Reliable fault diagnosis (FD) is important to ensure safety in nonlinear engineering systems. Modern engineering systems are often subject to unknown complex nonlinearities and varying operation conditions, therefore, one of the main challenges for FD of nonlinear systems is the robustness against these uncertainties. In this paper, a novel robust FD approach combining the reinforcement learning (RL) and the deterministic learning theory (DLT) is developed for a class of discrete-time nonlinear systems. The DLT is employed to pre-train the neural network (NN) aiming at approximating the unknown nonlinear complexity and obtaining dynamical fault models, then RL techniques are employed to adapt the NN parameters to improve the robustness of fault models. The stability of the learning process is rigorously analyzed using Lyapunov-based methods, and the effectiveness of the presented method is validated by a rotating stall warning experiment based on the data from Beihang University compressor test rig. Experiment results demonstrate that compared with other methods, the proposed method can achieve better performance in lead warning time and robustness.
Read full abstract