Modern clustering techniques are becoming more popular, but traditional clustering methods often heavily rely on artificial feature extraction, costing too much human labor resources. This task is especially important when it comes to finding surface defects on medical syringes. To solve this problem, our study comes up with a new way to detect syringe defects without defective samples using deep fuzzy subspace clustering in a smooth way. In particular, our method combines conventional fuzzy c-means and deep autoencoder into a joint framework. During the clustering process, the autoencoder part automatically extract features from dataset itself and generates the representative feature embeddings. After the feature extraction, the fuzzy clustering is performed to find the optimal data structure. This “feature learning, fuzzy clustering” process is iteratively performed until the objective function converges. This new method improves the accuracy and reliability of clustering by a lot, which makes it a great choice for finding scale defects on medical syringes. The efficacy of our method is evaluated on four publicly accessible datasets and one medical syringe dataset. The results demonstrate its potential to revolutionize the field of surface defect detection in this critical domain.