This thesis considers the application of changepoint detection methodology for the analysis of acoustic sensing signals. In the first part, we propose a detection procedure for changes in the second-order structure of a univariate time series. This utilises a penalised likelihood based on Whittle’s approximation and allows for a non-linear penalty function. This procedure is subsequently used to detect changes in acoustic sensing data which correspond to external disturbances of the measuring cable.
The second part shifts focus to multivariate time series, and considers the detection of changes which occur in only a subset of the variables. We introduce the concept of changepoint vectors which we use to model such changes. A dynamic programming scheme is proposed which obtains the optimal configuration of changepoint vectors for a given multivariate series. Consideration of pruning techniques suggests that these are not practically viable for this setting. We therefore introduce approximations which vastly improve computational speed with negligible detrimental impact on accuracy. This approximated procedure is applied to multivariate acoustic sensing data.