Data-driven semi-supervised and supervised learning algorithms for health monitoring of pipes

• Published in: Mechanical systems and signal processing 2019-09, Vol.131, p.524-537
• Main Authors: Sen, Debarshi, Aghazadeh, Amirali, Mousavi, Ali, Nagarajaiah, Satish, Baraniuk, Richard, Dabak, Anand
• Format: Article
• Language: English
• Subjects: Algorithms; Cast iron; Cluster analysis; Clustering; Damage detection; Damage localization; Data-driven structural health monitoring; Flaw detection; Fracture mechanics; Hierarchical clustering; Machine learning; Maintenance costs; Multinomial logistic regression; Piezoelectric transducers; Piezoelectricity; Pipes; Sensors; Statistical analysis; Wave propagation in pipes
• ISSN: 0888-3270; 1096-1216
• DOI: 10.1016/j.ymssp.2019.06.003

•A semi-supervised and a supervised learning algorithm for health monitoring of pipes is proposed.•The proposed approaches are data-driven, hence circumventing the need of computationally prohibitive model-based approaches.•The proposed approaches can perform damage detection in pipes with only two actuator-sensor pairs.•Efficacy of the proposed approaches are demonstrated using experimental data from two cast iron pipes. The use of guided ultrasonic waves (GUWs) for SHM of pipelines has been a popular method for over three decades. The superiority of GUWs over traditional vibration-based techniques lie in its ability to detect small damages (cracks and corrosion) over a satisfactory length of a pipeline. The physics of the system, however, is extremely involved that renders model-based techniques computationally prohibitive. Data-driven approaches, based on statistical learning algorithmsare far more suitable in such scenarios. In this paper, we propose two data-driven techniques, involving a semi-supervised and a supervised learning approach, for damage detection in pipes. In addition to circumventing the use of a model-based approach, the proposed approaches also aid in reducing the number of sensors deployed, leading to reductions in maintenance costs. The semi-supervised learning-based approach detects the presence of damage using a hierarchical clustering-based algorithm. The supervised learning-based approach performs damage localization in a multinomial logistic regression framework. We validate the proposed algorithms by acquiring guided wave responses from experimental pipes in a pitch-catch configuration using low-cost piezoelectric transducers. We demonstrate that our fully data-driven techniques accurately detect and localize cracks on two cast iron pipes of different lengths using a combination of two sensors.