Research On Health Indicator Assessment Method Of Rolling Bearings Based On Life Information And Deep Time Series Features

In recent years,equipment fault diagnosis and life prediction have attracted great attention among all kinds of fields of society under the impetus of a new round of digital technological revolution of artificial intelligence.It has become a popular trend in the Prognostics and Health Management(PHM)field to use machine and deep learning methods to evaluate the Health status of equipment.Feature engineering preprocessing is often the most critical pre-step of prediction algorithm model.In this paper,rolling bearings,a key component of the most common rotating equipment,are taken as the research object.Deep learning technology and transfer learning skills are used to solve the problem that the degradation time scale and the different distribution of degradation data are ignored.A series feature extraction method was proposed for vibration signals of life cycle under different working conditions,and a feature engineering preprocessing algorithm was constructed to measure the health index of rolling bearings.The main work and contributions of this paper are as follows:(1)CMIE-F&E signal preprocessing method based on continuous mutual information entropy is proposed.Aiming at the task of building health indicators of rolling bearings in the whole life cycle,a selective filtering enhancement algorithm was designed based on the degradation law of the whole life cycle.Different frequency component in the whole development process of degradation of life contains different degradation information,the traditional filtering algorithms are usually based on a component of the vibration signal is studied under static moment,based on some rules will weed out some of those components,and the rule applies to rolling bearing the whole life cycle,the final signal filtering.The proposed CMIE-F&E signal preprocessing method is based on the signal changes in the whole life cycle.The principle,process and necessity of signal processing are explained in detail.The results of this method are compared with those of other filtering methods through experiments,and the effectiveness of this method is verified.(2)SAM sequential feature extraction method based on attention mechanism is proposed.Traditional feature extraction methods usually take the signal statistical features of continuous sampling values at a certain time as input features,ignoring the problem that the time scale of degradation which is smaller than the sampling time scale.The proposed SAM sequence feature extraction method studies the information changes inside the sampled samples,and realizes the extraction of the degraded information inside the sampled samples through the self-attention vector recombination of vibration signals at the sampling time.On the basis of introducing the basic ideas of relevant feature extraction methods,the model structure and super parameter selection of the proposed algorithm are explained in detail.By comparing the health index construction and life prediction results based on this feature and those not based on this feature,the superiority of SAM timing feature extraction is illustrated at different levels.(3)An adaptive method of SAM-CTR health indicator construction based on joint distribution is proposed.To solve the inconsistent across the data distribution under the condition of rolling bearings and data,model the reuse of difficult problems,improve the accuracy of the model prediction and generalization ability,will be based on joint distribution depth migration of the adaptive learning classification method is introduced to health indicators to build regression problems,put forward a kind of health indicators to construct model of combination of classification and regression.In this chapter,the basic theory of the domain-distributed adaptive method of deep transfer learning is described in detail.Then,the structure,optimization strategy,training process and key hyperparameter selection of the proposed SAM-CTR health indicator construction network model are described in detail.The advantages of joint distribution adaptive in the task of health index construction and the effectiveness of the proposed SAM-CTR algorithm are fully illustrated by the comparative experimental results in training and testing stages by selecting the appropriate location of adaptive layer.