Fault Detection & Diagnosis for Small UAVs

via
Machine Learning
Elgiz Baskaya Murat Bronz Daniel Delahaye
ENGIE Ineo - Groupe ADP - UAS Lab, ENAC Laboratoire MAIAA, ENAC
SAFRAN RPAS Chair Toulouse, FRANCE Toulouse, FRANCE
UAS Lab, ENAC Email: [email protected] Email: [email protected]
Toulouse, FRANCE
Email: [email protected]

Abstract—The new era of small UAVs necessitates intelligent on cost effective systems, resulting in the smaller sensors
approaches towards the issue of fault diagnosis to ensure a and actuators with less accuracy, impose the software to
safe flight. A recent attempt to accommodate quite a number achieve even more. The expectation that UAVs should be less
of UAVs in the airspace requires to assure a safety level. The
hardware limitations for these small vehicles point the utilization expensive than their manned counterparts might have a hit on
of analytical redundancy rather than the usual practice of reliability of the system. Cost saving measures other than the
hardware redundancy in the conventional flights. In the course of need to support a pilot/crew on board or decrement in size
this study, fault detection and diagnosis for aircraft is reviewed. would probably lead to decrease in system reliability.
An approach of implementing machine learning practices to Systems are often susceptible to faults of different nature.
diagnose faults on a small fixed-wing is selected. The selection
criteria behind is that, data- driven fault diagnosis enables Existing irregularities in sensors, actuators, or controller could
avoiding the burden of accurate modeling needed in model-based be amplified due to the control system design and lead to
fault diagnosis. failures. A fault could be hidden thanks to the control action
In this study, first, a model of an aircraft is simulated. This (1).
model is not used for the design of Fault Detection and Diagnosis The widely used method to increase reliability is to use
(FDD) algorithms, but instead utilized to generate data and test
the designed algorithms. The measurements are simulated using more reliable components and/or hardware redundancy. Both
the statistics of the hardware in the house. Simulated data is requires an increase in the cost of the UAS conflicting one
opted instead of flight data to isolate the probable effects of the of the main reasons of UAS design itself band consumer
controller on the diagnosis, which will complicate this preliminary expectations (2). To offer solutions for all different foreseen
study on FDD for drones. categories of airspace, a variety of approaches should be
A supervised classification method, SVM (Support Vector
Machines) is used to classify the faulty and nominal flight considered. While hardware redundancy could cope with the
conditions. The features selected are the gyro and accelerometer failure situations of UAVs in the certified airspace, it may
measurements. The fault considered is loss of effectiveness in not be suitable for UAVs in open or some subsets of specific
the control surfaces of the drone. Principle component analysis categories due to budget constraints. Analytical redundancy
is used to investigate the data by reducing the feature space is another solution, may be not as effective and simple as
dimension. The training is held offline due to the need of
labeled data. The results show that for simulated measurements, hardware redundancy, but relies on the design of intelligent
SVM gives very accurate results on the classification of loss of methods to utilize every bit of information on board aircraft
effectiveness fault on the control surfaces. wisely to deal with the instances.
There are three approaches to achieve safe FTC in standard
I. I NTRODUCTION flight conventions. First one is the fail operational systems
The cost effectiveness and reachability of COTS elements, which are made insensitive to any single point component
shrinking size of electronics serve as a perfect environment failure. The second approach is the fail safe systems where
for small flying vehicles to emerge. This accelerating trend a controlled shut down to a safe state is practiced whenever
towards small but capable flying vehicles is pushing the limits a critical fault is pointed out by a sensor. The level of
of both hardware and software potentials of industry and degradation assures to switch to robust (alternate) or direct
academia. Increasing usage of these vehicles for a variety of (minimal level of stability augmentation independent of the
missions pushes a further liability to secure the flight. nature of the fault) mode. Switching from nominal mode
To achieve a safe flight is not an easy task considering the to the robust and direct modes leads to a decrease in the
unknowns of the systems hardware, environment and possible available GNC functions. This causes a degradation in ease
system faults and failures to emerge. Also, increasing demand of piloting. And also some optimality conditions could have

978-1-5386-0365-9/17/$31.00 ©2017 IEEE

 should not only be sensitive to the faults but also robust to
 


the model uncertainties and external disturbances.
" Two distinct options to proceed in analytical redundancy
are the model based approaches and data-driven approaches.
They form the two ends of a continuous solution set line,
so utilizing them in a combination might end up with better

 
 solutions. Model based fault diagnosis highlights the compo-
nents of a system and the connections in-between, and their
corresponding fault modes. Data driven fault diagnosis rely on
the observational data and prefers dense, redundant and with
 

a frequency larger than the failure rate.

  


  !





A. Model Based

! In model based approaches, relations between measure-

ments and estimated states are exploited to detect possible
dysfunction. The most common ways to implement a model
based approach is to estimate the states, estimate the model
parameters, or parity-space. The accuracy of the results depend
Fig. 1. Variations of fault tolerant control systems
on the type of faults (additive or multiplicative). Additive faults
affects the variables of the process by a summation whereas
the multiplicative faults by a multiplication. When only output
been compromised. The third approach is fault tolerant control signal can be measured, signal model based methods can
systems in which redundancy in the plant and the automation be employed for fault detection such as Bandpass filters,
system is employed to design software that monitors the Spectral analysis(FFT) and maximum entropy estimation. For
components and takes in action whenever needed. The strategy the case, both the input and output signals are available, the
is most probably to try to keep plant availability and accept utilized methods for fault detection are called the process
reduced performance (3). based methods: State and output observers(estimators), Parity
II. M ETHODS FOR FTCS equations and Identification and parameter estimation. They
generate residuals for state variables or output variables. When
A common categorization of FTCS is passive and active
previous works investigated, it is concluded that the most
FTCS. In passive FTCS, the flight controller is designed in
widely used technique for sensor and actuator faults is the
such a way to accommodate not only the disturbances but
state and output observers (estimators) and for process faults,
also the faults. Active FTCS first distinguishes the fault via
identification and parameter estimation (4).
fault detection and diagnosis module and then switch between
the designed controllers specific to the fault case or design a The output of the model based fault detection methods
new one online (2). While active FTCS requires more tools is the stochastic behaviour with mean values and variances.
to handle faults as seen in Fig. 1, for faults not predicted and With the use of change detection methods, deviations from
not counted for during the design of the robust controller, this the normal behavior can be detected. For that purpose, three
method most probably fails. available methods considered are, mean and variance estima-
Even with a long list of available methods, aerospace tion, likelihood-ratio-test and Bayes decision, run-sum test and
industry has not implemented FTC widely, except some space two-probe t-test. Fault detection is only supported by simple
systems, due to the evolving nature of the methods, the tricks threshold logic or hypothesis testing in most of the applications
coming with the nonlinear nature of the problem, design (4).
complexity and high possibility of wrong alarms in case A bunch of studies discovers the band of different ap-
of large disturbances and/or modeling uncertainties. So the proaches for model-based fault detection. Detecting sensor and
already carried reliability measures concerning the hardware actuator faults via state estimation, utilizing an EKF is applied
redundancy is now the preferred way because of its ease and to a F-16 model in (5). Parameter identification via H∞ filter
maturity being implemented on various critical missions with is used to indicate icing in (6).
considering human lives. A drawback of model-based approaches is that they re-
quire accurate model of the aircraft for successful detection.
III. FAULT D ETECTION AND D IAGNOSIS In a small UAV system susceptible to various uncertain-
FDD is handled in two main steps; fault detection and fault ties/disturbances and most of the cases does not have an
diagnosis. Fault diagnosis encapsulates fault isolation and fault accurate model, leading a model-based approach might fail.
identification. The methods for detection and diagnosis are And also, a mathematical model of a UAV is constructed
investigated for their frequency of utilization separately for within the flight envelope, and does not necessarily describe
sensor, actuator, process and controller faults in (4). FDD the possible dynamics invoked by a failure on board.
 A way to handle that is to offer solutions to cope with
the uncertainties. A fairly old study in 1984, investigates the
design problem FDI systems robust to uncertainties within
the models. One of the two steps of FDI, two steps being
the residual generation and decision-making, is targeted. They
offer to handle model uncertainties, by designing a robust
residual generation process (7). Another study deals with
model uncertainties by determining the threshold of the resid-
ual in a novel way with an application to detect aileron actuator
fault (8). (9) utilize two cascade sliding mode observers state
estimation and fault detection to guarantee staying in sliding
manifold in the presence of unknown disturbances and faults.
B. Machine Learning
Model-based approaches had various successful applications
Fig. 2. Common actuator faults (1)
until now, most of them assuming accurate model is available
on board. With the new era of UAVs, the airspace is expected
to be populated by an abrupt increase in the number of
UAVs. The variety of UAVs, expense of accurate modeling thanks to its improvement in accuracy of detection (19).
practices, the difficulty in modeling the behavior of UAV Application of SVM on fault detection is mostly held in
in case of failures, call for alternative approaches for the mechanical machinery, such as roller bearings, gear box, turbo
quite challenging problem of FDD. The increased efficiency of pump rotor and sometimes other systems; semi-conductors,
sensors on board, the increase in the computational capabilities refrigeration systems and chemical processes. Its application
of autopilot processors, and the advances in machine learning on complex systems has not been very widely adopted yet and
techniques in the last decade may offer efficient data-driven forms the basis of study for our research.
solutions to FDD.
IV. S YSTEM M ODELING
In data driven methods, a detailed knowledge about the
internal dynamics of the system is not necessary. The data In this study, first, a model of an aircraft is simulated. This
available is the source of information with regard to the model, will not be used for the design of FDI algorithms, but
behavior of the system. Supervised learning, which requires instead will be utilized to test them. Nonlinear aircraft flight
to label the fault cases previously in the training data, is dynamics for translational and attitude motion can be given as
usually utilized for data-centric inference of causes. In case a system of first order partial differential equations
of an unlabeled fault, the result is expected as a probability
distribution of the available normal modes, identified fault
ẋN ED = Cbn v b (1)
labels and a probable unknown fault. What is needed at that
1 b  b × b
point is to first detect and localize the fault and then to consult v̇ b = mg + Ftb + Fab − ωb/i v (2)
domain experts for labeling for further integration of this fault m
1
into the diagnosis scheme (10). q̇0 = − qνT ωb/i
b
(3)
2
1 × 
Amidst data driven methods for FDD, such as Neural
b
Networks (11) and Principal Component Qnalysis (PCA) (12), q̇ν = qν + q0 I3 ωb/i (4)
Support Vector Machines (SVM) appear more recently in the 2
b b × b
literature. (13) argues artificial intelligence methods for fault J ω̇b/i = M − ωb/i J ωb/i (5)
detection of complex systems. Comparison between PCA and
where xN ED ∈ IR3 is the position of the center of mass of
model based stochastic parity space approaches is given in
UAV with respect to inertial frame I expressed in the body
(14). In (15), the authors argues to use dynamic PCA since
frame B, v b is the velocity of the center of mass of UAV with
UAV flight controls is a dynamic system itself and DPCA can
to I expressed in B, q = [q0 , qvT ]T ∈ IR3 × IR is the unit
reflect unknown disturbances, while model-based approaches
quaternion representing the attitude of the body frame B with
can only model typical disturbance.
respect to inertial frame I expressed in the body frame B,
SVM is introduced in 1964 in the statistical learning theory
J ∈ IR3×3 is the positive definite inertia matrix of the drone,
domain and relies on structural risk minimization principle
M ∈ IR3 represents the moments acting on the drone. The
(16). Although the theory has old roots, its application to
notation x× for a vector x = [x1 x2 x3 ]T represents the
classification as a machine learning algorithm is recent and
skew-symmetric matrix
originally offer solutions for two-class classification (17; 18).
⎡ ⎤
SVM’s first application as a classifier was mainly on object 0 −x3 x2
classification in images and followed by fault detection lately. x× = ⎣ x 3 0 −x1 ⎦ (6)
The use of SVM on fault detection has gained popularity −x2 x1 0
 5 6
4
normal
fault
3
4
2

0 2
0 20 40 60 80 100 120 140 160 180

z2
3
0
2

0 -2
-1

-2

-3 -4
0 20 40 60 80 100 120 140 160 180
-6 -5 -4 -3 -2 -1 0 1 2
z1
Fig. 3. Loss of effectiveness fault simulation in aileron command and
corresponding accelerometer x axis measurement Fig. 4. Principal component analysis for visualization of faulty and normal
data in reduced dimensional feature space

The stability and aerodynamic force coefficients are gener-

V. C LASSIFICATION OF FAULT VIA SVM
ated by AVL. The input vector can be written as u (t) ∈ IR3
 T SVM is a relatively new approach for classification offering
u (t) = δa δe n (7) better generalization property thanks to its foundations on the
structural risk minimization principle (21; 22) while other
Here δa aileron deflection angle in degrees, δe elevator classifiers usually only minimizes the empirical risk. This
deflection angle in degrees, n engine speed in rev/s. advances the capacity of generalization even with a small
When the actuators are healthy, actual control input signal number of instances by reducing the risk of overfitting for a
will be equal to the given input signal. In case of a fault the nicely tuned parameters setting. It can be applied to nonlinear
actual signal can be modeled as systems and problems offering a vast number of features. Fur-
thermore, taking advantage of convex optimization problems
u (t) = Euc + uf (8) in the solution of SVM models, another attractive reason to
use SVM rises as avoidance of global minimas, while Neural
where uc is the desired control signal, E = diag(e1 , e2 , e3 ) Networks is inherently prone to local minimas.
is the effectiveness of the actuators where 0 ≤ ei ≤ 1 with The idea behind SVM is to find an optimal hyperplane
(i = 1, 2, 3) and uf additive actuator fault. This model makes that will linearly separate the classes. This is achieved with
it possible to simulate all four types of actuator faults shown the introduction of maximum margin concept which is the
in Fig. 2. Most of the FDI algorithms are implemented to distance in between the boundaries when they are extended
open-loop systems, ignoring the probable influences of the until hitting the first data point as in Fig. 5. The points closest
controller might cause on the detection performance (20). Here to the hyperplane (decision boundary) are called the support
the system is open-loop as well. Further implementation of a vectors and are the representatives of the data sets to be used
controller is foreseen to understand the effect of the selected for the decision process. This helps to decrease the data to
controllers. So we follow a step by step approach and hope handle abruptly, enhancing the ability to cope with the curse
to end with a more realistic case, in which real flight data is of dimensionality and reducing the computational complexity.
utilized and diagnosis is achieved online aside a functioning SVM has other tricks to deal with not linearly seperable
controller. problems such as using kernels to map data into higher
Since it is not possible to see all features, we take advantage dimensional feature spaces where they can be separated with
of the dimensionality reduction technique called Principle a linear hyperplane.
Component Analysis (PCA) for visualization. Here what we A binary classifier is used in this work to classify two
do is to map the feature vector, x ∈ IRn to a lower dimensional classes, faulty and nominal. The fault considered in this study
space where the new feature set will be represented by is the loss of effectiveness of the control surfaces. SVM being
z ∈ IRk . Fig. 4 shows the resulted most significant elements a supervised classification algorithm has two main phases as
for a mapped feature space from six dimensional feature vector shown in Fig. 6. In the training phase, the model is learned as
to two. a fit to the labeled data that is fed to the SVM algorithm. This
 
x2




 




      








  



Fig. 6. Supervised learning basics



 
5

x1 3

Fig. 5. SVM working principle

1

0
0 20 40 60 80 100 120 140 160 180

phase is usually followed with a tuning phase where some of

the parameters of SVM is changed and results are compared 2

to have the best fit via cross validation to avoid overfitting.

1
The last phase is the prediction, where for a new instance
the classifier predicts if it corresponds to a faulty or nominal 0
condition.
-1
Training data is comprised of labeled data where the label
can belong to one of two possible cases. This data set is saved -2
in X ∈ IRm×n where m, n correspond to number of instances 0 20 40 60 80 100 120 140 160 180

and features respectively. The label information corresponding

to the measurement instances is also fed to the SVM algorithm Fig. 7. Posterior probability of loss in effectiveness fault for test set when a
during the training phase as output vector y ∈ {−1, 1}. The fault is injected at t = 120s.
aim of SVM is to find an optimal hyperplane maximizing the
margin by solving the optimization problem for non-linearly
separable datasets the classifier is tested on the test set. This parameter also
tuned for the outliers to generalize the distribution of the

m
1 2
data rather than resulting in fine fits for each individual data
minγ,ω,b ω + C ξi (9) in the training set. With a satisfactory result of the training
2 i=1 & tuning is followed by the prediction where the classifier
s.t. y i (ω T x( i) + b) ≥ 1 − ξi , i = 1, · · · , m (10) predicts if the new measurement data belongs to the faulty or
ξi ≥ 0, i = 1, · · · , m (11) nominal class. The output of the SVM classification is not the
(12) probability that the new measurement belongs to one class as is
in the traditional classification problems, but directly the class
To avoiding overfitting, which is the main problem of information it belongs to. For investigating the performance of
parametric discrimination approaches such as neural networks, the classifier on the test set, a method (23) is used to calculate
parameter C is tuned to result in the optimal fit for the cross the posterior probabilities giving the probability that the new
validation set. The data set available is first divided to two measurements belongs to faulty mode. Results shows as in
portions with a percentage of %20, %80 where the bigger Fig. 7 that prober tuning achieves very accurate and instant
chunk is the training set and the remaining is the test set. detection for the drone fault.
Further, the training set is divided as cross-validation and
training sets. The idea to split data is to avoid overfitting. VI. C ONCLUSION
Overfitting means that the models trained being very accurate Integration of drones into airspace needs the introduction of
fit for the data they are trained to but fail to generalize with indigenous designs that will serve safe solutions for drones.
new inputs resulting in bad prediction performance for the One of the aspects of the problem is to assure a safe flight by
new data. To assess the performance of the classifier trained designing fault detection and diagnosis with cheaper avionics
with the training data is tuned to give a better performance common in a vast number of drones projected. This work aims
with the cross validation data. And then the final ability of to design a classifier via SVM to solve FDD of drones with
actuator faults. This problem possess various challenges. This [9] R. Sharma and M. Aldeen, “Fault detection in nonlinear
work focuses on a loss of effectiveness fault which is more systems with unknown inputs using sliding mode ob-
difficult than a stuck fault to diagnose, but easier to mitigate. server,” in 2007 American Control Conference. IEEE,
A model of a MAKO UAV is simulated to generate data 2007, pp. 432–437.
and test the designed algorithms. The simulated data of gyro [10] J. Stutz, “On data-centric diagnosis of aircraft systems,”
and accelerometer measurements are given to classifier to train IEEE Transactions on Systems, Man and Cybernetics,
for the two class labeled data set. A supervised classification 2010.
method, SVM (Support Vector Machines) is used to classify [11] M. Schlechtingen and I. F. Santos, “Comparative anal-
the faulty and nominal flight conditions. Principle component ysis of neural network and regression based condition
analysis is used to investigate the data by reducing the feature monitoring approaches for wind turbine fault detection,”
space dimension. The training is held offline due to the need of Mechanical systems and signal processing, vol. 25, no. 5,
labeled data but prediction is envisioned be held real time. The pp. 1849–1875, 2011.
results show that for simulated measurements, SVM gives very [12] X. Sun, H. J. Marquez, T. Chen, and M. Riaz, “An
accurate results on the classification of loss of effectiveness improved pca method with application to boiler leak
fault on the control surfaces. detection,” ISA transactions, vol. 44, no. 3, pp. 379–397,
Further study is envisaged to deal with the controller diag- 2005.
nosis interaction and classification of multiple faults. Also dis- [13] W.-h. Gui and X.-y. Liu, “Fault diagnosis technologies
cussion of SVM for online training might be addressed since based on artificial intelligence for complex process,”
SVM is in need for labeled data which requires generating the Basic Automation, vol. 4, p. 000, 2002.
labeled data during flight. [14] A. Hagenblad, F. Gustafsson, and I. Klein, “A compar-
ison of two methods for stochastic fault detection: the
ACKNOWLEDGMENT parity space approach and principal component analysis,”
This work was supported by the ENGIE Ineo - Groupe ADP 2004.
- SAFRAN RPAS Chair. [15] M. Li, G. Li, and M. Zhong, “A data driven fault
detection and isolation scheme for uav flight control sys-
R EFERENCES tem,” in Control Conference (CCC), 2016 35th Chinese.
[1] G. J. Ducard, Fault-tolerant flight control and guidance TCCT, 2016, pp. 6778–6783.
systems: Practical methods for small unmanned aerial [16] V. Vapnik and A. Chervonenkis, “A note on one class
vehicles. Springer Science & Business Media, 2009. of perceptrons,” Automation and remote control, vol. 25,
[2] P. Angelov, Sense and avoid in UAS: research and no. 1, p. 103, 1964.
applications. John Wiley & Sons, 2012. [17] B. E. Boser, I. M. Guyon, and V. N. Vapnik, “A training
[3] M. Blanke, C. W. Frei, F. Kraus, R. J. Patton, and algorithm for optimal margin classifiers,” in Proceedings
M. Staroswiecki, “What is fault-tolerant control,” in of the fifth annual workshop on Computational learning
Preprints of 4th IFAC Symposium on Fault Detection theory. ACM, 1992, pp. 144–152.
Supervision ans Safety for Technical Processes, SAFE- [18] V. Vapnik, “The nature of statistical learning theory
PROCESS, 2000, pp. 40–51. springer new york google scholar,” 1995.
[4] R. Isermann and P. Ballé, “Trends in the application of [19] N. Laouti, N. Sheibat-Othman, and S. Othman, “Support
model-based fault detection and diagnosis of technical vector machines for fault detection in wind turbines,”
processes,” Control engineering practice, vol. 5, no. 5, IFAC Proceedings Volumes, vol. 44, no. 1, pp. 7067–
pp. 709–719, 1997. 7072, 2011.
[5] C. Hajiyev and F. Caliskan, “Sensor and control sur- [20] R. Pandita, J. Bokor, and G. Balas, “Closed-loop perfor-
face/actuator failure detection and isolation applied to f- mance metrics for fault detection and isolation filter and
16 flight dynamic,” Aircraft Engineering and aerospace controller interaction,” International Journal of Robust
technology, vol. 77, no. 2, pp. 152–160, 2005. and Nonlinear Control, vol. 23, no. 4, pp. 419–438, 2013.
[6] J. W. Melody, T. Hillbrand, T. Başar, and W. R. Perkins, [21] S. R. Gunn et al., “Support vector machines for classifi-
“H∞ parameter identification for inflight detection of air- cation and regression,” ISIS technical report, vol. 14, pp.
craft icing: The time-varying case,” Control Engineering 85–86, 1998.
Practice, vol. 9, no. 12, pp. 1327–1335, 2001. [22] S. Yin, X. Gao, H. R. Karimi, and X. Zhu, “Study on
[7] E. Chow and A. Willsky, “Analytical redundancy and support vector machine-based fault detection in tennessee
the design of robust failure detection systems,” IEEE eastman process,” in Abstract and Applied Analysis, vol.
Transactions on Automatic control, vol. 29, no. 7, pp. 2014. Hindawi Publishing Corporation, 2014.
603–614, 1984. [23] J. Platt et al., “Probabilistic outputs for support vec-
[8] H. Rotstein, R. Ingvalson, T. Keviczky, and G. J. Balas, tor machines and comparisons to regularized likelihood
“Fault-detection design for uninhabited aerial vehicles,” methods,” Advances in large margin classifiers, vol. 10,
Journal of guidance, control, and dynamics, vol. 29, no. 3, pp. 61–74, 1999.
no. 5, pp. 1051–1060, 2006.