Considering the diversity and distributed requirements of safety production processes, a fault detection based on knowledge graph is proposed. A knowledge graph based on mutual information and embedded variable similarity is constructed. A global feature extraction based on graph attention networks and a local feature extraction based on bidirectional long short-term memory networks is proposed, along with the design of a sub-block fusion collaborative prediction model and fault detection. The experiments show that the proposed model not only has good practical applicability, but also outperforms existing benchmark methods in terms of detection accuracy and false alarm rate. Moreover, it has more effective parameter sensitivity, providing technical support for fault detection in safety production processes.