Correlation Analysis And Prediction Of Alarm Data Of Refining Device Based On Deep Learning

With the production equipment and equipment of domestic refining and chemical enterprises become more complex and intelligent,enterprises pay more attention to production process safety and production efficiency,then there are more and more alarm points are set in the industrial process,resulting in a large number of alarm information.Some of the alarms are invalid alarms,and some are caused by the unreasonable design of the alarm system and the defect of the alarm location.Such alarms further cause a sharp increase in the number of alarms.For alarm management,the influx of a large amount of alarm information in a short period of time will interfere with the operator’s judgment,resulting in the failure of important alarms to be dealt with in time,which is likely to cause accidents.Therefore,it is necessary to carry out correlation analysis on the alarm information,optimize the alarm system,and make full use of the alarm log for alarm prediction to reduce the occurrence of accidents.This thesis mainly carries out the following research:(1)In view of the large amount of alarm log data and many attributes,the method of manually analyzing alarm correlation is time-consuming and labor-intensive,and the alarm information is difficult to be represented by binary sequences,this paper proposes an alarm data association analysis method based on word embedding and negative sampling optimization.The alarm variables in the alarm log are mapped to vectors,and the vectors are used for cluster analysis after dimensionality reduction.In the case analysis,the alarm correlation analysis is carried out according to the cluster analysis,and the clustering result is basically consistent with the actual correlation of the alarm.Compared with other methods,the accuracy and similarity of the alarm vector obtained by word embedding in the word analogy task are 83.33% and 98.82%,which are 5.55%and 7.57% higher than other vector representation methods,respectively.This shows that the quality of the alarm vector obtained by the method in this thesis is better,the semantic and temporal relationship between the alarms can be preserved to the greatest extent,and the cluster analysis results are reliable.The correlation between the alarms can be used to make reasonable suggestions for the optimization of the alarm management system.(2)Aiming at the problem that the alarm log cannot be directly input into the prediction and early warning model for identification,and there is no reference standard for parameter selection and model optimization in the process of abnormal condition prediction and modeling,this paper solves the problem of reading alarm log information through the optimized CBOW model,a bidirectional LSTM model is constructed for alarm prediction.In the case study,parameters are selected according to the predicted performance of the model under different conditions to obtain the optimal model.In the case study,parameters are selected according to the predicted performance of the model under different conditions to obtain the optimal model.In the comparative analysis,compared with the fully connected RNN and one-way LSTM methods,the method proposed in this thesis has the best effect,the highest prediction accuracy,and an increase of 8.51%;the average early warning time is the longest,which can provide operators with more adequate processing time.(3)In view of the alarm optimization strategy obtained from the alarm correlation analysis in the above research,combined with the model prediction effect,this thesis uses the Qtpy graphic design tool to design the interface,uses the Python programming language as the development tool,and uses the database technology to develop an alarm log based alarm log.This module mainly realizes three functions: alarm monitoring,related alarm query,and abnormal working condition prediction.The optimized alarm monitoring interface is set to display by partition,which can find key alarms more clearly.In the relevant alarm query interface,the petrochemical enterprise alarm management platform is connected to the enterprise alarm management system,and the alarm log is preliminarily processed and stored in the database.The word embedding model is used to process the vectorization of the alarm and calculate the similarity to realize the query of the relationship between the alarms.In the abnormal working condition prediction interface,the bidirectional LSTM model is used to predict possible alarms,and output the abnormal working condition corresponding to the alarm.Through this function,operators can predict possible abnormal situations more quickly,and take action to eliminate hidden dangers in a timely manner.