Remaining Useful Life Prediction for Complex Electro-Mechanical System Based on Conditional Generative Adversarial Networks

Abstract

Remaining Useful Life (RUL) prediction is of significance to provide valuable information for implementing condition-based maintenance and repair. Except for the difficulty on formulating the physical model of the complex electro-mechanical system, another challenge is how to utilize the sparse samples to achieve accurate prediction results. To address this issue, this paper proposes a novel RUL prediction method based on the sample augmentation by the improved Conditional Generative Adversarial Networks (CGAN). The aircraft Auxiliary Power Unit (APU) is taken as a typical complex electro-mechanical object. Two-dimensional condition monitoring samples of the aircraft APU contain the potential degradation information, which bring difficulty for formulating an accurate and stable RUL prediction model. First, its two-dimension condition monitoring samples are augmented by the improved CGAN. Then, the augmented samples and the original samples are both utilized as the input of the RUL prediction method. Through comparison experiments on a practical sample set, the effectiveness of the proposed method is evaluated by different RUL prediction methods and combinations of samples.