Machine learning-enhanced acoustic emission technique for impact source identification and classification in steel pipes.

Fatukasi, Oluseyi S.; Abolle-Okoyeagu, Judith; Prathuru, Anil; Solademi, Oludare

Abstract

External impacts represent a prevalent source of structural damage in pipes/piping systems. Detecting and identifying the nature, and the severity of these impacts would enhance the reliability and operational safety of such critical infrastructure. This study investigates the integration of machine learning algorithms into acoustic emission (AE) techniques, for the identification and the classification of external impact sources in carbon steel pipes. The research methodology comprised two phases: a reference experiment and a main experimental study. The reference experiment utilized a 100 cm carbon mild steel pipe with Hsu-Nielsen pencil lead breaks at three source points. AE parameters (energy, peak amplitude, rise time) were compared between open-ended and damped cross-section conditions. Results validated the AE wave attenuation concept and confirmed the effects of boundary conditions on AE wave propagation. The preliminary calibration study established a baseline for experimental validation and ensured the integrity of the main experiment. The main experiment involved the controlled dropping of steel balls of varying masses (9g and 17g), from different heights onto a steel pipe. AE signals were recorded at 25cm, 40cm, and 55cm distances between the attached piezoelectric sensor and impact locations. The AE wave signal was transformed using Fast Fourier Transform (FFT). The frequency spectral was further analyzed using a Short-time Fourier Transform (STFT). Subsequently, key AE parameters, including amplitude, energy, rise time, and frequency content, were extracted from the recorded waveforms. These parameters were subsequently utilized to train and evaluate seven supervised machine learning models. Performance evaluation based on precision, F-1 score, and recall metrics revealed that the gradient boost algorithm achieved the highest accuracy of 0.74, while the support vector machine (SVM) model demonstrated the lowest accuracy of 0.10. The study exhibited the robustness of machine learning classifiers and AET’s effectiveness in distinguishing various impact scenarios in steel pipes towards the advancement of structural health monitoring of piping systems.

Citation

FATUKASI O.S., ABOLLE-OKOYEAGU, J., PRATHURU, A. and SOLADEMI, O. 2024. Machine learning-enhanced acoustic emission technique for impact source identification and classification in steel pipes. Presented at the 12th Annual conference of Society of structural integrity and life (DIVK12), 17-19 November 2024, Belgrade, Serbia.

• Presentation Conference Type: Presentation / Talk
• Conference Name: 12th Annual conference of Society of structural integrity and life (DIVK12)
• Start Date: Nov 17, 2024
• End Date: Nov 19, 2024
• Acceptance Date: Sep 18, 2024
• Publication Date: Nov 17, 2024
• Deposit Date: Sep 20, 2024
• Publicly Available Date: Nov 21, 2024
• Peer Reviewed: Peer Reviewed
• Keywords: Acoustic emission testing; Machine learning; Impact detection; Steel pipes; Structural health monitoring; Signal processing
• Public URL: https://rgu-repository.worktribe.com/output/2481725