A Semi-Nested Point Cloud Autoencoder for Solder Joint Inspection

As a One-class learning method, autoencoder has excellent performance in a variety of anomaly detection tasks, including one-dimensional sequences and two-dimensional image data. However, there are few studies on one-class classification methods for 3D data. In particular, in actual working conditions, due to improper assembly and other reasons, the target point cloud pose is often random within a certain range, it is difficult for the autoencoder to fit the generated point cloud to all poses of the target sample, which leads to performance degradation and thus brings additional challenges. To improve this disadvantageous situation, a uniquely designed Semi-Nested Point Cloud Autoencoder (SN-PAE) is proposed. SN-PAE first uses a rotation-invariant point cloud autoencoder to learn rotationinvariant descriptive latent variables, and then measures the latent variables by a semi-nested Latent Autoencoding Module (LAM). This avoids point cloud measurement and makes SN-PAE a more robust measurement method. In addition, we implement a set of experiments using solder joints as an example, and the results demonstrate the feasibility and effectiveness of the proposed method.