In this paper we tackle the problem of eNodeB failure detection in LTE networks using Binary Classification techniques under smart-cities Machine-to-Machine (M2M) traffic. We train 20 different classifiers with data from two 24 hrs simulations with different traffic volume levels. Input features for the classification models are built aggregating packet generation and access collisions from the eNodeB on which failures are being detected, as well as from its closest neighbors, by computing statistics for each time-bin. Network service providers generally maintain network performance data by processing real-time data to produce periodic aggregated summaries, which in practice constitutes a filter on the data, reducing the quantity of information available for inference. We explore the effect of different levels of granularity in data aggregation and their effect on our ability to detect failures. We gathered data from M2M traffic along an LTE network simulated using publicly available geographic city data on Montreal, Canada. With Linear Support Vector Machine and Bagged Decision Trees, failure detection rates above 97.5% were achieved, with false positive rates under 2.8%, showing that, even in 30 minutes aggregations, it is feasible to extract meaningful data from aggregations of data from LTE networks with M2M traffic.