Coarse Exogenous Fault Detection in Multi-Robot Systems using Variational Autoencoders and Normalizing Flows
Öffentlich DepositedMulti-Robot Systems (MRS) are a fast-growing, widely researched facet of robotics that involve multiple robots that can cooperate and communicate to complete a given task. Many tasks envisioned for MRS are set in dynamic environments that are dangerous. In such cases, the likelihood of the occurrence of a fault is high. Many unpredictable factors such as falling debris and fire become potential sources of faults in the system. Therefore, fault detection algorithms are an essential component of autonomous MRS. Fault detection in MRS is classified as endogenous and exogenous. Endogenous algorithms are used by a robot to detect faults in itself, whereas exogenous algorithms are used by robots to detect faults in other robots. In this work we focus on exogenous fault detection. Development of exogenous algorithms so far mainly addresses fine-grained fault detection. Such algorithms require detailed information about specific components, such as the output of specific sensors and actuators. Though such methods perform well with detecting specific types of faults, they are computation-intensive since they involve multiple sequential steps to generate inferences. This can be solved by using machine-learning-based approaches. We are the first to use the Variational Autoencoder (VAE), an unsupervised-deep-learning-based algorithm, for fault detection in MRS. We show how VAEs improve fault detection performance using only information about the overall behavior of a robot such as its position and heading, as perceived by its peers. We show that this approach is fast and can generalize well to a wide range of faults as well as a wide range of MRS tasks. We also show how our approach can be used to monitor multiple robots simultaneously, thus reducing the computation resources required to monitor the MRS. We also discuss the limitations of our approach and propose that our algorithm is a promising solution to fault detection when used to complement existing state-of-art techniques. We compare multiple design choices for the VAE such as different input representations, architecture, and loss functions. We further improve the fault detection performance of the VAE using Normalizing Flows and provide a comparison of multiple variants of flows.
- Creator
- Mitwirkende
- Degree
- Unit
- Publisher
- Identifier
- etd-72391
- Stichwort
- Advisor
- Orcid
- Committee
- Defense date
- Year
- 2022
- Date created
- 2022-08-17
- Resource type
- Source
- etd-72391
- Rights statement
- Zuletzt geändert
- 2023-09-28
Beziehungen
- In Collection:
Objekte
Permanent link to this page: https://digital.wpi.edu/show/41687m812