Research On Transformer Fault Diagnosis Technology Based On ANN

Realizing the mapping from fault symptom space to fault space is the key of transformerfault diagnosis. It is a complex nonlinear problem and it can realize the identification anddiagnosis of the fault. ANN (Artificial Neural Network) is an important tool in the field offault diagnosis. But because of complex working condition and various fault types, someproblems such as network vast, long learning time and falling into local extreme value oftenoccur in diagnosing. ANN combined with other artificial intelligence theories is becoming thetrend of fault diagnosis. In this paper, ANN is combined with ACA (Ant Colony Algorithm)and EX (Expert System) to improve the properties of network, and then used for transformerfault diagnosis. The details are as follows:(1) Beginning with the reason of transformer faults, the sources of dissolved gas intransformer oil are analyzed, and then the fault mechanism and diagnostic reasons are alsodeeply researched. Finally the theory that different faults produce different kind and amountof gases in the oil is confirmed as the diagnostic way.(2) A new way to improve ACA is proposed. In order to reduce the amount ofcalculation, the local updating procedure of pheromone is deleted. One kind of convergencefactor is designed to achieve that during the global updating procedure of pheromone, in eachiteration the best route would be given additional increment of pheromone. The simulationresults show that the improved ACA has higher searching ability and faster convergent ability.Then the improved ACA is used to optimize ANN. In order to lower the complexity, heuristicfactors are saved. The selection range of the main parameters of amount, pheromonevolatility and quantity are discussed. The optimized ANN has some advantages in avoidingthe problems such as local extreme values and long learning time, also has a high accuracy.(3) A hybrid model of integrating these two technologies is proposed. It uses a screeningstructure to transmit two different types of knowledge between ANN and EX. If diagnosisprocedure calls new rule knowledge, the screening structure will send the data to sample datasets to train ANN, in this way, ANN gets the new rule knowledge and makes the utilizationefficiency of knowledge in a higher level.(4) Based on this model, a transformer fault diagnosis system is made in matlab and C#programming language. And this system has been successfully applied as a support system inthe diagnosis of transformer fault.