Evaluation And Optimization Of Performance Degradation Characteristics For Bearing Remaining Life Prediction

Bearings are widely used in various rotating machinery,such as motors,fans,running parts of high-speed trains,etc.Under harsh operating conditions,bearings can fail,causing mechanical systems to fail to function properly.Condition monitoring and health management of bearings play a positive role in maintaining the safety and reliability of mechanical systems.Data show that nearly half of all motor failures are related to bearing degradation.Predicting the remaining service life of a bearing enables monitoring of bearing operating status and early warning of failures,enabling timely replacement of failed bearings and reduced downtime.Therefore,it is of great significance to predict the remaining service life of the bearing.Advances in deep learning and industrial big data have fueled a boom in data-driven forecasting methods.Feature extraction is a key step in data-driven methods.The quality of features is directly related to the accuracy of subsequent prediction models.Among them,feature evaluation methods are analytic and versatile,and are widely used in extraction work.However,the current bearing feature evaluation system has certain limitations,and the correlation between the selected features is not considered enough.At the same time,as a feature post-processing method,the effect of feature evaluation depends heavily on the quality of the initial features.The existing feature extraction methods rarely consider the individual differences of bearings,and the degradation characteristics of different bearings have certain differences,which ultimately lead to poor prediction results.In view of the above problems,the main research results of this paper are as follows:(1)Aiming at the problem of how to use the feature evaluation method to optimize the features and predict the remaining life,a study on the optimization method of bearing performance degradation feature evaluation is proposed.Firstly,the convolutional autoencoding network is used to extract the initial features in the frequency domain signal of the bearing;then the features are scored with the help of three evaluation indicators of correlation,monotonicity and robustness,and the scores of the three evaluation indicators are linearly weighted as each The final score of dimensional features is normalized,and good features are selected from the initial features according to the threshold of 0.5;finally,a long short-term memory network prediction model is built,and the input data is used to train the prediction model.The effectiveness of the feature evaluation method is verified by experiments.(2)Aiming at the problem that the current bearing feature evaluation system does not consider the correlation between the selected features enough,a method for optimizing the bearing performance degradation feature based on clustering and evaluation is proposed.In order to overcome the original shortcomings of the feature evaluation system,the clustering method is optimized on the basis of feature evaluation.The specific steps are: first construct an improved Kmeans clustering algorithm based on correlation,and cluster the features according to the correlation;after the clustering is completed,perform feature evaluation on each type of feature,screen out the excellent features,and perform feature fusion as a followup prediction feature.Experiments show that the feature optimization method of feature clustering and evaluation is better than the single feature evaluation method.(3)Aiming at the problem that the feature evaluation method depends heavily on the goodness of the initial features,taking correlation as an example,a temporal correlation enhancement optimization method for bearing performance degradation features is proposed.A feature correlation constraint is constructed within the convolutional autoencoder learning framework,and a feature correlation constrained convolutional autoencoder model is proposed by improving the loss function of the convolutional autoencoder.The model realizes the adaptive extraction of lifetime high-correlation features in the frequency domain,which enhances the temporal correlation of the proposed features.Experiments show that the feature correlation constrained convolutional auto-encoding model has higher prediction accuracy than the auto-encoding model and the correlation feature selection method.(4)Aiming at the problem of bearing individual differences that are not considered in the feature extraction process,a feature optimization method to reduce individual differences in bearing performance degradation is proposed.The sample data of the same label is input in parallel,the trend consistency is calculated,and the difference of the same degradation feature on different bearing individuals is quantified.By constructing trend consistency constraints and combining with convolutional autoencoders,a feature extraction model of trend consistency constrained convolutional autoencoders is built.The model can automatically enhance the trend consistency of features in the process of feature extraction,thereby reducing the individual differences of bearing features.The feasibility of the method is verified on the bearing public dataset.