| To ensure the reliability and safety of modern large-scale industrial processes,data-driven process monitoring methods have received a lot of attention and application,especially multivariate statistical methods.However,real industrial processes and systems are often characterized by dynamic and nonlinear features.In addition,process data also often have non-Gaussian characteristics.Using traditional multivariate statistical methods,satisfactory process monitoring performance cannot be obtained.This topic focuses on data-driven process monitoring and fault detection methods and proposes a series of process monitoring and fault detection methods based on the analysis of typical variables to provide assurance for the safe operation of the actual blast furnace ironmaking process and wind turbine.The main research contents of this thesis are as follows:1)To address the dynamic nature of industrial processes,this paper proposes a fault detection method based on a combination of Canonical Variate Analysis(CVA)and Support Vector Data Description(SVDD).First,the variables that have a large impact on the blast furnace production process are selected using the Difference in Coefficient of Variation(DCV)method.Then,CVA was used to extract the dynamic features of the process data.Based on the extracted dynamic features,SVDD is used to construct statistical information to solve the non-Gaussian problem of the process.For the Lagrangian optimization problem in the SVDD algorithm,the proposed method uses the Sequential minimum optimization(SMO)algorithm to calculate the Lagrangian multipliers,which decreases the computational complexity.Finally,the effectiveness of the proposed method is verified by using the collected operational data of the actual blast furnace ironmaking process.2)To address the nonlinearity of industrial systems,a new fault detection method for Floating Offshore Wind Turbines(FOWT)is proposed based on the Kernel Canonical Variate Analysis(KCVA)method.First,KCVA is used to extract the dynamic and nonlinear features of the floating offshore wind turbine to solve the dynamic and nonlinear problems inherent in the FOWT model.Then,based on the features extracted by KCVA,T~2and Square Prediction Error(SPE)statistics are established to monitor the systematic variation of the FOWT model.To address the non-Gaussian characteristics of the FOWT operational data,the paper uses the Kernel Density Estimation(KDE)method to calculate the thresholds for the T~2and SPE statistics.Finally,the Fatigue,Aerodynamics,Structures,and Turbulence(FAST)v8simulator designed by the National Renewable Energy Laboratory(NREL)is used to construct a 10 MW FOWT.The 10 MW FOWT simulation model was constructed to verify the superiority of the proposed fault detection method.3)To address the dynamic and nonlinear problems of industrial processes,this paper combines ensemble learning with kernel canonical variate analysis and proposes a new fault detection method of Ensemble Kernel Canonical Variate Analysis(EKCVA).First,multiple KCVA sub-models are constructed using the ensemble learning approach to better characterize the nonlinear and dynamic features of the process.For each of the constructed sub-KCVA models,the corresponding process monitoring statistics and the corresponding thresholds are constructed.Then,a Bayesian inference strategy is used to integrate the fault detection results of individual models,thus improving the process monitoring performance.Finally,the effectiveness and superiority of the proposed method are proved using two commonly used industrial datasets,Continuous Stirred-Tank Reactor(CSTR)and Tennessee Eastman Process(TEP). |
