| Rolling element bearing is a key component in mechanical equipment,and its prognosis can predict its remaining service life,which helps to optimize production and maintenance management and provides an important basis for prognostics and health management of the whole equipment.In recent years,the applications of deep learning and industrial big data have proved their great potential in the research and application of prognostics.However,some open problems,such as insufficient feature extraction capability of deep network,unsatisfied accuracy and stability of prediction models,are still needed to improve.To solve these problems,this thesis focuses on bearing prognostics and studies the following three aspects: convolutional neural network(CNN)improvement,multi-task learning model and multi-model prediction with ensemble learning.The main research contents and contributions in this study are summarized as follows:(1)In view of the limited capability of convolutional neural network for complex data analysis and feature extraction,an improved CNN is proposed based on the efficient channel attention and Resnet.The former is used to automatically adjust the channel weights and enhance the feature extraction capability of the CNN,and the latter is used to solve the problem of performance decline caused by deep structure.On this basis,the information about the degradation time is added into the CNN and then a prediction procedure is designed based on multi-stage degradation of bearings.The prediction model is trained according to degradation stages to enhance its robustness.Experimental results indicate that the proposed improved CNN enhance the feature extraction capability and further improve the accuracy of bearing remaining useful lifetime(RUL)prediction.(2)Considering that there is a strong correlation between bearing degradation trend and the existing incipient defect(s),and it is difficult to make full use of failure information in bearing data sets,a multi-task learning model for RUL prediction is developed based on the deep CNN with wide first-layer kernel(WDCNN)and progressive layered extraction(PLE).The WDCNN can improve the efficiency of feature extraction on complex data sets,and the PLE can provide necessary information sharing for multi-task learning.Moreover,a method based on Wasserstein distance is presented to automatically identify the incipient defect.Fault diagnosis and health stage division are then used as auxiliary tasks for RUL prediction.The experimental results demonstrate that the constructed deep network can make full use of monitoring data and labels and use auxiliary tasks to improve the accuracy of main task,the RUL prediction,meanwhile,it also reduces the influence caused by the overfitting during network training.(3)Using a single prediction model,its prediction results always fluctuate and show different performances for different degradation stages.To solve this problem,a weight assignment method based on multi-task score(MT-Score)are designed for ensemble learning and then an improved multi-model stacking method(IMSM)is presented for bearing prognosis.Four deep networks with different structures or tasks are first selected as base models,and their weights are automatically assigned according to multi-task label identifications and prediction scores.Some of models are heterogeneous and aim to enhance the diversity of the ensemble learning,and the other models are homogeneous that can be used as auxiliary tasks to improve the performance of the main task,i.e.,the RUL prediction of bearings.The experimental results show that the IMSM can synthesize the advantages of multiple deep-base models.After combining with the multi-task learning,the prediction model can obtain accurate and stable RUL prediction results.Therefore,the developed deep networks and related methods in this study are suitable for intelligent feature extraction and RUL prediction of complex bearing sets,which is also a valuable reference for prognostics of other rotating machinery. |
PDF Full Text Request