Research On Remaining Useful Life Of Long And Short Term Memory Network Considering Uncertainty Quantification

With the development of smart manufacturing,smart devices have become more and more widely used because of their significant cost savings and labor gains.However,Due to the outside environment and long time running,it is easy for the components and materials to be worn out,which will influence the security and stability of the equipment.In order to ensure the normal operation of the equipment system and reduce the economic losses caused by unexpected accidents,the remaining useful life is used to determine the failure time,and the equipment is replaced and repaired at an appropriate time to improves the security and stability of the operation of the facility and reduces the loss due to downtime.The rapid development of sensors has made state monitoring data exhibit characteristics such as large amounts of data and multi-dimensional features.Deep learning algorithms are widely used due to their ability to automatically learn features from data and describe the degradation process of complex nonlinear systems.Due to the unbalanced state of the data in deep learning and the uncertainty in neural network models,a new approach is proposed to predict the residual life based on long-term and short-term memory network systems.The PHM2012 dataset and CMAPSS dataset were used to experimentally verify the established network model.The main research contents include:(1)When predicting the remaining life of equipment,the collected data is one-dimensional degradation time series data,with high correlation between data,and high requirements for the selection of neural networks.In this paper,a longterm and short-term memory network that can well learn and handle the dependencies between sequential data is used to extract equipment degradation feature information from the data set for residual life prediction.(2)In order to solve the problem of unbalanced equipment monitoring data,which results in poor accuracy in the forecast of residual life a residual life prediction method based on generating adversarial domain long and short-term memory networks is proposed.First,replace the generator module in the Wasserstein generation countermeasure network with a short-term and long-term memory network,construct a new Wasserstein generation countermeasure network framework,then expand the synthesized data to the original data to construct a new dataset,and finally send the data set to the long and short-term memory network for remaining useful life.The validity of the proposed method has been verified by experiments.,(3)In order to solve the uncertainty of the constructed network,this paper presents a long-term and short-term memory network with residual lifetime prediction based on Bayesian reasoning.The LSTM network model parameters are expressed as random variables under a Bayesian framework,A posterior distribution is obtained by using variational inference to approximate the parameters of the network.The posterior distribution is obtained by means of variational reasoning.The RUL prediction probability distribution function is obtained by means of model training,and then residual life prediction is performed.A new residual life prediction framework has been constructed to reduce the uncertainty of the model,thereby enhancing the ability to predict degradation trends and improving prediction accuracy.