Dissertation Defence: Fusion of multimodal nondestructive inspection data for improved pipeline defect characterization
December 3 at 8:00 am - 12:00 pm
Jiatong Ling, supervised by Dr. Zheng Liu, will defend their dissertation titled “Fusion of multimodal nondestructive inspection data for improved pipeline defect characterization” in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Electrical Engineering.
An abstract for Jiatong Ling’s dissertation is included below.
Examinations are open to all members of the campus community as well as the general public. Please email zheng.liu@ubc.ca to receive the Zoom link for this exam.
Abstract
Pipelines are economical and efficient modes of transporting oil and gas. These systems inevitably confront various risk factors throughout their lifespan, which can lead to defects. These defects compromise the integrity of pipeline systems and may result in catastrophic accidents. Accurate characterization of defects through nondestructive inspection (NDI) is essential for maintaining pipeline integrity.
Despite the strengths of various inspection techniques, individual inspection tools are limited in their ability to quantify all types of defects accurately. Therefore, this research aims to fuse multiple inspection resources, specifically the axial magnetic flux leakage (MFL) and ultrasonic testing (UT) pipeline in-line inspection data, to improve defect characterization results. The developments were carried out in three steps, i.e., uncertainty quantification of defect characterization using an individual NDI measurement, registration of multimodal NDI measurements, and fusion of multimodal NDI measurements.
Firstly, an uncertainty quantification framework for defect characterization from individual NDI measurements is developed. The proposed framework integrates the uncertainty estimate technique and a task‑guided conformal predictor to produce both point estimation and statistically valid prediction intervals. The comprehensive experiments on real-world inspection data demonstrate that the proposed framework constructs intervals that achieve target coverage while adapting to local defect variability.
Secondly, an automated registration method is proposed for MFL and UT inspection data to spatially align the multimodal NDI measurements. A translation model based on a generative adversarial network (GAN) is trained to translate UT data into the MFL domain, and the registration parameters are optimized by applying a difference metric that minimizes the gradient of the difference map between the translated and measured MFL images. The resulting alignment of multimodal inspection data enables reliable subsequent data integration.
Thirdly, a deep-learning fusion framework is designed to integrate MFL and UT for improved defect characterization. The proposed multimodal fusion network employs a dual-branch encoder for modality-specific feature extraction and a cross-modal enhancement module for feature interaction. Guided by UT-derived edge priors, the decoder reconstructs ambiguity-reduced and structurally faithful depth profiles. Experiments on real inspection data demonstrate that the proposed method achieves higher characterization accuracy compared with single-modality and conventional fusion baselines.
The above developments provide effective and efficient tools for pipeline integrity management. In particular, the accurate characterization of defects is critical for downstream pipeline residual strength, probability of failure, and residual lifetime assessment. This enables the implementation of predictive-maintenance technology in pipeline operations, facilitating optimal inspections and ensuring the maximization of the pipeline system’s long-term utility.
