On Novel Machine Learning Approaches for Acoustic Emission Source Localisation: A Probabilistic Perspective

With the objective of making engineering infrastructure safer and more cost-effective to operate and maintain, the use of automated strategies for monitoring damage in structures and high value assets are becoming increasingly common. A critical component in the assessment of a structure’s condition is the localisation of defects, with a promising solution the monitoring of acoustic emissions, a technique concerned with passively listening to ultrasonic signals generated by damage mechanisms. With that said, a significant barrier to a more widespread adoption of techniques of this nature are their use in structures with intricate geometrical features and anisotropic materials. In these structures, propagation paths are complex, material parameters often unknown, with stochasticity and a deficiency in complete physical understanding introducing sources of uncertainty that are often unaccounted for. The work contained in this thesis develops and extends a probabilistic framework for localising acoustic emissions in complex structures, handling uncertainty in a principled manner through Bayesian inference. A forward mapping of expected arrival time information is first learnt through the use of Gaussian process regression. For an event with an unknown origin, it is shown that these maps can be used to quantify a likelihood of emission location, providing probable damage locations on the structure. Next, the use of a heteroscedastic noise model is presented, allowing predictions made by the localisation model to be locally-weighted such that sensors contribute to the prediction relative to the quality of coverage offered, returning a more accurate, confident and robust localisation methodology. On the topic of the practicality of implementing the proposed approach, the inclusion of physical insight is considered within a grey-box framework to constrain the Gaussian process to abide by known physical laws. It is demonstrated that the constraints improve performance where the availability of training data reduces, increasing the feasibility of implementing the developed methodology. Finally, localisation is extended to cases where the geometry is not most appropriately characterised in Euclidean space, such as for roller-element and many other types of bearings. It is demonstrated how localisation may also be performed in a condition monitoring setting, as well as demonstrating the ability of the method to handle measurements that are contaminated with significant noise levels.