22‐1: Automated Malfunction Detection for Robotic Arms in Panel Manufacturing Using Deep Latent State Space Model

The paper describes a novel deep‐learning‐based system for malfunctions detection for the robotic arms used in display panel manufacturing process. Our method automates the preventive inspection of the manufacturing robotic arms by using a deep learning model that replaces the laborious human inspection. More importantly, our method achieves continuous monitoring of robotic arm condition, and alarms the maintenance team ≥2 days prior to the eventual failure, thus eliminating wastes of production materials and time consumed in the unexpected machine downtime. To achieve this automated and continuous malfunction detection, we leverage the deep latent state space model to learn the normal behavior of robotic arms from multivariate time‐series sensor data collected in daily production, and then identify various types of malfunctions by looking at the difference between the expected normal signals and the measured signals. The deep latent state space model incorporates specialized latent embeddings formulation that is tailored to capture the full dynamics in time‐series data. On an internal dataset that is directly sampled from production line, our approach achieves 100% malfunction detection (i.e. recall rate) while maintaining low false alarm rate (~2% measured in the span of a year). Our method can significantly advance the efficiency of preventive maintenance and reduce the cost in display panel manufacturing line.