Intelligent Anomaly Detection of Robot Manipulator based on Energy Consumption Auditing

Robot manipulators have become a key component in smart manufacturing, and play instrumental roles in automated production lines. It is of paramount importance of monitoring their operational health and security. Nevertheless, manufacturing systems are vulnerable to both cyber and physical threats/attacks, necessitating a side-channel monitoring mechanism. Recently, the machine learning techniques to detect and identify threats in Internet of Things by auditing energy consumption have been developed. This paper presents a holistic framework that combines the deep autoencoder (DAE) and energy consumption auditing for real-time health and security monitoring and diagnosis of manufacturing systems. The key idea is to train DAEs with energy consumption data of normal operation only without presence of threats/anomaly, which is then employed to reconstruct energy consumption online. This semi-supervised approach allows to detect faults by evaluating the magnitude of reconstruction errors. That is, a larger error indicates anomalous events. The method is demonstrated using a delicately designed experiment, in which a robot manipulator is programmed to perform different tasks in a circular pattern. Cyber and physical threats are implemented to trigger anomalous events. For all anomalies, the path of the robot manipulator remains unchanged, which makes impossible health and security monitoring through simple inspection of completed products/tasks. By auditing the DAE reconstruction error of energy consumption during task execution, most of the anomalies are detected with the accuracy of greater than 80%. Having relatively small false positives, the precision is found to be above 73%. Unfortunately, threats are not detected every time instance during the entire period of a single attack, causing many false negatives and a low (23%) recall rate. However, having a high recall rate is not necessary because often anomalies are counted as events, which does not need to be detected for every instance. Our study shows that energy consumption auditing in conjunction with DAE is a promising approach for independent, side-channel anomaly detection.