可解释性
计算机科学
残余物
组分(热力学)
可靠性(半导体)
过程(计算)
方位(导航)
特征(语言学)
编码器
信号(编程语言)
数据挖掘
人工智能
模式识别(心理学)
算法
功率(物理)
语言学
物理
哲学
量子力学
热力学
程序设计语言
操作系统
作者
Qile Wu,Ke Xiao,Wanquan Xiong,Yanfeng Han,Bin Song
标识
DOI:10.1088/1361-6501/ad128a
摘要
Abstract Rolling bearings serve as indispensable core components in modern industrial equipment and they are critical for safety and reliability. Consequently, accurate prediction of their remaining useful life (RUL) is essential and has far-reaching implications. This paper proposes a novel RUL prediction model, referred to as the dynamic rectified linear unit-based residual additive attention ConvGRU (DReLU-RA-ConvGRU) model, which is built upon the encoder–decoder structure to accurately predict the RUL of bearings. To overcome the limitation of the original signal, characterized by a single feature and limited degradation information, three-domain features are employed and filtered as inputs to the model. The DReLU component in the proposed RUL prediction model effectively captures variable feature information within the degraded signal, while the ConvGRU component learns both temporal and spatial information with fewer parameters. Additionally, the RA component captures the significant contributors to RUL prediction, and the inclusion of residuals facilitates easier network learning. Furthermore, a three-dimensional visualization of the attention weights was conducted to enhance the interpretability of the network’s prediction process. In order to verify the effectiveness of the method, RUL prediction was conducted using vibration data from the PRONOSTIA platform and compared against several existing methods. The results demonstrate the method’s superior performance and feasibility, as indicated by high scores.
科研通智能强力驱动
Strongly Powered by AbleSci AI