Background: In this era of mobile wallet platforms, to ensure key requirements like high availability and performance, the company must have mechanisms in place to detect anomalies at any given point in time. Ericsson Wallet Platform(EWP), a mobile wallet platform, is facing the problem of manually analyzing all the logs and reports and taking comprehensive action decisions accordingly. Therefore, a need for more reliable techniques in nailing down the exact root cause of a given trouble hasarisen.

Objectives: The objective of this thesis work is to automate the process of troubleshooting through Root Cause Analysis(RCA) to the maximum possible extent given our experiences and resources provided. The idea is to help domain experts save a lot of effort and time while performing RCA on the application. Every step through the process of RCA in EWP will be considered and the complex and time-consuming step is to be automated with the help of data-driven Machine Learning(ML) techniques.

Methods: Various formats of datasets from the EWP application are analyzed to understand the application data functionality. A literature review is conducted to comprehend which type of ML algorithms will work best for performing anomaly detection on the datasets of the application. We compared various unsupervised ML algorithms in the process of achieving accurate anomaly detection for each dataset. An interface is built to investigate the inter-relation between the anomalies in time series to find the root cause and to provide easier access.

Results: The performance results of each kind of ML algorithm for anomaly detection are evaluated using metrics Accuracy, Precision, Recall, and F1 Score. This proved K-means(optimized using Isolation Forest) and One-class Support Vector Machine(OCSVM) to work as the best models for Wait Reports and Performance logs respectively. The efficiency of these models is improved by using Random Forest Classifier. The interactive Graphical User Interface(GUI) built for our combined model has successfully placed the output graphs for correlating the anomalies in the EWP application.

Conclusions: A detailed study of the EWP application data logs was conducted and repeated experiments were performed to determine the best anomaly detection models for different data sets. An interactive GUI has been built to visualize and correlate these anomalies in time series in order to find out the chain of events causing the problems. The scope to continue our work has been penned down in the future work section.