Deep Learning Based Fault Diagnosis And Life Prediction For Ball Screw

As one of the key components of mechanical equipment,ball screw subs contain rolling bearings and ball screw components,which are widely used in processing and manufacturing,chemical metallurgy,aerospace and other fields,it is of great engineering value to accurately evaluate the operating condition and accurately predict the remaining life of ball screw subs.However,the accurate capture of fault characteristics from sensor data,multi-sensor information fusion strategy,fault diagnosis and remaining life prediction model construction are recognized technical difficulties in the industry.In this paper,we aim at the key technical bottlenecks in this field,and carry out the research on deep learning-based fault diagnosis and life prediction of ball screw subsets,aiming at providing technical support for ball screw subsets fault warning and predictive maintenance.The main research contents are as follows.1.For the problem of fault feature extraction in high-precision intelligent diagnosis of ball screw sub,two kinds of multi-type diagnosis methods combining feature extraction method and deep learning are proposed.One is the feature extraction method based on frequency domain information: the time domain information is Fourier transformed into frequency domain information.The other is a multi-domain feature extraction method based on multi-domain features: time domain,frequency domain and time-frequency domain features are extracted as fault state information.Both methods are input to the convolutional capsule neural network for fault diagnosis under multiple scenarios,and the experiments show that both feature extraction methods can better represent the fault state information,which lays the foundation for realizing highprecision fault diagnosis of ball screw sub.2.In order to enhance the correspondence between the degradation information of vibration signal and life label,two life prediction methods are proposed.One is the life prediction method based on integral correction and global attention mechanism: firstly,the time domain features are transformed into integral features by integral correction;finally,they are input to the long and short-term memory network optimized based on global attention mechanism.The other is the life prediction method based on the degradation stage division and attention mechanism: first,the degradation stage division is performed,and the wavelet packet energy spectrum of the fast degradation stage is selected as the degradation information;finally,it is input to the temporal convolutional network optimized based on the attention mechanism.After the life prediction experiments of multiple scenarios,the effectiveness of the two life prediction methods is shown.3.For the problem of incomplete single-sensor monitoring,two lifetime prediction methods are proposed.One is the life prediction method based on the convolutional gated recurrent unit network: first,the time domain,frequency domain and time-frequency domain features of the multi-sensor are extracted;then,the multi-sensing information is input to the convolutional gated recurrent unit network to fuse the feature layer.The other is the life prediction method based on parallel convolutional neural network: first,wavelet noise reduction is applied to the multi-sensor information;then,the noise reduced single sensor information is input to the parallel convolutional neural network to achieve the multi-sensor information fusion in the data layer.The experimental results show that the fusion of the two multi-sensor information avoids the problem of inadequate perception of single sensor information and shows a better life prediction effect.4.In order to realize the fast reconfiguration of life prediction models under different working conditions and scenarios,a life prediction method based on TCA migration learning is proposed.Firstly,the time-domain features are extracted on the source and target domains;secondly,TCA migration learning is used to realize the feature migration between the source and target domains;then,the TCA-migrated source domain data are modeled in the life prediction model,and finally,the TCA-migrated target domain data are directly input to the life prediction model of the source domain.The method has been experimentally verified to achieve the migration of life prediction models under different working conditions and scenarios,and has high industrial application value.5.A Lab VIEW-based real-time intelligent fault diagnosis and life prediction system for ball screw subs has been developed for the actual industrial demand of ball screw sub research.The system provides users with real-time status data analysis service of ball screw subs,which is important for reducing the maintenance cost of ball screw subs.