Font Size: a A A

Fault Detection And Identification For Tensor Data Based On Multilinear Principal Component Analysis

Posted on:2020-01-05Degree:MasterType:Thesis
Country:ChinaCandidate:Y L WangFull Text:PDF
GTID:2428330623963594Subject:Control engineering
Abstract/Summary:PDF Full Text Request
The process monitoring of industrial system is of great significance for ensuring the safe operation of equipments,improving production efficiency and product quality.With the popularity of distributed control systems and intelligent instruments,a large number of sensor data can be recorded.Datadriven fault detection and identification methods can mine hidden features representing process states from a large number of historical data,and has a wide range of applications in the actual process.As a natural representation of multidimensional arrays,tensor can record numerical information while retaining the structural information of data.In this paper,based on the topological structure and dynamic characteristics of the system,process data in tensor form is constructed to store more industrial process information.However,in the face of tensor data,it is usually necessary to expand it into vector form in order to apply to most kinds of learning algorithms.In order to reduce the loss of structural information,this paper introduces multi-linear principal component analysis(MPCA)and support high-order tensor machine(SHTM)methods to establish a tensor space fault detection and identification model.The specific contents of this paper are as follows:(1)The concept of tensor and its related operations are introduced from the geometric and algebraic levels.Then,according to the characteristics of some industrial systems,such as structure and time series,an appropriate high-order tensor form of data representation is established.(2)Tensor distance(TD)can reasonably measure the location relationship between high-order data.In order to realize the fault detection of tensor samples in industrial process and reduce training time,a fault detection model based on tensor distance is proposed.In the input space,the process data can be quickly detected by measuring the distance between samples and the data center.(3)In order to reduce manual participation and improve the efficiency of fault detection,multi-linear principal component analysis(MPCA)algorithm is used to extract the main features of data in high-dimensional variable space to avoid damaging the high-order structure and intrinsic correlation of the original tensor.A fault detection model based on MPCA is proposed,meanwhile,two monitoring statistics R and E and corresponding control limits are designed to solve the fault detection problem.(4)On the basis of PCA-SVM fault identification model of first-order tensor,a MPCA-SHTM fault identification model suitable for any order tensor is proposed.Without expanding the data into vector form,the numerical records and structural features of tensor samples are directly learned,which reduces the risk of dimension disaster and over-fitting,to realize the classification and identification of industrial system faults.
Keywords/Search Tags:Tensor Distance, Multilinear Principal Component Analysis, Supporting High-Order Tensor Machines, Fault Detection and Recognition
PDF Full Text Request
Related items