NSF Award
2427909
The problem of real-time monitoring and detecting changes in the statistical properties of multi-stream data has many applications in science and engineering. These include monitoring public health for possible pandemic onset and recognizing abnormal events in multi-camera video surveillance. Often the data of interest is non-stationary, i.e., its statistical properties that change with time. Traditional change detection algorithms are not optimized to process non-stationary data. This project will develop algorithms that are provably robust against uncertainty in data distribution and easily implementable in practice. The algorithms will be further developed to apply in settings where privacy, energy efficiency, and high-dimensionality of data come into play. The developed algorithms will be applicable to solve a wide class of spatiotemporal change detection problems in public health and cyber-physical systems. The algorithms will be validated on several publicly available datasets and the code will be made publicly available. Students will have opportunities to participate in the research and efforts will be made to recruit participants from underrepresented groups. <br/><br/>The algorithms developed in this project will be optimized to detect changes in the statistical properties of multi-stream non-stationary data with the minimum possible delay, subject to a constraint on the rate of false alarms. These quickest change detection algorithms will be designed to be robust against uncertainty in the distribution of data before and after the change. The project is divided into four technical thrusts. The first thrust will develop robust algorithms for quickest change detection when there are multiple streams of non-stationary data with unknown pre- and post-change distributions and the change can occur in any subset of the streams. The algorithms will also be designed to identify the affected stream, at the time the change is declared, and an alarm is raised. The algorithms will be based on the least favorable pair of distributions in the pre- and post-change uncertainty classes. Procedures to analytically characterize or numerically calculate the least favorable pair will also be provided. The second thrust will develop robust algorithms for non-stationary high-dimensional data. The algorithms will be based on the gradient of the logarithm of the density of the data which can be learned using deep neural networks. The third thrust will develop algorithms for data-efficient quickest change detection in non-stationary data. A data-efficient procedure utilizes adaptive sampling techniques to control the average number of observations used before the change. The fourth thrust will develop optimal algorithms for some special classes of non-stationary processes encountered in traffic safety, satellite safety, and public health applications.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
