A Process Fault Detection Method On The Low Dimensional Manifolds Obtained By Diffusion Maps

In the process of chemical industry,failure can be detected and eliminated in time and accurately can avoid serious consequences.The condition of the device in the modern industrial production is monitored by a variety of sensors,it results a large number of high-dimensional,non-linear data.In view of the characteristics of data,more and more scientists pay attention to manifold learning methods in the field of industrial process fault diagnosis.Considered the non-linear characteristics of the chemical process data and the noise interference in the process of data acquisition,diffusion mapping algorithm is investigated.The main research contents include the following aspects:1)Considering the linear fault detection method has low accuracy for nonlinear data,fault detection based on diffusion maps and support vector machine(DM-SVM)is researched.Diffusion maps is used to compress the data to reduce the storage cost of data,and the obtained low-dimensional feature vector is used as the input of SVM to detect faults,so as to promote the speed and accuracy of faulty detection globally.And using the standard data of TE process simulation to detect the fault,and the validity of the method is verified.2)The recognition effect of SVM is greatly influenced by its parameters.In view of this problem,the artificial fish swarm algorithm is used to optimize its parameters,it can improve the fault detection accuracy of the algorithm.Compare it with particle swarm optimization and genetic algorithm,TE process data is utilized to prove the rationality of the choice of algorithm.3)The Gaussian kernel function is used in the diffusion mapping algorithm to measure the similarity between the sample points by calculating the Euclidean distance between the sample points.How to measure the similarity accurately between samples is considered,a metric learning method based on Mahalanobis distance is researched.The goal is to learn a Mahalanobis matrix such that the distances of the point pairs with same type are as small as possible,while those with different type are as large as possible.Then,a more accurate and more realistic weight matrix is acquired to improve fault detection.Finally,the validity of the method is verified by experiments,and it can improve the level of fault detection.