Research On Deep Learning For The Construction Of Health Indicator And The Prediction Of Life Of Bearings

In the construction of bearing health indicators(HI)and the prediction of remaining useful life(RUL),for purpose of solving the problems of inadequacy of extracting features,low performance of HI,insufficient generalization of models,and the prediction results vulnerable to the impact of signal intensity and change amplitude,the HI construction and the RUL prediction based on deep learning have been studied.The paper puts forward a HI construction approach on account of Principal Component Weighted Analysis(PCWA),constructs a HI on account of Multiple Features Fusion Depth Network(MFFDN),and improves a RUL prediction approach on account of Weighted Temporal Convolution Network(WTCN),which raise the ability of HI reflecting the degradation process and the RUL predicted precision.The primary coverage and research results are listed as follows:(1)A method of constructing HI based on PCWA is proposed.Aiming at the problems of lack of standards for bearing parameter selection,less information on feature degradation,and deficiency of small components when constructing HI by traditional statistical analysis methods,the PCWA-based HI construction method was studied.Using monotonicity,correlation and robustness to construct comprehensive assessment criteria to supply foundation for features screening and HI evaluation,which reduces the interference of expert experience and prior knowledge;On account of principal component analysis,PCWA is put forward to decrease the dimensionality of multiple features,which lower the complexity of data and the impact of noise interference;The percent contribution is regarded as the weight of the weighted sum of the corresponding principal components to fuse the dimensionality reduction vectors,and a principal component weighted Analysis HI is constructed,which overcome the problem of missing small components and increases the key degradation information contained in HI.Verification on the published bearing datasets proves that the proposed method can validly raise the ability of HI to reflect the degradation process.(2)A model of HI construction model based on MFFDN is constructed.Aiming at the problems of conventional deep learning approaches like weak generalization capacity,inadequate feature extraction,and unreasonable division of bearing health status,a HI construction approach based on MFFDN is put forward.The FPT detection method is introduced to divide the health stages,so as to set labels for the network input data,reasonably map the health status of bearing degradation and reduce the difficulty of network learning;MISH function acts as a substitute for Re LU function as the activation function of the network to prevent neurons from being forced to die due to negative values and avoid information loss;Using the capacity to extract local and deep features by CNN,a parallel feature extraction module is built to extract the degradation information of multiple parameters concurrently,and the Concatenate layer is used to fuse that to construct a multiple feature fusion HI,which raises the generalization and feature mining capacity of the network.The results of verification manifest the constructed HI have good monotonicity,correlation and robustness.(3)A RUL prediction method based on WTCN is improved.Aiming at the problems of hard model selection,low forecast precision,complex sample data,and the influence of signal intensity and variation amplitude on the prediction results in RUL prediction,a RUL prediction model based on WTCN is proposed.The multiple feature fusion HI with superior performance are used as the sample data,which lower the difficulty of learning the complex mapping relationship between degraded features and RUL by network;Using temporal convolutional network as the model of RUL prediction and taking advantage of the characteristics of dilated causal convolution and residual connection,which strengthens the capacity of the network to process temporal information,and alleviates the problem of information leakage and gradient disappearance;By time weighting the mean square error function,the time loss function is proposed,which relatively increases the weight of network error in the degradation stage and makes the model tend to reduce the error of network pickup when the vibration is large and raise the precision of prediction.The contrastive trials manifest this approach has a high prediction precision,which can validly capture the degradation mode and is suitable for rolling bearing RUL prediction.