Construction of Interpretable Health Indicators for System-level Fault Prognosis Driven by Inductive Generative Learning

The construction of health indicators (HIs) is critical for the system-level fault prognosis. Due to its high interpretability and credible identifiability, the model-driven approaches are more welcome. However, for a complex system, it exists always some inter-component coupling mechanism among model-driven HIs, which makes the fault representations difficult to distinguish based on these HIs. Therefore, this paper proposes a model-driven construction method of interpretable HI driven by inductive generative learning, aiming to optimize the accuracy of the system-level fault prognosis via the decoupling among the model-driven HIs. The inductive architecture attempts to treat automatically the high-coupling performance knowledge inside the mathematical model to extract the correspondent identifiable HIs, while the integrated generative architecture realizes an efficient identification of the extracted HIs in order to guarantee a real-time fault prognosis. Experiments on a simplified electrical hydrostatic actuator model and a high-quality electromechanical actuator model are conducted to demonstrate the state-of-arts level of the proposed approach on the HI construction for the system-level fault prognosis under inter-component coupling circumstances.