SVDD Driven By Data And Space Structure Synergistically And Its Application In Fault Diagnosis Of Rolling Bearing

As one of the most commonly used components in modern industrial systems,rolling bearings may cause great economic losses and personal injury due to minor failures.Therefore,accurate fault diagnosis of rolling bearings can ensure the normal operation of the equipment,which has important practical significance.Due to the high cost of fault data acquisition,the rolling bearing data usually contains a large amount of normal data and only a small amount of fault data,which belongs to one-class classification problems.To this end,starting from the Support Vector Data Description(SVDD)method,the SVDD algorithm is deeply studied.Aiming at the problem that only the correlation of the data itself is considered in the modeling process,while the spatial structure of the data is ignored,this paper proposes the following three methods:1.A global and local joint regularized SVDD algorithm is proposed.SVDD does not consider the data distribution and assigns the same penalty parameter to each data,resulting in a high sensitivity in the selection of the penalty parameter.To this end,the algorithm proposes three distance measurement methods for Gaussian kernel space based on the global and local structure of the data,and designs the probability density relative to the global and local,then based on this density,the penalty parameter of information entropy for regularizing SVDD is designed,the algorithm enables the use of more information and reduced information uncertainty in fault diagnosis.Experiments on various UCI public datasets including bearing datasets show that this algorithm has better detection performance than other existing algorithms.2.A distribution entropy penalized SVDD algorithm is proposed.Generally,in the learning task of anomaly detection and fault diagnosis,the number of normal samples is far more than that of negative samples,but negative samples can also provide corresponding information,making the classification more accurate.To this end,the algorithm designs the information entropy of normal samples and negative samples by using the probability densities relative to the global,respectively regularizes the penalty parameters of SVDD for normal samples and the penalty parameters for negative samples,and extends the data structure distribution to include in the SVDD model of a small number of negative samples.Experiments on various anomaly detection and rolling bearing datasets show that the algorithm improves the accuracy of anomaly data detection to a certain extent.3.A feature entropy regularized SVDD algorithm is proposed.SVDD directly uses the original data in the modeling process,and treats each feature attribute equally,without considering the subspace structure of the data.To this end,the algorithm considers the contribution of each feature of the sample,uses the feature weight to improve the robustness of the data,and uses the feature entropy of the weight to regularize the SVDD,and uses the alternate update strategy to design an optimization method.Feature weights enhance SVDD's utilization of data subspace structure information.The effectiveness of the optimized feature weights is verified using artificial two-dimensional data and various real datasets,at the same time,experiments on various anomaly detection data sets show that the algorithm effectively improves the detection ability of negative samples.