Data-driven Based Condition Monitoring And Fault Diagnosis Of Low Temperature Waste Heat Generation System

Low temperature waste heat generation system transforms waste heat from the boiler exhaust gas into mechanical energy, finally high grade electrical energy via organic Rankine cycle (ORC). It has great significance in the aspect of saving energy, water and reducing harmful gas emission. With the long-time operation, some faults may occur in some components of the system. This leads to the efficiency decline, performance deterioration, huge economic loss, and even greatly threatens personnel safety. Therefore, developing the condition monitoring and fault diagnosis system which guarantees the safe, steady and efficient operation of ORC process, is paid more and more attention.The contents of this paper mainly include three parts. In the first part, the faults, possible causes and consequences of each component are described and analyzed in detail. Then the fault diagnosis simulation platform of ORC system is built here by monitoring process parameters and some typical faults are selected to realize the fault diagnosis from the system frame. In the second part, considering that the random disturbances in the mass rate and temperature of exhaust gas in the inlet of evaporator do not necessarily obey Gaussian distribution, kernel entropy component analysis (KECA) is used to detect the faults of nonlinear ORC system. The introduction of Renyi quadratic entropy extends the data from Gaussian distribution assumption to any distribution. Kernel density estimation method takes the place of traditional methods to determine the control limits of monitoring statistics. Then the proposed method is compared with kernel independent component analysis- kernel pricinpal component analysis (KICA-KPCA) method, and the simulation results indicate that KECA can detect the faults better than KICA-KPCA. In the third part, support vector machine (SVM) is adopted to identify the faults in ORC system. On account of the disadvantages of SVM method, multiclass relevance vector machine (M-RVM) is proposed to classify the different types of faults at one time. The comparison results between SVM and M-RVM show the higher classification accuracy of M-RVM method.