Experimental study on transient stress wave event detection and pattern recognition of acoustic emission in underwater structures

Abstract

Monitoring runner cracks within the flow passages of hydraulic turbines has always been a significant challenge for the safe operation of the hydropower industry. Acoustic emission (AE) is an emerging dynamic non-destructive testing technology capable of real-time monitoring of structural damage by capturing stress waves generated during material deformation. To investigate the effectiveness of AE and deep learning technologies for online monitoring of hydraulic turbine runner cracks, this study designed a comprehensive simulation experiment that encompassed underwater structural stress-wave propagation, signal acquisition, and signal feature analysis. For the acquired AE signals, an end-to-end network was developed by cascading a parameter-adaptive time-frequency analysis neural network layer with a convolutional neural network (CNN) to perform feature recognition and classification diagnosis. The results indicate that the proposed network exhibits an exceptional capability in interpreting AE signals, accomplishing signal classification and recognition under limited-sample supervised learning. Furthermore, the trained network can accurately locate transient stress wave events within long waveform signals. This verifies that the neural network, while better adapting to the constraints of realworld edge deployment scenarios, possesses the potential to locate and identify abnormal AE signals. Furthermore, it preliminarily explores the theoretical feasibility of a fully automated hydraulic turbine runner crack-monitoring system based on AE and deep learning technologies.