터보팬의 잔여 유효 수명 예측을 위한 다수 소수 특징 추출 병렬 연결 합성곱 신경망

초록

The prediction of remaining useful life (RUL) plays a crucial role in assessing the condition of a machine before it completely fails, ensuring performance by the execution of preventive maintenance beforehand. Recently, various deep learning models have been frequently used for RUL estimation, and they have shown good performance. However, these deep learning models face several challenges such as inefficiency owing to the selection of complex preprocessing methods, overfitting owing to model complexity, and other unresolved issues. Therefore, this study proposes a new deep-learning-based approach to address these issues by constructing a novel structure that includes a simple preprocessing step, minority feature extraction module, and majority feature extraction module. First, it explains the relatively simple preprocessing and assumptions regarding the target data of an undefined training set. Second, we describe the design of a convolution-based model using minority and majority feature extraction modules created through 2D convolutional layers. This model can learn the relationships between minority and majority sensors over time. By connecting the modules in parallel, it aggregates various types of information using multiple features from a single dataset. Finally, we present various experiments on the proposed algorithm and compare it with the latest existing methods using the NASA C-MAPSS dataset.

키워드