This paper proposes an approach to analyze operation records of axle counters, a core part of railway infrastructure. Our aim is to introduce an efficient way to automatically extract knowledge regarding failures of such devices. As the data provided does not contain a ground truth regarding causes of failures, failure information and their causes should be extracted from underlying relations between recorded events. After a data pre-processing step, the recorded events are clustered with respect to the relationships that can be highlighted among them. As a result, classes of events can be highlighted, from which a classification system can be proposed. Beyond this specific application, the approach is a novel way to tackle reliability analysis problems.