Dynamic Event Knowledge Graph–Based Fault Diagnosis for Multisensor Electromechanical Systems of Aviation Equipment and Its Application
TL;DR
An event knowledge graph is leveraged for UAV sensor fault diagnosis: the graph’s architecture inherently structures the array of distinct sensors, and causal logic derived from event progression boosts the graph nodes’ ability to extract meaningful features, and a relationship-focused attention mechanism captures spatial–temporal patterns within sensor data.
Abstract
Aviation equipment, such as unmanned aerial vehicles (UAVs) rely on a suite of sensors to support their control and navigation functions. Yet, in challenging operational conditions, these sensors are prone to malfunction, posing risks of serious safety incidents and financial harm. Although neural network–based fault diagnosis techniques have gained traction in mechanical systems, they typically struggle to synthesize data from multiple sources and disregard the value of event-based system insights. The rise of large language models (LLMs) and multimodal foundation models is transforming risk assessment and predictive maintenance in engineering systems. These models excel at processing heterogeneous data including maintenance logs, equipment manuals, time-series sensor data, and visual reports, yet suffer from inherent cognitive uncertainty, model hallucinations, black-box opacity, and poor adaptability to dynamic working environments, which limit their use in safety-critical aviation systems. To address this, this work leverages an event knowledge graph (EKG) for UAV sensor fault diagnosis: the graph’s architecture inherently structures the array of distinct sensors. Concretely, causal logic derived from event progression boosts the graph nodes’ ability to extract meaningful features, and a relationship-focused attention mechanism (built on gated convolution) captures spatial–temporal patterns within sensor data. Additionally, this study incorporates a UAV dynamic model (rooted in an event-triggered framework) into EKG construction to improve diagnostic precision, interpretability, and safety margins. When tested on three custom-built data sets, the approach achieved diagnostic accuracies of 94.93%, 98.71%, and 92.97%, respectively. Relative to standard data-driven methods and pure LLM-based diagnostic schemes, the proposed technique also exhibits stronger feature extraction performance, fewer hallucination risks, full decision traceability, and delivers the top overall diagnostic accuracy.
