Investigations On Cell Outage Detection Techniques For Wireless Networks

To meet the demands on service diversity and efficiency,future wireless networks are propelled to fulfill more intelligence on operation and maintenance.Cell outage detection as an important use case for intelligence on operation and maintenance is a hot research topic.Machine learning techniques provide solutions to leverage and utilize all information collected in the network.Hence,machine learning based approaches for cell outage detection for wireless network are investigated in this thesis.Since the key performance indicators(KPI)data in current networks are untagged,an approach for cell outage detection is proposed based on improved adaptive resonance theory with adaptive threshold.The KPI information reported in user measurement report is first learned competitively in the ART network using similarity as a criterion,and the competition result is compared with an adaptive threshold to determine its initial cluster.During the process,the threshold is adjusted dynamically according to the relationship between the learned samples and the current sample.Finally,a statistically test is adopted to determine the final number of clusters of the initial clustering results,and the outage cell is located by combining the location information in the report.Simulation results show that the proposed outage detection scheme can accurately detect and locate the outage base station by adopting data that contains large amount of information and can be conveniently acquired.Compared with the existing outage detection scheme based on classification method,the proposed algorithm can save the time,labor and material cost for acquiring labeled KPI,and the incremental learning method of ART which continuously learns while retaining old knowledge memory further enables its application in dynamic wireless networks.Since in femtocells,users are sparse and user data can be easily outdated,an approach for femtocell outage detection is proposed based on self-taught learning.Firstly,the sparse autoencoder is used to learn important features from a large number of wireless network KPI datasets with low utilization rate,then the extracted feature is migrated to the heterogeneous network KPI dataset,and then classified and detected by support vector machine method.Besides,considering the imbalance of the heterogeneous network KPI dataset,a pre-processing sampling method is adopted in the process.Simulation results show that compared with traditional femtocell outage detection approach based on classification,the proposed approach can detect femtocell outage more accurately,while the utilization of data in the network is also improved in feature extraction step.Furthermore,in networks of dynamic topology caused by re-deployment and switching on/off of femtocells,the approach based on feature migration method is much more efficient since it utilizes auxiliary data.