Industrial Alarm Analysis Model Based On Machine Learning And Pattern Recognition Algorithm

The "Alarm Flood" phenomenon is widespread in the industrial production field.In the systematic and automated industrial production process,a single abnormal event may trigger a series of chain alarm notifications.When the number of alarms generated is far greater than the actual number handled by system terminal operators,this situation can be called "Alarm Flood" phenomenon.In an "Alarm Flood"situation,operators lack sufficient time to properly process each alarm notification.When critical alarm notifications are mishandled,it may lay a significant hidden danger to the operation of the system.By designing an analysis model for alarm sequence data,it is possible to automatically recognize the "Alarm Flood"phenomenon.However,the cause of the "Alarm Flood" problem in the industrial field is complex,and relying only on knowledge-driven methods or data-driven methods for modeling cannot effectively solve the problem.Sequential pattern mining can effectively analyze frequent subsequences in the alarm notification sequence.However,the alarm sequence analysis model is based on the linear time information of the alarm notification,ignoring the local correlation of the alarm notification,and the analysis results of the model are mainly qualitative analysis,which lacks the intuitive expression of the internal connection between alarms.This paper introduces the word vector theory and the alarm spanning tree structure,proposes an improved alarm correlation analysis method,and designs a quantitative calculation formula for the internal correlation of combined alarms,to achieve efficient representation of the causal relationship between different alarms.Based on the proposed alarm correlation analysis method,this paper formulates the "Alarm Flood" identification strategy and the related industrial alarm analysis model.The workflow of this model includes data preparation stage and pattern recognition stage.In the data preparation phase,expert knowledge is introduced to train the alarm notification classifier,and a preliminary classification of all alarm notifications is achieved based on the small-scale label sample set and the "Alarm Flood" recognition strategy.In the pattern recognition stage,a single source frequent alarm analysis model and a multi-source frequent alarm analysis model are designed.The former uses the sequence analysis method to analyze and locate the "Alarm Flood" events caused by single factors such as unreasonable system settings and human interference;the latter uses the improved alarm correlation analysis method to analyze and locate the relevant events of redundant alarms or chain alarms.The model proposed in this paper has been applied to the analysis of the tower's dynamic environment monitoring data.The results show that the method can effectively locate the root cause alarm and efficiently characterize the degree of correlation between frequent alarm events.