In this paper, Bidirectional Mapping Relation Triple Extraction (BMRTE) is proposed to address the challenges of numerous overlapping triples and centralized knowledge in fault maintenance texts, enabling the construction of a fault knowledge graph. The pre-trained language model BERT is employed in BMRTE for the generation of an initial vector representation for each token in the texts. A binary tagging framework is used to identify base entities that may be related in the text on the basis of encoding tokens. To extract entity pairs from the text effectively, the bidirectional mapping framework is designed to map the base entities to their associated entities. Finally, multiple relation classification matrices are used to identify entity pairs and determine the relation triple. Our proposed model outperforms the compared baselines, as evidenced by the experimental results obtained from both a widely used public dataset and our labeled fault maintenance dataset.