Interest in damage detection and damage pattern recognition of engineering structures by non-destructive techniques has been increasingly growing. As a non-destructive technique, acoustic emission (AE) has developed rapidly to detect dynamic defects and their evolution behaviors of composite structures, based on the transient elastic waves produced by rapid energy release due to the geometry change of structures. In this paper, AE technology is utilized to monitor the real-time condition of the composite scarf joint (SJ) under tensile loading. First, after AE signal acquisition, dimensionality reduction of eight AE features is realized by employing principal component analysis such that the Curse of Dimensionality can be avoided. Second, feature selection is continued by introducing two evaluation indexes, i.e., correlation coefficient and Laplacian score. Third, after the optimal cluster number is determined, damage pattern recognition is accomplished by introducing k-means++ algorithm which explores the distribution of each pattern in the space constructed by four informative AE features. Based on the clustering results, damage initiation and evolution in SJ specimens under tensile loading are subsequently explored. The shear failure of the adhesive layer which is a characteristic damage pattern for SJ specimens shows a relatively-high activity after the early stage. Matrix cracking and fiber/matrix interface debonding are two fundamental damage patterns which keep active in the whole process.