Quantification of Leakage Flow in Industrial Valves by Acoustic Emission and Machine Learning

Abstract

Valve component tightness is critical for maintaining facility integrity and optimizing energy efficiency in power generation plants. Non-intrusive detection and quantification of internal leakages are essential for implementing condition-based maintenance. However, early detection and, in particular, accurate quantification—remains a major challenge. Conventional non-intrusive methods (such as vibration, acoustic emission, or temperature measurements) often exhibit limitations, primarily due to the lack of accurate quantification models and to the high levels of background noise in industrial environments. In this study, we propose using acoustic emission (AE) measurements to estimate leakage flow rates. A total of 1,500 tests were conducted on two experimental facilities designed to simulate industrial air, water, and steam leaks. Following signal preprocessing, statistical features—including the Average Signal Level (ASL)—were extracted to train two models: an ensemble learning model and a correlation-based mathematical model. Both approaches demonstrate a robust relationship between the acoustic signal and the measured leakage flow rate. The results show a prediction accuracy of 92% for the machine learning model and 95% for the mathematical model. This approach holds significant promise for condition-based —and potentially predictive maintenance, enabling the quantification of defect severity, its progression over time, and the optimization of maintenance scheduling.