Vibration-Based Fault Detection in Drone using Artificial Intelligence

Recent years have seen a huge increase in the study of drones. There is a
lot of published articles regarding drone, focusing on control
optimization, fault detection, safety mechanisms, etc. In fault detection,
most of the studies focused on the effects of faulty propellers
and rotors, and there is very limited academic research on drone arms. In
this paper, a fault detection based on the vibration of the multirotor
arms using artificial intelligence (AI) is proposed. There are some cases
where due to an accident, the arm of the multirotor crack or loose. This
is normally unnoticeable without disassembly and if not taken care of, it
would have likely resulted in a sudden loss of flight stability, which
will lead to a crash. Two types of AI methods are incorporated in this
study, namely, fuzzy logic and neuro-fuzzy, using the fuzzy logic and
ANFIS toolbox in the MATLAB software, respectively. For the neuro-fuzzy
approach, we use 100 and 1000 datasets to determine the effects
of the dataset size on the neuro-fuzzy performance. Both datasets
are divided into training (80%), testing (10%), and checking data (10%).
The guidelines for constructing the AI algorithms are based on the
experimental data for five experimental conditions; (i) (a) Original
multirotor condition without modifying the multirotor arms, (b) 100%
screwed multirotor arms condition (full tighten), (c) 50% screwed
multirotor arms condition (half tighten), (d) 10% screwed multirotor arms
condition, and (e) Unscrewed multirotor arm conditions. Their results are
compared to determine the best method in predicting the safety of the
multirotor. Both methods provided acceptable decision making but the
neuro-fuzzy approach depends on the dataset used as overfit model might
give incorrect decision making. Because the vibration data are collected
in an indoor environment, this framework is more suitable for early
prediction before flying the multirotor outdoor. A video
demonstrating the real-time deployment of our proposed method is included
in the mp4 format.