Research Of Fault Detection And Diagnosis Based On Generalized Non-negative Matrix Projection Algorithm

In the increasingly complex process industry, fault detection and diagnosis (FDD)technique is crucial to guarantee the process safety, efficiency, product quality andespecially avoiding the disastrous accident. Because of the rapid development of thedistributed control system and the computer technology, massive amounts of historicaldata are stored in the database. If these data could be transformed into valuableinformation reflecting the process status, the well performing FDD model could beesablished in someway. Hence, the data-driven method becomes an active research areain the FDD field.As an emerging multivariate statistical analysis technique, negative matrixfactorization (NMF) has very important application significance in the research field ofintelligent information processing. NMF methods are able to learn the localized featureby the implementation of data compression and dimension reduction because of theadditivity, sparsity and interpretability, which performs better than conventional methodssuch as principal component analysis (PCA).Because of the great significance, this dissertation aims to establish the FDD modelfor the complex industrial processes based on NMF method. A new algorithm namedgeneralized non-negative matrix projection (GNMP) is proposed to be the core algorithmin order to handle the process data, and a new class of FDD models is established basedon GNMP algorithm. Furthermore, this GNMP model is extended in the aspect ofenhancing the robustness, adding the adaptivity and introducing the supervisorymechanism by the class information.Specifically, the main contributions of this dissertation are as follows:(1) Based on the existing theory and practice of NMF methods, this dissertationproposes a new algorithm with an embedded linear projection, called generalizednon-negative matrix projection (GNMP), which could relax the non-negativity restrictionfor the original data, get more orthogonal and sparser basis vectors, and converge theoretically.(2) In view of the non-Gaussianity of the complex process data, a new FDDapproach is proposed based on the aforementioned GNMP algorithm. Two types ofmonitoring indices, N2statistic and SPE statistic, and the corresponding contributionplots are designed for detecting the process status and diagnosing the fault location.Kernel density estimation (KDE) is used to estimate the underlying probability densityfunction and calculate the upper control limits of N2and SPE, respectively.(3) The robustness of aforementioned GNMP model is enhanced for the datamissing phenomenon of the complex process data. A new method call robust GNMP(RGNMP) is proposed based on GNMP and EM algorithm. RGNMP could estimate themissing values when decomposing the original data. Then, a robust FDD model isestablished based on RGNMP method.(4) Against to the time-varying characteristics of the complex processes, an adaptiveFDD method is proposed based on GNMP combining moving window technique, whichis call MWGNMP algorithm. This new approach could update the FDD model adaptivelywhen the steady statuses of processes change in order to make the pocess statuses stillunder control.(5) A supervised classification algorithm is proposed based on GNMP algorithmintegrating the classification ability of Fisher discriminant analysis, which is calledFGNMP algorithm. This new method could translate fault diagnosis into the multi-classclassification problem by making the most of class information of data. FGNMP methodcould achieve the superior diagnosis performance even though there are multiple faulttypes occurring in the complex processes.