We use machine learning algorithms to detect the crystalline phase in undercooled melts simulated via molecular dynamics. Our classification method relies only on the analysis of local conformation and environmental fingerprints of individual monomers. We employ self-supervised autoencoders to compress the fingerprint information, coupled with a Gaussian mixture model to distinguish ordered states from disordered ones. The method does not require explicit information on the expected phases in the system, but automatically detects the local signatures of alignment and stretched conformations in the neighborhood of the monomer, and thereby determines a decision boundary between two classes. We demonstrate that the classes identified by the method correspond to a large extent with the result of classifiers based on human-intuitive order parameters such as the stem length [Luo, C.; Sommer, J.-U. Macromolecules 2011, 44, 1523]. We also show that the machine learning result may facilitate the determination of decision boundaries for simpler classifiers that are commonly used for such systems. In the present system, we thereby enhance the resolution of detailed time patterns of crystalline order before an apparent signature of the transition is visible in thermodynamic properties. At a time point before the transition is manifested in the specific volume and specific heat, a sudden stabilization of monomers in the crystalline phase is observed.