Transfer life prediction of gears by cross-domain health indicator construction and multi-hierarchical long-term memory augmented network

Abstract

The long-term remaining useful life (RUL) prediction of gears is crucial for the safe operation and maintenance of rotating machinery. However, most existing RUL prediction methods face great challenge under the variable working conditions due to the lack of enough prior run-to-failure data. Therefore, this paper addresses to explore a new transfer life prediction methodology for gears. A gear health indicator (HI) transfer construction framework named TQFMDCAE is first proposed by a quadratic function-based multi-scale deep convolutional auto-encoder and maximum mean discrepancy, and it can generate the cross-domain HIs under different working conditions. Next, a novel RNN-based network named multi-hierarchical long-term memory augmented network (MLMA-Net) is developed for the life prediction of gears based on the obtained HIs. In MLMA-Net, a new memory augmentation function is intended to increase the network's long-term memory capacity. The proposed multi-hierarchical mechanism then divides the sequence information of the network into three attention hierarchies and three cell hierarchies, respectively. Experiments on equipment indicate that the developed MLMA-Net has a remarkable predictive capacity, particularly for predicting the long-term life of an object. Meanwhile, comparative results demonstrate that the proposed RUL prediction methodology is superior to other typical RUL estimation methods.