Detecting Software Anomalies in Robots by Means of One-class Classifiers

Abstract

[Abstract] The growing dependence on collaborative robots in essential industrial and service sectors raises urgent concerns regarding their reliability and ability to handle faults. Undetected software issues can degrade performance, jeopardize safety, and result in expensive downtimes. Incorporating collaborative robots into daily life and industrial settings requires strong and dependable systems, especially concerning software. While most anomaly detection research has focused on hardware anomalies, this study addresses the underexplored challenge of software anomaly detection in component-based robotic systems. Leveraging a publicly available dataset with labeled software-induced anomalies, six one-class classification techniques were evaluated: Approximate Convex Hull, Autoencoder Neural Networks, K-Means, K-Nearest Neighbors, Principal Component Analysis, and Support Vector Data Description. Each classifier was assessed across preprocessing methods and hyperparameter configurations, using the Area Under the Curve (AUC) as the primary performance metric. The results show that Principal Component Analysis outperforms other methods in most scenarios, although the optimal performance varies depending on the anomaly type. The results confirm that the suggested one-class classification method is an efficient means of early identification of software anomalies in robotic systems, potentially improving operational reliability and reducing downtime.