16334

14570

كارشناسي ارشد

كنترل

اصفهان : دانشگاه صنعتي اصفهان

1399

چهارده، 75ص. مصور، جدول، نمودار

ايمان ايزدي

مرضيه كمالي، مهران صفاياني

1399/12/12

كتابنامه

برق

مهندسي برق و كامپيوتر

1399/12/12

2679533

Modeling the relationships between industrial process alarms using deep learning Negin Javanbakht negin javanbakht@ec iut ac ir February 17 2021 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language Farsi Supervisor Dr Iman IzadiAbstract Alarm management systems are an essential part of industrial units controlled by the operator using distributed controlsystems Alarms are raised to inform the operator of abnormal situations which could be due to the process variablesexceeding the specified threshold or a fault When an alarm is raised the operator must intervene immediately and find thealarm s root cause Otherwise the fault could propagate through the process causing damage to other process equipmentand even shutting down the entire plant On the other hand due to the propagation of faults between the equipment theoperator faces many alarms in a short time which causes the inability of the operator to handle the current situation Alarmsare the operator s guide in finding the cause of the faults But due to the interaction and complexity of the system thereis no one to one relationship between alarms and faults So we need a mechanism that can help the operator find the rootcause of the alarm Also by predicting the next alarm and its occurrence time the operator can manage the critical situation In this study the objective is to model the relationships between alarms to find their root cause and predict the next alarmwith the highest possibility and the time of its occurrence For this purpose deep neural networks have been used and theproposed models have been implemented on the well known Tennessee Eastman process For training the neural network the alarms are first pre processed and organized in sequences Since the neural network input must be numerical vectors andthe alarms are in textual sequences the word embedding model has been used The alarm sequences have been convertedto numerical vectors Two deep neural networks have been proposed to find the alarm s root cause with high accuracy usinga small amount of data The network training process is also fast Then two deep neural networks with three inputs andthree outputs were used to predict the next alarm and its occurrence time The sequence of alarms the hour and minute ofthe alarms are the neural network input and the next alarm with the highest probability of occurrence as well as the hourand minute of its occurrence are the outputs Acceptable accuracy was obtained for this section as well Key Words Alarm management system Root cause analysis Alarm prediction Word embedding Deep neural network

ايمان ايزدي

مرضيه كمالي، مهران صفاياني