Deep multisource parallel bilinear-fusion network for remaining useful life prediction of machinery

With the increasing demand for stability, safety, and reliability of commissioned machines, diverse types of sensors are positioned on key components. To exploit these multisource data, more and more deep learning-based remaining useful life (RUL) prediction approaches are developed recently. These approaches, however, still suffer from the following limitations: 1) Multisource data are tangled whilst being input into the network, which incurs information confusion and interference. 2) Valuable features of different sources which are sensitive to degradation states are not extracted adaptively. 3) The fusion methods prevent sufficient interaction between multisource features. To overcome the above drawbacks, a deep multisource parallel bilinear-fusion network (MPBFN) is proposed for RUL prediction of machines in this paper. The proposed MPBFN develops multiple parallel subnetworks to automatically extract deep features from different source data separately. Then, the extracted high-dimensional features are fused by a specially designed temporal compact bilinear fusion (TCBF) module. Finally, the RUL estimation module is used to perform RUL regression prediction. The proposed MPBFN is evaluated with multisource data collected from life testing of milling cutters and compared with several state-of-the-art RUL prediction approaches. Experimental results show that the proposed MPBFN outperforms other prognostic approaches in terms of accuracy and robustness.