Expert Recommendation Method for Fault Maintenance Based On Industrial Manufacturing Knowledge

• Published in: 2023 IEEE International Conference on Data Mining Workshops (ICDMW) pp. 137 - 146
• Main Authors: Fu, Jiacheng, Tian, Jin, Xu, Jiacheng, Fang, Zhijun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.12.2023
• Subjects: Data mining; Expert recommendation; Heterogeneous propagation; Industrial knowledge graph; Knowledge graphs; Maintenance engineering; Manufacturing; Manufacturing processes; Productivity; Recommendation system; Task analysis
• ISSN: 2375-9259
• DOI: 10.1109/ICDMW60847.2023.00025

In the industrial manufacturing process, equipment failure problems occur frequently and will have a negative impact on productivity if not resolved on time. Therefore, finding experts who can quickly deal with failure problems is a crucial task. To address this challenge, this research combines the task of recommending maintenance experts for industrial fault problems with a recommendation algorithm based on knowledge graph (KG), intending to meet the need for maintenance expert recommendations in the industry. Existing KG-based recommendation algorithms tend to ignore the association between the current hop triplet set, the initial seed and the previous hop triplet set in the knowledge propagation process. In addition, when constructing representations of experts and fault problems, existing methods also do not sufficiently distinguish the preference difference features that exist between them, resulting in inaccurate representations of the constructed features. The model Collaborative Prospective Knowledge-aware Attentive Network (CPKAN), which is based on a heterogeneous propagation strategy and uses the attention module to control the representation of each hop triplet set, is proposed in this paper as a solution to these issues. This model improves the association between the current hop triplet set, the initial seed, and the previous hop triplet set. Meanwhile, it adjusts the preference difference features between experts and fault problems separately to generate more accurate embedding representations of experts and fault problems, which serve as the basis for the subsequent expert recommendation tasks. Results from the experiment demonstrate that CPKAN outperforms the current state-of-the-art model in our dataset in terms of AUC and F1 performance by 1.03% and 4.89%, respectively.