Research On Remaining Useful Life Prediction Of Rolling Bearing Based On Fusion Neural Network

Rolling bearings are parts that frequently appear in mechanical equipment,and are also key components that have a good impact on equipment performance.The healthy operation of rolling bearings is closely related to the stable and efficient operation of mechanical equipment.Therefore,the remaining life prediction of rolling bearings is an important basic technology for equipment maintenance.The traditional bearing life prediction technology needs to involve more complex professional knowledge,and the model has a large generalization ability,while the data-driven prediction model can make full use of big data to achieve accurate and rapid bearing life prediction.To predict the remaining life of bearings,it is necessary to extract the characteristics that can reflect the life state from the historical operation data of bearings.In order to make these features objectively and comprehensively reflect the real performance degradation trend of the bearing,this paper proposes an improved adaptive variational mode decomposition method based on simulated annealing algorithm to extract time-frequency domain features.And the penalty factor has a great influence on the results and there is no mature method to determine it.The simulated annealing algorithm can adaptively determine the parameters of the variational mode decomposition,so that it can characterize the degradation trend of the bearing more comprehensively and objectively.At the same time,combined with the traditional analysis method,the feature set of the bearing is constructed as the input data.In order to solve the phenomenon of poor prediction stability,low prediction accuracy and prediction lag caused by the interaction between features.This paper proposes a fusion neural network based on convolutional autoencoder,long short-term memory network and selfattention mechanism.First,the convolutional autoencoder is used to extract the deep features of the high-dimensional feature vector,and the input data is reconstructed so that it can map the degradation trend from multiple dimensions.At the same time,for the remaining service life information mapped by the degradation trend,dimensionality reduction is performed to extract important features and reduce the amount of model calculation.Secondly,the reconstructed data is input to the long-short-term memory network for time-series modeling,and the selfattention mechanism is used to strengthen the internal correlation of the data,adjust the neuron weight configuration,and improve the network’s ability to learn the input data.Finally,this paper uses the PHM2012 bearing data set to conduct experiments on the proposed fusion neural network and competition model.The experimental results show that the network proposed in this paper is reasonable and effective.Compared with the competition model,it has lower prediction error and better simulation performance.The combination of goodness and faster prediction speed effectively reduces the hysteresis prediction phenomenon and realizes accurate,stable and fast life prediction.On the basis of the above research,a visual rolling bearing remaining life prediction system is constructed.The algorithm and software are combined through a modular design,and custom parameters and early warning functions are added to it,so that real-time monitoring and warning can be carried out for bearings that have reached the life warning threshold.