Building causal models for finding actual causes of unmanned aerial vehicle failures

Abstract

Read online

Finding actual causes of unmanned aerial vehicle (UAV) failures can be split into two main tasks: building causal models and performing actual causality analysis (ACA) over them. While there are available solutions in the literature to perform ACA, building comprehensive causal models is still an open problem. The expensive and time-consuming process of building such models, typically performed manually by domain experts, has hindered the widespread application of causality-based diagnosis solutions in practice. This study proposes a methodology based on natural language processing for automating causal model generation for UAVs. After collecting textual data from online resources, causal keywords are identified in sentences. Next, cause–effect phrases are extracted from sentences based on predefined dependency rules between tokens. Finally, the extracted cause–effect pairs are merged to form a causal graph, which we then use for ACA. To demonstrate the applicability of our framework, we scrape online text resources of Ardupilot, an open-source UAV controller software. Our evaluations using real flight logs show that the generated graphs can successfully be used to find the actual causes of unwanted events. Moreover, our hybrid cause–effect extraction module performs better than a purely deep-learning based tool (i.e., CiRA) by 32% in precision and 25% in recall in our Ardupilot use case.

Keywords

• unmanned aerial vehicle
• actual causality
• natural language processing
• causal graph
• root cause analysis
• automated diagnosis