Wear indicator construction of rolling bearings based on multi-channel deep convolutional neural network with exponentially decaying learning rate

Wear indicators (WIs) attempt to identify historical and ongoing degradation processes by extracting features from acquired data. The quality of the constructed WIs affects the validity of the data-driven prediction directly to a great extent. The main problems of the existing WI construction methods are as follows: (1) the existing WI construction methods are based on the single channel sensor signal, resulting in the insufficient use of the measured data; (2) the existing WI construction based on deep learning is using a fixed learning rate, leading to low training efficiency. To solve the above problems, a multi-channel deep convolutional neural network with exponentially decaying learning rate (EMDCNN) is proposed to evaluate the health of rolling bearings. In this paper, the original multi-channel signals are input to the proposed network. Exponentially decaying learning rate is proposed to train the neural network efficiently. Moreover, a weighted evaluation criterion is proposed in this paper. The validation results show that the proposed method is superior to the compared four WI construction methods in monotonicity, trendability, robustness, and the value of weighted criterion is 15.3%, 10.8%, 19.0%, 14.8% higher than that of ECNN-WI, FCNN-WI, NN-WI and SOM-WI respectively.