DISSOLVED GAS ANALYSIS (DGA) DATA INTERPRETATON METHOD FOR TRANSFORMER FAULT IDENTIFICATION USING COGNITIVE ARTIFICIAL-INTELLIGENCE

• Main Author: Octavianus, Karel
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39110
• Institution: Institut Teknologi Bandung

Fault in transformer can be detected using Dissolved Gas Analysis (DGA). Doernenburg Ratio Method (DRM) is one of the most common methods used to interpret DGA data. DRM has several limitations, it has low accuracy, furthermore it can only identify single fault. These limitations can be overcome using Cognitive Artificial-Intelligence (CAI). CAI is a new perspective in Artificial Intelligence that works based on Knowledge Growing Principle (KGS). Information from multi sources are fused using ASSA2010 (Arwin Sumari- Suwandi Ahmad) information fusion to obtain new information with Degree of Certainty (DoC). These information are used to make decision. Data is collected from IEC TC 10 labeled dataset. IEC TC 10 that comprised of labeled data that is put into groups based on the types of fault, Partial Discharge (PD), Low-Energy Discharge (LE), High-Energy Discharge (HE), Thermal-Low (TL), and Thermal-High (TH). These faults are verified by experts using visual inspection on the faulty equipments. The proposed method, CAI, works by mapping indications and Possible Faults (PF). At first, each PF is identified given a single indication, and proceeded with identification of PF considering all indications. identifies faults based on each indication, followed by the combination of all indications. The more similarity between the indications and the respected Possible Fault (PF), the more likely the respected fault had occurred. The PFs are arranged in the form of Degree of Certainty (DoC), which is used to determine faults. The proposed method is verified using the IEC TC 10 labeled dataset and compared with conventional DRM, FIS, and ANN. The proposed method is also tested using Single Gas Ratio (SGR) as supplement indication. Multi-fault identification is verified using the combination of different samples. Result shows that the proposed method implementation on conventional DRM increases its accuracy to 98.3%, while FIS-DRM increases the accuracy to 94.02%, and ANN-DRM increases the accuracy to 84.71%, furthermore, CAIDRM is capable of identifying multi-fault.