Integrated Error Detection at Software Launches with the Use of Machine Learning

Sammanfattning

During installation of firmware updates over the air, servers will usually collect data with statuscodes and information about the installation and update process. Given a large dataset with multiple installation steps, errors might occur occasionally dependant on outside factors such as internet connection, limited storage space or other issues that are not directly related to the stability of the update. Some updates however might have more issues in certain areas of the update process compared to the average. This thesis introduces an integrated system that detects anomalies on specific updates to determine the software stability of a released update. The system utilizes the unsupervised machine learning model isolation forest on device aggregated data to highlight anomaly devices and make determinations about an update’s software stability. Using the isolation forest model with device aggregated data a high accuracy, recall, and precision metric is achieved while computing within a reasonable time.