Anomaly detection for fabricated artifact by using unstructured 3D point cloud data
针对非结构化三维点云数据缺乏全局坐标排序、异常建模困难的问题，提出一种概率框架下的贝叶斯网络方法，将点分为参考面、异常和离群三类，用变分期望最大化算法推断类型，仿真和实例验证了准确性和鲁棒性。
3D point cloud data has been widely used in surface quality inspection to measure fabricated artifacts, allowing the high density and precision of measurements and providing quantitative 3D geometric characteristics for anomalies. Unlike structured 3D point cloud data, unstructured 3D point cloud data can capture the surface geometry completely. However, anomaly detection by using unstructured 3D point cloud data is more challenging, due to the nonexistence of global coordinate ordering and the difficulty of mathematically modeling anomalies and discriminating outliers. To deal with these challenges, this article formulates the anomaly detection problem into a probabilistic framework. By categorizing points into three types, i.e., reference surface point, anomaly point, and outlier point, a novel Bayesian network is proposed to model the unstructured 3D point cloud data. The variational expectation-maximization algorithm is used to estimate parameters and make inference on the unknown types of points. Both simulation and real case studies demonstrate the accuracy and robustness of the proposed method.