In the field of industrial anomaly detection, 3D data anomaly detection is gaining increasing attention. 3D data can detect anomalies, such as surface depressions and bumps, which cannot be identified by 2D images from multiple angles. For this, we propose an innovative anomaly detection method with knowledge distillation for 3D point cloud data. Specifically, our approach involves training a set of student-teacher models based on masked autoencoders and then detecting anomalies by evaluating the differences between the output features of the student and teacher models. To address the generalization issues caused by the structural similarity between the teacher and student models, we propose a hierarchical structure to make the teacher model and student model asymmetric. To improve the model’s capability to learn the irregular and challenging aspects of the point cloud data, we propose a novel contrastive loss. This loss function optimizes the model by constraining the parts of each sample with the greatest feature expression differences. Additionally, we adopt an anomaly score computation method that is consistent with the contrastive loss. Extensive experiments demonstrate that this method performs well on benchmarking MVTEC 3D-AD dataset.