Pipelines constitute critical infrastructure owing to the economic and environmental impacts of the catastrophic nature of component level failures. Hence, development of effective damage detection schemes for pipes is crucial for a robust structural health monitoring system. Use of guided ultrasonic waves (GUWs), for the purposes of damage detection in pipes, is a popular choice due to its damage localization capabilities with respect to traditional vibration based techniques. Development of high fidelity models, capturing the physics involved in GUWs, for damage detection becomes computationally exhaustive. An alternative is to use data-driven models that replace the high fidelity model with statistical learning based parametric models. The goal of such surrogate models is not necessarily to simulate the system behavior in all its complexity, but focus on the task of detecting damage efficiently. We propose a semi-supervised statistical learning approach for damage detection in pipes. This involves combining an unsupervised learning technique with minimal a priori information to aid level I damage detection (detection of presence of damage). Experiments are conducted on a steel pipe for demonstrating the efficacy of the proposed approach. Guided wave signals are acquired in a pitch-catch setting using piezoelectric sensors. The acquired signals, from both damaged and undamaged configurations of the pipe, are used for the proposed algorithm. The semi-supervised learning echnique is shown to be effective in detecting presence of damage with the use of minimal sensors.

Related Publications