Our member Elgiz Baskaya will defend her thesis on 16/05/2019 at 10:00 am in room G11 in ENAC. We welcome you all.
Titre : ‘Fault detection and diagnosis for drones using machine learning’
This new era of small UAVs currently populating the airspace introduces many safety concerns, due to the absence of a pilot onboard and the less accurate nature of the sensors. This necessitates intelligent approaches to address the emergency situations that will inevitably arise for all classes of UAV operations as defined by EASA (European Aviation Safety Agency). Hardware limitations for these small vehicles point to the utilization of analytical redundancy, rather than to the usual practice of hardware redundancy in manned aviation. In the course of this study, machine learning practices are implemented in order to diagnose faults on a small fixed-wing UAV to avoid the burden of accurate modeling needed in model-based fault diagnosis. A supervised classification method, SVM (Support Vector Machines) is used to classify the faults. The data used to diagnose the faults are gyro and accelerometer measurements. The idea to restrict the data set to accelerometer and gyro measurements is to check the method’s classification ability, with a small and inexpensive chip set and without the need to access the data from the autopilot, such as the control input information.
This work addresses the faults in the control surfaces of a UAV. More specifically, the faults considered are the control surface stuck at an angle and the loss of effectiveness. First, a model of an aircraft is simulated. This model is not used for the design of Fault Detection and Diagnosis (FDD) algorithms, but is instead utilized to generate data. Simulated data are used instead of flight data in order to isolate the probable effects of the controller on the diagnosis, which may complicate a preliminary study on FDD for drones. The results show that for simulated measurements, SVM gives very accurate results on the classification of the loss of effectiveness faults on the control surfaces. These promising results call for further investigation so as to assess SVM performance on fault classification with flight data. Real flights were arranged to generate faulty flight data by manipulating the open source autopilot, Paparazzi. All data and the code are available in the code sharing and versioning system, Github. Training is held offline due to the need for labeled data and the computational burden of the tuning phase of the classifiers. Results show that from the flight data, SVM yields an F1 score of 0.98 for the classification of control surface stuck faults. For the loss of efficiency faults, some feature engineering, involving the addition of past measurements is needed in order to attain the same classification performance. A promising result is discovered when spinors are used as features instead of angular velocities. Results show that by using spinors for classification, there is a vast improvement in classification accuracy, especially when the classifiers are untuned. Using spinors and a Gaussian Kernel, an untuned classifier gives an F1 score of 0.9555, which was 0.2712 when gyro measurements were used as features. In summary, this work shows that SVM gives a satisfactory performance for the classification of faults on the control surfaces of a drone using flight data.