Score-Based Anomaly Detection for Smart Manufacturing Systems

We propose an anomaly detection framework that generates an anomaly score for the current observation of a task. The proposed framework can overcome two major problems in smart manufacturing systems, where a robot learns task execution by human expert demonstrations. First, it can function as a warning system to alert human operators when the robot goes into a significantly different state than those of demonstrations. Second, in order to prevent redundant samples, the proposed framework can be used as a guiding system for human experts throughout the recording of demonstrations for collecting novel states, which are different from already collected samples. Unlike parametric models, the proposed framework is better suited for behavioral cloning-adapted smart manufacturing applications in which the training dataset is continuously extended. We also propose four real-world datasets for different applications, e.g., car plugging and refrigerator production line inspection. To the best of our knowledge, it is the most comprehensive real-world dataset for anomaly detection for smart manufacturing systems with a robotic arm. We evaluate our method quantitatively on the real-world datasets and compare it with several standard baselines from the literature.