A Data-Driven Framework for Ultrasonic Structural Health Monitoring of Pipes

Cylindrical shells serve important roles in broad engineering applications, such as oil and natural gas pipelines, and pressurized industrial piping systems. To ensure the safety of pipe structures, various inspection equipment and platforms have been developed based on nondestructive testing (NDT) technologies. However, most existing approaches are time and labor intensive, and are only conducted intermittently. Drawbacks of current NDT methods suggest a proactive, automated and long-term monitoring system. Structural health monitoring (SHM) techniques continuously assess structural integrity through permanently installed transducers, allowing condition-based maintenance to replace the current practice of economically inefficient schedule-based maintenance.

Ultrasonics is an appealing SHM technology in which guided waves interrogate long stretches of a pipe with high sensitivity to damage, and can be generated by a surface-mounted, small-size piezoelectric wafer transducer (PZT). The challenges of implementing ultrasonic SHM with PZTs as active sensing devices lie in: (1) the wave pattern is complex and difficult to interpret; (2) it is even more difficult to differentiate changes produced by damage from changes produced by benign environmental and operational variability

The ultimate goal of this research is to develop an ultrasonic sensing and data analysis system for continuous and reliable monitoring of pipe structures. The objective of this dissertation is to devise a data-driven framework for effective and robust analysis of guided wave signals to detect and localize damage in steel pipes under environmental and operational variations. The framework is composed of a three-stage SHM scheme: damage detection,damage localization and damage characterization, supported by a multilayer data processing architecture incorporating statistical analysis, signal processing, and machine leaning techniques.

The data-driven methodology was first investigated through laboratory experiments conducted on a pipe specimen with varying internal air pressure. The sensed ultrasonic data were characterized and mapped onto a high dimensional feature space using various statistical and signal processing techniques. Machine learning algorithms were applied to automatically identify effective features, and to detect and localize a weak scatterer on the pipe. The reliability and generality of the data-driven framework was further validated through field tests performed on an in-service hot-water pipe under large, complex and uncontrollable operating conditions.

This data-driven SHM methodology involves an integrated process of sensing, data acquisition, statistical analysis, signal processing, and pattern recognition, for continuous tracking of the structural functionality in an adaptive and cost-effective manner. The techniques developed in this dissertation are expected to have broader applications related to the regular inspection, maintenance, and management of critical infrastructures not just limited to pipes.