Combining Surrogate Digital Twins with Extended Kalman Filters for Model-based Robot Failure Diagnosis

Data-driven models like deep neural networks have become the state-of-the-art for fault diagnosis. A significant drawback of such models is that they lack explainability and are difficult to generalize. Model-based approaches like the multi-observer approach can address the drawbacks of data-driven models by capturing domain knowledge for fault diagnosis but they remain computationally expensive and time-consuming. In this study, we propose a surrogate digital twin-based system-level fault diagnosis algorithm to reduce the computational burdens of the model-based fault diagnosis methods while keeping their benefits of explainability and generalization capabilities. The developed approach incorporates a Long Short-Term Memory (LSTM) model as a surrogate model for a highly accurate digital twin of the system being diagnosed. While keeping the simulation accuracy, the computational complexity can be greatly reduced. Then, the developed algorithm relies on a Multi-Extended Kalman Filter (MEKF) for multi-class fault diagnosis, in which the state transition function is driven by the surrogate digital twin of the physical entity. Each potential failure mode is modelled in the surrogate digital twin, and an extended Kalman filter is used to estimate the system states under each failure model. The failure mode with the lowest root mean squared error compared to the observation data will be identified as the occurring failure. The proposed hybrid approach also allows leveraging system-level observations to determine the failure states of components. We apply this method to a robotic arm with an end-effector, whose position is determined by 4 rotating motors. Using only the motor control commands and end-effector position measurements, the introduced approach reliably identifies the operational status of the robot, specifying whether it functions properly or if one of its motors is stuck. The framework was trained using data generated by the digital twin and tested with data from the real robot. The experimental results demonstrate an accuracy of 94% compared to 61% obtained with an LSTM only, and a strong computational time reduction, with no observed discrepancies between training and tests.

Keywords: Model-based fault diagnosis, Multi Extended Kalman Filter, LSTM, Digital Twin, Robotic Arm.