Identifying and Assessing Damage in Infrastructure Using Topological Data Analysis and Machine Learning

Explore an innovative approach to identifying and assessing damage in infrastructure using topological data analysis and machine learning in this 43-minute conference talk by Stéphane Béreux. Discover a low-cost, imaging-based alternative to expensive laboratory diagnostics for assessing concrete damage. Learn about the application of a lightweight convolutional neural network (CNN) for crack recognition and the extraction of topological descriptors correlating with damage. Understand how topological data analysis is uniquely applied in the post-processing step. Examine the method's demonstration on real data samples of concrete affected by alkali-aggregate reaction. Gain insights into the potential of this proof-of-concept to become a fully functional and easy-to-use concrete damage assessment tool, pending more annotated data. Follow the presentation's journey from discussing the Genoa bridge disaster to exploring the chosen approach, CNN architecture, training data for segmentation, persistent histogram segmentation, crack segmentation process, and the application of concepts like Betti numbers and persistent homology in infrastructure damage assessment.

Intro
The Genoa bridge disaster
A growing risk
Damage Rating Index (DRI)
Necessity for a new method
The natural choice
Chosen approach
A revolutionary neural network architecture...
with groundbreaking performances
Main layers of a CNN I
Strengths of the CNN
Training data for segmentation
Segment the grey histogram of Pu
Failure of regular binarization methods
Persistent histogram segmentation
Alignment of the pictures
Cleaning of the homography artefacts
Result of the alignment
Finetuning
Crack segmentation process
Betti numbers
Interest of relative homology
Persistent homology
Total persistence dimension o
Maximal relative persistence dimension 1
Pipeline assessment