Bearing Health Monitoring Based on the Improved BiISTM-CRF

Abstract

Bearing Remaining Useful Life (RUL) prediction has important meaning in the mechanical maintenance. However, the existing RUL algorithms cannot achieve stable prediction. Therefore, an improved bearing health monitoring algorithm based on Bidirectional Long Short-Term Memory (BiLSTM) integrating Conditional Random Field (BiLSTM-CRF) is proposed. The empirical mode decomposition (EMD) algorithm is used to decompose the bearing diagnostic signal into several intrinsic mode function (IMF) components. Moreover, the effective IMF component is selected to reconstruct the signal by combining the crosscorrelation coefficient and kurtosis criterion. Through the reconstructed signal extracting the time-frequency features into a feature vector, the feature data with lower dimension can be got. Then, the feature with lower dimension as inputs and RUL status as the output are used to train the BiLSTM-CRF model, which can achieve more accurate predictions. Finally, the XJTU-SY bearing data is used to verify the effectiveness of the proposed algorithm. Experiments show that this proposed method can get the best performance comparing with the convolutional neural networks and the Long Short-Term Memory.