LAIK: Location-Specific Analysis to Infer Key Performance Indicators

Abstract

Key Performance Indicators (KPIs) are important measures of the quality of service in cellular networks. There are multiple efforts by cellular carriers and 5G standardization on the use of crowdsourcing to minimize drive tests (MDT) and self-organize the network while improving KPIs via a user feedback loop. Since propagation highly depends upon the environment, readily-available geographical data could be coupled with the crowdsourced user data to infer performance. In this paper, we build a framework to infer KPIs by establishing a relationship between geographical data and crowdsourced channel measurements via neural networks. In particular, for a specific user location, we leverage delay spread measurements in the region to design a cone-shaped filter for the geographical and user data extraction. Then, a location-specific received signal power prediction is obtained via the neural network trained using the extracted geographical and user data. We study the impact of the angle chosen for the cone and various features selected on location-specific KPI prediction. We then leverage the location-specific inference by repeating the prediction over a set of locations in a region to infer the path loss in a given environment. In both types of KPI inference, we compare against state-of-the-art solutions and show that significant improvement in KPI prediction accuracy is achieved using the proposed strategy. Furthermore, for network planners, we show that our framework can use only geographical information to predict KPIs with a negligible error in user locations that lack signal quality data. By employing the proposed framework to predict location-specific and regional KPIs, we achieve an accurate estimation of network coverage and a 7-fold reduction in throughput estimation error compared to a state-of-the-art solution.