A novel continuous delay hidden layer deep belief network and its application in life prediction of rolling bearings

Abstract

Abstract The traditional rolling bearing trend prediction method is difficult to predict the bearing degradation process comprehensively and scientifically, and the feature extraction is highly dependent on the manual experience method, and the prediction ability of the bearing measurement data with strong dynamic characteristics and nonlinear is very limited. Therefore, how to build a high-precision life prediction model has become an urgent problem and challenge. In this paper, a continuous delay hidden layer deep belief network (CDHLDBN), kernel principal component analysis (KPCA)-CDHLDBN, is proposed based on the KPCA preprocessing method. Firstly, a continuous deep belief network (CDHLDBN) with a delay layer is proposed, which makes the output of the hidden layer more fully consider the influence of the previous moment on the output of this moment, so as to avoid the defect of DBN model which is difficult to process the data with strong dynamic time series. Then, all visible layers of a continuous DBN with a delay layer are de-averaged, and then the simulated bearing sequence data is trained and predicted using the KPCA-CDHLDBN life prediction model. The results show that KPCA-CDHLDBN model can predict nonlinear dynamic time series data better than other deep learning models. Finally, it is applied to the performance degradation life prediction of rolling bearings. The performance of the proposed method was verified by the bearing accelerated life vibration data provided by PRONOSTIA experimental platform, and the performance of the proposed method was compared with that of the traditional DBN and its improved prediction model. The results show that KPCA-CDHLDBN has better prediction accuracy and faster prediction speed.