In robotic surgery, technical skill is a major challenge for surgeons and trainees. Although research in surgical skill assessment has made considerable progress, most of the methods are post-operative analysis, and only a few methods are intra-operative or real-time analysis. In this study, we introduce a method to extract the unusual movements which are rarely seen in Experts and identify the types of the unusual movements using IMU data and a semi-supervised learning approach. Our previously collected dataset includes 12 subjects (3 Experts: EX, 2 Fellows: FL, 3 Intermediates: IN, and 4 Novices: NO) performing surgical training tasks on a da Vinci Simulator. We used the data from Experts to train an Autoencoder to reconstruct the input data. An unusual movement will show a larger reconstruction error - a data pattern that Autoencoder has never seen from EX. NO showed significantly higher number of unusual movements and average error of the unusual movements than EX, FL, and IN, indicating NO had the worst performance, thus validating our method. Then we identified the types of unusual movements using clustering. We could see large wrist flexions, extensions, and deviations in these unusual movements. In this study, we introduced a method to extract the movements which were not observed in EX data, and we found that the use of wrist is a good indicator of surgical expertise level. We will seek for methods to help surgeons avoid these behaviors.