Deep learning for fault diagnosis based on multi-sourced heterogeneous data

As a representative of deep learning, Boltzmann machine is being widely studied to solve some complex problems. The model of deep Boltzmann machine has been successfully applied to unsupervised learning for single modality (e.g., text, images, video or audio). In this work, we focus on the model for multiple input modalities and apply it to fault diagnosis. At present, most work on fault diagnosis uses structured data and accuracy needs improving. For images, most work stays on the stage of manual analyses and has low level of intelligence. Moreover, the sources of data are multiple and its architecture should be scalable since data comes from many platforms and the type of platforms will keep increasing. In this case, existing approaches cannot solve the above problems. Hence comprehensive analysis of images and structured data using deep learning is proposed. Its goal is to extract a unified representation for both images and structured data with massive attributes. In the phase of feature learning, deep restricted Boltzmann machine auto-encoder is used. Through unsupervised learning, more abstract features are extracted and some useless information is removed, which can achieve layered and precise representation. In the phase of supervised training, linear classifier is employed. Extensive experiments on power transformers and circuit breakers show that the proposed model with two modalities based on deep learning theory has higher accuracy than the previous approaches with supervised learning or just one modality.