Research On Prognosis Of Roller Bearings Based On Sparse Coding

Roller bearings are the key components of rotating machinery for the safe and reliable operation.The failure of roller bearings will seriously affect the useful life and production efficiency of the equipment.Aiming at the defect frequently occurred parts in the rotating machinery,such as bearings,fault trend and remaining useful life prediction techniques have been conducted in this work,which can significantly reduce maintenance costs of equipment,improve the management of enterprises,avoid severe accidents and conduct condition-based maintenance instead of breakdown maintenance.Therefore,the study of the thesis has great theoretical and engineering significances.Based on the time-domain features of vibration signals,some new methods have been developed and applied to bearing fault prediction based on deep learning theory and time series analysis.First,the paper reviewed data-driven fault prediction techniques with machine learning.For example,the traditional Bagging-based decision tree method and support vector regression combined with particle swarm optimization have been used to predict the trend of bearing via time-domain features.Resutls showed that two machine learning methods not only require cumbersome parameter tuning,but also need to improve the robustness and prediction accuracy.Moreover,a novel deep learning method has been proposed in this paper in order to solve the shortcomings of traditional machine learning methods.Sparse coding method is one of deep learning techniques,which is emphatically utlized in bearing fault prediction.The proposed method use unsupervised dictionary learning and repeated iterations to achieve coefficients and can deeply mine signatures of the data.Moreover,sparse signal can be reconstructed that will effectively reduce the redundant information of the dataset.The vibration data in the bearing full lifetime has been collected on the bearing test platform in Intelligent Maintenance System(IMS).Results have damonstrated that sparse coding method has higher robustness and prediction accuracy than traditional machine learning methods.Next,a novel weight constraint sparse coding(WCSC)algorithm has been developed to further improve the prediction performance of sparse coding model.Basis Pursuit De Noising(BPDN)algorithm is used in the WCSC instead of the basis pursuit(BP)algorithm because BP is not suitable for practical applications,especially in the presence of noise.Since excessively sparse is caused by traditional sparse coding model will reduce the predicting accuracy,the local linear embedding(LLE)algorithm is adopted to improve the sparsity constraint rule,namely the weight constraint sparse coding.Results of bearing run-to-failure data originated from IMS shows the proposed WCSC is much more effective than the original one.Finally,a hierarchical sparse coding network framework with multi-layer structure is developed based on deep learning technique.The method increases the Max-Pooling layer to form a network unit based on WCSC.The new multi-layer sparse coding network combined with greedy learning features can effectively extract high-dimensional features in the signal.Experimental results well demonstrate the effectiveness of the hierarchical sparse coding for bearing fault trend and remaining useful life prediction,compared with two sparse coding models.