Research On Health Evaluation And Remaining Life Prediction Of Rolling Bearings Based On Deep Temporal Network

The Prognostics and Health Management(PHM)of mechanical systems is mainly aimed at fault diagnosis and remaining useful life prediction based on historical and current information from condition monitoring.With the development of sensor monitoring technology,industrial production has entered the era of big data.Mining and analysis of industrial field big data has become one of the main research directions of PHM technology.In recent years,deep learning has achieved great success in machine translation,image recognition,sequence prediction and other aspects due to its outstanding feature extraction and learning capabilities.Rolling bearings are one of the important parts of mechanical systems,which have been widely used in many fields such as aerospace,transportation,intelligent manufacturing and so on.Therefore,it is of great significance to study the degradation theory of rolling bearings and realize the health monitoring and remaining useful life prediction of rolling bearings.This paper takes rolling bearings as the research object,combines the deep learning theory,and carries out research work in the following aspects in view of the practical problems existing in the remaining useful life prediction of rolling bearings:(1)In order to accurately evaluate different stages of rolling bearing degradation process,a rolling bearing state assessment method based on energy entropy moving average crosscorrelation is proposed.This method first decomposes the vibration signal into different frequency bands through the maximum overlap discrete wavelet packet transform,and calculates the energy entropy value of the different frequency bands.The cross-correlation coefficient of the energy entropy is calculated by setting the time sliding window,and the degrading stage of the rolling bearing can be adaptive identified according to the transformation of the cross-correlation coefficient.The effectiveness of the proposed method is verified by rolling bearing dataset.(2)In order to realize online prediction of rolling bearing health indicator values,a rolling bearing health indicator values prediction method based on BiGRU and temporal pattern attention(BiGRU-TPA)is proposed.Firstly,a bidirectional gated recurrent unit structure is proposed to obtain temporal information from two directions.Then the contribution of information at each moment is calculated through temporal pattern attention mechanism to help the network capture important information.Finally,health indicators are input into BiGRUTPA network in the form of a sequence to learn temporal dependencies.The rolling bearing dataset is used to verify the validity of the proposed online prediction method for rolling bearing health indicator values.(3)In order to realize the prediction of the remaining useful life of rolling bearings,a remaining useful life prediction framework of rolling bearings based on temporal convolutional network(TCN-RSA)is proposed.In this framework,a causal dilated convolution is proposed to improve the traditional convolution method,and temporal convolutional network is constructed by the causal dilated convolution to learn the deep feature representation of rolling bearings.Then the residual self-attention mechanism is used to obtain the information of important moments and solve the problem of gradient disappearance and gradient explosion in the deep network.Finally,the framework establishes the mapping between the deep representation of features and the remaining useful life prediction labels through the regressor.The validity and superiority of the proposed prediction framework are verified by two rolling bearing datasets.(4)In order to realize the cross-condition remaining useful life prediction of rolling bearings,a cross-condition prediction framework based on the transferable temporal convolutional network(TTCN-RSA)is proposed in combination with the transfer learning theory.Firstly,an evaluation index of rolling bearing is proposed for feature transferability in this framework,and a rolling bearing transferable feature set can be constructed based on the evaluation index.Then the TCN-RSA network is used to learn the deep representation of features,the distribution difference of features under different working conditions is calculated by the maximum mean difference,and the domain adaptation method is used to minimize the distribution difference.The validity of the proposed cross-condition remaining useful life prediction framework is verified on rolling bearing datasets.