The Research On Modeling And Diagnosis Methods Based On Mass Data Flow For Thermal Process

The faults in thermal process not only have a negative impact on generating efficiency,but also would force serious equipment accident,disrupt plant operation,and cause environmental pollution.An effective fault diagnosis approach can detect and eliminate abnormal condition as soon as possible,which can effectively improve the safety and reliability of thermal process.In modern industrial power plants,real-time data flow is an important basis of process control and monitoring,and the corresponding mass history data make it possible to diagnose faults for thermal processes.Therefore,it is of a practical significance and a great realistic value to launch the research on data-driven modeling and process diagnosis methods.This thesis focuses on the research work of data-driven modeling and process diagnosis methods for thermal processes,and the main research content is as follows:(1)A nonlinear modeling method based on the modified input training neural network(MITNN)is proposed to improve the training speed with an adaptive adjustment mechanism of momentum constant and learning rate by using Nesterov’s accelerated gradient algorithm(NAG).(2)A new fusion input neural network(FITNN)by integrating the process prior knowledge and history data is proposed to overcome the “over-fitting” problems in the ITNN model training process.In FITNN,firstly,the process prior knowledge is translated as equality,inequality,and monotonic,and concavity constraints.Then,the constraints are introduced into the basic ITNN training objective function according to the restriction mechanism based on the Lagrange penalty function.Finally,the FITNN is trained by NAG algorithm.In the training process,the relaxation variable method is used to realize the continuity of objective function,and thus reduce the model training difficulty.(3)A reconstruction-based neural network method is proposed for fault diagnosis purposes to overcome the “smearing effects” problems.In this method,hypothetical fault directions are introduced into the diagnosis process of AANN and ITNN models,and then the corresponding reconstruction indexes are calculated by tuning the model inputs and the magnitudes of hypothetical fault directions.The optimal reconstruction index is selected to further pinpoint the source of the detected faults.Moreover,an efficient fault diagnosis strategy based on the reconstruction-based input training neural network(RBITNN)is implemented for nonlinear systems,which can reduce misdiagnosis rate for both single and multiple variable(s)faults without prior knowledge efficiently.(4)An improved reconstruction-based input training neural network(IRBITNN)is developed to enhance the diagnostic efficiency of RBITNN based on CMA-ES and BAB approaches.In this IRBITNN,the CMA-ES algorithm is used to reduce the reconstruction time of each hypothetical faulty variable combination,while the BAB algorithm is used to reduce the number of hypothetical faulty variable combinations.This IRBITNN is successfully used in a natural gas combined cycle power plant,and the results shows its effectiveness,satisfying the requirement of on-line fault diagnosis.