.png){kind=icon}
.png){kind=icon}
Return
CausalFPS: A causal discovery-based method for UAV flight parameter selection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ymssp.2026.113851.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The flight performance of Unmanned Aerial Vehicles (UAVs) is highly dependent on their complex sensing systems, and accurate detection and troubleshooting of their sensors is an important topic in the current development of the technology. The UAV ’s sensing system can capture various flight parameters during the mission. However, UAV is a highly coupled complex system, and multiple parameters interact with each other, which brings great challenges to the selection of input parameters based on artificial experience or statistical correlation. Therefore, from the perspective of causality, we propose a causal discovery method for UAV sensor fault parameter selection—CausalFPS. CausalFPS aims to deeply understand the propagation path of faults by capturing the nonlinear causal relationship between sensor parameters, and select the most discriminative core parameters for fault diagnosis. Specifically, firstly, the Structural Equation Modelling (SEM) and Variational Autoencoder (VAE) are fused to achieve the nonlinear causal inference of multi-sensor parameters, and the general causal relationship between parameters is initially constructed. Secondly, a fault information oriented causal discovery module is innovatively designed to guide the model to focus on learning the causal relationship of fault information by obtaining the strength of each parameter node in real time and weighting the input parameters. Finally, the PageRank algorithm is used to quantify the global influence of each parameter in the causal network and sort them, and then the most influential parameters are selected for fault diagnosis. The effectiveness and superiority of CausalFPS has been verified in real flight cases, where CausalFPS achieves the diagnostic accuracy of other algorithms selecting 20 parameters when only 5 parameters are selected.
Journal
IF:
8.9
Papers:
1.3W
Citations:
6.6W