[4Yin-B] Poster 7

[4Yin-B-05] Practical Integrated Design of Knowledge Graphs and Bayesian Networks for Fault Diagnosis

Tomoaki Ishikawa¹, 〇Tomoya Onishi¹, Nobuki Nemoto¹ (1. TOSHIBA Corporation)

Keywords:

Fault Diagnosis,Knowledge Graph,Bayesian Network

In the field of fault diagnosis, the scarcity of failure data often makes it difficult to apply machine-learning-based methods. Although knowledge graphs provide the advantage of systematically organizing expert knowledge, they are not sufficient for dynamically prioritizing potential fault causes according to observed sensor conditions. To address this limitation, this study proposes a practical diagnostic model that integrates knowledge graphs with Bayesian networks. The knowledge graph is used to narrow down plausible cause candidates, and a compact Bayesian network is constructed for the reduced subset, enabling probabilistic inference under uncertainty and prioritization of causes. Furthermore, we clarify that, in specifying conditional probability tables, the design of prior distributions based on expert knowledge rather than data-driven learning plays a central role, and we outline strategies for mitigating inference bias that arises when normal data dominate available records. This study demonstrates that effective fault-cause estimation is feasible even in data-poor environments by appropriately leveraging expert knowledge.