Long short-term memory neural network with weight amplification and its application into gear remaining useful life prediction

• Published in: Engineering applications of artificial intelligence 2020-05, Vol.91, p.103587, Article 103587
• Main Authors: Xiang, Sheng, Qin, Yi, Zhu, Caichao, Wang, Yangyang, Chen, Haizhou
• Format: Article
• Language: English
• Subjects: Attention mechanism; Feature fusion; Gear vibration signal; Recurrent neural network; Residual life prediction
• ISSN: 0952-1976; 1873-6769
• DOI: 10.1016/j.engappai.2020.103587

As an important component of industrial equipment, once gears have failures, they may cause serious catastrophes. Thus, the prediction of gear remaining life is of great significance. The health indicator of gears is first generated by fusing time-domain and frequency-domain features of gears vibration signals via the isometric mapping algorithm. Then a new type of long-short-term memory neural network with weight amplification (LSTMP-A) is proposed for accurately predicting gear remaining life. Compared with traditional LSTMs, LSTMP-A amplifies the input weights and the recurrent weights of the hidden layer to different degrees by the attention mechanism according to the contribution degree of the corresponding data, and a projection layer is added into the network. With LSTMP-A, we can predict the health characteristics of gears based on historical fusion features. With the monitoring data of a gear life cycle test, the comparative experiments show that the proposed gear remaining life prediction method has higher prediction accuracy than the conventional prediction methods.