Advances in Production Engineering & Management
Volume 20 | Number 3 | September 2025 | pp 380–390
https://doi.org/10.14743/apem2025.3.547
Manufacturing process quality prediction via temporal knowledge graph reasoning with adaptive multi-scale temporal path fusion and self-attention mechanism
Zong, H.; Shuai, B.
ABSTRACT AND REFERENCES (PDF) |
FULL ARTICLE TEXT (PDF)
A B S T R A C T
To tackle the pronounced temporal dynamics and intricate interdependencies within process manufacturing knowledge, this paper introduces an innovative framework: the Adaptive Multi-Scale Temporal Path Fusion Network (AMTPFNet). The method constructs short-term (high-frequency) and long-term (low-frequency) historical subgraphs to generate multi-scale temporal representations. It also employs a self-attention mechanism for query-aware temporal path modeling, enabling adaptive weight allocation based on varying time spans. Extensive experiments are conducted on benchmark datasets, including ICEWS18, GDELT, WIKI, and YAGO. Additionally, an application analysis is presented using electromechanical fault data. The results demonstrate that AMTPFNet exhibits remarkable effectiveness and robustness in temporal knowledge graph reasoning tasks, achieving MRR scores of 0.914 on YAGO and 0.838 on WIKI. It achieves high efficiency in predicting future production facts and assessing process quality in industrial workflows. Root causes of failures (e.g., insulation, friction) for motor components (stators, rotors) are accurately predicted, demonstrating the framework’s transferability to real-world manufacturing scenarios. Although electromechanical fault data are used as a case study, the framework generalizes to manufacturing quality prediction and is readily transferable to finance, healthcare, and social media analytics.
A R T I C L E I N F O
Keywords • Temporal knowledge graph reasoning (TKGR); Multi-scale temporal modeling; Temporal path fusion; Self-attention mechanism; Knowledge graph embedding; Manufacturing process analytics; Process quality prediction
Corresponding author • Zong, H. , Ying, K.-C.
Article history • Received 25 August 2025, Revised 1 October 2025, Accepted 13 October 2025
Published on-line • 31 October 2025
E X P O R T C I T A T I O N
» RIS format (EndNote, ProCite, RefWorks, and most other reference management software)
» BibTeX (JabRef, BibDesk, and other BibTeX-specific software)
» Plain text
