The Research On Fault Diagnosis Of Reaction-regeneration System In Catalytic Cracking Unit

In recent years,the task of preventing and resolving major safety risks of hazardous chemicals is still difficult,and standardized diagnosis has become an important link to improve the safety of petrochemical equipment.The traditional fault diagnosis method is poor in performance on unbalanced long time series data and has some limitations in real-time fault detection and fault tracing.Taking the reaction-regeneration system of catalytic cracking unit in a typical petrochemical process as the research object,the problems of feature extraction,fault detection and fault tracing in the actual production process were studied,and a complete set of fault diagnosis process was formed.Specific research work is described as follows:1.To solve the problem of high dimensionality and strong correlation of petrochemical process data,a variational mode decomposition method based on minimum center frequence-residual energy is proposed to eliminate the influence of variational mode decomposition on parameter dependence.Feature extraction of long time series data of reaction-regeneration system is carried out to obtain a complete representation of fault information.It is applied to numerical simulation signal and off-line signal of reaction-regeneration system,and its effectiveness is proved by comparing with other decomposition methods.2.To solve the problem of excessively long data series and unbalanced distribution in petrochemical process,a long short-term memory network integrating key-value pair attention mechanism was proposed.The basic features of the sequence are extracted from the long short-term memory network,and the interactive features are extracted from the attention mechanism.The integration of the two can better distribute the importance of the unbalanced data features,and complete the real-time fault detection task with high accuracy and high stability.It is applied to TE public data set and reaction-regeneration system real-time data set,and the effectiveness of the method is proved by comparing with other excellent algorithms in this field.3.To solve the problem that fault tracing is difficult in actual production,an event network analysis method based on bayesian inference is proposed.Based on heuristic search,the bayesian network nodes were selected by bat optimization random forest method,the network structure was trained by the shortest description length scoring function,the network parameters were estimated by the maximum posterior probability,and then the fault propagation path and fault source were determined by bayesian inference.At the same time,combined with mechanism process,a complete fault tree model is built based on the fault event,and the causal fault tree model is obtained by combining the data attribute characteristics.Applied to reaction-regeneration system,on the basis of causal fault tree fault source analysis,more accurate fault propagation path analysis and fault source search,to provide targeted guidance for field operations.