Remaining Useful Life Prediction Of Rolling Bearings Based On Graph Neural Network

Bearings are a commonly used component in mechanical equipment,providing support and rotational support for moving parts.However,as the operating time increases,bearings gradually wear and age,eventually leading to equipment failure and downtime.Predicting the remaining life of bearings can help companies detect bearing failure risks in advance,perform maintenance and replacement in a timely manner,and avoid equipment damage and downtime losses.Predicting the remaining life of bearings can greatly improve equipment reliability and operational efficiency,reduce maintenance and replacement costs,and also ensure production and job safety.Therefore,research on bearing remaining life prediction has significant importance in fields such as mechanical manufacturing,industrial automation,and transportation.This article mainly covers the following research topics:1.This article summarizes and introduces the mainstream methods for extracting features from vibration signals,and analyzes the features extracted.The analysis shows that the features extracted by the mainstream methods cannot fully represent the degradation of bearings and cannot accurately predict the remaining life of bearings.Therefore,based on the multi-step mean sampling method,combined with Long Short-Term Memory(LSTM)network,this article proposes the Multi-step Mean Sampling Long Short-Term Memory(MSSLSTM)network.The MSSLSTM network can perform mean sampling of vibration data with multiple step sizes and window sizes simultaneously,and fully leverage the ability of LSTM to process time-domain data.This network is used to extract the health features of rolling bearings,which have better trend,monotonicity,and robustness,providing reliable inputs for the subsequent prediction of bearing remaining life.2.This article proposes a Simulate Anneal Cuckoo Search Algorithm(SACS)based on the simulated annealing method to improve the Cuckoo Search Algorithm(CS)which is affected by the fixed probability set by experience and influences its global and local search capabilities.The SACS algorithm dynamically adjusts the probability of CS algorithm with the search rounds,replacing the fixed probability with a variable one.The algorithm is validated on four test functions,and the results show that SACS has advantages in both optimization ability and speed.This provides a better optimization algorithm for optimizing neural network hyperparameters,while saving a significant amount of human and material resources.3.This article proposes a Spearman-Graph Neural Network(SGNN)that uses Spearman correlation coefficient optimization to address the issue of time series data lacking graph structures and thus cannot be used as data for graph neural networks.The network first calculates the correlation between different variables using Spearman correlation coefficients to construct a variable correlation matrix.Then,this matrix is used as the input for the graph neural network and is adaptively adjusted.The article combines this network with the MSSLSTM feature extraction method to propose a complete MSSLSTM-SGNN bearing remaining life prediction method.This method predicts remaining life using the PHM2012 dataset and is compared with the prediction results of LSTM,BP neural network,and radial basis function neural network.The results show that the MSSLSTM-SGNN proposed in this article has the lowest root mean square error,mean absolute percentage error,and mean absolute error,proving its effectiveness.