Study On Algorithm For Rolling Bearing Remaining Useful Life Prediction And Development Of Monitor Software

The rolling bearing is called “the joint of industry”,whose health state has a great influence on the stability of the machine.Considered the contradiction between the lower risk of disaster caused by broken bearing and the lower cost of bearing maintenance,the best win-win strategy is constantly bearing monitoring and real time remaining useful life prediction.Besides,the first 10-year national plan for transforming China's manufacturing,entitled "Made in China 2025",also clearly pointed out that the technology and the application of machine reliability should be developed.Therefore,it is important to do research on the degradation of bearing and achieve accurate real time bearing remaining useful life prediction with modernization techniques.Considered the problem that the life-cycle data of bearing is too long to train a well recurrent neural network,this paper proposed to use convolution and attention mechanism.First,life-cycle data are compressed along the time axis to some extent via using convolution,then attention mechanism helps recurrent neural network to deal with a long sequence of data and make a prediction.After tested on the test data,the result proved that it makes a more accurate prediction than other methods,which also shows that artificial neural network is a promising tool in the problem of bearing remaining useful life prediction.However,such neural networks with complex structure always costs much time,especially the recurrent neural network that deals with over long sequences.This makes it unable to give a prediction in a second.For this reason,this paper suggests to use a neural network to optimize the process of feature extraction from bearing vibration signals.By comparing the self-supervised learning methods based on autoencoder the ones based on contrastive learning,we found that the extracted features from later ones are more representative of the health state of bearing.After that,the process sequence also should be shorter to reduce the running time of the recurrent neural network.Therefore,based on the degradation theory of bearing,this paper suggests to abandon some old data in the sequence,the older the more are abandoned.And use position encoding to keep the information of time.As a result,this sequence of cycle-life data is shortened but lost little information about the change of bearing health state.Finally,by using regularization strategies,reg LSTM is able to make an accurate prediction for a second.In the end,a software system for real time bearing health state monitoring and remaining useful life prediction is designed in C#.By using modularization programming method and interface,this software system is less depending on specific data acquisition card and allow rapid development in signal processing algorithms,including the ones based on deep learning thanks for the Tensor Flow.net package.Then a test is carried out with test data as simulation input,and the result identified that such neural networks are able to make a real time accurate prediction in bearing remaining useful life.