Automated visual inspection in high-mix bearing manufacturing is critical for quality assurance, yet practical deployment of deep learning models for inspection remains limited by labeled data scarcity, especially during frequent new part introduction. Frequent new part introduction also introduces the risk of contaminating a presumed normal training dataset with defective data while process parameters are still being optimized. State-of-the-art unsupervised anomaly detection (UAD) nearest-neighbor memory bank methods promise high image-level performance with minimal model training, but the latent representations lack interpretability. This work evaluates a probabilistic memory bank model tolerant of training data contamination by using Gaussian mixture models (GMM) for the latent representation of patch features. The GMM model is found to not only be a sensible representation for modeling features extracted from images of a multi-Gaussian surface, but also to perform as well or better than other state-of-the-art nearest-neighbor memory bank anomaly detection methods, evaluating effectiveness in the presence of training data contamination up to 34%, where other studies only evaluate effectiveness to 15% contamination. This is accomplished through full-covariance GMM, Mahalanobis scoring, and top-n patch score aggregation. This work also finds that iterative model training can be eliminated, saving up to 89% training time over other probabilistic methods while exceeding state-of-the-art UAD methods by 4% AUROC. The resulting probabilistic latent feature representation is well positioned to be used for continual learning or integration with existing statistical process control (SPC), which is commonly used in manufacturing. • Full covariance Gaussian Mixture Model (GMM) to capture correlated latent textures. • Robust to training data contamination common to bearing manufacturing processes. • Probabilistic scoring for seamless integration with statistical process control. • Up to 89% faster training than other probabilistic methods for in-process deployment. • Outperforms state-of-the-art memory bank-based anomaly detection methods by 4%.
Hopkins et al. (Mon,) studied this question.