Article Preview
Top1. Introduction
The Network Management System (NMS) of telecom network has made it possible to deal with everyday events. There are a large number of alarms collected by network devices in NMS, which is the basis of follow-up fault diagnosis. Ideally, every alarm should draw the attention of an operator, but many alarms may occur almost simultaneously to overwhelm experienced operators. The situation is well known as alarm floods or alarm showers. Alarms do not always contain accurate information about network faults clearly, that is, the fault may trigger a flood of alarms produced by other devices connected to the device when a fault occurs, and not all of these alarms can instruct the root cause of failure, resulting in that critical alarms are buried under unnecessary ones. Therefore, alarm management has also become a focus in industrial systems of different areas. Stern etc. proposed a method of reducing the size of alarm sets by taking advantage of redundant information and applied it to alarm data of three large Natural Gas Processing Plants (Soares, Pinto, & de Souza, 2016). Jacobs and Dagnino (2016) introduced a novel graph-based data mining approach that can be used to analyze industrial alarm data and used the experiment on a power generation station and oil refinery datasets to demonstrate scalability and efficiency of the method. Allan etc. proposed a set of practices for monitor alarm reduction without undermining patient safety in intensive care unit to improve patient safety by optimizing alarm systems. (Allan, Doyle, Sapirstein, & Cvach, 2017) Alarm analysis has also been recognized as an important problem in system monitoring and faults detection. (Kondaveeti, Shah, & Izadi, 2009) (Yang, Shah, Xiao, & Chen, 2012) (Akatsuka, Noda, & Sugimoto, 2013)
Researchers have focused on ‘alarm correlation’ to reduce the number of alarms and at the same time find out all possible faults. A typical thought of alarm correlation is to establish an alarm knowledge base and determine the maximum possible failure or faults. A few alarm correlation algorithms have been proposed, including Rule-Based Correlation, (Cronk, Callahan, & Bernstein, 1988) Case-Based Reasoning, (Slade, 1991) Fuzzy Logic, (Zadeh, 1996) Bayesian Networks, (Heckerman, Mamdani, & Wellman, 1995) and Neural Networks. (Gurer, Khan, Ogier, & Keffer, 1996) The above methods satisfy small-scale network and it is not enough to meet the needs of network maintenance only through expert knowledge. So, some researchers (Manganaris, Christensen, Zerkle, & Hermiz, 2000) (Julisch, 2002) focused on data mining to analyze alarm sequence and dig related rules, confidence is calculated to obtain the correlation rules. Weiss described how data mining can be used to uncover useful information buried within nuisance alarms, present several data mining applications and demonstrated that data mining can be used to identify network abnormal situations. (Weiss, 2005) Mannila presented the efficient algorithm (WINEPI) for detection of frequent episodes of a given number of episodes. (Mannila, Toivonen, & Verkamo, 1997) This algorithm is applied in telecommunication alarm management in TASA. The data mining method focuses on related rules to find out faults in the network but fails to show visual display to operators. Taniar proposed exceptions rules mining algorithm to generate candidate exception based on the knowledge about negative association rules in the database, which can also be applied to find out abnormal alarms and diagnose faults. (Taniar, Rahayu, Lee, & Daly, 2008) (Daly, & Taniar, 2004)