A sensor-based approach for fault detection and diagnosis for robotic systems

As we rely more on robots, thus it becomes important that a continuous successful operation is maintained. Unfortunately, these sophisticated, machines are susceptible to different faults. Some faults might quickly deteriorate into a catastrophe. Thus, it becomes important to apply a fault detection and diagnosis (FDD) mechanism such that faults will be diagnosed in time, allowing a recovery process. Yet, some types of robots require an FDD approach to be accurate, online, quick, able to detect unknown faults, computationally light, and practical to construct. Having all these features together challenges typical model-based, data-driven, and knowledge-based approaches. In this paper we present the SFDD approach that meets these requirements by combining model-based and data-driven techniques. The SFDD utilizes correlation detection, pattern recognition, and a model of structural dependencies. We present two different implementations of the SFDD. In addition, we introduce a new data set, to be used as a public benchmark for FDD, which is challenging due to the contextual nature of injected faults. We show the SFDD implementations are significantly more accurate than three competing approaches, on the benchmark, a physical robot, and a commercial UAV domains. Finally, we show the contribution of each feature of the SFDD.