Statistical algorithms in fault detection and prediction: Toward a healthier network
Benjamin Cheung, Gopal Kumar, Sudarshan A. Rao · Bell Labs Technical Journal · 2005
Very high reliability/availability at affordable cost requires a proactive approach to system faults and failures. This calls for sophisticated fault detection algorithms that ultimately could evolve into fault prediction strategies. This paper presents statistical algorithms — the Operational Fault Detection (OFD) class of algorithms — toward reaching these goals. OFD algorithms analyze system performance metrics to detect fault signatures. The concept behind OFD is to raise alarms for conditions that adversely impact customer revenue or system performance. Initial versions of OFD, deployed in the field, count meaningful events and raise alarms when a test statistic, based on the event counts, exceeds a predefined threshold. Setting the thresholds required human intervention. This is considered time consuming by our customers, even though the concepts of OFD have been well received. This paper suggests a new generation — the second-generation OFD — that is inherently adaptive and requires minimal human intervention. These new algorithms are designed to detect system performance degradations, paving the way to more mature fault prediction strategies. Detecting degradations is a precursor to fault predictions, as degradations are often early signatures of potentially catastrophic faults.