Electrical Fault Diagnosis From Text Data: A Supervised Sentence Embedding Combined With Imbalanced Classification
Xiao Jing, Zhiang Wu, Lu Zhang, Zhe Li, Dejun Mu
Abstract
Huge amounts of text data describing malfunction, defect, and safety hazard have been recorded in power maintenance sectors. Effectively mining such text data, and thus, classifying electrical fault types from text data bring the potential to considerably reduce the manpower in a variety of industrial applications such as power scheduling and periodic report generation. However, short sentences in verbal expressions, pervasive technical terms of electronics, and imbalanced fault type distribution put an enormous hindrance to fault diagnosis using unstructured data analytic approaches. It has conclusively been shown that deep learning is highly effective to learn representations for unstructured data. In this paper, we design and implement a deep-learning-based electrical fault identification framework, of which the core component is supervised Sentence Embedding combined with an Imbalanced Classification (SEIC) model. SEIC incorporates very little domain knowledge represented by class-related keywords into supervised sentence embedding. Meanwhile, both sentence embedding training and imbalanced multilabel classification training are guided by one unified objective. Experimental results on real-world dataset demonstrate that SEIC significantly improves the accuracy of electrical fault classification over existing deep models. Key factors affecting SEIC are carefully explored by an extensive ablation study.