Fault detection for serial robotic manipulators: A review

Robotic manipulators are an increasingly popular tool with a wide variety of applications. Over the course of their operating life, these systems are subject to wear and tear that can degrade their performance. To improve longevity, it is important that effective fault detection techniques be developed in order to identify and diagnose these faults before they escalate into catastrophic and potentially dangerous failures. Fault detection techniques can generally be classified as signal-based, model-based, data-driven, or as a hybrid thereof. Signal-based methods use only signal processing techniques and simple evaluation methods to detect fault signatures in operating signals. Model-based methods use system models to predict the manipulator’s behaviour, against which the actual behaviour is evaluated. Data-driven methods have no system knowledge, and use machine learning methods to uncover fault indicating patterns in real-time operating data. Hybrid methods, which incorporate elements from two or more of the above methods, are an emerging research area that may hold the key to advancements in manipulator fault detection. In this review, robotic manipulator fault detection research from the last ten years is analyzed, in particular with respect to issues such as validation, output specificity, and generalizability.