An Anomaly Detection System for Smart Manufacturing Using Deep Learning
Abstract
The smart manufacturing evolution enables financial and operational improvements across the manufacturing industry. However, smart manufacturing encompasses complex, interconnected systems which can fail at any time. To address this challenge, a novel, two-part anomaly detection system for robotic processes, with an application focus on robotic surface finishing, is presented. The first part proposes an unsupervised Attention-based Convolutional Long Short-Term Memory Autoencoder with Dynamic Thresholding (ACLAE-DT) framework for anomaly detection and diagnosis in multivariate time series of robotic surface finishing components. The second part proposes a deep residual Convolutional Neural Network-based triplet model for anomaly detection in the produced robotic surface finishes, where defective training samples are synthesized exclusively from non-defective samples via random erasing data augmentation. Evaluation results demonstrate the performance strength of ACLAE-DT over state-of-the-art time series methods and the performance strength of the proposed triplet model in detecting anomalies for known and novel surfaces.