Random forests‐enabled context detections for long‐term evolution network forrailway

Abstract

An ensemble learning-based approach for context detection for high-speed railway (HSR) is proposed, evaluated, and compared against various machine learning algorithms. The context information is collected and formatted from the realistic physical data, which are measured by the field test in a commercial 4G cellular network on-board high-speed train in an open area, HSR station, and typical urban area. The radio propagation knowledge implies that the channel model is implemented in feature extractions. The independent out-of-bag errors show ensemble learning approaches; especially, random forests-based algorithm achieves very accurate context detection (up to 93.5%), which is much higher than single tree (66%). Other ensembles with sub-spaces of discriminant (67%) and K-nearest neighbour (69%), as well as linear discriminant analysis (55.7%), and support vector machine (SVM) (27.7%). Furthermore, the features are selected based on the feature importance evaluation in the first round training. The selected predominant features and the reasonable number of individual trees construct the refined random forests, which obtain a high accuracy (92.6%) and 70% time reducing. This experimental study benefits from physical radio knowledge to support advanced long-term evolution (LTE) network for railway and future smart rail applications.