Fast Semiautomatic Change Point Detection in Deformation Monitoring Using Kullback–Leibler Divergence

Automated change point detection in deformation monitoring data is desirable for engineering applications (e.g., operation, maintenance, and early warning). This study proposes a novel change point detection strategy using the Kullback–Leibler divergence metric. Instead of calculating the true divergence (a computationally intensive task), the proposed strategy operates in Gaussian space, in which closed-form expressions for Kullback–Leibler divergence exist; therefore, it is fast. The proposed strategy requires only one manual input parameter (sliding window size); therefore, it is semiautomatic. The strategy can handle non-Gaussian data, multivariate data, variance change, and trend change. Four synthetic monitoring data are utilized first to evaluate the capability of the strategy. Benchmarked against a closely related method based on Pearson divergence, >1,000× acceleration of the run time is achieved with similar or better accuracy in terms of true and false positive detection rates. A monitoring data set of 2,000 data points can be processed in only ≈0.03 s (≈45 s for the benchmark method). Four real monitoring data were utilized to demonstrate the change point detection. In a drilling data set, distinct rock layers were identified correctly. In an interferometric synthetic-aperture radar (InSAR) data set, ground deformation due to volcanic eruption was identified. Finally, four slope deformation monitoring data sets from New Zealand, Japan, and Italy were utilized to identify the onset of acceleration point for early warning of slope failure. MATLAB codes are provided for future applications.