Causal‐Aware Dynamic Knowledge Graph for Wind Turbine Fault Diagnosis: Integrating Domain Knowledge With SCADA Data
Year: 2026
Type: article
Abstract: ABSTRACT This paper proposes a Causal‐Aware Dynamic Knowledge Graph (CA‐DKG) framework that systematically fuses FMEA domain knowledge with data‐driven causal discovery from SCADA time series. We first develop a heterogeneous knowledge graph construction method that unifies structured FMEA failure triples with sensor monitoring relationships, providing a unified representation framework for cross‐source fault information covering components, failure modes, root causes and operational variables. A multi‐stage causal discovery pipeline is further designed, integrating the PCMCI algorithm, expert physical rule audit and VCDF multi‐window stability verification, which effectively filters spurious correlations and extracts high‐confidence directed causal edges from SCADA time series. On this basis, a relational graph attention network embedded with EvolveGCN temporal modeling is constructed to learn dynamic graph representations, enabling interpretable causal link prediction and fault propagation reasoning instead of opaque black‐box fault detection. Validated on 161 days of SCADA data from a 5 MW offshore wind turbine with 23 sensor variables including physics‐based thermal residuals, CA‐DKG achieves 0.88 AUC and 0.61 F1‐score for causal edge prediction, outperforming the topology‐based PageRank baseline by 20% in AUC, while VCDF retains 66.7% of stable causal edges. This work provides an explainable knowledge‐data integrated framework for wind turbine intelligent fault diagnosis and predictive maintenance, supporting trustworthy causal reasoning for maintenance decisions.
DOI: 10.1002/qre.70440
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7222397427