Fault diagnosis method of critical industrial equipment based on knowledge graphs and multi-task learning
嘉楠 卞, 泽慧 冒, 斌 姜, 亚杰 马, 文静 刘 · Scientia Sinica Informationis · 2022
To resolve the problem of sparse fault data in critical industrial equipment and meet the demands of fault diagnosis, we propose a critical industrial equipment fault diagnosis model, termed multi-task learning for knowledge graph-enhanced fault diagnosis (MKFD), which is based on a knowledge graph and multi-task learning. The model realizes the inference of fault root causes. First, a multi-task learning framework is designed, and an improved cross-stitch network is constructed to realize the information sharing among subtasks in the framework. Then, the interaction matrix of fault phenomena and fault root causes is constructed using operation and maintenance data, and the knowledge graph embedding model is built using a multi-layer perceptron. The embedding of the interaction matrix and knowledge graph is regarded as two subtasks in the multi-task learning framework. Through the alternative learning of subtasks, the parameters of MKFD are optimized to infer the fault root cause, thus assisting the operation and maintenance personnel in fault diagnosis. Moreover, this scheme was verified by constructing two critical industrial equipment fault knowledge graphs based on the operation and maintenance data of a domestic industrial enterprise, and the results show that the proposed method has good performance.