Sensitivity of PCA and Autoencoder-Based Anomaly Detection for Industrial Collaborative Robots

Industrial robots are extensively used in the manufacturing industries due to their efficiency and precision. Concurrently, the use of industrial collaborative robots (often referred to as cobots) is on the rise because they are easy to reprogram and interact smoothly with humans. However, cobots are prone to different faults and malfunctions, which can lead to unplanned downtime; thus, the need for early fault detection (a key aspect of predictive maintenance in Industry 4.0). In this work, anomalous conditions were artificially introduced to an industrial collaborative robot by placing different weights on the shoulder arm of the robot. Subsequently, two distinct models, Principal Component Analysis (PCA) and sparse Autoencoder (AE), were constructed to identify those anomalies. The two models leverage multivariate operational data sourced from a universal robot (UR5e) and are individually trained to reconstruct the normal or baseline data. Based on the reconstruction error of the models, the Q and T² metrics are examined for the PCA model while the Mahalanobis distance (MD) is used for the sparse AE model to detect abnormal conditions. The results show the two methods are sensitive and robust in detecting the investigated anomaly; however, PCA is more effective from a computational perspective.