Remaining useful life estimation for aero-engine with multiple working conditions via an improved generative adversarial network

For aero-engine, due to the changeable working conditions and various failure modes, it is often difficult to estimate the remaining useful life (RUL) online. At the same time, it is difficult to calibrate the health status data of the engine. In this case, the traditional model-based prediction methods have poor adaptability. However, the data-driven methods have poor robustness due to the lack of a matching domain. We propose a transfer prediction method with an improved generative dversarial network (GAN). The sample data from different working conditions and different failure modes are transferred to the tested engine for life prediction, to solve the problem of training data sample acquisition. Firstly, a multi-source data fusion method for engine health indicators is proposed, which fuses multi-dimensional health features into one-dimensional health indicators. Then the sliding window method is employed to construct the time-series samples. Based on the idea of GAN, a dynamic adversarial domain adaptive transfer network is proposed to estimate the RUL of aero-engine. A weighted loss function is proposed and added to the network to improve the robustness. The experiment of C-MAPSS data is employed to test the proposed method at last.