Degradation detection for telecommunication networks using time varying binomial distribution models

We deal with the degradation detection problem in telecommunication gateways. The time series to be monitored is non-stationary but almost periodic (pseudo-periodic). The authors propose a technique called “optimization of partition model” which generates local stationary models for a pseudo-periodic time series. The optimization is based on the minimal Akaike´s (1974) information criterion (AIC) principle. The technique called the sequential probability test (SPRT) is also applied to make efficient decisions. Experiments to evaluate methods for optimization and incremental model update are conducted with real data. The result shows the proposed method is effective and makes more precise decision than the conventional one