Computação evolucionária aplicada ao diagnóstico de falhas incipientes em transformadores de potência utilizando dados de cromatografia

Abstract

In an electrical substation of hydro-electric power station, power transformers are the equipments responsible for raising the voltage values, with the goal of transmitting electrical energy at high voltages with reduction of losses by Joule effect. On the other hand, at consumers centers, transformers are responsible for reducing the voltage level to allow distribution in urban centers. The importance of such equipments is clear. A major fault generates a high cost of corrective maintenance, drop in performance indicators, in addition to fines by the regulatory agency, which in Brazil is called the National Agency of Electric Energy (Agência Nacional de Energia Elétrica - ANEEL). Therefore it is highly desirable evaluate the lifetime of a power transformer, or at least identify the incipient faults in the transformer before a catastrophic failure occurs. The life cycle of a power transformer is directly related to the thermal factor, i.e., when in operation it is subjected to high temperatures that impact, in a negative way, their insulation system. The insulation system of the transformer can be divided into solid (made of cellulose-based) and liquid (by immersing the transformer in a tank filled with insulating oil). The thermal and electrical stress, suffered by the power transformer in operation, causes the degradation of both insulating paper and oil. With the aging of the mineral oil, gases are produced in a more pronounced way when a fault occurs. To prevent damage to transformers, the companies have adopted preventive maintenance procedures, however the costs involved in this kind of procedure can be minimized through predictive maintenance. The predictive maintenance can indicate the best time to perform preventive maintenance, avoiding unnecessary expenses. This dissertation aims at developing, analyzing and implementing a novel methodology that uses an Evolutionary Algorithm based on Genetic Programming to detect rules from power transformers database to predict incipient faults. The fault diagnosis abilities based on dissolved gas analysis (DGA), is enhanced by developing a framework called MINERA to discovery rules from database. MINERA was developed based on a modified Genetic Programming Algorithm that uses concepts from Information Theory, such as Entropy and Information Gain. This framework also provides support to take full advantage of the architecture of multi-core processors.