The Research For Fault Diagnosis System Of Catalytic Cracking

Petrochemical industry is one of the pillars of our national economy, playing a very important role in promoting economic and social development. As the core of refinery equipment, FCC is a high degree of continuous production with complex process, many operating variables and equipments. The probability of fault and risk is very high. Improving the production process of the security, reliability and validity is especially important.In this issue, it was analyzed the characteristics of different fault diagnosis methods. Combining FCC process, equipments, operational characteristics, and utilizing the fusion strengths between artificial neural networks and expert system, TFCCFD system had been developed based on the study of the reaction-regenerator system of FCC in the south-west of Tarim Basin.By deeply analyzing and classifying the hidden danger of the reaction-regeneration system that may occur and the reasons as well as the corresponding measures, the faults may be generalized into two kinds of failures including the process faults and the equipment failures. According to the historical data and operating conditions of the reaction-regeneration system, identifying seven kinds of process faults and four kinds of the equipment failures, determining the application of BP network model and defining corresponding 15 symptoms parameters of the process fault and 8 symptoms parameters of the equipment failure. The network structure of the process and equipment faults was 15-12-7 and 8-8-4 respectively.Using the adaptive-adjustment learning rate BP algorithm, selecting the S-function as the activation function, the learning rate and indicated error were respectively 0.8 and 0.01 in the circumstances to ensure network convergence. The results show that, TFCCFD system can quickly diagnose, analyze the cause of the malfunction, judge the nature of failure, propose the related measures according to the input data including the actual operation data and simulation training data. For the two sub-networks of typical process and equipment failures, the accuracy of diagnosis was 75%, misdiagnosis rate was 25%, basically reached the expected requirements.