Acoustic emission source location based on machine learning and Bayesian inversion

Acoustic emission (AE) source location plays a crucial role in structural health monitoring, where P-wave travel time-based location methods are the most commonly employed. However, the modeling accuracy of P-wave travel time using uniform, 1D, or even simple 3D velocity model in complex structures remains limited, resulting in low location accuracy. To address this issue, a method combining machine learning (ML) and Bayesian inversion is proposed. Firstly, a Back-Propagation Neural Network (BPNN) model is used to establish the nonlinear relationship among AE source, sensor location, and P-wave travel time difference (TTD), ensuring accurate travel time estimation. Then, the TTD data is embedded into a Bayesian inversion framework, and the Markov Chain Monte Carlo (MCMC) algorithm is employed for global sampling and source location, effectively avoiding local optimum issues in traditional location methods. Synthetic tests on a circular hole-contained structure show that the proposed BPNN-Bayesian method achieves an average location error (ALE) of only 0.10 cm for noise free data, and 1.53 cm after 2 ms Gaussian noise is added. In pencil-lead break (PLB) experiments, the method achieves an ALE of 0.54 cm, outperforming traditional BPNN (ALE = 1.90 cm), Kriging (ALE = 0.62 cm), and Inverse Distance Weighting (IDW) (ALE = 1.52 cm)-based methods. It also surpasses shortest path algorithms like straight-line and A*-based methods. Moreover, field tests on eight blasting events yielded an average location error of 42.42 m. The proposed method offers a promising solution for AE source location in complex structures.