Fault detection in electromechanical systems plays a significant role in product quality and manufacturing efficiency during the transition to smart manufacturing. Because collecting a sufficient number of datasets under faulty conditions of the system is challenging in practical industrial sites, unsupervised fault detection methods are mainly used. Although fault datasets accumulate during machine operation, it is not straightforward to utilize the information it contains for fault detection after the deep learning model has been trained in an unsupervised manner. However, the information in fault datasets is expected to significantly contribute to fault detection. In this regard, this study aims to validate the effectiveness of the transition from unsupervised to supervised learning as fault datasets gradually accumulate through continuous machine operation. We also focus on experimentally analyzing how changes in the learning paradigm of the deep learning model and the output representation affect fault detection performance. The results demonstrate that, with a small number of fault datasets, a supervised model with continuous outputs as a regression problem showed better fault detection performance than the original model with one-hot encoded outputs (as a classification problem).