Fault Diagnosis of Power Systems Using Visualized Similarity Images and Improved Convolution Neural Networks

Fault diagnosis is important to stable operation of power systems, and the machine-learning-based fault-diagnosis models were widely studied because of their strong generalization ability. However, the model structures are generally designed according to the topology of power systems. Once there are changes in topology, the system fault characteristics might change, and the models structure and parameters are often required to be adjusted for applying to the new systems. To avoid frequent adjustment work of fault-diagnosis models when system topology changes, we propose a novel fault-diagnosis model for power systems. First, a new data preprocessing using gradient calculation and similarity assessment is presented, and the gradient similarities among the multichannel electrical signals are converted to the visualized similarity images, which are fed to the neural network for further processing. Second, the spatial pyramid pooling (SPP) and hashing classifier (HC) are used in the convolution neural network. With the aid of the SPP and HC techniques, the structure of the fault-diagnosis model can maintain unchanged even though there are topological changes in the power systems. To validate the effectiveness of the proposed model, several state-of-the-art fault-diagnosis models are used for comparison. The results show that the proposed model with unchanged structure is well performed in accuracy and noise immunity, and friendly to the parameters setting.