Locating acoustic emission sources in complex structures using Gaussian processes

A standard technique in the field of non-destructive evaluation is to use acoustic emissions to characterise and locate the damage events that generate them. The location problem is typically posed in terms of the times of flight of the waves and results in an optimisation problem, which can at times be ill-posed. A method is proposed here for learning the relationship between time of flight differences and damage location using data generated by artificially stimulated acoustic emission (AE)—a classic problem of regression. A structure designed to represent a complicated aerospace component was interrogated using a laser to thermoelastically generate AE at multiple points across the structure's surface. Piezoelectric transducers were mounted on the surface of the structure, and the resulting waveforms were recorded. A Gaussian process (GP) with RBF kernels was chosen for regression. Since during AE monitoring not all events can be guaranteed to be detected by all sensors, a GP was trained on data for all possible combinations ( subsets) of sensors. The inputs to the GPs were the differences in time of flight between sensors in the set, and the targets were the locations of the source of ultrasonic stimulation. Subsequent (test) data points were located by every possible GP, given the active sensors. It is shown that maps learned on a given structure can generalise effectively to nominally identical structures.