Abstract:
|
Monitoring processes of physical or administrative systems frequently include unusual and unexpected observations. In a large, expensive experimental setting or in a large throughput data processing center, it is important to identify the anomalies and the associated circumstances. We provide a general solution to flag and adjust anomalies in a regularly-observed process. Depending on the presence of prior information and the distributional characteristics of a process, we vary the required minimum length of the process. In application to a large experimental setting, the final goal is to provide just-in-time anomaly detections and alerts; in an administrative process control setting, a final outcome is to suggest adjustments to correct the surprises.
|