
- Sponsor
- Department of Civil and Environmental Engineering
- Views
- 45
- Originating Calendar
- CEE Seminars and Conferences
Advancing Noncontact Sensing and Asynchronous Multimodal Data Fusion for Miter Gates
Advisor: Professor Billie F. Spencer
Abstract
Navigational locks are critical nodes in the inland waterway transportation network which rely on the proper functioning of miter gates. To avert outages, Structural Health Monitoring (SHM) systems help prevent costly failures and optimize maintenance as miter gates age. However, the operational deployment of SHM on miter gates is hindered by data acquisition challenges. Contact sensor installation is costly and requires a dewatered structure, a rare occurrence. This dissertation advances miter gate SHM in three areas: expanding noncontact sensing capabilities in challenging environments, developing multimodal inference frameworks robust to camera motion, and handling misalignment between the multimodal data streams fed into inference frameworks.
First, to address the physical limitations of underwater contact sensors, this research proposes the use of imaging sonar as a noncontact sensing technology for miter gate monitoring. By demonstrating an upper bound of the detection threshold at approximately 0.4 mm in a laboratory setting, this work indicates controlled-condition viability.
Second, a multimodal model updating framework is presented to improve upon unimodal sensing approaches. This approach fuses local structural responses (strain) with global kinematics (vision-based displacement) to more tightly constrain and improve inference at jointly-observed parameters. To mitigate camera pose variation, the framework explicitly parameterizes camera motion to isolate and correct environmental distortions in the vision data.
Building upon the multimodal use of strain and vision data, the dissertation addresses the data asynchronicity problem, an issue in this real-world deployment where cameras and strain sensors rarely sample concurrently. To transition the multimodal techniques from hand-crafted datasets to direct use of field data, a probabilistic framework is introduced that recasts unaligned and sequential strain and vision observations as a stochastic observation process. The approach enables repeated assessment without strict temporal synchronization and provides near-real-time inference after offline training of a neural network.
The presented methodologies for noncontact detection of displacements, camera pose variation, and asynchronous multimodal fusion provide enabling methods for field-oriented SHM. The dissertation demonstrates these sensing and inference frameworks as method-development contributions, while validation of field condition-estimation accuracy remains for future work.