Intelligent Fault Diagnosis with Deep Architecture

The status judgment of converter valve equipment is an important part of ultra high voltage maintenance. However, accurate judgment of possible failures remains challenging. This paper proposes a novel multi-region deep architecture to improve abnormal judgement by considering long-range context information in increasingly finer spatial regions. Specifically, the specified characteristics of video sequence will be completely explored through C3D and recurrent neural network with LSTM cells, extracting high-level spatio-temporal features. In addition, spatial pyramid crop is proposed to characterize each video at different spatial pyramid levels. These sampling sub-videos can encode locally saliency among categories. Experimental results demonstrate the effectiveness of our model, compared with other state-of-the-art approaches.