Detection of periodic problems using non-parametric density estimation

A significant number of chronic GSM quality faults cause chronic quality alarms, which happen periodically, repeat often and disappear after a short time. There are a huge number of quality alarms and alarm quality experts who are responsible for the fault management try to determine chronic quality faults by examining quality alarms. The task of examining the history of a huge number of alarms and determining the chronic quality faults is a very time consuming and expensive task. In this study, a new non-parametric density estimation approach is proposed to determine quality faults which cause periodic alarms. We use the alarm history to do a nonparametric density estimation using histograms, based on the source, type and reoccurrence information for each alarm. These histograms were examined by the experts and they labeled the alarm histograms based on whether they point to chronic quality problems or not. We built a model to determine chronic quality problems and learned the parameters of this model using the alarm histogram data labeled by the experts. The learned model had a very good performance in determining the chronic quality faults.