Event Details
Lightweight Deep Learning Model for Nondestructive Evalulation of Crack Defects
Presenter: Yixiang Jia
Supervisor:
Date: Fri, July 26, 2024
Time: 08:00:00 - 00:00:00
Place: Zoom, link below.
ABSTRACT
Location: Remote via Zoom
Join Zoom Meeting
Meeting ID: 817 5353 4313
Password: 954905
One tap mobile
+17789072071,,81753534313# Canada
+16475580588,,81753534313# Canada
Dial by your location
+1 778 907 2071 Canada
+1 647 558 0588 Canada
Meeting ID: 817 5353 4313
Find your local number:
Note: Please log in to Zoom via SSO and your UVic Netlink ID
ABSTRACT
Ultrasonic nondestructive evaluation (NDE) is an essential tool in various industries, including aerospace, energy, and civil engineering, for assessing the structural integrity of manufactured products without damaging them. This thesis is focused on the automated analysis of ultrasonic NDE data by means of low-cost machine learning (ML) techniques, particularly in the context of inline pipeline inspection. We propose two lightweight neural network architectures for efficient multi-attribute classification to characterize surface-breaking crack defects in terms of their location, size, and tilt. Our networks have under 2M parameters and incorporate novel design elements inspired by the latest MobileNet models. Their computational footprint is also small, not exceeding 100M floating-point operations (FLOPs) per data sample. The proposed models process raw channel data acquired by a transducer array, as opposed to multi-view beamformed image patches utilized in related works, thus eliminating the computational burden associated with image reconstruction. Our evaluation results, based on a public-domain NDE dataset, demonstrate that our networks offer a balanced combination of their competitively high classification performance and low cost. These findings highlight the potential of lightweight deep learning models in ultrasonic NDE data analysis, which contributes to the development of more advanced and intelligent inspection systems. Our future research will focus on refining the proposed models to enhance their spatio-temporal feature learning, interpretability, generalization capability, and applicability to other fields, such as biomedical imaging and computer vision.