FAULT DETECTION AND EVENT PREDICTION IN NETWORKS USING FEATURE MATRIX.

A data features selection and organization scheme is presented for future event predictions using historical event data. Most of the reported event prediction methods use data features in the form of time series. The proposed approach uses simple data feature statistics. First an event is defined based on historical event data features for all discrete time instances, then all events are ordered chronologically and divided into N time- windows with an overlapping interval. The probabilities (relative frequencies) for occurrence of all sequences of two events are calculated for each time-window and stored in a 3D input feature matrix. A prediction technique is trained using probabilities from one time-window to learn to predict probabilities for the next time-window. Once the prediction technique is iteratively trained for all N time-windows of the training data, it is used to predict future probabilities. The accuracy is calculated by comparing with the known test data probabilities. The proposed approach is tested for fault prediction in telecommunication networks using various Artificial Intelligence Techniques.