Predicting residual life of rolling bearing using IMMFE and BiLSTM-GRU-LR

Noise interference in the collected signals results in nonlinear and non-stationary signals during rotating machinery operations. Therefore, the recognition rate of feature extraction and rolling bearing life prediction is extremely low. Recurrent neural network can effectively process multi-time series data. The residual life prediction of rolling bearings based on deep learning has become a promising tool. Deep learning tools are independent of characteristics of rolling bearings or multi-view sequence data. This paper presents a method of predicting the residual life of rolling bearings using an improved cyclic neural network. First, an improved mean multi-scale fuzzy entropy(IMMFE) integrating the complementary ensemble empirical decomposition with adaption noise is proposed to extract the classical rolling bearing characteristic values as a new performance degradation evaluation index to improve the correlation of residual life characteristics. Second, a lasso regression and cyclic bidirectional long short-term memory-gated recurrent unit-Lasso regression (BiLSTM-GRU-LR) neural network is established to predict the residual life of rolling bearings. The algorithm was experimentally validated using the American Intelligent Maintenance Center in the United States (IMS) bearing life data. The experimental results show that the proposed BiLSTM-GRU-LR method, using IMMFE feature sets, has good robustness and high identification accuracy when predicting the residual life of rolling bearings under fault conditions.