Login to the D&D KM-IT.

Not a registered user? click here for U.S. registration or here for international registration.      Forgot your password? Click here

User name:   Password:   Close

Search the D&D KM-IT

Welcome Guest
Try our mobile friendly tool.

Share page:  

printer friendly logo

FIU has successfully completed a Design & Development of Convolutional AutoEncoder algorithm to identify cracks in D&D mock-up facility

Published: December 23, 2020

ML Task.pngConvolutional AutoEncoder (CAE) architecture developed by the FIU team

Florida International University has successfully completed a Design & Development of Convolutional AutoEncoder algorithm to identify cracks in D&D mock-up facility. ​This task supports structural health monitoring of D&D facility to identify cracks and structural defects for surveillance and maintenance. Structural health monitoring is imperative to the ongoing surveillance and maintenance (S&M) across the DOE complex. Deep Learning algorithms provide state-of-start technologies capable of facilitating the assessment of structural integrity in aging nuclear facilities. The FIU team has implemented a robust solution for anomaly detection as part of the structural health monitoring system.

The anomalies are defects (cracks) on D&D concrete structures . The proposed architecture uses a Convolutional AutoEncoder (CAE) – deep learning approach followed by an image post processing routine to generate an anomaly heat map of the predicted defects in the input images. This approach provides a robust solution by training on defect-free wall images for anomaly detection in an unsupervised learning manner.


Abstract: In this paper, we have proposed a semi-supervised based anomaly detection mechanism, which involves a convolutional auto-encoder, to facilitate the structural healthcare monitoring of DOE infrastructure. In this regard, the cracks and spalling are considered as the anomalies on the concrete structures, whereas the normal/healthy surfaces of the concrete structures are considered and grouped as the normal class. We employ an unsupervised learning methodology for training the model, as it only takes the normal class image datasets as its input data. After the model is trained, the proposed “semi-supervised” anomaly detection approach does not require any prior knowledge of the concrete defects. It traces the possible anomalies/defects with the least involvement of the domain experts.

Click link below to read full paper.

Task 6 - Structural Health Monitoring Deliverable_2020_P3_D2_Report_FINAL.pdfTask 6 - Structural Health Monitoring Deliverable_2020_P3_D2_Report_FINAL.pdf

Back to Top
More Modules

Download Original    Management of D&D of Oak Ridge Building 3505 | After

Pre cache