Data Mining Visual Inspection Information in Electrical Machine Maintenance Reports

Abstract

Visual inspections from industrial electrical machine maintenance contain information regarding fault causality and early indications. Predictive maintenance paradigms include increasing applications of machine learning. Natural text proves costly for processing and feature extraction. This work applies state-of-the-art natural language text processors in a Synchronous Generator maintenance report corpus to extract features and observe limitations. Results indicate a high level of correlation between text features and faults and are encouraging for extended application in the scope of electric machine predictive maintenance and repair. Reverse engineering of the procedure and specialization on the task can provide literature with an unexplored avenue of information retrieval. These innovative models can be trained and used for decision-making during scheduled outages. Information from well-trained classifiers can be extrapolated to advance fault causality understanding in broader electric machine research.