As an effective mechanized harvesting equipment, the combine harvester can greatly improve the harvesting efficiency of crops. However, it is inevitable that some mechanical failures will occur during harvesting operations. Since the driver lacks specialized maintenance experience, he does not know the cause of the failure and how to repair the machine when the failure occurs. This will seriously affect the harvest of crops, and even it may also cause safety accidents. Since knowledge graphs can use graph databases to store unstructured data such as expert knowledge in a standardized manner, knowledge graphs have good application prospects in the field of fault diagnosis question and answer. Based on this, a set of knowledge graphs for combine harvester fault diagnosis is proposed. Firstly, the entities and entity relationship types required in the knowledge graph are clarified based on expert knowledge, the entity extraction model of the Bidirectional Gated Recurrent Unit (BiGRU) and the Transformer encoder is combined with the RoBERTawwmext pretraining model to extract entities from unstructured text. Secondly, the RoBERTawwmext pretraining model is again used to fuse the recurrent neural network (RNN) model to conduct entity review of the extracted entities. Thirdly, after the entity review is completed, the RoBERTawwmext pretraining model is used to extract the entity relationships existing between the head entity and the tail entity, by combining the relationship between the Bidirectional Gated Recurrent Unit (BiGRU) and the attention mechanism. Finally, the extracted entities and entity relationships are formed into triples, and the triples are used to build a knowledge graph, so that the knowledge graph can be used to implement intelligent question and answer.

联合收割机作为一种有效的机械化收割设备，可以极大地提高农作物的收获效率。然而在进行收割作业时不可避免地会发生一些机械故障，由于驾驶员缺乏专门的维修经验，无法确定故障发生的原因以及出现故障后应该如何维修机器，导致严重影响农作物的收获，甚至还可能引发安全事故。由于知识图谱能够利用图数据库将专家知识等非结化数据进行规范化的存储，所以在故障诊断问答领域，知识图谱有着良好的应用前景，基于此提出一套面向联合收割机故障诊断领域知识图谱的构建方法。根据专家知识明确知识图谱中所需要的实体和实体关系类型，利用RoBERTawwmext预训练模型融合双向门控循环单元(BiGRU)和Transformer编码器的实体抽取模型对非结构化文本进行实体抽取；利用RoBERTawwmext预训练模型融合循环神经网络(RNN)模型对抽取的实体进行实体审核；在实体审核完成后使用RoBERTawwmext预训练模型融合双向门控循环单元(BiGRU)和注意力机制的关系抽取模型对头实体和尾实体之间存在的实体关系进行抽取；将抽取到的实体和实体关系组成三元组，利用三元组构建知识图谱，从而可以利用知识图谱实现智能问答。