The characteristic acoustic emission (AE) signals generated as a result of different damage mechanisms in the material were analyzed. The signals were obtained from bending of glass (GFE) and carbon fibre epoxy (CFE) composites. Characterization of signals in the time or frequency domain was not sufficient for the formation of distinct clusters in the space of feature vectors. For this reason, new features with time-frequency domain content were extracted from deep convolutional autoencoder. After removing the isolated points by the DBSCAN method, the formation of more distinct clusters was achieved with respect to the original feature vectors. The unsupervised classification of the feature vectors was performed separately for each composite using the K-means method. Comparison of the extracted feature vectors of high-frequency characteristic signals of GFE and CFE composites revealed diverse clusters that were adequately sorted using the spectral clustering method.