Research On Data-driven Approach For Fault Diagnosis And Degration Process Assessment On Rolling Bearings

Rotating machinery is a kind of indispensable equipment in current industrial production processes,among which rolling bearings are one of the most important and vulnerable components.Rolling Bearing failure not only affects the operating accuracy of the equipment but also compromises production safety.Therefore,it is necessary to monitor the operating status of bearings.This paper mainly uses data-driven machine learning technology to study the problems of fault diagnosis and performance degradation assessment in bearing health management systems,following the mainline of "single working condition fault diagnosis-multi-working condition fault diagnosis-performance degradation assessment",and the specific content includes the following six aspects:(1)To address the issue of high parameter complexity in single working condition fault diagnosis models based on traditional neural networks,a non-iterative training method for fault diagnosis models based on uniformly distributed and normally distributed random functional connection networks are proposed.The multiplication congruence method and the Box Muller method are used to construct random factors as weights between input and hidden layers.Comparative experiments demonstrate that the normally distributed random parameter neural network has better diagnostic results.(2)To further enhance the application of the random parameter neural network model in bearing fault diagnosis,an improved rolling bearing fault diagnosis method based on a Stochastic Configuration Network(SCN)is proposed.By optimizing the constraint conditions of SCN’s random weights,the accuracy of fault diagnosis is improved.Experimental analysis compares the improved SCN fault diagnosis model with various traditional machine learning-based fault diagnosis models,verifying the effectiveness of the method.(3)To address the inconsistency between the model training condition and the actual working condition and the insufficiency of labeled sample data in the actual working condition,a multi-working condition rolling bearing fault diagnosis method based on domain adaptation Extreme Learning Machine(DA-ELM)is proposed.Introducing domain adaptation technology into ELM from the perspective of regularization terms,one ELM each for the source and target domain working conditions are allocated.The source domain ELM extracts the source domain working condition patterns,and the target domain ELM achieves a more reliable rolling bearing fault diagnosis under the target domain with a small number of labeled samples.Compared to other domain-adaptive ELM methods,this method has better generalization ability.(4)To address the difficulty in controlling the participation of the source and target domains in the domain-adaptive fault diagnosis model under multi-working conditions and the inability to automatically extract temporal features,a domain-adaptive method based on bilinear Convolutional Neural Network(CNN)is proposed.First,the CNN is used to automatically extract fault features in the source domain working condition.Then,the bilinear mechanism transfers the source domain pattern to the target domain,combining with a small amount of prior knowledge in the target domain,to conduct reliable rolling bearing fault diagnosis with a diagnostic accuracy rate of over 98%.(5)For the delay in determining the initial prediction time in rolling bearing remaining life prediction,the initial fault time from the perspectives of dynamic3-interval and voting mechanism are determined and the remaining life of the rolling bearing using an Long Short Term Memory-based method combined with a double exponential model is predicted,establishing a bearing performance degradation assessment method.Due to its excellent real-time performance,this method can be used for rapid online detection.(6)To address the issue of insufficient prediction reliability analysis in existing bearing performance degradation assessment methods,a rolling bearing performance degradation assessment framework based on Bayesian neural networks from the perspective of uncertainty modeling is proposed,further improving the reliability of the prediction.Multi-dimensional sensitive feature fusion is used to screen features,and an autoencoder is used to reconstruct features to determine the initial fault time.A one-dimensional Bayesian neural network is then used to predict the remaining life of the bearing.In comparison with various deep Bayesian network-based models,our method not only has excellent predictive performance but also provides a quantification method for the uncertainty of the predictions.