15704

14031

كارشناسي ارشد

كنترل

اصفهان : دانشگاه صنعتي اصفهان

1399

دوازده، 95ص. : مصور، جدول، نمودار

ايمان ايزدي

احسان يزديان

محسن مجيري، مريم ذكري

1399/06/03

كتابنامه

مهندسي برق

مهندسي برق و كامپيوتر

1399/06/03

2627212

Process Mining of Alarm Data for TopologyExtraction and Fault Diagnosis Using Deep Learning Amir Neshastegaran a neshastegaran@ec iut ac ir July 18 2020 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language Farsi Supervisor Assoc Prof Iman Izadi iman izadi@cc iut ac ir Abstract Thanks to technological developments in software and hardware fields the behavior of industrial process can be repre sented by a large volume of data These data make it possible for the researchers to study the behaviors of the industrialprocess using data based approaches Investigating plant faults fault diagnosis and fault propagation analysis are amongthe most popular subjects which have always been under special attention This issue is a matter of importance since thepropagation of a fault in plants can cause abnormal or even hazardous situations leading to considerable financial health or environmental costs Hence in this research we concentrate on the fault behavior in plants using alarm data Employingalarm data for fault analysis does not necessarily require the process model Therefore it is less dependent on the knowl edge of process experts Moreover alarm data is lower in volume and higher in time resolution in comparison with processdata In this research we consider two aspects of the fault analysis in the plants fault propagation and fault diagnosis Thisresearch firstly presents a framework for deriving a process topology indicating the fault propagation path using processmining methods When a fault occurs in some part of the plant it often propagates due to process variable interconnectionsand triggers the underlying alarms in each part of the plant Hence using process mining methods on the alarm data a modelof the fault behavior propagation can be extracted in the form of a process topology Then the conformity of the extractedmodel is measured with respect to the alarm data to evaluate its performance In the following using deep learning concepts two neural networks are proposed one for fault diagnosis and another for the next alarm prediction These independentnetworks both receive a sequence of alarm data in their inputs However in the outputs one network will diagnose thefault and another will predict the next alarm Fault diagnosis facilitates removing the fault in addition to referring to theunderlying topology illustrating the fault propagation path The next alarm prediction is of importance especially for therandom faults following no specific topology as it helps the operators to take precautionary actions The whole frameworkis studied and implemented on the well known Tennessee Eastman process and the detailed results are presented Key Words Fault Propagation Fault Diagnosis Alarm Management System Process Mining DeepLearning

ايمان ايزدي

احسان يزديان

محسن مجيري، مريم ذكري