Early detection of fatigue crack-growth in steel structures is an ongoing challenge. Furthermore, characterization of the different stages of the fatigue lifecycle using NDE techniques is particularly difficult. AE systems have been shown to serve as early damage detection mechanisms in bridge structures. This technology, however, is fraught with noise problems and complex datasets that are difficult to interpret. This paper attempts to design and implement a data mining scheme that can classify raw AE datasets into discrete clusters using an improved variant of the popular k-means clustering algorithm. The datasets are then augmented with the class label found during clustering, and a series of rules are inferred using a C4.5 decision tree classification algorithm. An implementation of the data mining scheme is coded in MATLAB®, with data from PAC® AE systems as the input. In order to validate this procedure, data from a pencil lead break test with a concurrent noise source is fed into the data mining program. Classification using the decision tree is compared to manual classification of the pencil lead break hits. The resulting decision tree is then applied to a similar dataset in order to evaluate the generality of the resulting rule sets. Once validated, the data mining program is applied to data belonging to a steel fatigue crack-growth test. Results of this classification are discussed, and possible improvements to the data mining scheme are suggested.