Abstract

Currently, telecommunications companies seek to optimize wireless communication networks. This work proposes a practical method for classifying the coverage of the Long Term Evolution (LTE) network based on measurements of the Reference Signal Received Power (RSRP) parameter. Initially, the RSRP is captured by an industrial router, and a Federated Learning (FL) approach based on The Federated Averaging (Fe dAvg) algorithm technique is applied. Three classes of coverage are established: “poor”, “good” and “regular”, represented by classes 0,1 and 2 respectively. Analyzing the influence of the RSRP probability distribution on the global model’s accuracy, it is shown that a uniform distribution allows for achieving higher accuracy in fewer update rounds and shorter convergence times. This is in contrast to a non-IID distribution, where each random variable does not have the same probability distribution and is not all mutually independent. The results obtained confirm the potential of federated learning for coverage classification in LTE networks, preserving data privacy by not sharing it directly between devices.