Fault Monitoring And Recognition Of High-sulfur Gas Sweetening Process

With the economic development and energy structure adjustment,China's high-sulfur natural gas(HSG)resources have been continuously developed,and the processing volume has also increased year by year.But at the same time,because the highly toxic components hydrogen sulfide in HSG is as high as dozens of that of ordinary gas fields,the safety threat to the collection facilities and personnel environment has multiplied.Therefore,real-time monitoring and fault identification of high-sulfur natural gas sweetening process is of great significance for ensuring reliable of the system and achieving safe production.With the popularization of distributed control system(DCS),a large amount of historical data of the sweetening process is collected and stored,which also provides a solid foundation for datadriven intelligent fault monitoring and recognition.So,this paper proposes a process monitoring and fault identification method based on high-order cumulant analysis(HCA)and long-short-time memory neural network(LSTM)for the actual production data of HSG sweetening process.(1)Aiming at the problem that HCA method can not effectively monitor and feedback the migration of working condition information in time series,this paper proposes a HSG sweetening process fault monitoring model based on dynamic high-order cumulant analysis(DHCA).The dynamic preprocessing of the data is performed by adding a time delay to the input matrix to increase the data in the time dimension.Then the extended dynamic matrix is input to the HCA model for monitoring.(2)Against the problems of HCA method in independent component space index construction and multi-indicator monitoring strategy,this paper proposes a fault monitoring method based on contribution-weighted high-order cumulant analysis(CW-HCA)joint index.The third-order cumulant of the independent component of the HCA method is improved,and the corresponding weight is given according to the contribution of the independent component,so as to mitigate the interference of the component with less influence.Then,the weighted independent component space index is combined with the residual space index to obtain a new joint indicator to achieve monitoring.(3)After the fault is detected,the fault type needs to be quickly and effectively identified.In this paper,we use LSTM's timing and powerful pattern recognition and training ability to train the original process data,adaptively learn the dynamic information,and obtain the diagnosis results.At the same time,because the sweetening process data is collected in an online time series manner,some data with low frequency and random failure type data are difficult to obtain,which further leads to data imbalance between the fault classes.Therefore,this paper proposes an LSTM fault identification model for data imbalance learning,which is used to supplement fault imbalance data by Synthetic Minority Oversampling Technique(SMOTE).