Multi-scale Dense Gate Recurrent Unit Networks for bearing remaining useful life prediction

Internet of thing (IoT), with the rapid development, is the systematic combination of physical process, information and communication technologies. Industry internet of thing (IIoT), as the extension of IoT in industry, makes the industrial production more intelligent and efficient. Remaining useful life prediction (RUL), as an essential application area of IIoT, plays an increasingly crucial role. In traditional data-based methods, the feature extraction methods depend on the prior knowledge and are separated from the RUL models. Though ensemble learning can be applied to prevent overfitting, the methods about ensemble learning are still separated from the RUL model. To overcome these drawbacks, a novel deep learning network, namely Multi-scale Dense Gate Recurrent Unit Network (MDGRU) is proposed in this paper, which is composed of the feature layers initialized by pre-trained Restricted Boltzmann Machine (RBM) network, multi-scale layers, skip gate recurrent unit layers, dense layers. By adding multi-scale layers and dense layers, the network can capture the sequence features and ensemble different time-scale attention information. Meanwhile it is an end-to-end network combining the feature extraction methods and RUL models only by pre-training the RBM model so it is more convenient for application. Our experiments with real bearings datasets show that proposed MDGRU network is able to achieve higher accuracy compared to other data-driven methods. (C) 2018 Elsevier B.V. All rights reserved.