With the increasing complexity of production scenarios, a large amount of production information is stored in the enterprises of industrial domain. How to in-depth tap the value of complex document information and establish information links are important questions. In this paper, a framework of knowledge graph construction in industrial domain based on document-level relation extraction is proposed. To improve the accuracy of named entity recognition, domain knowledge is added to the word embedding matrix initialization of BiLSTM-CRF. For the task of relation extraction, this paper proposes the Knowledge-Aided Graph Inference network (KAGI), a relation extraction method for long paragraphs in industrial domain, which captures the complex interactions among entities by constructing document graph and innovatively adds knowledge representation to node construction and path inference through TransR. At the application level, BiLSTM-CRF and KAGI are used to construct knowledge graph from knowledge representation model and Chinese fault reports for steel production line, SPOnto and SPFRDoc respectively. The quality of the extracted knowledge graph meets the requirements of actual production environment applications. The result shows that KAGI can deeply mine the production reports and extract rich knowledge and patterns from them, which providing a solution for production management.