factor-in Structural Attributes in Infrastructure Probabilistic Deterioration Models
http://profs.polymtl.ca/jagoulet/Site...
Outline:
00:00 Context
01:10 Data
03:35 Existing Model
07:50 Limitations
09:05 Challenges
11:15 BNN
12:50 Research Objectives
13:48 Proposed model
20:01 Joint Framework
24:49 Conclusion
Abstract: Transportation infrastructure plays a crucial role in fueling the economic growth by facilitating the movement of people, goods, and services. Over time, the condition of the infrastructure deteriorates due to aging, usage, and environmental factors. Although periodic visual inspections are the most common approach to monitoring structural health at a large scale, they are costly and subjective. This results in few and often unreliable data points for each structure. Despite this, most public infrastructure owners increasingly rely on deterioration models based on visual inspections for planning maintenance and rehabilitation activities. Yet, depending solely on a limited number of visual inspections for each structure is not sufficient to reliably model their deterioration. Factoring in structural attributes (e.g., age, location, etc.) can compensate for the lack of inspections by allowing information-sharing between structures. These attributes are good predictors of deterioration and are often readily available. A recent kernel-based regression method has successfully included structural attributes in an infrastructure probabilistic deterioration model while also quantifying the inspectors’ uncertainty. Despite being capable of reliably modeling deterioration, this method requires considerable computational resources and can only include few structural attributes. These issues make the kernel-based approach unfit for modeling the deterioration of large infrastructure networks; yet, finding an alternative method is not trivial, as it must be probabilistic, computationally efficient, and capable of including many attributes. Although Bayesian neural networks (BNN) are well-suited for large datasets and have many of the desired qualities, their integration into the existing deterioration model has been restricted by their inference mechanisms relying on sampling or gradient-based optimization. However, closed-form inference in BNNs was recently made possible by a probabilistic method called tractable approximate Gaussian inference (TAGI). This research aims to fuse a TAGI-trained BNN with a large-scale infrastructure probabilistic deterioration model to learn the relation between the structural attributes and the deterioration rates. The proposed method is verified on a synthetic visual inspection dataset and its performance is compared against the existing kernel-based approach using the bridge network data from the province of Quebec. The new method is shown to be orders of magnitude faster than the existing one without com- promising the predictive performance. It seamlessly incorporates all the available attributes, which removes the tedious and time-consuming task of identifying the most important ones each time new analyses are performed. These advantages prompted the extension of the pro- posed method to jointly estimate the parameters of inspector’ across all structural categories (beams, slabs, etc.), rather than having to rely on the current category-wise setup. They also enabled automating the end-to-end processing of all the data composed of thousands of structures encompassing millions of visual inspections. The effect of the joint framework on inspectors’ parameters is examined, and its performance is compared against the category- wise approach using several structural categories. Overall, the efficiency and scalability of the proposed joint and category-wise frameworks allow for reliably modeling the deterioration of large infrastructure networks.