A0-delta-T deep learning framework for reliable acoustic emission damage location in plate-like structures

Abstract

Acoustic emission (AE) damage location methods rely strongly on the estimation of wave arrival times. These arrival times corresponding to the fast symmetric modes are generally susceptible to noise. In contrast, the subsequent antisymmetric modes with higher out-of-plane energy could provide relatively accurate arrival times through time–frequency analysis, which still has inherent errors. Aiming at more effective damage location of large-scale plate-like structures under noise interference, an A0 mode arrival time difference (A0-delta-T) deep learning framework was developed by integrating continuous wavelet transform (CWT), position embedding multihead self-attention long short-term memory (PE-MHSA-LSTM) network, and genetic algorithm (GA). For data preprocessing, CWT extracted coefficients at the dominant frequency of A0 mode. Then, a PE-MHSA-LSTM model was trained to identify the A0-delta-T of the sensor pair from corresponding CWT coefficients feature segments. GA was utilized to optimize the error function related to A0-delta-T and estimate the damage source location. The experiment conducted on a large-scale steel plate specimen demonstrated that the proposed framework could achieve a reliable performance of AE source location at various noise levels. Such a robust noise resistance highlighted its potential for field applications in engineering structures.