| With the development of rotating machinery in the direction of large-scale,complex,high-speed,etc.,the reliability of rolling bearings has been put forward higher requirements.On the one hand,it is necessary to continuously improve the material quality and processing technology of the bearings;on the other hand,it is necessary to monitor and evaluate the bearing health status so as to maintain the equipment in time and avoid causing major accidents.Therefore,this thesis carries out the research on the method for the remaining useful life prediction of rolling bearings,which mainly includes:First,to address the problem that the traditional time-domain features cannot completely represent the degradation state change of rolling bearings,41-dimensional features from multiple domains are extracted,including: 10-dimensional time-domain features such as root mean square and rectified mean,16-dimensional time-frequency domain features of frequency band energy and energy ratio derived from the three-layer decomposition of wavelet packets and 15-dimensional features from the infogrambased negative entropy.Then,the redundant initial feature matrix is initially screened using the average correlation coefficient combining Pearson coefficients and Spearman coefficients to select the most optimal feature set.Based on the training set data under different working conditions from the open data of IEEE PHM 2012 Challenge,the screening results of two correlation coefficients of different bearings and their average coefficients are analyzed and compared to demonstrate that the features obtained based on the average coefficients are more balanced.Second,to address the problem of information redundancy caused by a large amount of feature data,a stacked auto-encoder-based feature fusion method is proposed,which utilizes the unsupervised learning feature of the auto-encoder to achieve compressed fusion of the initial optimal feature set and obtain the optimal health indicator curve of the bearing.The full training set of public data is used to verify the method.On the one hand,the health indicators obtained by directly inputting the original vibration signal data into the stacked auto-encoder have better monotonicity,robustness and trend performance than those obtained from the extracted and filtered feature matrix as input,which verifies the necessity of the previous work;on the other hand,compared with the health indicators obtained by dimensionality reduction using principal component analysis,the fused health indicators can completely express the change pattern of bearing health status and have good degradation consistency,which establishes a good foundation for the later prediction work.Then,for the problem that the life prediction method relies on mathematical models or a large amount of empirical data,a rolling bearing life prediction method combining particle filtering and fractional order gray model is proposed,which does not require complete modeling of the system and has more flexible application scenarios.Based on the traditional gray prediction,the fractional order cumulative operator is introduced to improve the accuracy of the prediction results.It is further combined with particle filtering algorithm to achieve accurate prediction of the remaining service life of rolling bearings.The test set data under different working conditions using publicly available data is validated,and the effectiveness and accuracy of the proposed method is demonstrated by comparing the prediction results with those derived from general exponential models with error analysis.Finally,in order to verify the validity and applicability of the proposed prediction model,two sets of data from the BPS rolling bearing life experiments were selected to complete the method control work in chapters 2,3 and 4,respectively.From the life prediction errors,it can be seen that the method in this paper has relatively better accuracy and applicability. |
