Research On Life Prediction Method Of Rolling Bearing Based On Unsupervised Deep Model Transfer

Rolling bearings are one of the key components in rotating machinery and equipment,which are widely used in different industrial fields.However,in actual work,the continuous changes of bearing operating conditions have made it difficult to obtain a large number of labeled data.At the same time,traditional methods of predicting life have the disadvantages of both difficulty in constructing health indicators and large life prediction errors.Therefore,being able to accurately predict the remaining useful life of rolling bearings is of great significance to industrial production.This paper proposes a method for predicting the remaining useful life of rolling bearings under different working conditions,which is divided into two parts:state identification and life prediction.In terms of state identification,it is proposed that a method based on deep transfer learning model for identifying the state of rolling bearings.This method first extracts the root mean square value of the full life cycle signal of rolling bearing and performs normalization processing,and to determine the optimal number of segments in the characteristic sequence as normal period,degraded period and recession period,it introduces a bottom-up time series segmentation algorithm and the coefficient of variation and the rate of change.The data in source domain is input into the improved fully convolutional neural network for training so as to extract deep features,obtaining a pre-training model.It is proposed to use the gradient of the pre-training model as a "feature" to participate in the training process of target domain network together with the traditional features of the pre-training model to achieve the purpose of transfer learning,establishing a model of identifying states and obtaining the probability of state identification.The experimental results prove that the proposed method can obtain higher accuracy without requiring the target domain data with labels when identifying bearing data under different working conditions.In terms of remaining useful life prediction,on the basis of state identification,the method of estimating state probability is introduced to realize remaining life prediction.This method first obtains the residence time of different degraded states of vibration data of rolling bearing in the source domain,at the same time,determines the remaining useful life corresponding to each degraded state.Then the method extracts the probability of each degraded state from the model of identifying states,and combines both the estimation of state probability and the residence time of different degraded states to obtain the remaining useful life of the bearing at the current time point.The experimental results prove that the proposed method of predicting remaining useful life of rolling bearing,which is based on the probability of state identification and estimation of state probability,does not need to construct health indicators and has a smaller prediction error.The prediction error is 19.37%and most bearings are advanced predictions.