Implementación de un sistema de predicción de medidas LTE utilizando la tarjeta de desarrollo Zedboard y Machine Learning

Abstract

The different conditions under which LTE mobile networks operate make it necessary to establish strategies that allow the interpretation of different variables involved in their design, construction, deployment and operation. One of these key variables within the LTE mobile context are the power ratio and signal quality measurements, RSSI, RSRP and RSRQ. The main methods used to quantify these measurements in a physical environment, beyond on-site measurement, may be propagation models, however, in addition to the mathematical construction involved in their use and implementation, they are not really tools that forecast or predict signal behavior in a specific environment, clearly showing a need that has not been adequately met in this area. In this perspective, a viable option to the prediction of these measurements are predictive models, but if in addition to the problem of predicting itself, the context of working with mobile equipment and outdoor environments is added, it makes sense to route these prediction models in a device that offers portability in some way. FPGA-type reprogrammable boards or embedded systems provide a moderate amount of processing resources that, in applications such as those required for this problem, turn out to be sufficient to implement predictive models. Thus, ZedBoard, together with a dedicated operating system that supports programming languages specialized in Machine Learning such as Python, is a suitable candidate for code prototyping in this field. We propose then the design and execution in an interactive Python code environment, such as Jupyter Notebook, the development of three predictive models based on the algorithms of Linear Regression, Polynomial Regression, and Random Forests, respectively, using as reference the real measurements taken from an LTE base station on a mobile terminal. Additionally, a predictive model based on Recurrent Neural Networks is developed outside the ZedBoard. The purpose of this is, firstly, to diversify the alternatives in terms of Machine Learning algorithms used; and also, to technically support that the effectiveness and usefulness of the use of other embedded models that apparently do not have the necessary power in this type of applications, are also capable of offering complete solutions in an area in which they are not usually considered. At the end of the design and deployment on the test data, the predictions made by each model are conclusive and reaffirm the real possibility of implementing predictive models based on Machine Learning techniques in embedded systems, and within the context of mobile telecommunications. For this specific problem, the Random Forests model proved to be far superior and adequately adapted to the characteristics of the LTE variables analyzed with an error margin of less than 6% in any case, even above the Recurrent Neural Network model that was not executed in ZedBoard, allowing to deduce that there are other options in relation to the prediction of mobile signals and offering new alternatives to the classical methods of planning, dimensioning and characterization of an LTE cellular network.