Multiscale Information Fusion-based Extreme Learning Machine for Power Transformer Fault Diagnosis

Power transformer fault diagnosis is one of the most important techniques in smart grid. Recently, extreme learning machine (ELM) has been proved effective in analyzing dissolved gas to diagnose transformer faults. However, due to the complexity and variability of dissolved gas data, it is challenging to choose an optimal number of hidden layer node in ELM. For previous ELM methods, the number of hidden layer node is fixed. It means that these methods may fail to exploit dissolved gas information to diagnose effectively. To tackle this problem, this paper introduces a novel multiscale information fusion-based extreme learning machine (MIF-ELM) technique for power transformer fault diagnosis. The proposed method contains two main steps. First, multiple ELM models with different node numbers are employed to produce initial diagnosis results. Different diagnosis results can reflect various statistics information of dissolved gas, which can provide complementary information for power transformer fault diagnosis. Then, an effective information fusion method is developed to fuse diagnosis results with different scales, so as to generate a fused diagnosis result. Experimental results demonstrate that our proposed approach can outperform some classical and outstanding methods including support vector machine (SVM) and back propagation (BP) neural network.