Fault tolerant control method for displacement sensor fault of wheel-legged robot based on deep learning

Abstract

In this paper, a fault-tolerant control method based on deep learning is proposed for multi displacement sensor fault of a wheel-legged robot with new structure. Unlike most methods that only detect a single sensor, the proposed method can detect a large number of sensors simultaneously and rapidly. The residual error is generated by sensor values and the prediction model which is established by deep belief network(DBN) in deep learning, to detect faults and locate faulty sensors. Then, by using other non-faulty sensor information to reconstruct the signal through the neural network and combining with the coupling relationship of the 6-DOF platform, the fault sensor signal can be estimated accurately and the error accumulation problem can be also solved. Comparing the two algorithms of neural network and support vector machine(SVM), the reconstruction signal of neural network has higher accuracy. So, the performance of the wheel-legged robot can be guaranteed within a safety range. It is proved that the proposed method has high reliability and stability.