| As an important part of common machinery equipment,rolling bearings are prone to various kinds of faults under complex working conditions like long-term heavy load and strong impact,etc.Faulty bearings will cause the performance deterioration of whole machinery.Successful detection of bearing fault at an initial stage will be helpful to make timely maintenance and avoid serious accident occurrence.Therefore,accurate and reliable detection and diagnosis at the early stage of fault occurrence is considered as a key step of fault prognostic and health management（PHM）.However,in actual engineering,the current early fault detection technology for rolling bearings is still immature,especially the robustness of the detection results.The main challenges are as follows: 1)There are differences in the distribution of bearing data under different working conditions,and it is difficult to build an offline model that can be adapted to online data;2)There is noise during normal operation of the bearing,and the abnormal fluctuations caused will interfere with the detection model;3)Due to the constraints of online scenarios,the amount of available data for the target bearing is insufficient,and the feature representation ability is not good.To solve the above problems,we start to improve data characterization capabilities,depth-based neural network algorithm framework,the introduction of technology to enhance learning and data migration are based on robust feature extraction based on the cross-bearing condition data distribution and multi-scale adaptation The effective representation of feature information can solve the problem of poor robustness of early fault detection models caused by inconsistent online data distribution and insufficient feature representation capabilities.The main research contents include:（1）Aiming at the problem of mismatch between online data and offline trained models caused by inconsistent bearing data distribution,from the perspective of transfer learning,a robust detection method for early fault of rolling bearings based on deep transfer learning is proposed.Firstly,by constraining the data distribution between the normal data of different bearings to train the domain adaptive Auto-encoder,so as to obtain the common feature representation of the normal state data of different bearings;Secondly,based on the extracted common feature representation,a method of bearing state assessment based on Robust Deep Auto-encoder（RDA）is proposed.This method can accurately identify the boundary between the normal state and the early fault state in the low-rank feature space of the bearing data;Finally,a support vector machine detection model is constructed to detect early faults according to the results of the state division of auxiliary bearings.Experimental results on the IEEE PHM Challenge 2012 bearing dataset and XJTU-SY bearing dataset show that this method can outperform existing detection methods in terms of detection accuracy and false alarm rate.（2）Aiming at the problem of insufficient data feature representation in normal state of the bearing caused by insufficient online detection of target bearing data,from the perspective of data augmentation technology,a multi-scale robust anomaly detection method for early failure detection of is proposed.This method first incorporates the data augmentation technology into the framework of the RDA and proposes a robust multiscale Deep Support Vector Data Description（Deep-SVDD）model,and constructs multiple SVDD hypersphere in the low-rank feature space of RDA,so that a robust multi-scale Deep-SVDD model is proposed,which integrates multi-scale feature information to comprehensively determine the anomaly probability of data samples;secondly,a data augmentation method based on vibration signals is proposed,so that the abnormal detection method based on data augmentation can be applied to bearings dataset;Finally,simulation experiments are carried out on the IEEE PHM Challenge 2012 bearing dataset and XJTU-SY bearing dataset.The results show that the method is sensitive to incipient fault and the detection results are stable.Compared with the existing detection methods,it has better robustness.In summary,this article focuses on the robustness of incipient fault detection,from the perspectives of constructing cross-condition domain adaptive models and improving the ability of data feature representation,which provides a way to improve the robustness of incipient fault detection methods.This article also provide a useful reference for real-time fault warning of various rotating machinery,and have significant academic research and engineering application value. |
