Hybrid Kriging and multilayer perceptron neural network technique for coverage prediction in cellular networks

摘要

Coverage prediction is a crucial issue in wireless network planning and design. Typically, coverage prediction techniques start by model identification given a set of measurements at specified locations. Then, this model is used to predict the signal strength at other locations using some machine-learning like approach. However, when the drive tests (or test-beds) become costly because of some environmental constraints, the performance of the machine learning-based model is questionable due to the insufficiency of the training dataset. In this context, we suggest exploiting the geostatistical interpolation technique named Ordinary Kriging to enrich the training data set. For this purpose, a set of received signal strengths from wireless transmitters has been collected at known locations through cellular network technology, which are then used to generate an enriched dataset according to the Ordinary Kriging interpolation technique. The results show that the hybrid Ordinary Kriging and machine learning model significantly enhances path loss accuracy and offers a new setting for data reproducibility.A database is constructed from received signal strength measurements which is not enough to produce accurate results when it is used to train a feed forward neural network for the task of coverage prediction. A geospatial interpolated technique named Kriging is used in order to produce a big amount of data. The neural network is trained by both of the original and the interpolated datasets. The coverage prediction results more reproducible and accurate than the results without the interpolation stage.[GRAPHICS].