Search

U istraživanju je razvijen klasifikacijski model koji za klasifikaciju stanja transformatora na osnovi analize otopljenih plinova koristi tri razvijene metode zaključivanja i interpretacijske metode, koje su u primjeni već niz godina. Kako su interpretacijske metode razvijane koristeći različite pristupe i ugađane na različitim podacima, one često pri interpretaciji DGA rezultata daju različite pa i kontradiktorne interpretacije. Zato je njihova primjena otežana i zahtijeva veliko znanje i iskustvo dijagnostičara pri primjeni interpretacije i određivanju dijagnoze. Novi matematički modeli metoda zaključivanja koji su razvijeni tijekom ovog istraživanja omogućuju donošenje zaključne dijagnoze na temelju odluka interpretacijskih metoda, s dovoljno visokom točnošću. Da bi model za klasifikaciju stanja transformatora bio funkcionalan implementirano je osam interpretacijskih metoda: dvije IEC metode, dvije Roger-ove metode, metoda Duval-ovog trokuta, Doernenburg-ova metoda, metoda ključnog plina i metoda logaritamskog nomograma. Pri tom je šest metoda interpretirano u izvornom obliku, metoda ključnog plina je implementirana koristeći tehniku neizrazite logike, dok je metoda logaritamskog nomograma implementirana koristeći model zaključivanja, koji se primjenjuje i na razini modela za klasifikaciju stanja. Klasifikacijski model određuje faktore podrške dijagnoza. Konačna dijagnoza je dijagnoza s najvećim faktorom podrške. Model je općenit u smislu da može provoditi klasifikaciju s promjenjivom razlučivošću, odnosno može s većom ili manjom razlučivošću dijagnosticirati stanje transformatora. Model za zaključivanje je općenit i mogao bi se primijeniti za donošenje bilo kakvih odluka na osnovi glasova više glasača, ako glasači imaju jasno definiranu proceduru glasovanja (donošenja odluka) i ako postoje podaci na kojima se model može učiti i testirati. Model ima i svoje nedostatke, a jedan od njih je njegova kompleksnost, što ga čini nepogodnim za „ručnu“ primjenu, već se praktički može primjenjivati isključivo uz pomoć računala. In this research a classification model for identification of transformer faults has been developed. The model provides diagnosis on basis of concentrations of dissolved gases in transformer oil, obtained by dissolved gas analysis (DGA) method. Traditionally, DGA data are interpreted using interpretation methods. There are many such methods. They were developed using different approaches, and they were tuned on different sets of examples. So, they often provide different, even contrary diagnoses for same DGA data. Each method can provide correct as well as incorrect diagnosis, so it is recommended to make assessment using more than one interpretation method. Therefore, using these methods for diagnosis assessment requires deep knowledge and experience of a diagnostician. Eight interpretation methods are implemented in the model: two IEC methods, two Rogers ratio methods, Duval triangle method, Doernenburg method, key gas method, and logarithmic nomograph method. Six methods were implemented as they are defined, while key gas method was implemented using fuzzy logic, and logarithmic nomograph is implemented using the same inference model used on the level of classification model. Each interpretation method in the model provides interpretation of DGA data. Inference methods are also implemented in the model, and they provide a final diagnosis on basis of said interpretations. Therefore, the interpretation methods are considered as voters, and their interpretations are considered as candidates for the final diagnosis. Three inference methods have been developed: method which validates all candidates VAC, method which validates supported candidates VSC, and method with voting of exquisite voters VEV. VAC method is designed so that any of possible candidates can win in voting process, regardless which candidates are supported in the voting process. This is the main drawback of the method, and therefore its classification properties are not high enough for practical application. VSC method is designed so that only candidates supported in the voting process can be selected for the final diagnosis. In this way the drawback of VAC method is avoided. VEV method is designed so that only voters with the best classification properties can participate in the voting process. Iinference methods have to be trained on set of DGA data with belonging diagnoses. In this research set of 100 DGA data was used for training and testing of the model. At training inference methods learn about classification properties of interpretation methods for each diagnosis, creating a specific classification patterns, which in fact are matrices. After training, inference methods select final diagnoses on basis of learned patterns and certain inference algorithm. The developed inference methods enable assessment of diagnosis with satisfactory accuracy. Classification model provides diagnosis with belonging supporting factor. Diagnosis with highest supporting factor is a final diagnosis. The model is general by means of providing diagnosis with variable resolution. Three levels of diagnosis resolution are supported: low, medium and high diagnosis resolution. At low resolution four different diagnoses can be classified (no fault NF, partial discharges PD, thermal fault T, and electrical discharges D). At medium resolution five different diagnoses can be classified (NF, PD, T, D, and mixed thermal and electrical fault DT). At high diagnosis resolution eight different diagnoses can be classified (NF, PD, low temperature thermal fault T1, medium temperature thermal fault T2, high temperature thermal fault T3, DT, electrical discharges of low energy D1, and electrical discharges of low energy D2). Classification properties of inference methods were tested. VAC method has inferior classification properties compared to classification properties of interpretation methods, while VSC and VEV methods have superior classification properties compared to classification properties of interpretation methods. Classification properties of the model were also tested on historical DGA data. This is important because model should provide consistent diagnoses even when input data change. This test confirmed satisfactory behaviour of the model. The inference model is general, so it may be used for any inference process on bases of votes of several voters, if the voters vote according to clearly defined procedure (decision making) and if there are data for training and testing of the model.