Edgealarm: Petrol rafinerisi için sınır bilişim destekli gerçek zamanlı ve akıllı alarm yönetim sistemi

Edge computing is changing the course of data processing in many industries as it enables the processing of data closer to its origin. Although various edge computing architectures exist for different industries, to the best of our knowledge, there is no extensive and practical utilization of edge computing for oil refineries and specifically for alarm management systems. Alarm management system is an integral part of every industry to carry out various operations safely. An ineffective alarm system can increase the operator workload, which can lead to catastrophic events. The factors such as alarm flooding, momentary alarms, and nuisance alarms impede the operator productivity and hence increase his/her workload. In this thesis, we propose an efficient real-time alarm management system based on the edge computing paradigm, namely EdgeAlarm, without altering the existing infrastructure of an oil refinery. The proposed EdgeAlarm system consists of three modules: Storage Reduction (SR) module, Operator Workload Reduction (OWR) module, and an Artificial Intelligence assistant (AI) module. The SR module is responsible for filtering the momentary alarms at the edge and only sends useful alarms to the cloud (Historian server). In the OWR module, we propose two novel techniques to reduce the operator workload by classifying the true and nuisance alarms. The AI module consists of a Recurrent Neural Network (RNN) to predict the future alarms to assist the operator in real-time. We evaluated the proposed EdgeAlarm system on the 13 months alarm dataset of an alarm management system of T¨upra¸s oil refinery plant. Our findings indicate that the SR module achieves a reduction in the storage utilization of the Historian server by approximately 10%. The OWR module reduces the operator workload approximately by 50% by finding the true and nuisance alarms. Finally, the AI module assists the operator to avoid or prepare for any abnormal situation in advance by predicting the majority of the upcoming alarms with the help of RNN with an accuracy of 55% per alarm source at the edge in real-time and some alarm sources are predicted with more than 90% accuracy.