System monitoring with real-time contrasts

Monitoring real-time data steams is an important learning task in numerous disciplines. Traditional process-monitoring techniques are challenged by increasingly complex, high-dimensional data, mixed categorical and numerical variables, non-Gaussian distributions, nonlinear relationships, etc. A new monitoring method based on real-time contrasts (RTC) between the reference and real-time data is presented. RTC assigns one class label to the reference data, and another class label to a window of real-time data, and, thus, transforms the monitoring problem to a dynamic series of classification problems. This differs from previous work that trained a classifier one time at the start of monitoring. Furthermore, a number of monitoring statistics based on the generalized likelihood-ratio principle are discussed. These include error rates and class probability estimates from classifiers. Typically, the window size for the real-time data is much smaller than the size of the reference data and this class imbalance is also considered in the approach. Variable contributor diagnostics can be simultaneously obtained from the approach and are briefly discussed. Both mean and variance shifts are illustrated. Experiments are used to compare monitoring statistics and to illustrate performance advantages of RTC relative to alternative methods.