Proposta de um método de identificação de anomalias em redes móveis com base na representação dimensional de KPIs usando PCA e clustering

Abstract

The complexity, coverage and amount of data produced by mobile networks have required new methods for a more automated analysis and searching for patterns of non-obvious anomalies. In this sense, this research proposed an anomaly identification method based on dimensionality reduction and clustering techniques, applied to a data set composed of performance indicators. This method employs the PCA and DBSCAN techniques, which are applied to a hour segmented real dataset. A second method, based on the structure of the proposed method, but with SOM and LoOP techniques was presented. This method can be seen as a reference method for comparing the proposed method, in order to validate the results found. The methods were evaluated using a data set with six indicators of an LTE network in operation. To investigate the performance and sensitivity of the methods, synthetic anomalies were inserted based on some degradation scenarios characteristic of mobile networks. In addition, network elements with a high number of anomalies were assessed by two qualified professionals. The methods performed well in the identification of the inserted synthetic anomalies, however, some limitations in the reference method regarding parameterization and the definition of a threshold of the techniques hinder their use with other data sets. Finally, the results encourage the use of methods as a tool to guide the correction of failures, optimize resources and indication of network capacity expansion.