Durability Prediction of Rolling Bearings Based on CNN–Mogrifier LSTM

In consideration of the limitations of traditional durability prediction models for rolling bearings in terms of prediction accuracy and the capability to mine the temporal dependence of full life cycle vibration signals, this study proposes a durability prediction model for bearings based on a convolutional neural network (CNN) and improved long short-term memory (LSTM). The model introduces the Mogrifier mechanism to optimize the LSTM unit, explicitly modeling the interaction between input data and hidden state to enhance the capability to capture global features and the dynamic evolution relationships of vibration signals. The data is based on the IMS Data Challenge Dataset. First, the original vibration signal is normalized with MinMaxScaler, and the labeled time-series tensor is constructed based on the annotated signal data as model input. Then, the CNN is used to extract the local features of the signal, while Mogrifier LSTM is used to model temporal features deeply to achieve high-precision remaining life prediction. Experimental results show that the proposed model exhibits significantly better prediction accuracy than the traditional LSTM model, with an accuracy improvement of 93.86%, which is 6.12% higher than that of the ordinary LSTM model. The root mean square error is reduced by 35.04%, and the coefficient of determination is increased by 6.5%. The results indicate the effectiveness of the Mogrifier LSTM mechanism in modeling complex temporal data, providing new ideas and technical support for the durability prediction of bearings.