ELM-based adaptive neural estimation for actuator faults detection and diagnosis of nonlinear uncertain systems

Abstract

Actuator, as a key component of control system, whose faults detection and diagnosis (FDD) is a complex problem due to system modeling uncertainty, so it is essential to propose advanced FDD scheme that accurately detects the faults. In this paper, we develop an actuator FDD scheme for a class of uncertain nonlinear systems based on extreme learning machine (ELM). In ELM, all parameters of hidden layer nodes need not be adjusted during learning, which may simply be assigned with random values, and the output weights only need to be adjusted. Within this scheme, two ELMs are employed to learn the unknown system function and unknown fault function. Firstly, a stable adaptive observer is designed to monitor faults in an online manner. Secondly, adaptive threshold is designed to make the fault detection, and deviation between the actual and the estimated system is known as residual. If the residual exceeds threshold at finite time denotes a fault occurrence. Different from the existing schemes, online computational efficiency and learning speed are improved considerably because ELM is introduced in this FDD scheme. Finally, a single-link robotic arm will be employed in simulation to illustrate the effectiveness of the proposed FDD scheme.