Machine-Learning-Based Non-Destructive Evaluation of Refractory Anchor Welds via Analysis of Percussion-Induced Acoustic Signals

Welding is a critical process in modern infrastructure, particularly for securing refractory anchors in high-temperature vessels used across refineries, power plants, and chemical facilities. Ensuring the quality of these welds is essential, but traditional inspection methods are often destructive, time-consuming, or limited in accuracy. This thesis investigates a novel non-destructive evaluation (NDE) technique that leverages machine learning (ML) and percussion-induced audio analysis to assess weld quality. A custom weld plate was fabricated containing both properly welded and intentionally flawed refractory anchors. Controlled mechanical impacts—using tools such as a hammer and chisel—were applied to the exposed ends of these anchors. The resulting audio signals were recorded and transformed into Mel-Frequency Cepstral Coefficients (MFCCs). These MFCCs served as input features for multiple machine learning models including supervised and unsupervised models. The supervised models used were support vector machines (SVM), logistic regression, recurrent neural networks (RNN). The unsupervised models used were k-means clustering. All models were evaluated using three progressively independent tests: a dependent 70:30 train-test split, a semi-independent test using newly recorded data from the same weld plate, and a fully independent test involving unseen anchors. Despite increased variability across tests, the supervised models achieved 100% classification accuracy, while the unsupervised clustering method reached 99.42%. These results demonstrate that audio-based machine learning offers a fast, cost-effective, and objective alternative for weld inspection, with strong potential for improving industrial quality control practices.

Name thesis:degree_name

Master of Science in Mechanical Engineering

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Houston

Year dc:date.issued

2025

Author dc:creator

• Chirigos, Mason

Advisor dc:contributor.advisor

• Song, Gangbing

Committee members dc:contributor.committeemember

• Chen, Xuemin
• Zhu, Weihang

• Mechanical engineering

Language dc:language.iso

English

Handle dc:identifier.uri

https://hdl.handle.net/10657/20627

OAI identifier oai:identifier

oai:uh-ir.tdl.org:10657/20627