Abstract
Industrial robots have continued to play important roles in automation for smart manufacturing. Effective maintenance strategies are essential for successful robot operation. However, maintenance can be costly which has led manufacturers to explore different ways to monitor and evaluate the health of their robot workcell during operation. Successful methods to assess robot health rely on capturing and analyzing significant amounts of continuously generated data from the workcell. Various data streams relevant to a robot workcell may be available such as robot controller data, programmable logic controller data, maintenance logs, and part quality data. Although these data streams can reveal information about the workcell, the large volume and complexity of the data can prove difficult to analyze. Researchers at the National Institute of Standards and Technology (NIST) have developed a test method and accompanying unsupervised learning anomaly detection procedure to assess the health of robot workcells. The approach involves defining a distance metric which is then passed to a robust outlier detection step. Performance of the approach demonstrated on several types of simulated robot failures and compared against different popular anomaly detection algorithms.
