Research On Prediction Method Of Bearing Performance Degradation Trend Based On Recurrent Neural Network

Rolling bearings play an important role in rotating machinery with their unique advantages.However,the actual working conditions are often very harsh,resulting in bearing degradation.Reliable and accurate prediction of bearing degradation trends has important guiding significance for the health management of mechanical equipment.In this paper,the research on the prediction method of bearing degradation trend based on recurrent neural network is carried out in the process of bearing degradation.The main contents are as follows:(1)For the problem of traditional bearing degradation trend prediction method does not fully consider the relationship between historical vibration state and future state.A prediction method based on recurrent neural network is proposed.By processing the bearing vibration signal,the time domain,frequency domain and time-frequency domain feature parameters of the bearing vibration signal are extracted respectively.Then the monotonic feature evaluation index,which is conducive to trend prediction,is constructed to select the monotonic optimal feature parameters.Finally,the recurrent neural network is used to predict the degradation trend of the bearing.Compared with the traditional BP neural network,the proposed method is better.(2)For the problem that single feature parameter cannot fully characterize the whole process of bearing degradation and the gradient instability existing in the training process of recurrent neural network,the prediction method of bearing degradation trend based on bottleneck feature and long short-term memory network is proposed.By constructing the consistency evaluation index,the sensitive feature parameter set is further screened out.Moreover,the stacked auto-encoder network learns the degradation information of the bearing with sensitive feature parameters,and the bottleneck feature that fully reflects the bearing degradation state is extracted.The gradient instability phenomenon of recurrent neural network structure is explored,and the bearing prediction model based on long short-term memory network is established.Finally,the effectiveness of the method is verified by single-step and multi-step prediction of bearing degradation trend.(3)For the problem that the bearing degradation feature parameters contain many detail components that are not conducive to trend prediction and the long short-term memory network ignore future information,a degradation trend prediction method based on multi-scale bottleneck feature and bidirectional long short-term memory network is proposed.By wavelet multi-scale decomposition of sensitive feature parameter set,the detailed components are eliminated,and the trend items of sensitive feature parameters are extracted.Then the multi-scale bottleneck feature is extracted by the stacked auto-encoder network.The time delay is introduced to construct a prediction model based on the bidirectional long short-term memory network.Finally,the accuracy and generalization ability of single-step and multi-step prediction are further improved by the accelerated degradation experiments.