A Scalable Transformer-based Framework for Fault Detection in Mission-Critical Systems

Chao Wang, Bill Zeng, Huaze Tang, Zhenpeng Shi, Qian Zhang, Wenbo Ding

Keywords: multi-agent systems, unmanned aerial vehicle, offline reinforcement learning, scenario-based testing

Abstract:

Detecting underlying faults is crucial in the development of mission-critical planning systems, such as UAV trajectory planning in Unmanned aircraft Traffic Management (UTM), which is vital to airspace safety.
Inevitably, there exists a small set of rare, unpredictable conditions where the UTM could suffer from catastrophic failures.
Most traditional fault detection approaches focus on achieving high coverage by random input exploitation.
However, random methods are struggling to detect long-tail vulnerabilities with unacceptable time consumption.
To tackle this challenge, we propose a scenario-oriented framework to search long-tail conditions, accelerating the fault detection process.
Inspired by in-context learning approaches, we leverage a Transformer-based policy model to capture the dynamics of the subject UTM system from the offline dataset for exploitation acceleration.
We evaluate our approach over 700 hours in a massive-scale, industry-level simulation environment.
Empirical results demonstrate that our approach achieves over 8 times more vulnerability discovery efficiency compared with traditional expert-guided random-walk exploitation, which showcases the potential of machine learning for fortifying mission-critical systems.
Furthermore, we scale the model size to 2 billion parameters, achieving substantial performance gains over smaller models in offline and online evaluations, highlighting the scalability of our approach.

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9720