Summary

Paper proposes a Time Line Modeling (TLM) method based on UAV SITL simulation environment to obtain and process on-board failure logs of drones.

TLM method includes two stages:

• Fault Time Point Anchoring Method
• Fault Time Window Stretching Method

Multiple flight missions executed in SITL simulation environment.

Failure of several common components of UAV:

• GPS
• Accelerometer
• Engine
• RC

Initial and end location of fault determined by Fault Time Point Anchoring.

In data processing, features that are not universal are removed.

Flight data of UAV is optimised by using data balance method of Time Window Stretching to balance normal and abnormal data.

Applied Sequential Minimal Optimisation, Random Forest, and Convolutional Neural Network to process data.

Rationale

Data volume is large in volume, but volume of abnormal data is less compared to normal data (data imbalance).

Based on Time Line Modeling method to acquire and process UAV fault flight data in sim environment.

Time related features and non-universal features are removed.

Initial and end positions of fault data are determined using Fault Time Point Anchoring method.

Time Window Stretching method used to supplement abnormal data so that the data becomes balanced.

Environment

SITL in ArduPilot to simulate flight of UAV.

QGroundControl to visualise real-time position of UAV.

Python interface of Dronekit used to control drone.

Ubuntu 18.04 based on VMware Workstation 16.2.3.

Ground station is on a Windows 11 laptop (16 GB RAM, 11th Gen Intel(R) Core(TM) i7-11800H @2.30 GHz)

Methodology

Flight path is generated and then the points are linearly interpolated to inject more points. This will be used later.

Simulated faults:

• Simulation of GPS Fault - GPS disabled.
• Simulation of Accelerometer Fault - Accelerometer 1 and 2 disabled.
• Simulation of Engine Faults - Engine switched off.
• Simulation of RC System Fault - RC Disabled

Data Processing

Logs converted to .csv.

Time dependent features are removed (referring to a particular time anchor that repeats periodically), and so are non-universal features (basically tethered to a particular location or time).

Evaluation

Producing some arcane strategy to stretch fault window. Doesn't make any sense. Should show some examples, but didn't.

No dataset provided.