Safe-to-fly: An On-board Intelligent Fault Diagnosis System With AutoML for Unmanned Aerial Vehicles

Ritik Bompilwar, Surya Pratap Singh Rathor, Aparna Sinha, Debanjan Das

2022 IEEE Delhi Section Conference (DELCON)

Abstract

Real-time fault diagnosis in Unmanned Aerial Vehicles (UAVs) is a challenging task. Data-driven intelligent diagnosis of faults ensures flight safety for UAVs. In this paper, real-time fault diagnosis on small scale fixed-wing UAVs has been shown by data of natural flight conditions with a wrapped wing structure that breaks the geometric symmetry. AutoML based approach was taken for multi-class fault classification. Two datasets were created from the combination of flight data of two days. The experimental results showed that the proposed Deep Learning AutoML model significantly improves performance over conventional Machine Learning methods such as the Decision Tree, KNN, and the Random Forest. The test accuracy of the proposed AutoML model was 74% and 100% on the first and second datasets, respectively. The AutoML model's capability in classifying low fault severity and complex faults demonstrated the method's usefulness and excellence.

Methodology

Safe-to-fly proposes an on-board AutoML-based fault diagnosis pipeline for fixed-wing UAVs. The system uses real flight data from two days and performs multi-class classification across normal flight and multiple fault conditions tied to left and right control-surface efficiency. Two datasets are constructed: the first uses 21st July data for training and 23rd July data for test, and the second combines both days under an 80:20 split. Decision Tree, KNN, and Random Forest baselines are compared against AutoML models built with Auto-Keras neural architecture search.

Results

Model	Accuracy	Precision	Recall	F1
AutoML	100%	99%	99%	99%
Decision Tree	76%	35%	44%	37%
KNN	93%	88%	86%	87%
Random Forest	76%	43%	44%	41%

On the combined dataset (21st + 23rd July, 80:20 split), the AutoML model reaches 100% accuracy with 99% precision, recall, and F1, clearly outpacing Decision Tree, KNN, and Random Forest. On the harder cross-day setting, AutoML still trains to 91% and tests at 74%.

Citation

@INPROCEEDINGS{9752852,
  author    = {Bompilwar, Ritik and Rathor, Surya Pratap Singh and Sinha, Aparna and Das, Debanjan},
  booktitle = {2022 IEEE Delhi Section Conference (DELCON)},
  title     = {Safe-to-fly: An On-board Intelligent Fault Diagnosis System With AutoML for Unmanned Aerial Vehicles},
  year      = {2022},
  doi       = {10.1109/DELCON54057.2022.9752852}
}