Most of the current available patent classification systems are too general to TRIZ users who pay more attention to find analogous patents in other fields that have solved the similar technical problems by using the same solutions. In this paper, we propose an approach to automatically classify patents oriented to TRIZ applications based on a personalized classification schema. Firstly, we construct a personalized classification schema in micro-meso-macro levels. Micro-level is composed of Subject-Action-Object (SAO) extracted from patent text, meso-level Problems solved and Solutions used (P&S) topics generated by Latent Dirichlet Allocation (LDA) based on SAO and macro-level technology domain generated by LDA based on P&S topics. Then, we choose an appropriate feature and classifier to preliminarily classify patents according to the personalized classification schema. Finally, the classifier is optimized by smoothing imbalanced data and reducing features dimensions of SAO. We evaluated the approach with Large Aperture Optical Elements (LAOE) patent documents set as a case study. The results of case study showed that this approach can classify patents with high accuracy and speed and facilitate TRIZ users to better utilize patents in medium size data set.