Research Of An Unsupervised Fault Diagnosis Method Based On Multi-manifold Spectral Clustering

The deep promotion of smart manufacturing has made the data generated by smart manufacturing system increase explosively.How to excavate the value of these data to serve for the operation optimization and decision-making of manufacturing system,has become a hot issue in current research.Machine learning(ML)is one of the most commonly used methods in manufacturing data analysis and processing.However,it is hard to obtain high quality labeled data in practical manufacturing systems.Therefore,it has great application prospects in studying unsupervised machine learning methods to handle massive unlabeled data.Spectral clustering is an unsupervised manifold learning method,which can uncover the intrinsic structure hidden in high-dimensional non-linear data and divide them into different clusters according to the intrinsic structure.However,spectral clustering assumes that the data are located on a single manifold,while manufacturing data are usually lied on multi-manifolds.Therefore,an improved spectral clustering algorithm is proposed for the data distributed on multi-manifolds,and applied to fault diagnosis in this research.Moreover,a fast fault diagnosis method for the new data which named “out of sample” is proposed.In this research,an iterative multi-manifold spectral clustering(IMMSC)is proposed for the data distributed on multi-manifolds.For the condition that the existence of intersection and overlap in the data distributed on multi-manifolds,the affinity matrix is constructed based on the similarities of local tangent spaces.A more precise data structure is learned by an iterative optimization process which makes the weights of adjacent points from different manifolds approach to 0,while the weights of adjacent points belonging to the same manifold relatively large,so that the performance of clustering is improved.The proposed method is tested on 5 simulated datasets.The results show that the proposed method can achieve better clustering performance while maintaining the stability of clustering results compared with traditional clustering methods and other spectral clustering methods.An unsupervised fault diagnosis based on IMMSC is proposed for the unlabeled characteristics of manufacturing data.The time domain analysis method is conducted to extract the original high-dimensional features from the vibration signal.Then,the IMMSC is utilized to obtain the mapping from the high-dimensional space to the low-dimensional feature space,and divide the data into different clusters.The low-dimensional features which contains the fault information are also the obtained.After that,the Local Outlier Factor(LOF)algorithm is applied on the low-dimensional features to identify the normal condition.Finally,the fault index obtained from the low-dimensional features which represents different faults is utilized to recognize different faults.The proposed fault diagnosis method is applied on the motor bearing dataset from Case Western Reserve University(CWRU).The results indicate that the proposed method can identify normal condition and different levels of faults accurately.For the new samples,an efficient solution that extract the low-dimensional features of new samples is proposed,and the fault types of new samples are identified based on the lowdimensional features.The unsupervised fault diagnosis based on IMMSC is conducted on the known data to identify the fault types,learn the intrinsic structure of data and the knowledge related to the faults.Then the low-dimensional features of the new samples are estimated based on the Nystr?m formula and the data structure.After that,the fault types of new samples are recognized by the low-dimensional features and the knowledge related to the faults.The proposed method is evaluated on the bearing data from CWRU.The results show that the proposed method can identify the faults of new samples effectively and accurately.Finally,a summary of this paper is given,and some future issues that worth to be studied are presented.