Bushings are key equipment in transformers. At present

operators have accumulated a large amount of cross-modal data for bushing operation

including text and images. How to effectively use the above data to predict bushing failures and identify causes is critical to improving bushing operation and maintenance efficiency. This paper proposes a method for constructing a knowledge graph of transformer bushing faults based on cross-modal data. First

a top-down method is used for knowledge modeling to construct the ontology layer. Second

the ALBERT (A Lite Bidirectional Encoder Representations from Transformers)-BiLSTM (Bidirectional Long Short Term Memory)-CRF (Conditional Random Field) model and the ALBERT-FC(Fully Connected) model are used to extract entities and relations from text

and their F1 scores reach 96.60% and 98.99% respectively. Then

the ResNet (Residual Network)-50 model is used to extract features from the bushing fault image

and combined with the BADGE (Batch Active learning by Diverse Gradient Embeddings)- based active learning sampling strategy to achieve accurate fault image classification with a small number of training samples. Its F1 score reaches 92.11%. Finally

the text is converted into word vectors

and the text knowledge and image knowledge are associated and fused by calculating the semantic similarity. A transformer bushing fault knowledge graph containing cross-modal data is constructed and applied in field cases to infer the causes and evolution process of faults.