Automatic multi-differential deep learning and its application to machine remaining useful life prediction

Different levels of characteristic information cannot be mined using various feature extraction modes in most neural networks, and thus, a novel method called the automatic multi-differential learning deep neural network (ADLDNN) is proposed in this work. First, a measurement-level division unit is designed for actively classifying multisource measurements into several levels. Then, a multibranch convolutional neural network (MBCNN), in which each branch can execute the corresponding feature extraction in accordance with the level of its input data, is constructed. Second, a multicellular bidirectional long short-term memory is proposed. A bidirectional trend-level division unit is used for actively classifying the output features of MBCNN into several levels of degradation trend along the forward and backward directions. Each cell unit implements the corresponding feature learning on the basis of the bidirectional trend level. Finally, the remaining useful life of a machine is predicted via a fully connected layer and the linear fitting of a regression layer. The effectiveness of the proposed method is validated on the widely used C-MAPSS dataset and an actual wind turbine gearbox bearing dataset. Comparative results show that the proposed ADLDNN is superior to state-of-the-art methods.

Automated summary

This paper proposes a novel method called ADLDNN, which addresses the limitation of neural networks in mining different levels of characteristic information. By introducing a measurement-level division unit, a multibranch convolutional neural network, and a multicellular bidirectional long short-term memory, the proposed method is able to effectively extract and learn features. Experimental results demonstrate that the proposed method outperforms state-of-the-art methods in predicting machine remaining life.