Core Concepts
A novel untrained anomaly detection method based on 3D point cloud data that can achieve accurate anomaly detection on complex manufacturing surfaces using a single sample without requiring any training data.
Abstract
The paper proposes a novel untrained anomaly detection method for complex manufacturing surfaces based on 3D point cloud data. The key challenges addressed are the direct modeling of complex surfaces represented by 3D point clouds and the inability to train many machine learning and statistical models when only a single sample is available.
The proposed approach transforms the input 3D point cloud into a collection of profiles along different directions, inspired by the design and manufacturing process of rotating bodies. These profiles are expected to have similar shapes, indicating that they can be represented by a low-rank matrix. However, the unstructured nature of 3D point clouds introduces error points that can deviate significantly from the expected profiles.
To address this, the paper introduces a Component Segmentation and Cleaning (CSC) module that segments the complex surface into multiple basic and simple components. This enables the removal of outlier profiles near component boundaries, and allows the Robust Principal Component Analysis (RPCA) algorithm to be applied to each clean component, effectively modeling the low-rank reference surface and detecting sparse anomalies.
Extensive experiments on different types of complex manufacturing parts demonstrate that the proposed method outperforms state-of-the-art approaches, achieving promising results for accurate 3D anomaly detection using only a single sample without any training data.
Stats
The depth d and radius r of hole anomalies are d ∈[0.9, 1.1], r ∈[0.2, 0.5] respectively, and the length l, width w and height h of the scratch anomalies are l ∈[0.6, 1.2], w ∈[0.5, 1.1], h ∈[0.1, 0.2].
The bounding box sizes of Part1, Part2 and Part3 are (12.8, 10.0, 5.1), (12.8, 8.0, 3.4) and (28.5, 8.8, 3.6) respectively.
Quotes
"The surface quality inspection of manufacturing parts based on 3D point cloud data has attracted increasing attention in recent years."
"Achieving accurate 3D anomaly detection is challenging, due to the complex surfaces of manufacturing parts and the difficulty of collecting sufficient anomaly samples."