Adaptive Correlation Analysis And Prediction Method For Industrial Process Alarm Based On Text Mining

Alarm system aims at guiding operators to notice abnormal process state.Because of unreasonable design of alarm systems,in abnormal conditions,hundreds of process alarms and even alarm flooding may occur,which seriously interfere with operators' judgment of current process state,thus contributing to the occurrence of industrial accidents.Therefore,using advanced alarm management to forecast dynamic process and identify correlated alarms,can improve the performance of alarm system and avoid continuous occurrence of alarm flooding.Considering untimely and correlated alarms in alarm systems,combining problems in existing data-driven methods such as lack of self-adaptability and insufficient data,an adaptive correlation analysis and prediction method for industrial process alarm based on text mining is proposed.According to process data and alarm log information,aiming at four topics including adaptive pre-warning with trend monitoring,adaptive risk propagation path identification,alarm system optimization based on alarm clustering and alarm prediction based on text mining,the main researches are summarized below:(1)An adaptive process pre-warning method based on trend monitoring is proposed according to the poor timeliness of alarm system pre-control,lack of self-adaptability and multivariate pre-warning system in existing trend monitoring and pre-warning methods.The trend characteristics of process variables are extracted adaptively,and the variables of non-steady trend are pre-warned.Then an adaptive weight calculation method for multivariate pre-warning is studied,which provides a priority of pre-warned variables for operators.The accuracies of the proposed method,applied on abnormal situations of atmospheric column and fractionator,for identifying increasing or decreasing trends are 17.9% and 12.1% higher than that of traditional trend extraction method.Compared with threshold-based alarm method,the proposed method has an average time in advance of 2min33 s and 2min35 s,thus improving the timeliness and effectiveness of alarm systems.(2)To prevent the occurrence of correlated alarms and alarm flooding,an adaptive analysis method of risk propagation path is proposed,considering the subjective and uncertain factors in existing causal reasoning methods and lack of adaptive quantitative reasoning mechanism.The propagation paths are adaptively predicted based on the correlation degree and trend changing information between variables of related equipment.The method is demonstrated by a risk process of the fractionator.Compared with traditional methods such as Petri net and fuzzy logic reasoning,the proposed method quantifies process causal reasoning mechanism and avoids subjectivity and uncertainty.(3)Most of the existing alarm correlation analysis methods are based on binary alarm sequences,but sometimes it is difficult to obtain representative binary alarm sequences.To that end,an alarm system optimization method based on alarm clustering is proposed considering the diversity and correlation of process alarm log information.The correlated alarms are grouped adaptively and visualized combining Word2 Vec and aggregated k-means clustering method.An optimization strategy of alarm system based on alarm clustering is then put forward to suppress correlated alarms and eliminate redundant and chattering alarms.The method is applied to a steam generation and distribution system.Compared with the threshold-based alarm system before optimization,the number of alarms is reduced by 61.2%.The proposed method is conducive to further identification of alarm root cause,so as to prevent the occurrence of correlated alarms and alarm flooding.(4)Considering that the existing alarm prediction methods are mostly based on process data but some variables usually do not generate an alarm frequently or have no process data,combining the correlation and continuation characteristics between alarm variables,a process alarm prediction model is established based on word embedding and deep learning.Correlated alarms that may occur are predicted through quantitatively processing alarm variables and modeling LSTM neural network.Through case analysis of a central heating and cooling plant,the time in advance of the predicted alarm is mainly concentrated in 0-10 min,and the average time in advance is 6min49 s.Compared with the N-gram model,the average prediction accuracy of proposed method is improved by 7.4%.