The Research Of Fault Detection For Telecommunication Network Based On Datamining

In the area of telecommunication network management, fault management is an important and challenging task. The sticking point in fault management is fault detection, which relies on the knowledge about faults, especially, about the relation between alarms and faults. This necessary knowledge can be acquired through analyzing and interpreting the alarm information. So fault location is the aim and alarm information analyzing is the approach. Currently the main measure for fault management is using alarm correlation system, which is an expertise system. But the complexity of the telecommunication network leads to acquiring necessary knowledge to construct a correlation system for a special net is very difficult. Consulting the application model of TASA system, the thesis based on the project "The research of telecommunication network fault management system based on knowledge discovery", which is a tackle-key-problem project of scientific research in ChongQing, chiefly studies datamining technology used in constructing an alarm correlation system for network fault detection. According to the proposed datamining model, the thesis builds the datamining system model for telecommunication alarm database. The whole process of datamining is also described in this paper, including data collection and preprocessing, rule discovery, rule post-processing and application. Some advices are given for certain existing problem.On the basis of fully analyzing current algorithms for mining frequent episodes from telecommunication alarm data, the thesis extends the current episodes rule discovery algorithm and makes it capable to find episodes rule with two-time bound and can be applicable for large set database. In order to make the discovered patterns fit for fault detection, mining multi-dimension episodes is proposed in the thesis. The discovered multi-dimension episodes rules have the partially ordered relation in time as single-dimension episodes do, at the same time, they have more information connecting with the devices with which have some thing wrong, so that multi-dimension episodes rules are much more fit for fault detection than single episodes rules do. The thesis reduces pattern matching time and memory required in the running process through converting the events properties from multiple dimensions to single dimensions.On the basis of fully analyzing classical algorithms for association rule discoveryand in order to mine the alarm data, the thesis puts forth a new algorithm fitting for mining longer patterns, which recursively dynamically decomposes the search space into small independent chunks. The algorithm bases on founded lattice theory, greatly reducing the number of candid pattern and avoiding pattern match of string. Besides, the algorithm utilizes binary bit vector compactly stored in memory. Operation '&' computes 32 bits once to produce candid itemset's binary bit vector and reduce the length of bit vector by heuristic method which advances capability a lot both in time and memory consuming. The new algorithm is experiment against FP-growth algorithm and the result is given. At last, the thesis analyzes in briefly the other technologies used in the proposed model.