Research On Alarm Fuzzy Association Rules Mining In Multi-domain Distributed Network

When the network fails to work, it will trigger and produce large amounts of alarms between adjacent nodes, there must be some correlation between these alarms, the core of network fault diagnosis is the alarm correlation analysis, compressing redundancy alarm to locate the root alarms, actually it’s the process of inference using the network alarm association rules. Data mining technology can solve the bottlenecks of the source of association rules in correlation analysis and reduce reliance on managers and experts. It exists a causal relationship between the network alarms and the cause of network failure, but due to the complexity and heterogeneity of network, this causal relationship is not a simple mapping to determine, on the contrary, it appears as random and fuzzy. The traditional Boolean logic is difficult to solve the fuzzy relationship between alarms and faults. On the other hand, considering the large communication networks tend to be divided into different domains of management, so in this distributed architecture, we not only need to mine the alarm fuzzy association rules in the sub-domain, but also to mining association rules between domains.Aiming at the above problems combined with the background of the multi-domain distributed network, this thesis combined association rule mining techniques in data mining with fuzzy theory to mine the alarm fuzzy association rules within and between domains of the large-scale communication network for alarm correlation analysis. Specific research and innovation are summarized as follows:1. Due to the feature that the original network Alarms don’t apply for the association rule mining, the thesis extracted alarm information fields to establish a unified information model of alarm. It presented a policy that each sub-domain using the time window and sliding step mechanism to synchronize the establishment of alarm transaction library combined with the large-scale communication network having multi-domain distributed structural characteristics.2. Quantified the alarms information in alarm transaction library, then aiming at the issue that the effect of the traditional fuzzy C-means clustering algorithm in dealing with large-scale alarm data is unsatisfactory and the clustering results and convergence speed of FCM clustering algorithm is sensitive to the initialization of clustering centers, this thesis proposed that we can cluster the data roughly using the Canopy clustering algorithm first to obtain a value of k and k Canopy center vectors,3. then sorted the k Canopy center vectors before a further and fine clustering to fuzz the quantized alarm data using CPFCM algorithm based on Hadoop platform, After further processing of the alarm items, then we got an effective fuzzy alarm transaction library consists of frequent 1-itemset. Simulation results show that: MapReduce of CPFCM algorithm has better clustering performance and faster running speed when compared with MapReduce of FCM clustering algorithm, and the fuzzy results using CPFCM algorithm is better when they are used in the alarm fuzzy association rule mining.4. Aiming at requirement of multiple scans about the database and the low time-efficient defect on Apriori-based fuzzy association rules mining algorithm, the thesis used the fuzzy association rules mining algorithm based on Fuzzy FP-tree(fuzzy frequent pattern tree) to mine alarm fuzzy association rules, combining the multi-domain distributed characteristics of the network it improved the mining algorithm ulteriorly, and then it proposed local-local Hadoop-based Fuzzy FP-tree parallel algorithm for mining alarm fuzzy association rules and global-local Fuzzy FP-tree-based parallel algorithm for mining alarm fuzzy association rules. the simulation verified that both algorithm has good time efficiency and can mine alarm fuzzy association rules of the multi-domain distributed network effectively accurately by providing alarm correlation data for fuzzy inference module of network fault diagnosis system. And then the thesis analyzed the advantages, disadvantages and usage scenarios of the two algorithms further.