Feature extraction for fault diagnosis in series elastic actuators

Series elastic actuators provide beneficial characteristics for safe human-robot interaction and energy efficient robotic motions. Yet, such actuators might have an increased probability of faults due to their higher complexity and operation in critical states, e.g., antiresonance. This contribution investigates feature extraction methods for fault diagnosis in such actuators. Stiffness and motion sensor faults are focused since those are assessed to have high occurrence probabilities with potentially severe consequences. To detect stiffness deviations, a recursive least squares estimator is implemented while Kalman-Bucy filters are applied to generate residuals that indicate encoder faults. The methods are examined using models of system and fault dynamics of a variable torsion stiffness actuator. The simulation results show that very distinct features for fault diagnosis can be extracted. The investigated feature extraction methods are very promising for interpretation by classification methods. Recommendations on how to implement those methods for diagnosis purposes are given.