Sensor Fault Detection and Compensation in Lunar/Planetary Robot Missions Using Time-Series Prediction Based on Machine Learning

Abstract :

Mobile robots operating in a lunar or planetary space mission can usually neither be repaired from nor brought back to earth. In case of sensor damages or drop outs an overhaul of the hardware leading to a properly working sensor is not possible in most cases. Instead, the system has to continue working as reliable as possible. In the special case of an autonomous space robot this means that the robot, first, needs to detect a sensor drop out automatically. Second, the missing sensor signal needs to be compensated. In typical mobile robot setups this is possible by using other sensor modalities. Presented is a method to detect single sensor faults by model-based predictions covering multiple sensor modalities. The methods are learned using one of two methods: tested and compared are a multi-layer perceptron (MLP) and a “Neural Gas (NG)” vector quantization method. The test use case is a turning skid-steered robot with four different sensor modalities (velocities left and right wheels, gyroscope Z-axis, and horizontal optical flow). With the collected training and test data the model predictions turns out to be accurate enough for the purpose of a sensor fault detection. Moreover, by using learned models a compensation in case of a sensor fault can be possible.