Research On Fault Diagnosis Of Industrial Imbalanced Data Based On Manifold Learning

With the progress of technology,industrial processes are developing towards intelligence.As the equipment generally works in a normal state and is only in a fault state for a short time,this causes an unbalanced character of the various types of fault data collected,which makes fault diagnosis difficult.At the same time,the collected data is also characterised by unbalance,nonlinearity,strong coupling and having a manifold structure.How to generate data and extract effective features for unbalanced fault samples becomes an important factor to improve the diagnostic accuracy.Compared to classic fault diagnosis methods,the manifold learning algorithm is able to retain the local structural information of the data and extract the main features related to the fault.Locality Sensitive Discriminant Analysis(LSDA),a local manifold learning method,is able to maintain the local geometric structure while making full use of class labels for discriminant analysis.Therefore,this thesis makes use of two imbalance fault diagnosis methods based on improved SMOTE with integrated LSDA to model chemical processes as follows.(1)This thesis first proposes a fault diagnosis method based on the Cluster Synthetic Minority Over-Sampling Technique(Cluster-SMOTE)with integrated LSDA--the c SMOTE-LSDA,in which c SMOTE clusters the minority classes by means of the K-Means algorithm,first dividing a number of minority clusters,and then applying the SMOTE algorithm separately.This algorithm not only removes outliers and noise,but also makes the synthesised new samples all lie within the same minority class,avoiding the generation of overlapping samples between classes and greatly reducing the computational effort.The balanced high-dimensional,non-linear samples are then dimensionally reduced using the LSDA algorithm of manifold learning to extract valid data features,and the Ada Boost.M2 multi-classifier is used for fault classification.(2)In an effort to further improve the quality of minority class fault sample generation and fault diagnosis accuracy,this thesis also proposes a fault diagnosis method based on Improved Synthetic Minority Over-Sampling Technique(ISMOTE)integrated LSDA--ISMOTE-LSDA.ISMOTE uses both Euclidean and cosine distances to jointly measure the proximity between two points,taking into account both linear and spatial geometric locations,resulting in a more rational selection of the k nearest neighbours of the sample,leading to more efficient instances.The algorithm successfully solves the problem that the synthesis of c SMOTE is monotonic and does not consider the spatial location,and improves the accuracy of fault diagnosis.(3)Finally,this thesis uses the Tennessee Eastman process dataset as the research object to verify the performance of these two methods separately.The results show that the c SMOTE-LSDA method proposed in this thesis has better fault diagnosis accuracy than the traditional fault diagnosis method,and the improved ISMOTE-LSDA further improves the quality of minority class sample generation and fault diagnosis accuracy,which is more suitable for unbalanced fault diagnosis problems.