Transformer Fault Diagnosis Based On Improved ART2 Network

Dissolved Gases Analysis (DGA) is an important method to diagnose incipient fault of power transformer. It has the abilities to detect the potential faults promptly in the transformer, prevents unexpected accidents and reduces unnecessary interruption maintenance, avoids huge losses that caused by excessive maintenance. So it extends equipment life and optimizing configuration of equipment maintenance. However, DGA technology exist some defects, for example, it has many steps in analysis process, operations are complex, can not be continuous monitoring, all of these make against to discover and analysis the latent fault. On-line monitoring technology as the supplement to DGA technology and makes a further development of DGA technology, but the gas component that can be detected are very limited, so it is not suitable for wide application. Currently synthesizes artificial neural network, fuzzy mathematics and rough set theory to search better methods in fault diagnosis of transformer has become a hot research filed.The interior faults of transformer are of diversity due to the complicated working environment. This paper analyzes diagnosis methods of power transformer fault deeply and collects a large number of DGA examples to build a new model of transformer fault diagnosis which is based on fuzzy C means (FCM) and subtractive clustering. The FCM algorithm defects that need a long time to initialize cluster centers and setting cluster types in advance are not exist in the new model. The simulation results shows that this model has a high correct rate.Fuzzy clustering is an unsupervised clustering method, although it can achieve real-time diagnosis of transformer faults, it need to hard impose a certain subordination for datas, so it is difficult to account for clustering results. The adaptive resonance network is introduced to solve these problems.but as one part of the adaptive resonance network,the traditional ART2 network's outputs are affected by the present order of inputs,so an improved ART2 network is proposesd, this improved algorithm adjust it's parameters a and b by active degree l to reache network's stability-plasticity balanced. And then establishes a transformer fault diagnosis model, simulation results show that this improved model has a very high accuracy and stability in the transformer fault diagnosis.