Bayesian Changepoint Modelling for Reference-Free Damage Detection with Acoustic Emission Data

Acoustic emission testing remains a popular technique in the SHM community due to its effectiveness in detecting and localising small occurrences of damage in a structure. One of the characteristics of acoustic emissions is that all structures will produce some level of emission even if only benign flaws are present. As such, it is often necessary to collect a baseline reference set of data against which newly collected measurements can be compared to establish the state of the structure. In this work, a methodology is proposed which eliminates the need for this “training phase” through use of a statistical model which learns and adapts online. The chosen technique is a Bayesian online changepoint detection method where bursts of acoustic emission are modelled as a Poisson point process. In this way the inherent stochasticity in the number of “hits” emitted in a given window is learnt and modelled online, then significant changes in the properties of the generating stochastic model are used to provide sensitivity to damage. This approach has a number of benefits which are demonstrated on an experimental dataset from a bearing test rig. The main benefit is removing the need to collect extensive data before the SHM system can become operational. The proposed solution also allows prior engineering knowledge to be exploited, for instance by specifying priors related to the expected number of hits in a given window, or through use of a Hazard Function which encodes prior belief about the expected time before damage will occur. Finally, the method is shown to characterise the full probability distribution over possible run lengths. This information provides not only an indication of if a significant change in the behaviour of the system has occurred, but also automatically quantifies a degree of confidence in that change.